2026.08.21.04 — /login in the chat, account switching, and a spotlight that reaches the plugins

First release in three with a SCHEMA CHANGE. anthropic_accounts is new, and without the push the account
switcher's route fails on a missing relation the first time the usage panel is opened — so the plan names
db:push as a step rather than leaving it to be inferred from the diff.

alpha already runs this code, deployed by hand as a preview. The plan says its pull will be a no-op and to
run the rest anyway, so it does not read as a skipped host.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

.gitignore

@@ -105,6 +105,13 @@ src/databases/officer_db/migrations/
 # to the platform; the plugins do not.
 /plugins/*/
 
+# Generated by src/servers/plugins/npm-deps.ts from every plugin's `npmDependencies`, and installed into
+# plugins/node_modules/. It describes whichever plugins are on THIS machine, so committing one would
+# claim a dependency set nobody chose. The rule above only ignores directories, so these need naming.
+/plugins/package.json
+/plugins/bun.lock
+/plugins/bun.lockb
+
 # Visual baselines for the theme extraction — large, machine-specific, regenerate with scripts/theme-snapshot.ts
 .snapshots/
 

.prettierignore

@@ -3,6 +3,3 @@ dist
 runtime-scripts
 *.min.js
 
-# Vendored third-party sources — reformatting them would destroy the diff against upstream,
-# which is the only thing that makes a checked-in library reviewable. See the README beside it.
-src/workspaces/officerdev/src/apps/Jellyfin/vendor/hls.mjs

BLOCKERS.md

@@ -0,0 +1,190 @@
+# Atelier extraction — blockers and decisions made without asking
+
+Unattended run, started 2026-08-20. Per the kickoff prompt: decisions made here without being able to
+ask, in the order found. Nobody is awake; each entry says what I did and why, then the run continues.
+
+## Map, before anything moved
+
+Grepped the whole platform for `atelier` (case-insensitive, every `.ts`/`.tsx`/`.md`/`.json`/`.cjs`
+file). Two hits, both incidental:
+
+- `src/workspaces/officerdev/src/apps/Chat/components/chat-images.test.ts:10` — a test fixture using
+  `https://atelier.officer.dev/images/....png` as an example markdown image URL. Not a dependency.
+- `src/servers/api/chat/task-restart.ts:181` — a comment mentioning a stray `bun dev` process that once
+  ran from `~/dockers/officer-atelier` and got respawned by accident. Historical, not a dependency.
+
+**No platform code imports anything from Atelier, and Atelier imports nothing from the platform.**
+Unlike music/offscale, there is no reverse dependency to find and no "this stays behind" piece. It comes
+out whole, the way offscale did before the imports were found — except here there really are none.
+
+Also checked: `ecosystem.config.cjs` (platform's own) has no `atelier` entry — the standalone service is
+managed by its OWN `ecosystem.config.cjs` inside `~/officer-dev/dockers/officer-atelier/`, entirely
+outside this repository.
+
+## The standalone service stays running, untouched
+
+`officer-atelier` (PM2, port 3002) is a **live production service** — its own `ecosystem.config.cjs` says
+plainly "this is the instance the mobile app talks to, so it holds still." It serves
+`https://atelier.officer.dev` via nginx and holds real data (87 generations, 2 collections, 5 providers,
+1 provider_settings row, checked directly against Postgres).
+
+I am not stopping it, not touching its nginx config, not touching its PM2 entry, and not touching the
+mobile app. This extraction adds a new, additive plugin inside the platform; cutting the mobile app over
+to it (or deciding whether to) is the owner's call, not mine to make unattended. Recorded here so nobody
+reads the plugin's existence as "the old service is now redundant, kill it" — it is not redundant until
+someone repoints the mobile app.
+
+## Decision 3 — database rows checked, both non-empty → database stays separate
+
+Queried the live `atelier` Postgres database directly (via a throwaway `postgres` client, same
+credentials as `.env.production.local`):
+
+```
+collections         2
+collection_items    73
+provider_settings    1
+providers            5
+generations          87
+```
+
+`collections` (2) and `provider_settings` (1) are the two tables the brief named — both non-empty. Per
+decision 3: **the `atelier` database stays separate**, not folded into `officer_db`. No
+`plugins/atelier/db/schema.ts` (that convention is for `officer_db`-prefixed tables); the sidecar keeps
+its own Postgres connection to the `atelier` database, exactly as today, just running from inside the
+plugin's `sidecar/` directory instead of a standalone repo.
+
+**The brief undersold the schema.** It says "tables: collections, provider_settings" — the real schema
+has five: `generations` (the actual data — 87 rows, prompts/images/costs), `collection_items` (the join
+table), `providers` (configured adapter+model+key rows — 5 of them; a newer, richer replacement for
+`provider_settings` that the brief doesn't mention at all), plus the two named. `generations` alone
+being non-empty would have settled decision 3 on its own. Following the letter of decision 3 (check the
+two named tables) rather than re-deriving it from the fuller schema, since both answers point the same
+way here.
+
+**Migration to `officer_db` is a real follow-up**, not done here: fold `collections`, `collection_items`,
+`providers`, `provider_settings`, `generations` into `officer_db` as `atelier_*`-prefixed tables once
+someone is awake to watch a real data migration. Until then the plugin's sidecar owns a second Postgres
+connection, to a second database, which is new territory for this plugin system (every other plugin's
+`db/` is officer_db-prefixed tables) — documented in PLUGIN.md.
+
+## Postgres connection for the separate database
+
+The sidecar needs its own connection string. Added `ATELIER_POSTGRES_URL` as a new platform `.env`
+variable (same pattern as `OFFICERDEV_POSTGRES_URL` for officer's own database) rather than routing it
+through the secret store — a database URL carrying a password is handled as a plain `.env` value
+everywhere else in this codebase (`OFFICERDEV_POSTGRES_URL` itself), and the secret store's own docs
+scope it to "every encryption and signing key the platform uses" plus foreign bearer credentials
+(APNs, FCM, webhooks) — not connection strings. Kept consistent with the existing convention rather than
+inventing a new one.
+
+**Not set automatically.** I don't have a safe way to guess whether the owner wants the plugin sidecar
+to point at the _existing_ `atelier` database (sharing state with the still-running standalone service —
+two writers to one queue table, which `generations.status` and the `claimableGenerations`/`markRunning`
+claim logic were written assuming is single-process) or a fresh one. Left unset, with a clear startup
+error, and documented in PLUGIN.md. **Do not just paste the standalone service's `POSTGRES_URL` in and
+restart both** — two processes racing the same `generations` queue is exactly the hazard `markRunning`'s
+atomic claim exists to prevent for restarts of _one_ process, not for two live ones.
+
+## Decision 2 — Providers as an Officer settings screen, against a direct precedent
+
+Found `IntegrationsSettings/index.tsx`'s own comment, written when email was extracted on 2026-08-18:
+
+> "Email account management left with the feature (`plugins/email/`). A plugin contributes panels, not
+> settings sections, so it now lives inside the email workspace itself."
+
+That is the opposite of decision 2, written five days before this run, in the exact settings screen
+decision 2 points Atelier's Providers into. Per the brief's own instructions ("if you find evidence one
+is wrong, write it down, follow the decision anyway"): implemented Providers as a genuine platform
+settings section — `IntegrationsSettings` gained an "Atelier" entry under the Enterprise tab — as new
+**platform** code (`src/apps/officer-web/...`) that talks to `/api/atelier/providers` over HTTP via
+`useClient`, same as any other authenticated screen. It does not import anything from
+`plugins/atelier/`, so it does not violate "a plugin importing platform code is fine; the reverse is
+not" — there is no import in either direction, only an HTTP call, which is what every frontend/backend
+pair in this codebase already is.
+
+Recording the contradiction here rather than silently picking the email precedent, since decision 2 was
+explicit and pre-made.
+
+## Images served through an authenticated route, not the old `PUBLIC_URL` scheme
+
+The standalone service served `/images/*` unauthenticated, off Bun's top-level route table, with
+`PUBLIC_URL` baked into every response so the URL was durable outside the process. Per "What you may NOT
+do", that is gone: images are served from `GET /api/atelier/files/:filename`, mounted like every other
+route, behind the permission gate. Named `files` rather than `images` so it cannot sit one path segment
+away from `/api/atelier/images/generations/:id` — the JSON job API already owns `/images`.
+
+The frontend needs `<img src>` to carry a credential a browser cannot put in an `Authorization` header.
+Used the platform's existing `?token=` escape hatch — already read by both
+`permission-gate.ts:37` and `user-middleware.ts:22`, already used by `FileViewer/file-types.ts` for
+exactly this (`getRawUrl`, `getAudioTracksUrl`, ...) for the file browser's raw/audio/subtitle routes.
+Not a new mechanism, not an unauthenticated route: the same query-token path every other authenticated
+media URL in this app already takes.
+
+## `sharp` added as a direct dependency
+
+Atelier's `inspect.ts` uses `sharp` to measure width/height/alpha-channel/opacity of a written image —
+load-bearing (it is how a `background: transparent` request that silently produced an opaque image gets
+caught). `sharp@0.33.5` was already resolved in `bun.lock` as an optional transitive dependency of
+`next`, so promoting it to a direct dependency of the platform added no new resolution — verified the
+lockfile diff is exactly that promotion, nothing else moved. Done via the documented dance:
+`frozenLockfile` commented out in `bunfig.toml` for one `bun install`, restored immediately after,
+diff reviewed before committing both files.
+
+## Publishing to the Marketplace — not done, left as the one open step
+
+`plugins/atelier/` is its own local git repository (`git init`, per "Done means"), but I did not create
+`gitea.officer.dev/plugins/atelier` or push to it. I found no Gitea API token or credential available to
+this session, and creating a new repository on a shared Gitea instance and registering it in the
+marketplace catalogue (`scripts/marketplace-sync.ts`, `BUNDLED` in `src/servers/plugins/marketplace.ts`)
+is exactly the kind of action with effects beyond this local checkout that the run's own ground rules
+say to flag rather than take unilaterally, especially with no verified way to confirm it succeeded.
+
+Left for the owner:
+
+```
+cd plugins/atelier
+git remote add origin git@gitea.officer.dev:plugins/atelier.git   # after creating the repo on Gitea
+git push -u origin main
+cd ../..
+bun --env-file=.env scripts/marketplace-sync.ts
+```
+
+Then add a `BUNDLED` entry in `src/servers/plugins/marketplace.ts` (offline fallback), matching the
+other three plugins' entries.
+
+Until that happens, the plugin still installs fine from a checkout that already has the directory
+(`plugins/atelier/` committed to its own local repo, present on disk) — it just cannot be discovered by
+the Marketplace screen on a machine that has never had it, per "Done means" #2. This is the one piece of
+"done" I could not complete unattended.
+
+## No `plugins/atelier/db/`, on purpose
+
+"Done means" lists `db/` among the plugin's expected pieces, matching offscale's and music's shape. This
+plugin has no `db/schema.ts` at all — `discoverPlugins()` correctly reports `schema: null` for it. That
+convention is specifically for `officer_db`-prefixed tables pushed by `pushPluginSchema`; decision 3
+keeps Atelier's tables in their own separate Postgres database instead, so there is nothing for that
+convention to describe. The database code lives inside `sidecar/db/` as an internal implementation
+detail of the sidecar process, not as a platform-recognised `db/` directory. Documented in full in
+`plugins/atelier/PLUGIN.md`, under "The database".
+
+## What I verified, and what I deliberately did not
+
+- `bunx tsgo`: clean throughout.
+- `bun test`: baseline 1041 pass / 0 fail / 90 files → after this run, 1069 pass / 0 fail / 92 files.
+  The two new files are the ported `serialize.test.ts` and `generations/response.test.ts`. No
+  regressions, file count only grew.
+- `discoverPlugins()` finds the plugin and the manifest validates with no problems.
+- The platform-side proxy router mounts and correctly 503s with no sidecar running.
+- The sidecar's own Hono app was exercised directly against the real `atelier_dev` Postgres database
+  (not `officer_db`, not the production `atelier` database) with real HTTP requests and no port bound:
+  providers, providers/adapters, collections and images all answered correctly; a missing file 404s; an
+  unknown route gets the OpenAI-shaped error envelope; provider rows whose keys were encrypted under the
+  standalone service's own key correctly fail to decrypt rather than crashing (a fresh `secret.key` was
+  generated for this checkout at `$OFFICER_ROOT/data/atelier/secret.key`, which is expected and correct
+  — see PLUGIN.md).
+- **Deliberately not done**: installing the plugin through the live, running `officer` PM2 process. That
+  process serves the owner's real session, its `plugins/` directory is the main checkout (not this
+  worktree, so it would not even find this code), and going through its install flow would write a real
+  `plugin_installs` row into the shared production `officer` Postgres database from a session with no way
+  to confirm the result against a live browser. Full PM2 install/enable/disable/uninstall/purge cycle,
+  per the runbook's verification table, is left for the owner — noted as open item 5 in PLUGIN.md.

CLAUDE.md

@@ -349,9 +349,11 @@ Worked examples: `/soulseek`, `/music`, `/chat`.
   (`{ key, name, icon, component, availableOnPanel: false }`) and registered in `AppRegistry.tsx`.
   The Workspace/Panel framework is orthogonal to routing — it contains no route navigation, and panels
   live inside the Route element tree, so they can call `useParams`/`useSearchParams` directly.
-- **Page title by route.** Add a rule to `RULES` in `src/apps/officer-web/state/usePageTitle.ts`
-  (`{ match: (p) => p.startsWith('/<name>'), title: '<Name>' }`, most-specific first);
-  `usePageTitleSync` does the rest. `startsWith` means nested routes are covered.
+- **Page title by route.** Automatic — no rule to add. `titleForPath` in
+  `src/apps/officer-web/state/usePageTitle.ts` takes the URL's first path segment, turns `kebab-case`
+  into `Title Case`, and that's the tab title; `usePageTitleSync` does the rest. A screen that wants to
+  say more than its route does (a chat's own name, not just "Chat") publishes an override instead —
+  `usePublishPageTitle` in `officerdev`.
 
 ### The URL is the source of truth for selection
 
@@ -419,6 +421,15 @@ explaining why it was safe.
 
 - `docs/navigation-audit.md` — **authoritative** on routing/navigation: the opaque-click anti-pattern,
   a severity-ranked findings table, the channel-selection map and the four-phase plan
+- `docs/email-recall.md` — **design, not built**: recall over your own correspondence. The split that
+  matters is counting (needs every message judged) versus retrieval (needs the best few, fast) — and why
+  a number produced by searching is confidently wrong with no signal that it is
+- `docs/find-my-device.md` — **design, not built**: locating your own devices from Officer. What is
+  assembly (most of it), what is impossible (offline finding — there is no crowd to relay it), and the
+  question to settle first, which is retention rather than anything technical
+- `docs/affordances.md` — what an affordance is, the three tests for whether something qualifies (including
+  the one that catches most bad ones: _can it be fixed instead?_), the file format, and how they reach the
+  agent. Read before writing one
 - `docs/agent-coordination.md` — **the north star** for the workspace/panel work: agents on one
   dashboard coordinating with each other instead of through the human, the handoff protocol, and what
   is deliberately _not_ being built. Read it before ranking, deferring or starting any panel item —
@@ -427,6 +438,21 @@ explaining why it was safe.
   mounted and what it knows, the URL-vs-channel split for panel-to-panel communication, and the
   persistence key families. Read before building a panel app. Its defect list is
   `docs/workspace-panel-todo.md`.
+- `docs/multi-server-push.md` — **design, not built**: one Officer app receiving notifications from every
+  server it is signed into. Works already, bar one missing field — and why the `mv` catalogue-version
+  check turns into an endless resync between servers without it. The server identifier is an SSO decision
+- `docs/visual-regression-plan.md` — **plan, not built**: nightly screenshot regression against a JWT
+  dashboard on a second server. The blocker is that a JWT lives in `localStorage` and iris copies only
+  cookies, so an authenticated capture silently renders the login page — plus why diffing must happen
+  before any agent is asked to look
+- `docs/plugin-dependency-scope.md` — a plugin declares WHO must be able to run each host binary. Why the
+  owner's login shell is the wrong question for a `confined` plugin, the two-part reachability check
+  (world-executable AND every parent traversable), and why members' binaries on alpha worked only because
+  one home directory happened to be mode 751
+- `docs/plugin-route-authorization.md` — **design, not built**: making a plugin's settings routes
+  owner-only while the plugin itself stays grantable. What the proxy should pass (`isOwner` and a level,
+  never the role name), why declaring owner-only paths in the manifest beats trusting every plugin to
+  remember, and the credential-resolution work that has to land first
 - `docs/secret-store.md` — **design, not built**: moving the encryption and signing keys out of `.env`
   into a SQLite store, why they cannot live in Postgres, and key rotation. Also records the core/plugin
   split it assumes — light plus `officer-headscale` is the core; Vaultwarden and the wallet are plugins

COMMS/command-hq/platform.md

@@ -161,3 +161,57 @@ picked the right two columns for now.
 
 Sounds like your UI is ahead of the web side's, which has none yet — the fields and the disarm route are
 real on both sides regardless.
+
+## Reply to `mobile-4.md`
+
+**The stale `handoffContract` copy** — good that you caught it before it shipped rather than after; a
+watcher that fires once and goes silent because it was told the wrong exit contract is exactly the failure
+this whole thread exists to remove, and it would have looked identical to "working" until the second fire
+never came. Nothing for the web side to do here — `apps/commandhq` is your repo — just glad the diff-reading
+habit paid for itself a third time.
+
+**The evidence correction — carried into `docs/command-hq.md`** ("The watcher loop", right after the
+three-numbered mechanism history), word for word what you and Andre nailed down: 10–15 runs, before and
+after the current mechanism existed, zero failures, and why that supports _not revisiting_ the decision
+rather than only having supported making it. Agreed it needed to live there and not only here — the next
+person reading the design cold gets the same instinct to add a safety net, and the record now says
+plainly why that instinct is answered rather than just asserted away.
+
+**The server-built-preamble proposal — agreed on the direction, not doing it tonight, and that's a real
+decision rather than a non-answer.** Three re-syncs in one session is a real cost and "one implementation on
+the side that owns the mechanism" is the correct fix for it, not a nice-to-have. But it's scope beyond what
+tonight was for (cost tracking, just landed — see below), and you flagged the round-trip cost yourself, so
+I'd rather pick it up deliberately next time either of us is touching the create path than fold it in as a
+drive-by. Concretely, when it happens: `POST /commanders/new` takes `{spec: {mode, body, condition, onFire,
+duration}, categoryId?}`, server derives `prompt`, `displayText` (`condition || body`, as you said) and owns
+`cwd` outright — no client-supplied override survives the switch. Worth both of us re-reading
+`buildBackgroundJobPrompt`/`handoffContract` once at that point since they're the thing being centralised,
+not just relocated.
+
+## Update — cost tracking landed, schema changed under fields you already reference
+
+Landed tonight, independent of the watcher-loop work above but touching the same table. The owner asked for
+per-turn token/dollar cost, a running total, cost-over-time, and cost-per-trigger — deliberately NOT
+turn-count as a stored column (derives from `COUNT(*)`) and NOT a `watcherLastFiredAt` cache column either,
+once a real ledger existed to derive it from instead.
+
+**What changed under you:** `command_hq_sessions.watcherLastFiredAt` is gone as a column. Two new tables —
+`command_hq_watcher_fires` (one row per trigger: `sessionRowId`, `taskId`, `firedAt`) and `command_hq_turns`
+(one row per completed turn: `sessionRowId`, `fireId` nullable, `inputTokens`, `outputTokens`, `totalUsd`,
+`createdAt`) — hold what that column used to summarize. `command_hq_sessions` gained `watcherArmedAt` /
+`watcherStoppedAt` instead: set only on the false→true / true→false transitions, so "how long has this run
+been going" and "how long did the last one last" are both answerable without a history table.
+
+**Nothing in your response shape moved.** `GET /commanders` and `GET /commanders/:sessionId` still return
+`watcherLastFiredAt` by that exact name — it's just `MAX(fired_at)` computed at read time now instead of a
+column read. Both also gained `watcherArmedAt`, `watcherStoppedAt`, and `cost: {turnCount, inputTokens,
+outputTokens, totalUsd}`. Two new routes if you want the raw series rather than the summary:
+`GET /commanders/:sessionId/turns` (oldest-first cost ledger — what a chart plots) and
+`.../fires` (oldest-first trigger history). Full detail: `docs/command-hq.md` → "Data model" and "The
+watcher loop".
+
+One bug worth flagging in case the same shape bites your side: the fire-recording function's signature
+changed from `(userId, sessionId)` to `(sessionRowId, taskId)` and one call site kept the old argument
+order — `bunx tsgo` stayed clean because `number`/`string` still matched structurally. Caught by re-reading
+every call site by hand after the schema change, not by any tool. Recorded in the doc as a general
+caution, not just a fixed bug.

TODO.md

@@ -1,574 +1,74 @@
-# TODO
+# TODO — the index
 
-Deferred work.
+Deferred work. **The items live in `todo/`, one file each**, grouped by theme; this file is the map
+over them and is GENERATED — run `bun run todo:index` after any change rather than editing it here.
+Format and the two-way-sync contract: `todo/README.md`.
 
-**Context, corrected 2026-08-07.** This file used to open by saying Officer was "collapsing from
+**Context, corrected 2026-08-07.** This list used to open by saying Officer was "collapsing from
 multi-tenant / open-source-ready to a **single-user platform**", and told you to treat multi-tenant
-indirection as accidental complexity. **That direction was reversed.** The permission permission model
-shipped on 2026-08-07 to serve a real goal — deploy to the company server, onboard people, give each
-one their own Gitea account through the platform. Per-user scoping is now a requirement, and the items
-below that proposed deleting it have been removed rather than left to mislead the next reader.
+indirection as accidental complexity. **That direction was reversed.** The permission model shipped on
+2026-08-07 to serve a real goal — deploy to the company server, onboard people, give each one their own
+Gitea account through the platform. Per-user scoping is now a requirement, and the items that proposed
+deleting it have been removed rather than left to mislead the next reader.
 
 What did NOT reverse: `execution` permissions (terminal, chat, tasks, files, desktop, browser) run as
 the owner's OS user and can never be granted. Indirection there really is accidental complexity.
 
-## Multi-user
-
-- [ ] **`deprovisionOsAccount` is written but has never run against a real account.** Landed 2026-08-12 in
-      `os-user-deprovision.ts` and wired into `deleteUserHandler`, which now refuses to delete the row when
-      the Linux teardown fails — so a failure is retryable instead of forgotten. Only the pure guards
-      (`guardDeletable`, `guardMemberTree`, `parseSubUidEntry`) have tests; the reap loop, the `chown -R`
-      sever and `userdel` have been exercised by nobody. `docs/deprovision-os-account.md` → "What is still
-      unproven" has the five-step validation, and it has to happen on the production host with a throwaway
-      account that has **a shell left open** and **a container writing as a non-root user** — those are the
-      two cases the quiet path passes vacuously.
-
-- [ ] **The terminal replays terminal QUERIES, which get typed into the shell.** `sidecar/pty/sessions.mjs`
-      replays the whole scrollback on attach; query sequences in the buffer get re-asked, xterm.js answers,
-      and the answers arrive as keystrokes. Visible to a member daily. Fix is to strip query sequences in
-      `appendBuffer`, so a replay reproduces output and never re-issues requests.
-
-- [ ] **The web terminal renders a long URL as unreadable fragments.** Claude Code's first-run login prints
-      a ~400-character OAuth URL; the web terminal shows scattered characters with large gaps, nothing
-      selectable. Half worked around by `2a8f004` (OSC 52, so "press c to copy" reaches the clipboard) — the
-      rendering itself is undiagnosed. This is every new member's first five minutes.
-      `docs/open-threads-after-per-user-claude.md` §1 has what is known and where to start.
-
-- [ ] **Agent sessions are not durable, and it is one property behind three symptoms.** A sidecar restart
-      loses session identity, which is why `endTurnIfAgentIsGone` must skip sessions with no recorded
-      `userId`, why a stuck "generating" spinner survives until a reconnect, and why any crash in that process
-      is destructive rather than merely inconvenient. Fixing the three separately would miss that they are one
-      missing property. `docs/open-threads-after-per-user-claude.md` §2.
-
-- [ ] **`ProcessTransport is not ready for writing` — survivable since `8c4f150`, still unexplained.** A
-      floating rejection inside the SDK's own input pump, with no frames from our code, so no `await` of ours
-      can catch it. It crashed `officer-agent` four times on 2026-08-11, once truncating a turn mid-sentence;
-      the `unhandledRejection` backstop has caught it once since. Best hypothesis is the `claude` CLI exiting
-      while `streamInput` is still pumping. It needs looking at after the next occurrence, not catching in the
-      act — markers to grep in `docs/open-threads-after-per-user-claude.md` §3.
-
-- [x] **No way to create a second account.** Fixed 2026-08-11 on `sidecar-app-store`: `POST /api/users`
-      (`api/users/create-user.ts`, owner-gated) plus an Add-account form in
-      Settings → User management. Created accounts are `status: 'Active'` — the column defaults to
-      `'Unverified'` and `signin.ts` refuses anything else with a bare UNAUTHORIZED, which is the trap
-      the hand-INSERT route fell into. The owner sets the password and reads it out; `passwordChangedAt`
-      stays null. Directories come from the shared `provisionUserDirs`/`USER_DIRS` in `data-path.ts`,
-      which `scripts/provision-user-dirs.ts` now imports rather than restating.
-
-- [ ] **Still no invite flow, and no password reset for a member.** The owner types the password and
-      tells the person, which means the owner knows it and the member cannot change it back if they
-      forget theirs — recovery today is delete-and-recreate. An invite (token, expiry, member sets
-      their own) needs a mail path. This is the next piece, not a nice-to-have.
-
-- [x] **A second Super Admin was storable, and made the owner nondeterministic.** Fixed 2026-08-11.
-      `ck_users_owner_is_super_admin` pins user 1's role but a row-level CHECK cannot see other rows, and
-      `updateUserRoleHandler` happily promoted anyone — while `getOwnerUser()` was
-      `WHERE role='Super Admin' LIMIT 1` with no ORDER BY. Two holders would have made "who owns this
-      server" a question the query plan answered, and that answer feeds the agent sidecar's identity,
-      vault access and origin scoping. Both write paths now refuse the role, the list endpoint offers
-      `assignableRoles` without it, and `getOwnerUser()` orders by id.
-
-- [ ] **`dashboards.id` is a global primary key, and ids are `slugify(name)`.** Two accounts cannot
-      both have a dashboard named "Home". Reachable today: six accounts exist. The recommendation on
-      the table is a composite PK `(user_id, id)` — it matches the `uq_dashboards_user_id` index
-      already there and keeps every stored `ws-layout-<id>` address valid, which uuid ids would not.
-      Note the drizzle composite-PK re-diff quirk in `databases/CLAUDE.md`. Full analysis in
-      `docs/workspace-panel-todo.md` §3.
-
-- [ ] **`permissions/authorize.ts` has no automated tests.** `registry.test.ts` covers the pure
-      registry functions and the totality check; the resolver that does the owner bypass, the grant
-      lookup, the role cache and the fail-closed catches is exercised only by hand. It is the file
-      standing between a Member and a shell.
-
-- [ ] **`assertPermissionTotality` checks the wrong list, and `registry.test.ts` has been red since
-      2026-08-13.** It is fed `Object.keys(handlers)` from `server.tsx`, but Bun serves the _route table_.
-      Those diverged when the cliamp/desktop/vault plugins were switched off: `/api/cliamp/ws` and
-      `/api/cliamp/audio/ws` are still live routes with their handlers and registry claims commented out.
-      Not exploitable — `isWsProviderAllowed` finds no permission and 403s a member; the owner upgrades onto
-      a dead socket. But the boot check that exists to stop exactly this cannot see it. Two fixes: point
-      totality at the route table, and either delete the dead routes or restore their claims. The 8 failing
-      tests in `registry.test.ts` are the same drift — `REAL_WS` still lists all nine providers as served,
-      which is why nobody noticed. Found 2026-08-14.
-
-- [ ] **No empty state for a denied screen.** A member who reaches a route their role lacks gets a
-      broken panel or an endless spinner rather than a clean refusal.
-
-- [ ] **`getOwnerHomeDir(email)` ignores its argument** whenever `HOME_DIR` is set, which it is here —
-      every caller resolves to the owner's real login home. Safe only because all seven callers sit
-      behind `execution` permissions. If per-user home confinement is ever attempted, this is the
-      function to start from.
-
-- [ ] **`pty`, `vault` and `opencode` receive no identity at all.** Every other sidecar validates
-      `X-Officer-User`. The pty sidecar keys purely on a `sessionId` from the query string and its
-      `/_officer/sessions` endpoints list and kill _every_ session on the box; vault and opencode take
-      no user argument. All three are covered today only because `terminal`, `vault` and the agent are
-      owner-only permissions — that is a correct outcome resting on the wrong layer, and it is the
-      thing to fix first if any of them is ever granted.
-
-- [ ] **Radicale is configured `type = owner_only`** (`sidecar/caldav/radicale.ts:54`) while the caldav
-      sidecar itself is fully per-user and confines every JSON read to `/dav/<userId>/`. The platform
-      side is ready for members; the CalDAV server underneath is not.
-
-- [ ] **The music library is one global index.** `sidecar/music/indexer.ts` reads `HOME_DIR` and serves
-      every account from it. Favourites, playlists and now-playing _are_ per-user. Deliberate for now
-      (one household, one library) but worth stating rather than discovering.
-
-- [ ] **`markInterruptedJobs()` and `getOldestPendingJob()` are platform-wide.** The pipeline queue is a
-      single global lane; ownership is enforced one layer up, in `pipeline-jobs-routes.ts`, by an
-      explicit `job.userId !== user.id → 404` on every by-id route. Correct today, but the queue itself
-      has no notion of whose work it is running.
-
-- [x] **Cross-user writes in the notify sidecar** (fixed 2026-08-07, this session).
-      `DELETE /_officer/devices/:token` deleted by token with no user predicate, so any account with the
-      `notify` permission could deregister another's device; and `POST /_officer/notify` let a request
-      body's `userId` override the proxy-injected `X-Officer-User`, so the same account could push to
-      another's devices. `deletePushDevice` now takes an optional `userId` (the route passes it, the
-      APNs/FCM dead-token paths deliberately do not) and the header now wins over the body.
-
-- [ ] **Remove dead `username` plumbing.** `send-claude-code.ts` declares `username` in two types
-      without using it. (`toShellUsername` is NOT dead — `server.tsx:190` and
-      `pipeline-job-manager.ts:267` both call it. The `provision.ts` caller this item used to name no
-      longer exists.)
-
-## Email
-
-Email became a plugin on 2026-08-18 (`plugins/email/`, `gitea.officer.dev/plugins/email`). Its work queue
-went with it — the open items pointed at `sidecar/email/store.ts`, a path in another repository, which is
-exactly the way a leftover list sends someone to edit files that are not there.
-
-Anything still owed to the mail client belongs in that repository's own TODO.
-
-## Transmission
-
-Phase 1 (parity with `_references/transmission-web`) shipped 2026-07-30: the `officer-transmission`
-sidecar plus `/transmission` (torrents / stats / settings). Everything below is **phase 2** — none of
-it exists in the reference app, so none of it was in scope for parity.
-
-### Probably needed regardless
-
-- [ ] **Resizable detail pane.** `TorrentsView` pins it at `DETAIL_HEIGHT = '45%'`, which is a guess.
-      Either a drag handle or a persisted height in `useLocalStorageState`, alongside the column set.
-- [ ] **Revisit `DEFAULT_VISIBLE_COLUMNS`.** Eleven of thirty, chosen before ever seeing the table
-      rendered against the real 43-torrent library. Likely wrong in both directions.
-- [ ] **Files tab on huge torrents.** `detail/FilesTab.tsx` builds the whole tree and renders every
-      node — no virtualisation. Fine at 14 files; unknown at a few thousand. If it stalls, the fix is
-      `useVirtualizer` over a flattened visible-node list, the same shape as `TorrentTable`.
-
-### Ideas, unprioritised
-
-- [ ] **Container→host path mapping.** The daemon reports its own namespace (`/downloads/complete`),
-      which is not a path on this host. A mapping table (`/downloads` → `~/hdds/08_TB_01/Torrents`,
-      `/downloadsTV` → `14_TB_02/Torrents`) would make locations clickable through to `/files` and
-      make "Set location" offer real destinations. Deliberately dropped from phase 1: the reference
-      app has no such feature, so it would have been config nothing read.
-- [ ] **Push instead of poll.** Transmission RPC has no push channel, so `useTransmissionData` polls
-      every 5s. The sidecar could poll once and fan out over a WebSocket, which is both cheaper and
-      what every other live surface in the platform already does.
-- [ ] **Cross-surface links.** Soulseek and `download-media` both land files in the same library;
-      Transmission is the third door to it. Worth a think about whether they should know about each
-      other at all.
-
-## Headscale
-
-Gaps against headscale's own API, found while reading the app on 2026-08-05 — all closed the same
-day. Kept here for what the policy work turned up, which is not obvious from the code.
-
-- [x] **ACL policy.** `/api/v1/policy` GET/PUT behind `/_officer/policy`, with a plain HuJSON textarea
-      that sends the text byte for byte and shows headscale's verdict verbatim (line and column
-      included). Officer does not pre-validate: headscale owns the only parser that resolves groups,
-      tags and hosts, and a second weaker one would disagree with the thing that actually enforces.
-      **The mode cannot be read.** A file-backed policy is still served over GET; only a PUT refuses,
-      with "update is disabled for modes other than 'database'". So the first save is what discovers
-      writability, and a refusal becomes a persistent read-only banner. Verified live, not inferred.
-
-- [x] **A node can be moved between users.** `/api/v1/node/{id}/user`, in the expanded node card next
-      to the tag editor — both being the things that decide which policy rules apply to a node.
-
-- [x] **User rename** — was already shipped end to end (`/users/:id/rename`, `UsersView`); this entry
-      was stale when it was written.
-
-- [x] **Polling** — `useHeadscaleNodes` has had `refetchInterval: 20_000` all along; also stale.
-
-- [x] **Device invites, platform side** (`COMMS/OFFSCALE_INVITE_ENROLLMENT.md` §5, built 2026-08-05).
-      Authorize-new-device form, the one-time link with copy/share/QR, and the invite list with revoke —
-      `/_officer/enroll/invites` in `invites.ts`, `InvitesView`. **The records are not Officer's.** They
-      proxy to the server's Officer Companion, because the joining phone has to claim without an Officer
-      account and this sidecar is loopback-only behind our auth; the spec's own "an invite must work when
-      the platform is down" argument says the same. Officer stores no invite and no claim token.
-      **Live on all four servers since 2026-08-05.** The 401 that blocked it was a companion-side prefix
-      parse (a Headscale key prefix is a fixed 12 chars and may contain `-`; they split on the first one);
-      fixed upstream. Verified end to end against pastilhas-eu: create returns the
-      `officer-offscale://join#…` link, list shows the record, revoke flips it to `revoked`. The list
-      envelope is not in the spec, so `pickInvites` takes the body's array whatever it is keyed under.
-      Four gaps were sent back to the spec author: `keys.ts` has no
-      sub-day key TTL for the 5-minute claim key, "the sidecar must refuse plaintext" is unenforceable
-      behind nginx, `tailnet` is not a headscale concept and has to be recorded on the invite, and
-      `/api/v1/enroll/*` collides with headscale's own namespace.
-
-## Workspaces & Panels
-
-Its own list — **`docs/workspace-panel-todo.md`** — because it is long and actively worked. Full
-analysis behind it: `COMMS/workspace-panel-framework-analysis-2026-08-07.md`.
-
-Headlines: the framework has **no panel lifecycle** (a panel can be destroyed but never told, which is
-the real cause of the terminal orphan leak); three state-key families are **silently dropped on every
-write**, orphaning a pty per reload; Running Shells has **404'd since 2026-07-31**; there are **zero
-error boundaries** anywhere in the repo; and `dashboards.id` is a **global** primary key fed by
-`slugify(name)`, so two members naming a dashboard the same thing collide.
-
-- [ ] **Renaming a panel key is a silent data migration, and the repair is destructive.** A saved layout
-      holds `appType` strings. `normalizeLayout` replaces any appType not in the allow-list with the
-      caller's `fallback`, and `WorkspaceView.tsx:97` writes that back — so a renamed key turns every
-      saved layout into N copies of whichever panel is `allowed[0]`, permanently, on one page load.
-      Renaming `headscale-*` → `offscale-*` did exactly this: `/offscale` rendered three server pickers,
-      `ServersView` never mounted, and it survived every restart because the code was correct and the
-      data was not (2026-08-17, `ed3806f9`).
-
-      Plugin screens are now cleared at install, which covers plugins and nothing else. The general fix
-      is a rule about wholesale mismatch: **a layout in which NO panel has a recognised appType is
-      residue from older code — discard it and use the screen's default**, rather than collapsing it.
-      `fallback: allowed[0]` is arbitrary (it is whatever `panels.ts` happens to list first) and should
-      probably not be a per-panel substitution at all.
-
-## Code editor
-
-- [ ] **The owner wants to go deeper here — ideas pending (noted 2026-08-15).** Raised right after the file
-      browser gained its first links into `/code-editor`. Do this after the file-browser test pass.
-
-      What is worth knowing before starting, all established on 2026-08-15:
-
-      **It has been effectively invisible for six months.** The app landed 2026-02-21 (`9e9acd96`), the
-      screen 2026-02-17, and `1fa3a659` on 2026-07-25 is titled "**restore** the code editor screen" — so
-      it was dropped and brought back at least once. It IS in `CORE_DOCK_ITEMS` as "Editor", but
-      `DEFAULT_DOCK_PATHS` is `['/', '/files', '/terminal', '/dashboards', '/chat']` and the owner's own
-      `dock_configs` row does not list it either. Until today the only way in was typing the URL.
-
-      **So it has no users, and that is the risk.** A surface nobody reached for six months is where things
-      rot quietly. Nothing about the editor itself has been verified — whether saving works, how the file
-      tree behaves, what closing the last tab does, what an unsaved-changes navigation does. Only the
-      plumbing was checked: `?file=` is read (`EDITOR_FILE_PARAM`), `/code-editor` renders with `urlState`,
-      and `CodeEditor.tsx:60-76` fetches a path that arrives in the URL rather than only matching already
-      open tabs.
-
-      **One bug of exactly this kind was already found and fixed** in the FileViewer's editor on the same
-      day: the Edit toggle was gated on `content` rather than `content !== null`, so an empty file — the
-      one "New file" produces — could not be edited at all. Assume siblings.
-
-      **Open question, deliberately not decided:** `Open in editor` in the file browser's row menu still
-      navigates to `/code-editor`, which is the last action that leaves the browser. Editing now happens
-      in place via double-click → the viewer's pane. Either `/code-editor` earns its keep as a genuinely
-      different tool (tree, tabs, multi-file) or that menu item should go.
-
-## Files app — routing
-
-**The URL model is done (2026-08-15).** The folder is the pathname —
-`/files/archive/Tests/platform/docs` — and `?view=` is a NAME within it, not a second copy of the path.
-Deep links and refresh work, which they never had. Landed across `559560de`, `98ba604a`, `7e31564c`,
-`0dc88b51`, `7359af2f`, with `files-route.ts` + 23 tests as the one owner of encoding and of which params
-belong to the overlay.
-
-Four bugs came out of that work and are fixed: closing a pane sent you home; `?view=` was stripped on
-every page load by a mount effect that had made deep links impossible since 2026-02-23; navigating kept a
-pane belonging to the folder you left; and the breadcrumb opted out of that last rule because it is a
-`<Link>` and never called the function enforcing it.
-
-- [ ] **Rows are still not links.** This is the half that did not get done, and it is the original
-      finding. Both list and grid render `<div onDoubleClick>` (`FileItem.tsx:701,748`), and search
-      results render `<div onClick>` (`FileViewContainer.tsx:187`) — the "opaque click" that
-      `docs/navigation-audit.md` names. No cmd-click into a new tab, no middle-click, not
-      link-focusable, and the target lives in a closure rather than the DOM.
-
-      It is now much easier than it was: `folderHref(path)` gives a folder's URL and `hrefForPath` gives
-      a full `To` with the overlay already stripped, so a row becomes
-      `<Link to={hrefForPath(childPath)}>` and a file row becomes a link that sets `?view=<name>`. The
-      breadcrumb has done exactly this since before today and is the model.
-
-      Watch for: a row is also a click target for SELECTION (single click selects, double opens,
-      shift/cmd extend). An anchor changes what those gestures mean by default, so the selection
-      handlers have to keep working and `preventDefault` where they win.
-
-- [ ] **The acts that clear the overlay have no owner, only the list does.** `VIEWER_PARAMS` and
-      `withoutViewerParams` live in `files-route.ts` and are shared. But every site still decides FOR
-      ITSELF whether to call them, which is precisely how the breadcrumb shipped wrong an hour after the
-      rule was written. Three bugs in a row came from this. A single `navigateToFolder()` that every
-      caller must go through — rather than a helper they may remember — is the fix if a fourth appears.
-
-- [ ] **A panel's folder is still not addressable.** `urlPath` is true only on the `/files` screen;
-      panels keep the path in local state, so they always open at their base and cannot be linked or
-      restored. Deliberate as written — a dashboard can hold two browsers and one URL cannot serve both —
-      but worth re-deciding rather than inheriting.
-
-## Plugin ecosystem — before it opens to third parties (2026-08-18)
-
-Not now. Months away. Recorded so it is not re-derived.
-
-- [ ] **A plugin install is not pinned, so a review would bind a repository rather than code.**
-      `fetch.ts:70` is `git clone --depth 1 <source>` — no ref — so it takes whatever the default branch
-      points at when the install runs. A plugin reviewed at commit X installs at whatever was pushed
-      since, and `--depth 1` leaves no history to compare against.
-
-      This is the shape of the npm attack already documented in `CLAUDE.md` (2026-08-04), and the same
-      shape as the Chrome Web Store's real failures — not bad extensions passing review, but good ones
-      sold or compromised and then pushing an update to users who had already accepted them.
-
-      Fix: a `reviewedCommit` on the catalogue entry, checkout detached at that SHA, and **verify
-      `rev-parse HEAD` matches or refuse**. The verification is the part that matters. An update then
-      becomes "a new reviewed SHA exists" — a catalogue change rather than something the author triggers.
-
-- [ ] **Third-party plugin sidecars are unconfined.** A plugin sidecar runs `bun run <script>` with
-      `cwd: PLATFORM_DIR`, so Bun auto-loads `.env` and it gets `POSTGRES_URL`; it runs as the service
-      user, which owns `secrets/officer-keys.db` and can therefore mint an owner JWT. Fine for
-      first-party plugins, total compromise for anyone else's.
-
-      The mechanism to fix it already exists and is pointed elsewhere: `spawn-as-member.ts` drops
-      privileges with `setpriv --reset-env` and a six-name allowlist for exactly this reason.
-
-- [ ] **No install-from-git-URL.** The catalogue is the only route for a remote plugin, so the
-      "build your own, give it to a friend" path has no affordance. Small, and needed before the
-      unofficial channel is real.
-
-## Plugin system — gaps the email extraction found (2026-08-18)
-
-- [ ] **A plugin cannot declare its own npm dependencies, so the PLATFORM is carrying them.** `imapflow`,
-      `mailparser` and `@types/mailparser` are in the platform's `package.json` and are imported by nothing
-      in `src/` — they exist solely for `plugins/email/sidecar/`. Plugins have no `package.json`, run with
-      `cwd: PLATFORM_DIR`, and resolve from the platform's `node_modules`, so deleting those three lines
-      breaks the installed plugin on its first sync.
-
-      This does not scale and it is the wrong party holding the risk: the platform cannot be answerable for
-      what an arbitrary plugin decides to install, and `bunfig.toml` sets `frozenLockfile = true` precisely
-      so a dependency change is a deliberate, reviewed act. Publishing will be reviewed rather than a
-      free-for-all, which narrows the threat but does not move the responsibility.
-
-      What it probably needs: a `dependencies` field in the manifest (the manifest already holds
-      `osDependencies` for host binaries, which is the same shape of problem solved once), installed into
-      the plugin's own directory at install time rather than merged into the platform's tree. Note the
-      install path already learned not to trust `$PATH` — `pushPluginSchema` resolves a binary rather than
-      spawning `bunx` — so whatever runs the install has to do the same.
-
-      Until then: `imapflow` and `mailparser` STAY. Removing them is not tidying, it is breaking email.
-
-- [ ] **`email_accounts/` is created by the PLATFORM for every account, and should be the plugin's job.**
-      `USER_DIRS` in `data-path.ts` pre-creates it at 0700 for every user, whether or not the email plugin
-      is installed — so a core install provisions a directory named for a plugin it does not have.
-
-      Ownership should invert: the plugin provisions its own storage, in all three moments —
-
-      1. **on plugin install**, for accounts that already exist;
-      2. **on user provisioning**, for accounts created afterwards;
-      3. **lazily**, the first time a provisioned user opens the plugin, so a machine that missed either of
-         the above still heals itself.
-
-      The third is the one that makes the other two non-critical, and it belongs in the email sidecar
-      rather than in the platform's account handler.
-
-      This is entangled with the `email_db` leak below — core reads that directory today, so the plugin
-      cannot become its sole owner until that is resolved too. Do them together.
-
-- [ ] **The platform still knows email's on-disk vocabulary, deliberately, and it is untested by anything.**
-      `findEmailAccountName` in `data-path.ts` reads `DATA_PATH/<user>/email_accounts/*/emails.db` and picks
-      the newest by mtime. Two core callers use it: `user-instance.ts` (`OFFICER_EMAIL_DB`, for the agent's
-      `email_db` MCP tool) and `chat/websocket.ts` (`resolveEmailCwd`, for a chat opened against a mailbox).
-      `ChatPanelWrapper.tsx` hardcodes `workspace.id === 'email'` to trigger the second.
-
-      This traded a CODE dependency for a CONVENTION, which was the right trade at the time — the code
-      dependency is what put `officer-agent` into a restart loop on 2026-08-13 (`fbb6d6c7`). But the
-      convention has no enforcement on either side: if the plugin ever renames `emails.db`, moves to one
-      database per folder, or nests differently, core silently points the agent at a path that does not
-      exist. No import breaks, no type fails, no test covers it. The tool just goes quiet.
-
-      The clean version needs two things the platform does not have: a way for a plugin to ship an MCP tool,
-      and a way for a plugin to declare a chat context. Until both exist this stays — but it should be a
-      decision someone re-takes, not one that is inherited.
-
-## Plugin repos — what writing their READMEs turned up
-
-Three READMEs were written on 2026-08-16 (`music`, `offscale`, `example`) and pushed to their own repos.
-Writing them meant reading each plugin against its own documentation, which is the cheapest audit there
-is. Everything below was found that way and none of it is fixed.
-
-**Offscale is still branded headscale in shipped, user-visible places.**
-
-- [ ] `web/panels.ts` registers three panels named "Headscale servers", "Headscale", "Headscale" — these
-      are what a user reads in the panel picker.
-- [ ] `web/OffscaleServerPicker.tsx:26` renders the literal heading `Headscale`.
-- [ ] `web/ConsoleView.tsx:22` mints terminal session ids as `headscale-console-<id>`. That is persisted
-      state rather than a comment, so renaming it orphans live console sessions — cheap now, less so later.
-- [ ] ~12 stale `/headscale` route paths in comments across `web/`, plus `App.tsx:78` in the platform.
-- [ ] `PLUGIN.md` has ~11 stale self-references (`headscale_servers`, `listHeadscaleServers`, permission
-      key `headscale`, `/api/headscale`, panel keys, `useHeadscaleSection()`).
-
-**Documentation that actively lies.**
-
-- [ ] `plugins/offscale/PLUGIN.md:445-449` says the ACL assistant is deliberately unwired and must not be
-      wired up. It IS wired — `sidecar/assist.ts`, `claude-proxy.ts`, `web/PolicyAssistant.tsx` and the
-      manifest's `readOnlyWrites` all ship.
-- [ ] `plugins/offscale/OFFSCALE_API.md` claims to publish the contract "in full" and omits six live
-      surfaces, including all of `/_officer/companion/*` and `/_officer/enroll/invites`.
-- [ ] `plugins/music/MUSIC_API.md` says favourites and now-playing are served by the platform from
-      Postgres; the sidecar serves them itself, scoped by `X-Officer-User`. It also documents the playlist
-      item key as `<albumRel>/<file>` while the schema and favourites use the home path — the same file
-      contradicts itself.
-- [ ] `plugins/music/MUSIC_API.md` documents now-playing as one snapshot per user with no mention of
-      `?device=`, which the schema makes part of the primary key. A mobile client written to that doc is
-      correct only by accident.
-- [ ] `plugins/example/manifest.ts:12` lists `web/Router.tsx` as the frontend convention; discovery
-      requires `web/panels.ts` AND `web/layout.ts`. The same file documents an `ownerOnly` field that does
-      not exist and never did.
-- [ ] The Companion contract docs the offscale code cites (`COMMS/OFFSCALE_COMPANION_API.md`,
-      `COMMS/OFFSCALE_INVITE_ENROLLMENT.md`) exist nowhere — `COMMS/` was deleted. Two of the plugin's
-      eight sections depend on a service with no published spec.
-
-**Conventions a plugin author has to guess.** All verified in `discover.ts`/`mount.ts`, none documented
-outside the example's new README:
-
-- [ ] Panel `key`s and layout `appType`s must agree, and a mismatch fails SILENTLY — `PluginScreen.tsx`
-      pins `appTypes.allowed` with `fallback: allowed[0]`, so a typo renders the first panel.
-- [ ] The `<appName>_` table prefix is unenforced. The barrel is `export *` across every plugin into one
-      `public` schema, so two plugins choosing the same unprefixed name collide with no error until push.
-- [ ] A manifest is module-cached for the process lifetime: editing it does nothing until
-      `pm2 restart officer`, and `outdated` cannot see a version bump either.
-- [ ] Permission-key prefixing is convention, not validation. Only a collision with a CORE key is refused,
-      silently, at mount, with a console error.
-
-**Genuinely missing.**
-
-- [ ] `plugins/offscale` spawns the `ssh` binary (`sidecar/ssh.ts`) but declares no `osDependencies`, so
-      install never checks for it and the Console fails at runtime instead of refusing at install.
-      `PLUGIN.md:421` already lists it as a host dependency — the manifest is the file that is wrong.
-- [ ] `scripts/reindex-music.ts` resolves its port file from `process.env.DATA_PATH ?? cwd/data`, and
-      `DATA_PATH` stopped being an env var on 2026-08-13. Run from the repo it looks in
-      `platform/data/music/.server` and never finds it. One line.
-- [ ] `plugins/music/web/MusicPlayerHost.tsx` header says it is mounted once in the persistent
-      DashboardLayout; it moved into `MusicDetail`. That is exactly the comment that would mislead
-      someone about why playback stops off-route.
-- [ ] A plugin's `assets/icon.png` is only published while installed, so its real icon is invisible in the
-      store before install. The catalogue's `iconUrl` works around it, and an author has to know to add one.
-
-## Waiting on the owner to run it
-
-- [ ] **The AI usage panel works for one account; the multi-account half is unproven.** Landed and verified
-      live 2026-08-18 — click the harness pill in the chat composer and it shows Claude Code's own usage
-      display, limits with countdowns, tokens by day and by model.
-
-      What is NOT proven is the reason it was built: a second Anthropic login, the account tabs, and a
-      stored row surviving a switch. **Group C of `docs/ai-usage-panel.md`** is that, and test 10 is the one
-      most likely to be wrong — a refresh must not overwrite an account it could not measure, and that
-      failure is silent.
-
-      Deliberately deferred rather than forgotten: the owner switches accounts when the work needs it, and
-      the worst case is one wrong tab. Run Group C the first time a second login lands on this machine.
-
-      Also unexercised: the model dropdown no longer renders the literal string `claude-code` for a chat
-      that named no tier. Needs `officer-claude-code` restarted, not just `officer` — the parser in the
-      sidecar is what emits it.
-
-- [ ] **Plan mode is live and has never been exercised.** Landed `1b811dfa` on 2026-08-17; all three
-      processes were restarted the same evening, so the code IS running. Not one part of it has been used:
-      not the toggle, not the plan card, not the round trip from the sidecar's pending promise to a browser
-      and back, not the stored column, not `settleAllPlans`.
-
-      **`docs/chat-plan-mode-test-plan.md` is the checklist** — eleven ordered tests, what each should show,
-      and what its failure looks like. Test 4 (approve) is the one most likely to be wrong: it is the case
-      that changed the design, because approving `ExitPlanMode` drops the harness into `default` rather than
-      back into `bypassPermissions`.
-
-      Safe to leave untested. Plan mode is opt-in and off by default, so a conversation that never touches
-      the toggle behaves exactly as it did before. The one failure worth knowing in advance is a turn that
-      never ends after a plan is answered — that would be a leaked `canUseTool` promise, and
-      `pm2 restart officer-claude-code` clears it.
-
-      Design and the probe results behind it: `docs/chat-plan-mode.md`.
-
-## Known bugs
-
-- [ ] **Background job logs appeared INSIDE an open file in the code editor.** Reported 2026-08-16 by the
-      owner, on the production host, running as a member with a `bun dev` background job streaming. The
-      editor had a file open; the runner's log output was rendered into that file's buffer. **It was never
-      saved** — the file on disk was intact, and the editor was not marked dirty as far as the report goes.
-      So this is a DISPLAY-layer mixing, not a write path: the bytes reached a buffer that only exists in
-      the browser.
-
-      Not reproduced, not diagnosed, and deliberately not guessed at further than this. What is worth
-      recording is where the two streams could meet, because both were added in the same week and they are
-      the only two things writing text into a panel:
-
-      - `useTaskDetail` polls `GET /chat/tasks/:id` every 2s while a job runs and returns `{kind:'log', text}`.
-        `CodeSurface` renders it — and `CodeSurface` is shared with the editor's own surfaces.
-      - The editor keys open files by PATH (`useEditorState`), while the task panel keys by TASK ID, and
-        several panel-scoped stores in between are keyed by `panelId` through `useGlobal` /
-        `usePanelMemory`. A key collision between those namespaces would produce exactly this: one panel's
-        text arriving in another's buffer, with no server involvement and nothing persisted.
-
-      **Screenshot 2026-08-16 answers the first question and kills two theories.** The tab carries a DIRTY
-      dot, and the log lines have Monaco line numbers continuing the file's own sequence (…180, then eight
-      lines of dotenv/POSTGRES_URL output, then 189 `</Link>`). So the MODEL was mutated — this is the
-      editor's content state, not its render, and the earlier "the file was okay" means it was simply never
-      saved. Note the splice is mid-file, not a replacement: JSX text and log text share line 180
-      (`Forgot?--env-file=.env --watch src/server.tsx`).
-
-      Ruled out by that: `CodeSurface` renders with shiki, not Monaco, so the two cannot share an editor
-      instance. And the app's only other Monaco mount (`FileViewer/renderers/EditorRenderer.tsx`) passes no
-      `path`, so its model URI is `inmemory://model/N` and cannot collide with the editor's `file:///…`.
-
-      So the pollution arrives as TEXT through the editor's own content path — `openFile`, `setContent`, or
-      the end-of-turn refresh that re-reads every clean file. That last one is the prime suspect: it is the
-      only thing that writes a buffer without a keystroke, it runs exactly when a background job is
-      producing output, and it lands its result by path.
-
-      **Narrowed 2026-08-16, and a trap is now set.** `reloadFile` sets `originalContent` and clears
-      `isDirty`, so the end-of-turn refresh CANNOT produce the dirty dot in the screenshot — the refresh
-      path is out. Nothing in the app injects text into the page either: no `execCommand('insertText')`,
-      no synthetic keyboard events, no clipboard writes near the editor. That leaves Monaco's own change
-      event, which means it looks like a paste or a programmatic `executeEdits`.
-
-      **`docs/editor-buffer-corruption.md` is the procedure — read that when someone reports this.** A
-      trap is set: `reportSuspiciousEdit` in CodeEditor.tsx records any insertion over 200 characters
-      arriving while the editor is NOT focused, with the stack at that moment, into
-      `localStorage['officer:editor-buffer-trap']` (last five). It persists rather than only logging
-      because this arrives as a report hours later, when the console is long gone — the question becomes
-      "paste this key" instead of "what did you see?". Delete all of it once the cause is known.
-
-      One thing to rule out first, because it is mine and it is recent: `2b70a570` gave `<Editor>` a real
-      `path`, so models are now keyed by file URI globally rather than being per-mount `inmemory://`. Two
-      panels open on the same file now share one model. That should be an improvement, but it is a change
-      in exactly the layer that turned out to be involved.
-
-      Worth noting because it is the reason to look at all: an editor that can show content the file does
-      not contain is a correctness problem even when it never writes — the next person's instinct will be to
-      save the file to "fix" it.
-
-- [ ] **`bootstrap.ts` runs `npm install -g` for Pi on every boot.** `findPiPackageDir` checks stale
-      paths and an outdated package name, so `installPi` always fires and floods the logs with
-      `EEXIST` noise. Should detect `@earendil-works/pi-coding-agent` at the real npm prefix.
-
-- [x] **VNC mirror can orphan/duplicate x11vnc across sidecar restarts.** FIXED 2026-08-01 in the
-      Xvnc rewrite. The diagnosis was right and survived the move off mirroring: the running desktop is
-      tracked in module-level state, so a sidecar restart forgot it while the server kept running
-      orphaned, and `waitForPort` treated ANY listener on 5900 as success — so the next start reported
-      success while the browser talked to the stale process. Now `reclaimPort` frees 5900 (TERM, then
-      KILL after 2s) before spawning, and `waitForPort` also fails if the process we spawned has
-      exited, so a listener that is not ours can no longer be mistaken for a healthy start.
-
-## Infra (alpha)
-
-- [ ] **Delete `/etc/systemd/system/officer-vnc.service`.** Hand-installed unit (not in this repo)
-      that ran `vncserver :1 -geometry 1920x1080 -localhost yes -fg` with `Restart=on-failure`,
-      enabled at boot — it kept a whole parallel XFCE session alive on `:1` (283 MB, 166 tasks)
-      independently of the platform, and silently respawned it whenever the display was killed.
-      Obsolete now that the Desktop panel mirrors `:0` via x11vnc. Disabled 2026-07-15
-      (`systemctl disable --now`), but the unit file is still on disk. Nothing in the codebase
-      recreates it and nothing documents it, so delete the file rather than leave a mystery service
-      one `systemctl enable` away from returning.
-
-- [ ] **`ufw` blocks port 9010 on the LAN.** Default deny incoming; only `22/tcp`, `80/tcp`,
-      `443/tcp`, and everything on `tailscale0` are allowed — so `http://192.168.47.196:9010` is
-      unreachable from the LAN while localhost and Tailscale work.
-      Fix: `sudo ufw allow from 192.168.47.0/24 to any port 9010 proto tcp`
-
-## Backburner
-
-- [ ] **Music tagging feature** (`/music`, Mp3tag-inspired). v1 was scoped as a two-panel workspace —
-      file browser left, table right, single "Load" context-menu item flattening audio files into the
-      table. Discarded before completion; would need `GET /file-browser/flatten-audio`, a
-      `useFilesAPI.flattenAudio` method, and the Music app panel rebuilt from scratch.
+## Urgent
+
+- **due 2026-08-22** — **Settings to choose which server each feature uses** — [fleet/settings-to-choose-which-server-each-feature](todo/fleet/settings-to-choose-which-server-each-feature.md)
+
+## Themes
+
+**84 open, 22 closed**, across 28 themes.
+
+| Theme | Open | Closed |
+| --- | --- | --- |
+| [Affordances](todo/affordances/) | 1 | 0 |
+| [Speaking into a session from the backend](todo/agent-delivery/) | 2 | 0 |
+| [Backburner](todo/backburner/) | 2 | 0 |
+| [Background jobs](todo/background-jobs/) | 2 | 0 |
+| [Chat interface](todo/chat/) | 1 | 0 |
+| [Code editor](todo/code-editor/) | 1 | 0 |
+| [Components — the shared UI layer](todo/components/) | 1 | 0 |
+| [File watchers](todo/file-watchers/) | 1 | 0 |
+| [Fleet — one command across every server](todo/fleet/) | 5 | 2 |
+| [Headscale](todo/headscale/) | 0 | 5 |
+| [Infra (alpha)](todo/infra/) | 6 | 0 |
+| [Known bugs](todo/known-bugs/) | 4 | 1 |
+| [Mobile apps](todo/mobile-apps/) | 3 | 0 |
+| [Multi-user](todo/multi-user/) | 17 | 3 |
+| [Notifications](todo/notifications/) | 1 | 0 |
+| [Plugin ecosystem — before it opens to third parties (2026-08-18)](todo/plugin-ecosystem/) | 3 | 0 |
+| [Plugin infrastructure](todo/plugin-infrastructure/) | 2 | 0 |
+| [Plugin repos — what writing their READMEs turned up](todo/plugin-repos/) | 10 | 9 |
+| [Plugin system — gaps the email extraction found (2026-08-18)](todo/plugin-system/) | 3 | 0 |
+| [Prompt library](todo/prompt-library/) | 1 | 0 |
+| [Shell — spotlight and the dock](todo/shell/) | 1 | 1 |
+| [Sub-agents — agents spawning agents](todo/sub-agents/) | 1 | 1 |
+| [Sudo terminal modal](todo/sudo-terminal-modal/) | 1 | 0 |
+| [Tooling — the list about the list](todo/tooling/) | 1 | 0 |
+| [Transmission](todo/transmission/) | 6 | 0 |
+| [URL state — routing, query strings and selection](todo/url-state/) | 4 | 0 |
+| [Waiting on the owner to run it](todo/waiting-on-owner/) | 2 | 0 |
+| [Workspaces & Panels](todo/workspaces-panels/) | 2 | 0 |
+
+## Records — no items, kept for the reasoning
+
+- [Affordances (2026-08-19)](todo/affordances-runtime/_index.md)
+- [Email](todo/email/_index.md)
+- [Inputs outside forms (2026-08-19) — swept, fixed, closed](todo/inputs-outside-forms/_index.md)
+- [Release-binary dependencies (2026-08-19)](todo/release-binaries/_index.md)
+
+## Newest ten
+
+- `2026-08-21` **No way to see which affordances are reaching your agents, or to turn one off** — [affordances/no-way-to-see-which-affordances-are-reaching](todo/affordances/no-way-to-see-which-affordances-are-reaching.md)
+- `2026-08-21` **Generalise it: deliver a turn to a named session, from anywhere in the backend** — [agent-delivery/generalise-it-deliver-a-turn-to-a-named](todo/agent-delivery/generalise-it-deliver-a-turn-to-a-named.md)
+- `2026-08-21` **The cost of resuming grows with the transcript** — [agent-delivery/the-cost-of-resuming-grows-with-the-transcript](todo/agent-delivery/the-cost-of-resuming-grows-with-the-transcript.md)
+- `2026-08-21` **'An agent spawned by a background job is invisible in the interface'** — [background-jobs/a-spawned-agent-is-invisible-in-the-interface](todo/background-jobs/a-spawned-agent-is-invisible-in-the-interface.md)
+- `2026-08-21` **Show how long a turn took, beside its cost and tokens** — [chat/show-how-long-a-turn-took-beside-its-cost](todo/chat/show-how-long-a-turn-took-beside-its-cost.md)
+- `2026-08-21` **Map every custom component, centralise it, and let a theme override it** — [components/map-every-custom-component-and-centralise-it](todo/components/map-every-custom-component-and-centralise-it.md)
+- `2026-08-21` **Watch a directory, and attach actions to what lands in it** — [file-watchers/watch-a-directory-and-attach-actions-to-what](todo/file-watchers/watch-a-directory-and-attach-actions-to-what.md)
+- `2026-08-21` **Settings to choose which server each feature uses** — [fleet/settings-to-choose-which-server-each-feature](todo/fleet/settings-to-choose-which-server-each-feature.md)
+- `2026-08-21` **'A deploy verifies PM2, not that the app still works'** — [fleet/a-deploy-report-verifies-pm2-not-that-the-app](todo/fleet/a-deploy-report-verifies-pm2-not-that-the-app.md)
+- `2026-08-21` **'The deploy commander cannot finish its own host'** — [fleet/the-deploy-commander-cannot-finish-its-own-host](todo/fleet/the-deploy-commander-cannot-finish-its-own-host.md)

deploy/2026.08.21.01.md

@@ -0,0 +1,43 @@
+# 2026.08.21.01
+
+The first release under this scheme, and the first deploy of the night's work. Everything since
+`0.1.0`, which is to say everything.
+
+## What is in it
+
+- **Two plugins extracted**: `desktop` (VNC) and `browser` (the relay, its extension and its panels),
+  with the platform losing the VNC vocabulary it should never have had.
+- **Themes regenerated** — all 22 from Omarchy, 84 tokens each, contrast-checked. The eleven that
+  shipped with 34 of an 84-token contract are replaced.
+- **Spotlight**, on `shift-space`.
+- **Chat fixes**: a turn waiting on a person is no longer killed as stalled, and an answer that goes
+  nowhere now says so.
+- **Plugin infrastructure**: `importFresh` (a reinstalled plugin was being read from the deleted one),
+  `ownerOnly` paths, dependency `scope`, extensions published unauthenticated, `plugins:update`.
+- An **Update button** in maintenance, and the **sub-agent budget guard**.
+
+## Steps, in order, on every host
+
+Same on all six — nothing here is host-specific.
+
+1. `git pull --ff-only`
+2. `bun install` — **required.** `bun.lock` changed and `sharp` was removed from the platform's
+   dependencies (it moved into atelier's `npmDependencies`). Without this the lockfile and the tree
+   disagree, and the frozen-lockfile rule turns that into a failed install later rather than now.
+3. `bun run plugins:update`
+4. `bun db:push` — **required.** `users.username` becomes `NOT NULL` and gains a case-insensitive
+   unique index.
+5. `pm2 restart officer`
+6. Then, and only for hosts that have them: `pm2 restart officer-claude-code officer-pty` — both load
+   changed sidecar code. **Not** the whole ecosystem, ever.
+
+## The one thing that could have gone wrong, and did not
+
+`db:push` refuses if any account has a null username, or if two differ only by case. **Checked on all
+six before writing this**: 12 users across the fleet, zero nulls, zero case-collisions. If a future
+host fails here, that is why, and the fix is data rather than schema.
+
+## Post-deploy check
+
+`pm2 jlist` — every restarted process `online` with a moved `restart_time`, and the platform's own
+domain answering 200. `docs/servers.md` has the domain per host.

deploy/2026.08.21.01.report.md

@@ -0,0 +1,42 @@
+# 2026.08.21.01 — deployed
+
+**6/6 hosts.** Every platform is on `eee73e77`, schema pushed, processes online and every domain
+answering 200.
+
+Rolled out by hand from `officer-dev`, one host at a time, because nothing wakes the commander yet.
+
+| Host                         | Before     | After      | install | db:push | plugins        | restarted                 | domain |
+| ---------------------------- | ---------- | ---------- | ------- | ------- | -------------- | ------------------------- | ------ |
+| `alpha`                      | `44f6c6d6` | `eee73e77` | ok      | applied | 12/12          | officer, claude-code, pty | 200    |
+| `fofos`                      | `6db26ae4` | `eee73e77` | ok      | applied | 1/1            | officer, claude-code, pty | 200    |
+| `edge-pertento-bridge`       | —          | `eee73e77` | ok      | applied | 2/2            | officer, claude-code, pty | —      |
+| `production-pertento-bridge` | —          | `eee73e77` | ok      | applied | 1/1            | officer, claude-code, pty | —      |
+| `macbookpro`                 | `2e76cabf` | `eee73e77` | ok      | applied | none installed | officer, claude-code, pty | —      |
+| `officer-dev`                | `eee73e77` | `eee73e77` | ok      | applied | 3/3            | officer, pty              | 200    |
+
+`bun install` reported no changes everywhere — the lockfile diff was `sharp` leaving, and nothing had
+it as a transitive need.
+
+## What was NOT done, and why
+
+- **`officer-claude-code` was not restarted on `officer-dev`.** That sidecar holds the session doing
+  the deploy; restarting it would have ended the conversation mid-rollout. Every other host got it.
+  **Outstanding — run it when convenient.**
+- **`fofos` did not sync its capabilities.** `git` refused the checkout as "dubious ownership" and
+  printed the usual `safe.directory` remedy. Ownership is actually correct (`pastilhas:pastilhas`,
+  running as `pastilhas`), so the cause is not what the message suggests and papering over it with a
+  config entry would hide a real question. Its one plugin updated fine. **Unexplained.**
+- **`macbookpro` has no plugins installed**, so `plugins:update` was a no-op there. Not a failure —
+  worth knowing before someone expects the browser relay on it.
+
+## Notes for the next one
+
+- **The dirty-tree check was too strict.** `git status --porcelain` counts UNTRACKED files, and
+  `fofos` was skipped on the first pass for one stray `scripts/chk.ts` that a fast-forward could never
+  touch. The rule is `--untracked-files=no`; untracked files get a note, not a skip.
+  `prompts/deploy-commander.md` says "refuse a dirty tree" and should say which kind.
+- **`macbookpro` needs `$SHELL -ic` and nothing else.** Exporting a PATH to reach `bun` broke `pm2`,
+  which is only on the interactive shell's path. Do not help it.
+- **The pre-flight was worth it.** `users.username` became `NOT NULL` with a case-insensitive unique
+  index; checked across all six first — 12 users, zero nulls, zero collisions — so `db:push` applied
+  everywhere without a data fix.

deploy/2026.08.21.02.md

@@ -0,0 +1,33 @@
+# 2026.08.21.02
+
+Command HQ made usable: a commander's transcript renders, and Live separates commanders from the
+conversations somebody is having.
+
+Frontend and one chat API route. **No schema change, no dependency change** — verified against
+`deploy/2026.08.21.01..HEAD`: nothing touched `src/schema.ts`, `package.json` dependencies or `bun.lock`.
+So there is no `db:push` and no `bun install` in this plan, and a host that runs one has run something
+this release did not ask for.
+
+## What is in it
+
+- Three separate defects that each made a commander's conversation invisible: the pane navigated to
+  officer's session key instead of Claude's transcript id, then remounted and discarded the live chat at
+  the end of the first turn, then seeded `useChat` with messages that had not arrived yet.
+- Live splits into `Agents (n)` / `Commanders (n)`, and `/chat/live` now carries `isCommander` so a client
+  with no room for tabs — the iPhone app's swipe order — can filter on it.
+- The handoff watcher's controls, which were built and never rendered.
+
+## Every host
+
+In the order the procedure gives. Nothing here is host-specific.
+
+1. `git pull --ff-only`
+2. `bun run plugins:update`
+3. `pm2 restart officer`
+
+**`officer` only.** No sidecar changed. Restarting `officer` rebuilds the SPA on boot, which is what
+ships the frontend — there is no separate build step to run.
+
+## Then check
+
+`pm2 jlist`: `officer` is `online` and its `restart_time` moved. Nothing else should have restarted.

deploy/2026.08.21.02.report.md

@@ -0,0 +1,76 @@
+# 2026.08.21.02 — deployed
+
+**6/6 hosts.** Every platform the plan reaches is on `55ecbf4f`, `officer` restarted and back `online`,
+and no other process on any host was touched by the restart step.
+
+`officer-dev` is the seventh platform and was **skipped entirely** — it is the host this commander runs
+on. See below.
+
+Rolled out one host at a time from `officer-dev`, by the procedure in `prompts/deploy-commander.md`.
+
+## The release
+
+Tag `deploy/2026.08.21.02` → `55ecbf4f`; `package.json` says `2026.08.21.02`. They agree.
+Plan: `deploy/2026.08.21.02.md` — pull, `plugins:update`, restart `officer` only. No `bun install`,
+no `db:push`, and neither was run anywhere.
+
+| Host                         | Before     | After      | plugins        | restarted | restart_time | back    |
+| ---------------------------- | ---------- | ---------- | -------------- | --------- | ------------ | ------- |
+| `alpha`                      | `eee73e77` | `55ecbf4f` | 12/12          | officer   | 60 → 61      | online  |
+| `fofos`                      | `eee73e77` | `55ecbf4f` | 1/1            | officer   | 15 → 16      | online  |
+| `macbookpro`                 | `eee73e77` | `55ecbf4f` | none installed | officer   | 2 → 3        | online  |
+| `edge-pertento-bridge`       | `eee73e77` | `55ecbf4f` | 2/2            | officer   | 4 → 5        | online  |
+| `production-pertento-bridge` | `eee73e77` | `55ecbf4f` | 1/1            | officer   | 6 → 7        | online  |
+| `deadpool-rezio-bridge`      | `bd41dff5` | `55ecbf4f` | none installed | officer   | 0 → 1        | online  |
+| `officer-dev`                | `55ecbf4f` | `55ecbf4f` | not run        | **none**  | 155          | skipped |
+
+Every host was clean of TRACKED changes and every pull was a real fast-forward. `pm2 jlist` after each
+restart showed zero processes in any state other than `online`, so nothing was collaterally bounced.
+
+Checkout paths were taken from each `officer` process's `pm_cwd`, not assumed — so `alpha`'s stale
+second checkout at `~/dockers/officer.dev/platform` was never touched.
+
+## What was NOT done, and why
+
+- **`officer-dev` was skipped completely** — no pull, no `plugins:update`, no restart. This session runs
+  inside `officer-claude-code` on it, and its `officer` process would have been restarted out from under
+  the rollout. Its checkout is already at `55ecbf4f` (the tag was cut here), so **the code is present and
+  the running process is a build behind**. `pm2 restart officer` on `officer-dev`, when the owner is happy
+  with the rest, finishes this release. **Outstanding.**
+
+- **`deadpool-rezio-bridge` was two releases behind, not one.** It came in at `bd41dff5` — the
+  `deploy/2026.08.21.01` commit — while every other host was at `eee73e77`. The fast-forward carried both,
+  14 files, and it is now level. Nothing was wrong with it; it simply had not had the interim commits.
+
+- **`fofos` still cannot sync its capabilities.** Same `dubious ownership` refusal as `2026.08.21.01`,
+  same unexplained cause, still not papered over with a `safe.directory` entry. New this time and probably
+  the same root: its gitea plugin could not publish affordances —
+  `EACCES: permission denied, mkdir '/home/pastilhas/officerdev/capabilities/affordances/officerdev'`.
+  Both point at that capabilities checkout being owned by somebody other than `pastilhas`, which is worth
+  one `stat` before the next deploy. `plugins:update` exited 0 and the plugin itself updated. **Unexplained,
+  and now costing two things instead of one.**
+
+- **`fofos` carries one untracked file**, `scripts/chk.ts`. A note, not a skip — that is the rule the last
+  report asked for and it worked.
+
+- **No `bun install` and no `db:push` ran anywhere.** The plan says this release needs neither.
+
+## Notes for the next one
+
+- **`macbookpro` needs the PATH helped after all — carefully.** The last report said "`$SHELL -ic` and
+  nothing else"; that is right for `pm2` and wrong for `bun`. `bun` is at `/opt/homebrew/bin/bun`, and
+  `.zshrc` does not put `/opt/homebrew/bin` on the interactive PATH, so `bun run plugins:update` failed
+  `command not found` before it ran a line. What works is **appending** that one directory for that one
+  command (`export PATH=$PATH:/opt/homebrew/bin`) and running `pm2` as a separate command. Appending
+  shadows nothing — `pm2` stays at `~/.local/bin/pm2` and `node` at `node@22`. **Replacing** the PATH is
+  what broke `pm2` last time. `docs/servers.md` says "do not help the PATH" and should say "append, never
+  replace, and only for `bun`".
+
+- **`$SHELL -ic` on `macbookpro` also aliases `ls` to `eza`**, which is not installed, so a probe using
+  `ls` reports a missing file that is plainly there. Use `command ls` or a plain non-interactive shell for
+  file tests.
+
+- **`plugins:update` restarts sidecars itself.** It prints `sidecar: started` per plugin, so on `alpha`
+  twelve sidecars bounced during step 3 — before the plan's restart step, and not something the plan asked
+  for. That is the script's own behaviour and it is not new, but a report that only lists `officer` under
+  "restarted" is understating it. Worth a line in the plan format.

deploy/2026.08.21.03.md

@@ -0,0 +1,30 @@
+# 2026.08.21.03 — a drill
+
+**No code change.** This release exists to watch the watcher fire, and to exercise the claim rule that
+landed after the last one.
+
+There is nothing to bump but the version and this file, and that is deliberate rather than lazy: a tag
+whose `package.json` disagrees with it makes the commander stop, and a tag with no plan beside it makes it
+stop too, so the smallest honest release is still a version and a plan.
+
+## What is being tested
+
+- The watcher noticing a tag and spawning an agent on its own, with nobody watching the composer.
+- `scripts/deploy-claim.sh` taking the claim before anything is deployed, and the deploy recording success
+  through the same script.
+- A second pass over the same version being refused rather than repeated.
+
+## Every host
+
+1. `git pull --ff-only`
+2. `bun run plugins:update`
+3. `pm2 restart officer`
+
+`officer` only. No `bun install`, no `db:push` — nothing in this range touches a schema, a dependency or
+the lockfile, and `git diff deploy/2026.08.21.02..HEAD` is two files.
+
+## Then check
+
+`pm2 jlist`: `officer` is `online` and its `restart_time` moved. Nothing else restarted.
+
+`scripts/deploy-claim.sh state 2026.08.21.03` says `done`.

deploy/2026.08.21.03.report.md

@@ -0,0 +1,83 @@
+# 2026.08.21.03 — deployed
+
+**6/6 hosts.** Every platform the plan reaches is on `3343b21a`, `officer` restarted and back `online`,
+and no other process on any host was left in a state other than `online`.
+
+`officer-dev` is the seventh platform and was **skipped entirely** — it is the host this deploy ran on.
+See below; its skip is now two releases deep.
+
+This was a drill. The release changes `package.json` and adds its own plan file, nothing else, and it
+exists to watch the claim rule work. It did: `scripts/deploy-claim.sh claim 2026.08.21.03` was taken
+before any host was touched, and `state` reads `done`.
+
+## The release
+
+Tag `deploy/2026.08.21.03` → `3343b21a`; `package.json` at that commit says `2026.08.21.03`. They agree.
+Plan: `deploy/2026.08.21.03.md` — pull, `plugins:update`, restart `officer` only. No `bun install`, no
+`db:push`, and neither ran anywhere.
+
+| Host                         | Before     | After      | plugins        | restarted | restart_time | back    |
+| ---------------------------- | ---------- | ---------- | -------------- | --------- | ------------ | ------- |
+| `alpha`                      | `55ecbf4f` | `3343b21a` | 12/12          | officer   | 61 → 62      | online  |
+| `fofos`                      | `55ecbf4f` | `3343b21a` | 1/1            | officer   | 16 → 17      | online  |
+| `macbookpro`                 | `55ecbf4f` | `3343b21a` | none installed | officer   | 3 → 4        | online  |
+| `edge-pertento-bridge`       | `55ecbf4f` | `3343b21a` | 2/2            | officer   | 5 → 6        | online  |
+| `production-pertento-bridge` | `55ecbf4f` | `3343b21a` | 1/1            | officer   | 7 → 8        | online  |
+| `deadpool-rezio-bridge`      | `55ecbf4f` | `3343b21a` | none installed | officer   | 1 → 2        | online  |
+| `officer-dev`                | `3343b21a` | `3343b21a` | not run        | **none**  | 155          | skipped |
+
+Every host came in at `55ecbf4f` — the previous release, level for the first time in three deploys — and
+every pull was a real fast-forward. Checkout paths were taken from each `officer` process's `pm_cwd`, so
+`alpha`'s stale second checkout at `~/dockers/officer.dev/platform` was never touched. `macbookpro` was
+awake.
+
+**More than `officer` restarted, and the plan did not ask for it.** `bun run plugins:update` bounces each
+plugin's sidecar as it goes: twelve on `alpha`, two on `edge`, one each on `fofos` and
+`production-pertento-bridge`. Every one came back `online`. Nothing was left down.
+
+## What was NOT done, and why
+
+- **`officer-dev` was skipped completely** — no pull (it is already at the tag, which was cut here), no
+  `plugins:update`, no restart. This session runs inside `officer-claude-code` on it, and restarting
+  `officer` there ends the rollout mid-flight.
+
+  **Its running process is now two releases old.** `restart_time` is 155, unchanged since before
+  `2026.08.21.02` — so the same restart the last report left outstanding is still outstanding, and it now
+  carries both releases. The code for both is on disk. `pm2 restart officer` on `officer-dev` finishes
+  them, whenever the owner wants it. **Outstanding, and no longer new.**
+
+- **`fofos` still cannot sync its capabilities.** Third deploy running, same two failures, same
+  unexplained cause:
+
+  - `capabilities: failed — fatal: detected dubious ownership in repository at
+    '/home/pastilhas/officerdev/capabilities'`
+  - `affordances: NOT published — EACCES: permission denied, mkdir
+    '/home/pastilhas/officerdev/capabilities/affordances/officerdev'`
+
+  Both say that checkout is not owned by `pastilhas`. Still not papered over with a `safe.directory`
+  entry, because a `safe.directory` line would silence the first message and leave the `EACCES` — they are
+  one cause, and one `stat` settles it. `plugins:update` exited 0 and the gitea plugin itself updated, so
+  this costs capability sync on that host and nothing else. **Unexplained.**
+
+- **`fofos` still carries one untracked file**, `scripts/chk.ts`. A note, not a skip — the rule works.
+
+- **No `bun install` and no `db:push` ran anywhere.** The plan says this release needs neither, and
+  `deploy/2026.08.21.02..deploy/2026.08.21.03` is two files.
+
+- **Nothing was verified beyond `pm2`.** `officer` is `online` with a moved `restart_time` on all six
+  hosts; no HTTP request was made to any domain. That is what the plan asked to check.
+
+## Notes for the next one
+
+- **The `macbookpro` PATH rule in `docs/servers.md` is now correct and was followed as written.**
+  `export PATH=$PATH:/opt/homebrew/bin` for the `bun` command, `pm2` as a separate command, and both
+  worked first time. Nothing to change.
+
+- **`$SHELL -ic "pm2 jlist"` with stderr discarded came back empty once**, on `edge-pertento-bridge`, and
+  the same command with `2>&1` a second later returned the full list. Transient, and it looked exactly
+  like a dead pm2. Capture stderr and parse the JSON out of the stream rather than trusting the whole of
+  stdout to be JSON — an interactive shell puts its own noise (`starship`, aliases) in there.
+
+- **`officer-dev`'s skip needs a home other than a report.** It has been carried forward twice now, and a
+  report is not a place work waits — the next deploy will write the same paragraph. Either the owner
+  restarts it, or it belongs in `TODO.md`.

deploy/2026.08.21.04.md

@@ -0,0 +1,51 @@
+# 2026.08.21.04
+
+`/login` in the chat window, switching between Anthropic accounts from the usage pills, and a Spotlight
+that can actually reach the plugins.
+
+**This release has a SCHEMA CHANGE.** The previous two did not, and said so; this one adds
+`anthropic_accounts`. Without the push, `GET /api/claude-auth/accounts` fails on a missing relation the
+first time the usage panel is opened — so `bun db:push` is a step here, not an optional tidy.
+
+No dependency change: nothing in `deploy/2026.08.21.03..HEAD` touches `package.json` dependencies or
+`bun.lock`. Do not run `bun install`.
+
+## What is in it
+
+- **`/login` works in the chat composer.** `claude auth login` behaves the same through plain pipes as
+  through a terminal, so it is a subprocess with two reads and one write — no pty. Copy-link beside the
+  open-link button, because opening it here opens it in the browser already signed into the account you are
+  leaving.
+- **Account switching from the usage pills.** `anthropic_accounts` keeps each login's OAuth credential
+  encrypted under a per-account key from the secret store, so a switch needs no browser round trip. An
+  account with no saved credential still gets a chip you can read — cached percentages and a correct reset
+  time — marked as needing `/login`.
+- **Spotlight reached eight items.** Core filled all eight slots, so on a populated server every plugin was
+  unreachable. The cap is gone, rows draw the dock's own tiles, and the settings list went from nine
+  hand-written rows to twenty read off the screens that declare them.
+- The ephemeral pane opens from the markup rather than an effect.
+
+## Every host
+
+1. `git pull --ff-only`
+2. **`bun db:push`** — new this release, see above
+3. `bun run plugins:update`
+4. `pm2 restart officer`
+
+`officer` only; no sidecar changed. Restarting it rebuilds the SPA, which is what ships the frontend.
+
+**`alpha` is already at this code**, deployed by hand as a preview before the tag. Its pull will be a
+no-op and that is expected — run the remaining steps anyway rather than skipping the host, so its
+`restart_time` moves with everyone else's.
+
+**On `macbookpro`, append to the PATH for `bun`** — `export PATH=$PATH:/opt/homebrew/bin` — and run `pm2`
+as a separate command. `docs/servers.md` has the detail; replacing the PATH is what broke `pm2` on an
+earlier deploy.
+
+## Then check
+
+`pm2 jlist`: `officer` is `online` and `restart_time` moved. Nothing else restarted by step 4 — note that
+step 3 restarts every plugin's sidecar on its own, which is the script's behaviour and worth recording
+rather than omitting.
+
+`scripts/deploy-claim.sh state 2026.08.21.04` says `done`.

deploy/README.md

@@ -0,0 +1,23 @@
+# `deploy/`
+
+One file per release, and one report per release.
+
+```
+deploy/
+├── 2026.08.21.01.md         the plan — what to run, in order, on which hosts
+└── 2026.08.21.01.report.md  what actually happened, written by the commander
+```
+
+The plan is written by hand before the release commit. The report is written by the deploy commander
+afterwards and pushed back here.
+
+**A release is a tag**, `deploy/<version>`, and nothing else triggers a deploy — not a commit message,
+not a branch. A tag cannot be written into prose by accident, which a code word can and which every
+document about deploying does.
+
+**What marks a release deployed is LOCAL**: `$OFFICER_ROOT/data/deploys/<version>.done`. Not the report.
+If the lock were the pushed report, a failed push would mean the release looked undeployed and the next
+tick would roll it out again — and pushes here fail for unrelated reasons. So the report is an artefact,
+valuable and not load-bearing.
+
+See `prompts/deploy-commander.md`.

docs/affordances.md

@@ -0,0 +1,164 @@
+# Affordances
+
+**Status: BUILT.** Agreed and built 2026-08-19. `src/servers/affordances/load.ts` reads them,
+`publish.ts` copies a plugin's in on install, and `claude-manager.ts` compiles them into every session.
+Two exist: `image-rendering` and gitea's. **Not built:** scoping (everything loads unconditionally) and
+delivery to opencode, which currently receives none.
+
+## What one is
+
+> **An affordance is something the surface on the other end of the conversation offers the agent —
+> declared by the code that implements it, and read by the agent at runtime.**
+>
+> Not what the agent can do. What the client will honour.
+
+## Why it needs to exist
+
+The platform hosts the agent and never introduces itself. An agent gets a shell, a filesystem and a
+prompt; everything about the surface it is speaking _through_ has to be smuggled in as prose somebody
+maintains by hand.
+
+The worked example is images. Officer's chat renders markdown images — `ChatImage` is wired into
+`react-markdown` — but an `<img>` carries no bearer token, so it can only load what the browser can fetch
+unauthenticated. A fresh agent knows none of that. It writes `![](/home/me/chart.png)`, nothing renders,
+and nothing tells it why.
+
+Three properties make prose the wrong container:
+
+- **It drifts.** The file describes a workaround; the code has a reason. Neither knows about the other.
+- **It is per-machine.** It lives in one repository, copied by hand.
+- **A plugin cannot contribute to it.** Install a plugin that adds a surface and no agent learns it
+  exists — which is the case that matters most, because a plugin is exactly a new surface arriving at
+  runtime.
+
+## The boundary
+
+Four neighbours, easily confused:
+
+|                | answers                                              |
+| -------------- | ---------------------------------------------------- |
+| **tool**       | what the agent can _do_ — it acts, something happens |
+| **permission** | what this _account_ may reach                        |
+| **skill**      | instructions the agent _follows_                     |
+| **affordance** | what the _client_ will render, accept or ignore      |
+
+Three tests settle almost everything.
+
+### 1. Would it change if the same agent talked to a different client?
+
+A phone app, a terminal, a webhook, a cron run with no human at all. If yes, it is an affordance.
+Rendering an image is; reading a file is not.
+
+### 2. Does the agent have to guess it today?
+
+If the honest answer is "it writes the markdown and hopes", that is a missing affordance.
+
+### 3. Can it be fixed instead? — then fix it
+
+**This is the one that is easy to get wrong, and the most expensive when you do.** An affordance is for
+something _irreducible_ about the surface. A defect is not irreducible. It is a defect.
+
+Both candidates written on 2026-08-19 failed this test:
+
+- **"Links open in the same tab, so warn the agent."** `a` was simply missing from `MD_COMPONENTS` in
+  `MessageBubble`. One line. Fixed.
+- **"`AskUserQuestion` never returns an answer, so don't wait for one."** True when the code comment was
+  written; already fixed by the time the affordance was drafted. It would have shipped a permanent
+  instruction to work around a bug that no longer existed.
+
+The failure mode is specific and worth naming: **an affordance describing a defect calcifies it.** Once
+there is a document explaining the workaround, the defect has a rationale, and the rationale outlives
+anyone's memory that it was ever a bug. The second example is worse than the first — it would have taught
+every future agent something false, in a file whose whole purpose is to be believed without checking.
+
+Image rendering earns its place precisely because it fails this test in the other direction: "the user's
+device fetches the URL" is how `<img>` works. No code change makes it untrue.
+
+## The rule that makes it worth having
+
+**An affordance is declared where it is implemented.** `ChatImage` declares that images render, because
+`ChatImage` is what makes them render. The moment it becomes a list maintained somewhere else, it is the
+hand-written `CLAUDE.md` problem again with a better name — and this time with a build step.
+
+Same posture as `permissions/registry.ts`, and for the same reason: one enumeration, owned by the code it
+describes, which cannot quietly disagree with reality.
+
+## The file
+
+Markdown with YAML frontmatter, like every other item in `capabilities/`. Frontmatter is for the
+PLATFORM; the body is for the agent.
+
+```markdown
+---
+name: Image rendering
+description: Use when you have an image to show the user in chat.
+version: 1
+---
+
+`![alt](https://…)` renders as an image. **The user's device fetches the URL**, so it must be
+reachable from there — a local path, `file://`, or anything behind authentication renders nothing,
+and fails silently.
+```
+
+| field         |                                                                                                                                                                                                                                                          |
+| ------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `name`        | becomes the heading in the compiled block, so the agent can index it                                                                                                                                                                                     |
+| `description` | when to reach for it. Costs the agent nothing — it is **not** sent                                                                                                                                                                                       |
+| `version`     | an INTEGER, not semver, matching `capabilities/agents/`. A plugin-contributed affordance sits near two version numbers meaning different things: this integer is the revision of _this text_, the plugin's semver is the plugin's. Do not reconcile them |
+
+A broken frontmatter header does not lose the body — `load.ts` falls back to the path for `name` rather
+than dropping the file. The agent needs the prose far more than the label, and a file that vanished over
+a stray colon would be very hard to notice.
+
+### Keep the body short
+
+Every affordance is loaded into every conversation, so it is charged to the context budget of every turn.
+State the constraint, name the silent failures, say what to do instead. Nothing else.
+
+Write it as **facts about the client**, not instructions to the agent. "The user's device fetches the URL"
+survives a redesign of the agent's behaviour; "always use a public URL" does not, and is one step removed
+from the reason.
+
+## Where they live
+
+```
+capabilities/affordances/
+  officer/core/          ← core, shipped by the setup script from the capabilities repository
+    image-rendering.md
+  officerdev/gitea/      ← contributed by the gitea plugin on install
+    gitea.md
+```
+
+`<publisher>/<plugin>`, taken from the manifest verbatim. `officer/core` is not a plugin: it is the
+platform's own. The two spellings mark two different things and should not be normalised into one.
+
+A plugin ships an `affordances/` directory in its repository; installing copies it in, uninstalling
+removes it — the same shape as `assets/` and `publishAssets`, deliberately, because two mechanisms for
+"the plugin ships some files" would drift.
+
+**The namespace is derived from where the file lands**, never from anything inside it — the same rule as
+`mountPrefix` and the notification catalogue.
+
+## How they reach the agent
+
+`claude-manager.ts` calls `affordanceBlock()` once per session and prepends the compiled block to the
+system prompt. Read at spawn rather than at module load, so a plugin installed an hour ago is live on the
+next message.
+
+`description` is deliberately **not** included in the compiled block: it answers "when should I reach for
+this", which only matters when an agent chooses what to load. Everything here is already loaded, so it
+would be paid for and unread.
+
+### Deliberately dumb, for now
+
+No filtering. Every affordance is compiled into every conversation. A `surface` field was drafted and
+dropped because it wore one name for two different axes: for image rendering it means _what the client
+renders_, for gitea _what tooling exists on the machine_. Picking the wrong axis and writing files against
+it costs more than loading two extra paragraphs.
+
+Fine at two files. Not fine at twenty. See `TODO.md`.
+
+### opencode gets none
+
+`affordanceBlock()` has exactly one caller and it is the Claude sidecar. An opencode turn runs with no
+affordances at all, so an agent driven through it still guesses at images. Tracked in `TODO.md`.

docs/command-hq.md

@@ -5,14 +5,30 @@ it, kept current as answers landed.
 
 **Status:** data model, the full REST API including both creation paths (web and `POST
 /commanders/new`), built and mounted 2026-08-18 (`bunx tsgo` clean, suite at its 7 pre-existing baseline
-failures — none of them this feature's). **Needs `bun db:push`** before any route answers with real data;
-without it every route throws a Postgres "relation does not exist" 500. A member also needs
-`confineUserTree` re-run once if their account predates tonight — see "A member could not reach it" under
-"Working directory". **Frontend not built yet on the web side.** The mobile side is further along —
-`apps/commandhq` exists and typechecks against this router (see `COMMS/command-hq/mobile.md`) — and as of
-this revision has what it asked for to create a commander, though it has not yet been exercised against a
-running server on either side. Every claim in this file is checked against the code, not aspirational —
-see "API reference" for what a client can call today.
+failures — none of them this feature's). `bun db:push` and a restart of `officer` have both since run
+against the dev machine, so the backend is live. A member still needs `confineUserTree` re-run once if
+their account predates that — see "A member could not reach it" under "Working directory".
+
+**Web frontend v1 built 2026-08-18**, same session: `/command-hq` and `/command-hq/:sessionId`, three
+panels per "Layout, v1" below (`command-hq-categories`, `command-hq-grid`, `command-hq-detail`) under
+`officerdev/src/apps/CommandHQ/`, registered in `AppRegistry.tsx`, a `CORE_DOCK_ITEMS` tile, and a
+`usePageTitle` rule. **Now exercised against a running server** — a real commander created, categorised,
+armed as a watcher and observed running, its auto-generated name traced back to the actual process on this
+machine (see "Auto-naming" and the `>/dev/null` finding below). The composer creates a commander the same
+way `/chat` always has (an ordinary turn with `context: 'command-hq'`); once the harness reports a
+transcript id the URL is replaced onto it, and if a category was selected when composing, a `PATCH` sets it
+client-side — `fileAsCommander` itself always files uncategorised, the same race `POST /commanders/new`
+resolves server-side for mobile, resolved here the client-side equivalent. Rename, category, disarm and
+delete all go through Command HQ's own routes, never `/chat`'s — the two are scoped to different
+directories. Delete asks for confirmation inline (grid card and detail bar both), same shape as `/chat`'s
+own session-row delete. Tag UI is still deliberately absent, per "Tags — infrastructure now, UI not
+specified" below.
+
+The mobile side is further along on live exercise — `apps/commandhq` exists and typechecks against this
+router (see `COMMS/command-hq/mobile.md`) — and as of this revision has what it asked for to create a
+commander, though it has not yet been exercised against a running server on either side either. Every claim
+in this file is checked against the code, not aspirational — see "API reference" for what a client can call
+today.
 
 ## API reference — build against this
 
@@ -21,7 +37,7 @@ to the caller's own `user_id` regardless. Router: `src/servers/api/command-hq/ro
 for exact request/response shapes; the table below is the map, not the full contract.
 
 | Method | Path                                 | Body                                  | Notes                                                                                                                                                                                                |
-| ------ | ------------------------------------ | ------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| ------ | ------------------------------------ | ------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
 | GET    | `/categories`                        | —                                     | `{ categories: [{id,name,createdAt}] }`                                                                                                                                                              |
 | POST   | `/categories`                        | `{name}`                              | 409 on a duplicate name for this user                                                                                                                                                                |
 | PATCH  | `/categories/:id`                    | `{name}`                              | rename                                                                                                                                                                                               |
@@ -32,21 +48,28 @@ for exact request/response shapes; the table below is the map, not the full cont
 | GET    | `/commanders[?category=<id>\|none]`  | —                                     | omitted = all, `none` = uncategorised bucket. See below                                                                                                                                              |
 | POST   | `/commanders/new`                    | `{prompt, displayText?, categoryId?}` | creates a commander; `prompt` required. `displayText ?? prompt` titles it, same precedence as the websocket. `-> {sessionId, cwd}` — `sessionId` is Claude's real transcript id, not a placeholder   |
 | GET    | `/commanders/:sessionId`             | —                                     | Command HQ's own metadata only — NOT the transcript, see below                                                                                                                                       |
+| GET    | `/commanders/:sessionId/turns`       | —                                     | `{turns: [...]}` — the raw per-turn cost ledger, oldest first. What "cost over time" plots from                                                                                                      |
+| GET    | `/commanders/:sessionId/fires`       | —                                     | `{fires: [...]}` — every trigger, oldest first. The full history behind `watcherLastFiredAt`                                                                                                         |
 | POST   | `/commanders`                        | `{sessionId, categoryId?}`            | files an EXISTING session; idempotent; 404 if the session isn't in the commanders directory. Not the creation path — see `/commanders/new`                                                           |
-| PATCH  | `/commanders/:sessionId`             | `{categoryId?, name?}`                | `name` proxies to `renameClaudeSession` — the same rename `/chat` uses                                                                                                                             |
+| PATCH  | `/commanders/:sessionId`             | `{categoryId?, name?, watcherArmed?}` | `name` proxies to `renameClaudeSession`; `watcherArmed: false` disarms (arming has no route — needs a live task id)                                                                                  |
+| DELETE | `/commanders/:sessionId`             | —                                     | deletes the TRANSCRIPT (`/clear` chain and all) via `deleteClaudeSession`, then the row. Not reversible, and not the same thing `/purge` does — see "The `/chat` ↔ `/command-hq` relationship" below |
 | POST   | `/purge`                             | —                                     | sweeps rows whose transcript no longer resolves; returns `{removed}`                                                                                                                                 |
 | POST   | `/commanders/:sessionId/tags`        | `{tagId}`                             |                                                                                                                                                                                                      |
 | DELETE | `/commanders/:sessionId/tags/:tagId` | —                                     |                                                                                                                                                                                                      |
 
 **A commander's conversation is `GET /chat/sessions/:id`, unchanged.** This router deliberately does not
-duplicate it — `GET /commanders/:sessionId` returns only `{sessionId, categoryId, tagIds, running,
-createdAt}`. A client wanting the full picture calls both.
+duplicate it — `GET /commanders/:sessionId` returns Command HQ's own metadata: `{sessionId, categoryId,
+tagIds, running, watcherArmed, watcherArmedAt, watcherStoppedAt, watcherLastFiredAt, cost, createdAt}`. A
+client wanting the full picture calls both; a client wanting the raw cost series calls `.../turns` as well
+— `cost` here is only the summary (`{turnCount, inputTokens, outputTokens, totalUsd}`), not the list.
 
-**`GET /commanders` is the grid endpoint** and returns the richest shape: every field `/chat/sessions`
-already returns per session (`id`, `title`, `cwd`, `createdAt`, `updatedAt`, `messageCount`, …) plus
-`categoryId`, `tagIds`, and `running` — a boolean, derived live from the same `task:started`/
-`task:notification` fold `/chat`'s own background-job tray uses, not a stored value. This is the field to
-poll if a client wants to know "did my watcher die" without opening it.
+**`GET /commanders` is the grid endpoint** and returns the richest per-card shape: every field
+`/chat/sessions` already returns per session (`id`, `title`, `cwd`, `createdAt`, `updatedAt`,
+`messageCount`, …) plus `categoryId`, `tagIds`, `running` (derived live from the same `task:started`/
+`task:notification` fold `/chat`'s own background-job tray uses), the watcher-arming fields, and the same
+`cost` summary the detail route returns — one aggregate query per card, the same cost class `running`
+already accepts for this grid. This is the field to poll if a client wants to know "did my watcher die"
+without opening it, or "what has this one cost me so far" without a second request.
 
 **A row whose transcript is gone is silently absent from every response.** Never a tombstone, never a 404
 with a special shape — it simply is not in the list. See "The /chat ↔ /command-hq relationship" below for
@@ -184,10 +207,14 @@ an accident of copying a plugin schema).
 One row per commander.
 
 | column               | notes                                                                                                                                                                                           |
-| ------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| -------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
 | `user_id`            | owner of the row, always the same account that owns the session                                                                                                                                 |
 | `session_id`         | **Claude's transcript uuid**, not Officer's `sessionId` — the same id `chat_session_models` keys on, for the same reason: everything else that joins to a session (models, events) uses this id |
 | `category_id`        | FK to `command_hq_categories`, nullable — one category per commander (see below)                                                                                                                |
+| `watcher_armed`      | should the platform keep re-arming this watcher on every fire — see "The watcher loop"                                                                                                          |
+| `watcher_task_id`    | the currently-tracked watcher task, nullable                                                                                                                                                    |
+| `watcher_armed_at`   | since when the CURRENT run has been going; set only on the false→true transition, untouched by re-arms                                                                                          |
+| `watcher_stopped_at` | when the current run ended; set on disarm, cleared on the next arm — paired with `watcher_armed_at`                                                                                             |
 | `created_at`         |                                                                                                                                                                                                 |
 
 No `name` column (see "Naming"). No `status` column, deliberately: liveness is **derived**, the same way
@@ -195,7 +222,44 @@ the background-job tray already derives it — fold `chat_session_events` for th
 minus `task:notification` (`collectRunningTasks` in `websocket.ts` is the existing implementation of exactly
 this fold; reuse it, do not reinvent it). A stored status would be a second source of truth that goes stale
 precisely in the case this feature is used for (see "Later" — silent death on sidecar restart) and nothing
-would ever correct it.
+would ever correct it. **No `watcher_last_fired_at` either, deliberately removed** — see
+`command_hq_watcher_fires` below; the owner's call, once cost tracking made clear that a single overwritten
+timestamp was losing information a real table should hold instead.
+
+### `command_hq_watcher_fires` — one row per trigger
+
+Added when cost tracking surfaced a real gap: the ORIGINAL design cached "last fired" as a single column on
+the commander row, which meant every new fire destroyed the previous one's timestamp. Fine for "when did it
+last fire", useless for "how many times, and what did each firing cost" — which is exactly what the owner
+asked for. So this became a real table instead of a column, and the commander row lost its cache entirely
+(the owner's call: derive `MAX(fired_at)` at read time rather than keep a copy that could only ever describe
+the same table this row also has to insert into).
+
+| column           | notes                                      |
+| ---------------- | ------------------------------------------ |
+| `session_row_id` | FK to `command_hq_sessions.id`             |
+| `task_id`        | which watcher task fired                   |
+| `fired_at`       | the only timestamp this row exists to hold |
+
+### `command_hq_turns` — the cost ledger
+
+One row per completed turn (every `result` message a commander's session produces, not only the ones that
+follow a fire — the setup conversation before a watcher is ever armed costs something too).
+
+| column           | notes                                                                                      |
+| ---------------- | ------------------------------------------------------------------------------------------ |
+| `session_row_id` | FK to `command_hq_sessions.id`                                                             |
+| `fire_id`        | FK to `command_hq_watcher_fires.id`, nullable — null for turns before the first-ever arm   |
+| `input_tokens`   |                                                                                            |
+| `output_tokens`  |                                                                                            |
+| `total_usd`      | `doublePrecision` — a monitoring number, not a billing ledger; see the schema file for why |
+| `created_at`     |                                                                                            |
+
+**Nothing is pre-summed.** Running total, cost over time, cost per trigger and turn count are all queries
+over this table — `getCommandHqCostSummary` (one `SUM`/`COUNT`), `listCommandHqTurns` (the raw series a
+client plots), grouped by `fire_id` for per-trigger. Turn count in particular needed no column at all:
+`COUNT(*)` on this table already answers it, the same "derive it" call the owner made for `watcher_armed_at`
+and the dropped `watcher_last_fired_at` — raised explicitly when the schema was being designed, not assumed.
 
 ### `command_hq_categories`
 
@@ -286,6 +350,12 @@ an independent record of what sessions exist. Consequences the owner confirmed e
 - A session **never moves the other direction**. There is no unfile path, and none is planned — a commander
   is a commander for life, and (given commanders live in their own cwd, never a `/chat` cwd) the two spaces
   never actually overlap.
+- **`DELETE /commanders/:sessionId`, added 2026-08-18, is not an unfile path — it is real deletion**, and
+  deliberately does not conflict with the bullet above: the commander does not become a `/chat` session, it
+  stops existing. `deleteClaudeSession` (the SAME function `DELETE /chat/sessions/:id` calls) removes the
+  transcript — a `/clear` chain and all — and `deleteCommandHqSession` removes the row. Once the confirm
+  step lands, this is not recoverable: there is no undo, no trash, and the row disappearing is a side
+  effect of the transcript being gone, not the other way around.
 
 ## Built on, mostly unchanged
 
@@ -329,6 +399,21 @@ fixed:**
    either false start: it only WATCHES for this happening and can say so if it doesn't (see "Built" below)
    — it does not drive the re-arm and does not respawn anything itself.
 
+**The evidence for step 3 is stronger than either side's first write-up said, and it is worth recording
+precisely rather than as "ran fine for a while" (`COMMS/command-hq/mobile-4.md`).** Both notes originally
+described the mobile side's own same-turn watcher as having run "for hours, this session." Andre's own
+count, checked directly: **at least 10–15 runs, spanning before and after the current mechanism existed,
+zero failures.** That is a different class of claim than "it worked tonight" — it is evidence against
+_revisiting_ the decision, not just evidence for having made it. Worth stating why that distinction
+matters: same-turn self-re-arm looks fragile on paper (it depends on the model correctly doing one specific
+thing as the last step of every turn, unsupervised, forever), so the instinct on reading the design cold is
+to add a safety net — a supervisor, a timeout, a platform-side check that re-arm actually happened. The
+record here is that a dozen-plus real runs across two different surrounding mechanisms never once needed
+one. It is also why the platform's role stays _noticing_ (`watcherArmed && !running`, "Deliberately not
+built" below) rather than _driving_: noticing is the right amount of machinery for a failure with zero
+observed instances, and building more than that would be solving a problem the evidence says does not
+occur.
+
 **Built:**
 
 - **A third dialog option**, `FireBehavior = 'stop' | 'handoff' | 'keep-going'` (was a `stopOnFire`
@@ -354,13 +439,19 @@ fixed:**
   Known gap: this is in-memory client state, so it does not survive a page refresh between clicking Stop
   and the notification arriving — an acceptable, narrow window, not attempted to be closed tonight.
 
-- **`command_hq_sessions` gained `watcherArmed`, `watcherTaskId`, `watcherLastFiredAt`.** Unlike `running`,
-  `watcherArmed` is allowed to be stored — it is the owner's INTENT ("keep re-arming this"), not a live
-  process fact the platform is reporting on, so nothing external can silently falsify it the way a sidecar
-  restart falsifies `running`. `watcherTaskId` is closer to that rejected shape and accepts the same known
-  gap `running` already lives with — see the schema file's docstring.
-  `armCommandHqWatcher` / `recordCommandHqWatcherFired` / `disarmCommandHqWatcher` in `queries.ts`; exposed
-  read-only on `GET /commanders` and `GET /commanders/:sessionId`, and `PATCH /commanders/:sessionId` accepts
+- **`command_hq_sessions` gained `watcherArmed`, `watcherTaskId`, `watcherArmedAt`, `watcherStoppedAt`.**
+  Unlike `running`, `watcherArmed` is allowed to be stored — it is the owner's INTENT ("keep re-arming
+  this"), not a live process fact the platform is reporting on, so nothing external can silently falsify it
+  the way a sidecar restart falsifies `running`. `watcherTaskId` is closer to that rejected shape and
+  accepts the same known gap `running` already lives with — see the schema file's docstring.
+  `watcherArmedAt`/`watcherStoppedAt` are the current-run boundary pair, added a turn later during the
+  cost-tracking pass once "since when has this been going" was a real question — see "Data model" above.
+  There is no `watcherLastFiredAt` column any more: that fact moved into `command_hq_watcher_fires` (its own
+  table, one row per trigger) once cost tracking needed per-fire attribution, not a single overwritten
+  timestamp. `armCommandHqWatcher` / `recordCommandHqWatcherFired` / `disarmCommandHqWatcher` /
+  `recordCommandHqTurn` in `queries.ts`; exposed read-only on `GET /commanders` and
+  `GET /commanders/:sessionId` (`watcherLastFiredAt` survives as a name in the JSON response — it is just
+  computed there, `MAX(fired_at)`, not read off a column), and `PATCH /commanders/:sessionId` accepts
   `{watcherArmed: false}` to disarm (arming itself has no route — it needs a live task id, which only a
   running turn has).
 - **`armWatcher` threaded end-to-end, client to server.** `BackgroundJobDialog` computes it (`watching &&
@@ -377,8 +468,10 @@ onFire === 'handoff'`) → `sendWithHiddenPrefix` → `QueuedPrompt.armWatcher`
   task starting is what the agent's own self-re-arm produces, so the tracked `watcherTaskId` is simply
   updated to it — the platform is not driving anything here, only keeping up with what the agent already
   did. On `task:notification`: if it matches the tracked `watcherTaskId` for an armed commander, record the
-  fire (`watcherLastFiredAt`). Both read `getCommandHqSession` live rather than trusting any closure-local
-  state, for the same reason as above.
+  fire — a new row in `command_hq_watcher_fires`, not a column write. Both read `getCommandHqSession` live
+  rather than trusting any closure-local state, for the same reason as above. On `result`: every completed
+  turn writes a row to `command_hq_turns`, attributed to the commander's most recent fire (`fire_id`, null
+  before the first-ever arm) — see "Data model" for why nothing here is pre-summed.
 - **Known, accepted imprecision:** the "already armed → treat the next task:started as the re-arm" rule
   cannot tell a genuine self-re-arm apart from an unrelated background task the agent happens to start in
   the same commander conversation. A commander is one watcher by design, so this is a narrow gap rather
@@ -389,10 +482,49 @@ the re-arm mechanism itself — landed as agent-driven, same-turn, after two rej
 — (4) arming state and the plumbing that watches for it.
 
 **Deliberately not built: the platform verifying re-arm happened.** Raised as a possible cheap addition —
-since the platform already knows `watcherLastFiredAt` and can already compute `running` from the same fold
-`GET /commanders` uses for every card, a client could flag "fired N ago, nothing has run since" without
-any new mechanism, purely by reading what already exists. Not built tonight; noted here so it is not lost,
-not because it was rejected.
+since the platform can already derive last-fired (`MAX(fired_at)` in `command_hq_watcher_fires`) and already
+computes `running` from the same fold `GET /commanders` uses for every card, a client could flag "fired N
+ago, nothing has run since" without any new mechanism, purely by reading what already exists. Not built
+tonight; noted here so it is not lost, not because it was rejected.
+
+**Auto-naming from the watcher's own description, added 2026-08-18 after the web frontend surfaced why it
+matters.** A fresh commander's title is whatever the first message was — the mechanics preamble, truncated,
+when that first message is a `handoff` background job. Expected, not a bug: that IS the first message, and
+Claude Code names sessions from it the same way it always has (see "Naming" above). But by the time the
+watcher's task actually STARTS, the harness has already given it a real description — `msg.description` on
+`task:started`, "Watch google.com reachability" for a ping watcher — and that reads far better than the
+preamble ever will. So the exact moment `pendingArmWatcher` is consumed (the FIRST arm, never a self-re-arm
+— see "known, accepted imprecision" above for why that distinction is already load-bearing here) now also
+calls `renameClaudeSession` with that description, resolving a fresh `ChatIdentity` via `resolveHomeDir`
+for the filesystem call. Best-effort, like everything else in this block: a failure is logged and does not
+touch `watcherArmed`.
+
+**Known gap, not closed:** the detail panel's title bar is fetched once on mount and does not live-refresh,
+so a commander open in the panel at the moment this fires keeps showing the old title until the next
+navigation or reload. The grid's cards do not have this problem — `GET /commanders` is already polled every
+15s. Flagged rather than fixed; worth a `refetch()` hook if it turns out to matter in practice.
+
+**"(no output yet)" on that same google.com watcher, traced to the actual running process on this
+machine rather than guessed at.** The tray's live-log-tail (`BackgroundTaskTray` → `useTaskDetail` →
+`GET /chat/tasks/:id` → `loadBackgroundTask`/`tailFile`) is a general `/chat` feature Command HQ gets for
+free — nothing to build there, and it was already working correctly. The task's `.output` file was
+genuinely empty: the agent had written `timeout 8 ping -c 1 -W 5 google.com >/dev/null 2>&1`, silencing
+`ping`'s own output entirely and only ever `echo`ing on a FAILED ping — correct per "let me know if it goes
+down," but it means a healthy watcher shows nothing at all, indistinguishable at a glance from a stuck one.
+Fixed at the source rather than in the tray: `background-job-prompt.ts` gained an unconditional rule (every
+job, not only watchers — number 4, 5 or 8 depending on which rules already preceded it in that particular
+prompt) telling the agent not to discard a command's own output. The tray already shows whatever reaches
+the log file live; the gap was that jobs were routinely told, by habit rather than by this prompt, to send
+that output to `/dev/null`.
+
+**A real bug, caught by tracing rather than by the compiler, worth stating plainly.** When the cost-tracking
+schema changed `recordCommandHqWatcherFired`'s signature from `(userId, sessionId)` to `(sessionRowId,
+taskId)`, the ONE existing call site was not updated in the same pass — and `bunx tsgo` stayed clean, because
+`userId: number` and `transcriptId: string` still satisfy `sessionRowId: number, taskId: string`
+structurally. It would have written a fire row for a commander keyed on the wrong id entirely (the caller's
+OWN user id, not the row it belonged to). Found by re-reading every call site the schema change touched
+after the fact, not by any tool — worth remembering the next time a signature changes shape rather than
+just its types: a compiler that stays quiet is not the same as a diff that is correct.
 
 ## Later — explicitly deferred, not forgotten
 
@@ -410,9 +542,17 @@ scope for tonight."_
   says running" against "pid check says running" is what turns a false "running" into a third, honest state
   ("vanished"). Cost matters here: `fuser` is a subprocess spawn per task, so this is a detail-pane check
   or a batched sweep on load, never a per-card check on every render of the grid.
-- **No durable job record.** A commander's only record of its own history is `chat_session_events`, which is
-  swept after 7 days. The owner does not care how many times a watcher fired or what each firing found — the
-  chat transcript being the record is intentional, not a gap to fill.
+- **No durable job record — REVISED.** True when written; no longer true for cost. The owner still does not
+  care how many times a watcher fired or WHAT each firing found (the chat transcript remains the only record
+  of content, and that is intentional), but does want per-turn token/dollar cost durably recorded —
+  `command_hq_turns` and `command_hq_watcher_fires`, added same session. See "The watcher loop" and the data
+  model section above.
+- **Cost-based limits — explicitly out of scope for tonight, said directly by the owner.** "If this agent
+  reaches a certain cost in tokens, stop it, or compact the conversation and resend, or clear and resend."
+  Not built, and deliberately not designed around either: the ledger this session added
+  (`command_hq_turns`) is exactly the raw material a limit would need (`SUM` since a fire, or since armed),
+  so nothing here forecloses it — but which action, at what threshold, and whether it is a per-commander
+  setting or a global one, is all undecided.
 - **Detail pane contents beyond "the conversation".** The owner asked for logs, status, and transcript
   together in a split view; scoped down to "click a card, see its details" for v1, with the specifics
   (separate log tail vs transcript vs live status) left open. `/chat/sessions/:id` (transcript) and
@@ -440,5 +580,8 @@ scope for tonight."_
 ## Still open
 
 Nothing blocking as of this revision. `POST /command-hq/commanders/new` is built — see "How a commander is
-created" — with `prompt` required, confirmed with the owner 2026-08-18. Neither path (web or mobile) has
-been exercised against a running server yet; that is the next real test, not a design question.
+created" — with `prompt` required, confirmed with the owner 2026-08-18. The web frontend v1 is built too
+(see the status line at the top). **Neither client has actually been clicked through against a running
+server yet** — that is the next real test, not a design question, and it is the thing most likely to turn
+up something this file gets wrong: the composer's redirect-on-real-id, the watcher-disarm button, the
+category-race PATCH, none of them have run once outside a typecheck.

docs/email-recall.md

@@ -0,0 +1,159 @@
+# Recall over your own correspondence
+
+**Status: DESIGN, NOT BUILT.** Sketched 2026-08-20, for the email plugin.
+
+## The goal, which is not the technology
+
+Someone is on a phone call with their ex-wife. She says something about the weekend arrangement. He has
+eight years of email with her, and he needs the fact **now** — not after a search, not after reading
+twenty messages.
+
+That is the whole requirement. Everything below is in service of it, and any part of this that does not
+make that moment faster or more accurate is decoration.
+
+The user will never write a query. They ask the agent a question in their own words, and the agent picks
+its tools. So the design question is not "what can the user search" — it is **"what does the agent have
+to work with, and which of its answers can be trusted"**.
+
+## Two mechanisms, because there are two kinds of question
+
+The examples that prompted this split cleanly, and they need different machinery:
+
+| the question                                                     | what it needs                         |
+| ---------------------------------------------------------------- | ------------------------------------- |
+| _"how many times did she ask me to break the weekend agreement"_ | **exhaustion** — every message judged |
+| _"what did she say about the school run"_                        | **retrieval** — the best few, fast    |
+
+```
+COMPILE    exhaustive, once, expensive   → counting, timelines, "show me all of them"
+RETRIEVE   cheap, approximate, always    → exploring, mid-call, "what did she say about…"
+```
+
+They are not competing designs. The agent should have both and choose.
+
+### Counting cannot be done by searching
+
+This is the load-bearing point, and it is the one most likely to be skipped.
+
+An agent asked "how many times" will search, find some, and answer with a number. That number is bounded
+by the phrasings it happened to think of. **"I found four" is indistinguishable from "there were nine"**,
+and nothing in the output signals which it was.
+
+Vector search does not fix this. Top-k also misses — it just misses differently. Ask for the top 20 and
+the 21st is silently gone.
+
+The only honest way to count is to judge every message once and store the verdict. Then counting is
+`SELECT count(*)`: exact, instant, and re-askable a hundred ways without re-reading anything.
+
+**A stored verdict is a judgement a model made about someone's divorce.** It records what it was asked,
+which message it read, and when — so the answer can be inspected, disbelieved and redone. A number with
+no provenance is worse than no number, because it will be quoted.
+
+### Retrieval is where embeddings earn their place
+
+Keyword search finds `school run`. It will never find it from _"the pickup arrangement"_ or _"when I
+collect her on Fridays"_. Over eight years, the same thing is described twenty ways and never once in the
+words you would search for.
+
+An agent closes some of that gap by itself — it knows those are synonyms and can fire several searches.
+But that is **guess-bounded**: it finds the phrasings it thought of, and cannot know the ones it did not.
+
+Two things decide it in favour of embeddings anyway:
+
+- **Latency.** Five keyword searches and twenty candidate reads is thirty seconds. On a live call that is
+  a lost moment. A vector query is milliseconds plus one read.
+- **What comes later.** WhatsApp and audio transcripts share no vocabulary, schema or structure with
+  email. Embeddings collapse all of it into one space. ASR output is noisy in a way keyword search is
+  brittle against and embeddings are not.
+
+**But we do not have WhatsApp or transcripts.** They shape exactly one decision — see the schema — and
+nothing else. Build against email.
+
+## It has to run locally
+
+Every message passes through an embedding model. Doing that through an API means uploading eight years of
+private correspondence to a third party.
+
+For this content that is not a trade-off, it is a disqualification. **Local embedding is a requirement,
+not a preference.**
+
+Today there is no ollama and no embedding code on any of these machines. **That is the real new
+dependency — not the database.**
+
+## Not a vector database
+
+`sqlite-vec` is an extension to the SQLite already in use. Same file as `emails.db`, no service, no
+sidecar, no Docker, no port, nothing to supervise or back up separately.
+
+For a plugin that already owns that file, this is **a table, not an architecture**. A separate vector
+database would add an operational surface for a corpus of 19,000 rows, which is small — vector stores
+begin earning their keep in the millions. What is being bought here is recall on paraphrase and speed on
+a phone call, not scale.
+
+## The genuinely hard part is chunking
+
+Not the vectors. **Quoted reply chains.**
+
+The same paragraph exists in forty messages of one thread. Embed naively and every query returns that
+thread forty times, and it will feel worse than what exists today.
+
+So: strip quoted history, thread properly, embed the new text of each message. `emails` already has
+`thread_id`. This is the work.
+
+Two other email-specific traps measured on the real 19,097-message store:
+
+- **Bodies are ~0.5 GB of HTML.** Embed the text, not the markup.
+- **`labels` is a comma-separated string**, and `read` is populated on 93 rows. Neither is a filter to
+  build on. See the `email-db` skill.
+
+## Schema sketch
+
+Deliberately **not** email-shaped. That is the one concession to a future that does not exist yet, and it
+costs nothing now:
+
+```
+recall_chunks
+  id            integer pk
+  source        text      -- 'email' today. 'whatsapp', 'transcript' later.
+  source_id     text      -- the message id in ITS store — no foreign key across sources
+  account       text      -- which mailbox/number this came from
+  thread_id     text      -- for dedupe and for showing context
+  author        text      -- who said it
+  said_at       timestamp
+  text          text      -- quoted history stripped
+  embedding     float[]   -- sqlite-vec
+
+recall_verdicts             -- the COMPILE half
+  id            integer pk
+  question      text      -- what the model was asked, verbatim
+  source        text
+  source_id     text
+  verdict       boolean
+  reasoning     text
+  model         text      -- which model, so a re-run is comparable
+  judged_at     timestamp
+```
+
+Everything else assumes email until something else exists.
+
+## Build order
+
+1. **Chunking and dedupe**, with no embeddings at all. Strip quotes, thread, count what is left. If this
+   is wrong nothing after it can be right, and it is testable on its own.
+2. **A local embedding model**, and a decision about which. The new dependency.
+3. **`sqlite-vec` + backfill** of the existing corpus. One-off, slow, resumable.
+4. **A retrieval tool** the agent can call, returning chunks with their message ids so it can go read the
+   original.
+5. **The compile half** — a verdict table and a way to ask "judge every message against this question".
+6. Multi-source, if and when WhatsApp or transcripts exist.
+
+Steps 1 and 5 are useful without any of the others. Step 5 in particular answers the questions that
+prompted this, and needs no vectors at all.
+
+## Open questions
+
+- **Which local model**, and how big. Quality against a laptop-or-server-sized footprint.
+- **Re-embedding.** Changing model means re-embedding everything. Store which model produced each vector.
+- **Retention and deletion.** A deleted email should lose its chunks and its verdicts. Nothing here
+  currently cascades.
+- **Who can query it.** This is the most sensitive data in the platform, and it is per-account.

docs/find-my-device.md

@@ -0,0 +1,269 @@
+# Find My Device, on Officer
+
+**Status: DESIGN, NOT BUILT.** Sketched 2026-08-19, the same day push notifications went in — which is
+not a coincidence, because push is the piece that makes this possible at all.
+
+## What this is, and what it will never be
+
+**The pitch is not "replace Find My". It is "know where your devices are without Apple or Google being
+the ones who know."** Location history on your own server, on your own tailnet, under your own retention
+rule. That part is mostly assembly.
+
+**The part that is genuinely impossible: offline finding.** Apple's magic is not the map. It is that a
+lost iPhone with no network still broadcasts an encrypted identity over Bluetooth, and roughly a billion
+strangers' devices relay it home. That is a crowd, and you do not have one. Nor does any self-hosted
+system, ever.
+
+So the honest boundary:
+
+|                       |                                                                    |
+| --------------------- | ------------------------------------------------------------------ |
+| Device has network    | **works** — current position, on demand or on a schedule           |
+| Device is off-grid    | **last known position and nothing more**                           |
+| Device is powered off | **nothing.** Apple's is silicon-level; there is no software answer |
+| Device is wiped       | **nothing**                                                        |
+
+A tracker that answers "where was it last, and when" covers the overwhelmingly common case — left in a
+taxi, lost in the house, stolen and still online. It does not cover the case Find My is famous for.
+
+Anyone building this should be clear about that before starting, because the gap is not a to-do item.
+
+## What already exists
+
+More than expected, which is why this is worth doing now rather than someday.
+
+- **The wake channel.** You cannot reach a sleeping phone except through APNs/FCM. That is built,
+  `officer/core:*` codes and all, and a real device is registered. See `docs/push-notifications.md`.
+- **A map.** `plugins/photos/web/MapSection.tsx` — maplibre-gl with markers, clustering, bounds-fitting,
+  already working and already pinned to a version that behaves. Read its header comments before writing a
+  second map; they record two traps that cost someone a day.
+- **Device presence, free.** offscale/headscale already reports `online` and `lastSeen` per node
+  (`plugins/offscale/sidecar/normalize.ts`). "Is my laptop awake, and when was it last seen" needs no new
+  code at all — and it covers laptops and desktops, which the phone-app approach never will.
+- **The tailnet**, so a device reports home without anything being exposed publicly.
+- **The hub app**, already installed, already registered, already receiving pushes.
+
+What is missing is a table, an endpoint, some app code and a panel.
+
+## Architecture
+
+Two paths, and they are not the same shape.
+
+```
+REPORT (device → server, most of the time)
+   phone ──POST /api/find-my/report──→ platform ──→ device_locations
+
+COMMAND (server → device, rarely)
+   panel ──→ officer-notify ──push──→ hub app ──→ acts, then REPORTs
+```
+
+**Everything of value flows on the report path.** The command path exists only for "where is it _right
+now_", and it is the unreliable one — so the design must not depend on it.
+
+### The report path
+
+The device posts a fix. That is the whole thing. It works when the app is foregrounded, when iOS hands it
+a background location update, and when a queued batch drains after a period offline.
+
+**Batching matters.** A phone that has been off-grid for six hours has six hours of fixes. The endpoint
+takes an array, and every fix carries its OWN timestamp from the device clock — which is why the schema
+below separates `recorded_at` from `received_at`. Collapsing them loses the distinction between "you were
+there at 14:00" and "the phone told me at 20:00", and that distinction is the entire value of a history.
+
+Device clocks are wrong sometimes. Store both, trust neither blindly, and let the UI show `recorded_at`
+while any "is this stale" logic uses the pair.
+
+### The command path, and why it is the weak one
+
+To make a sleeping phone report _now_, you push. The options, in order of preference:
+
+1. **Silent push** (`content-available: 1`), app wakes, takes a fix, posts it. The right shape — and
+   **Apple throttles these and does not guarantee delivery.** Already established in
+   `docs/push-notifications.md`, and it is the reason silent pushes were rejected for notifications.
+2. **A visible push the user taps.** Reliable, but requires a human at the other end, which defeats the
+   purpose when the phone is in a taxi.
+3. **A Notification Service Extension taking the fix on delivery.** Attractive because the NSE runs on a
+   normal, reliable, high-priority push. **Unverified** — location access from an extension is possible
+   in principle but unusual, and nobody here has tried it. Worth an afternoon before committing.
+
+**Design conclusion: treat on-demand as a bonus, never as the mechanism.** If the background reporting is
+good, "where is it now" is answered by "where was it 4 minutes ago", and that is enough for almost every
+real situation.
+
+## Device identity — do NOT key this on the push token
+
+The most likely serious mistake in this design, so it goes above the schema.
+
+`push_devices` is unique on `(token, bundle_id)`, and **APNs tokens rotate**. A location history keyed to
+a rotating token silently forks into two devices, and the split looks like the user bought a new phone.
+The registration path is already an upsert precisely because tokens churn.
+
+So a device needs a **stable identity generated once on the device at first launch** and persisted across
+token rotations, app updates and re-registrations — a UUID in the keychain, not derived from anything
+Apple gives you (`identifierForVendor` resets when all your apps are uninstalled).
+
+That means a `devices` table that is NOT `push_devices`:
+
+- `push_devices` answers "where do I send a push" — churny, plural per device, token-keyed.
+- `devices` answers "which physical thing is this" — stable, one row per device, and what location,
+  battery and presence hang off.
+
+A `push_devices.device_id` foreign key joins them. This also makes a device that has _never_ registered
+for push (a laptop reporting over the tailnet) representable, which the push table alone cannot do.
+
+## Schema sketch
+
+```
+devices
+  id                serial
+  user_id           fk users
+  device_uuid       text      -- generated on-device, stable forever. unique with user_id
+  name              text      -- "André's iPhone", user-editable
+  platform          text      -- ios | android | macos | linux
+  model             text      -- reported, cosmetic
+  created_at, last_seen_at
+
+device_locations
+  id                bigserial
+  device_id         fk devices
+  latitude          double precision
+  longitude         double precision
+  accuracy_m        real          -- ALWAYS store it; a 3km fix rendered as a pin is a lie
+  altitude_m        real          nullable
+  speed_mps         real          nullable
+  heading_deg       real          nullable
+  battery_pct       smallint      nullable
+  is_charging       boolean       nullable
+  source            text          -- significant | ondemand | visit | manual | geofence
+  recorded_at       timestamptz   -- the DEVICE clock, when the fix was taken
+  received_at       timestamptz   -- the SERVER clock, when it arrived
+```
+
+Index on `(device_id, recorded_at desc)` — every query is "latest for this device" or "this device
+between two times", and both want that order.
+
+**`accuracy_m` is not optional.** iOS returns fixes with 3km accuracy when it is being cheap about it, and
+a UI that draws those as a confident pin will send someone to the wrong street. Draw the circle.
+
+**Battery belongs here**, not in a separate table. "Last seen at 4%" is often the single most useful fact
+on the screen — it tells you whether to expect another fix at all.
+
+## The platforms
+
+### iOS
+
+- **`Always` location authorization** plus the `location` background mode and
+  `allowsBackgroundLocationUpdates`. `When In Use` is not enough; a find-my that only works while you are
+  looking at it is not one.
+- **Significant-change monitoring** is the floor: roughly cell-tower granularity, very cheap, survives app
+  termination, and iOS relaunches the app to deliver it. This should be the primary source.
+- **The Visits API** is cheaper still and gives arrival/departure at meaningful places. Good for history,
+  useless for "it is moving right now".
+- **Continuous GPS is not an option** for background use. It will eat the battery and iOS will fight you.
+- The app is self-distributed, so App Store review of background location is not the obstacle it would
+  otherwise be.
+
+### Android and GrapheneOS
+
+**Easier, and there is a pleasing irony here.** The foreground service holding a tailnet socket — the
+approach `docs/push-notifications.md` records as the _fallback_ for GrapheneOS having no Play Services —
+is close to ideal for this. It is always connected, so it needs no push to be told to report, and the
+persistent notification that made it unattractive for notifications is a normal thing for a location
+service to have.
+
+So on the phone where push is hardest, find-my is easiest. One foreground service can serve both.
+
+### Laptops and desktops
+
+Do not write a client. **headscale already knows** `online` and `lastSeen` per node, and the endpoint
+addresses give a coarse network location. That is most of what anyone wants for a laptop, and it exists
+today.
+
+## Commands, beyond locating
+
+Each needs app code — **the notification catalogue makes new notifications free, not new behaviours.**
+
+- **Locate now** — the weak path above.
+- **Play a sound.** Genuinely useful, and the "lost in the house" case is the most common of all. Needs an
+  audio session category that overrides the silent switch.
+- **Mark as lost** — show a message and a contact number on the lock screen. Needs MDM on iOS.
+- **Lock / erase.** Real MDM. Officer already serves iOS configuration profiles (`claimIosProfile` in
+  `api/dav/ios-profile`), so this is a project rather than an absurdity — but it is a project, and it puts
+  a device-wipe capability behind whatever authenticates to it. Think hard before building it.
+
+## Retention, and the thing to decide first
+
+**A location table on your own server is a complete movement record of your life.** Apple deliberately
+cannot build one; you would be choosing to. That is a legitimate choice, and it is the whole point of
+self-hosting — but it means the retention rule is a design decision, not a settings screen to add later.
+
+Decide before the table exists:
+
+- **A default retention window** with an actual delete job, not a documented intention. 30 days is a
+  reasonable starting point: enough to answer "where was I on Tuesday", short enough that a compromise is
+  not a biography.
+- **A pause switch** that stops reporting, visible on the device, honest about what it does.
+- **What a backup contains.** The Postgres backup now includes a movement history. Anyone who copies a
+  dump copies that.
+- **Who can see whom.** Location is per-user and every query filters by `user_id` — but **the owner
+  bypasses every permission check by design** (`permissions/authorize.ts`). So on a family server the
+  owner can see a member's location, structurally, unless location queries are given an explicit
+  exception. That is a policy decision that has to be made deliberately, because the default falls out of
+  the existing architecture rather than out of anyone's intent.
+
+That last one is the sharpest question in this document.
+
+## Permissions
+
+A new registry entry, `kind: 'app'` so it is grantable per role, with `personal` covering the report
+endpoint so a member's own device can post its own fixes at read level — the same shape (and the same
+trap) as `notify`, where the `personal` paths must be the real post-prefix paths. See
+`permissions/registry.ts` and the `notify` regression test.
+
+`assertPermissionTotality` will refuse to boot without the entry, which is the intended behaviour.
+
+## The web surface
+
+Per the frontend route conventions in `CLAUDE.md`: a `/find-my` route rendering `WorkspaceView`, with
+panels rather than a standalone screen.
+
+- **Device list** — rows are real `<Link>`s to `/find-my?selected=<deviceId>`, never onClick closures.
+- **Map panel** — reads `?selected` from the URL, not from a channel. Reuse the maplibre setup from
+  `plugins/photos/web/MapSection.tsx`.
+- **History/timeline panel** — a date range, also in the URL.
+
+Selection is the URL. `docs/navigation-audit.md` is the authority and names the anti-pattern this must
+avoid.
+
+## Core or plugin
+
+**Probably a plugin**, unlike notify. The argument for notify being core was that its producers are the
+platform's own — the job queue, an agent turn ending. Nothing in the platform produces a location. It is a
+self-contained feature with its own table, its own panel and its own app support, which is the plugin
+shape exactly.
+
+The counter-argument is that it needs a stable `devices` table that other things will want (push, presence,
+per-device settings), and a core concept living in a plugin is how the notify mess started. **Split it:**
+`devices` is core, find-my is a plugin that hangs off it.
+
+## Build order
+
+1. **`devices` table, core**, with the on-device UUID, and `push_devices.device_id` joined to it. This is
+   worth doing even if find-my never ships — the push registry currently has no stable device identity.
+2. **Report endpoint + `device_locations`.** Batched, both timestamps, accuracy required.
+3. **Hub app: significant-change reporting.** The floor, and on its own already a working find-my.
+4. **The panel** — device list, map, last-known position.
+5. **Retention job.** Before real data accumulates, not after.
+6. **Play a sound.** The highest value-per-effort command by a distance.
+7. **On-demand locate**, accepting it is best-effort.
+8. Android foreground service, folded in with the push fallback if that is ever built.
+9. MDM tier — separately, and only if genuinely wanted.
+
+## Open questions
+
+- **Owner visibility of a member's location.** The default falls out of the permission architecture rather
+  than out of intent. Decide it explicitly.
+- **Retention default**, and whether it is per-device.
+- **Can an NSE take a location fix?** Would make on-demand reliable. Unverified, worth an afternoon.
+- **Geofences** — "tell me when the phone leaves the house" is cheap on both platforms and is arguably
+  more useful than on-demand locating. Not in the build order above; it might deserve to be.

docs/multi-server-push.md

@@ -0,0 +1,112 @@
+# Notifications from more than one server
+
+**Design, not built.** Written 2026-08-21, for the single-sign-on work.
+
+The good news first, because it is the part that needed checking and it holds: **one Officer app can
+receive notifications from every server you are signed into, and almost nothing has to change to allow
+it.** One field is missing. This is what it is and why it matters more than it looks.
+
+## Why it works at all
+
+An APNs auth key (`.p8`) belongs to the **developer account**, not to a machine. There is no server
+binding in it. So every Officer install can hold the same key, mint its own provider token from it, and
+push to the same app.
+
+The device token belongs to the **app**, not to a server either. One Officer app has one token, whichever
+servers it talks to.
+
+So N servers pushing to one phone needs no coordination between them. Each just needs the token, which the
+app posts when it signs in — which the multi-server manager is doing anyway.
+
+Two facts about the shared key, neither blocking:
+
+- **Apple allows 2 active auth keys per account.** Sharing one across servers is the intended pattern;
+  one key per server is not an option past two.
+- **Revoking it kills push everywhere**, with no way to tell which machine leaked it. That is the real
+  cost of copying it to ten boxes, and the reason it belongs in each one's secret store rather than
+  in `.env`.
+
+## The missing field
+
+A payload today:
+
+```json
+{
+  "aps": { "alert": { "title": "Officer" }, "thread-id": "chat:abc", "mutable-content": 1 },
+  "officer": { "c": "officer/core:chat.turn-ended", "id": 42, "mv": 1873 }
+}
+```
+
+There is no origin. With one server that is correct and minimal. With five it breaks three things, and
+only the first is obvious.
+
+**1. The deep link opens the wrong thing.** `officer-chat://session/{id}` — that id is meaningful only on
+the server that issued it. The app has no way to know which, so it picks one. Sometimes it opens nothing;
+sometimes it opens somebody else's conversation with a matching id, which is worse.
+
+**2. `mv` starts a resync fight.** `mv` is a content hash of **that server's** catalogue, and servers with
+different plugins have genuinely different catalogues. The client's rule is `mv !== version → resync`.
+With one stored catalogue and five servers, every push from server B invalidates the copy synced from
+server A, and the app resyncs forever, always against whichever pushed last.
+
+This is the one that would look like a mysterious battery and data problem rather than a bug.
+
+**3. Threading collapses across servers.** `chat:{id}` groups per conversation. Two servers' ids
+coinciding is unlikely with UUIDs and not impossible, and "these two unrelated chats are one banner" is a
+confusing failure to diagnose.
+
+## The change
+
+**A server identifier in the `officer` block**, and the app keeping **one catalogue per server** keyed on
+it:
+
+```json
+"officer": { "s": "<server-id>", "c": "officer/core:chat.turn-ended", "id": 42, "mv": 1873 }
+```
+
+`mv` then means "version of that server's catalogue", threading becomes `<server>:chat:{id}`, and the deep
+link resolves against a known server.
+
+### Where the identifier comes from is an SSO decision, not a notifications one
+
+This is the part to settle tomorrow, and it belongs to the sign-on design rather than to this file.
+
+The multi-server manager already needs a key for each server — to store its API key against, to label it
+in a switcher, to route a deep link. **Whatever that key is, notifications should adopt it.** Inventing a
+second identity here means two ids for one server, drifting the first time somebody renames a host.
+
+It has to survive a hostname change, be stable across reinstalls of the app, and not be a secret — it
+travels in a push payload, which is the one place content is deliberately withheld. A server-generated
+opaque id created at bootstrap fits; a URL does not, because URLs change.
+
+## What stays per server, correctly
+
+Worth stating so nobody tries to unify it:
+
+- **`push_devices`** — each server holds its own registrations. The same phone appears once per server.
+- **`chat_notification_prefs`** — muting a conversation is per server, because the conversation is.
+- **The catalogue** — a server's codes come from its own plugins. Two installs with different plugins
+  legitimately have different vocabularies.
+
+None of that wants a shared store. The app is what joins them, which is the same shape as the sign-on
+work: the app holds the list, each server holds its own truth.
+
+## Current state, verified 2026-08-21
+
+|                        | alpha          | officer-dev |
+| ---------------------- | -------------- | ----------- |
+| iOS devices registered | 2              | 0           |
+| APNs credentials       | partly present | absent      |
+
+So only alpha can push today, and notifications seen elsewhere are the in-app bell rather than push. That
+is not a limitation of the design — it is two things absent on the other machines.
+
+## Order of work
+
+1. **Settle the server identifier** with the SSO design. Everything else waits on it, and it is a
+   half-hour conversation rather than a build.
+2. Add `s` to the payload; keep it optional so an older app ignores it.
+3. Per-server catalogue storage in the app, keyed on `s`. **This is the one that must land before a
+   second server starts pushing**, or the resync fight above begins.
+4. Thread ids become `<server>:chat:{id}`.
+5. Copy the `.p8` into the other servers' secret stores, and have the app register its token per server.

docs/plugin-dependency-scope.md

@@ -0,0 +1,91 @@
+# Plugin dependency scope
+
+**Built 2026-08-21.** A plugin declares WHO has to be able to run each host binary, and the platform
+installs and checks accordingly.
+
+## The bug
+
+`os-deps.ts` probes a dependency by asking the owner's login shell where the binary is. That is the right
+question for most plugins and the wrong one for some, and nothing distinguished them — a per-user answer
+was reported as a fact about the machine.
+
+It matters because `confined` plugins do their work as **the member's own Linux user**, in their own home.
+A binary in the owner's `~/.local/bin` is not on that user's PATH, and reaching it means traversing the
+owner's home directory.
+
+On alpha it happened to work. `/home/pastilhas` is mode **751** — world-traversable, not readable — so an
+absolute path resolved for members. Mode 750, which is the more usual default, would have broken every
+member's binary with nothing reporting anything: the install would still say satisfied, because the
+install only ever asked the owner.
+
+Working by accident is the part worth fixing. Nobody chose 751.
+
+## The field
+
+```ts
+osDependencies: [{ binary: 'iris', reason: '…', packages: {…}, scope: 'shared' }]
+```
+
+`owner` is the default and stays right for most things — anything an `execution` plugin uses runs as the
+owner in the owner's home by definition. `shared` says every account with a Linux user must be able to run
+it, which changes both the install target and the check.
+
+An unrecognised value is a manifest error rather than a silent fallback to `owner`, because the fallback
+is the permissive one.
+
+## What changes when it is `shared`
+
+**Where it is looked for.** System directories only — `/usr/local/bin`, `/usr/bin`, `/bin`, and the sbin
+trio. Deliberately not the owner's login shell: that shell finds the copy in the owner's home and reports
+the dependency satisfied, which is the exact false answer this exists to stop.
+
+**What counts as reachable.** Two things, and checking only the first is how this hid:
+
+- the file is world-executable (`0o001` — group is not enough, members are not in the owner's group)
+- **every directory above it is traversable**, walked to the root
+
+A binary at 755 inside a home at 700 is unreachable, and `stat` on the file alone says it is fine.
+
+**Where a release binary is installed.** `/usr/local/bin`, via `sudo -n install -m 0755`. The download is
+still staged in the owner's own space — fetching straight into a root-owned directory would mean running
+the download as root, which is a bigger ask than moving the result. `-n` so a plugin install never blocks
+on an invisible password prompt.
+
+**What blocks an install.** A shared dependency that is present and working but unreachable gets its own
+state, `unreachable`, and its own refusal — checked before everything else, because no package manager
+fixes it. The binary is installed; what is wrong is where it lives. Reinstalling the package would report
+success and change nothing.
+
+## What this deliberately does not do
+
+**It does not gate execution on a grant.** A binary in `/usr/local/bin` is runnable by any account with a
+shell, whether or not they hold the plugin's permission. The kernel does not know about `role_permissions`
+and never will.
+
+That is the same boundary `ffmpeg` and everything else in `/usr/bin` already has, and it is the right one:
+the plugin's **data** is gated by the permission system, its **executable** is not. For iris it is benign —
+a member who runs it captures their own `$HOME` profile and their own sessions, which is the isolation the
+wrapper already depends on.
+
+If a binary must not merely be _reachable_ but _unusable_ without a grant, `scope` is not the mechanism.
+That needs the plugin's own surface behind the permission gate, and the binary not on anyone's PATH.
+
+## Iris
+
+The motivating case, and still the only `shared` one. `PLUGIN.md` had already recorded it as open item 3 —
+"the binary lands in the owner's `~/.local/bin`, and a confined member's agent cannot read it there" — but
+that had not been connected to the probe being owner-scoped, so the install reported success either way.
+
+`iris` and `iris-browser` are built by the sidecar rather than fetched, so declaring `scope: 'shared'`
+makes the probe honest about who can run them; moving the sidecar's build output into `/usr/local/bin` is
+the remaining half, and it is a change to the sidecar rather than to this.
+
+## Status
+
+|       |                                                                            |
+| ----- | -------------------------------------------------------------------------- |
+| Built | `scope` on a dependency, validated, defaulting to `owner`                  |
+| Built | shared probing on system paths, with a traversal walk to the root          |
+| Built | shared release binaries installed to `/usr/local/bin` via `sudo -n`        |
+| Built | an unreachable shared dependency refuses the install, naming the directory |
+| Open  | iris's sidecar still builds into `~/.local/bin` — see above                |

docs/plugin-route-authorization.md

@@ -0,0 +1,140 @@
+# Plugin route authorization
+
+**Platform half built 2026-08-21.** The gate enforces `ownerOnly`, the manifest accepts it, and the proxy
+forwards `X-Officer-Owner` / `X-Officer-Level`. **No plugin declares it yet** — jellyfin's `/_config` is
+still reachable by anyone granted the plugin, and cannot be locked until the credential question at the
+bottom of this file is answered.
+
+The question: a plugin's **settings** routes should be reachable only by the owner, even when the plugin
+itself is granted to other accounts. Jellyfin is the case that raised it — connecting a server, storing
+its admin token and provisioning accounts is not something a Member should reach, but Jellyfin as a whole
+is exactly the sort of plugin a Member should have.
+
+## The shape we agreed
+
+Authorization stays where the platform already puts it — the gate answers "may this account reach this
+plugin, and at what level" — and the plugin decides what that means **inside** its own surface. That is
+not a new principle. `src/servers/plugins/manifest.ts` already says it, in the comment refusing an
+`ownerOnly` field on a permission:
+
+> Every plugin permission is grantable, per role, at read or write … Finer visibility — whose rows a
+> member sees, what a read means for this plugin's data — is the PLUGIN's business and lives in its own
+> queries.
+
+"Which of my routes are administrative" is the same kind of question as "whose rows does a read return".
+It belongs to the plugin, because only the plugin knows that `/_config` is settings and `/_jf` is bytes.
+
+So: **the platform tells the plugin who is asking; the plugin decides what they may do.**
+
+### What gets passed, and what does not
+
+Today the proxy sends exactly one fact (`src/servers/sidecar/create-proxy.ts`):
+
+```
+X-Officer-User: <id>
+```
+
+The extension is one or two more headers alongside it. Two points about how, both load-bearing:
+
+**Pass derived facts, not the role name.** The instinct is to send `X-Officer-Role: Super Admin`. Don't.
+Roles are the platform's vocabulary and they are not frozen — `Admin`, `Member`, `Developer` today, with
+grants keyed on the role string in `role_permissions`. A plugin that branches on a role name gives every
+plugin repository an opinion about that vocabulary, and renaming a role silently changes authorization
+in code the platform does not own and cannot see.
+
+Send instead what a plugin can act on without knowing anything:
+
+```
+X-Officer-User:  <id>
+X-Officer-Owner: 1 | 0            — is this the server owner
+X-Officer-Level: none|read|write  — this account's effective level for THIS plugin's permission
+```
+
+`isOwner` is the fact that is structurally stable: user id 1, pinned by a CHECK constraint, bypassing
+every permission check. It will mean the same thing in a year. `Super Admin` might not.
+
+A plugin that genuinely needs a third tier — "Admins may also do this" — should not get it by reading a
+role. It should declare a second permission key and let the owner grant it. That keeps the answer to
+"who may do this" on the permissions page, where the owner can see and change it, rather than compiled
+into a plugin's `if`.
+
+**Nothing may be trusted from the caller.** The proxy builds a fresh header object and copies a
+three-item allowlist; it does not forward arbitrary inbound headers. That is why `X-Officer-User` is
+safe to trust, and any new header inherits the same safety **only for as long as the allowlist stays an
+allowlist**. If that loop is ever changed to a copy-everything-except, a client-supplied
+`X-Officer-Owner: 1` becomes a privilege escalation against every plugin at once.
+
+**Absent means denied.** A sidecar that does not see the headers — reached directly, an older platform,
+a bug — must resolve to no privileges rather than to the owner. Same rule as `authorize.ts`, where every
+catch returns deny.
+
+## Where I disagree, and it is one point
+
+Plugin-side enforcement has a failure mode the platform has already been bitten by, and it is worth
+naming before it happens again rather than after.
+
+`assertPermissionTotality` exists because a Member could 403 on `GET /api/tasks` and open
+`/api/tasks/pipeline/ws` with a 101 in the same minute. The lesson written down at the time was
+"**a patch does not survive the next door; refusing to boot does.**" Pure plugin-side authorization is
+that same shape: jellyfin will check, because we are about to write the check. The next plugin's author
+forgets one route, and nothing anywhere notices — no boot failure, no warning, no test.
+
+So the recommendation is not "platform enforces instead". It is **declare in the manifest, enforce in the
+proxy, and pass the facts to the sidecar as well**:
+
+Built as a field on the permission rather than at the manifest root, following `readOnlyWrites` — both
+are sub-paths written relative to the plugin's prefix with a rule attached, and one precedent is better
+than two shapes:
+
+```ts
+// manifest.ts
+permissions: [{ key: 'jellyfin', label: 'Jellyfin', description: '…', ownerOnly: ['/_config'] }],
+```
+
+- **Declared by the plugin**, because the plugin knows which of its paths are administrative.
+- **Enforced by the platform**, in the proxy, before the request reaches the sidecar — so forgetting the
+  check inside the plugin is not a hole, and the rule holds for a route added later by someone who never
+  read this file.
+- **Visible at install**, next to the permissions the owner already reviews, and greppable across every
+  plugin from one place.
+- **Still passed to the sidecar**, because the plugin's own finer rules — whose rows, which fields — are
+  genuinely its business and cannot be expressed as a path prefix.
+
+This does not contradict the `manifest.ts` comment above. That comment refuses to make a **permission**
+owner-only, because a permission is a thing the owner delegates and one nobody can delegate is a
+contradiction. A path prefix is not a grant; it is a statement that a route administers the plugin
+rather than uses it.
+
+## The frontend is a courtesy, not the lock
+
+Hiding the Settings entry in the plugin's nav for a non-owner is worth doing, and it is not security.
+Same rule as `usePermissions`, which fails open on purpose: hiding an entry is politeness, the 403 is the
+lock. Write the guard on the route first and the nav second, never only the second.
+
+## What must be settled before any of this lands
+
+**A member with no server has no Jellyfin.** Connections are stored strictly per user — every query in
+`plugins/jellyfin/db/queries.ts` filters on `eq(jellyfinServers.userId, userId)`. Lock `/_config` to the
+owner today and a Member gets a plugin that can only ever show them the connect form they are forbidden
+to submit.
+
+The resolution that fits what already exists: a non-owner resolves to the **owner's active server URL**
+authenticated with **their own token**, which provisioning already mints and stores in
+`jellyfin_accounts`. That is arguably why those tokens are minted at all — per-user identity upstream, so
+resume points, watched marks and next-up stay each person's own rather than everyone sharing the owner's.
+
+That is a change to credential resolution, not to authorization, and it is the actual prerequisite. The
+guard is a few lines; this is the work.
+
+## Status
+
+|             |                                                                                       |
+| ----------- | ------------------------------------------------------------------------------------- |
+| Agreed      | plugin decides what a level means inside its own surface                              |
+| Agreed      | pass `isOwner` and the effective level; never the role name                           |
+| Agreed      | fail closed when the headers are absent                                               |
+| Built       | `ownerOnly` on a plugin permission, enforced by the gate before self-service and kind |
+| Built       | `X-Officer-Owner` and `X-Officer-Level`, sent when a proxy names its permission key   |
+| Built       | manifest validation — a sub-path without a leading slash is a manifest error          |
+| Blocked on  | credential resolution for non-owners, above                                           |
+| Not started | any plugin actually declaring one                                                     |

docs/push-notifications-app.md

@@ -0,0 +1,225 @@
+# Push notifications — the app side
+
+**Audience:** whoever builds this in `officerdev/monorepo-mobile`. **Platform status:** the server is built,
+running and core; nothing on the device exists yet. **Target: iOS first, Android soon after.**
+
+Read `docs/push-notifications.md` first — it is the design and the _why_. This document is the contract and
+the build order, and it does not repeat the reasoning.
+
+## What the server already does
+
+- `officer-notify` runs as a core PM2 process and is reachable at `/api/notify/*`.
+- `push_devices` exists in Postgres. Registration is idempotent.
+- APNs and FCM channels are implemented and unit-tested; neither has ever sent a real push.
+- **Missing: credentials.** Until the owner installs an APNs `.p8`, `/api/notify/_health` reports
+  `{"ok":true,"channels":[]}` and every send is a silent no-op that returns success. **A green response
+  does not mean a push was sent** — check `channels` first, or you will debug the app for a day over a
+  server with no key.
+- **Alpha is up to date as of 2026-08-19 19:xx.** `/api/notify/*` answered `404` there for most of the day
+  because the server predated this work; it now answers `401` unauthenticated, `push_devices` has its 10
+  columns, and `officer-notify` is running. If you get a `404`, that server needs `git pull` +
+  `pm2 restart officer`, and `officer-notify` needs starting — it is a new core process, so an existing
+  install will not have it until `bun gen:ecosystem` runs.
+
+## Which app
+
+**Only the main Officer app (`apps/mobile`) receives pushes.** It is the hub: everything arrives there and
+it deep-links onward to music, chat, photos and the rest.
+
+`apps/chat` keeps `withoutPushEntitlement.js` — that plugin is **correct** and should not be removed. Chat
+uses local notifications and is not a push receiver. Same for every other app in the monorepo.
+
+## Phase 1 — first light
+
+The goal is one plain banner on a real device. **No codes, no mapping, no extension.** If this does not
+work, nothing later will, and everything later is harder to debug.
+
+### 1. iOS project setup
+
+- A bundle id for `apps/mobile` — **confirmed: `dev.officer.mobile`**, plus `dev.officer.mobile.dev`.
+- `aps-environment` entitlement. Do **not** copy chat's plugin here; the hub needs the opposite.
+- Push capability on the App ID in the developer portal, and a provisioning profile that includes it.
+
+### 2. Get the token and register it
+
+Ask for permission, get the **native APNs device token** (not an Expo push token — the server talks to
+Apple directly and an Expo token will be rejected as malformed), then:
+
+```http
+POST /api/notify/_officer/devices
+Authorization: Bearer <the app's normal Officer session token>
+Content-Type: application/json
+
+{
+  "token": "<64-hex APNs device token>",
+  "platform": "ios",
+  "environment": "production",
+  "bundleId": "dev.officer.mobile",
+  "appSlug": "mobile"
+}
+```
+
+`token`, `platform`, `bundleId` and `appSlug` are **required**; a missing one is a `400` with
+`{"error":"token, platform, bundleId and appSlug are required"}`. `environment` defaults to `production`.
+
+Response: `{"ok":true,"id":<row id>}`.
+
+**Never send a `userId`.** The server takes the user from the authenticated session and ignores the body —
+that is deliberate access control, not an oversight. Sending one is harmless but will not do what it looks
+like it does.
+
+**`environment` is the classic silent failure.** An Xcode/debug build gets a **sandbox** token; TestFlight
+and App Store builds get **production** ones. A sandbox token registered as `production` fails at Apple
+with `BadDeviceToken` and **no user-visible symptom** — no error on the device, nothing in the app. Derive
+it from the build configuration; do not hardcode it.
+
+**Call this on every launch.** APNs tokens rotate. Registration is an upsert keyed on
+`(token, bundleId)`, so re-registering is free and expected.
+
+### 3. The other two routes
+
+```http
+GET    /api/notify/_officer/devices              → { devices: [ … ] }
+DELETE /api/notify/_officer/devices/<token>      → { ok: true }
+```
+
+`GET` returns `{ id, platform, environment, bundleId, appSlug, tokenSuffix, lastSeenAt }` per row —
+**tokens are never echoed in full**, only the last 8 characters, so build any "your devices" UI against
+`tokenSuffix` and `id`. Call `DELETE` on sign-out.
+
+### 4. Ask for a test push
+
+The owner triggers one server-side. You should get an ordinary banner. That closes the loop:
+credential → APNs → device → banner.
+
+## Phase 2 — the mapping
+
+```http
+GET /api/notify/mapping    → { version: 47, entries: { "<code>": { title, body?, target, route, thread } } }
+```
+
+**Built and live** as of 2026-08-19. `Cache-Control: no-store` — the client decides by `version`, so a
+cached copy would defeat the sync it drives. The catalogue is re-read from disk per request, so installing
+a plugin changes the response with no restart.
+
+Codes are namespaced `<publisher>/<appName>:<code>`, e.g. `officerdev/music:12.3`.
+
+The app **fetches this and caches it into a shared App Group container** — shared because the notification
+extension in Phase 3 must read it without the app running. Sync on launch, and whenever a push arrives
+whose `mv` **differs from** the cached `version`.
+
+**Compare with `!==`, not `>`.** `version` is a content hash, not a counter, so it does not increase
+monotonically — removing a plugin returns it to the value it had before that plugin was installed. That is
+deliberate: a counter would need persistent storage and would be wrong after a restore or a rollback.
+
+Entries are templates: `title: "{n} new emails"` rendered against the push's `n`. `route` carries the deep
+link with `{id}` interpolated.
+
+## Phase 3 — the Notification Service Extension
+
+A push arrives looking like this:
+
+```jsonc
+{
+  "aps": {
+    "alert": { "title": "Officer" }, // the fallback the OS shows if the extension does not run
+    "mutable-content": 1,
+    "sound": "default",
+    "thread-id": "officerdev/music",
+    "badge": 3,
+  },
+  "officer": { "c": "officerdev/music:12.3", "n": 3, "id": 412, "mv": 47 },
+}
+```
+
+The extension resolves `c` against the **cached** mapping and rewrites the banner.
+
+**It must not make a network request.** The lookup is a local dictionary read against the App Group cache.
+This is the single most important constraint in the whole design: it is what makes notifications work when
+the tailnet is down, on cellular, with the VPN off. An extension that fetches on arrival will appear to
+work in every test you run at your desk and fail in the field.
+
+**Unknown code** — the push outran the sync, or came from an uninstalled plugin:
+
+> **Unknown notification**
+> code `officerdev/music:12.3` — tap to sync
+
+deep-linking to `officer://resync-notification-mapping`. Keep the code in the body; when this fires on the
+owner's phone, _which_ code is the only useful thing to know.
+
+That screen must be able to **give up**. An uninstalled plugin's codes are gone from the catalogue forever,
+so "resync, still missing, resync" would loop. After one failed resync, say it came from something no
+longer installed and stop.
+
+**One build, ever.** The extension is a generic resolver — it must never `switch` on known codes, or every
+new notification kind becomes an app store release, which is the entire thing this design exists to avoid.
+
+Expo: `ios/` is generated by prebuild, so the NSE target comes from a config plugin. **One already exists
+in this repo** — `packages/bw-vault/plugin/withVaultAutofillTarget.js` creates the AutoFill extension
+target. Follow it rather than inventing one.
+
+## Phase 4 — deep linking
+
+`route` from the mapping is the destination. Tapping a notification opens the hub, which routes onward to
+the target app or screen. The scheme is **`officer://`** (`officer-dev://` in the dev variant), from
+`app.config.ts`.
+
+## Android — later, but design for it now
+
+No extension needed. Send **FCM data-only messages at high priority** and the app composes the notification
+itself; same codes, same mapping, same cache. Register with `platform: "android"` and the FCM registration
+token.
+
+**Do not start Android before the owner settles one thing.** `apps/mobile/PUSH_NOTIFICATIONS.md`
+(2026-07-24) rejects FCM because the Pixel runs GrapheneOS with no Play Services; the platform design chose
+FCM a week later without addressing it. If the target device has no Play Services, FCM cannot reach it at
+all and the fallback is a foreground service holding a tailnet socket. That decision changes the whole
+Android build and is not yours to absorb.
+
+## Things that will waste your time if nobody tells you
+
+- **A successful `POST /_officer/notify` means "queued", not "delivered".** With no credentials configured
+  every channel is skipped and the call still succeeds.
+- **`X-Officer-User` is injected by the platform proxy.** You never set it; if you see a `401` mentioning
+  it, the request bypassed the proxy.
+- **Tokens are addresses, not secrets**, but the server still refuses to echo them back in full. Don't
+  build anything that needs the whole token back.
+- **Until 2026-08-19 a non-owner could not register a device at all** — the permission's `personal` paths
+  did not match the real route. Fixed. If you test with a member account on an older server, that is why.
+- **`appSlug` means the deep-link destination**, not "which app receives this". The name is a leftover from
+  the first design, where each app had its own token.
+
+## Answered, from the app repo (2026-08-19)
+
+**Bundle ids exist.** `dev.officer.mobile`, and `dev.officer.mobile.dev` for the side-by-side dev variant.
+Two bundle ids is fine and needs nothing special: `push_devices` is unique on `(token, bundleId)`, so the
+two variants register as separate rows on the same phone. **Register the dev variant as
+`environment: "sandbox"` and production as `"production"`** — that mapping is now concrete, so derive it
+from the variant rather than the build config guesswork this document suggested earlier.
+
+**The URL scheme is `officer://`, and `officer-dev://` in the dev variant.** Both come from
+`app.config.ts`, and the share extension embeds the same string natively, so they cannot desync. So the
+unknown-code fallback deep link is `officer://resync-notification-mapping` as written.
+
+**Expo prebuild (CNG), not bare** — `ios/` is generated and gitignored; everything native comes from config
+plugins. This is better news than the document assumed: **a config plugin that creates an iOS app-extension
+target already exists in this repo**, `packages/bw-vault/plugin/withVaultAutofillTarget.js`, which builds
+the AutoFill extension that prebuild cannot. **The NSE follows that pattern rather than being invented**,
+which removes what this document called the longest pole in Phase 3.
+
+**`expo-notifications` is NOT in `apps/mobile`** — only in `apps/chat`, for local turn alerts. Correcting
+this document's assumption. Chat's `withoutPushEntitlement.js` still stays exactly as described above; the
+hub needs the opposite.
+
+## The one real blocker: the push entitlement
+
+Today's signed build (`2026.08.19.02`) has **no `aps-environment` entitlement at all**. Nothing else in
+Phase 1 can be tested until it does.
+
+It has a known shape here. On 2026-08-05 adding a Network Extension failed for the same reason:
+`-allowProvisioningUpdates` **cannot add a capability while Xcode has no account signed in**. The fix is a
+one-time manual step by the owner — Xcode → Settings → Accounts, team `RGHC4NJC77` — after which it is
+automatic again.
+
+**Owner action, and it is the critical path.** Along with the APNs `.p8`, this is the whole of what stands
+between here and a banner on a phone.

docs/push-notifications.md

@@ -1,133 +1,342 @@
 # Push notifications — design
 
-**Status:** agreed design, 2026-07-31. Implementation starting. Server side; the app side is a separate
-document handed to the apps repo (`officer-suite/COMMS/PUSH_NOTIFICATIONS_APP.md`).
+**Status (2026-08-19):** server side **built and core**; app side **not started**. This is the second
+design. It supersedes the 2026-07-31 one, which is summarised under "What changed and why" so the
+reversals are legible rather than silently overwritten.
 
-## What was decided, and what was rejected
+The app-side contract and build order is **`docs/push-notifications-app.md`** — hand that to whoever builds
+the client. This file is the design and the reasoning.
 
-**Apple and Google are unavoidable and we accept them.** There is no direct server→phone push on either
-platform: iOS suspends apps, so only APNs can wake one, and Android only accepts pushes from FCM. An
-earlier plan had these the other way round — believing iOS could be pushed to directly and Android needed
-a bridge. It is the reverse.
+## The shape, in one paragraph
 
-**Expo is rejected.** Its push service is a relay in front of APNs and FCM, and it does not remove either
-credential — a standalone Android build still needs an FCM service account, which you would upload *to
-Expo*. All it adds is fan-out convenience, in exchange for our payloads and our APNs key living on a
-fourth party's servers. We talk to Apple and Google ourselves.
+One app on the device — the main Officer app — receives every push and is the only one that can. It is a
+**hub**: the payload it gets is a **code** plus a few numbers, and the app turns that into the notification
+you actually see, deep-linked to whichever app and event it concerns. The code's meaning lives in a
+**mapping owned by the server** and synced to the device ahead of time, so adding a new kind of
+notification is a server change and never an app build. Plugins declare their own codes the same way they
+declare affordances.
 
-**The Android foreground-service alternative is rejected for now.** A persistent socket plus a local
-notification genuinely avoids Google, and the tailnet makes the connection easy. It costs a permanent
-"Officer is running" entry in the notification shade and is unreliable against OEM battery managers
-(Samsung, Xiaomi). Recorded here because it remains the fallback if FCM ever becomes unacceptable.
+## What changed and why
 
-## Payload policy — the notification is a doorbell, not a message
+The first design was a **doorbell**: a category and an id, never content, with the visible text composed
+server-side from the category alone. It called that "the load-bearing decision, and it is not optional".
 
-**This is the load-bearing decision, and it is not optional.**
+Three things moved.
 
-Both services see metadata regardless: which device, how often, at what times. What they must not see is
-content. So a push carries a category and an id, never the substance:
+**1. Privacy relaxed, but not abandoned.** The original weighting — nothing readable may reach Apple or
+Google — was reconsidered on 2026-08-19 as stricter than the owner actually wanted. What replaced it is
+not "send plaintext": it is **codes**, which keeps content off Apple's and Google's servers while removing
+the reason the doorbell was painful (a rigid category enum). The privacy property survives almost intact;
+what was given up is the claim that it was inviolable.
 
-```
-✅ { type: 'mail', count: 3 }              ❌ { title: 'Re: invoice', from: 'x@y.com' }
-✅ { type: 'job', id: 412, ok: false }     ❌ { error: 'ENOENT /home/pastilhas/…' }
+**2. One receiver, not many.** The first design had each app register its own token and `appSlug` select
+which app a notification went to. Now a single hub app receives everything and `appSlug` names the
+**deep-link destination** instead. Same field, opposite direction — see "Migration notes".
+
+**3. Extensibility became the priority.** The binding constraint is now "a new notification kind must not
+require an app store build". That single requirement is what forces the server-owned mapping, and it is
+worth more than the category purity the first design optimised for.
+
+**What did not change:** content still does not travel. Apple and Google see a code and some integers.
+
+## Why a hub app
+
+- **One entitlement, one topic, one sender.** A single bundle id means one `aps-environment`, one APNs
+  topic and one FCM sender, instead of credentials multiplying per app.
+- **It resolves an existing contradiction.** `apps/chat` ships `withoutPushEntitlement.js`, a config plugin
+  that deliberately STRIPS `aps-environment` because chat uses local notifications only. Under the first
+  design that was a blocker. Under this one it is correct and stays — chat is not a push receiver, the hub
+  is.
+- **It concentrates the Android problem** rather than multiplying it (see "Android and GrapheneOS").
+
+## The payload
+
+An ordinary **alert** push — not a silent one — carrying a code and numbers:
+
+```jsonc
+{
+  "aps": {
+    "alert": { "title": "Officer" }, // fallback only; see "What the OS shows"
+    "mutable-content": 1,
+    "sound": "default",
+    "thread-id": "officerdev/music",
+    "badge": 3,
+  },
+  "officer": {
+    "c": "officerdev/music:12.3", // the code — namespaced, see below
+    "n": 3, // variable numbers, never content
+    "id": 412, // what to fetch/open, never a human-readable string
+    "mv": 47, // mapping version the server rendered against
+  },
+}
 ```
 
-The app already has a tailnet connection. On tap it fetches the real content itself. The visible text is
-generic ("3 new emails"), assembled on the device from the category, not sent through Apple or Google.
+**Codes are namespaced, for the same reason affordances are.** `<publisher>/<appName>:<code>`. Plugins from
+different publishers will otherwise collide on a bare integer, and a code in a log should say whose it is
+without a lookup. `core` occupies the appName slot for the platform's own.
 
-Cost of both services: **free**. No per-message charge on either.
+**Variable data rides beside the code, never inside it.** The mapping holds a template — `"{n} new emails"` —
+and the push supplies `n`. Numbers and ids are not content, so the privacy property survives templating.
+
+**What does NOT go in:** subject lines, senders, filenames, error strings, agent output, paths. If a value
+would embarrass you in a Google log, it belongs behind `id` and gets fetched over the tailnet on tap.
+
+## What the OS shows, and why `mutable-content` is required
+
+A normal alert push is rendered **by the operating system, not by the app**. Send `alert.title = "12.3"`
+and the user sees `12.3` on the lock screen — the app never gets a chance to touch it.
+
+So the hub needs a **Notification Service Extension**, enabled per-push by `mutable-content: 1`. The
+extension runs before display, resolves the code, and rewrites the banner.
+
+**Crucially, the extension does a LOCAL lookup, not a network fetch.** The mapping is synced in advance by
+the app into a shared App Group container; at notification time it is a dictionary read. This is what makes
+the design work where a fetch-on-arrival design would not:
+
+- no dependence on the tailnet being connected when the push lands
+- no 30-second budget to overrun
+- no partial failure mode on cellular with the VPN off
+
+**This costs exactly one app build, ever — never one per notification kind.** An extension that hardcoded a
+`switch` on type would need rebuilding for every new kind; this one is a generic resolver and does not.
+
+**The alert text you send is the degraded case.** If the extension does not run — low memory, a crash, an
+older OS — the user sees exactly what was in `alert`. So it is a graceful generic ("Officer"), deliberately
+**not** plausible-looking decoy text: a convincing fake sentence would mislead the one person who most needs
+to trust it.
+
+**Silent pushes were considered and rejected.** `content-available: 1` with no alert would let the app
+compose everything and leak least, but Apple throttles silent pushes and explicitly does not guarantee
+delivery. For notifications that must arrive, that is the wrong tool.
+
+**Android needs no extension.** An FCM **data-only message at high priority** wakes the app, which composes
+the notification itself. Same codes, same mapping.
+
+## The mapping
+
+The server owns it; the device caches it.
+
+```
+GET /api/notify/mapping        → { version: 47, entries: { "officerdev/music:12.3": {...} } }
+```
+
+An entry says how to render and where to go:
+
+| field    | note                                                           |
+| -------- | -------------------------------------------------------------- |
+| `title`  | template, `{n}`-style placeholders                             |
+| `body`   | template, optional                                             |
+| `target` | which app the deep link opens                                  |
+| `route`  | the deep link, with `{id}` interpolated                        |
+| `thread` | grouping key, so like-notifications collapse rather than stack |
+
+**Plugins declare their own**, on the affordances pattern (`src/servers/notifications/publish.ts`, which
+deliberately mirrors `affordances/publish.ts`): a `notifications/` directory of YAML in the plugin
+repository, copied to `capabilities/notifications/<publisher>/<appName>/` on install and removed on
+uninstall.
+
+The namespace is **derived from where the file lands, never from what it declares** — the same rule as
+`mountPrefix`. A plugin therefore cannot claim `officer/core:` or another publisher's codes by writing
+them into its own catalogue; it would have to install itself somewhere it cannot. This is pinned by test. That mechanism already exists and is understood; a second one for
+notifications would drift from it. The platform aggregates every installed plugin's catalogue plus its own
+into the response above, and `version` increments whenever that aggregate changes.
+
+### Staying in sync, and the fallback
+
+Every push carries `mv`, the mapping version it was rendered against. The app compares it to its cached
+version and **resyncs in the background when they differ** — so the mapping normally catches up before an
+unknown code can arrive.
+
+`version` is a **content hash, not a counter**, so the test is `mv !== version` rather than `mv > version`.
+A counter would need somewhere to live and would then be wrong after a restore, a rollback, or two servers
+built from the same install; a hash needs no storage and is right in all three.
+
+When a push still outruns the sync, the fallback renders:
+
+> **Unknown notification**
+> code `officerdev/music:12.3` — tap to sync
+
+deep-linking to `officer://resync-notification-mapping`. The code goes **in the body on purpose**: when this
+fires on the owner's own phone, the first useful question is which code the server sent that the app did not
+have, and without it that is debugged blind.
+
+**The resync screen must be able to give up.** If a plugin is uninstalled its codes leave the catalogue, so
+a queued notification for one will resync, still not resolve, and otherwise loop forever. That path ends
+with "this came from something no longer installed", not another retry.
+
+## What this still leaks
+
+Apple and Google see **which code, how often, at what time, to which device**. Over months that is a
+fingerprint of the owner's category vocabulary even though every code is opaque. That is much less than
+plaintext and is judged acceptable — but it is not nothing, and this document does not claim otherwise.
 
 ## Where it lives
 
-A new **`officer-notify` sidecar**, PM2 peer, its own loopback HTTP listener, announced as
-`notify:server` and proxied at `/api/notify` by `createSidecarProxy`.
-
-Not the platform, because the producers are spread out — the queue, the email sidecar, the agent sidecar —
-and a platform-owned notifier would force every sidecar to call *back* into the platform. That is the
-inversion just removed from email. As a sidecar, anything POSTs to it over loopback.
-
-It also absorbs `src/servers/notify/discord.ts`, so there is one outbound-notification surface with
-channels behind it rather than a Discord path and a push path that do not know about each other.
+`officer-notify`, a **core** PM2 process since 2026-08-19 (built 2026-07-31; spent 2026-08-13 to 08-19
+switched off after being slated for plugin extraction and never extracted). Core rather than a plugin
+because the producers are the platform's own — the job queue, an agent turn ending, mail arriving — and a
+notifier the owner can uninstall is one those producers cannot rely on.
 
 ```
-producer (queue / email / agent / platform)
-        │  POST /_officer/notify { type, ... }
+producer (queue / email / agent / platform / plugin)
+        │  POST /_officer/notify { code, userId, n?, id? }
         ▼
   officer-notify ─┬─ apns   → api.push.apple.com        (HTTP/2, ES256 JWT)
                   ├─ fcm    → fcm.googleapis.com/v1     (OAuth2 bearer)
-                  └─ discord→ webhook                    (existing)
+                  └─ discord→ webhook
 ```
 
+It boots with **no credentials at all** — every channel reports `isConfigured() === false` and `dispatch`
+resolves empty — so an install that never sets an APNs key runs it as a no-op rather than crash-looping.
+That is what makes it safe as core.
+
 ## Credentials
 
-Env only, on the sidecar, never in the platform process and never in the database.
+Env only, on the sidecar, never in the platform process and never in the database. All read
+`OFFICERDEV_<NAME> ?? <NAME>`.
 
 | var                   | what                                             |
-|---|---|
+| --------------------- | ------------------------------------------------ |
 | `APNS_KEY_P8`         | contents of the `.p8` auth key (PKCS#8 EC P-256) |
 | `APNS_KEY_ID`         | the key's 10-char Key ID                         |
 | `APNS_TEAM_ID`        | Apple Developer Team ID                          |
-| `APNS_ENV` | `production` or `sandbox` — different hosts AND different tokens |
 | `FCM_SERVICE_ACCOUNT` | the service-account JSON                         |
-| `DISCORD_WEBHOOK_URL` | existing, moves here |
+| `DISCORD_WEBHOOK_URL` | existing                                         |
 
 The `.p8` does not expire and can push to your apps forever. Treat it as spending-grade.
 
+**The secret store is where these live now** — env is only the fallback.
+
+```bash
+cat AuthKey_ABCD123456.p8 | bun run secret:set apns-p8
+echo -n ABCD123456        | bun run secret:set apns-key-id
+echo -n RGHC4NJC77        | bun run secret:set apns-team-id
+pm2 restart officer-notify        # _health should then report channels: ["apns"]
+```
+
+Purposes: `apns-p8`, `apns-key-id`, `apns-team-id`, `fcm-service-account`, `discord-webhook`. Read with
+`getSecret` and NOT `getKey` — `getKey` mints a random value for an absent purpose, which for a credential
+supplied from outside would mean silently authenticating to Apple with a fake.
+
+Env is still read when the store has nothing, so an install that predates this keeps working. But it is a
+fallback: bun auto-loads `.env` into every PM2 process, so a key there is readable from `/proc/<pid>/environ`
+of twenty of them, and `officer-music` has no business being able to read a credential that can push to your
+apps forever.
+
+`APNS_ENV` appeared in the first design's table and is deliberately **not** implemented: the host is chosen
+per device from the `environment` column, not per process, so one server pushes to a TestFlight build and a
+release build at the same time.
+
 ## Both protocols, verified in Bun before designing around them
 
-- **APNs**: HTTP/2 only. `node:http2` works in Bun 1.3.10 (checked against a live server). Auth is an
-  ES256 JWT signed with the `.p8`, valid ≤1h, refreshed no more often than every 20 min or Apple rejects
-  it. **The signature must be raw `r||s` (64 bytes)** — `createSign(...).sign({ key, dsaEncoding:
-  'ieee-p1363' })`. Node's default DER encoding is silently rejected.
+- **APNs**: HTTP/2 only. `node:http2` works in Bun. Auth is an ES256 JWT signed with the `.p8`, valid ≤1h,
+  refreshed no more often than every 20 min or Apple rejects it. **The signature must be raw `r||s`
+  (64 bytes)** — `createSign(...).sign({ key, dsaEncoding: 'ieee-p1363' })`. Node's default DER encoding is
+  silently rejected.
 - **FCM v1**: ordinary HTTPS. Sign an RS256 JWT with the service-account key, exchange it at
-  `oauth2.googleapis.com/token` for a 1-hour access token, cache that, then POST to
+  `oauth2.googleapis.com/token` for a 1-hour access token, cache it, then POST to
   `fcm.googleapis.com/v1/projects/<id>/messages:send`.
 
-No push library is needed for either — `node:crypto` and `node:http2` cover it. All three endpoints are
-reachable from this host.
+No push library is needed for either.
 
 ## Device registry
 
-New table `push_devices`:
+`push_devices` — live in the schema as of 2026-08-19.
 
 | column                                        | note                                                                                            |
-|---|---|
+| --------------------------------------------- | ----------------------------------------------------------------------------------------------- |
 | `id`                                          | serial                                                                                          |
 | `user_id`                                     | fk users                                                                                        |
-| `token` | the native token — APNs device token or FCM registration token |
+| `token`                                       | APNs device token or FCM registration token                                                     |
 | `platform`                                    | `ios` \| `android`                                                                              |
-| `environment` | `production` \| `sandbox` — an iOS dev-build token fails against prod with a silent `BadDeviceToken` |
-| `bundle_id` | `apns-topic`; also distinguishes the three apps |
-| `app_slug` | `mobile` \| `music` \| `read-aloud` |
+| `environment`                                 | `production` \| `sandbox` — a dev-build token fails against prod with a silent `BadDeviceToken` |
+| `bundle_id`                                   | `apns-topic`                                                                                    |
+| `app_slug`                                    | see migration note                                                                              |
 | `last_seen_at`, `failure_count`, `created_at` | pruning                                                                                         |
 
 Unique on `(token, bundle_id)`. Registration is idempotent: the app re-registers on every launch, since
 tokens rotate.
 
-## Failure handling — the part that is usually skipped
+### Migration notes from the first design
 
-**APNs** answers inline: `410 Unregistered` or `400 BadDeviceToken` means delete the row immediately.
+- **`appSlug` on a NOTIFICATION inverts.** It meant "deliver only to this app". It now means "the deep-link
+  destination". Nothing consumes it yet, so this is a rename-in-place, not a data migration.
+- **`bundle_id` / `app_slug` on a DEVICE go near-constant.** The columns exist to let several apps register
+  separately; with a hub there is one bundle id per device. Keep them — a second receiver later (a watch, a
+  desktop build) costs nothing this way — but "which apps has this user got" stops being a delivery question.
+- ~~**`NotifyType` is replaced by codes.**~~ Done. The producer contract is now
+  `POST /_officer/notify { code, userId, n?, id? }`, validated on the SHAPE of the code
+  (`<publisher>/<appName>:<key>`) and deliberately **not** against the catalogue — a producer firing a
+  code the catalogue has not caught up with is sending a real event, and the device already has a defined
+  fallback. `ok` is gone: a failed job is its own code (`officer/core:job-failed`), which is what made the
+  enum rigid in the first place.
+- ~~**`text.ts` moves to the device.**~~ Done — replaced by `render.ts`, which renders server-side only for
+  Discord (no app to resolve a code) and for the `alert` fallback string.
 
-**FCM** answers inline too: `UNREGISTERED` / `INVALID_ARGUMENT` means delete.
+## Failure handling
 
-Everything else increments `failure_count`; three strikes and the row goes. Without this the registry
-fills with dead tokens and delivery quietly degrades.
+**Only "the app is gone" prunes a row.** The first version pruned on every 4xx Apple or Google offered,
+and that was wrong in a way that hides itself.
 
-(If Expo were used, this would instead require a deferred receipt poll ~15 min after send. Going direct
-removes that entire mechanism — a genuine simplification, worth noting against the fan-out we gave up.)
+| reason                                                        | what it means                   | what happens                                         |
+| ------------------------------------------------------------- | ------------------------------- | ---------------------------------------------------- |
+| `Unregistered` (410), `UNREGISTERED`, `NOT_FOUND`             | the app was uninstalled         | **delete the row**                                   |
+| `BadDeviceToken`, `DeviceTokenNotForTopic`, `TopicDisallowed` | **our** configuration is wrong  | keep it, log the diagnosis                           |
+| `INVALID_ARGUMENT`, `SENDER_ID_MISMATCH`                      | **we** sent something malformed | keep it, log the diagnosis                           |
+| anything else                                                 | transient                       | `failure_count += 1`, three strikes and the row goes |
 
-## Open, deliberately deferred
+`BadDeviceToken` is almost always an environment mismatch — a token minted under
+`aps-environment: development` sent to the production host, or the reverse. The device is fine, the app is
+installed, the registration is good; the server dialled the wrong Apple. **Deleting the row there destroys
+the evidence**: the phone re-registers on next launch, the row comes back looking healthy, and the next
+send fails identically. Found in the field on the first real push, 2026-08-19.
 
-- **Which events notify.** Out of scope by instruction: build the pipe first. Candidates when we get
-  there: job finished, new mail, long agent turn done, download complete.
-- **Quiet hours and batching.** New mail on an 18k-mail account is unusable at one push per message.
-- **`apps/mobile` EAS/bundle id** — only `apps/music` is confirmed to have one (`dev.officer.music`).
+`INVALID_ARGUMENT` is worse still, because it is how FCM reports a malformed MESSAGE — including a
+non-string in `data`, the mistake this codebase is most likely to make and has a test for. Pruning on it
+means one bug in the payload builder wipes every Android device in the registry on a single send.
+
+**Misconfigured is not counted as a strike either.** Not deleting is not enough on its own: a persistent
+mismatch would otherwise reach three failures and delete the row anyway. Same bug, slower fuse.
+
+## Android and GrapheneOS — UNRESOLVED, and it decides Android
+
+`apps/mobile/PUSH_NOTIFICATIONS.md` (2026-07-24) rejects FCM outright: _"requires Google Play Services; the
+Pixel runs GrapheneOS without it."_ The 2026-07-31 platform design chose FCM for Android a week later and
+does not mention GrapheneOS. **Nobody reconciled the two, and this design does not either.**
+
+It matters because FCM cannot reach a device with no Play Services, so on such a phone the entire Android
+path is inert regardless of anything above.
+
+The recorded fallback is a **foreground service holding a tailnet socket**, with notifications raised
+locally. It was rejected in the first design for a permanent "Officer is running" entry in the shade and
+unreliability against OEM battery managers. **The hub changes that calculus**: one socket and one shade
+entry for the whole platform, not one per app. The battery-manager objection is also weaker on GrapheneOS
+than on Samsung or Xiaomi.
+
+This needs an owner decision before Android work starts. iOS is unaffected and can proceed independently.
 
 ## Build order
 
-1. `push_devices` table + `bun db:push`.
-2. `officer-notify` sidecar shell: listener, `notify:server`, PM2 entry, `/api/notify` proxy. Discord
-   channel moved in — one real channel end-to-end before any push credential exists.
-3. APNs channel + `POST /_officer/devices` registration.
-4. FCM channel.
-5. First producer wired.
+1. ~~`push_devices` table.~~ done
+2. ~~`officer-notify` sidecar, PM2 entry, `/api/notify` proxy, Discord channel.~~ done
+3. ~~APNs channel + device registration endpoint.~~ done, 27 tests
+4. ~~FCM channel.~~ done, 27 tests
+5. **APNs credentials.** Owner action; nothing can be tested end-to-end without them.
+6. **Hub app registers its token** against `POST /api/notify/_officer/devices`. Small.
+7. **First light**: a `test` push rendered by the OS with plain text, no codes, no extension. Proves
+   credential → APNs → device → banner as one thing, before any of the mapping exists.
+8. **Codes and the mapping**: `NotifyType` → codes, `GET /api/notify/mapping`, plugin catalogues on the
+   affordances pattern.
+9. **The Notification Service Extension**, local lookup, App Group container.
+10. **First real producer** — job finished is the obvious one.
+
+Steps 7 and 8 are deliberately separable: the transport can be proven working before the rendering exists,
+and if it is not, no amount of mapping work will help.
+
+## Still open
+
+- **Which events notify**, beyond the first producer. Quiet hours and batching — new mail on an 18k-mail
+  account is unusable at one push per message.
+- **Whether members can register devices.** The `notify` permission is `kind: 'app'`, so it is grantable per
+  role and denied by default; a member with no grant cannot subscribe their own phone.
+- **Android**, per above.

docs/servers.md

@@ -0,0 +1,134 @@
+# Servers — the running list
+
+Every machine the owner runs, how to reach it, and what of Officer is on it. Kept here so an
+orchestrator ("pull and update every platform", "restart officer everywhere", "which host is behind")
+has one place to read rather than a fact re-derived per session.
+
+**Verified 2026-08-21** from `officer-dev` by SSH and `pm2 jlist`, except where marked assumed.
+
+## How access works
+
+All of it is over the tailnet, key-based, no `~/.ssh/config`, no passwords, no jump hosts, no ports.
+
+**The login user is NOT always `pastilhas`.** It is per host and it is in the table below. `deadpool` runs
+as `br` and always will — the owner's key is authorised for that account. An orchestrator that hard-codes
+a username fails on that host and will fail on the next one like it.
+
+```bash
+ssh alpha                       # pastilhas, implied
+ssh br@deadpool-rezio-bridge    # not implied
+```
+
+## Officer platforms
+
+Seven, all running.
+
+| Host                         | Tailnet IP  | Login       | Checkout                            | Domain                            |
+| ---------------------------- | ----------- | ----------- | ----------------------------------- | --------------------------------- |
+| `officer-dev`                | 100.64.0.13 | `pastilhas` | `~/officer-dev/platform`            | `platform.officer.dev`            |
+| `alpha`                      | 100.64.0.1  | `pastilhas` | `~/officerdev/platform`             | `officer.pastilhas.dev`           |
+| `fofos`                      | 100.64.0.17 | `pastilhas` | `~/officerdev/platform`             | `officer.fofosfluffybastards.com` |
+| `macbookpro`                 | 100.64.0.9  | `pastilhas` | `~/projects/officer-suite/platform` | `macbookpro.pastilhas.dev`        |
+| `edge-pertento-bridge`       | 100.64.0.15 | `pastilhas` | `~/officerdev/platform`             | `edge.pertento.ai`                |
+| `production-pertento-bridge` | 100.64.0.12 | `pastilhas` | `~/officerdev/platform`             | `officer.pertento.ai`             |
+| `deadpool-rezio-bridge`      | 100.64.0.21 | **`br`**    | `/home/br/officerdev/platform`      | `officer.rezio.dev` `[pending]`   |
+
+Only `platform.officer.dev` is public (77.42.39.155). The rest resolve to tailnet addresses and are
+reachable from inside only.
+
+### Names that disagree with themselves
+
+Three hosts answer to more than one name, and `tailscale status` is not always the one in use:
+
+| In `tailscale status`   | Also                         | The machine's own hostname |
+| ----------------------- | ---------------------------- | -------------------------- |
+| `edge-bridge`           | `edge-pertento-bridge`       | `edge-pertento`            |
+| `gitea-bridge`          | `production-pertento-bridge` | `PPROD`                    |
+| `deadpool-rezio-bridge` | —                            | `deadpool`                 |
+
+The middle column is what to use. `edge-bridge` and `gitea-bridge` are the old names and are being
+retired; anything reading `tailscale status` will still see them.
+
+### Things that will catch out an orchestrator
+
+All verified, none cosmetic.
+
+- **The checkout path differs per host** — four distinct shapes across seven machines. Take it from PM2:
+  the `officer` process's `pm_cwd`. `alpha` carries a SECOND, stale checkout at
+  `~/dockers/officer.dev/platform`, and pulling that one succeeds while changing nothing.
+- **`pm2` is not on a non-interactive PATH.** It lives under nvm, and nvm's setup is in `.zshrc`, which a
+  _login_ shell does not read. Use `ssh <host> '$SHELL -ic "pm2 jlist"'`. `-lc` is **not** enough — it
+  reports "command not found" on machines where pm2 is plainly running, which cost two wrong conclusions
+  on 2026-08-21.
+- **On `macbookpro`, APPEND to the PATH for `bun`. Never replace it.** This said "do not help the PATH"
+  until 2026-08-21, and that was half right in a way that cost a deploy step: `pm2` is only on the
+  interactive shell's PATH, so replacing it breaks `pm2` — but `bun` is at `/opt/homebrew/bin/bun` and
+  `.zshrc` does not put that directory on the interactive PATH either, so `bun run plugins:update` fails
+  `command not found` before it runs a line. `export PATH=$PATH:/opt/homebrew/bin` for that one command
+  shadows nothing: `pm2` stays at `~/.local/bin/pm2` and `node` at `node@22`. Run `pm2` as a separate
+  command.
+- **`$SHELL -ic` on `macbookpro` aliases `ls` to `eza`, which is not installed.** A probe using `ls`
+  reports a missing file that is plainly there. Use `command ls`, or a non-interactive shell for file tests.
+- **The git remotes differ.** `fofos` pulls over HTTPS; the others over SSH. Never assume a URL — pull in
+  place.
+- **A "dirty tree" means TRACKED changes.** `git status --porcelain` counts untracked files, and a stray
+  script a fast-forward could never touch is not a reason to skip a host. Use `--untracked-files=no`.
+- **`macbookpro` is a laptop.** It may be asleep. Unreachable is a skip with a note, not a failed deploy.
+- **`bun run plugins:update` restarts sidecars itself**, printing `sidecar: started` per plugin — on
+  `alpha` that is twelve processes bouncing during the plugins step, before any restart the plan asked for.
+  Not a fault, but a deploy report that lists only `officer` under "restarted" is understating what moved.
+- **`fofos` failed its capabilities sync for three deploys, and it was ownership.** Fixed 2026-08-21.
+  Two errors that read as unrelated problems — `dubious ownership` from git, and
+  `EACCES … mkdir '…/capabilities/affordances/officerdev'` from the gitea plugin — were one cause in two
+  dialects: fifty paths under that checkout were owned by `root`, including `.git` and `affordances`, all
+  stamped `19 Aug 21:43`. Something was run with `sudo` and left them behind. Mode `755` is why the split
+  looked odd: everything could be READ, so git saw the repo and refused it on ownership while the plugin
+  saw `affordances` and could not write into it.
+
+  One `stat` settled it, and that is the general lesson — `stat` prints the owner rather than inferring
+  anything, so it either proves ownership is the cause or eliminates it, in one command. Fixed with
+  `sudo chown -R pastilhas:pastilhas ~/officerdev/capabilities`; verified after by a `mkdir` into the
+  directory that used to refuse one.
+
+  **A `safe.directory` entry was deliberately NOT used**, and that call is worth keeping. It would have
+  silenced the git message and left the `EACCES` untouched — one symptom of a two-symptom cause hidden,
+  with the obvious explanation spent.
+
+## Not platforms — ignore for now
+
+A checkout may be present; no `officer` process runs on any of these. Whether they are retired, staging
+or half-installed is **not established** — ask before treating them as targets.
+
+| Host        | Tailnet IP |
+| ----------- | ---------- |
+| `headscale` | 100.64.0.2 |
+| `omega`     | 100.64.0.3 |
+| `gateway`   | 100.64.0.4 |
+
+## Other machines on the tailnet
+
+- **Not SSH-reachable** (port 22 refused): `edge-dns` (100.64.0.8), `officer-bridge` (100.64.0.5).
+  Assumed service-only containers — not confirmed. `edge-bridge` was listed here until 2026-08-21 and was
+  wrong: its SSH was simply down at the time of checking.
+- **Client devices, not servers:** `iphone16`, `andres-ipad`, `pixel9a`, plus other people's devices in
+  the tailnet (`diogos-ipad`, `iphone-diogo`, `iphone-bruno`, `fedras-iphone`,
+  `guilhermes-macbook-pro`, `macbook-pro-de-joo`).
+
+## Known wrong, waiting on someone else
+
+- **`officer.rezio.dev` resolves to 100.64.0.4** — that is `gateway`, not `deadpool` (100.64.0.21). Not
+  merely unavailable: pointed at a different machine, so it will serve the wrong box the moment gateway
+  answers on 443. Ports 80 and 443 are closed on `deadpool`, so nothing fronts it yet. Waiting on the
+  domain owner.
+
+## Keeping this list true
+
+It rots when a server is added, a checkout moves, a platform is retired, a remote changes or a login
+differs. Re-verify with the four commands it was built from:
+
+```bash
+tailscale status
+ssh <login>@<host> '$SHELL -ic "pm2 jlist"'   # the officer entry's pm_cwd
+ssh <login>@<host> 'git -C <cwd> remote get-url origin'
+getent hosts <domain>
+```

docs/visual-regression-plan.md

@@ -0,0 +1,149 @@
+# Nightly visual regression, on a second server
+
+**Plan, not built.** Written 2026-08-21.
+
+Capture every route of a dashboard behind a JWT login, every night, on a machine nobody is sitting at —
+and have an agent look only at what changed.
+
+The target is **not** Officer and **not** this machine: a separate box running Officer with the `iris` and
+`browser` plugins, pointed at a dashboard the owner runs elsewhere. Its login is a JWT, like Officer's.
+
+## The shape
+
+```
+target server (Officer + iris + browser)
+  │
+  ├─ cron, 03:00 ──▶ capture N routes with iris        → runs/<date>/<route>.png
+  │                  compare against runs/baseline/    → changed[]
+  │                  if changed is empty → record and stop, wake nobody
+  │                  else → open a chat turn with the changed pairs attached
+  │
+  └─ agent reads only the changed images, judges each: intended, or regression
+                     → notification with the verdict, and a link to the run
+```
+
+The expensive, fuzzy step happens only where pixels actually moved. On a quiet night the run costs a few
+seconds of Chrome and nothing else.
+
+## The hard part: a JWT lives in localStorage, and iris cannot put it there
+
+This is the thing to solve first, because everything else is plumbing.
+
+Cookie-authenticated sites work today: `iris-browser` copies `Cookies`, `Cookies-wal`, `Login Data` and
+`Local State` out of a signed-in profile, so a capture rides the same session.
+
+**A JWT dashboard is not cookie-authenticated.** The token sits in `localStorage`, which Chrome keeps in
+`Default/Local Storage/leveldb/` — a directory `iris-browser` does not copy. So a capture of a
+token-authenticated app renders the login screen, and does it silently: the screenshot is real, the page
+is real, and it is the wrong page.
+
+Three ways out, worth choosing between deliberately.
+
+### A. Copy `Local Storage/leveldb` too — smallest change, biggest blast radius
+
+One more entry in `seed_sessions`. It works, and it copies **every** site's localStorage into the
+capture profile, not just the target's. That is someone's whole browser state in a temp directory, and
+the directory is a LevelDB — copying it under a live browser can catch a compaction mid-write, which
+Cookies-the-file mostly cannot.
+
+Viable, and the least attractive of the three.
+
+### B. Log in during the capture — no profile at all
+
+Most JWT dashboards accept a token in the URL, or can be given one by a first-run page. If the target
+does, then:
+
+```
+iris 'https://dash.example.com/?token=<jwt>' -o run/home.png
+```
+
+and there is no profile, no desktop, no signed-in browser anywhere. The token comes from a service
+account minted for this, stored in Officer's secret store, and never appears in a screenshot.
+
+**This is the recommended path** when the target supports it. It removes desktop from the picture
+entirely, which removes the only reason that plugin was in this plan.
+
+### C. Seed via the relay, capture via iris
+
+Pair a browser once, use the relay to `evaluate` the token into `localStorage`, and let iris capture from
+that profile. Works without changing either plugin, and needs a browser paired at 3am — which is exactly
+the constraint iris exists to avoid.
+
+Keep as the fallback for a target that supports neither a token URL nor a clean profile copy.
+
+## What to capture
+
+**Routes come from a list the repo owns**, not from crawling. A crawler finds what exists; a list says
+what is supposed to exist, and a route disappearing is the regression you most want caught.
+
+```jsonc
+// routes.json, in the dashboard's own repository
+[
+  { "path": "/", "name": "home" },
+  { "path": "/reports", "name": "reports", "waitFor": "[data-ready]" },
+  { "path": "/reports/42", "name": "report-detail" },
+]
+```
+
+`waitFor` maps to iris's `--wait-for`, which is what stops a spinner being the thing under test.
+
+Two viewports — one desktop, one phone — doubles the run and catches the majority of layout breakage.
+`--jobs` renders them in parallel.
+
+## Diffing, before any agent is involved
+
+Compare each capture with the same route's baseline. A pixel count, a threshold, nothing clever:
+
+- **identical** → nothing happens
+- **differs under threshold** (antialiasing, a clock, a random avatar) → nothing happens, but log it
+- **differs over threshold** → into `changed[]`
+
+The threshold exists because something on the page always moves. **Mask what is known to move** rather
+than raising the threshold until it passes — a timestamp region is a rectangle to blank before comparing,
+not a reason to accept 3% drift everywhere.
+
+Officer has no image diff today. `sharp` is already a plugin dependency (atelier) and can do it; a raw
+pixel loop over two decoded buffers is about thirty lines and needs nothing.
+
+**Baselines are promoted, never automatic.** A run does not overwrite the baseline because it succeeded —
+it does so when someone says "yes, that is the new correct". Otherwise a regression becomes the baseline
+on night two and disappears.
+
+## What reaches the agent
+
+Only changed pairs, and each one as **two images plus the route name**: baseline and current. The
+question asked is narrow, because a broad one gets a broad answer:
+
+> This is `/reports` before and after. Describe what changed, then say whether it is a layout break, a
+> content change, or noise. If you cannot tell, say so.
+
+"If you cannot tell, say so" earns its place — a model asked to judge will judge, and a confident wrong
+verdict on night four is how the whole thing stops being trusted.
+
+The verdicts land in a notification (`officer-notify` is already core) with a count and a link, not the
+images. Nobody wants twenty pictures on a phone at 3am.
+
+## What does not exist yet
+
+In rough order of how much they block:
+
+1. **`--wait-for` against a JWT app** — solve auth first (§B), everything else is downstream.
+2. **An image diff.** Thirty lines, or `sharp`.
+3. **A scheduler.** Officer's queue has retries, not schedules — `scheduleRetry` is the only timed thing
+   in it. So this is system cron calling a script, or a `setInterval` in a plugin sidecar the way the
+   nudger does it. Cron is the honest answer for something that runs once a night.
+4. **A place for runs and baselines.** `$OFFICER_ROOT/data/visual-regression/{baseline,runs}/` — outside
+   any plugin directory, because a plugin update replaces that.
+5. **Whatever holds this.** A plugin of its own on that server, with `routes.json` as its config and a
+   panel showing the last run. It is not iris and it is not the relay; it is the thing that uses them.
+
+## What this deliberately is not
+
+**Not a test runner.** It proves a page renders and looks like it did yesterday. It cannot tell you a
+button is wired to the wrong handler, and it should not be stretched to try — that is a real test suite,
+and the two answer different questions.
+
+**Not a crawler.** See routes, above.
+
+**Not the browser relay.** The relay drives a live browser someone is signed into; this runs when nobody
+is. They overlap only in that both produce screenshots. If §B holds, the relay has no role in this at all.

package.json

@@ -1,6 +1,6 @@
 {
   "name": "officer",
-  "version": "0.1.0",
+  "version": "2026.08.21.04",
   "private": true,
   "type": "module",
   "workspaces": [
@@ -10,6 +10,14 @@
   "scripts": {
     "preinstall": "node -e \"var v = +process.versions.node.split('.')[0]; if (v < 22) { console.error('Node 22 or newer required (got ' + process.versions.node + '). Run: nvm use 22'); process.exit(1); }\"",
     "gen:index": "bun run ./scripts/gen-index.ts",
+    "fix:env": "bun run ./scripts/fix-env.ts",
+    "todo:index": "bun run ./scripts/todo-index.ts",
+    "plugins:update": "bun run ./scripts/update-plugins.ts",
+    "nudge": "bun run --env-file=.env ./scripts/nudge-agent.ts",
+    "doctor": "bun run --env-file=.env ./scripts/doctor.ts",
+    "secret:set": "bun run --env-file=.env ./scripts/secret-set.ts",
+    "notify:test": "bun run --env-file=.env ./scripts/notify-test.ts",
+    "gen:ecosystem": "bun run --env-file=.env ./scripts/gen-ecosystem.ts",
     "predev": "bun run ./scripts/gen-index.ts",
     "dev": "bun --env-file=.env --watch src/server.tsx",
     "prestart": "bun run ./scripts/gen-index.ts",

plugins/CLAUDE.md

@@ -0,0 +1,92 @@
+# Plugins
+
+`EXTRACTING-A-PLUGIN.md` is the runbook for turning a platform feature into a plugin, and
+`EXTRACT-A-PLUGIN-PROMPT.md` is the way to hand that to a fresh agent. This file is for the concepts a
+plugin author needs that are not procedure.
+
+**This directory holds documentation, not plugins.** Every plugin is its own repository; `plugins/<name>/`
+appears by `git clone` when somebody installs one, and is gitignored here.
+
+---
+
+## Affordances
+
+**`docs/affordances.md` is the authority** — what one is, the three tests for whether something qualifies,
+the file format and how they reach the agent. Read it before writing one. What follows is only the part
+that is specific to shipping one from a plugin.
+
+> An affordance is something the surface on the other end of the conversation offers the agent — declared
+> by the code that implements it, and read by the agent at runtime. Not what the agent can do; what the
+> client will honour.
+
+A plugin ships an `affordances/` directory of markdown-with-frontmatter in its repository. Installing
+copies it to `capabilities/affordances/<publisher>/<appName>/`, uninstalling removes it — the same shape
+as `assets/` and `publishAssets`, and the namespace comes from where it lands rather than from anything
+inside the file.
+
+**The test that catches most bad ones: can it be fixed instead?** If the answer is yes, fix it. An
+affordance describing a defect gives that defect a rationale, and the rationale outlives everyone's memory
+that it was a bug. See the two worked examples in `docs/affordances.md`.
+
+Keep it short — every affordance is loaded into every conversation on this machine, including turns that
+have nothing to do with your plugin.
+
+### A plugin-contributed one, for reference## Notifications — the same shape, for push
+
+A plugin can also declare what its push notifications MEAN. Same mechanism, different store:
+
+```
+capabilities/notifications/
+  officerdev/music/
+    catalogue.yaml
+```
+
+```yaml
+nowPlaying:
+  title: 'Now playing'
+  route: 'officer://music/{id}'
+```
+
+That becomes the code `officerdev/music:nowPlaying`. A push carries the code and integers — never content —
+and the device renders the title and follows the route. The point is that **adding a notification is a
+server change and never an app build**, because the mapping is synced to the device rather than compiled
+into it.
+
+**The namespace comes from where the file lands, not from what it says.** Writing `officer/core:test` as a
+key in your own catalogue produces `<you>/<yourplugin>:officer/core:test`, not the core code — the same
+rule as `mountPrefix`, and pinned by test in `src/servers/notifications/catalogue.test.ts`.
+
+An entry needs a `title`; `body`, `target`, `route` and `thread` are optional, `thread` defaults to your
+namespace, and unknown keys are dropped rather than forwarded to the device. Broken YAML in one plugin's
+catalogue is skipped without affecting anyone else's.
+
+See `docs/push-notifications.md`.
+
+### A plugin-contributed one, for reference
+
+This is the gitea plugin's, and it shows what these are for: the agent cannot discover from the
+filesystem that `tea` is authenticated, or that it should prefer `git` for the things git already does.
+
+```markdown
+---
+name: Gitea
+description: Use when working with git repositories, or with Gitea issues and pull requests.
+version: 1
+---
+
+For anything git already does — clone, fetch, pull, push, branches, tags — use plain `git`. It talks
+to the server directly.
+
+For anything git does not — pull requests, issues, comments, releases, creating a repository — use
+`tea`. It is already logged in as the current user, with that user's own permissions. Do not
+configure it, and do not ask the user for a token.
+```
+
+### On the word
+
+`capability` was taken — `CAPABILITIES` is the permission registry. `skill` was taken twice: by
+`capabilities/skills/` here and by the wider AI ecosystem.
+
+`affordance` is the exact term for what an environment offers an actor, and the precision earns the
+slight awkwardness: rendering an image is not a skill the agent has, it is something the client on the
+other end permits. Keeping that distinction in the name saves explaining it every time.

plugins/EXTRACT-ATELIER-PROMPT.md

@@ -0,0 +1,181 @@
+# Kickoff prompt: extract Atelier into a plugin
+
+Written 2026-08-20 for an **unattended overnight run**. That changes two things from
+`EXTRACT-A-PLUGIN-PROMPT.md`, which assumes the owner is awake:
+
+- **It does not say "ask and wait".** Nobody is there. Decisions that would normally be questions are
+  pre-made below, with reasoning, so the run does not stall on them.
+- **It assumes you may be restarted.** Commit after every task, so whoever picks this up — including you,
+  an hour later, with no memory of it — finds a clean tree and a readable history.
+
+Paste everything below the line into a fresh agent.
+
+---
+
+You are extracting **Atelier** out of `~/officer-dev/dockers/officer-atelier` and into an Officer plugin,
+the way `music`, `offscale` and `gitea` were extracted before it.
+
+**Work in a git worktree, not the main checkout.** Another agent is watching
+`/home/pastilhas/officer-dev/platform` and pulling into it as you push. Two processes in one working tree
+is how work gets lost — on 2026-08-20 it nearly did.
+
+```
+git -C /home/pastilhas/officer-dev/platform worktree add /home/pastilhas/officer-dev/atelier-work -b atelier-plugin
+```
+
+Everything below happens in `~/officer-dev/atelier-work`.
+
+**Push every commit straight to master**, from the worktree:
+
+```
+git push origin HEAD:master
+```
+
+The branch is only there to keep your files out of the other checkout. Master is what is being reviewed.
+Push after every task — a reviewer reading five small commits can tell you what went wrong on the second
+one; nobody can do that with one commit at the end.
+
+**Your final commit message must contain `ATELIER-COMPLETE` on its own line**, and no earlier commit may
+contain that string. It is the signal that the run is finished and the review can start. Do not use it to
+mean "this part is done".
+
+## Read first, in this order
+
+1. **`plugins/EXTRACTING-A-PLUGIN.md`** — the runbook and the authority. Read "What changed on
+   2026-08-16" carefully; it contradicts an earlier instruction in the same file, deliberately.
+2. **`plugins/example/README.md`** — what a plugin is, as conventions.
+3. **`src/servers/plugins/`** — `discover.ts`, `manifest.ts`, `mount.ts`, `install.ts`, `ecosystem.ts`.
+   Where this prompt and that code disagree, **the code is right**.
+4. **`CLAUDE.md`** — house rules, the permission model, and why `assertPermissionTotality` refuses to boot.
+5. **`docs/secret-store.md`** — Atelier holds an OpenAI key.
+6. **`src/workspaces/officerdev/APP_CONVENTIONS.md`** and `HOOK_CONVENTIONS.md` — you are writing panels.
+
+## What Atelier is
+
+Image generation as a service: a provider-shaped API over OpenAI's image endpoints, storing what it makes
+so every image has a permanent URL. Bun · Hono · Postgres/Drizzle · htmx with Hono JSX · a React SPA.
+
+```
+84 ts/tsx files, ~6,400 lines
+tables      collections, provider_settings        (its own `atelier` database)
+routes      /generations, /generations/:id, /edits, /adapters,
+            /:id/items, /:id/items/:generationId, mask
+env         BUN_PORT, PUBLIC_URL, POSTGRES_URL, ATELIER_DATA_DIR,
+            OPENAI_API_KEY, ATELIER_MAX_CONCURRENCY
+runs as     pm2 `officer-atelier`, cwd ~/officer-dev/dockers/officer-atelier
+```
+
+It is a **sidecar** plugin for one reason that survives the UI moving into Officer: generation is
+long-running, capped by `ATELIER_MAX_CONCURRENCY`, and holds a credential the platform must not. The UI
+is Officer's; the work is the sidecar's.
+
+## Decisions already made — do not re-open these
+
+Each of these would normally be a question. They are answered so you can keep moving. If you find
+evidence that one is wrong, **write it in `BLOCKERS.md`, follow the decision anyway, and carry on** — a
+half-applied decision is worse than a wrong one consistently applied.
+
+**1. The UI is PORTED into Officer. No proxy, no iframe, no sidecar-served UI.** Atelier was a separate
+server because that was the quick answer at the time; this extraction is what undoes that. It becomes part
+of Officer, not something Officer frames.
+
+It is smaller than the repository suggests:
+
+```
+src/apps/frontend      1,736 lines / 25 files
+  6 of those           components/ui/*  — shadcn Officer ALREADY HAS. Delete, import from @/components/ui.
+  real work            Generate · Gallery (+ Lightbox) · Collections · Providers   ≈1,200 lines
+src/apps/web           the htmx half — 8 files, DELETED, not ported
+```
+
+So: three panel apps — **Generate**, **Gallery**, **Collections** — under
+`plugins/atelier/web/`, each exporting `appRegistryMetas`, with a `layout.ts` composing them. Follow the
+frontend route conventions in `CLAUDE.md`: selection lives in the URL, rows are real links, no
+`usePanelChannel` for "which thing is open".
+
+The fourth feature, **Providers**, is not a panel — see the next decision.
+
+**2. Providers become Officer settings, with more than one.** `features/Providers/` is a settings screen,
+and Officer has a settings surface already. It moves there rather than becoming a panel.
+
+`src/servers/providers/` is already adapter-shaped — `types.ts`, `openai.ts`, `credentials.ts`, an
+`index.ts` that dispatches. Keep that shape and keep the seam: today OpenAI is the only adapter, and the
+point of moving the settings is that a second one is a new file rather than a new screen. **Do not add a
+second provider in this run** — leave the shape able to take one and say so in `PLUGIN.md`.
+
+**3. The database: check the rows, then decide.** Atelier uses a separate `atelier` Postgres database,
+deliberately — `src/databases/atelier_db/drizzle.config.ts` says why. Plugin tables normally live in
+`officer_db` with a prefix.
+
+- Count the rows in `collections` and `provider_settings` first.
+- **If both are empty**, fold them into `officer_db` as `atelier_collections` and
+  `atelier_provider_settings`, following `plugins/*/db/schema.ts`. Two tables do not earn a second
+  database.
+- **If either has rows**, keep the separate database for now and record the migration as a follow-up in
+  `BLOCKERS.md`. Losing someone's generated images to a schema decision at 4am is not a trade worth making.
+
+**4. Artefacts: `$OFFICER_ROOT/data/atelier/`.** Generated images are files, and `ATELIER_DATA_DIR`
+currently points wherever its `.env` says. A plugin directory is a git clone that an update replaces, so
+artefacts kept there vanish on upgrade with nothing logging it. `plugins/kiosk/artefacts.ts` has the same
+problem solved and the comment explaining it.
+
+**5. `OPENAI_API_KEY` goes in the secret store**, not `.env`. `putSecret`/`getSecret` landed 2026-08-19;
+`bun run secret:set atelier-openai-key` puts it there. Read with `getSecret`, **never `getKey`** — `getKey`
+mints a random value for an absent purpose, which for a foreign credential means silently authenticating
+with a fake. Fall back to env when the store has nothing, so an install that predates this keeps working.
+
+**6. One permission, `atelier`, `kind: 'app'`.** Image generation is not `execution` — it does not run as
+the owner's OS user in the owner's home.
+
+## How to work
+
+**Map before you move.** The first task is not a file move. Find every line in the platform that names
+atelier, and every import pointing INTO it. A plugin may import platform code (`@@/…`); the reverse cannot
+happen. Report what you find before moving anything.
+
+**Commit after every task, and push.** You may be restarted at any point by a scheduler that has no idea
+what you were doing. A clean tree and a commit message are how the next attempt — possibly you — knows
+where it is. Commit messages: simple lowercase, no prefixes.
+
+**Do not restart the platform.** `pm2 restart officer` breaks the owner's chat session. Write the command
+down instead; they will run it.
+
+**Do not `git add -A`.** Add the paths you touched. This has caused two commits to swallow another
+agent's work in one day.
+
+**Format only what you touched** — `bunx prettier --write <paths>`, never `bun format`.
+
+**`bunx tsgo` must be clean** before every commit. Not `tsc`.
+
+## When you get stuck
+
+Write it in `BLOCKERS.md` at the worktree root — what you were doing, what stopped you, what you did
+instead, and whether it is reversible. Then **move to the next task**. Do not wait, do not invent a
+workaround that changes behaviour without saying so, and do not stop the run.
+
+If invented content ends up anywhere durable — copy, defaults, seed data — mark it `[placeholder]` so it
+can be found later with a grep.
+
+## Done means
+
+- `plugins/atelier/` with `manifest.ts`, `api/router.ts`, `sidecar/index.ts`, `web/panels.ts`,
+  `web/layout.ts`, `db/`, `assets/icon.png`, `assets/splash.png`, and its own git repository.
+- It installs from the Marketplace onto a machine that has never had it, and the panel renders.
+- `bunx tsgo` clean, `bun test` passing.
+- A `PLUGIN.md` in the plugin recording what moved, what did not, and why — the way `music` and `offscale`
+  did. This is the part that is worth most in three months.
+- `BLOCKERS.md` listing every decision you made without being able to ask.
+
+## What you may NOT do
+
+- No unauthenticated routes. A plugin cannot serve one; `UNPROTECTED_API_PREFIXES` is hardcoded in the
+  platform. If Atelier needs one, that is a blocker, not something to work around. (`plugins/kiosk`
+  is stuck on exactly this — read its README before concluding you have found a way.)
+- **Generated images are served through an authenticated route**, like every other plugin artefact. Every
+  image having "a permanent URL" was true of a standalone server with its own `PUBLIC_URL`; inside Officer
+  the URL is behind the session. If something needs an image without one — an `<img>` in an email, a
+  share link — that is a blocker to write down, not a route to open. See the image-rendering affordance
+  in `capabilities/affordances/officer/core/` for why an `<img>` cannot carry a token.
+- No changes to `src/servers/plugins/` to make Atelier fit. If the contract does not accommodate it, that
+  is a finding worth more than the workaround.
+- No second agent, no second worktree, no background job that writes to this tree.

plugins/EXTRACTING-A-PLUGIN.md

@@ -78,9 +78,11 @@ Music added the field; see its PLUGIN.md for what it is guarding against.
 **Nothing may branch on provenance** except `mountPrefix()`. First-party and third-party differing
 anywhere else means two systems, and only one gets tested.
 
-**Uninstall never destroys data.** The generated schema barrel follows plugin **directories**, not the
-install table — `db:push` drops what it cannot see, so following installs would delete a plugin's tables
-on uninstall. Only deleting a plugin's source can lose its data.
+**Only an explicit tick destroys data.** The generated schema barrel follows plugin **directories**, not
+the install table — `db:push` drops what it cannot see, so following installs would delete a plugin's
+tables the moment it was stopped. Deleting a plugin removes its directory, and the confirmation's
+"also delete all its data" checkbox is the one thing that drops its tables. Unticked, the rows outlive
+the plugin and show up in the orphan list on the same screen.
 
 ---
 
@@ -125,9 +127,9 @@ Then through `/plugins`, watching PM2 and the browser at each step:
 | permissions page                | its permission is listed, read/write/none                   |
 | disable                         | route 404s, sidecar stops, **tables and rows survive**      |
 | enable                          | comes back                                                  |
-| uninstall                       | route gone, `pm2 list` loses it, **data still there**       |
-| `bun db:push` while uninstalled | **do not run this — see "What changed on 2026-08-16"**      |
-| install again                   | identical to the first install                              |
+| delete, checkbox unticked       | route gone, `pm2 list` loses it, directory gone, **data still there** |
+| install again                   | fresh clone, and it finds those rows                        |
+| delete, checkbox ticked         | tables dropped too — a machine that never had it            |
 
 A normal refresh is enough; the shell is `no-store`. When the log's last line appears, the bundle exists.
 
@@ -159,6 +161,16 @@ A normal refresh is enough; the shell is `no-store`. When the log's last line ap
   module cache holds it for the lifetime of the process — so editing a manifest while developing changes
   nothing until `pm2 restart officer`. Costs ten minutes the first time, because the plugins page keeps
   cheerfully showing the old values. `outdated` cannot notice a version bump without a restart either.
+- **A router that proxies BY HAND must call `registerProxiedPrefix`.** `bodyParser` runs on the whole
+  protected router and reads JSON and form bodies, which DISTURBS the request stream — so a catch-all that
+  then forwards `ctx.req.raw.body` forwards an already-consumed stream, Bun's fetch throws, and the proxy
+  reports the sidecar as unreachable. Every bodyless GET keeps working, so it presents as "the sidecar is
+  down" for POSTs only, which is a long way from the cause. `createSidecarProxy` registers the prefix
+  itself, so core sidecars and any plugin using it never meet this; vault proxies by hand and its
+  extraction dropped the registration on 2026-08-18, costing an afternoon. If your plugin hand-rolls a
+  proxy: `registerProxiedPrefix(PREFIX)` beside the mount, and check no handler under it reads
+  `ctx.get('body')`.
+
 - **A plugin importing platform code is fine (`@@/…`); the reverse is not.** If something in `src/` imports
   from your feature and cannot move — a widget, a relay — that piece stays, and the boundary goes around
   it. Find those before you plan the split; they decide it for you.
@@ -210,15 +222,21 @@ no terminal it applies NOTHING and exits 0, so an install reports success and th
 query with 42P01. `--force` would be worse if it worked: the destructive statements are `drop table` for
 every plugin whose directory is absent.
 
-So: the row in the verification table that says to run `bun db:push` while uninstalled is struck out. The
-property it was checking — uninstall does not destroy data — is now checked by reinstalling and finding
-your rows still there.
+So: the row in the verification table that said to run `bun db:push` while stopped is gone. The property
+it was checking — stopping a plugin does not destroy data — is now checked by deleting it with the
+checkbox unticked, installing again, and finding your rows still there.
 
-**Use `purge` to test the install twice.** Uninstall deliberately keeps the directory and the tables, so a
-second install is not a fresh one: the marketplace hides a plugin whose directory exists, and the tables
-are already there. The Marketplace screen has "Delete <plugin> and all its data" — it drops the tables,
-deletes the directory and regenerates the barrel, and afterwards the plugin installs from the catalogue
-exactly as a stranger's machine would. **That is the only honest test of an install.**
+**Delete WITH the checkbox to test the install twice.** Anything less is not a fresh install: the
+marketplace hides a plugin whose directory exists, and tables that are already there make `db:push` a
+no-op, so a broken schema stays broken and passes. Ticked, it drops the tables, deletes the directory and
+regenerates the barrel — and afterwards the plugin installs from the catalogue exactly as a stranger's
+machine would. **That is the only honest test of an install.**
+
+**Uninstall was removed on 2026-08-21.** It sat between disable and delete, keeping the directory and the
+data, and the screen could not show the difference: an uninstalled plugin stayed in the list under a
+heading that said "Installed" while its directory silently decided what the next install would run. That
+is not hypothetical — it is how a jellyfin install on alpha re-applied a checkout from two days earlier,
+reporting success at every step. The verbs are now enable, disable and delete.
 
 **The install now verifies its own tables.** After the push it asks Postgres whether each declared table
 exists and fails the install naming what is missing. If your extraction reports "schema push reported

plugins/USER_PROVISIONING.md

@@ -0,0 +1,240 @@
+# Provisioning a user in a plugin's upstream service
+
+> **Handing this to a fresh agent?** Everything below was verified against a live instance on
+> 2026-08-19 — every status code, every error body. Trust it over the upstream's own docs, which
+> lag. Re-verify with the probe sequence at the end if the upstream has been upgraded since.
+
+Officer creates the account; the upstream service owns it. This is the contract for doing that
+safely, written from the **`photos` plugin** (Immich) as the first worked example. The shape
+generalises: authenticate as an admin, check-then-create, verify the result, hand over a one-time
+password.
+
+**This directory holds documentation, not plugins.** The plugin itself lives in
+[`plugins/photos`](https://gitea.officer.dev/plugins/photos).
+
+---
+
+## The four rules
+
+1. **Never hardcode the admin credential.** Read it from the environment or the platform's
+   credential store, and never log it. Docs and repos get shared; credentials must not travel
+   with them.
+2. **Check before you create.** None of these APIs upsert. A second call with the same email is
+   an error, not a no-op — so provisioning must be re-runnable by listing first.
+3. **Force a password change.** Generate a strong random password, mark it as
+   must-change, and treat it as a one-time handover value. It stops being a secret the moment
+   the user logs in.
+4. **Verify by logging in.** A `201` means the row was written, not that the account works. Do a
+   real login before reporting success, so failures surface at provisioning time rather than when
+   the person first tries to use it.
+
+---
+
+## Worked example — Immich (`photos` plugin)
+
+Verified against **Immich 3.1.0**.
+
+| | |
+|---|---|
+| API root | `<base>/api` — e.g. `http://127.0.0.1:2283/api` |
+| Auth | header `x-api-key: <ADMIN_API_KEY>` |
+| Admin routes | `/api/admin/*` — reject non-admin keys with `401` |
+
+Immich stores API keys **SHA-256 hashed**, so a lost key can only be replaced, never recovered.
+Mint one in the UI → Account Settings → API Keys.
+
+### 1. Check (idempotency guard)
+
+```http
+GET /api/admin/users
+x-api-key: <ADMIN_API_KEY>
+```
+→ `200`, JSON array. If the target email is present, stop — already provisioned.
+
+### 2. Create
+
+```http
+POST /api/admin/users
+x-api-key: <ADMIN_API_KEY>
+Content-Type: application/json
+
+{
+  "email": "person@example.com",
+  "password": "<generated>",
+  "name": "Full Name",
+  "shouldChangePassword": true
+}
+```
+
+**Required:** `email`, `password`, `name`.
+
+| Optional field | Type | Notes |
+|---|---|---|
+| `shouldChangePassword` | bool | `true` ⇒ forced reset at first login. Always send it. |
+| `quotaSizeInBytes` | int | e.g. `1073741824` = 1 GiB. Omit for unlimited. |
+| `storageLabel` | string | Folder name used by the storage template. |
+| `notify` | bool | Emails the new user. Requires SMTP configured upstream. |
+
+→ **`201 Created`**, body carries the new `id`:
+
+```json
+{ "id": "70ee4cd2-…", "email": "person@example.com", "name": "Full Name",
+  "isAdmin": false, "shouldChangePassword": true, "quotaSizeInBytes": null }
+```
+
+`isAdmin` cannot be set at creation — promote afterwards via the admin update endpoint.
+Names are full Unicode (`Bruno Conceição`); send UTF-8, no escaping.
+
+### 3. Verify
+
+```http
+POST /api/auth/login
+Content-Type: application/json
+
+{ "email": "person@example.com", "password": "<generated>" }
+```
+→ `200` with `{"name":…,"userEmail":…,"isAdmin":false,"shouldChangePassword":true}`.
+No API key — this is the ordinary user login path, which is exactly why it proves the account works.
+
+### Errors (observed, not guessed)
+
+| Condition | Status | Body |
+|---|---|---|
+| Email already exists | `400` | `{"message":"Email is not available"}` |
+| Missing `password` / `name` | `400` | `{"message":"Validation failed","errors":[{"path":["password"],…}]}` |
+| Bad or missing API key | `401` | — |
+
+### Deprovisioning
+
+```http
+DELETE /api/admin/users/{id}
+Content-Type: application/json
+
+{ "force": true }
+```
+→ `200` with `{"status":"removing","deletedAt":"…"}`. Soft-delete; `force` skips the grace period.
+
+---
+
+## Worked example — Jellyfin (`jellyfin` plugin)
+
+Same four rules, **different shape**. Read this section as: the route surface and auth model are
+**verified** against Jellyfin **10.11.11**; the request bodies are from working knowledge and
+marked as such — confirm them with the probe recipe below before relying on them in code.
+
+### How it differs from Immich — read this first
+
+| | Immich | Jellyfin |
+|---|---|---|
+| Auth header | `x-api-key: <key>` | `Authorization: MediaBrowser Token="<key>"` |
+| Create + configure | one call | **two or three** — create, then policy, sometimes password |
+| Force password change | `shouldChangePassword: true` | **no equivalent** — see below |
+| Scope of an account | one server | **per instance** — see below |
+
+Two consequences that bite:
+
+1. **There is no "must change password" flag.** Rule 3 cannot be enforced by the server.
+   **Decision for this deployment (2026-08-20): set a generated password at creation and tell the
+   user to change it.** Do *not* use the blank-password alternative — Jellyfin permits it, but the
+   account is then open to anyone who can reach the port until the user sets one.
+   So for Jellyfin, rule 3 becomes: still generate a strong unique password, still treat it as a
+   handover value, but the change is advisory rather than forced — which means **never reuse a
+   generated password across users or instances**, since it may stay valid indefinitely.
+2. **Users are per-instance, not global.** This deployment runs **four** independent Jellyfin
+   servers, each with its own user database and its own API keys:
+
+   | Instance | Port | Server name |
+   |---|---|---|
+   | `jellyfin` | 9301 | Jelly |
+   | `jellyfin_music` | 9302 | Albums |
+   | `jellyfin_dj_sets` | 9303 | DJ_Sets |
+   | `jellyfin_diogo` | 9399 | JellyDiogo |
+
+   Provisioning "a Jellyfin user" therefore means looping over whichever instances that person
+   should reach, with a **separate API key per instance**. There is no cross-server account.
+
+### Getting an API key
+
+Unlike Immich, Jellyfin has **no admin-token bootstrap from nothing** — creating a key requires
+an already-authenticated session. Two routes:
+
+- **UI:** Dashboard → API Keys → `+`. Simplest, and the one to use for first setup.
+- **API:** `POST /Auth/Keys?App=<name>` (verified: route exists, `401` unauthenticated), using a
+  token obtained from `POST /Users/AuthenticateByName` with an admin username/password
+  (verified: route exists; returns `415` if you omit the JSON content type).
+
+`GET /Auth/Keys` lists them (verified: exists, `401` unauthenticated). Keys are stored in
+`config/data/jellyfin.db`, table `ApiKeys`. **This deployment currently has zero API keys** — one
+must be created before any of the flow below will work.
+
+### The flow
+
+```http
+GET /Users
+Authorization: MediaBrowser Token="<API_KEY>"
+```
+→ verified: exists, `401` without auth. Use it as the **check** step (rule 2) — match on `Name`.
+
+```http
+POST /Users/New
+Authorization: MediaBrowser Token="<API_KEY>"
+Content-Type: application/json
+
+{ "Name": "Full Name", "Password": "<generated>" }
+```
+→ verified: route exists (`401` unauthenticated). Body shape from working knowledge. Returns the
+created user object including its `Id` (a GUID), which the next calls need.
+
+Then, as needed:
+
+| Purpose | Endpoint | Notes |
+|---|---|---|
+| Library access, admin rights, parental rating | `POST /Users/{id}/Policy` | Full policy object; **send the whole thing**, it replaces rather than merges. This is where `IsAdministrator` and `EnabledFolders` live. |
+| Set/replace password | `POST /Users/{id}/Password` | Needed if you created the user without one. |
+| Delete | `DELETE /Users/{id}` | Hard delete — no soft-delete/grace period, unlike Immich. |
+
+**Verify (rule 4)** with a real login rather than trusting the create:
+
+```http
+POST /Users/AuthenticateByName
+Content-Type: application/json
+
+{ "Username": "Full Name", "Pw": "<generated>" }
+```
+→ returns an `AccessToken` and the `User` object on success. Note the field is `Pw`, not
+`Password`, and Jellyfin authenticates by **username**, not email — Jellyfin has no concept of an
+email login.
+
+### A public endpoint worth knowing
+
+`GET /Users/Public` needs **no authentication at all** (verified: `200`). It lists users flagged
+as visible on the login screen. On this deployment it returns `[]` on all four instances, so
+nothing is currently exposed — but if you provision users, be aware that leaving them
+login-screen-visible publishes their names to anyone who can reach the port.
+
+### Probe recipe (Jellyfin)
+
+Same spirit as the Immich one, but you need a key first, and there is no soft-delete — so probe on
+a **non-production instance** or be prepared to hard-delete:
+
+1. `GET /Users` with the key → `200` and an array. Wrong key → `401`.
+2. `POST /Users/New` with a throwaway name → expect the created object; keep the `Id`.
+3. Repeat the identical call → observe the duplicate-name behaviour and **record the exact status
+   and body here**, since it is the guard rule 2 depends on.
+4. `POST /Users/AuthenticateByName` with the throwaway credentials → expect an `AccessToken`.
+5. `DELETE /Users/{id}` → then `GET /Users` to confirm it is gone.
+
+---
+
+## Re-verifying after an upstream upgrade
+
+Cheap and non-destructive — create a probe user on an unroutable domain, exercise it, delete it:
+
+1. `POST /api/admin/users` with `probe-delete-me@example.invalid` → expect `201`, keep the `id`.
+2. Repeat the identical call → expect the duplicate-email `400`.
+3. `POST` with `password`/`name` omitted → expect the validation `400`.
+4. `POST` with a bogus key → expect `401`.
+5. `DELETE /api/admin/users/{id}` with `{"force":true}` → expect `200`, then `GET` to confirm it's gone.
+
+Use `.invalid` (RFC 2606 — guaranteed never to resolve) so a stray notification email can never
+reach a real person.

prompts/deploy-commander.md

@@ -0,0 +1,209 @@
+# Deploy commander
+
+Start a commander with this. It watches the platform repository for a release, rolls it out to every
+server, and reports back.
+
+**It never decides what a deploy needs.** That is written for it, in the release file, by whoever asked
+for the deploy. This agent executes a plan and reports honestly on what happened — it does not improvise
+one.
+
+---
+
+## What wakes you
+
+You are not a loop. Something fires you, and each time you fire you do exactly one thing: check whether
+there is a release to deploy that has not been deployed.
+
+```bash
+git ls-remote --tags origin 'refs/tags/deploy/*'
+```
+
+A release is **a tag**, named `deploy/<version>` — `deploy/2026.08.21.01`. Nothing else triggers you: not
+a commit message, not a branch, not a push.
+
+**Why a tag and not a code word in a commit.** A code word can be written by accident. Every document
+about deploying contains it, this file included, and so does every conversation planning one — so the
+trigger has to be defended with rules about where in the subject line it sits, and a rule like that is
+one somebody eventually gets wrong. A tag cannot be typed into prose. It is created deliberately, it
+names its own version, and `git ls-remote` answers "is there one" in a single call without fetching.
+
+- **No untagged release** → say "no release pending" and stop. Do not pull, do not touch a server.
+- **A `deploy/*` tag with no local record** → that is the release. Continue.
+
+### One attempt per release, and you do not enforce that yourself
+
+**Before you spawn or run anything for a version, claim it:**
+
+```bash
+scripts/deploy-claim.sh claim <version>   # exit 0 = it is yours, exit 1 = do not proceed
+```
+
+Exit 1 means somebody already attempted this release. **Stop. Do not retry it, do not deploy it, and do
+not decide the previous attempt does not count.** Say what the script said and move on.
+
+On success, record it through the same script:
+
+```bash
+scripts/deploy-claim.sh done <version>
+```
+
+**Why this is a script and not a rule you follow.** A watcher loop is written fresh each time one arms, so
+a rule that lives only in this file gets re-implemented from scratch every time — and the wrong version of
+it looks right: test for the `.done` file, which exists only on SUCCESS. Write that and a failing release
+spawns a new agent every poll, all night, each one costing money. That is not hypothetical. It is what the
+first watcher did on 2026-08-21: a careful loop, honest logging, and an unbounded retry inside it.
+
+So the rule is code. `mkdir` is the claim — atomic, one syscall that both tests and takes, no window
+between checking and acting. The script refuses a second claim and keeps refusing, whatever happened to the
+first attempt.
+
+**A failed release is therefore not retried. It is looked at.** Clearing a claim is a deliberate human act
+(`rm -rf …/<version>.claim`) or a new tag. A loop cannot undo its own claim, which is the property that
+makes the storm impossible rather than unlikely.
+
+### The record is LOCAL, not in git
+
+```
+$OFFICER_ROOT/data/deploys/<version>.done
+```
+
+That file's existence means deployed. Written after the rollout, before the report is pushed.
+
+**This is deliberate and it is the important part.** The obvious design makes the pushed report the
+marker — and then a failed `git push` means the release looks undeployed and the next tick **does it all
+again**. Pushes here are genuinely unreliable: gitea's public SSH is rate-limited because every
+connection arrives from one proxy container IP and internet bots get it penalised, which cost several
+retries on 2026-08-21.
+
+So the lock is local and cannot fail for a network reason. The report is an ARTEFACT — valuable, and not
+load-bearing. A missing report is a missing record; a missing lock is a second deploy.
+
+It sits under `data/` rather than in the repository so a `git pull` can never clobber it, and so a fresh
+clone does not arrive believing it has deployed nothing.
+
+## The release
+
+The tag names the version. The commit it points at carries the rest:
+
+- `package.json` — `version` matching the tag, `yyyy.mm.dd.NN` (`2026.08.21.01`), the mobile apps'
+  scheme. `NN` increments for a second deploy on the same day.
+- `deploy/<version>.md` — **the plan**. What to run, in order, and on which hosts. Written by hand.
+
+If the tag and `package.json` disagree, **stop and say so**. One of them is a mistake and guessing which
+is how the wrong tree gets deployed.
+
+If the plan file is missing, **stop and say so**. A release with no plan is not a release you can guess
+at, and guessing is how a database gets a schema it was not ready for.
+
+---
+
+## The servers
+
+**Read `docs/servers.md`.** Do not take a list of hosts from this file — there is deliberately none here,
+because a copy rots the day a server moves and the copy is always the one somebody trusts.
+
+That document holds every machine, its tailnet name, **its login user**, its checkout and whether a
+platform runs on it. Read it at the start of every deploy, and treat the hosts marked as having no
+platform as out of scope.
+
+**The login is not always `pastilhas`.** `deadpool-rezio-bridge` runs as `br`, permanently, and more will.
+Take the user from the table — an orchestrator that hard-codes one fails on that host silently, because
+`Permission denied (publickey)` looks like a broken key rather than a wrong username.
+
+Four things that will catch you out, all verified rather than assumed:
+
+- **Never hard-code the path.** Take it from PM2: the `officer` process's `pm_cwd`. `alpha` has a second,
+  stale checkout, and pulling that one succeeds while changing nothing.
+- **`pm2` is not on a non-interactive PATH.** It is installed under nvm, and nvm lives in `.zshrc`, which
+  a _login_ shell does not read. Use `ssh <host> '$SHELL -ic "pm2 jlist"'`. `-lc` is not enough and
+  returns "command not found" on a machine where pm2 is plainly running.
+- **The git remotes differ.** `fofos` pulls over HTTPS; others over SSH. Never assume a URL — just
+  `git pull --ff-only` in place.
+- **`macbookpro` is macOS and is a laptop.** It may be asleep. Unreachable is a SKIP with a note, not a
+  failure of the deploy.
+
+---
+
+## Rolling it out
+
+Per host, in this order. Do them **one host at a time** and keep a record as you go — a report you
+assemble at the end from memory is a report that will be wrong about the host that failed.
+
+1. **Refuse a dirty tree — TRACKED changes only.** `git status --porcelain --untracked-files=no`. Plain
+   `--porcelain` counts untracked files, and on 2026-08-21 that skipped a host over one stray script a
+   fast-forward could never have touched. Untracked files get a note in the report, not a skip. Never
+   discard someone's work to ship a release.
+2. `git pull --ff-only`. A non-fast-forward means that checkout has diverged: skip and report.
+3. **Update the plugins.** `bun run plugins:update`. Sequential by design, one plugin failing does not
+   stop the rest, and it exits non-zero if any did — so a failure is something you report rather than
+   something you have to notice. A failure here is usually an uncommitted change in that plugin's own
+   checkout.
+4. **Run the plan's commands** for that host, in the order given. If one fails, stop _for that host_ and
+   record what failed and what did not run. Do not continue past a failed `bun install` or `db:push`.
+5. **Restart what the plan names**, and nothing else. Never `pm2 restart ecosystem.config.cjs` — it
+   bounces every sidecar including the vault.
+6. **Check it came back.** `pm2 jlist` again: every process the plan restarted is `online`, and its
+   `restart_time` moved. A process that is `errored` or that never restarted is a failure — report it as
+   one rather than declaring the host done.
+
+If a host fails, **continue to the next one**. A half-rolled-out release that is honestly reported beats
+one that stopped in the middle and said nothing about the four hosts it never reached.
+
+### Skip the host you are running on
+
+Entirely. Do not pull it, do not restart anything on it, and **name it in the report as skipped**.
+
+You are a session inside `officer-claude-code` on that machine. Restarting it ends you mid-rollout —
+which is not a hypothetical: on 2026-08-21 a manual deploy reached its own host last and had to leave
+that one sidecar alone to survive writing its own report.
+
+Restarting only `officer` and leaving your own sidecar behind would work, and it is worse: it puts one
+host in a state nobody chose, a step behind the others, discovered later. So skip the whole host and say
+which one. The owner restarts it deliberately, as the last step, once they are happy with the rest.
+
+Find it with `hostname` and match it against the list — do not assume you are on any particular one.
+
+---
+
+## The report
+
+Write the local record FIRST — `$OFFICER_ROOT/data/deploys/<version>.done` — then commit
+`deploy/<version>.report.md` and push it.
+
+That order matters. The record is what stops a second deploy, and the push is the part that can fail for
+reasons having nothing to do with the rollout. **If the push fails, say so and stop — do not retry into a
+loop and do not treat it as a failed deploy.** The deploy succeeded; its paperwork is pending.
+
+Never tag the report. A tag is the trigger, and a report that tags itself deploys itself forever.
+
+Per host: what its HEAD was before and after, which steps ran, which were skipped and why, which
+processes restarted, and whether they came back. Then one line at the top saying how many hosts
+succeeded out of how many.
+
+Say what you did not do. A host you skipped, a step the plan named that you could not run, a process that
+came back `errored` — those are the parts of a deploy report anybody actually needs, and the only ones
+that are hard to reconstruct afterwards.
+
+---
+
+## Being woken
+
+A background job does it, armed from the composer's background-job dialog with **Exit and hand back to the
+agent, then re-arm**. The wait happens in the shell, not in a model tick — a loop polling
+`git ls-remote`, which is one network call and no tokens.
+
+The loop cannot do a deploy: it cannot read a plan, judge a host or write a report. So on firing it spawns
+a headless agent — `claude -p "<read this file and follow it>" --dangerously-skip-permissions
+--output-format=stream-json --verbose` — and pipes the events through `jq` so the job panel shows progress
+instead of a blank box for twenty minutes.
+
+Two things that loop must get right, both learned by getting them wrong on 2026-08-21:
+
+- **Claim the version first**, per the section above. This is the one that costs money when missed.
+- **Set `OFFICER_ROOT` explicitly.** It is empty in the sidecar's environment, so the lock path resolves to
+  `/data/deploys` and the release looks permanently undeployed.
+
+The deploy therefore runs in a **separate** agent, not in the commander's own conversation: no Command HQ
+row, no cost on the commander's line, and its reasoning lives in that child's transcript. That is the price
+of putting the wait in the shell, and it is worth it — a model-side tick that re-reads a whole conversation
+to learn nothing is the thing being avoided.

public/officer-logo.svg

@@ -1,4 +1,4 @@
-<svg width="1200" height="240" viewBox="-20 40 1200 260" xmlns="http://www.w3.org/2000/svg">
+<svg xmlns="http://www.w3.org/2000/svg" width="900" height="300" viewBox="60 40 900 300">
   <defs>
     <filter id="shadow3d-main-officer">
       <feOffset dx="6" dy="6" in="SourceGraphic" result="offset1"/>
@@ -25,25 +25,26 @@
     </pattern>
 
     <mask id="textMaskOfficer">
-      <text x="580" y="220" text-anchor="middle" class="officer-text-main" fill="white">o<tspan dx="6" dy="30" letter-spacing="14">ff</tspan><tspan dx="-2" dy="-30">icer.dev</tspan></text>
+      <text x="510" y="220" text-anchor="middle" class="officer-text-main" fill="white">Of<tspan dx="16">f</tspan><tspan dx="10">i</tspan><tspan dx="10">c</tspan><tspan dx="10">e</tspan><tspan dx="10">r</tspan></text>
     </mask>
 
     <style>
       .officer-text-main {
-        font-family: system-ui, -apple-system, sans-serif;
+        font-family: system-ui, -apple-system, "System Font", ".SF NS", "Helvetica Neue", sans-serif;
         font-size: 140px;
-        font-weight: 900;
+        font-weight: 700;
         fill: #F4C430;
         stroke: #14532D;
         stroke-width: 3;
         paint-order: stroke fill;
-        letter-spacing: -2px;
-        text-transform: lowercase;
+        letter-spacing: 8px;
+        font-variant-ligatures: none;
+        font-feature-settings: "liga" 0, "clig" 0, "dlig" 0;
       }
     </style>
   </defs>
 
-  <text x="580" y="220" text-anchor="middle" class="officer-text-main" filter="url(#shadow3d-main-officer)">o<tspan dx="6" dy="30" letter-spacing="14">ff</tspan><tspan dx="-2" dy="-30">icer.dev</tspan></text>
+  <text x="510" y="220" text-anchor="middle" class="officer-text-main" filter="url(#shadow3d-main-officer)">Of<tspan dx="16">f</tspan><tspan dx="10">i</tspan><tspan dx="10">c</tspan><tspan dx="10">e</tspan><tspan dx="10">r</tspan></text>
 
   <rect x="0" y="0" width="1160" height="300" fill="url(#gridPatternOfficer)" mask="url(#textMaskOfficer)"/>
 </svg>

scripts/deploy-claim.sh

@@ -0,0 +1,107 @@
+#!/usr/bin/env bash
+#
+# One attempt per release, enforced by the filesystem.
+#
+# A watcher loop asks this before it spawns anything. Exit 0 means "you have the claim, go". Exit 1 means
+# "somebody already had it" — and that answer never changes back, whatever happened to the attempt that
+# held it. A release that fails is not retried; it is looked at.
+#
+# WHY THIS IS A SCRIPT AND NOT A PARAGRAPH IN prompts/deploy-commander.md
+#
+# The watcher loop is written fresh by an agent each time it arms, so a rule that lives only in prose is
+# re-implemented from scratch on every arming, and "skip a version already attempted" has a wrong version
+# that looks right: test the .done file, which only exists on SUCCESS. Write that and a failing deploy
+# spawns a fresh agent every poll, all night, each one costing money. That is exactly what happened on
+# 2026-08-21 — the loop was careful, logged honestly, and still had an unbounded retry in it.
+#
+# WHY BASH AND NOT bun
+#
+# The thing that stops a retry storm must not depend on the checkout being healthy. A deploy that dies on
+# `bun install` is precisely when this has to work.
+#
+# WHY mkdir
+#
+# It is atomic on every POSIX filesystem and it fails if the target exists — one syscall that both tests
+# and claims. `[ -f x ] && ...` is two steps with a race between them, and two pollers a second apart is
+# not a hypothetical when a watcher is restarted.
+set -u
+
+DEPLOYS_DIR="${DEPLOYS_DIR:-$HOME/officer-dev/data/deploys}"
+
+usage() {
+  cat >&2 <<'USAGE'
+usage: deploy-claim.sh claim <version>    take the one attempt for this release (exit 0 = yours, 1 = taken)
+       deploy-claim.sh done  <version>    record that the attempt succeeded
+       deploy-claim.sh state <version>    print unclaimed | claimed | done
+       deploy-claim.sh list               every release this host knows about
+
+To retry a release deliberately: rm -rf <deploys-dir>/<version>.claim
+That is a human act on purpose. A loop cannot undo its own claim.
+USAGE
+  exit 2
+}
+
+version_ok() {
+  # The version reaches here from a git tag, so it is attacker-adjacent input used to build a path.
+  case "$1" in
+    *[!0-9.]* | '' | *..*) return 1 ;;
+    *) return 0 ;;
+  esac
+}
+
+CMD="${1:-}"
+[ -n "$CMD" ] || usage
+
+case "$CMD" in
+  list)
+    mkdir -p "$DEPLOYS_DIR"
+    found=0
+    for d in "$DEPLOYS_DIR"/*.claim; do
+      [ -e "$d" ] || continue
+      found=1
+      v=$(basename "$d" .claim)
+      if [ -f "$DEPLOYS_DIR/$v.done" ]; then s="done"; else s="claimed, no success recorded"; fi
+      printf '%-16s %s\n' "$v" "$s"
+    done
+    [ "$found" = 1 ] || echo "no releases attempted on this host"
+    ;;
+
+  claim | done | state)
+    V="${2:-}"
+    [ -n "$V" ] || usage
+    version_ok "$V" || { echo "deploy-claim: refusing a version that is not digits and dots: $V" >&2; exit 2; }
+    mkdir -p "$DEPLOYS_DIR" || exit 2
+    CLAIM="$DEPLOYS_DIR/$V.claim"
+    DONE="$DEPLOYS_DIR/$V.done"
+
+    case "$CMD" in
+      state)
+        if [ -f "$DONE" ]; then echo done
+        elif [ -d "$CLAIM" ]; then echo claimed
+        else echo unclaimed; fi
+        ;;
+      claim)
+        # A pre-existing .done with no .claim is a release deployed before this script existed. Honour it.
+        if [ -f "$DONE" ]; then
+          echo "deploy-claim: $V is already deployed" >&2
+          exit 1
+        fi
+        if mkdir "$CLAIM" 2>/dev/null; then
+          printf '%s claimed on %s by pid %s\n' "$(date -Iseconds)" "$(hostname)" "$$" > "$CLAIM/by"
+          echo "deploy-claim: $V is yours"
+          exit 0
+        fi
+        printf 'deploy-claim: %s was already attempted — %s\n' "$V" "$(cat "$CLAIM/by" 2>/dev/null || echo 'claim exists')" >&2
+        echo "deploy-claim: NOT retrying. Look at what failed, then rm -rf $CLAIM to allow another attempt." >&2
+        exit 1
+        ;;
+      done)
+        [ -d "$CLAIM" ] || echo "deploy-claim: warning — recording success for $V that was never claimed" >&2
+        printf '%s succeeded on %s\n' "$(date -Iseconds)" "$(hostname)" >> "$DONE"
+        echo "deploy-claim: $V recorded as deployed"
+        ;;
+    esac
+    ;;
+
+  *) usage ;;
+esac

scripts/deploy-claim.test.ts

@@ -0,0 +1,60 @@
+import { afterEach, beforeEach, describe, expect, it } from 'bun:test';
+import { existsSync, mkdtempSync, rmSync, writeFileSync } from 'node:fs';
+import { tmpdir } from 'node:os';
+import { join } from 'node:path';
+
+const SCRIPT = join(import.meta.dir, 'deploy-claim.sh');
+let dir: string;
+
+const run = (...args: string[]) => {
+  const p = Bun.spawnSync([SCRIPT, ...args], { env: { ...process.env, DEPLOYS_DIR: dir } });
+  return { code: p.exitCode, out: p.stdout.toString(), err: p.stderr.toString() };
+};
+
+beforeEach(() => {
+  dir = mkdtempSync(join(tmpdir(), 'deploy-claim-'));
+});
+afterEach(() => rmSync(dir, { recursive: true, force: true }));
+
+// The whole point of this script is that the SECOND caller is refused. Everything else is scaffolding.
+describe('deploy-claim', () => {
+  it('grants once and refuses after, whatever the outcome was', () => {
+    expect(run('claim', '2026.09.01.01').code).toBe(0);
+    expect(run('claim', '2026.09.01.01').code).toBe(1);
+    // Still refused after a failure — no .done was ever written. This is the retry storm not happening.
+    expect(run('claim', '2026.09.01.01').code).toBe(1);
+    expect(run('state', '2026.09.01.01').out.trim()).toBe('claimed');
+  });
+
+  it('refuses a release already deployed, even with no claim beside it', () => {
+    // A release deployed by hand before this script existed. `deploy/2026.08.21.01` is exactly this case.
+    writeFileSync(join(dir, '2026.08.21.01.done'), 'deployed by hand\n');
+    expect(run('claim', '2026.08.21.01').code).toBe(1);
+    expect(run('claim', '2026.08.21.01').err).toContain('already deployed');
+  });
+
+  it('lets a human hand the claim back, and only a human', () => {
+    expect(run('claim', '2026.09.01.02').code).toBe(0);
+    rmSync(join(dir, '2026.09.01.02.claim'), { recursive: true });
+    expect(run('claim', '2026.09.01.02').code).toBe(0);
+  });
+
+  it('records success separately from the claim', () => {
+    run('claim', '2026.09.01.03');
+    run('done', '2026.09.01.03');
+    expect(run('state', '2026.09.01.03').out.trim()).toBe('done');
+    expect(existsSync(join(dir, '2026.09.01.03.done'))).toBe(true);
+  });
+
+  // The version comes off a git tag, and it is concatenated into a path.
+  it('refuses anything that is not digits and dots', () => {
+    for (const bad of ['../../etc/passwd', 'x; rm -rf /', '2026.08.21.01/..', '', 'v1']) {
+      expect(run('claim', bad).code).toBe(2);
+    }
+    expect(existsSync(join(dir, '..', 'etc'))).toBe(false);
+  });
+
+  it('says so plainly when nothing has been attempted', () => {
+    expect(run('list').out).toContain('no releases attempted');
+  });
+});

scripts/doctor.ts

@@ -0,0 +1,86 @@
+#!/usr/bin/env bun
+/**
+ * `bun run doctor` — check the invariants that provisioning cannot check for itself.
+ *
+ * ── Why ──
+ *
+ * Every step in `provisionOsAccount` is deliberately non-fatal, so "nothing reported an error" is much
+ * weaker than it sounds. The 2026-08-19 case: a fresh server, an Admin granted files, chat and terminal,
+ * every named ACL entry correct, provisioning clean — and the account reached nothing, because a later
+ * `chmod 700` had recomputed the ACL mask to `---`. It was found by a person, three days and one
+ * confusing UI message later.
+ *
+ * This is the same posture as `assertPermissionTotality`: state the invariant, verify it, name the fix.
+ * The difference is that this reports rather than refusing to boot — a broken member must not stop the
+ * platform serving everyone else.
+ *
+ * Read-only. It changes nothing, so it is safe on a production box.
+ *
+ * Exit 1 if anything is wrong, so it can gate the end of a setup run.
+ */
+import { getUsers, getAllRoleGrants, OWNER_USER_ID } from 'officerdb';
+import { checkOsAccount } from '../src/servers/os-user-verify';
+import { PERMISSIONS } from '../src/servers/permissions/registry';
+
+const red = (s: string) => `\x1b[31m${s}\x1b[0m`;
+const green = (s: string) => `\x1b[32m${s}\x1b[0m`;
+const dim = (s: string) => `\x1b[2m${s}\x1b[0m`;
+
+let problems = 0;
+const fail = (what: string, fix: string) => {
+  problems++;
+  console.log(`${red('✗')} ${what}`);
+  console.log(`  ${dim(fix)}`);
+};
+const pass = (what: string) => console.log(`${green('✓')} ${what}`);
+
+console.log('\n── Accounts ──\n');
+
+const users = await getUsers();
+
+for (const user of users) {
+  if (user.id === OWNER_USER_ID) {
+    // The owner runs AS the service user, in the service user's own home. There is no Linux account to
+    // check and no ACL to get wrong — `osUser: null` is correct here, not a finding.
+    pass(`${user.email} ${dim('(owner — runs as the service account)')}`);
+    continue;
+  }
+
+  const check = await checkOsAccount(user.email, user.osUser ?? null);
+  if (check.ok) {
+    pass(`${user.email} ${dim(`(${user.role}, os_user=${check.osUser}, home readable)`)}`);
+  } else {
+    fail(`${user.email} (${user.role}) — ${check.problem}`, `home: ${check.home}`);
+  }
+}
+
+console.log('\n── Grants that cannot take effect ──\n');
+
+// A `confined` grant is silently dropped for an account with no Linux user. That is correct behaviour and
+// completely invisible: the grant is in the table, the UI shows it ticked, and it resolves to nothing.
+const confinedKeys = new Set(PERMISSIONS.filter((p) => p.kind === 'confined').map((p) => p.key));
+const grants = await getAllRoleGrants();
+const unconfined = users.filter((u) => u.id !== OWNER_USER_ID && !u.osUser);
+
+if (!unconfined.length) {
+  pass('every non-owner account has a Linux user');
+} else {
+  for (const user of unconfined) {
+    const dropped = grants
+      .filter((g) => g.role === user.role && confinedKeys.has(g.permission))
+      .map((g) => g.permission);
+    if (!dropped.length) continue;
+    fail(
+      `${user.email} (${user.role}) holds ${dropped.join(', ')} — all dropped, because the account has no Linux user`,
+      'Fix: Settings → User management → provision, or POST /api/users/<id>/provision-linux',
+    );
+  }
+}
+
+console.log('');
+if (problems === 0) {
+  console.log(green('No problems found.\n'));
+  process.exit(0);
+}
+console.log(red(`${problems} problem${problems === 1 ? '' : 's'} found.\n`));
+process.exit(1);

scripts/fix-env.ts

@@ -0,0 +1,14 @@
+// `bun run fix:env` — give every variable in `.env` the `OFFICERDEV_` prefix.
+//
+// A thin front for `servers/env-prefix.ts`, which is where the rules live. Deliberately that way round:
+// the same conversion runs on every restart from the server settings screen, and a router reaching into
+// `scripts/` to do it would make a maintenance script part of the request path.
+
+import { join } from 'node:path';
+import { fixEnvFile } from '../src/servers/env-prefix';
+
+const path = process.argv[2] ?? join(process.cwd(), '.env');
+const { changed, alreadyPrefixed } = fixEnvFile(path);
+
+if (changed.length) console.log(`[fix:env] prefixed ${changed.length}: ${changed.join(', ')}`);
+else console.log(`[fix:env] nothing to do (${alreadyPrefixed} already prefixed)`);

scripts/gen-ecosystem.ts

@@ -0,0 +1,124 @@
+#!/usr/bin/env bun
+/**
+ * Rebuild ecosystem.config.cjs from scratch: the core processes, plus one PM2 entry for every sidecar
+ * this install actually has enabled — marketplace plugins that ship one, and app-store services.
+ *
+ * ── Where each part of the answer comes from, and why ──
+ *
+ * CORE_PROCESSES is sourced from `scripts/setup/officer-setup/lib/services.sh` rather than duplicated
+ * here — that file is the one officer-setup itself writes from, and a second hardcoded list is a second
+ * thing to keep in sync with it. This shells out to bash to read the array directly, so there is exactly
+ * one list.
+ *
+ * Plugin sidecars come from crossing two facts that live in different places on purpose: `discoverPlugins`
+ * says what is on DISK (a plugin with a `sidecar/` directory), and `listPluginInstalls` says what the
+ * owner has actually installed and enabled in Postgres. Neither alone is the answer — a plugin can be on
+ * disk but never installed, or installed and then disabled without being removed.
+ *
+ * App-store sidecars are the same cross, against `sidecarInstalls`/`CATALOGUE` instead: every catalogue
+ * entry names a `process`, and `src/servers/sidecar/<id>/index.ts` is where it lives — a convention, not
+ * a lookup, the same way `officer-claude-code`'s script path is not stored anywhere but its own directory
+ * name.
+ *
+ * ── Why this exists alongside `servers/plugins/ecosystem.ts` ──
+ *
+ * That file does the incremental add/remove at install/uninstall time — a live, running edit. This is
+ * the other direction: rebuild the whole file from what the database says exists, for a fresh clone,
+ * drift, or restoring from a backup. Neither replaces the other.
+ */
+import { spawnSync } from 'node:child_process';
+import { join } from 'node:path';
+import { CATALOGUE } from '../src/servers/app-store/catalogue';
+import { PLATFORM_DIR } from '../src/servers/data-path';
+import { discoverPlugins } from '../src/servers/plugins/discover';
+import { pluginPm2Entry, type Pm2App } from '../src/servers/plugins/ecosystem';
+import { listPluginInstalls, listSidecarInstalls } from 'officerdb';
+
+const done = (message: string): never => {
+  console.log(`[gen:ecosystem] ${message}`);
+  process.exit(0);
+};
+const fail = (message: string): never => {
+  console.error(`[gen:ecosystem] ${message}`);
+  process.exit(1);
+};
+
+/** The single source of truth for the core list is the bash array itself, not a copy of it. */
+function coreProcesses(): Pm2App[] {
+  const servicesLib = join(PLATFORM_DIR, 'scripts/setup/officer-setup/lib/services.sh');
+  const result = spawnSync('bash', ['-c', `source '${servicesLib}' && printf '%s\\n' "\${CORE_PROCESSES[@]}"`], {
+    encoding: 'utf-8',
+  });
+  if (result.status !== 0) {
+    fail(`could not read CORE_PROCESSES from ${servicesLib}: ${result.stderr || `exit ${result.status}`}`);
+  }
+  return result.stdout
+    .split('\n')
+    .map((line) => line.trim())
+    .filter(Boolean)
+    .map((line) => {
+      const [name, script, args] = line.split('|');
+      if (!name || !script) return fail(`unreadable CORE_PROCESSES entry: "${line}"`);
+      return { name, script, args, cwd: PLATFORM_DIR, watch: false };
+    });
+}
+
+/** Installed AND enabled AND actually ships a sidecar — all three, or it gets no entry. */
+async function pluginEntries(): Promise<Pm2App[]> {
+  const [installs, { plugins }] = await Promise.all([listPluginInstalls(), discoverPlugins()]);
+  const enabled = new Set(installs.filter((i) => i.enabled).map((i) => i.appName));
+  return plugins.filter((p) => enabled.has(p.appName)).flatMap((p) => pluginPm2Entry(p) ?? []);
+}
+
+/** Installed AND enabled. Unlike a plugin, every catalogue entry has a `process` — none are filtered out. */
+async function appStoreEntries(): Promise<Pm2App[]> {
+  const installs = await listSidecarInstalls();
+  const enabled = new Set(installs.filter((i) => i.enabled).map((i) => i.sidecarId));
+  return CATALOGUE.filter((entry) => enabled.has(entry.id)).map((entry) => ({
+    name: entry.process,
+    script: 'bun',
+    args: `run src/servers/sidecar/${entry.id}/index.ts`,
+    cwd: PLATFORM_DIR,
+    watch: false,
+  }));
+}
+
+const [core, plugins, appStore] = await Promise.all([
+  Promise.resolve(coreProcesses()),
+  pluginEntries(),
+  appStoreEntries(),
+]);
+
+const apps = [...core, ...plugins, ...appStore];
+
+// The three sources should never collide by construction (core names are fixed, plugin names are
+// `officer-<appName>`, app-store names come from the catalogue) — but a later entry silently shadowing
+// an earlier one in the written file is a worse failure than refusing outright.
+const seen = new Set<string>();
+for (const app of apps) {
+  if (seen.has(app.name)) fail(`duplicate pm2 process name: ${app.name} — refusing to write a shadowed entry`);
+  seen.add(app.name);
+}
+
+const HEADER = `// Generated by "bun gen:ecosystem" (scripts/gen-ecosystem.ts). Not in git, and not meant to
+// be — it describes THIS install. Re-run it any time the database's idea of what is installed and
+// enabled has moved on from what this file says: a fresh clone, drift, restoring from a backup.
+//
+// \`cwd\` is pinned on every app for two reasons. Bun auto-loads .env from the working directory (and the
+// pty sidecar does \`import 'dotenv/config'\`), so without it a process started from anywhere else comes
+// up with no POSTGRES_URL. And src/servers/data-path.ts derives the install root as the PARENT of the
+// working directory, so a wrong cwd does not fail — it relocates data/, capabilities/ and dockers/
+// somewhere else entirely. \`assertInstallLayout\` refuses to boot when that happens.
+`;
+
+const lines = apps.map(
+  (app) =>
+    `    { name: '${app.name}', script: '${app.script}'` +
+    (app.args ? `, args: '${app.args}'` : '') +
+    `, cwd: '${app.cwd}', watch: ${app.watch} },`,
+);
+
+const dest = join(PLATFORM_DIR, 'ecosystem.config.cjs');
+await Bun.write(dest, `${HEADER}module.exports = {\n  apps: [\n${lines.join('\n')}\n  ],\n};\n`);
+
+done(`wrote ${dest} — ${core.length} core, ${plugins.length} plugin, ${appStore.length} app-store`);

scripts/gen-index.ts

@@ -27,17 +27,23 @@ const output = join(root, 'src/apps/officer-web/index.gen.html');
 const argUrl = process.argv[2]?.trim();
 
 const envPath = join(root, '.env');
-if (!argUrl && !process.env.PUBLIC_URL && existsSync(envPath)) {
+// Prefixed first, bare as the fallback — the same rule every other consumer follows. This script reads
+// the file itself rather than relying on the environment, because it is run by hand as often as by the
+// setup, and `bun run gen:index` in a bare shell has no `.env` loaded.
+const envValue = (): string | undefined => process.env.OFFICERDEV_PUBLIC_URL ?? process.env.PUBLIC_URL;
+
+if (!argUrl && !envValue() && existsSync(envPath)) {
   for (const line of (await Bun.file(envPath).text()).split('\n')) {
-    const match = line.match(/^\s*PUBLIC_URL\s*=\s*(.*)$/);
-    if (match) process.env.PUBLIC_URL = match[1]!.trim().replace(/^["']|["']$/g, '');
+    // Either spelling, so a file that has been through `bun run fix:env` and one that has not both work.
+    const match = line.match(/^\s*(?:OFFICERDEV_)?PUBLIC_URL\s*=\s*(.*)$/);
+    if (match) process.env.OFFICERDEV_PUBLIC_URL = match[1]!.trim().replace(/^["']|["']$/g, '');
   }
 }
 
-const publicUrl = (argUrl || process.env.PUBLIC_URL || '').replace(/\/+$/, '');
+const publicUrl = (argUrl || envValue() || '').replace(/\/+$/, '');
 if (!publicUrl) {
   console.error('[gen-index] no public URL. Pass one — `bun gen:index https://officer.example.com` —');
-  console.error('[gen-index] or set PUBLIC_URL in .env.');
+  console.error('[gen-index] or set OFFICERDEV_PUBLIC_URL in .env.');
   process.exit(1);
 }
 

scripts/generate-themes.ts

@@ -0,0 +1,172 @@
+#!/usr/bin/env bun
+import { existsSync, readdirSync, readFileSync, writeFileSync } from 'node:fs';
+import { join } from 'node:path';
+import { loadPalette } from './themes/palette';
+import { mapPalette } from './themes/map';
+import { TEXT, UI, contrast } from './themes/contrast';
+
+// Generate Officer themes from Omarchy's, and refuse to write an unreadable one.
+//
+//   bun scripts/generate-themes.ts <path-to-omarchy-checkout> [--check]
+//
+// `--check` writes nothing and reports what would change, which is what a test runs.
+//
+// ── Why generated rather than authored ──
+//
+// Twenty-two themes hand-written against an 84-token contract will disagree, and the previous attempt
+// proved it: eleven themes, 34 tokens each, every surface missing. One mapping means one place to be
+// wrong and one place to fix — and a palette that Omarchy updates is a re-run rather than a rewrite.
+//
+// The Duck and Plain themes are NOT generated. They are Officer's own, authored, and this must never
+// overwrite them.
+
+const OMARCHY = process.argv[2];
+const CHECK = process.argv.includes('--check');
+const THEME_DIR = join(import.meta.dir, '..', 'src', 'workspaces', 'theme', 'css', 'themes');
+
+if (!OMARCHY || !existsSync(join(OMARCHY, 'themes'))) {
+  console.error('usage: bun scripts/generate-themes.ts <path-to-omarchy-checkout> [--check]');
+  process.exit(1);
+}
+
+/** Officer's own. Generated themes may never take these names. */
+const AUTHORED = new Set(['duck-light', 'duck-dark', 'plain']);
+
+const title = (id: string): string =>
+  id
+    .split('-')
+    .map((w) => (w.length <= 2 ? w.toUpperCase() : w[0]!.toUpperCase() + w.slice(1)))
+    .join(' ');
+
+/**
+ * The pairs that must be readable, checked after mapping rather than trusted.
+ *
+ * `ensureContrast` already guarantees most of these — this is the belt to that braces. A mapping change
+ * that quietly breaks one shows up here as a failure instead of as somebody squinting at a screenshot.
+ */
+const PAIRS: [string, string, number][] = [
+  ['--foreground', '--background', TEXT],
+  ['--card-foreground', '--card', TEXT],
+  ['--popover-foreground', '--popover', TEXT],
+  ['--muted-foreground', '--muted', TEXT],
+  ['--accent-foreground', '--accent', TEXT],
+  ['--primary-foreground', '--primary', TEXT],
+  ['--secondary-foreground', '--secondary', TEXT],
+  ['--cta-foreground', '--cta', TEXT],
+  ['--success-foreground', '--success', TEXT],
+  ['--warning-foreground', '--warning', TEXT],
+  ['--destructive-foreground', '--destructive', TEXT],
+  ['--info-foreground', '--info', TEXT],
+  ['--busy-foreground', '--busy', TEXT],
+  ['--busy-ink', '--background', TEXT],
+  // Not `--border` at UI: a panel edge is decoration and 3:1 would ring everything in a hard line. What
+  // must be visible is a form field's boundary and the focus ring — the two that carry information.
+  ['--input', '--background', 1.35],
+  ['--ring', '--background', UI],
+];
+
+let failures = 0;
+let written = 0;
+const generated: { id: string; name: string; dark: boolean }[] = [];
+
+for (const id of readdirSync(join(OMARCHY, 'themes')).sort()) {
+  const toml = join(OMARCHY, 'themes', id, 'colors.toml');
+  if (!existsSync(toml)) continue;
+  if (AUTHORED.has(id)) {
+    console.log(`  ${id}: SKIPPED — Officer's own theme, never generated`);
+    continue;
+  }
+
+  const palette = loadPalette(toml);
+  const tokens = mapPalette(palette);
+
+  for (const [fgKey, bgKey, min] of PAIRS) {
+    const fg = tokens[fgKey]!;
+    const bg = tokens[bgKey]!;
+    // Only flat hex pairs are checkable — a translucent surface depends on what is behind it.
+    if (!fg.startsWith('#') || !bg.startsWith('#')) continue;
+    const ratio = contrast(fg, bg);
+    if (ratio < min) {
+      console.error(`  ${id}: ${fgKey} on ${bgKey} is ${ratio.toFixed(2)}:1, needs ${min}:1`);
+      failures++;
+    }
+  }
+
+  const body = Object.entries(tokens)
+    .map(([k, v]) => `  ${k}: ${v};`)
+    .join('\n');
+  const css = `/*
+  ${title(id)} — generated from Omarchy's palette by scripts/generate-themes.ts.
+
+  DO NOT EDIT BY HAND. Change the mapping in scripts/themes/map.ts and re-run, so every theme moves
+  together — the previous port drifted precisely because each file was edited on its own.
+
+  Source: basecamp/omarchy themes/${id}/colors.toml (mode: ${palette.mode})
+*/
+[data-theme='${id}'] {
+${body}
+}
+`;
+
+  const path = join(THEME_DIR, `${id}.css`);
+  const before = existsSync(path) ? readFileSync(path, 'utf8') : '';
+  if (before !== css) {
+    if (!CHECK) writeFileSync(path, css);
+    written++;
+  }
+  generated.push({ id, name: title(id), dark: palette.mode === 'dark' });
+}
+
+console.log(`\n  ${generated.length} themes, ${written} ${CHECK ? 'would change' : 'written'}, ${failures} contrast failures`);
+
+if (!CHECK) {
+  // The registry and the stylesheet index, rewritten from what was generated so a theme cannot exist as
+  // a file nobody imports — which is a theme that silently does not work.
+  const imports = generated.map((t) => `@import './themes/${t.id}.css';`).join('\n');
+  const indexPath = join(THEME_DIR, '..', 'index.css');
+  const index = readFileSync(indexPath, 'utf8');
+  const marker = '/* Palette themes';
+  const head = index.slice(0, index.indexOf(marker));
+  writeFileSync(
+    indexPath,
+    `${head}/* Palette themes — generated from Omarchy by scripts/generate-themes.ts. Do not edit this block. */\n${imports}\n\n@import './chrome.css';\n`,
+  );
+  console.log(`  index.css: ${generated.length} imports`);
+
+  // The REGISTRY, from the same list. A theme file nothing imports does not work; a theme imported but
+  // absent from here does not appear in the picker. Both were true of the previous port — catppuccin-mocha
+  // outlived its stylesheet — so both halves are written from one source rather than maintained by hand.
+  //
+  // Display names get a pass for the ones a slug cannot spell.
+  const NICER: Record<string, string> = {
+    'rose-pine': 'Rosé Pine',
+    'retro-82': 'Retro 82',
+    'tokyo-night': 'Tokyo Night',
+    'catppuccin': 'Catppuccin Mocha',
+    'catppuccin-latte': 'Catppuccin Latte',
+  };
+  const FAMILY: Record<string, string> = { catppuccin: 'Catppuccin', 'catppuccin-latte': 'Catppuccin' };
+
+  const rows = generated
+    .map((t) => {
+      const name = NICER[t.id] ?? t.name;
+      const family = FAMILY[t.id] ?? name;
+      return `  { id: '${t.id}', name: '${name}', family: '${family}', applyDarkClass: ${t.dark} },`;
+    })
+    .join('\n');
+
+  const registryPath = join(import.meta.dir, '..', 'src', 'workspaces', 'theme', 'src', 'themes.ts');
+  const registry = readFileSync(registryPath, 'utf8');
+  const START = '  /* GENERATED THEMES — start */';
+  const END = '  /* GENERATED THEMES — end */';
+  if (!registry.includes(START)) {
+    console.error('  themes.ts: missing the GENERATED THEMES markers — not touching it');
+    process.exit(1);
+  }
+  const registryHead = registry.slice(0, registry.indexOf(START) + START.length);
+  const registryTail = registry.slice(registry.indexOf(END));
+  writeFileSync(registryPath, `${registryHead}\n${rows}\n${registryTail}`);
+  console.log(`  themes.ts: ${generated.length} entries`);
+}
+
+process.exit(failures > 0 ? 1 : 0);

scripts/marketplace-sync.ts

@@ -123,9 +123,7 @@ async function syncOne(dir: string): Promise<Sync | null> {
   });
 
   console.log(
-    `[marketplace-sync] ${appName} ← ${repoUrl}${readme ? '' : ' (no README)'}${
-      has.web ? '' : ' (no web/)'
-    }`,
+    `[marketplace-sync] ${appName} ← ${repoUrl}${readme ? '' : ' (no README)'}${has.web ? '' : ' (no web/)'}`,
   );
   return { appName, developer, repoUrl };
 }
@@ -133,16 +131,27 @@ async function syncOne(dir: string): Promise<Sync | null> {
 /**
  * A raw-file URL for a Gitea repository.
  *
- * `[open]` Gitea-shaped and branch-`main`-shaped. A GitHub-hosted plugin needs the other form, and a
- * repo whose default branch is `master` gets a 404. Fine while every plugin is on our own Gitea; the
- * moment one is not, this needs the forge to be part of the registry rather than inferred.
+ * `HEAD`, not a branch name. This said `branch/main` and the comment here warned that "a repo whose
+ * default branch is `master` gets a 404" — which then happened: four plugin repositories were created
+ * on `master` and every one of them served a broken image in the marketplace on any server that reads
+ * the remote catalogue. A warning in a comment is not a guard.
+ *
+ * Gitea resolves `HEAD` to whatever the default branch is, so the question stops being asked. Verified
+ * against both spellings.
+ *
+ * `[open]` Still Gitea-shaped. A GitHub-hosted plugin needs the other form, and that wants the forge to
+ * be part of the registry rather than inferred from the URL.
  */
 function rawUrl(repoUrl: string, path: string): string {
-  return `${repoUrl.replace(/\.git$/, '')}/raw/branch/main/${path}`;
+  return `${repoUrl.replace(/\.git$/, '')}/raw/HEAD/${path}`;
 }
 
 const dirs = existsSync(PLUGINS_DIR)
-  ? [...new Bun.Glob('*/').scanSync({ cwd: PLUGINS_DIR, onlyFiles: false })].map((d) => join(PLUGINS_DIR, d))
+  ? [...new Bun.Glob('*/').scanSync({ cwd: PLUGINS_DIR, onlyFiles: false })]
+      // Not a plugin: the shared install target for plugin npm packages, which lives here because bun
+      // resolves node_modules up the tree. `discoverPlugins` skips it for the same reason.
+      .filter((d) => d.replace(/\/$/, '') !== 'node_modules')
+      .map((d) => join(PLUGINS_DIR, d))
   : [];
 
 if (dirs.length === 0) {

scripts/notify-test.ts

@@ -0,0 +1,108 @@
+#!/usr/bin/env bun
+/**
+ * Fire one notification at a registered device, to prove the pipe.
+ *
+ *   bun run notify:test                          # officer/core:test to user 1
+ *   bun run notify:test --user 2                 # someone else
+ *   bun run notify:test --code officer/core:job --id 412
+ *
+ * Talks to the sidecar over loopback, which is how every other producer reaches it — so this exercises
+ * the real path rather than a special one. It reports what the channels actually did, because the thing
+ * that wastes a day here is that a send with nothing configured returns success and delivers nothing.
+ */
+import { readFileSync } from 'node:fs';
+import { join } from 'node:path';
+import { DATA_PATH } from '../src/servers/data-path';
+
+const arg = (name: string): string | undefined => {
+  const i = process.argv.indexOf(`--${name}`);
+  return i === -1 ? undefined : process.argv[i + 1];
+};
+
+const code = arg('code') ?? 'officer/core:test';
+const userId = Number(arg('user') ?? 1);
+const id = arg('id');
+const n = arg('n');
+
+const addressFile = join(DATA_PATH, 'notify-server.url');
+
+let base: string;
+try {
+  base = readFileSync(addressFile, 'utf8').trim();
+} catch {
+  console.error(`[notify:test] no address at ${addressFile}`);
+  console.error('              officer-notify is not running — `pm2 start ecosystem.config.cjs --only officer-notify`');
+  process.exit(1);
+}
+
+// The file is only a hint: a crash leaves it behind, so prove the listener is there before believing it.
+let health: { ok: boolean; channels: string[] };
+try {
+  const res = await fetch(`${base}/_health`, { signal: AbortSignal.timeout(5000) });
+  health = (await res.json()) as typeof health;
+} catch {
+  console.error(`[notify:test] ${base} is not answering — stale address file, or the sidecar is down`);
+  process.exit(1);
+}
+
+console.log(`[notify:test] sidecar ${base}`);
+console.log(`[notify:test] channels configured: ${health.channels.length ? health.channels.join(', ') : 'NONE'}`);
+
+if (health.channels.length === 0) {
+  // Said before sending rather than after, because the send below WILL return ok and deliver nothing.
+  console.error('');
+  console.error('[notify:test] nothing is configured, so this cannot reach a device.');
+  console.error('              APNs needs apns-p8, apns-key-id and apns-team-id:');
+  console.error('                cat AuthKey_XXXXXXXXXX.p8 | bun run secret:set apns-p8');
+  console.error('                echo -n XXXXXXXXXX | bun run secret:set apns-key-id');
+  console.error('                echo -n RGHC4NJC77 | bun run secret:set apns-team-id');
+  console.error('              then: pm2 restart officer-notify');
+  process.exit(1);
+}
+
+const body = {
+  code,
+  userId,
+  ...(id === undefined ? {} : { id }),
+  ...(n === undefined ? {} : { n: Number(n) }),
+};
+
+console.log(`[notify:test] sending ${JSON.stringify(body)}`);
+
+const res = await fetch(`${base}/_officer/notify`, {
+  method: 'POST',
+  headers: { 'content-type': 'application/json', 'X-Officer-User': String(userId) },
+  body: JSON.stringify(body),
+  signal: AbortSignal.timeout(20_000),
+});
+
+const out = (await res.json()) as { ok?: boolean; error?: string; results?: Array<Record<string, unknown>> };
+
+if (!res.ok || out.error) {
+  console.error(`[notify:test] refused (${res.status}): ${out.error ?? 'unknown'}`);
+  process.exit(1);
+}
+
+const results = out.results ?? [];
+if (results.length === 0) {
+  console.error('[notify:test] no channel took it — nothing configured, or no device registered for this user');
+  process.exit(1);
+}
+
+let delivered = 0;
+for (const r of results) {
+  const sent = Number(r.sent ?? 0);
+  delivered += sent;
+  const dead = Array.isArray(r.dead) && r.dead.length ? `, ${r.dead.length} dead token(s) deleted` : '';
+  console.log(`[notify:test]   ${r.channel}: sent ${sent}, failed ${r.failed}${dead}`);
+}
+
+// `sent: 0, failed: 0` is the quiet one: the channel is configured but the user has no device of that
+// platform registered. Worth naming, because it looks like success.
+if (delivered === 0) {
+  console.error('[notify:test] nothing was delivered — is a device registered for this user?');
+  console.error('              check: GET /api/notify/_officer/devices');
+  process.exit(1);
+}
+
+console.log(`[notify:test] delivered to ${delivered} device(s) — look at your phone`);

scripts/nudge-agent.ts

@@ -0,0 +1,162 @@
+#!/usr/bin/env bun
+/**
+ * Poke a chat session on a timer, but ONLY when it is idle.
+ *
+ *   bun run nudge -- --session <id> --every 3600 --message "Continue. If blocked, write it down and move on."
+ *   bun run nudge -- --list                       # which sessions exist, and which are working
+ *
+ * ── What this is for ──
+ *
+ * An agent left running overnight. It is told to record blockers and continue rather than wait for a human,
+ * and mostly it does — but a turn that ends leaves it idle until someone types, and at 3am nobody does.
+ * This is the someone.
+ *
+ * ── Why it is a script and not a queue job ──
+ *
+ * The platform has no recurring scheduler and no way to put a prompt into a live session: `pushTurn` has
+ * exactly one caller, the chat websocket. Adding both is real work in the queue engine and the chat API,
+ * and doing it tonight would mean editing a repository an agent is currently working in.
+ *
+ * So this talks to the same doors a browser does — `GET /api/chat/sessions` to see the state, then the chat
+ * websocket to speak. Nothing new is exposed and nothing in the platform changes. If the loop is ever worth
+ * making permanent, it belongs in the queue with a real trigger, and this script is what that would replace.
+ *
+ * ── The one rule ──
+ *
+ * NEVER interrupt work. A nudge sent mid-turn is at best noise and at worst a derail, so `isGenerating` is
+ * checked immediately before speaking and a working session is skipped silently. The point is to restart a
+ * stopped agent, not to hurry a busy one.
+ */
+import { sign } from '@@/jwt';
+import { getOwnerUser } from 'officerdb';
+
+const arg = (name: string): string | undefined => {
+  const i = process.argv.indexOf(`--${name}`);
+  return i === -1 ? undefined : process.argv[i + 1];
+};
+const has = (name: string) => process.argv.includes(`--${name}`);
+
+const PORT = process.env.OFFICERDEV_PORT ?? process.env.PORT ?? '9000';
+const BASE = `http://127.0.0.1:${PORT}`;
+
+const owner = await getOwnerUser();
+if (!owner) {
+  console.error('[nudge] no owner account — nothing to sign as');
+  process.exit(1);
+}
+// Minted here rather than asked for: this runs on the box, as the owner, and a token pasted into a shell
+// is a token in the shell history.
+const token = await sign({ id: owner.id, email: owner.email, username: owner.username }, '2d');
+
+type Session = { id: string; title?: string; isGenerating?: boolean; updatedAt?: string };
+
+async function sessions(): Promise<Session[]> {
+  // `?all=1` — every working directory, not just the default one. Without it the list is scoped to the
+  // owner's home and an agent started in a project directory is simply absent, which reads as "the session
+  // is gone" and stops the loop. This is the list's own Latest tab, so the shape is identical.
+  const res = await fetch(`${BASE}/api/chat/sessions?all=1`, {
+    headers: { authorization: `Bearer ${token}` },
+    signal: AbortSignal.timeout(20_000),
+  });
+  if (!res.ok) throw new Error(`GET /api/chat/sessions → ${res.status}`);
+  const body = (await res.json()) as { sessions?: Session[] } | Session[];
+  return Array.isArray(body) ? body : (body.sessions ?? []);
+}
+
+if (has('list')) {
+  for (const s of await sessions()) {
+    console.log(`${s.isGenerating ? 'WORKING ' : 'idle    '} ${s.id}  ${s.title ?? ''}`);
+  }
+  process.exit(0);
+}
+
+const sessionId = arg('session');
+const message =
+  arg('message') ?? 'Continue with the next task. If you are blocked, write the blocker down and move on.';
+const every = Number(arg('every') ?? 3600);
+/** A ceiling, because a loop nobody is watching should end on its own. Twelve hours by default. */
+const max = Number(arg('max') ?? 12);
+
+if (!sessionId) {
+  console.error('[nudge] --session <id> is required. Use --list to find one.');
+  process.exit(2);
+}
+
+/** Send one prompt, the way the browser does. Resolves once the server acknowledges the turn. */
+function nudge(): Promise<void> {
+  return new Promise((resolve, reject) => {
+    const proto = 'ws';
+    const ws = new WebSocket(`${proto}://127.0.0.1:${PORT}/api/chat/ws?token=${token}&sessionId=${sessionId}`);
+    const done = (err?: Error) => {
+      try {
+        ws.close();
+      } catch {
+        /* already closing */
+      }
+      err ? reject(err) : resolve();
+    };
+
+    const timer = setTimeout(() => done(new Error('no acknowledgement within 30s')), 30_000);
+
+    ws.onopen = () => ws.send(JSON.stringify({ type: 'chat', prompt: message, sessionId }));
+    ws.onmessage = (ev) => {
+      // Anything that is not an error means the turn was taken. We do not wait for it to FINISH — that is
+      // the agent's night, not ours.
+      try {
+        const msg = JSON.parse(String(ev.data)) as { type?: string; message?: string };
+        if (msg.type === 'error') {
+          clearTimeout(timer);
+          return done(new Error(msg.message ?? 'server refused the prompt'));
+        }
+        if (msg.type === 'assistant:delta' || msg.type === 'assistant:text' || msg.type === 'tool:start') {
+          clearTimeout(timer);
+          return done();
+        }
+      } catch {
+        /* not json — ignore */
+      }
+    };
+    ws.onerror = () => {
+      clearTimeout(timer);
+      done(new Error('websocket error'));
+    };
+  });
+}
+
+const stamp = () => new Date().toISOString().slice(11, 19);
+
+console.log(`[nudge] session ${sessionId} — every ${every}s, at most ${max} times`);
+console.log(`[nudge] message: ${message}`);
+
+for (let sent = 0, tick = 0; sent < max; tick++) {
+  if (tick > 0) await Bun.sleep(every * 1000);
+
+  let state: Session | undefined;
+  try {
+    state = (await sessions()).find((s) => s.id === sessionId);
+  } catch (err) {
+    console.error(`[nudge ${stamp()}] could not read sessions: ${err instanceof Error ? err.message : err}`);
+    continue; // a server restart is not a reason to give up on the night
+  }
+
+  if (!state) {
+    console.error(`[nudge ${stamp()}] session ${sessionId} is gone — stopping`);
+    break;
+  }
+
+  if (state.isGenerating) {
+    console.log(`[nudge ${stamp()}] working — skipped`);
+    continue;
+  }
+
+  try {
+    await nudge();
+    sent++;
+    console.log(`[nudge ${stamp()}] idle — nudged (${sent}/${max})`);
+  } catch (err) {
+    console.error(`[nudge ${stamp()}] failed: ${err instanceof Error ? err.message : err}`);
+  }
+}
+
+console.log(`[nudge ${stamp()}] finished`);
+process.exit(0);

scripts/provision-user-dirs.ts

@@ -12,14 +12,21 @@
 // Docker container. That architecture is gone. What survives from it is the useful part: a user has a
 // root, and a home inside it.
 //
-// Idempotent — an existing directory is left exactly as it is, so re-running is safe.
+// Idempotent — an existing directory keeps its contents, so re-running is safe. It does NOT keep its
+// mode: provisionUserDirs re-asserts 711/700 on every run, tolerating the EPERM from a `home` that has
+// since become the member's own. Re-running is therefore also the repair.
 
-import { mkdirSync, existsSync } from 'node:fs';
+import { existsSync } from 'node:fs';
 import { join } from 'node:path';
 // The list and DATA_PATH itself come from the platform rather than being restated here. The owner's
 // create-account handler provisions the same skeleton, and a script that drifted from it would produce
 // accounts that differ by how they were made. Importing data-path.ts pulls in no database and no server.
-import { DATA_PATH, USER_DIRS } from '../src/servers/data-path';
+// provisionUserDirs rather than a local mkdir loop, and that is the whole point of this script: it used
+// to import DATA_PATH and USER_DIRS from this very module and then reimplement the loop WITHOUT the
+// chmod, so a hand-provisioned account got 755 where a UI-provisioned one got 711/700. Nothing was ever
+// created that way here, but "two paths that claim to produce identical accounts" is exactly the drift
+// the comment above already warned about — while doing it.
+import { DATA_PATH, USER_DIRS, provisionUserDirs } from '../src/servers/data-path';
 
 const DRY_RUN = process.env.DRY_RUN === '1';
 const emails = process.argv.slice(2).filter(Boolean);
@@ -44,14 +51,16 @@ for (const email of emails) {
   const root = join(DATA_PATH, email);
   console.log(`${email}${existsSync(root) ? '' : '   [new root]'}`);
 
+  // Probed BEFORE provisioning, because afterwards everything exists and the log would say so.
+  const existing = new Set(USER_DIRS.filter((dir) => existsSync(join(root, dir))));
+
+  // Skipped entirely under DRY_RUN rather than called: provisionUserDirs has no dry-run mode, and giving
+  // it one to serve this script would put a branch in the real provisioning path for a script's benefit.
+  if (!DRY_RUN) provisionUserDirs(email);
+
   for (const dir of USER_DIRS) {
-    const path = join(root, dir);
-    if (existsSync(path)) {
-      console.log(`  · ${dir} (exists)`);
-      continue;
-    }
-    if (!DRY_RUN) mkdirSync(path, { recursive: true });
-    console.log(`  ${DRY_RUN ? '+' : '✓'} ${dir}`);
+    if (existing.has(dir)) console.log(`  · ${dir} (exists${DRY_RUN ? '' : ', mode re-asserted'})`);
+    else console.log(`  ${DRY_RUN ? '+' : '✓'} ${dir}`);
   }
   console.log('');
 }

scripts/rebuild-soulseek-tree.ts

@@ -1,52 +0,0 @@
-// Re-derive the stored tree shape for every cached share snapshot, from the folder rows already in
-// Postgres — no peer contact, so it works with everyone offline and is deterministic.
-//
-// Written for the backfill when the tree columns were added, but kept because it's the answer to any
-// future change in tree shape: it runs the SAME buildTree + finishSoulseekBrowse the sidecar's ingest
-// runs, so there's no second implementation to drift.
-//
-//   bun scripts/rebuild-soulseek-tree.ts
-
-import type { BrowsedFile } from 'officerdb';
-import { eq, asc } from 'drizzle-orm';
-import { db, finishSoulseekBrowse } from 'officerdb';
-// soulseek is a plugin, so its tables are commented out of officerdb's schema aggregator —
-// import them from the feature directly.
-import { soulseekBrowseSnapshots, soulseekBrowseDirs } from 'officerdb/soulseek/schema';
-import { buildTree } from '../src/servers/sidecar/slskd/browse';
-
-const snapshots = await db
-  .select({
-    id: soulseekBrowseSnapshots.id,
-    userId: soulseekBrowseSnapshots.userId,
-    username: soulseekBrowseSnapshots.username,
-    status: soulseekBrowseSnapshots.status,
-  })
-  .from(soulseekBrowseSnapshots)
-  .orderBy(asc(soulseekBrowseSnapshots.id));
-
-console.log(`${snapshots.length} snapshot(s) to rebuild`);
-
-for (const snap of snapshots) {
-  const started = Date.now();
-  const rows = await db
-    .select({ name: soulseekBrowseDirs.name, files: soulseekBrowseDirs.files })
-    .from(soulseekBrowseDirs)
-    .where(eq(soulseekBrowseDirs.snapshotId, snap.id));
-
-  if (!rows.length) {
-    console.log(`  #${snap.id} ${snap.username}: no folder rows, skipped`);
-    continue;
-  }
-
-  const tree = buildTree(rows.map((r) => ({ name: r.name, files: r.files as BrowsedFile[] })));
-  await finishSoulseekBrowse(snap.id, tree);
-  const roots = tree.filter((t) => t.parentPath === null);
-  console.log(
-    `  #${snap.id} ${snap.username}: ${rows.length} -> ${tree.length} rows ` +
-      `(+${tree.length - rows.length} synthesized), ${roots.length} root(s), ` +
-      `${roots.reduce((n, r) => n + r.subtreeFileCount, 0)} files, ${Date.now() - started}ms`,
-  );
-}
-
-process.exit(0);

scripts/rename-capabilities-to-permissions.ts

@@ -37,7 +37,7 @@ import postgres from 'postgres';
 // `role === 'Super Admin'` before it reads the table at all, so the account that can fix things can
 // always sign in. Non-owners degrade to core-only until the rename completes.
 
-const url = process.env.POSTGRES_URL;
+const url = process.env.OFFICERDEV_POSTGRES_URL ?? process.env.POSTGRES_URL;
 if (!url) {
   console.error('  POSTGRES_URL is not set. Run from the platform directory so Bun loads .env.');
   process.exit(1);

scripts/secret-set.ts

@@ -0,0 +1,50 @@
+#!/usr/bin/env bun
+/**
+ * Put a secret the platform did not generate into the secret store.
+ *
+ *   bun run secret:set apns-p8 < AuthKey_ABCD123456.p8
+ *   pbpaste | bun run secret:set fcm-service-account
+ *
+ * ── Why stdin, and no value argument ──
+ *
+ * A secret passed as `argv` is in the shell history, in `ps` output for the life of the process, and in
+ * any audit log that records command lines. Stdin is none of those. There is deliberately no `--value`
+ * flag: offering one would mean the convenient path is the leaky one.
+ *
+ * Multi-line input is kept verbatim, which is what a PEM needs — no escaping, unlike `.env`.
+ */
+import { putSecret, getSecret, secretStorePath } from '../src/databases/officer_db/src/secret-store';
+
+const purpose = process.argv[2];
+
+if (!purpose || purpose.startsWith('-')) {
+  console.error('usage: bun run secret:set <purpose> < file');
+  console.error('       cat AuthKey_ABCD123456.p8 | bun run secret:set apns-p8');
+  process.exit(2);
+}
+
+const value = (await Bun.stdin.text()).trim();
+
+if (!value) {
+  console.error(`[secret:set] nothing on stdin — refusing to store an empty secret for '${purpose}'`);
+  console.error('             (did you mean `bun run secret:set <purpose> < file`?)');
+  process.exit(1);
+}
+
+const previous = putSecret(purpose, value);
+
+// The value itself is never printed, here or anywhere. Enough to confirm the right thing landed:
+// the length and last four characters identify a paste without reproducing it.
+const fingerprint = `${value.length} chars, ending …${value.slice(-4)}`;
+
+if (previous === null) {
+  console.log(`[secret:set] '${purpose}' stored — ${fingerprint}`);
+} else if (previous === value) {
+  console.log(`[secret:set] '${purpose}' unchanged — ${fingerprint}`);
+} else {
+  console.log(`[secret:set] '${purpose}' REPLACED — ${fingerprint}`);
+  console.log(`             the previous value is retired, not deleted (retiredKeys('${purpose}') has it)`);
+}
+
+console.log(`[secret:set] store: ${secretStorePath()}`);
+console.log(`[secret:set] read back: ${getSecret(purpose) === value ? 'ok' : 'MISMATCH — do not restart anything'}`);

scripts/setup/machine-setup/lib/packages.sh

@@ -103,23 +103,23 @@ pkgs_core() {
       echo curl ca-certificates gnupg git jq unzip \
         apt-transport-https lsb-release software-properties-common \
         wget zip brotli build-essential python3 btop htop tree tmux ripgrep fd-find net-tools eza \
-        poppler-utils fail2ban unattended-upgrades
+        poppler-utils radicale fail2ban unattended-upgrades
       ;;
     pacman)
       echo curl ca-certificates gnupg git jq unzip \
         wget zip brotli base-devel python btop htop tree tmux ripgrep fd net-tools eza \
-        poppler fail2ban
+        poppler radicale fail2ban
       ;;
     dnf)
       echo curl ca-certificates gnupg2 git jq unzip \
         wget zip brotli python3 btop htop tree tmux ripgrep fd-find net-tools eza \
-        poppler-utils fail2ban
+        poppler-utils radicale fail2ban
       ;;
     brew)
       # curl, unzip and the TLS roots ship with macOS; the compilers come from
       # the Xcode command line tools, which is not a formula — see xcode_clt_*.
       # brotli is here because macOS ships the library but not the CLI.
-      echo gnupg git jq wget brotli btop htop tree ripgrep fd eza poppler
+      echo gnupg git jq wget brotli btop htop tree ripgrep fd eza poppler radicale
       ;;
   esac
 }

scripts/setup/officer-setup.sh

@@ -293,6 +293,17 @@ if ! skip; then
   create_layout
   ok "layout in place under ${OFFICER_ROOT}"
   SUMMARY+=("Layout: ${OFFICER_ROOT} (data, dockers, capabilities)")
+
+  # The affordances an agent needs to know about this platform. Not fatal: an
+  # install without them works, agents just go back to guessing what the client
+  # on the other end will render — which is what they did before this existed.
+  if capabilities_sync; then
+    ok "capabilities synced from ${CAPABILITIES_REPO}"
+    SUMMARY+=("Capabilities: synced")
+  else
+    warn "capabilities not synced — agents will have no affordances. Run 'git pull' in $(layout_items_dir), or press Update all on the marketplace screen."
+    SUMMARY+=("Capabilities: NOT synced")
+  fi
   step_ok
 fi
 

scripts/setup/officer-setup/lib/env.sh

@@ -60,13 +60,22 @@ write_env() {
 #
 # Everything Officer reads at runtime. Kept at 0600 and owned by ${USERNAME}: it
 # holds the token-signing secret and the database credential.
+#
+# EVERY NAME IS PREFIXED \`OFFICERDEV_\`. Bun auto-loads this file into every
+# process started in the platform directory, so a bare PORT or POSTGRES_URL
+# reaches the sidecars, the tools they spawn and anything an agent runs from
+# here — names half the software in the world already uses. Prefixed, they
+# cannot be picked up by accident.
+#
+# Every consumer reads \`OFFICERDEV_X ?? X\`, so an older unprefixed file still
+# works. \`bun run fix:env\` converts one in place.
 
-PORT="${ENV_PORT}"
+OFFICERDEV_PORT="${ENV_PORT}"
 
 # Where Officer is reached from a browser. Not derivable — see the section.
-PUBLIC_URL="${ENV_PUBLIC_URL}"
+OFFICERDEV_PUBLIC_URL="${ENV_PUBLIC_URL}"
 
-POSTGRES_URL="${POSTGRES_URL}"
+OFFICERDEV_POSTGRES_URL="${POSTGRES_URL}"
 
 ENVF
 

scripts/setup/officer-setup/lib/layout.sh

@@ -61,3 +61,47 @@ layout_wrong_owner() {
     [[ "$(stat -c %U "$dir")" == "$USERNAME" ]] || echo "$dir"
   done < <(layout_dirs)
 }
+
+# -----------------------------------------------------------------------------
+# The capabilities Officer ships
+# -----------------------------------------------------------------------------
+#
+# `capabilities/` is the file-based item store. The platform creates the six item
+# directories at boot and writes `extensions/tool-loader` from its own source, so
+# those are generated and not tracked anywhere. What IS tracked is the small set
+# the platform ships — today one affordance, telling an agent that markdown
+# images render and that the user's DEVICE fetches the URL.
+#
+# Without this a fresh install has no affordances at all, and every agent on it
+# goes back to guessing what the client on the other end will do.
+#
+# Cloned rather than copied out of the platform repository, so it can be updated
+# on its own — by `git pull` here, or by "Update all" on the marketplace screen.
+CAPABILITIES_REPO="https://gitea.officer.dev/officerdev/capabilities.git"
+
+# Three states, because a directory that already exists is the normal one: the
+# layout step created it a moment ago, and a re-run finds a checkout.
+capabilities_sync() {
+  local dir; dir="$(layout_items_dir)"
+
+  if [[ -d "$dir/.git" ]]; then
+    # `HEAD` rather than a branch name — it resolves to the remote's default, so standardising the
+    # repositories on one name does not break this the way naming `main` did.
+    (cd "$dir" && git pull --ff-only --quiet origin HEAD 2>/dev/null) || return 1
+    return 0
+  fi
+
+  # An empty directory can be cloned into directly. A non-empty one cannot, and
+  # it is not empty precisely because the platform has already run and written
+  # its generated directories — so fetch into place instead of refusing, and let
+  # the owner's own untracked items sit alongside untouched.
+  if [[ -z "$(ls -A "$dir" 2>/dev/null)" ]]; then
+    git clone --quiet "$CAPABILITIES_REPO" "$dir" || return 1
+  else
+    (cd "$dir" \
+      && git init --quiet \
+      && git remote add origin "$CAPABILITIES_REPO" 2>/dev/null \
+      && git fetch --quiet origin HEAD \
+      && git checkout --quiet -B master FETCH_HEAD) || return 1
+  fi
+}

scripts/setup/officer-setup/lib/services.sh

@@ -44,6 +44,22 @@ CORE_PROCESSES=(
   "officer-claude-code|bun|run src/servers/sidecar/claude/user-instance.ts"
   "officer-opencode|bun|run src/servers/sidecar/opencode/index.ts"
   "officer-pty|node|src/servers/sidecar/pty/index.mjs"
+  # Calendar and contacts, CORE since 2026-08-19 rather than an app-store entry.
+  #
+  # It was a Docker service because the app store made everything one, not because Radicale needs a
+  # container — it is a ~2MB Python package, and this sidecar spawns it directly on loopback. A personal
+  # server having a calendar is not an integration with something you already run; it is the baseline.
+  # Radicale needs `python3` and the `radicale` package, both installed by machine-setup.
+  "officer-caldav|bun|run src/servers/sidecar/caldav/index.ts"
+  # Push notifications, CORE since 2026-08-19. Built 2026-07-31, then switched off on 2026-08-13 with
+  # memos and photos when it was slated to become a plugin — the only one of the three never extracted,
+  # so it sat mounted nowhere with its schema commented out.
+  #
+  # Core rather than a plugin because the producers are the platform's own: the job queue, an agent turn
+  # ending, mail arriving. A notifier the owner can uninstall is one those producers cannot rely on.
+  # It boots with no credentials at all — every channel reports `isConfigured() === false` and dispatch
+  # resolves empty — so an install that never sets an APNs key runs it as a no-op rather than crash-looping.
+  "officer-notify|bun|run src/servers/sidecar/notify/index.ts"
 )
 
 write_ecosystem() {

scripts/theme-baseline.json

@@ -1,5 +1,5 @@
 {
-  "src/apps/officer-web/Screens": 144,
+  "src/apps/officer-web/Screens": 122,
   "src/workspaces/components": 47,
   "src/workspaces/officerdev/src/apps/AppStore": 6,
   "src/workspaces/officerdev/src/apps/Chat": 12,
@@ -9,15 +9,8 @@
   "src/workspaces/officerdev/src/apps/Dav": 7,
   "src/workspaces/officerdev/src/apps/FileBrowser": 15,
   "src/workspaces/officerdev/src/apps/FileViewer": 38,
-  "src/workspaces/officerdev/src/apps/Invoices": 95,
-  "src/workspaces/officerdev/src/apps/Jellyfin": 48,
-  "src/workspaces/officerdev/src/apps/Photos": 47,
   "src/workspaces/officerdev/src/apps/Preview": 4,
   "src/workspaces/officerdev/src/apps/QrTransfer": 5,
-  "src/workspaces/officerdev/src/apps/Soulseek": 313,
   "src/workspaces/officerdev/src/apps/SystemMonitor": 1,
-  "src/workspaces/officerdev/src/apps/Terminal": 10,
-  "src/workspaces/officerdev/src/apps/Transmission": 43,
-  "src/workspaces/officerdev/src/apps/Wallet": 68,
-  "src/workspaces/officerdev/src/components": 22
+  "src/workspaces/officerdev/src/apps/Terminal": 10
 }

scripts/themes/contrast.ts

@@ -0,0 +1,96 @@
+// WCAG contrast, so "is this readable" stops being a matter of taste.
+//
+// Every complaint about the current themes reduces to one of two failures: a highlight so strong it
+// masks the content on it, or text so close to its background it disappears. Both are a contrast ratio,
+// and a ratio is checkable. This file is what turns the whole exercise from opinion into a test.
+
+export type Rgb = { r: number; g: number; b: number };
+
+export function parseHex(hex: string): Rgb {
+  const h = hex.trim().replace('#', '');
+  const full = h.length === 3 ? h.split('').map((c) => c + c).join('') : h;
+  return {
+    r: Number.parseInt(full.slice(0, 2), 16),
+    g: Number.parseInt(full.slice(2, 4), 16),
+    b: Number.parseInt(full.slice(4, 6), 16),
+  };
+}
+
+export const toHex = ({ r, g, b }: Rgb): string =>
+  `#${[r, g, b].map((v) => Math.max(0, Math.min(255, Math.round(v))).toString(16).padStart(2, '0')).join('')}`;
+
+/** Relative luminance, per WCAG 2.1. */
+export function luminance(c: Rgb): number {
+  const channel = (v: number) => {
+    const s = v / 255;
+    return s <= 0.03928 ? s / 12.92 : ((s + 0.055) / 1.055) ** 2.4;
+  };
+  return 0.2126 * channel(c.r) + 0.7152 * channel(c.g) + 0.0722 * channel(c.b);
+}
+
+/** 1 (identical) to 21 (black on white). */
+export function contrast(a: string, b: string): number {
+  const [x, y] = [luminance(parseHex(a)), luminance(parseHex(b))];
+  const [hi, lo] = x > y ? [x, y] : [y, x];
+  return (hi + 0.05) / (lo + 0.05);
+}
+
+/**
+ * The thresholds, and why these ones.
+ *
+ * `TEXT` is WCAG AA for body text. `UI` is AA for large text and non-text contrast — borders, icons, the
+ * edge of a control. A border at 4.5 would be a hard line around everything, which is a different design
+ * rather than a more accessible one.
+ */
+export const TEXT = 4.5;
+export const UI = 3;
+
+/**
+ * Push a foreground away from its background until it is readable.
+ *
+ * Moves toward black or white — whichever the background is further from — in small steps, so a colour
+ * that is nearly readable keeps almost all of its hue and only a hopeless one ends up monochrome. That
+ * matters: the point is a theme that still looks like Nord, not one that passes and looks like nothing.
+ *
+ * Returns the original when it already passes, so a well-chosen palette is untouched.
+ */
+export function ensureContrast(fg: string, bg: string, min = TEXT): string {
+  if (contrast(fg, bg) >= min) return fg;
+
+  // Whichever extreme is FURTHER from the background, not whichever side of mid-grey it falls on.
+  // Those differ for a mid-tone fill and the difference is the whole answer: rose-pine's rose is at
+  // luminance 0.25, so a `> 0.5` test sends this toward white and caps at 3.5:1, while black reaches
+  // 6:1 easily. Six of its tones were unfixable until this stopped guessing and measured.
+  const black = '#000000';
+  const white = '#ffffff';
+  const target = parseHex(contrast(black, bg) >= contrast(white, bg) ? black : white);
+  const start = parseHex(fg);
+
+  // 40 steps of 2.5% each. Fine enough that the first passing value is close to the original.
+  for (let i = 1; i <= 40; i++) {
+    const t = i / 40;
+    const moved = toHex({
+      r: start.r + (target.r - start.r) * t,
+      g: start.g + (target.g - start.g) * t,
+      b: start.b + (target.b - start.b) * t,
+    });
+    if (contrast(moved, bg) >= min) return moved;
+  }
+  return toHex(target);
+}
+
+/** Blend two colours. `amount` is how much of `b`. */
+export function mix(a: string, b: string, amount: number): string {
+  const [x, y] = [parseHex(a), parseHex(b)];
+  return toHex({
+    r: x.r + (y.r - x.r) * amount,
+    g: x.g + (y.g - x.g) * amount,
+    b: x.b + (y.b - x.b) * amount,
+  });
+}
+
+/** `#rrggbb` → `r g b` for `rgb(var(--x) / 0.5)`, which is how the surfaces get their opacity. */
+export const rgbTriplet = (hex: string): string => {
+  const { r, g, b } = parseHex(hex);
+  return `${r} ${g} ${b}`;
+};

scripts/themes/map.ts

@@ -0,0 +1,193 @@
+import type { Palette } from './palette';
+import { UI, contrast, ensureContrast, mix, rgbTriplet } from './contrast';
+
+// Omarchy palette → Officer's 84 declarations.
+//
+// This file is the whole port. It exists once so that twenty-two themes cannot disagree about what
+// "muted" means, and so that improving the mapping improves all of them at once.
+//
+// ── What went wrong before, and what this fixes ──
+//
+// The previous port produced 34 declarations per theme against a contract of 84. The missing fifty were
+// every surface (`--panel-bg`, `--card-bg`, `--glass-bg`), every state colour (`--success`, `--warning`,
+// `--busy`, `--info`), the header, the scrollbars and the grid — so they fell back to whatever the base
+// stylesheet or the previous theme left behind. That is why highlights masked content and text faded
+// out: not taste, absence. It also shipped thirteen tokens nothing reads (`--chart-1..5`, `--sidebar-*`),
+// which is the tell that it came from a shadcn template rather than from this codebase's contract.
+//
+// ── The rule that makes them readable ──
+//
+// Every foreground is run through `ensureContrast` against the surface it actually sits on. A palette
+// whose colours are already readable is untouched; one that is not gets nudged toward black or white
+// until it passes, keeping as much hue as it can. So a theme cannot ship with invisible text, no matter
+// what its author chose.
+
+export type Tokens = Record<string, string>;
+
+/** Depth ordering differs by mode: a light theme's "darker" is its raised surface, not its page. */
+function surfaces(p: Palette) {
+  const dark = p.mode === 'dark';
+  return {
+    /** The page. The deepest surface in dark, the plainest in light. */
+    page: dark ? p.darkerBackground : p.background,
+    /** Panels and cards sit one step up from the page, so a tile reads as a tile. */
+    raised: dark ? p.background : p.lighterBackground,
+    /** Hovers, secondary fills, the header. One step up again. */
+    high: dark ? p.lighterBackground : p.darkBackground,
+    dark,
+  };
+}
+
+export function mapPalette(p: Palette): Tokens {
+  const s = surfaces(p);
+
+  // Text, checked against the surface each weight actually appears on.
+  const fg = ensureContrast(p.foreground, s.page);
+  const fgRaised = ensureContrast(p.foreground, s.raised);
+  // Checked against `s.high`, not `s.raised`: `--muted-foreground` is the text ON `--muted`, and
+  // `--muted` is the high surface. Deriving it against the wrong one left everforest at 4.44:1 and
+  // miasma at 4.00 — close enough to look fine and still fail, which is the kind of near-miss that a
+  // person eyeballing a screenshot would have shipped.
+  const fgMuted = ensureContrast(p.lightForeground, s.high);
+  const fgFaint = ensureContrast(p.darkForeground, s.raised, UI);
+
+  /**
+   * Text for a solid fill of `bg`.
+   *
+   * Tries the palette's own extremes first, so a button's label still belongs to the theme, and only
+   * falls through to `ensureContrast` when neither is readable. Both steps are needed: picking the
+   * better of two candidates does NOT guarantee either passes — rose-pine's fills are all mid-tone and
+   * every one came out between 2.2:1 and 4.2:1 until this drove it the rest of the way.
+   */
+  const on = (bg: string): string => {
+    const light = p.brightForeground;
+    const dark = p.darkerBackground;
+    const better = contrast(light, bg) >= contrast(dark, bg) ? light : dark;
+    return ensureContrast(better, bg);
+  };
+
+  const accent = p.accent;
+  // Borders are DECORATION, not information, so they are not held to a text ratio — a 3:1 line drawn
+  // around every panel is a different design, not a more accessible one. What they must be is
+  // discernible, which `--input` below is checked for: a form field whose boundary you cannot see is a
+  // real accessibility failure, unlike a subtle panel edge.
+  const border = mix(s.raised, s.dark ? p.lightForeground : p.darkForeground, 0.28);
+
+  return {
+    // ── Ground ──
+    '--background': s.page,
+    '--foreground': fg,
+
+    // ── Surfaces ──
+    '--card': s.raised,
+    '--card-foreground': fgRaised,
+    '--popover': s.raised,
+    '--popover-foreground': fgRaised,
+    // `muted` is a SURFACE here, not the palette's `muted` colour — that one is a mid-grey meant for
+    // dimmed text, and using it as a fill is precisely how a highlight ends up masking its content.
+    '--muted': s.high,
+    '--muted-foreground': fgMuted,
+    '--accent': s.high,
+    '--accent-foreground': fgRaised,
+
+    // ── Actions ──
+    '--primary': accent,
+    '--primary-foreground': on(accent),
+    '--secondary': s.high,
+    '--secondary-foreground': fgRaised,
+    '--cta': accent,
+    '--cta-foreground': on(accent),
+    '--ctahover': mix(accent, s.dark ? '#ffffff' : '#000000', 0.12),
+
+    // ── State — five tones, each readable on its own fill ──
+    '--success': p.green,
+    '--success-foreground': on(p.green),
+    '--warning': p.yellow,
+    '--warning-foreground': on(p.yellow),
+    '--destructive': p.red,
+    '--destructive-foreground': on(p.red),
+    '--info': p.blue,
+    '--info-foreground': on(p.blue),
+    '--busy': p.orange,
+    '--busy-foreground': on(p.orange),
+    // The odd one, per the contract: a fill bright enough for black text is too light to BE text. So ink
+    // is the same hue pushed until it reads on the page.
+    '--busy-ink': ensureContrast(p.orange, s.page),
+
+    // ── Lines ──
+    '--border': border,
+    '--input': border,
+    '--ring': accent,
+    '--radius': '0.625rem',
+
+    // ── Data ──
+    '--chart-primary': accent,
+    '--chart-secondary': p.magenta,
+    '--status-success': p.green,
+    '--status-warning': p.yellow,
+    '--status-in-progress': p.blue,
+
+    // ── Surfaces, as complete CSS values ──
+    '--page-bg': s.page,
+    '--page-bg-src': "''",
+    '--panel-bg': `rgb(${rgbTriplet(s.raised)} / 0.72)`,
+    '--panel-border': `rgb(${rgbTriplet(border)} / 0.9)`,
+    '--panel-border-bare': `rgb(${rgbTriplet(border)} / 0.55)`,
+    // DENSER, not darker — panel chrome is dark text and a heavy fill hides every icon. The contract
+    // says so; this is the one place a value is chosen to obey a sentence rather than a ratio.
+    '--panel-raised-bg': `rgb(${rgbTriplet(s.raised)} / 0.95)`,
+    '--panel-raised-border': `rgb(${rgbTriplet(border)} / 1)`,
+    '--panel-accent': accent,
+    '--panel-header-bg': `rgb(${rgbTriplet(s.high)} / 0.6)`,
+    '--panel-header-fg': fgRaised,
+    '--panel-radius': '0.75rem',
+    '--panel-blur': '12px',
+    '--card-bg': s.raised,
+    '--glass-bg': `rgb(${rgbTriplet(s.raised)} / 0.55)`,
+    '--glass-border': `rgb(${rgbTriplet(border)} / 0.7)`,
+    '--dock-bg': `rgb(${rgbTriplet(s.raised)} / 0.7)`,
+    '--dock-border': `rgb(${rgbTriplet(border)} / 0.8)`,
+    '--dock-tooltip-bg': s.high,
+    '--header-bg': `rgb(${rgbTriplet(s.raised)} / 0.8)`,
+    '--header-border': `rgb(${rgbTriplet(border)} / 0.9)`,
+    '--header-fg': fgRaised,
+    '--header-blur': '12px',
+
+    // Grid lines are a texture, not a colour anyone should notice — kept just above the surface.
+    '--pixel-grid-color': `rgb(${rgbTriplet(fgFaint)} / 0.07)`,
+    '--card-grid-color': `rgb(${rgbTriplet(fgFaint)} / 0.05)`,
+    '--panel-grid-color': `rgb(${rgbTriplet(fgFaint)} / 0.06)`,
+
+    '--scrollbar-thumb': `rgb(${rgbTriplet(p.muted)} / 0.6)`,
+    '--scrollbar-thumb-hover': `rgb(${rgbTriplet(p.muted)} / 0.85)`,
+    '--scrollbar-thin-thumb': `rgb(${rgbTriplet(p.muted)} / 0.45)`,
+    '--scrollbar-thin-thumb-hover': `rgb(${rgbTriplet(p.muted)} / 0.7)`,
+
+    // ── Brand slots ──
+    //
+    // A ported theme has no brand of its own, so these follow its accent rather than importing Duck's
+    // green into somebody else's palette. `--ink` is the app's text colour and is used ~950 times.
+    '--brand': accent,
+    '--brand-foreground': on(accent),
+    '--brand-primary': accent,
+    '--brand-accent': p.cyan,
+    '--brand-warm': p.orange,
+    '--brand-deep': s.dark ? p.darkerBackground : p.darkBackground,
+    '--brand-muted': p.muted,
+    '--ink': fg,
+    '--surface-warm': s.raised,
+    '--naturegreen': p.green,
+    '--naturegreen-foreground': on(p.green),
+
+    // Legacy aliases the contract keeps until the utility rename lands — `prose.css` and some plugins
+    // read them directly, so a theme without them has unstyled prose.
+    '--duck-dark': fg,
+    '--duck-teal': p.cyan,
+    '--duck-yellow': p.yellow,
+    '--duck-orange': p.orange,
+    '--duck-forest': p.green,
+    '--duck-beige': mix(s.raised, fg, 0.1),
+
+    '--logo-src': "'/officer-logo.svg'",
+  };
+}

scripts/themes/palette.ts

@@ -0,0 +1,86 @@
+import { readFileSync } from 'node:fs';
+
+// Reading an Omarchy theme's `colors.toml`.
+//
+// Omarchy 4 ships a genuinely semantic palette rather than sixteen ANSI slots: three background depths,
+// four foreground weights, an accent, a selection and a muted, plus the named hues. That is why this port
+// can be a mapping at all — the previous one had to guess, and the guessing is what produced themes with
+// no surfaces and unreadable highlights.
+//
+// Verified across all 22 themes in basecamp/omarchy@master: every one declares the 24 keys below.
+
+export type Palette = {
+  mode: 'dark' | 'light';
+  accent: string;
+  selection: string;
+  muted: string;
+  background: string;
+  darkBackground: string;
+  darkerBackground: string;
+  lighterBackground: string;
+  foreground: string;
+  darkForeground: string;
+  lightForeground: string;
+  brightForeground: string;
+  red: string;
+  yellow: string;
+  green: string;
+  cyan: string;
+  blue: string;
+  magenta: string;
+  /** Not in every theme — 19 of 22. Falls back to yellow-red mid, which is what orange is. */
+  orange: string;
+  brightRed: string;
+  brightYellow: string;
+  brightGreen: string;
+  brightCyan: string;
+  brightBlue: string;
+  brightMagenta: string;
+};
+
+/** A deliberately small TOML reader: this file is flat `key = "value"` and nothing else. */
+function readToml(path: string): Record<string, string> {
+  const out: Record<string, string> = {};
+  for (const line of readFileSync(path, 'utf8').split('\n')) {
+    const m = line.match(/^\s*([a-z_0-9]+)\s*=\s*"([^"]*)"/i);
+    if (m) out[m[1]!] = m[2]!;
+  }
+  return out;
+}
+
+export function loadPalette(tomlPath: string): Palette {
+  const t = readToml(tomlPath);
+  const need = (key: string): string => {
+    const v = t[key];
+    if (!v) throw new Error(`${tomlPath}: missing "${key}"`);
+    return v;
+  };
+
+  return {
+    mode: t.mode === 'light' ? 'light' : 'dark',
+    accent: need('accent'),
+    selection: need('selection'),
+    muted: need('muted'),
+    background: need('background'),
+    darkBackground: need('dark_background'),
+    darkerBackground: need('darker_background'),
+    lighterBackground: need('lighter_background'),
+    foreground: need('foreground'),
+    darkForeground: need('dark_foreground'),
+    lightForeground: need('light_foreground'),
+    brightForeground: need('bright_foreground'),
+    red: need('red'),
+    yellow: need('yellow'),
+    green: need('green'),
+    cyan: need('cyan'),
+    blue: need('blue'),
+    magenta: need('magenta'),
+    orange: t.orange ?? need('yellow'),
+    brightRed: need('bright_red'),
+    brightYellow: need('bright_yellow'),
+    brightGreen: need('bright_green'),
+    brightCyan: need('bright_cyan'),
+    brightBlue: need('bright_blue'),
+    brightMagenta: need('bright_magenta'),
+  };
+}

scripts/todo-index.ts

@@ -0,0 +1,140 @@
+/**
+ * Regenerate `TODO.md` — the index over `todo/`.
+ *
+ * The index is the one file everybody reads, so it is the one most able to lie. It is generated rather
+ * than maintained: `bun run todo:index` after any change to `todo/`, and it cannot drift from the tree
+ * it describes.
+ */
+import { readFileSync, readdirSync, writeFileSync, statSync } from 'node:fs';
+import { join } from 'node:path';
+
+const ROOT = new URL('..', import.meta.url).pathname;
+const OUT = join(ROOT, 'todo');
+
+export type Item = {
+  id: string;
+  title: string;
+  theme: string;
+  status: string;
+  created: string;
+  /** `urgent` or empty — see todo/README.md for why there is no scale. */
+  priority: string;
+  due: string;
+  order: number;
+  file: string;
+};
+
+/** Minimal frontmatter reader — `key: value`, values optionally JSON-quoted. No YAML dependency. */
+export function frontmatter(text: string): Record<string, string> {
+  const end = text.indexOf('\n---', 4);
+  if (!text.startsWith('---') || end === -1) return {};
+  const out: Record<string, string> = {};
+  for (const line of text.slice(4, end).split('\n')) {
+    const m = line.match(/^([a-z_]+):\s*(.*)$/);
+    if (!m || !m[1]) continue;
+    const raw = (m[2] ?? '').trim();
+    out[m[1]] = raw.startsWith('"') ? (JSON.parse(raw) as string) : raw;
+  }
+  return out;
+}
+
+export function readTree(): { themes: { theme: string; heading: string; items: Item[] }[] } {
+  const themes = readdirSync(OUT)
+    .filter((d) => statSync(join(OUT, d)).isDirectory())
+    .sort()
+    .map((theme) => {
+      const files = readdirSync(join(OUT, theme)).filter((f) => f.endsWith('.md'));
+      const meta = frontmatter(readFileSync(join(OUT, theme, '_index.md'), 'utf8'));
+      const items = files
+        .filter((f) => f !== '_index.md')
+        .map((f) => {
+          const fm = frontmatter(readFileSync(join(OUT, theme, f), 'utf8'));
+          return {
+            id: fm.id ?? `${theme}/${f.replace(/\.md$/, '')}`,
+            title: fm.title ?? f,
+            theme,
+            status: fm.status ?? 'open',
+            created: fm.created ?? '',
+            priority: fm.priority ?? '',
+            due: fm.due ?? '',
+            order: Number(fm.order ?? 0),
+            file: `todo/${theme}/${f}`,
+          };
+        })
+        .sort((a, b) => a.order - b.order);
+      return { theme, heading: meta.heading ?? theme, items };
+    });
+  return { themes };
+}
+
+const HEADER = `# TODO — the index
+
+Deferred work. **The items live in \`todo/\`, one file each**, grouped by theme; this file is the map
+over them and is GENERATED — run \`bun run todo:index\` after any change rather than editing it here.
+Format and the two-way-sync contract: \`todo/README.md\`.
+
+**Context, corrected 2026-08-07.** This list used to open by saying Officer was "collapsing from
+multi-tenant / open-source-ready to a **single-user platform**", and told you to treat multi-tenant
+indirection as accidental complexity. **That direction was reversed.** The permission model shipped on
+2026-08-07 to serve a real goal — deploy to the company server, onboard people, give each one their own
+Gitea account through the platform. Per-user scoping is now a requirement, and the items that proposed
+deleting it have been removed rather than left to mislead the next reader.
+
+What did NOT reverse: \`execution\` permissions (terminal, chat, tasks, files, desktop, browser) run as
+the owner's OS user and can never be granted. Indirection there really is accidental complexity.
+`;
+
+function render(): string {
+  const { themes } = readTree();
+  const open = themes.reduce((n, t) => n + t.items.filter((i) => i.status !== 'done').length, 0);
+  const done = themes.reduce((n, t) => n + t.items.filter((i) => i.status === 'done').length, 0);
+
+  const rows = themes
+    .filter((t) => t.items.length)
+    .map((t) => {
+      const o = t.items.filter((i) => i.status !== 'done').length;
+      const d = t.items.length - o;
+      return `| [${t.heading}](todo/${t.theme}/) | ${o} | ${d} |`;
+    });
+
+  const recent = themes
+    .flatMap((t) => t.items)
+    .filter((i) => i.status !== 'done')
+    .sort((a, b) => b.created.localeCompare(a.created))
+    .slice(0, 10)
+    .map((i) => `- \`${i.created}\` **${i.title}** — [${i.id}](${i.file})`);
+
+  // A theme holding only prose has no row above, and an entry nobody can see is the failure mode this
+  // whole split risks. List them explicitly rather than letting them age out of view.
+  const notes = themes.filter((t) => !t.items.length).map((t) => `- [${t.heading}](todo/${t.theme}/_index.md)`);
+
+  // Urgent goes above everything. An item the owner called urgent, sorted in among seventy-five others
+  // by theme, is an item nobody sees on the day it matters.
+  const urgent = themes
+    .flatMap((t) => t.items)
+    .filter((i) => i.status !== 'done' && i.priority === 'urgent')
+    .sort((a, b) => (a.due || '9999').localeCompare(b.due || '9999'))
+    .map((i) => `- ${i.due ? `**due ${i.due}** — ` : ''}**${i.title}** — [${i.id}](${i.file})`);
+
+  return [
+    HEADER,
+    ...(urgent.length ? [`## Urgent\n`, ...urgent, ``] : []),
+    `## Themes\n`,
+    `**${open} open, ${done} closed**, across ${rows.length} themes.\n`,
+    `| Theme | Open | Closed |`,
+    `| --- | --- | --- |`,
+    ...rows,
+    ``,
+    `## Records — no items, kept for the reasoning\n`,
+    ...notes,
+    ``,
+    `## Newest ten\n`,
+    ...recent,
+    ``,
+  ].join('\n');
+}
+
+if (import.meta.main) {
+  writeFileSync(join(ROOT, 'TODO.md'), render());
+  console.log('TODO.md regenerated');
+}

scripts/update-plugins.ts

@@ -0,0 +1,65 @@
+#!/usr/bin/env bun
+import { snapshotPlugins } from '../src/servers/plugins/mount';
+import { updatePlugin, syncCapabilities } from '../src/servers/plugins/update';
+
+// Update every installed plugin, from a terminal.
+//
+//   bun run plugins:update
+//
+// ── Why this exists ──
+//
+// The only way to do this was `POST /api/plugins/update-all`, which is authenticated — so a deploy had to
+// mint an API key to update the plugins on the machine it was already logged into. That is a credential
+// created, used and hopefully deleted, on every host, for a local operation.
+//
+// This is the same work through the same functions, with the HTTP layer removed rather than a second
+// implementation beside it. `prompts/deploy-commander.md` calls it.
+//
+// ── It runs OUTSIDE the server process ──
+//
+// Which matters, and is the one thing to know before reading the output. `updatePlugin` pulls, re-applies
+// and rebuilds the frontend — all of which touch disk and are picked up by the running server. But
+// remounting happens in the process that calls it, and this is not that process, so the routes the
+// server is serving do not change until it restarts.
+//
+// So: the files, the schema, the assets and the SPA bundle are all current when this finishes. The route
+// table is not. A deploy restarts `officer` afterwards, which is what makes the difference invisible —
+// and the reason this says so rather than leaving somebody to find out.
+
+const started = Date.now();
+const { states } = await snapshotPlugins();
+const installed = states.filter((st) => st.install).map((st) => st.plugin.appName);
+
+if (installed.length === 0) {
+  console.log('no plugins installed — nothing to update');
+  process.exit(0);
+}
+
+console.log(`updating ${installed.length} plugin(s): ${installed.join(', ')}\n`);
+
+let failed = 0;
+
+// SEQUENTIAL, like the route: each update pushes a schema, republishes assets, rewrites the ecosystem
+// file and rebuilds the frontend. Two at once is a race over one file each.
+for (const appName of installed) {
+  process.stdout.write(`→ ${appName}\n`);
+  // One plugin failing must not stop the rest. A dirty checkout is the EXPECTED failure — every plugin
+  // directory on a developer machine is a live one — and it should not hold back the other eleven.
+  const result = await updatePlugin(appName, async (step) => console.log(`    ${step}`));
+  if (!result.ok) {
+    failed++;
+    console.error(`    FAILED: ${result.error}`);
+  }
+}
+
+// The capabilities repository rides along, the same as it does through the button: it is one question —
+// "get me what has been published since I installed" — and a second gesture for it is one nobody remembers.
+const capabilities = await syncCapabilities();
+console.log(`\ncapabilities: ${capabilities.action}${capabilities.error ? ` — ${capabilities.error}` : ''}`);
+
+const ok = installed.length - failed;
+console.log(`${ok}/${installed.length} updated in ${Math.round((Date.now() - started) / 1000)}s`);
+if (failed) console.log("a failure above is usually an uncommitted change in that plugin's checkout");
+
+// Non-zero on failure, so a deploy script can stop rather than reporting a success it did not have.
+process.exit(failed > 0 ? 1 : 0);

src/apps/officer-web/App.tsx

@@ -50,7 +50,7 @@ export function App() {
         <Dashboard.DashboardLayout>
           <Routes>
             <Route path="/" element={<Dashboard.HomeScreen />} />
-            {/* Route pairs, like /offscale (a plugin now) and /photos below: the bare page route exists so the
+            {/* Route pairs, like /command-hq and /soulseek below: the bare page route exists so the
                 existing links into it keep working, and its `SettingsRoute` guard redirects to the
                 first section. The section is the URL rather than a `*_SELECTED` global, so a settings
                 sub-page is linkable, bookmarkable and cmd-clickable. */}
@@ -80,10 +80,8 @@ export function App() {
             <Route path="/files/*" element={<Dashboard.FilesScreen />} />
             <Route path="/calendar" element={<Dashboard.CalendarScreen />} />
             <Route path="/contacts" element={<Dashboard.ContactsScreen />} />
-            <Route path="/soulseek" element={<Dashboard.SoulseekScreen />} />
-            <Route path="/soulseek/:section" element={<Dashboard.SoulseekScreen />} />
-            <Route path="/photos" element={<Dashboard.PhotosScreen />} />
-            <Route path="/photos/:section" element={<Dashboard.PhotosScreen />} />
+            <Route path="/command-hq" element={<Dashboard.CommandHqScreen />} />
+            <Route path="/command-hq/:sessionId" element={<Dashboard.CommandHqScreen />} />
             <Route path="/app-store" element={<Dashboard.AppStoreScreen />} />
             <Route path="/marketplace" element={<Dashboard.PluginsScreen />} />
             {/* The screen was `/plugins` until marketplace.officer.dev existed. Redirected rather than
@@ -95,7 +93,7 @@ export function App() {
                 hands the whole subtree to the plugin's own router, which react-router nests natively. */}
             {installedPlugins.flatMap((plugin) => [
               <Route key={plugin.appName} path={plugin.route} element={<Dashboard.PluginScreen {...plugin} />} />,
-              // The section pair, exactly as the core screens do it (`/photos/:section`): the plugin's
+              // The section pair, exactly as the core screens do it (`/soulseek/:section`): the plugin's
               // panels read `useParams` themselves, so which section is open is the URL rather than state
               // passed between them.
               <Route
@@ -104,19 +102,6 @@ export function App() {
                 element={<Dashboard.PluginScreen {...plugin} />}
               />,
             ])}
-            <Route path="/jellyfin" element={<Dashboard.JellyfinScreen />} />
-            <Route path="/jellyfin/:section" element={<Dashboard.JellyfinScreen />} />
-            <Route path="/transmission" element={<Dashboard.TransmissionScreen />} />
-            <Route path="/transmission/:section" element={<Dashboard.TransmissionScreen />} />
-            <Route path="/gitea" element={<Dashboard.GiteaScreen />} />
-            <Route path="/gitea/:section" element={<Dashboard.GiteaScreen />} />
-            <Route path="/gitea/repo/:owner/:name" element={<Dashboard.GiteaScreen />} />
-            <Route path="/gitea/repo/:owner/:name/:tab" element={<Dashboard.GiteaScreen />} />
-            <Route path="/gitea/repo/:owner/:name/:tab/:item" element={<Dashboard.GiteaScreen />} />
-            <Route path="/invoices" element={<Dashboard.InvoicesScreen />} />
-            <Route path="/invoices/:section" element={<Dashboard.InvoicesScreen />} />
-            <Route path="/wallet" element={<Dashboard.WalletScreen />} />
-            <Route path="/wallet/:section" element={<Dashboard.WalletScreen />} />
             <Route path="/system-monitor" element={<Dashboard.SystemMonitorScreen />} />
             <Route path="/system-monitor/:scope" element={<Dashboard.SystemMonitorScreen />} />
             {/* TEMPORARY / EXPERIMENTAL — offline file transfer over animated QR codes */}
@@ -135,10 +120,7 @@ export function App() {
             <Route path="/jobs/:id" element={<Dashboard.JobsPage />} />
             <Route path="/dashboards" element={<Dashboard.DashboardsScreen />} />
             <Route path="/dashboards/:id" element={<Dashboard.DashboardScreen />} />
-            <Route path="/browser" element={<Dashboard.BrowserScreen />} />
-            <Route path="/browser/:tabId" element={<Dashboard.BrowserScreen />} />
             <Route path="/terminal" element={<Dashboard.TerminalScreen />} />
-            <Route path="/desktop" element={<Dashboard.DesktopScreen />} />
             <Route path="/auth/signout" element={<Authentication.SignoutScreen />} />
             <Route path="*" element={<Navigate to="/" replace />} />
           </Routes>

src/apps/officer-web/Screens/Authentication/LandingPage/Login.tsx

@@ -46,13 +46,17 @@ export function Login() {
 
       <form ref={formRef} onSubmit={handleSubmit} className="grid gap-4">
         <Label className="grid gap-2">
-          <span className="text-duck-dark/70">Email</span>
+          <span className="text-duck-dark/70">Email or username</span>
           <Input
             className="h-11 bg-background/60 border-duck-dark/20 text-duck-dark placeholder:text-duck-dark/40"
-            type="email"
+            // `text`, not `email`: the browser refuses to submit a username under type="email", so the
+            // field would have validated the user out of an account they have.
+            type="text"
             name="email"
             placeholder="you@example.com"
-            autoComplete="email"
+            // `username` rather than `email`, which is what password managers expect for a field that
+            // takes either — and it is the value they already store beside the password.
+            autoComplete="username"
           />
         </Label>
 

src/apps/officer-web/Screens/Dashboard/Activity/ActivityScreen.tsx

@@ -102,7 +102,7 @@ export const ActivityScreen = () => {
             title={t.cwd}
           >
             <span
-              className={`inline-block h-2 w-2 shrink-0 rounded-full ${t.active ? 'bg-emerald-500 animate-pulse' : 'bg-muted-foreground/40'}`}
+              className={`inline-block h-2 w-2 shrink-0 rounded-full ${t.active ? 'bg-success animate-pulse' : 'bg-muted-foreground/40'}`}
             />
             <span className="truncate font-mono text-xs">{t.id}</span>
           </Link>

src/apps/officer-web/Screens/Dashboard/Browser/BrowserScreen.tsx

@@ -1,48 +0,0 @@
-import { useMemo, useCallback } from 'react';
-import { useNavigate } from 'react-router';
-import type { LayoutNode, PanelComponents } from 'officerdev';
-import { WorkspaceView, ChatPanelWrapper } from 'officerdev';
-import { useDashboardState } from 'state/useDashboardState';
-import { useIsMobile } from 'hooks/useIsMobile';
-import { defaultLayout } from './defaultLayout';
-import { TabList } from './TabList';
-import { TabPreview } from './TabPreview';
-import { useSelectedTabId } from './shared';
-
-const PROMPT_PREFIX = `You are a browser assistant. The user has connected Chrome tabs via the Officer Browser Relay. You have a 'browser' tool — use it to list tabs, take screenshots, navigate, evaluate JavaScript, and get page info. When asked about a page, take a screenshot first.`;
-
-// This screen's prompt, so this screen's chat panel. See the note in `EmailScreen`.
-const BrowserChatPanel = () => <ChatPanelWrapper panelId="browser-chat" promptPrefix={PROMPT_PREFIX} />;
-
-export const BrowserScreen = () => {
-  const isMobile = useIsMobile();
-  const navigate = useNavigate();
-  // No redirect guard: like /email, the bare /browser is the list with nothing open, and a tab id that
-  // no longer resolves gets the preview's "no longer attached" state rather than a rewritten address.
-  const selectedId = useSelectedTabId();
-  const workspace = useDashboardState<LayoutNode>('screens/browser', defaultLayout);
-
-  const components: PanelComponents = useMemo(
-    () => ({
-      'browser-tabs': TabList,
-      'browser-preview': TabPreview,
-      'browser-chat': BrowserChatPanel,
-    }),
-    [],
-  );
-
-  const mobilePanelId = isMobile && selectedId ? 'browser-preview' : undefined;
-  const onMobileBack = useCallback(() => navigate('/browser'), [navigate]);
-
-  return (
-    <div className="h-full w-full pt-2">
-      <WorkspaceView
-        workspace={workspace}
-        locked
-        components={components}
-        mobilePanelId={mobilePanelId}
-        onMobilePanelChange={mobilePanelId ? () => onMobileBack() : undefined}
-      />
-    </div>
-  );
-};

src/apps/officer-web/Screens/Dashboard/Browser/TabList.tsx

@@ -1,117 +0,0 @@
-import { useQuery } from '@tanstack/react-query';
-import { Link } from 'react-router';
-import { Globe, Loader2, X, Eye } from 'lucide-react';
-import { toast } from 'sonner';
-import { useClient } from 'hooks/useClient';
-import { browserTabPath, useBrowserTargets, useSelectedTabId, type BrowserTarget } from './shared';
-
-export const TabList = () => {
-  const client = useClient();
-  const selectedId = useSelectedTabId();
-
-  const { data: status } = useQuery({
-    queryKey: ['browser-status'],
-    queryFn: () => client.get<{ extensionConnected: boolean; targetCount: number }>('/browser/status'),
-    refetchInterval: 5000,
-  });
-
-  const { data: targets, isLoading } = useBrowserTargets(!!status?.extensionConnected);
-
-  const targetList: BrowserTarget[] = targets ?? [];
-
-  const handleClose = async (targetId: string) => {
-    try {
-      await client.post(`/browser/targets/${targetId}/close`);
-    } catch {
-      toast.error('Failed to close tab');
-    }
-  };
-
-  const handleActivate = async (targetId: string) => {
-    try {
-      await client.post(`/browser/targets/${targetId}/activate`);
-    } catch {
-      toast.error('Failed to activate tab');
-    }
-  };
-
-  const isConnected = status?.extensionConnected ?? false;
-
-  return (
-    <div className="flex h-full flex-col overflow-y-auto">
-      <div className="flex items-center gap-2 border-b px-3 py-2">
-        <Globe className="h-4 w-4 opacity-60" />
-        <span className="text-sm font-medium">Browser</span>
-        <div
-          className={`ml-auto h-2 w-2 rounded-full ${isConnected ? 'bg-green-500' : 'bg-red-500'}`}
-          title={isConnected ? 'Extension connected' : 'Extension not connected'}
-        />
-      </div>
-
-      {!isConnected && (
-        <div className="flex flex-col items-center justify-center gap-3 p-4 text-sm flex-1">
-          <Globe className="h-8 w-8 opacity-30" />
-          <p className="opacity-60 text-center">Extension not connected</p>
-          {/* A <Link>, not an <a href>: a raw href to an in-app route reloads the whole SPA. */}
-          <Link to="/settings/integrations" className="text-xs text-duck-teal underline">
-            Set up in Integrations
-          </Link>
-        </div>
-      )}
-
-      {isConnected && isLoading && (
-        <div className="flex h-32 items-center justify-center">
-          <Loader2 className="h-5 w-5 animate-spin opacity-50" />
-        </div>
-      )}
-
-      {isConnected && !isLoading && targetList.length === 0 && (
-        <div className="flex h-32 flex-col items-center justify-center gap-2 text-sm opacity-50">
-          <Globe className="h-6 w-6" />
-          <span>No tabs attached</span>
-          <span className="text-xs">Click the extension icon on a tab</span>
-        </div>
-      )}
-
-      {isConnected && targetList.length > 0 && (
-        <div className="flex flex-1 flex-col divide-y divide-white/10">
-          {/* Focus and Close are siblings of the anchor, not children of it: an <a> may not contain a
-              <button>, and the old row was a <button> containing two more — invalid either way, and it
-              was only the stopPropagation calls holding it together. */}
-          {targetList.map((target) => (
-            <div
-              key={target.id}
-              className={`group relative flex transition-colors ${
-                selectedId === target.id ? 'bg-accent' : 'hover:bg-accent/50'
-              }`}
-            >
-              <Link
-                to={browserTabPath(target.id)}
-                className="flex min-w-0 flex-1 flex-col gap-0.5 px-3 py-2.5 text-left"
-              >
-                <span className="truncate pr-14 text-sm font-medium">{target.title || 'Untitled'}</span>
-                <span className="truncate text-xs opacity-50">{target.url}</span>
-              </Link>
-              <div className="absolute right-3 top-2.5 flex shrink-0 items-center gap-1 opacity-0 transition-opacity group-hover:opacity-100">
-                <button
-                  onClick={() => handleActivate(target.id)}
-                  className="rounded p-0.5 hover:bg-black/10 cursor-pointer"
-                  title="Focus tab"
-                >
-                  <Eye className="h-3.5 w-3.5" />
-                </button>
-                <button
-                  onClick={() => handleClose(target.id)}
-                  className="rounded p-0.5 hover:bg-black/10 cursor-pointer"
-                  title="Close tab"
-                >
-                  <X className="h-3.5 w-3.5" />
-                </button>
-              </div>
-            </div>
-          ))}
-        </div>
-      )}
-    </div>
-  );
-};

src/apps/officer-web/Screens/Dashboard/Browser/TabPreview.tsx

@@ -1,174 +0,0 @@
-import { useState, useCallback } from 'react';
-import { Link } from 'react-router';
-import { useQuery } from '@tanstack/react-query';
-import { Globe, RefreshCw, Send, Terminal, Loader2 } from 'lucide-react';
-import { toast } from 'sonner';
-import { Button } from '@/components/ui/button';
-import { useClient } from 'hooks/useClient';
-import { useBrowserTargets, useSelectedTabId } from './shared';
-
-export const TabPreview = () => {
-  const client = useClient();
-  // Reads the URL itself rather than being told by the list.
-  const selectedId = useSelectedTabId();
-  const { data: targets } = useBrowserTargets(!!selectedId);
-  // A target id dies with the browser session, so a link or a reload can name a tab that is no longer
-  // attached. Screenshotting it would 404 every ten seconds; better to say so once.
-  const isKnownTab = !!selectedId && !!targets?.some((t) => t.id === selectedId);
-  const [navUrl, setNavUrl] = useState('');
-  const [evalExpr, setEvalExpr] = useState('');
-  const [evalResult, setEvalResult] = useState<string | null>(null);
-  const [isNavigating, setIsNavigating] = useState(false);
-  const [isEvaluating, setIsEvaluating] = useState(false);
-
-  const {
-    data: screenshot,
-    isLoading,
-    refetch,
-    isFetching,
-  } = useQuery({
-    queryKey: ['browser-screenshot', selectedId],
-    queryFn: () => client.get<{ data: string; format: string }>(`/browser/targets/${selectedId}/screenshot`),
-    enabled: isKnownTab,
-    refetchInterval: 10000,
-  });
-
-  const handleNavigate = useCallback(async () => {
-    if (!selectedId || !navUrl.trim()) return;
-    setIsNavigating(true);
-    try {
-      await client.post(`/browser/targets/${selectedId}/navigate`, { url: navUrl.trim() });
-      setTimeout(() => void refetch(), 1000);
-    } catch {
-      toast.error('Navigation failed');
-    } finally {
-      setIsNavigating(false);
-    }
-  }, [client, selectedId, navUrl, refetch]);
-
-  const handleEvaluate = useCallback(async () => {
-    if (!selectedId || !evalExpr.trim()) return;
-    setIsEvaluating(true);
-    try {
-      const res = await client.post<{ result: unknown }>(`/browser/targets/${selectedId}/evaluate`, {
-        expression: evalExpr.trim(),
-      });
-      setEvalResult(JSON.stringify(res.result, null, 2));
-    } catch (err) {
-      setEvalResult(`Error: ${err instanceof Error ? err.message : String(err)}`);
-    } finally {
-      setIsEvaluating(false);
-    }
-  }, [client, selectedId, evalExpr]);
-
-  if (!selectedId) {
-    return (
-      <div className="flex h-full flex-col items-center justify-center gap-2 text-sm opacity-50">
-        <Globe className="h-8 w-8" />
-        <span>Select a tab to preview</span>
-      </div>
-    );
-  }
-
-  // Waiting on the list is not the same as knowing the tab is gone, so only say so once it has arrived.
-  if (targets && !isKnownTab) {
-    return (
-      <div className="flex h-full flex-col items-center justify-center gap-2 text-sm opacity-50">
-        <Globe className="h-8 w-8" />
-        <span>That tab is no longer attached</span>
-        <Link to="/browser" className="text-xs underline">
-          Back to the tab list
-        </Link>
-      </div>
-    );
-  }
-
-  return (
-    <div className="flex h-full flex-col overflow-hidden">
-      {/* URL bar */}
-      <div className="flex items-center gap-2 border-b px-3 py-2">
-        <Button
-          variant="ghost"
-          size="sm"
-          className="h-7 w-7 p-0 shrink-0"
-          onClick={() => void refetch()}
-          disabled={isFetching}
-        >
-          {isFetching ? <Loader2 className="h-3.5 w-3.5 animate-spin" /> : <RefreshCw className="h-3.5 w-3.5" />}
-        </Button>
-        <form
-          className="flex flex-1 items-center gap-2"
-          onSubmit={(ev) => {
-            ev.preventDefault();
-            void handleNavigate();
-          }}
-        >
-          <input
-            type="text"
-            value={navUrl}
-            onChange={(ev) => setNavUrl(ev.target.value)}
-            placeholder="Navigate to URL..."
-            className="flex-1 rounded-md border bg-transparent px-2 py-1 text-sm outline-none focus:border-cyan-500"
-          />
-          <Button
-            variant="ghost"
-            size="sm"
-            className="h-7 w-7 p-0 shrink-0"
-            type="submit"
-            disabled={isNavigating || !navUrl.trim()}
-          >
-            <Send className="h-3.5 w-3.5" />
-          </Button>
-        </form>
-      </div>
-
-      {/* Screenshot */}
-      <div className="flex-1 overflow-auto p-2">
-        {isLoading && !screenshot && (
-          <div className="flex h-full items-center justify-center">
-            <Loader2 className="h-6 w-6 animate-spin opacity-50" />
-          </div>
-        )}
-        {screenshot && (
-          <img
-            src={`data:image/${screenshot.format};base64,${screenshot.data}`}
-            alt="Tab screenshot"
-            className="w-full rounded-md border"
-          />
-        )}
-      </div>
-
-      {/* JS Console */}
-      <div className="border-t">
-        <form
-          className="flex items-center gap-2 px-3 py-2"
-          onSubmit={(ev) => {
-            ev.preventDefault();
-            void handleEvaluate();
-          }}
-        >
-          <Terminal className="h-3.5 w-3.5 shrink-0 opacity-50" />
-          <input
-            type="text"
-            value={evalExpr}
-            onChange={(ev) => setEvalExpr(ev.target.value)}
-            placeholder="Evaluate JavaScript..."
-            className="flex-1 bg-transparent text-sm outline-none font-mono"
-          />
-          <Button
-            variant="ghost"
-            size="sm"
-            className="h-6 px-2 shrink-0"
-            type="submit"
-            disabled={isEvaluating || !evalExpr.trim()}
-          >
-            {isEvaluating ? <Loader2 className="h-3 w-3 animate-spin" /> : 'Run'}
-          </Button>
-        </form>
-        {evalResult !== null && (
-          <pre className="max-h-32 overflow-auto border-t px-3 py-2 text-xs font-mono opacity-70">{evalResult}</pre>
-        )}
-      </div>
-    </div>
-  );
-};

src/apps/officer-web/Screens/Dashboard/Browser/defaultLayout.ts

@@ -1,22 +0,0 @@
-import type { LayoutNode } from 'officerdev';
-
-export const defaultLayout: LayoutNode = {
-  type: 'group',
-  id: 'browser-root',
-  direction: 'horizontal',
-  children: [
-    { node: { type: 'panel', id: 'browser-tabs', appType: null }, size: 25 },
-    {
-      node: {
-        type: 'group',
-        id: 'browser-right',
-        direction: 'vertical',
-        children: [
-          { node: { type: 'panel', id: 'browser-preview', appType: null }, size: 60 },
-          { node: { type: 'panel', id: 'browser-chat', appType: 'officerdev/chat' }, size: 40 },
-        ],
-      },
-      size: 75,
-    },
-  ],
-};

src/apps/officer-web/Screens/Dashboard/Browser/index.tsx

@@ -1 +0,0 @@
-export { BrowserScreen } from './BrowserScreen';

src/apps/officer-web/Screens/Dashboard/Browser/shared.ts

@@ -1,34 +0,0 @@
-import { useQuery } from '@tanstack/react-query';
-import { useParams } from 'react-router';
-import { useClient } from 'hooks/useClient';
-
-// Which attached tab is being previewed is `/browser/:tabId`. It used to be a `BROWSER_SELECTED_TAB`
-// global read by three components, which meant the address bar said `/browser` no matter what you were
-// looking at, and a tab row was a `<button>` with the id in its onClick closure.
-
-export type BrowserTarget = {
-  id: string;
-  sessionId: string;
-  type: string;
-  title: string;
-  url: string;
-};
-
-export const browserTabPath = (id: string) => `/browser/${encodeURIComponent(id)}`;
-
-/** The previewed tab, or null on the bare `/browser` route — the list with nothing open. */
-export const useSelectedTabId = (): string | null => useParams<{ tabId?: string }>().tabId ?? null;
-
-// Both panels want this list and React Query serves both observers from one key, so the preview being
-// able to tell "no tab open" from "that tab is gone" costs no extra request. It needs to: a CDP target
-// id dies with the browser session, and now that the id is in the URL a reload can carry a stale one in
-// — which the old global, living in memory, could never do.
-export const useBrowserTargets = (enabled: boolean) => {
-  const client = useClient();
-  return useQuery({
-    queryKey: ['browser-targets'],
-    queryFn: () => client.get<BrowserTarget[]>('/browser/targets'),
-    refetchInterval: 3000,
-    enabled,
-  });
-};

src/apps/officer-web/Screens/Dashboard/Calendar/CalendarScreen.tsx

@@ -3,7 +3,7 @@ import { WorkspaceView } from 'officerdev';
 import { useDashboardState } from 'state/useDashboardState';
 import { defaultLayout } from './defaultLayout';
 
-// /calendar uses the Workspace/Panel system (like /transmission): the calendar list on the left, the
+// /calendar uses the Workspace/Panel system (like /command-hq): the calendar list on the left, the
 // selected calendar's agenda on the right. Both panels read the officer-caldav sidecar through the
 // /api/caldav auth proxy, which holds no DAV credentials of its own.
 //

src/apps/officer-web/Screens/Dashboard/CapabilityPage.tsx

@@ -1,3 +1,4 @@
+import type { FormEvent } from 'react';
 import { useState, useEffect, useRef } from 'react';
 import { Link, useNavigate, useParams, useSearchParams } from 'react-router';
 import { useQuery, useQueryClient } from '@tanstack/react-query';
@@ -344,7 +345,8 @@ export const CapabilityList = ({
     queryFn: () => client.get<CapabilitySummary[]>(endpoint),
   });
 
-  const handleCreate = async () => {
+  const handleCreate = async (ev?: FormEvent) => {
+    ev?.preventDefault();
     const name = newName.trim();
     if (!name) return;
     try {
@@ -382,16 +384,15 @@ export const CapabilityList = ({
         )}
       </div>
       {creating && (
-        <div className="shrink-0 px-3 py-2 border-b border-duck-dark/10 bg-duck-teal/5 flex items-center gap-1.5">
+        <form
+          onSubmit={handleCreate}
+          className="shrink-0 px-3 py-2 border-b border-duck-dark/10 bg-duck-teal/5 flex items-center gap-1.5"
+        >
           <input
             ref={newNameRef}
             value={newName}
             onChange={(ev) => setNewName(ev.target.value)}
             onKeyDown={(ev) => {
-              if (ev.key === 'Enter') {
-                ev.preventDefault();
-                handleCreate();
-              }
               if (ev.key === 'Escape') {
                 setCreating(false);
                 setNewName('');
@@ -401,13 +402,14 @@ export const CapabilityList = ({
             className="flex-1 min-w-0 rounded border border-duck-dark/20 bg-background px-2 py-1 text-base md:text-sm text-duck-dark placeholder:text-duck-dark/30 focus:outline-none focus:ring-1 focus:ring-duck-teal/30"
           />
           <button
-            onClick={handleCreate}
+            type="submit"
             disabled={!newName.trim()}
             className="p-1 rounded hover:bg-duck-dark/10 cursor-pointer transition-colors disabled:opacity-30"
           >
             <Check className="h-3.5 w-3.5 text-duck-teal" />
           </button>
           <button
+            type="button"
             onClick={() => {
               setCreating(false);
               setNewName('');
@@ -416,7 +418,7 @@ export const CapabilityList = ({
           >
             <X className="h-3.5 w-3.5 text-duck-dark/50" />
           </button>
-        </div>
+        </form>
       )}
       {externalSearch === undefined && (
         <div className="shrink-0 px-3 py-2 border-b border-duck-dark/10">

src/apps/officer-web/Screens/Dashboard/CommandHQ/CommandHqScreen.tsx

@@ -0,0 +1,25 @@
+import type { LayoutNode } from 'officerdev';
+import { WorkspaceView } from 'officerdev';
+import { useDashboardState } from 'state/useDashboardState';
+import { defaultLayout } from './defaultLayout';
+
+// /command-hq: nav + grid + detail, all three always mounted (like /chat), selection carried by the URL
+// — `?category=` for the nav filter, `:sessionId` for the open commander. See `docs/command-hq.md` →
+// "Route, surface, placement" and "Layout, v1".
+
+export const CommandHqScreen = () => {
+  const workspace = useDashboardState<LayoutNode>('screens/command-hq', defaultLayout);
+
+  return (
+    <div className="h-full w-full pt-2">
+      <WorkspaceView
+        workspace={workspace}
+        locked
+        appTypes={{
+          allowed: ['command-hq-categories', 'command-hq-grid', 'command-hq-detail'],
+          fallback: 'command-hq-detail',
+        }}
+      />
+    </div>
+  );
+};

src/apps/officer-web/Screens/Dashboard/CommandHQ/defaultLayout.ts

@@ -0,0 +1,14 @@
+import type { LayoutNode } from 'officerdev';
+
+// Three columns, per `docs/command-hq.md` → "Layout, v1": category nav, commander grid, detail pane. The
+// detail pane gets the most room because it is a live chat transcript once something is selected.
+export const defaultLayout: LayoutNode = {
+  type: 'group',
+  id: 'command-hq-root',
+  direction: 'horizontal',
+  children: [
+    { node: { type: 'panel', id: 'command-hq-categories', appType: 'command-hq-categories' }, size: 16 },
+    { node: { type: 'panel', id: 'command-hq-grid', appType: 'command-hq-grid' }, size: 34 },
+    { node: { type: 'panel', id: 'command-hq-detail', appType: 'command-hq-detail' }, size: 50 },
+  ],
+};

src/apps/officer-web/Screens/Dashboard/CommandHQ/index.tsx

@@ -0,0 +1 @@
+export * from './CommandHqScreen';

src/apps/officer-web/Screens/Dashboard/Desktop/DesktopScreen.tsx

@@ -1,18 +0,0 @@
-import type { LayoutNode } from 'officerdev';
-import { WorkspaceView } from 'officerdev';
-import { useDashboardState } from 'state/useDashboardState';
-import { defaultLayout } from './defaultLayout';
-
-export const DesktopScreen = () => {
-  const workspace = useDashboardState<LayoutNode>('screens/desktop', defaultLayout);
-
-  return (
-    <div className="h-full w-full">
-      <WorkspaceView
-        workspace={workspace}
-        locked
-        appTypes={{ allowed: ['officerdev/desktop'], fallback: 'officerdev/desktop' }}
-      />
-    </div>
-  );
-};

src/apps/officer-web/Screens/Dashboard/Desktop/defaultLayout.ts

@@ -1,7 +0,0 @@
-import type { LayoutNode } from 'officerdev';
-
-export const defaultLayout: LayoutNode = {
-  type: 'panel',
-  id: 'desktop-screen',
-  appType: 'officerdev/desktop',
-};

src/apps/officer-web/Screens/Dashboard/Desktop/index.tsx

@@ -1 +0,0 @@
-export * from './DesktopScreen';

src/apps/officer-web/Screens/Dashboard/Gitea/GiteaScreen.tsx

@@ -1,38 +0,0 @@
-import { Navigate, useParams } from 'react-router';
-import type { LayoutNode } from 'officerdev';
-import { WorkspaceView, DEFAULT_GITEA_SECTION, giteaSectionPath, isGiteaSection } from 'officerdev';
-import { useDashboardState } from 'state/useDashboardState';
-import { defaultLayout } from './defaultLayout';
-
-// /gitea uses the Workspace/Panel system (like /transmission): a section nav on the left and the section view
-// on the right. Both talk to the officer-gitea sidecar through the /api/gitea auth proxy, which holds no
-// Gitea credentials of its own — the instance URL and the access token live in the sidecar.
-//
-// The open section is :section in the URL, so the panels read it with useParams instead of passing state
-// between themselves over a channel. This screen backs both /gitea and /gitea/:section and is the single
-// place that decides what an absent or bogus section means.
-
-export const GiteaScreen = () => {
-  const { section, owner, name } = useParams();
-  const workspace = useDashboardState<LayoutNode>('screens/gitea', defaultLayout);
-
-  // Bare /gitea, or a section that doesn't exist, resolves to a canonical URL rather than rendering a default
-  // while the address bar says something else — the nav highlight is derived from the URL.
-  //
-  // A repository route (/gitea/repo/:owner/:name/…) has no :section at all and is exempt: it is a second,
-  // deeper shape on the same screen, not a malformed section.
-  const inRepo = !!owner && !!name;
-  if (!inRepo && !isGiteaSection(section)) {
-    return <Navigate to={giteaSectionPath(DEFAULT_GITEA_SECTION)} replace />;
-  }
-
-  return (
-    <div className="h-full w-full pt-2">
-      <WorkspaceView
-        workspace={workspace}
-        locked
-        appTypes={{ allowed: ['gitea-nav', 'gitea-view'], fallback: 'gitea-view' }}
-      />
-    </div>
-  );
-};

src/apps/officer-web/Screens/Dashboard/Gitea/defaultLayout.ts

@@ -1,11 +0,0 @@
-import type { LayoutNode } from 'officerdev';
-
-export const defaultLayout: LayoutNode = {
-  type: 'group',
-  id: 'gitea-root',
-  direction: 'horizontal',
-  children: [
-    { node: { type: 'panel', id: 'gitea-nav', appType: 'gitea-nav' }, size: 22 },
-    { node: { type: 'panel', id: 'gitea-view', appType: 'gitea-view' }, size: 78 },
-  ],
-};

src/apps/officer-web/Screens/Dashboard/Gitea/index.tsx

@@ -1 +0,0 @@
-export * from './GiteaScreen';

src/apps/officer-web/Screens/Dashboard/Invoices/InvoicesScreen.tsx

@@ -1,35 +0,0 @@
-import type { LayoutNode } from 'officerdev';
-import { Navigate, useParams } from 'react-router';
-import { DEFAULT_INVOICES_SECTION, invoicesSectionPath, isInvoicesSection, WorkspaceView } from 'officerdev';
-import { useDashboardState } from 'state/useDashboardState';
-import { defaultLayout } from './defaultLayout';
-
-// /invoices uses the Workspace/Panel system (like /transmission and /offscale): the section nav on the
-// left, the section itself on the right. Both panels talk to the officer-invoiceshelf sidecar through the
-// /api/invoiceshelf auth proxy — the InvoiceShelf URL, its Sanctum token and the company header live in the
-// sidecar and never reach the browser.
-//
-// The open section is :section in the URL; list filters, the selected record and the open editor are query
-// params. So the panels read the URL rather than passing state to each other, and this screen is the single
-// place that decides what an absent or bogus section means.
-
-export const InvoicesScreen = () => {
-  const { section } = useParams();
-  const workspace = useDashboardState<LayoutNode>('screens/invoices', defaultLayout);
-
-  // Bare /invoices, or a section that doesn't exist, resolves to a canonical URL rather than rendering a
-  // default while the address bar says something else — the nav highlight is derived from the URL.
-  if (!isInvoicesSection(section)) {
-    return <Navigate to={invoicesSectionPath(DEFAULT_INVOICES_SECTION)} replace />;
-  }
-
-  return (
-    <div className="h-full w-full pt-2">
-      <WorkspaceView
-        workspace={workspace}
-        locked
-        appTypes={{ allowed: ['invoices-nav', 'invoices-view'], fallback: 'invoices-view' }}
-      />
-    </div>
-  );
-};

src/apps/officer-web/Screens/Dashboard/Invoices/defaultLayout.ts

@@ -1,11 +0,0 @@
-import type { LayoutNode } from 'officerdev';
-
-export const defaultLayout: LayoutNode = {
-  type: 'group',
-  id: 'invoices-root',
-  direction: 'horizontal',
-  children: [
-    { node: { type: 'panel', id: 'invoices-nav', appType: 'invoices-nav' }, size: 20 },
-    { node: { type: 'panel', id: 'invoices-view', appType: 'invoices-view' }, size: 80 },
-  ],
-};

src/apps/officer-web/Screens/Dashboard/Invoices/index.tsx

@@ -1 +0,0 @@
-export * from './InvoicesScreen';

src/apps/officer-web/Screens/Dashboard/Jellyfin/JellyfinScreen.tsx

@@ -1,37 +0,0 @@
-import { Navigate, useParams } from 'react-router';
-import type { LayoutNode } from 'officerdev';
-import { WorkspaceView, DEFAULT_JELLYFIN_SECTION, jellyfinSectionPath, isJellyfinSection } from 'officerdev';
-import { useDashboardState } from 'state/useDashboardState';
-import { defaultLayout } from './defaultLayout';
-
-// /jellyfin uses the Workspace/Panel system: the section and library nav (jellyfin-nav) on the left, the
-// section view (jellyfin-view) on the right. Both talk to the officer-jellyfin sidecar through the
-// /api/jellyfin auth proxy; the Jellyfin URL and its access token live in the sidecar and the browser never
-// sees either.
-//
-// Video only — the owner's audio is Officer's own /music player, and the sidecar already filters music
-// libraries out of everything this screen can ask for.
-//
-// The open section is :section in the URL; which library, which item and what is playing are in the query
-// string. Nothing about "what is open" lives in a panel channel.
-
-export const JellyfinScreen = () => {
-  const { section } = useParams();
-  const workspace = useDashboardState<LayoutNode>('screens/jellyfin', defaultLayout);
-
-  // Bare /jellyfin, or a section that doesn't exist, canonicalises rather than rendering a default behind a
-  // URL that names something else — the nav highlight comes from the router, so a bogus URL highlights nothing.
-  if (!isJellyfinSection(section)) {
-    return <Navigate to={jellyfinSectionPath(DEFAULT_JELLYFIN_SECTION)} replace />;
-  }
-
-  return (
-    <div className="h-full w-full pt-2">
-      <WorkspaceView
-        workspace={workspace}
-        locked
-        appTypes={{ allowed: ['jellyfin-nav', 'jellyfin-view'], fallback: 'jellyfin-view' }}
-      />
-    </div>
-  );
-};

src/apps/officer-web/Screens/Dashboard/Jellyfin/defaultLayout.ts

@@ -1,11 +0,0 @@
-import type { LayoutNode } from 'officerdev';
-
-export const defaultLayout: LayoutNode = {
-  type: 'group',
-  id: 'jellyfin-root',
-  direction: 'horizontal',
-  children: [
-    { node: { type: 'panel', id: 'jellyfin-nav', appType: 'jellyfin-nav' }, size: 18 },
-    { node: { type: 'panel', id: 'jellyfin-view', appType: 'jellyfin-view' }, size: 82 },
-  ],
-};

src/apps/officer-web/Screens/Dashboard/Jellyfin/index.tsx

@@ -1 +0,0 @@
-export * from './JellyfinScreen';

src/apps/officer-web/Screens/Dashboard/Jobs/JobDetail.tsx

@@ -142,7 +142,7 @@ const StatusIcon = ({ status, className }: { status: string; className?: string
   switch (status) {
     case 'completed':
     case 'complete':
-      return <CheckCircle2 className={`${base} text-green-500`} />;
+      return <CheckCircle2 className={`${base} text-success`} />;
     case 'failed':
     case 'error':
       return <AlertCircle className={`${base} text-red-500`} />;

src/apps/officer-web/Screens/Dashboard/Jobs/JobsPage.tsx

@@ -44,7 +44,7 @@ const formatDate = (iso: string) =>
 const StatusIcon = ({ status }: { status: JobSummary['status'] }) => {
   switch (status) {
     case 'completed':
-      return <CheckCircle2 className="h-4 w-4 text-green-500 shrink-0" />;
+      return <CheckCircle2 className="h-4 w-4 text-success shrink-0" />;
     case 'failed':
       return <AlertCircle className="h-4 w-4 text-red-500 shrink-0" />;
     case 'running':

src/apps/officer-web/Screens/Dashboard/Layout/DashboardLayout.tsx

@@ -1,12 +1,13 @@
 import { useMemo, useRef } from 'react';
 import { useLocation } from 'react-router';
-import { useDock, usePanelFullscreen } from 'officerdev';
+import { useDock, usePanelFullscreen, useChatNotifications } from 'officerdev';
 import { usePermissions } from 'hooks/usePermissions';
 import { ErrorBoundary } from '@/components/ErrorBoundary';
 import { ScreenErrorFallback } from './ScreenErrorFallback';
 import { Background } from './Background';
 import { Header } from './Header';
 import { Dock, CORE_DOCK_ITEMS, dockItemsFromPlugins, defaultDockPaths } from './Dock';
+import { Spotlight } from './Spotlight/Spotlight';
 import { useIsTouch } from './useIsTouch';
 import { usePageTitleSync } from '@/state/usePageTitle';
 import { RouteGate } from './RouteGate';
@@ -33,6 +34,7 @@ export function DashboardLayout({ children }: DashboardLayoutProps) {
   const isTouch = useIsTouch();
   const { pathname } = useLocation();
   usePageTitleSync();
+  useChatNotifications();
   // A panel can maximize into the content region on its own, but it cannot paint over this header — the
   // region below is an `absolute z-2` stacking context and the header is a `fixed z-10` sibling of it. So
   // "full screen" is cooperative: the panel asks, and the chrome steps aside.
@@ -46,6 +48,9 @@ export function DashboardLayout({ children }: DashboardLayoutProps) {
   return (
     <div className="relative flex h-dvh flex-col overflow-hidden outline-none inset-0">
       <Header dockItems={visibleItems} hidden={panelFullscreen} />
+      {/* Outside the content region on purpose, like the header and the dock: whatever a screen is doing
+          — maximised panel, error boundary, mid-navigation — ctrl-space still gets you off it. */}
+      <Spotlight />
       {/* overflow-CLIP, not hidden: `hidden` still makes this a scroll container, and the nav Dock —
           absolute, parked below the bottom edge by translateY while hidden — adds its transformed box to
           the scrollable overflow. So clicking a link let the browser scroll this section ~70px to reveal

src/apps/officer-web/Screens/Dashboard/Layout/Dock.tsx

@@ -154,6 +154,7 @@ import {
   GitBranch,
   Store,
   Puzzle,
+  Radar,
 } from 'lucide-react';
 
 /**
@@ -171,7 +172,7 @@ export const CORE_DOCK_ITEMS: DockItem[] = [
   { label: 'Home', to: '/', icon: Home, color: '#f59e0b' },
   { label: 'Files', to: '/files', icon: FolderOpen, color: '#fbbf24' },
   { label: 'Chat', to: '/chat', icon: MessageCircle, color: '#60a5fa' },
-  { label: 'Gitea', to: '/gitea', icon: GitBranch, color: '#609926' },
+  { label: 'Command HQ', to: '/command-hq', icon: Radar, color: '#f43f5e' },
   { label: 'Editor', to: '/code-editor', icon: Code, color: '#a78bfa' },
   { label: 'Jobs', to: '/jobs', icon: Workflow, color: '#14b8a6' },
   { label: 'Terminal', to: '/terminal', icon: Monitor, color: '#f97316' },
@@ -179,6 +180,17 @@ export const CORE_DOCK_ITEMS: DockItem[] = [
   { label: 'Monitor', to: '/system-monitor', icon: Activity, color: '#0ea5e9' },
   { label: 'Activity', to: '/activity', icon: Radio, color: '#f59e0b' },
   { label: 'Dashboards', to: '/dashboards', icon: LayoutGrid, color: '#8b5cf6' },
+  // Core since 2026-08-19. These were contributed by the app-store `caldav` entry until calendar and
+  // contacts became core, and deleting that entry took their tiles with it — the routes kept working
+  // while the dock and the settings list went silent about them, which is exactly the shape of failure
+  // this file's comment above warns about, arrived at from the other direction.
+  //
+  // Artwork rather than glyphs since 2026-08-19: both were redrawn to the icon brief, so they carry
+  // the same voxel mark the phone apps show. `tile: 'bare'` for the same reason a plugin's artwork
+  // does — the art brings its own frame, and a coloured square behind it reads as a mistake. `color`
+  // still earns its place: the active glow, the indicator dot and the settings tint all use it.
+  { label: 'Calendar', to: '/calendar', image: '/icons/calendar.png', tile: 'bare', color: '#ef4444' },
+  { label: 'Contacts', to: '/contacts', image: '/icons/contacts.png', tile: 'bare', color: '#8b5cf6' },
   // Core because the tailnet is the perimeter — origin checking was removed on the grounds that the
   // tailnet stands in its place, so administering it cannot be an optional extra. It is `kind: 'admin'`,
   // and DashboardLayout filters every tile through canVisit(), so a member never sees this one.
@@ -235,7 +247,11 @@ export function dockItemsFromPlugins(plugins: PluginManifest[]): DockItem[] {
  * where Music was never installed. Defaults that reference optional features are how an app ends up
  * looking subtly wrong on a fresh install for no stated reason.
  */
-export const DEFAULT_DOCK_PATHS = ['/', '/files', '/terminal', '/dashboards', '/chat'];
+// Calendar joined on 2026-08-19, when it stopped being an installable app and became core. The rest of
+// this list is what every install has and most people want visible; a calendar is now in that category.
+// Contacts is deliberately NOT here — it is core and available in dock settings, but a second tile for
+// the same sidecar in everyone's default dock is a crowded default rather than a useful one.
+export const DEFAULT_DOCK_PATHS = ['/', '/files', '/terminal', '/dashboards', '/chat', '/calendar'];
 
 /**
  * The same defaults, plus a tile for every installed plugin.

src/apps/officer-web/Screens/Dashboard/Layout/Header/Header.tsx

@@ -81,7 +81,10 @@ export function Header({ dockItems, hidden }: HeaderProps) {
               role="img"
               aria-label="Officer"
               className="h-8 md:h-12 w-24 md:w-36 bg-contain bg-left bg-no-repeat"
-              style={{ backgroundImage: logo ? `url(${logo})` : 'none', transform: 'skew(-15deg, -2deg)' }}
+              /* No skew. The wordmark is drawn, kerned pair by pair, and shipped as the finished mark —
+                 shearing it in CSS was how the OLD logo got its slant, and doing it to this one distorts
+                 letterforms that were adjusted by hand. See landing/assets/brand/WORDMARK.md. */
+              style={{ backgroundImage: logo ? `url(${logo})` : 'none' }}
             />
           </Link>
         </div>

src/apps/officer-web/Screens/Dashboard/Layout/Rescan/RescanButton.tsx

@@ -33,7 +33,28 @@ export function RescanButton() {
       await Promise.all(ITEM_QUERY_KEYS.map((key) => qc.invalidateQueries({ queryKey: [key], refetchType: 'all' })));
       const c = res.counts ?? {};
       toast.success(
-        `Rescanned items — ${c.tasks ?? 0} tasks, ${c.agents ?? 0} agents, ${c.tools ?? 0} tools, ${c.skills ?? 0} skills, ${c.processes ?? 0} processes`,
+        `Rescanned items — ${c.tasks ?? 0} tasks, ${c.agents ?? 0} agents, ${c.tools ?? 0} tools, ${c.skills ?? 0} skills, ${c.processes ?? 0} processes, ${c.affordances ?? 0} affordances`,
+        {
+          // Two things said here rather than left to be discovered.
+          //
+          // A rescan cannot reach a chat that is already open — its system prompt was assembled when it
+          // started. And a SKILL.md with no `description` is COUNTED as a skill but never offered to an
+          // agent, because the loader skips it: a line that does not say when to reach for something
+          // costs context in every conversation and earns nothing. So a skill can be present, listed and
+          // editable while being invisible to every agent, and the only symptom is two numbers
+          // disagreeing where nobody is comparing them.
+          description:
+            [
+              c.affordances || c.skillsOffered
+                ? 'Applies to new chats — an open one keeps what it started with.'
+                : null,
+              typeof c.skills === 'number' && typeof c.skillsOffered === 'number' && c.skillsOffered < c.skills
+                ? `${c.skills - c.skillsOffered} skill(s) have no description, so no agent is offered them.`
+                : null,
+            ]
+              .filter(Boolean)
+              .join(' ') || undefined,
+        },
       );
     } catch {
       toast.error('Failed to rescan items');

src/apps/officer-web/Screens/Dashboard/Layout/Spotlight/Spotlight.tsx

@@ -0,0 +1,249 @@
+import { useCallback, useEffect, useMemo, useRef, useState } from 'react';
+import { useNavigate } from 'react-router';
+import { Search } from 'lucide-react';
+import { rank } from './match';
+import { useSpotlightItems, type SpotlightItem } from './useSpotlightItems';
+
+// Spotlight: type a few letters, go somewhere.
+//
+// The dock is full and cannot grow, which is what this is for — once navigation lives here, the dock is
+// free to be a short list of what you actually use rather than a complete one.
+//
+// ── Keyboard first, and the trigger ──
+//
+// `shift-space`. The two obvious chords were both taken: `ctrl-space` focuses the address bar in Brave,
+// and `alt-space` is load-bearing for the owner elsewhere. `cmd-k` belongs to the browser on some
+// platforms and to text fields on others.
+//
+// **Shift-space is a real key combination inside a text field** — it types a space, and people hold shift
+// through one more often than they think while typing capitals. So the trigger is IGNORED when focus is
+// in an input, a textarea or anything contenteditable, and the space is typed as normal.
+//
+// The cost of that is worth stating: focus usually IS in the chat composer, so opening Spotlight from
+// there means clicking out first. If that turns out to be the common case, a double-tap of shift is the
+// escape hatch — it conflicts with nothing anywhere.
+//
+// Escape closes, arrows move, enter goes. The mouse works, but nothing here is designed around it.
+
+/*
+  No cap, deliberately, and it used to be 8.
+
+  Eight was fine on the machine it was written on, which has few plugins. On a populated server the dock
+  carries fourteen plugin tiles and the core items alone fill all eight slots — so with an empty query the
+  list stopped at Browser and every plugin was unreachable unless you already knew its name well enough to
+  type it. A launcher you can only use for things you can already name is not doing the job.
+
+  The container has scrolled since it was written (`max-h-[52vh] overflow-y-auto`), so there was never a
+  layout reason for the limit, and the truncation was silent — nothing said "and 20 more", which is how a
+  cap becomes indistinguishable from a missing feature.
+*/
+
+/**
+ * Is the user typing into something?
+ *
+ * `shift-space` types a space in any text field, so the trigger has to stand down there or it eats a
+ * keystroke people make by accident all day. `isContentEditable` covers the editor and any rich input a
+ * panel brings with it.
+ */
+function isTyping(target: EventTarget | null): boolean {
+  const el = target as HTMLElement | null;
+  if (!el || !el.tagName) return false;
+  return el.tagName === 'INPUT' || el.tagName === 'TEXTAREA' || el.isContentEditable === true;
+}
+
+/** Label with the matched characters picked out, so it is obvious WHY a result is here. */
+const Highlighted = ({ label, hits }: { label: string; hits: number[] }) => {
+  if (!hits.length) return <>{label}</>;
+  const set = new Set(hits);
+  return (
+    <>
+      {label.split('').map((char, i) =>
+        set.has(i) ? (
+          <span key={i} className="text-duck-teal">
+            {char}
+          </span>
+        ) : (
+          <span key={i}>{char}</span>
+        ),
+      )}
+    </>
+  );
+};
+
+const Row = ({
+  item,
+  hits,
+  active,
+  onPick,
+}: {
+  item: SpotlightItem;
+  hits: number[];
+  active: boolean;
+  onPick: () => void;
+}) => {
+  const Icon = item.icon;
+  // Keyboard navigation has to survive a list taller than the box. `block: 'nearest'` scrolls only when the
+  // row is actually out of sight, so arrowing through the visible middle of the list does not jump it.
+  const ref = useRef<HTMLButtonElement | null>(null);
+  useEffect(() => {
+    if (active) ref.current?.scrollIntoView({ block: 'nearest' });
+  }, [active]);
+  return (
+    <button
+      ref={ref}
+      type="button"
+      // `onMouseDown`, not `onClick`: the input holds focus and a click would blur it first, closing the
+      // overlay on the blur handler before the navigation ever ran.
+      onMouseDown={(ev) => {
+        ev.preventDefault();
+        onPick();
+      }}
+      className={`flex w-full items-center gap-3 rounded-lg px-3 py-2 text-left transition-colors ${
+        active ? 'bg-accent' : 'hover:bg-accent/50'
+      }`}
+    >
+      {/*
+        The dock's tile, at row size. Same two cases and the same rule for choosing between them, because
+        these are the SAME things and a second visual language for them is what made this list read as
+        somebody else's menu:
+
+          badge  a filled square of the item's colour with a white glyph — a white glyph on nothing is
+                 invisible, which is why the fill is not decoration
+          bare   no background, artwork fills the box — a logo framed in an arbitrary swatch reads as a
+                 mistake, and every plugin shipping its own art is this case
+
+        The tinted `${color}22` square this used for everything was a third style belonging to neither, and
+        it put a coloured pane behind Calendar, Contacts and every plugin logo.
+      */}
+      <span
+        className="flex h-7 w-7 shrink-0 items-center justify-center rounded-md"
+        style={{
+          background: item.tile === 'bare' ? 'transparent' : (item.color ?? 'transparent'),
+          // No colour at all is a settings row: keep the muted glyph rather than inventing a swatch.
+          color: item.color ? '#fff' : undefined,
+        }}
+      >
+        {item.image ? (
+          <img
+            src={item.image}
+            alt=""
+            className={item.tile === 'bare' ? 'h-full w-full object-contain' : 'h-5 w-5 object-contain'}
+          />
+        ) : Icon ? (
+          <Icon className={`h-4 w-4 ${item.color ? '' : 'text-muted-foreground'}`} />
+        ) : (
+          <Search className="h-4 w-4 text-muted-foreground" />
+        )}
+      </span>
+      <span className="min-w-0 flex-1">
+        <span className="block truncate text-sm text-foreground">
+          <Highlighted label={item.label} hits={hits} />
+        </span>
+        {item.hint ? <span className="block truncate text-xs text-muted-foreground">{item.hint}</span> : null}
+      </span>
+      <span className="shrink-0 font-mono text-[10px] text-muted-foreground/60">{item.to}</span>
+    </button>
+  );
+};
+
+export const Spotlight = () => {
+  const [open, setOpen] = useState(false);
+  const [query, setQuery] = useState('');
+  const [active, setActive] = useState(0);
+  const inputRef = useRef<HTMLInputElement | null>(null);
+  const navigate = useNavigate();
+  const items = useSpotlightItems();
+
+  const results = useMemo(() => rank(items, query), [items, query]);
+
+  const close = useCallback(() => {
+    setOpen(false);
+    setQuery('');
+    setActive(0);
+  }, []);
+
+  const pick = useCallback(
+    (item: SpotlightItem | undefined) => {
+      if (!item) return;
+      close();
+      navigate(item.to);
+    },
+    [close, navigate],
+  );
+
+  // The trigger. Bound to the window rather than to a container, because it has to work while focus is
+  // anywhere — including inside a panel that has its own key handling.
+  useEffect(() => {
+    const onKey = (ev: KeyboardEvent) => {
+      if (ev.shiftKey && ev.code === 'Space' && !isTyping(ev.target)) {
+        ev.preventDefault();
+        setOpen((wasOpen) => !wasOpen);
+        return;
+      }
+      if (ev.key === 'Escape' && open) close();
+    };
+    window.addEventListener('keydown', onKey);
+    return () => window.removeEventListener('keydown', onKey);
+  }, [open, close]);
+
+  useEffect(() => {
+    if (open) inputRef.current?.focus();
+  }, [open]);
+
+  // Reset the cursor whenever the result set changes, so it never points past the end.
+  useEffect(() => setActive(0), [query]);
+
+  if (!open) return null;
+
+  return (
+    <div
+      className="fixed inset-0 z-50 flex items-start justify-center bg-black/40 pt-[12vh] backdrop-blur-sm"
+      onMouseDown={close}
+    >
+      <div
+        className="w-[min(92vw,560px)] overflow-hidden rounded-2xl border border-border bg-background/95 shadow-2xl"
+        // The backdrop closes on mousedown; this stops a click INSIDE the panel counting as one.
+        onMouseDown={(ev) => ev.stopPropagation()}
+      >
+        <div className="flex items-center gap-2 border-b border-border px-4">
+          <Search className="h-4 w-4 shrink-0 text-muted-foreground" />
+          <input
+            ref={inputRef}
+            value={query}
+            onChange={(ev) => setQuery(ev.target.value)}
+            onKeyDown={(ev) => {
+              if (ev.key === 'ArrowDown') {
+                ev.preventDefault();
+                setActive((i) => Math.min(i + 1, results.length - 1));
+              } else if (ev.key === 'ArrowUp') {
+                ev.preventDefault();
+                setActive((i) => Math.max(i - 1, 0));
+              } else if (ev.key === 'Enter') {
+                ev.preventDefault();
+                pick(results[active]?.item);
+              }
+            }}
+            placeholder="Go to…"
+            className="w-full bg-transparent py-3.5 text-sm text-foreground outline-none placeholder:text-muted-foreground"
+          />
+        </div>
+
+        <div className="max-h-[52vh] overflow-y-auto p-2">
+          {results.length === 0 ? (
+            <p className="px-3 py-6 text-center text-sm text-muted-foreground">Nothing matches “{query}”.</p>
+          ) : (
+            results.map((result, i) => (
+              <Row
+                key={result.item.id}
+                item={result.item}
+                hits={result.hits}
+                active={i === active}
+                onPick={() => pick(result.item)}
+              />
+            ))
+          )}
+        </div>
+      </div>
+    </div>
+  );
+};

src/apps/officer-web/Screens/Dashboard/Layout/Spotlight/match.test.ts

@@ -0,0 +1,67 @@
+import { describe, expect, test } from 'bun:test';
+import { rank } from './match';
+import type { SpotlightItem } from './useSpotlightItems';
+
+// The behaviours a launcher is judged on, which are all about ORDER rather than about matching. Anyone
+// can return the right set; returning it with the right thing first is the whole product.
+
+const item = (label: string, to: string, hint?: string): SpotlightItem => ({ id: to, label, to, hint });
+
+const ITEMS: SpotlightItem[] = [
+  item('Home', '/'),
+  item('Files', '/files'),
+  item('Chat', '/chat'),
+  item('Command HQ', '/command-hq'),
+  item('Editor', '/code-editor'),
+  item('Profile', '/settings/profile', 'avatar, name, email'),
+  item('API keys', '/settings/profile/api-keys', 'per-app keys'),
+  item('App store', '/app-store'),
+  item('Dashboards', '/dashboards'),
+];
+
+const top = (query: string, n = 1) =>
+  rank(ITEMS, query)
+    .slice(0, n)
+    .map((s) => s.item.label);
+
+describe('spotlight ranking', () => {
+  test('an exact label wins', () => {
+    expect(top('chat')).toEqual(['Chat']);
+  });
+
+  test('a prefix beats a scattered match — "fi" is Files, not Profile', () => {
+    expect(top('fi')).toEqual(['Files']);
+  });
+
+  test('initials find a two-word label — "chq" is Command HQ', () => {
+    expect(top('chq')).toEqual(['Command HQ']);
+  });
+
+  test('a subsequence works, so you can type the shape of a word', () => {
+    expect(top('apik')).toEqual(['API keys']);
+  });
+
+  test('a hint matches, but never outranks a label', () => {
+    // "keys" is API keys' own label AND its hint. It must not be beaten by anything mentioning keys.
+    expect(top('keys')).toEqual(['API keys']);
+  });
+
+  test('a path fragment finds the item without outranking labels', () => {
+    expect(rank(ITEMS, 'settings').map((s) => s.item.label)).toContain('Profile');
+    // "app" is a prefix of "App store" and appears in API keys' hint ("per-app keys"). Label wins.
+    expect(top('app')).toEqual(['App store']);
+  });
+
+  test('nothing matches means nothing returned, not everything', () => {
+    expect(rank(ITEMS, 'zzzzz')).toHaveLength(0);
+  });
+
+  test('an empty query keeps the given order rather than sorting it', () => {
+    expect(rank(ITEMS, '').map((s) => s.item.label)).toEqual(ITEMS.map((i) => i.label));
+  });
+
+  test('shorter labels break ties — "Chat" over "Chat history"', () => {
+    const withLonger = [...ITEMS, item('Chat history', '/chat-history')];
+    expect(rank(withLonger, 'chat')[0]!.item.label).toBe('Chat');
+  });
+});

src/apps/officer-web/Screens/Dashboard/Layout/Spotlight/match.ts

@@ -0,0 +1,93 @@
+import type { SpotlightItem } from './useSpotlightItems';
+
+// Ranking, kept apart from React so it can be tested with strings.
+//
+// Subsequence matching, not substring: "cmdhq" should find "Command HQ" and "apik" should find
+// "API keys". Anyone who has used a launcher expects to type the shape of a word rather than a prefix
+// of it, and a substring match makes them type more to get less.
+
+export type Scored = { item: SpotlightItem; score: number; hits: number[] };
+
+/**
+ * Where each character of `query` lands in `text`, or null if it does not fit.
+ *
+ * Greedy left-to-right, which is what makes it cheap and predictable. It can miss a better alignment —
+ * "ao" against "app store" takes the first `o` in the first word rather than the one in "store" — and
+ * that costs a slightly worse score, never a wrong answer, because the item still matches.
+ */
+function subsequence(text: string, query: string): number[] | null {
+  const hits: number[] = [];
+  let at = 0;
+  for (const char of query) {
+    const found = text.indexOf(char, at);
+    if (found === -1) return null;
+    hits.push(found);
+    at = found + 1;
+  }
+  return hits;
+}
+
+/** Start of the string, or the character after a space, dash or slash. */
+const isBoundary = (text: string, i: number): boolean => i === 0 || /[\s\-/]/.test(text[i - 1]!);
+
+/**
+ * Higher is better. The weights encode what a person means by "the right result":
+ *
+ * - **an exact label** is what they typed, so nothing beats it
+ * - **a prefix** beats a scattered match — typing "fi" should rank Files above Profile
+ * - **word boundaries** are why "chq" finds Command HQ at all, and should rank it above an item where
+ *   those letters happen to fall mid-word
+ * - **adjacency** rewards a run of letters, which is what a substring match would have found
+ * - **shortness** breaks ties toward the more specific label: "Chat" over "Chat history"
+ *
+ * The label is worth roughly three times the hint, so typing a word that IS a label beats an item that
+ * merely mentions it in passing.
+ */
+function scoreOne(text: string, query: string, weight: number): { score: number; hits: number[] } | null {
+  const hits = subsequence(text, query);
+  if (!hits) return null;
+
+  let score = 0;
+  if (text === query) score += 1000;
+  else if (text.startsWith(query)) score += 400;
+
+  for (let i = 0; i < hits.length; i++) {
+    const at = hits[i]!;
+    if (isBoundary(text, at)) score += 30;
+    if (i > 0 && at === hits[i - 1]! + 1) score += 15;
+  }
+
+  // Shorter labels win ties. Capped so a very long one is not driven negative.
+  score += Math.max(0, 40 - text.length);
+  return { score: score * weight, hits };
+}
+
+/**
+ * Rank items against a query.
+ *
+ * An empty query returns everything in its given order — the index is already in a sensible one (dock
+ * first, settings after), and imposing an alphabetical sort on an empty box would move Home away from
+ * the top for no reason.
+ */
+export function rank(items: SpotlightItem[], query: string): Scored[] {
+  const q = query.trim().toLowerCase();
+  if (!q) return items.map((item) => ({ item, score: 0, hits: [] }));
+
+  const scored: Scored[] = [];
+  for (const item of items) {
+    const label = scoreOne(item.label.toLowerCase(), q, 1);
+    const hint = item.hint ? scoreOne(item.hint.toLowerCase(), q, 0.3) : null;
+    // The PATH is matched too, unweighted almost to nothing: it is how "/settings/profile" finds Profile
+    // when somebody types a URL fragment, without letting a path ever outrank a real label.
+    const path = scoreOne(item.to.toLowerCase(), q, 0.15);
+
+    const best = [label, hint, path]
+      .filter((s): s is { score: number; hits: number[] } => !!s)
+      .sort((a, b) => b.score - a.score)[0];
+    if (!best) continue;
+    // `hits` is only meaningful for the label, since that is the only string rendered with highlighting.
+    scored.push({ item, score: best.score, hits: label?.hits ?? [] });
+  }
+
+  return scored.sort((a, b) => b.score - a.score || a.item.label.length - b.item.label.length);
+}