it has never run: db:push has not been applied and neither process has been
restarted. docs/ai-usage-panel.md records where every number comes from and
why officer cannot derive any of them itself, the two things the live
payload settled (every model-scoped bucket is null while limits[] carries a
real fable window; the window is named by kind, never by the model name),
the measured costs, and the staleness rules — which are the part most likely
to be wrong, because they assert things about data they did not measure.
sixteen ordered tests in five groups. group C is the actual feature. test 10
is the one that would break silently: a refresh must not overwrite an
account it could not measure.
also records what the scan measurement found, since it decides the deferred
work — 74.7% of message ids appear in more than one transcript file, so
per-file aggregates cannot be summed and an incremental scan is not the
one-liner it looks like.
TODO.md gains an entry under "waiting on the owner to run it", beside plan
mode, which is also still unexercised.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
half of the primitive kiosk needs. the manifest field and its validator are in;
the mounting is documented and not wired, so nothing serves anything publicly
today.
paths are relative to the plugin's own mount and refused rather than corrected:
no leading-slash-missing, no .., no //, no query or fragment, wildcard only as a
trailing /*, and never the bare / or /*. the plugin is trusted with what those
paths DO, never with which paths.
the mounting was measured rather than reasoned about, and two obvious designs
are wrong:
mounting the public router at the plugin prefix gives correct fall-through but
lets the router serve any route it defines - the manifest becomes documentation
rather than enforcement.
adding a use('*') guard blocks that leak and BREAKS the plugin's authenticated
api: a use('*') middleware always matches, so the parent never continues.
measured - an authenticated route 404'd with a valid token.
what works is registering only the declared patterns, delegating into a
pre-mounted app. undeclared paths never register a handler, so they fall through
naturally. enforcement by registration rather than by checking, which is why a
plugin cannot widen it.
docs/plugin-unauthenticated-paths.md has the probe results and the six
remaining steps.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
a sweep for remnants after 2026-08-18. nothing was broken - no dangling
import, symbol, route or permission key anywhere in src/ - but the platform
was still speaking the plugin's vocabulary in a dozen places.
the one that mattered is live code rather than prose: ChatPanelWrapper
hardcodes workspace.id === 'email' to send context: 'email', which is the
frontend half of the pair whose backend half (resolveEmailCwd) carries a
fifteen-line justification. the two ends now both say so, including that
adding a SECOND id there is the signal the platform needs a real way for a
plugin to declare a chat context.
the rest: two dead imports left by removing the settings section, an empty
'Email' type-section header, a protocol.ts section header with no members
naming the plugin's SSE design, and a queue comment pointing at
sidecar/email/sync-runner.ts - a path in another repository now.
CLAUDE.md, AGENTS.md and working-on-officer.md described email as a platform
feature, an app permission and a pm2 process. those are the files a fresh
agent reads first, so they were the highest-risk staleness in the tree.
TODO.md's whole Email section moved to the plugin's own repository. it held
OPEN items pointing at sidecar/email/store.ts, which is exactly how a leftover
list sends someone to edit files that do not exist.
the delete-account dialog promised it would remove email accounts and
playlists. both tables belong to plugins now, so that cascade only happens
where the plugin is installed - it now says 'anything an installed plugin
keeps for them' instead of naming two it cannot guarantee.
imapflow and mailparser STAY in package.json. the plugin has no package.json
of its own and resolves from the platform's node_modules, so removing them
would break the installed plugin's sidecar. the fix is giving plugins their
own dependency manifest, not deleting these.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the owner has no time to test any of it, so the checklist has to work for
someone with none of today's context: eleven ordered tests, what each should
show, and what its failure looks like.
test 4 is called out as the one most likely to be wrong - approving
ExitPlanMode drops the harness into default rather than back into
bypassPermissions, which is the finding that changed the design and therefore
the thing the implementation is most likely to have got wrong.
also separates what was verified from what was not: the probe runs against a
real claude, the turn-stream unit tests, and then everything end to end, which
is nothing.
the inline 'what to watch' list in chat-plan-mode.md is replaced by a pointer
so there is one source of truth, and TODO.md now says live-but-unexercised
rather than built-not-run - the restarts are done.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the owner confirmed the additional accounts are test rows - none has ever
signed in, no member has opened a shell or a chat. so the proxy secret was
readable by accounts that do not exist in practice, and there is nothing to
rotate.
the mode fix stands: it is about the install that gets used, not this one.
dropping the TODO and recording the reasoning instead, because a rotation item
left sitting there reads as outstanding work forever, and the next person to
find a world-readable secret should ask who uses the box before reaching for
the alarm.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
data/sidecar/claude-state.json holds proxySecret, and that string IS the
owner's Claude Max subscription: officer-anthropic-proxy accepts it as
x-api-key, swaps it for the owner's OAuth bearer and forwards to
api.anthropic.com. it listens on 127.0.0.1, which every local account can
reach.
it was 664 in a 775 directory. nothing had ever set a mode - mkdirSync and
Bun.write at ubuntu's default umask of 002. confirmed by reading the live
secret as 'nobody', an account with no relationship to the platform, and by
confirming the proxy 401s without the key so the secret is the only control.
this does NOT mean members shared the subscription. spawnClaudeAsMember runs
their claude through setpriv --reset-env with a six-name allowlist, so
ANTHROPIC_BASE_URL and ANTHROPIC_API_KEY cannot cross into it and their turns
go straight to anthropic on their own OAuth. that design is intact - it could
just be walked around by anyone with a shell.
fixed twice on purpose. state.ts sets 0600/0700 on every write, which is the
authoritative fix, and on every write rather than at creation because
writeFileSync leaves an existing file's mode alone. secureRuntimeSecrets()
sweeps at boot because officer-anthropic-proxy is a separate pm2 process that
can recreate the file hours after officer started, and because every existing
install has the old mode until something repairs it.
it repairs rather than refusing to boot, unlike assertSecretsClosed beside it.
the line is who wrote the file: .env is the installer's and we must not guess
at it; this is our own output at the machine's umask, so refusing to start
would punish the user for our bug.
the per-account copies are swept too. no live secret in them today, but they
map session keys to transcript uuids for named accounts.
tests verified to bite: removing the chmod while still reporting success fails
four of them.
a chmod does not undo exposure - the secret still needs rotating on the
production host. TODO.md has the sequence, including that officer-claude-code
caches it and must be restarted alongside the proxy.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
opusplan is in the picker and does nothing on this install: it is a per-MODE model switch (Opus while in
plan mode, Sonnet at rest — the CLI's own strings call them the upgrade and resting models), and
claude-manager hardcodes permissionMode: 'bypassPermissions', so plan mode is never entered. Not
orchestration, which is what the name suggests and what we both assumed.
Making the mode reachable is the fix. setPermissionMode is a control request on the live query, same channel
as the setModel that landed in 4c24df93, so the mechanism is proven. The hard part is ExitPlanMode: it is a
permission decision, and bypassPermissions auto-approves it, so a naive toggle produces a planner that
instantly approves its own plan. canUseTool is the hook, and whether it can be supplied after session
creation is the one thing to verify before designing any UI.
Records the landmines too — bypassPermissions is deliberate per CLAUDE.md and must be restored on exit, and
a canUseTool promise that never resolves wedges the turn.
`createMessageHandler` is built per turn, so `transcriptId` started null on every one — and
`applyModelChange` runs before the turn emits anything, so the `model:switched` it produces arrived while
there was nothing to key the row on. A resumed conversation therefore recorded its model only on the closing
`result`, and a turn that died took the switch with it.
Observed rather than reasoned: the sidecar log said `claude-code/fable -> claude-code/opus[1m]` while the
stored row still said `claude-code/opus`.
Seeded from the resume id, which the Claude path already has in scope for the spawn. A brand-new
conversation still waits for `session:claude` — it has no id yet by definition — but resumed is the case
where switching actually happens.
Also lands docs/chat-sessions-list.md: the plan of record for the sessions-list revamp, agreed with the owner
today. Six numbered targets, four decisions with their reasoning, and the landmines — chiefly that merging
the two left-column panels strands `chat-live` in every persisted layout, which is the same stale-appType
shape that made /offscale render three server pickers this morning (ed3806f9). Written down because the shape
came out of a conversation, not out of the code.
Asked for a thorough check that nothing of offscale is left in the platform. The
extraction was sound — no router, no permission entry, no app-store entry, no
screen, no panel app, no env var anything reads. Two things were genuinely wrong,
and both were in setup rather than in the server.
── The secret store minted a key for a plugin ──
`officer-setup` bootstrapped two purposes: `jwt` and `headscale`. The second was a
key for a PLUGIN, under a name the rename had already replaced, on a machine where
the plugin is not installed and nothing would ever ask for it. offscale encrypts
under `offscale`, and `getKey` creates on demand — which is precisely what lets
core know nothing about plugins. Now core bootstraps `jwt` alone.
── `.env.example` documented variables for a deleted router ──
HEADSCALE_URL, HEADSCALE_API_KEY and HEADSCALE_USER were described as driving
`/api/vpn`. That router was deleted on 2026-08-14 and nothing has read any of them
since. Left as a note saying where the tailnet's credentials live now, because
somebody with an older `.env` will otherwise wonder why setting them does nothing.
── Names that outlived the thing ──
Eleven comments referred to the plugin's route, panels, console or server registry
as "headscale": three sidecar config files describing their own shape by pointing
at it, two screens citing `/headscale` as the example of a Workspace/Panel route,
the terminal's note about the Console, and the photos types. Also `catalogue.ts`,
which still claimed headscale "is CORE now", and `working-on-officer.md`, which
told the reader a core install runs six processes including `officer-headscale`.
Upstream references are deliberately untouched: "Headscale's API", "the Headscale
admin credential", "is this a Headscale server?" all name the thing the plugin
drives, and renaming those would make the sentences false.
`docs/secret-store.md` argued headscale must be core because the tailnet is the
perimeter. Marked superseded rather than deleted, with the distinction that
settles it: the perimeter is the tailnet EXISTING, which machine-setup still
arranges on a core install — not the platform shipping the screen that
administers it.
Verified by counting: live code, setup provisioning, env references, permission
entries, app-store entries, hono mounts, dock defaults, screens and panel apps all
zero. What is left is generated files (Plugins.gen, plugin-schemas.gen), the
bundled catalogue entry, test fixtures using "offscale" as an example app name,
and history.
tsgo clean, frontend builds, 873 pass / 7 fail unchanged, setup scripts parse.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This bug will not be reproduced on demand. It will arrive as somebody's report,
hours later, from a browser whose console is long gone — so a `console.warn` is
seen by nobody, and "what did you see?" gets an answer built from memory and a
screenshot.
So the trap now persists. Up to five occurrences in
`localStorage['officer:editor-buffer-trap']`, each carrying the stack at the moment
of the edit, the first 200 characters that arrived, how many, which file, and when.
The question becomes "paste this key", and the answer is complete.
Recording a bug must never become one: private mode, a full quota and a disabled
storage API all throw at that write, and none of them is worth breaking an edit
over. The console line survives either way.
`docs/editor-buffer-corruption.md` is the procedure — what it looks like, the four
candidates already eliminated and why, the one command to ask a reporter for, how
to read each field, what an EMPTY key would mean (either the editor was focused,
making it a paste, or the insertion came in chunks under the threshold), and the
instruction to delete every piece of this once the cause is known. Linked from
TODO.md and from CLAUDE.md's reading list, so somebody handed a report finds it
without grepping for it.
Stated there rather than left implicit: a fix invented without knowing the cause
would hide the symptom on the one screen where it is visible while the mechanism
carried on doing whatever else it does. The corruption is never saved and reopening
the tab clears it, so waiting is cheap and a wrong fix is not.
tsgo clean, frontend builds, 873 pass / 7 fail unchanged.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Split out of docs/marketplace.md so the design document stays about how it works.
Ranked by what blocks something rather than by size: submissions first (until a
stranger's manifest can be read without executing it, the catalogue is ours only
and everything below is cosmetic), then plugin upgrades not migrating, then
staleness, then scale.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The site is up on HTTPS with a Let's Encrypt certificate, proxied to :9100, and
serving the three plugins out of `officer_marketplace`. An Officer install with no
configuration now reads its catalogue from it.
── What is running ──
PM2 entry added to THIS machine's ecosystem only. Deliberately not in
officer-setup's CORE_PROCESSES: the marketplace is one website officer.dev hosts,
not something every install should run — a copy on every machine would serve a
stale catalogue on a port nobody asked to open.
nginx (NPM host id 7) → 172.20.0.1:9100, with a ufw rule for the docker bridge.
That rule is not optional and its absence does not look like a firewall problem:
without it nginx HANGS until proxy_connect_timeout rather than being refused,
which is how the apptest hosts wasted an hour a few days ago.
── READMEs ──
`music`, `offscale` and `example` now have proper READMEs, written against their
own code and pushed to their own repos. They render on the plugin's page in the
store and are what a self-hoster reads before installing. The example's is a
tutorial as much as a shop window — it is the document a plugin author reads
first, and every convention in it was verified in `discover.ts` rather than
restated from memory.
`music`'s remote was HTTPS with no credentials on this box, so it could not be
pushed; switched to SSH like the other two. The catalogue still advertises HTTPS —
`marketplace-sync` converts — which is the property that matters, since an
`ssh://` clone URL only works on the machine that published it.
── Writing them was the cheapest audit available ──
Reading each plugin against its own documentation turned up ~20 real defects, all
recorded in TODO.md and none fixed here. The pattern worth naming: offscale is
still branded "headscale" in user-visible panel names and in a persisted terminal
session id, and its PLUGIN.md states outright that the ACL assistant is
deliberately unwired — while `sidecar/assist.ts`, `claude-proxy.ts` and
`PolicyAssistant.tsx` all ship. Documentation that confidently describes the
opposite of the code is worse than none.
Also fixed the docs on the frozen lockfile: both CLAUDE.md and bunfig.toml named
`bun install --no-frozen-lockfile` as the escape hatch, and bun 1.3.14 has no such
flag — it is ignored and the install fails anyway.
tsgo clean, 873 pass / 7 fail unchanged.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
44 files of prose — CLAUDE.md, AGENTS.md, TODO.md, 20 docs, both plugin design
documents, and the comment surface the earlier steps could not reach.
Applied against an explicit keep-list, not swept, because the word turned out to
have SIX meanings in this repository rather than the three the offscale doc
recorded:
permissions renamed (steps 1–2)
$OFFICER_ROOT/capabilities/ KEPT — the item store, and now the only thing
the word means that is ours
sidecar routing keys renamed to `handles` (step 3)
Lightning wallet KEPT — a domain term, and on the wire to the mobile apps
terminfo queries KEPT — XTGETTCAP, in the pty sidecar
InvoiceShelf KEPT — per-resource { write, bulkDelete } flags
The sweep still falsified two things, both caught by checking rather than by
review, and both in prose that discusses more than one meaning at once:
CLAUDE.md began claiming the item store lives at `$OFFICER_ROOT/permissions`.
It does not; that directory is on disk and full of skills and tools.
And the offscale doc's own note about the collision became
"Named `permissions`, NOT `permissions`" — a sentence that had eaten the thing
it existed to warn about.
Both restored, and the note rewritten to say what is now true: capability means
one thing of ours, and three that belong to somebody else's vocabulary.
Verified live after restart: self and admin permission endpoints 200, gated
route 200, agent-status 200, 9 grants intact with 6 permissions offered.
tsgo clean, 797 tests, 787 pass, same 7.
The rename is done. Four steps, no data lost, no client break that survived
the step it was introduced in.
docs/offscale-plugin.md becomes plugins/offscale/PLUGIN.md. most of it is about
the plugin system generally rather than about offscale, which is exactly why it
belongs with the worked example — the reasoning is most useful beside the code
it produced, and the platform's docs should not carry the history of something
it no longer knows exists.
every reference updated.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
offscale had no dock tile: that field comes from the app store's catalogue, so a
plugin installed through the plugin system was reachable only by typing its url
or following the link on /plugins.
built at runtime rather than baked into the generated bundle, deliberately —
WHO sees a tile is a permission question, and a grant takes effect on the next
request rather than the next build. presentation comes from the manifest and the
route from mountPrefix, so there is one source for both, and is the
plugin's first permission so the endpoint can filter a tile out for an account
that cannot reach the screen.
two sources for tiles today, because the app store still has its own catalogue.
one when it is rebuilt on this.
the doc now records the system as complete rather than half-stubbed, including
the three bugs the extraction found — the sidecar-before-mount ordering, the
missing tailwind, and the build that could delete its own shell — and what is
genuinely still open.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
an exclusionary rule, made structural. a plugin does not render a screen: it
contributes panels and says how they are arranged, and the shell renders
WorkspaceView around them.
web/panels.ts appRegistryMetas — at least one panel
web/layout.ts defaultLayout — how they are arranged
both required the moment web/ exists, and missing either is refused at discovery
by name and with the reason. tested:
probeplug: has a web/ directory but is missing web/layout.ts.
Every plugin route renders a Workspace: contribute panels and a layout,
not a screen.
there is deliberately no way to export a component. one that could would be free
to render a bare div, a full-page form, or its own navigation, and the platform
would become a shell hosting strangers' layouts rather than one application.
non-compliance is not so much refused as unrepresentable — there is nowhere to
put a screen.
the shell registers <prefix> and <prefix>/:section, exactly as the core screens
do, so a plugin's sections stay addressable and cmd-clickable, and panels read
useParams independently rather than passing state between themselves.
appTypes.allowed is pinned to that plugin's own keys, so a persisted layout
naming something else falls back instead of rendering another plugin's panel
inside this screen.
the example plugin is rebuilt to model it — two panels, a layout, one of them
calling its own /api/example/ping through useClient — because the reference
implementation is what everyone copies.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the plugin system works end to end for a plugin with an api/router.ts, at
runtime, with no restart. what is wired, what is not (schema push, the sidecar's
pm2 entry, websocket providers, totality across plugin routes), and what was
deliberately left: offscale is not extracted, because moving it deletes working
code across ~50 files and that wants someone watching.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
offscale is deliberately the simple case — one shared resource, read or write.
music is the next extraction and the right place to build the in-plugin
visibility system, because it has real per-user data (favourites, playlists,
now-playing) on top of a real shared one (a single global library index). so
'whose is this row' has a non-uniform answer there, where offscale's is just
'the owner's'.
not designed yet and deliberately not designed here. recorded so the intent
survives the gap.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the platform's contract is what it already has: a role holds read or write on a
capability, stored in role_capabilities and enforced by the gate. anything
beyond — who sees whose rows, per-user isolation, record ownership, visibility
of any kind — is the plugin author's job, inside the plugin. the platform should
not grow machinery for it. a plugin knows what its data means; the platform only
knows whether this account got through the door.
offscale v1 uses that exactly. one shared resource: read sees what the owner
sees, write can change it including deleting a server the owner registered. that
is dangerous on purpose — the stored credential is a headscale admin key with no
read-only equivalent, so write is close to full control of the tailnet, and that
is the owner's call. expected use is read for most roles.
two consequences, both inside the plugin. the queries stop scoping by the caller
and resolve to the owner's id, leaving the per-user shape in the table unused as
the seam if isolation is ever wanted. and two POSTs are really reads —
/ssh-test probes and /policy/assist explicitly never saves — so they need
readOnlyWrites, or a read-level account finds a broken feature where a withheld
permission should be.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
offscale needs none of the install fields the catalogue carries — no modes, no
existingFields, no configFields, no composeTemplate, no members. nothing to
provision, nothing to point at. install is put the code there, push the schema,
start the sidecar, swap the routes, and it is available.
configuration happens afterwards inside the app, which is already how headscale
works: a server is registered at runtime and lands in offscale_servers.
so no-questions is the default rather than offscale's special case, and the
prompting machinery gets designed against the first extracted plugin that
actually needs docker or a remote instance. part of why this was the right
pilot — it exercises mounting, schema and sidecar without install being a
variable at the same time.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the last gap. six ws providers live in bun's route table rather than hono's, so
the app swap does not reach them — but server.reload({routes}) does, and in both
directions: refused before, connected after install, refused again after
uninstall, with core routes untouched throughout.
so a plugin can own a socket from the start, and no part of an install needs the
process restarted.
still untested: whether connections already open across a reload survive it.
that matters before an install is allowed to interrupt somebody's terminal.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
this went round twice — runtime dynamic, then generated-plus-restart on the
belief that hono could not mount after serving, then back once that was actually
tested. the doc keeps the route rather than just the destination.
tested: SmartRouter (hono's default) and RegExpRouter both throw 'Can not add a
route since the matcher is already built'. TrieRouter and PatternRouter accept
it. so runtime adding is possible but costs the fast matcher, and hono has no
remove-route api at all, which uninstall needs.
what solves both is not adding routes but rebuilding: construct a fresh app from
the current plugin set and reassign the variable. the fetch closure reads it per
request, so the reassignment is the swap — atomic, no dropped connections, no
server.reload, and the default SmartRouter is kept. verified 404 before install,
200 after, 404 again after uninstall, with core routes unaffected throughout.
the mechanical cost is one line: server.tsx:322 is '/api/*': honoServer.fetch, a
bound method evaluated once at serve(), and has to become a closure or the swap
does nothing.
websockets stay open: six providers live in bun's route table rather than
hono's, so a plugin owning a socket needs server.reload({routes}), untested.
offscale has none.
and totality stops being a boot check — buildApp() is now the single place
routes are mounted, so it is where the assertion belongs, refusing the swap
rather than refusing the boot.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
this reverses the earlier decision for C (mount and unmount at runtime) and says
so rather than quietly overwriting it.
the requirement behind C was that the platform must not need to know a plugin in
advance. that is met either way: what it reads is a generated file listing the
installed routers, analogous to Plugins.tsx on the frontend — nothing hardcoded,
nothing read from a table at boot, the imports made concrete at install. C would
have bought only the absence of a restart.
and a restart is close to free here, because sidecars are pm2 peers rather than
children — a property that was fought for, since officer used to spawn the agent
and pm2's tree-kill took the owner's chat down on every restart. what a restart
costs is websockets, which reconnect, and in-memory session records, which
claude:list already recovers.
the happy consequence is that assertCapabilityTotality stays a boot check
instead of becoming a per-mount transaction. it does need to be fed the route
table rather than Object.keys(handlers) first — generated mounts widen that gap
rather than closing it, so that is a prerequisite and not a tidy-up beside it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
lean it to identity facts and human choices: publisher, version, platform range,
the four presentation fields, and permissions. everything structural becomes
convention, where presence is the declaration — sidecar/, api/router.ts,
db/schema.ts, web/Router.tsx, web/panels.ts. appName comes from the directory
name, so the id cannot disagree with where the code sits.
the dock tile and page title needed no fields at all: the tile is label + icon +
color + mountPrefix, and the title is label. writing them again was duplication
that could only drift.
runtime is the file extension. index.mjs is node, index.ts is bun — implicit,
but already the rule here, since officer-pty runs under node for node-pty's abi
and everything else is bun. better than a field that can contradict the file.
dependsOn is gone; nothing read it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
'capabilities' already means three things in this codebase — the permission
registry, the officer-items store, and the sidecar's routing keys. a fourth
would be one too many, and the field is really just permissions. the name is
free: the old permissions table went in 044aacf4.
and the kind enum is gone with it. the first draft handed a plugin the
platform's own five-value CapabilityKind and then forbade three of them. those
five exist because the platform has five sorts of surface; a plugin has two —
grantable to members, or owner-only. a boolean says it, and says it without
needing a prohibition: a plugin cannot claim core if core is not a word it can
say.
offscale is ownerOnly: true, which is what kind: 'admin' meant.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
written against the real plugin rather than invented as a field list, on the
theory that an abstract one includes what nothing needs and misses what is
awkward. that paid off on the first field that mattered.
the rule here said a plugin may declare `app` and nothing else. offscale's
capability is `admin` — owner only — and should stay that way, so the rule was
wrong. the distinction is direction, not privilege: `core` means every account
and not deniable, so claiming it grants yourself to everyone; `admin` means
owner only, which is a plugin restricting itself. corrected table in the doc.
core, execution and confined stay the platform's to assign.
`publisher` is the only input to the mount prefix, through one function, so
first-party and third-party cannot drift into two code paths.
sidecar.runtime is a field because officer-pty needs node for node-pty's abi
while everything else is bun — one plugin already needs it, so not speculative.
dependsOn is informational and unenforced. code dependencies need no declaration
now that a plugin builds inside the workspace, and service dependencies already
degrade; this exists so the store can say the console section wants the terminal
plugin, rather than the section silently doing nothing.
health is marked deferred rather than open, with the reasoning, so it does not
get re-raised. migrations likewise.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
everything moves to the plugin, frontend included, so federation stopped being
a later problem and had to be answered. it is answered by not needing it: bun
builds the spa into build/ at start and rebuilds it on install, serving from
that directory instead of compiling through the html import. Bun.build is a
runtime call, so an install needs no restart — just a refresh. same origin
throughout, which is why there is no cors work and no rewrite of useClient.
App.tsx keeps core routes and gains one map over `plugins`, each mounted at a
wildcard delegating to the plugin's own router. that list comes from a generated
Plugins.tsx, because a bundler cannot follow import(runtimeString) — the
specifier has to be concrete before the build. the six places the shell
currently hardcodes headscale collapse into that one file, dock included; the
runtime dockItemsFromPlugins path follows rather than competing with it.
presentation moves to build time, permission stays runtime.
dependencies turned out to be two different problems wearing one word. a service
dependency (assist → anthropic-proxy) is a wire call and already degrades. a
code dependency (ConsoleView → TerminalView) is in the bundle and cannot. rule:
may depend, must degrade. service calls go through the api carrying the user's
token, with the user's own permissions, which also deletes the state-file read
claude-proxy uses today to lift the proxy's secret.
no per-plugin permission list: a plugin is part of the app and bounded by the
account calling it. that makes marketplace review a security boundary rather
than a naming one, which is worth knowing rather than discovering.
and the developer environment is a platform checkout — clone it, run dev, build
the plugin inside. the 13 workspace packages resolve by name because bun links
them, so `import { useClient } from 'hooks/useClient'` just works with no
registry and no versioning. dev-time and build-time become the same mechanism.
also writes down the headscale inventory now that it has been read end to end,
including that assist.ts travels unwired as a marker and must not be tidied away
as dead code.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
the name looks like branding on someone else's project, which is exactly how it
gets 'corrected' back later. it is not: offscale is the stock headscale server
plus the companion that ships beside it, and the invite flow is the first thing
that only exists there.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
closes the one open item left by deleting /api/vpn. removing the vpn capability
leaves no member-grantable headscale surface and that is correct: the owner
mints an invite from the headscale app, the companion turns it into the redirect
the phone claims, and the device joins. no per-member permission on officer is
involved at any step.
recorded as decided rather than open so nobody reintroduces a member-facing
enrolment route believing something was lost. nothing was — /api/vpn/enroll was
the design the invite flow replaced, and it never had a UI.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
it had no caller. verified three ways before removing: nothing in the mobile
monorepo reaches it (enrollVpn's only call site is behind `if (embedded)`, and
the one app rendering VpnScreen never passes embedded), nothing in the officer
web app references it, and the live database holds no vpn grants. the companion
was checked separately by its own author — zero references there either.
and it will not come back. offscale is permanently standalone: the thing that
gets you to the platform cannot itself need the platform, or a broken tailnet
locks you out of both.
gone: api/vpn/router.ts, its mount, and the `vpn` capability. the registry keeps
a comment where the capability was, because its removal has a cost worth
recording — headscale is admin-only, so no member-grantable headscale surface
remains, and reintroducing one is a deliberate act rather than an oversight.
kept: the sidecar's enroll.ts. its bare POST /_officer/enroll handler is now
unreachable, but the file is also the dispatcher for /enroll/invites, which is
live and fundamental. the header comment now says so, so nobody deletes it
looking for dead code.
also records the third component in the doc. two of the three have an "enroll"
surface and only one is ours: /api/v1/enroll/* belongs to the companion, is
where the phone actually goes, and must not be collapsed into /api/offscale/*.
capabilities tests: 17 pass / 8 fail both before and after, stash-verified — the
8 are pre-existing, in totality and path-to-capability, which is precisely the
machinery dynamic mounting will rework.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
first step of extracting headscale into a plugin. CORE_PROCESSES is five now,
and the catalogue.test CORE[] mirror follows it — not optional, since that list
asserts "the catalogue must not offer a core process" and would have blocked
adding offscale to the catalogue later.
the local generated ecosystem file lost its entry too, and officer-headscale was
stopped and deleted from pm2 by hand. the platform still mounts /api/headscale
and still declares the headscale and vpn capabilities, so the feature is
present-but-unavailable rather than gone.
docs/offscale-plugin.md is a live document for the rest of it. what it records
that nothing else does: core is now `officer` alone and everything else is a
plugin; routes are /api/<app-name> for ours and /api/p/<creator>/<app-name> for
third parties, derived by one function so the two can never become two systems;
tables stay in public with an app-name prefix; mounting becomes genuinely
dynamic, which retires the "every route stays mounted" premise and relocates
assertCapabilityTotality from a boot check to a per-mount transaction.
it also records a rejected experiment with evidence — a postgres schema per
plugin works completely, including cross-schema FK, idempotent push and
DROP SCHEMA CASCADE as uninstall — and the reason not to: drizzle-kit 0.31.8
needs schemaFilter naming every schema, contradicting its own docs, and without
it push reports "No changes detected" and creates nothing. a plugin install that
reports success and makes no tables is the exact failure shape we have hit three
times this week.
and /api/vpn is dead: no caller in the mobile monorepo, none in the web app, no
grants in the database. offscale is permanently standalone, so it never comes
back. the invite flow is unaffected — the phone claims from the Companion, not
from officer.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
42 documents, 13,000 lines, and no way to tell from a filename which describe the
system as it is and which record an afternoon in July. This sorts them: living,
stale, historical, and two clusters that want consolidating.
Says plainly how much was verified — mostly filenames, status lines and greps for
what changed today — so it reads as a starting point rather than a verdict.
Names the two obvious consolidations without performing them. Nine opencode
documents for one migration that has landed (verified: `opencode serve` is in the
sidecar, so the plan's "nothing here is implemented" is false), and three
mobile-dav documents that are one correspondence. Both need all of them read
first, which is not a 4am job.
Marks the historical ones as not-to-be-rewritten. claude-sidecar-isolation.md
records the officer-claude to officer-agent rename that preceded tonight's rename
to officer-claude-code; editing it to match today's code would destroy the
reasoning it exists to hold.
And notes what most of them share: they were written when the estate was twenty
processes and everything was simply present. A core install is six. The fix is
usually one line — say whether the thing is core or a plugin — not a rewrite.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
It described four capability kinds and said terminal, chat and files can never be
granted. There are five, and those three moved to `confined` on 2026-08-11 — the
kernel enforces the boundary because the account has its own Linux user, and a
grant means nothing without one.
The layout diagram was missing dockers/ and secrets/, and implied the paths are
configured. They are derived from the working directory, which is why the pm2 cwd
pin and assertInstallLayout exist.
Adds what is switched off as of tonight: six core processes, every plugin router
commented out beside its capability claim, the ecosystem files now generated, and
.env down to three values with the keys in the secret store.
First of a documentation sweep. 42 docs; this one first because it is the
operational guide somebody actually reaches for.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Ubuntu 24.04, Debian 12, Arch and Fedora 41. OS and package-manager detection
correct on all four; --help works unprivileged; and the install report is written
end to end in a container that had never seen this code — task 1's mechanism
confirmed off the machine it was written on.
Three real findings.
--only does not isolate a step. Running --only "Core utils" still created a user
account, because ask_username and the account creation sit in the preamble above
the step framework, so everything before the first `step` runs every time. It is
defensible and it is not what the flag appears to promise.
.setup-answers travels with a copy of the tree. Correctly gitignored and 0600,
but it lives inside the repository directory, so `cp -r` carries it — a container
that had never run setup came up already knowing the username and created that
account. Nothing secret in it; it is a surprise, which in an installer is the
expensive kind.
adduser leaks its own interactive prompt ("Try again? [y/N]") on the
account-creation path. Harmless here because the run had already stopped, but a
hang on a real unattended install.
Also records what containers cannot reach: no init means systemd, netplan, ufw
and the sshd drop-ins are only verifiable as "wrote the right file"; Docker and
Postgres are untested; macOS is unreachable entirely and everything about it is
reasoned rather than executed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Enumerates the forty-seven prompts the two scripts actually ask and sorts them
into branches, consents and values — the distinction that decides what a
generated leaf can remove. Four real branches (OS, role, tailnet state, which
half), fourteen consents that let a leaf omit a section entirely, and a set of
values that must stay prompts because baking them in would mean publishing
somebody's hostname.
Names the two things that need deciding rather than deciding them:
Whether a leaf strips dead code or sets constants and calls the base. They are
different artifacts and the plan rests on which one is meant — the first is what
makes it auditable by being short, the second is what keeps it maintainable.
And the combinatorics: 4 OS x 3 roles x 3 tailnet states is 36 leaves before
consents, so the tree cannot be the full product. Publishing a few opinionated
leaves keeps the static-file-anyone-can-diff property; generating on demand does
not, which is the property per-leaf scripts existed for.
Also notes that --unattended and a generated leaf are the same mechanism seen
twice, and that install_config's existing behaviour — keep the user's file when
there is no tty — is the conservatism every unattended answer needs.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
navigator.clipboard is secure-context only, like crypto.randomUUID before it —
over plain http on a tailnet address the object does not exist. Twenty call
sites across eighteen files, in three states that all looked fine in review:
bare calls that threw and killed the handler, optional-chained calls that
silently did nothing, and one carrying the comment "Officer is always behind
HTTPS", which it is not.
The optional-chained ones are the worst of the three: a copy button that reports
success and copies nothing is indistinguishable from a working one until someone
pastes.
helpers/clipboard.ts falls back to document.execCommand('copy') over an
off-screen textarea — deprecated, and it works on any origin because it predates
the secure-context rule. Off-screen rather than hidden, because display:none and
visibility:hidden elements cannot be selected and the copy fails silently.
Reading the clipboard has no equivalent: execCommand('paste') was never permitted
from script. The file browser's paste-a-file path now checks canReadClipboard()
and explains itself instead of throwing.
docs/http-secure-context-audit.md is the full sweep the owner asked for: what was
fixed, what cannot be, and what was checked and found clear. crypto.subtle is
used nowhere in the frontend, which was the one worth confirming since it has no
cheap fallback. Notification's six matches are type names, not the API.
geolocation and navigator.share are already guarded. getUserMedia is in four
files and is being removed — but QrTransfer uses it for the CAMERA, not a
microphone, so "remove audio" does not cover it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The old name said nothing about what the process runs, and it sits directly
beside officer-anthropic-proxy — a different process doing a different job — so
"the agent" was ambiguous exactly where it mattered. CLAUDE.md already had to
spend a paragraph insisting the two are not the same thing. It spawns `claude`;
the name says so now.
Only two references were functional: the generator's CORE_PROCESSES and the CORE
list in catalogue.test.ts. Everything else was prose or comments.
Left alone deliberately: `x-officer-agent-token`. It looks like the same string
and is not — it is the agent-handoff HTTP header, naming a per-panel bearer
token, unrelated to any pm2 process. Renaming it would have changed a wire
protocol to tidy a label.
Historical docs keep the old name. claude-sidecar-isolation.md and
open-threads-after-per-user-claude.md are dated investigations that record the
PREVIOUS rename, from officer-claude to officer-agent, and rewriting them would
make that history unreadable. CLAUDE.md notes the change instead, where somebody
reading those will be looking.
Also worth recording, from the owner: merging this with officer-anthropic-proxy
into one sidecar was investigated tonight and rejected. They stay separate.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
src/databases/officer_db/src/<feature>/{schema.ts,queries.ts}, replacing the
parallel schema/ and queries/ trees. 24 feature directories, 46 files moved with
git mv so history follows.
The parallel trees had drifted, which is what the restructure is really fixing:
four features were named differently on each side — app-store/sidecar-installs,
email/email-accounts, server/server-config
operations had a schema and NO query file: its task_logs is reached directly
from src/servers/api/task-logger.ts, bypassing this package's own boundary
integrations had queries and NO schema, because it spans two features'
tables — server_integrations and user_integrations
Both lopsided cases survive as directories holding one file, which states the
problem instead of hiding it across two trees.
Nothing outside the package changed how it imports. `officerdb`, `officerdb/types`
and `officerdb/db` resolve exactly as before; index.ts absorbed the path changes.
Added `"./*": "./src/*"` so the new layout is reachable — `officerdb/soulseek/schema`
— which one script needed, because soulseek is a plugin and therefore commented
out of the aggregator.
schema/index.ts became src/schema.ts, keeping the core/plugin split from earlier
tonight. drizzle.config.ts and the package's "./schema" export follow it.
Verified rather than assumed: all 52 files in the package parse, every relative
import resolves against the new layout (checked by walking each specifier to a
real file, since parsing does not check paths), and everything in the tree
importing officerdb still parses. Not typechecked — empty node_modules, frozen
installs.
One rewrite bug worth recording: the rule mapping a query module's sibling import
also matched the './schema' this pass had just written, turning it into
'../schema/queries' in 22 files. Caught by the resolver check, not by parsing —
both spellings parse fine.
Also corrects every path reference the move invalidated: src/databases/CLAUDE.md's
layout diagram, the root CLAUDE.md data section, three docs, and seven sidecar
comments naming queries/<x>.ts.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
.env now holds PORT and POSTGRES_URL. Every encryption and signing key lives in
$OFFICER_ROOT/secrets/officer-keys.db — 0600, 0700 directory, owned by the
service user, created on first use.
The design doc planned to move ONE at-rest key into the store. What shipped
splits it: headscale, wallet, photos, jellyfin, invoiceshelf, vault and
service-connections each get their own, plus jwt. VAULT_STORE_KEY encrypted all
seven, so one leak opened all of them — and it was named after whichever plugin
needed it first, which is why it read as safe to change if you did not run a
vault. A core install bootstraps two, jwt and headscale; the rest appear when
their plugin first asks.
The file IS the secret. No second key unlocks it, because a key beside the store
it opens buys nothing. The gain was never secrecy, it is blast radius: bun
auto-loads .env into all twenty pm2 processes, so a key there is readable from
/proc/<pid>/environ of twenty processes — officer-music held the key that
decrypts wallet seed envelopes.
Two defects found by testing the store rather than reading it, both of which
would have shipped:
The WAL was 0644. Enabling WAL creates -wal and -shm at 0644 rather than
inheriting the database's mode, and a freshly written key lives in the WAL
before checkpoint — so the 0600 on the database was decorative. The 0700
directory covered it, but only until someone loosened the directory.
PRAGMA journal_mode = WAL takes an exclusive lock, and busy_timeout was set
AFTER it. With twelve concurrent openers, six died on that line with
SQLITE_BUSY. Every sidecar opens this store at boot, so they open it
simultaneously by definition: most of them would have failed to start on a cold
boot and none on a warm one. Fixed by ordering the pragmas; re-tested with
twelve racing processes, one key, one row.
crypto.ts takes a purpose as its first argument now, which the design doc had
explicitly promised would not happen — 32 call sites across seven query modules.
That promise is corrected in the doc rather than quietly dropped.
Also live, not just comments: wallet/upstream.ts gated wallet storage on
process.env.VAULT_STORE_KEY and would have reported "unconfigured" forever. It
asks the store now, and the question it answers changed — not "did somebody set a
variable" but "can this process open the store", since the key is created on
demand.
assertSecretsClosed covers the store, its directory and its WAL. The jwt key
mints owner tokens, so a member's shell reading it is strictly worse than the
.env leak that check was written for.
Not typechecked: node_modules is empty and installs are frozen, so the
officerdb/secret-store subpath could not be resolved at runtime here — verified
that officerdb/types fails identically, so it is the empty tree and not the new
export. The store module itself was tested directly: creation, idempotence across
processes, hasKey not creating, permissions, and the twelve-way race. Every
changed file parses; the setup section runs and degrades correctly when the
import is unavailable.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
ALLOW_ANY_ORIGIN, ALLOW_ANY_ORIGIN_MUSIC, and everything they gated. The flag
defaulted to ON, so none of it ran on a real install — what comes out is
documented defence in depth that was already switched off. The file said so
itself: "Both flags and their call sites come out once the tailnet is the
perimeter."
Origin was never authentication here in any case. An app's `officer://<hex>`
origin is chosen by the client, forgeable outside a browser, and extractable from
a shipped binary.
Gone: the two flags, isOriginAllowed, isOriginCheckDisabled, isMusicOriginExempt,
originValidationMiddleware, ORIGIN_RULES and the whole OFFICER_<APP>_ORIGIN
scheme, PUBLIC_URL's origin/host derivation, and origin-validation.test.ts, which
existed only to pin them. CORS now echoes whatever Origin it is given, which is
what every install already did.
What SURVIVES is the reason this needed care. origin-validation.ts held two
unrelated things, and the second was the global authorization gate — a valid
non-owner token reaches only what its role grants, deliberately NOT under the
flag because it is account-based rather than origin-based. Its own comment called
it "the airtight half". Deleting the file wholesale would have deleted
authorization.
So it moves to _middlewares/capability-gate.ts as capabilityGateMiddleware, with
the name matching what it does: nothing in it reads an Origin header any more.
hono.ts mounts it in the same position, ahead of every router.
origin-middleware.ts stays and is untouched — it extracts the Origin for six auth
handlers that log it, and for passkeys. Extraction, not validation.
Also updates every claim that rested on the old model: CLAUDE.md's security
section and repo map, docs/secret-store.md, docs/mobile-api-keys.md, and five
messages in machine-setup's Tailscale section which told the owner to set
ALLOW_ANY_ORIGIN=false when declining a tailnet. That advice is now impossible to
follow, and the honest version is different: with no tailnet the token is the
whole lock, so put a proxy in front and restrict who can reach it.
Not typechecked (empty node_modules, frozen installs). Every changed file parses;
the setup section was run and writes four variables now.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
OFFICER_OS_USERS is gone. The platform behaves as it always would have with the
flag on, and there is nothing to enable.
Six conditionals, five of which were dead weight — provisionOsAccount,
deprovisionOsAccount and the create/delete paths each opened with an early
"not enabled on this server" return, and the API told the frontend whether to
render the Linux controls at all. Those go, along with the 'disabled'
DeprovisionResult stage, which nothing can produce now.
The sixth is the one with teeth. assertSecretsClosed opened with
`if (!OS_USERS_ENABLED) return`, described in its own comment as "a no-op when
the feature is off, so an existing install is unaffected until the owner opts
in". It is now unconditional: the server refuses to boot while any .env in the
project root is group- or world-readable. A member's shell reading .env and
printing JWT_SECRET was confirmed exploitable when this check was written, and a
prerequisite that only holds when somebody remembers to set a variable is not a
prerequisite.
Nothing to remove on the environment side — the flag was never in .env.example
or in the setup script.
Not typechecked: node_modules is empty in this tree and installs are frozen, so
tsgo could not run. All six files parse under `bun build --no-bundle`, and the
changes are deletions of dead branches plus one removed early return. Formatted
with prettier 3.9.6 via bunx rather than the pinned resolution, for the same
reason; its one unrelated reformat was reverted by hand.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Third value for the store, agreed in conversation. Neither an encryption nor a
signing key — a bearer credential the agent presents to the proxy on localhost —
but it qualifies on the same properties: generated once, shared between two core
processes, fatal to regenerate silently.
It makes the case better than the other two, because it is not in .env. It is in
$DATA_PATH/sidecar/claude-state.json, which is the exact location decision 3
rules out by name: DATA_PATH is what gets backed up.
Also records the rename. ANTHROPIC_API_KEY is wrong in both halves — not
Anthropic's, not an API key; Anthropic's real credential is the OAuth token in
~/.claude/.credentials.json that the proxy swaps this one for. It is
anthropic-proxy-secret everywhere we control, and keeps the CLI's name only on
the assignment `claude` itself reads.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Written from the conversation of 2026-08-12. Nothing implemented; every claim
about the current tree was checked on that date.
The short version: secrets stay in Postgres, the keys move out of .env into a
small SQLite store, and rotation becomes an operation instead of data loss.
What is actually wrong today is narrower than "secrets in .env" and worth stating
precisely, because the separation that exists is correct and must survive a
refactor: secrets live in Postgres and the key that opens them does not. The
problem is blast radius across processes — Bun auto-loads .env, so
VAULT_STORE_KEY sits in the environment of all twenty pm2 processes, and
officer-music holds the key that decrypts wallet seed envelopes for no reason.
The document records the decisions and, more usefully, what was ruled out:
Keys cannot go in Postgres. A dump would carry the ciphertext and the thing
that opens it. Encrypting the key with a second key only moves the question —
one secret has to be readable without any other, and the only decision is where
it lives.
The store is not encrypted at rest, and this was tested rather than assumed:
stock SQLite silently ignores unknown pragmas, so `PRAGMA key` succeeds,
encrypts nothing, and the value is readable with `strings`. bun:sqlite ships
stock SQLite 3.53.0. Whole-file encryption needs SQLCipher, which is a second
native dependency, and this project already knows what one of those costs.
SQLite rather than a flat file for rotation, not secrecy: rotation needs key
VERSIONS, since an interrupted rotation needs the old key and the new one to
both exist.
The file must not live in $OFFICER_ROOT/data/ — that is what people back up,
and a key store in the same tarball as a database dump rebuilds the problem.
It also records the core/plugin split the design assumes: light plus
officer-headscale is the core, because CLAUDE.md rests the security model on the
tailnet and a model that rests on the tailnet cannot treat administering it as
optional. Vaultwarden and the wallet become plugins. Moving headscale into light
removes it from the app store automatically, since catalogue.test.ts asserts the
catalogue equals full minus light.
Five open questions are left open rather than guessed at.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Expands the communications section from a list of what worked into the actual convention: the
directory's lifetime and the rule that anything durable must move to docs/ before the merge;
numbering, parity as attribution, non-consecutive numbers; slugs; reply-in-a-new-file and the
one case where editing your own is right; referring to commits by sha because three remotes
carried the same branch names.
Records what a handoff must contain, with the verified/assumed split named as the rule that
carried the most weight — a handoff confident about something untested is worse than none,
because the reader builds on it. Adds a skeleton to copy.
Documents termination as the four attempts it actually took, ending at the only checkable
version: the exchange pauses when no open item is actionable by a participant. Adds the third
state, deferred-with-a-reason, since a two-state protocol forces an agent to lie in one
direction. Notes that a stall must be detectable because the human spotted both before either
agent did.
Adds a review-discipline section — check the enforcement rather than the description, run it
against a real machine, a check never seen failing is not evidence, distrust vacuous passes,
expect stacked bugs, distrust "inert today", and look at which way unknown resolves. Adds a
failure-mode table to pattern-match against, and the git hygiene that bit us, including
merge-verify-then-delete, which I got wrong.
Closes with session economics, an ordered list of what to build, and the one thing not to
automate: agents may coordinate on what is true and must not decide what is permitted.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The owner could not convey this to the second agent, who launched it differently and got
something that looked identical and did not work. The script was never the hard part; the
mechanism is.
States the requirement so it survives a different harness — a detached shell process owned by
the agent's harness, which exits when it has something to say, and whose exit re-invokes the
agent — and notes that dropping any one of those three breaks it invisibly.
Then the four wrong ways, each of which looks correct while running. Backgrounding with nohup
or & produces a process that polls correctly, detects the push, exits, and never tells the
agent, because the harness is not tracking it; I made that exact mistake and caught it only by
re-reading my own command. A model-driven interval is functionally correct and pays a full
context re-read per tick to learn nothing — the intuitive design, and the expensive one, which
is why it is the first thing to warn a new agent about. A loop that does not exit on detection
has no path to the agent at all. And per-tick logging is deferred cost that lands all at once
on wake.
Also records why 30s polling is free in a shell and ruinous in the model, including the
five-minute prompt-cache TTL that makes any model-side wake beyond it pay for a full uncached
read.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
pastilhasandClaude Opus 5