platform/docs/command-hq.md

Command HQ — design record

Spec conversation 2026-08-18, owner + agent, back and forth. This is the write-up asked for partway through 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). 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. Not yet exercised against a running server — typechecks and the suite is at baseline, but nobody has clicked through it live. 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 and disarm all go through Command HQ's own PATCH route, never /chat's — the two are scoped to different directories. 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

Base path /command-hq, core permission (every signed-in account, no grant to hold), every query scoped to the caller's own user_id regardless. Router: src/servers/api/command-hq/router.ts — read it directly 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
DELETE	/categories/:id	—	uncategorises its commanders, does not delete them
GET	/tags	—	schema exists; no UI planned for v1
POST	/tags	{name}
DELETE	/tags/:id	—
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?, watcherArmed?}	name proxies to renameClaudeSession; watcherArmed: false disarms (arming has no route — needs a live task id)
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 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 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 why that is the design rather than an oversight.

Two creation paths, one underlying mechanism. The web composer sends an ordinary /chat websocket turn in the commanders directory and never calls POST /commanders/new — the row is written server-side the moment a real transcript id exists. A client that cannot open that websocket with the right parameters — mobile — calls /commanders/new instead, which drives the identical spawn path (startClaudeTurn in websocket.ts, extracted so there is exactly one) with no live socket, and waits for the same transcript id before answering. POST /commanders (bare) is neither of these — it is the low-level upsert both paths ultimately rely on, not something a client calls to create anything new. See "How a commander is created".

What it is

A home for commanders — long-running background jobs (watchers) the owner wants to find again, organise, and check on. Deliberately separate from /chat: no cross-pollination. A commander is started from inside Command HQ, not from an existing /chat conversation, and a /chat conversation never becomes a commander after the fact.

Worked example the owner gave: a watcher that polls Gmail via IMAP every 10–60 minutes, compiles what came in, and pushes a summary — so the owner can uninstall the Gmail app and stop getting a notification per email. Command HQ is where that watcher lives, is named, is categorised, and is checked on.

Route, surface, placement

• /command-hq — new dock item, new route, core (kind: 'core' in permissions/registry.ts, not app). Every account gets it; there is no grant to hold. Core still needs the registry entry — assertPermissionTotality refuses to boot on an unregistered router regardless of kind.
• Per-user, not shared. Every table carries user_id; every query is scoped to the caller. Nothing about this is visible across accounts. The owner named a possible future (a super-admin view of who is running what) — do not build toward it, it costs nothing later because user_id is already there.
• Workspace/Panel framework, per the standing rule for every new dashboard route: <WorkspaceView workspace={ws} locked />, panels under officerdev/src/apps/CommandHQ/, a rule added to usePageTitle.ts.
• API-first. /command-hq will also be a mobile app surface eventually — not this task's concern, but every capability must be a real HTTP endpoint, not something only the web panel can reach.

Layout, v1

Two columns.

• Left: category list. Simple for v1 — a flat list of the user's categories, click to select. Filtering and search on this list are explicitly deferred (see "Later").
• Right: commander grid, scoped to the selected category. One card per commander.
• Click a card → detail pane. The right side splits to show it. What the detail pane contains beyond "the conversation" is deferred — see "Later". For v1 it can be as simple as the chat transcript.
• New commander starts from inside Command HQ's own composer — not a repurposed /chat panel reached from elsewhere. Almost certainly EmbeddableChat (officerdev/src/apps/Chat/EmbeddableChat/), which already exists as a chat surface embeddable outside the /chat screen.

How a commander is created

Resolved and built, both paths — there is no separate "file this into Command HQ" step on either one. Starting a new commander is the creation flow.

Web:

1. The composer sends an ordinary /chat websocket turn with context: 'command-hq'.
2. resolveChatCwd (websocket.ts) reads that context and resolves getCommandersDir(email) itself — msg.cwd is never consulted on this branch. The client does not choose, or even need to know, where commanders live.
3. The moment the session has a real transcript id, createMessageHandler writes the command_hq_sessions row — the same session:claude-then-result seed chat_session_models' record() already waits on, with a sibling closure (fileAsCommander) hung off the same two triggers, gated on a commandHq flag threaded in from msg.context === 'command-hq' at the call site.

Mobile needed a different door. COMMS/command-hq/mobile.md — read by the platform side 2026-08-18, after the mobile client had already been built against the router — found the web mechanism does not transfer: mobile sends chat turns with its own parameter set, so a context flag would have to be carried correctly by two independent clients, and a client that gets it wrong silently creates an ordinary /chat conversation in the wrong directory rather than failing loudly. Decided with the owner, and now built:

POST /command-hq/commanders/new
  body:  { prompt: string, displayText?: string, categoryId?: number | null }
  ->     { sessionId, cwd }

prompt is required — confirmed with the owner 2026-08-18. The actual usage pattern it needs to support: a commander often starts as an ordinary back-and-forth (figuring out the terms of the job with the agent) long before a background job is ever armed via the existing dialog — sometimes the first message really is just "we've already discussed this, do it." Either way there is always a real first message, the same as any conversation; there is no "empty commander" case to design around.

displayText was missing from the first version of this route, and the mobile side caught it before it was exercised against a real server. A commander created to become a background job sends the FULL buildBackgroundJobPrompt mechanics as prompt — a page of preamble, the owner's actual sentence only at the end — and command_hq_sessions has no name column, so the title comes from session.meta.title = (displayText ?? prompt).slice(0, 100), exactly the precedence the websocket path already uses (both clients already send displayText there — sendWithHiddenPrefix on web, useSessionView.ts on mobile). Without it, every commander created through this endpoint would have titled itself with the mechanics text, and the grid would have been a column of identical, unreadable cards. Fixed same session, before anyone had run it for real.

How it actually gets the id. Claude assigns a transcript id only once a turn has run far enough to report system.init — there is no way to reserve one in advance. So the endpoint genuinely starts the turn: startClaudeTurn (websocket.ts), the turn-starting half of handleClaudeCodeChat extracted so both the websocket path and this endpoint drive ONE spawn function rather than two that drift, called with ws: null. It resolves as soon as the harness's session:claude arrives — not once the whole turn finishes — via an onOutcome callback that mirrors what a live socket would already have been told (a refusal reason, a spawn failure, or the transcript id), so an HTTP caller gets a real answer instead of a generic timeout. The turn then keeps streaming exactly like any other session; the endpoint's job ends the moment there is something real for the client to attach to.

The category race, and why it resolves deterministically anyway. fileAsCommander (the web path's same hook, which this endpoint's turn ALSO passes through since context: 'command-hq' is set either way) always files with no category — it has no way to know one was requested. So the endpoint does its own createCommandHqSession (idempotent, upsert-on-conflict-do-nothing — safe whichever of the two insert attempts actually lands first) followed by an unconditional setCommandHqSessionCategory UPDATE. The UPDATE always runs last and always wins, so the category the client asked for is what's left standing regardless of ordering.

No name is stored at creation, on either path — see "Naming", next.

Naming — resolved, and it changes the schema

/chat never persists a title to Postgres. renameClaudeSession (claude-sessions.ts:646) appends a summary entry straight into the transcript's own JSONL — Claude's own on-disk format is the title's only home. An unrenamed session's title is the first prompt, truncated, computed on the fly. There is no title column anywhere in Postgres for a /chat session, ever.

Confirmed with the owner: command_hq_sessions carries no name column. The grid reads a commander's title live, the exact same way /chat's list already does. Renaming a commander is renameClaudeSession — the same function, same route, same file on disk — not a second mechanism that could disagree with the first. One title, one place it can live, zero drift possible by construction.

Data model

Tables prefixed command_hq_ per the house rule (purge finds a feature's tables by prefix; this one keeps the platform's naming convention even though it lives in core, a deliberate choice made explicit rather than an accident of copying a plugin schema).

command_hq_sessions

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 the background-job tray already derives it — fold chat_session_events for the session, task:started 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. 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

User-created, not a fixed enum. One category per commander — a plain FK column on command_hq_sessions, not a join table.

column	notes
user_id
name
created_at

Tags — infrastructure now, UI not specified

Resolved: build the tag infrastructure alongside categories, not deferred. Many-to-many, unlike categories:

• command_hq_tags — user_id, name, created_at.
• command_hq_session_tags — join table, (session_row_id, tag_id).

Nothing in the spec describes a tag picker or tag filter chips for v1's UI (only the category list was described). Build the schema; treat the UI as open until asked for.

Working directory — every commander shares one directory per user

Resolved, and it came with an explicit reversal worth keeping: the owner's first instinct was a directory per commander, each with its own isolated cwd. Retracted on the spot — creating a fresh directory at the moment a commander starts races the harness's own spawn (the cwd has to exist before or exactly when the first turn fires, and centralising that creation moment is exactly the kind of timing hazard Claude Code's own directory-per-project convention was not designed to be driven by). Simpler and safer: no per-commander directory at all.

DATA_PATH/<email>/commanders/ — one fixed directory, shared by every commander a user has. Every commander's cwd is this same path. This is precedented, not new: getEmailAccountsDir in data-path.ts already follows the identical shape (join(DATA_PATH, ownerEmail, 'email_accounts')) for a different feature's per-user directory. A getCommandersDir(email) alongside it is the natural addition. Needs to exist once, idempotently (mkdirSync(..., { recursive: true }), the pattern already used in websocket.ts), created lazily before the first commander's first turn — never per-commander, which is the exact race that was rejected.

This is a deliberate departure from ordinary /chat cwd resolution for the owner. A normal owner chat runs in getOwnerHomeDir — the owner's real login home — while getHomeDir (DATA_PATH/<email>/home) is otherwise reserved for non-owner accounts and the pipeline executor. Commanders use the managed, DATA_PATH-rooted directory even for the owner, which is exactly what keeps them out of the owner's own project transcripts and out of /chat's history for the same underlying reason categories exist: a commander is not a project conversation, and should not be filed as if it were.

A member could not reach it, and this is now fixed — but only going forward

Found while wiring the creation flow, not guessed at: getCommandersDir is a sibling of home, and confineUserTree (os-user.ts) confines every sibling of home to 700, owned by the SERVICE user, "unreachable even by name" — correct for platform-written directories like attachments/, and exactly wrong here, because a member's commander turn runs as the MEMBER's own Linux account via setpriv. This is the identical shape of bug resolveChatCwd's own comment already tells a story about (general_chat_sessions, abandoned for the same reason) — a working directory the turn cannot cd into, surfacing as an agent that fails on its own first command.

Fixed in confineUserTree itself, not by abandoning the directory: it now chowns commanders/ to the member exactly as it already chowns home, and excludes it from the loop that confines every other sibling. No ACL, unlike home — the file browser needs one because it reads a member's home from inside the platform process as the service user; Command HQ has no equivalent, since it reads a member's commander transcripts through listTranscriptsAs, the same as-the-member read /chat already uses for anyone's transcripts. Plain ownership is the whole requirement, same as ~/.claude/projects itself.

The gap this leaves: confineUserTree runs at provisioning time. An account provisioned before this fix landed will not have commanders/ chowned until it is provisioned again — the existing admin reprovision action (provision-linux-route.ts / manage-users.ts) already re-runs confineUserTree idempotently, so no new machinery was needed, but a member who existed before tonight needs that one manual re-run before their first commander turn will actually spawn. Worth a release note, not a code change.

The /chat ↔ /command-hq relationship

One predicate, opposite sign, both directions read from the same join.

• /chat's session lists (listClaudeSessions, listAllClaudeSessions, and the ChatHistory app) exclude any session with a command_hq_sessions row.
• /command-hq's card grid includes only sessions with a row, joined against the live session list.

The source of truth is always the live session list (transcripts on disk, exactly what /chat reads today), never the table. command_hq_sessions is a filter and a set of annotations on top of that list, not an independent record of what sessions exist. Consequences the owner confirmed explicitly:

• A row whose transcript is gone (deleted, home reset, whatever) is simply absent from both lists — it is not rendered as a tombstone, and nothing needs to notice. It is "someone else's problem" by design; the owner is not developing for people who will be surprised by this.
• A purge button can sweep command_hq_sessions for rows whose session_id no longer resolves to a live session and delete them. Housekeeping only — correctness never depended on it running.
• 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.

Built on, mostly unchanged

Everything about how a background job runs was out of scope for the original build — a place to browse, name and categorise the sessions the existing system produces. That held until real testing found the mechanism itself wanting; see "The watcher loop" below for what changed and why. Otherwise, unmodified:

• The whole lifecycle: run_in_background/Monitor, task:started/task:notification, chat_session_events as the durable log, collectRunningTasks as the fold, the tray.
• Stop/pause/restart routes under /chat/tasks/:id.
• renameClaudeSession — Command HQ's rename action IS this, not a new one.
• officer-notify — assume it works and is wired up separately; do not build a parallel notification path for this feature.

The watcher loop — exit, act, drain, re-arm

COMMS/command-hq/watchers-what-broke.md and watcher-loop-proposal.md, both 2026-08-18: the mobile side actually ran a watcher (poll git ls-remote, react to each push) through the dialog, and the mechanism broke in ways no amount of reading the code would have found — two watchers alive at once because a new one was armed before the old was reaped, a watcher firing on the agent's own push to the branch it was watching, "keep going" silently meaning the agent would never see a firing again because a background process reaches the model only when it EXITS. The owner's call, confirmed 2026-08-18: implement the proposal as written, before the first real tests — this governs the agent's behaviour generally, not only Command HQ, so it is not optional groundwork the way most of this document is.

The re-arm mechanism, settled 2026-08-18, after two false starts recorded here rather than quietly fixed:

1. First instinct: the PLATFORM silently respawns the watch command when it fires, no agent turn involved. Wrong — the SDK only ever notifies the agent about a process it is itself tracking via Monitor. A platform-side respawn exists outside that tracking, so it would fire once, work, and then go dark forever — the agent is never told about the second fire. Not a cost tradeoff, a dead end.
2. Second instinct: keep the agent in the loop, but have the PLATFORM drive it — detect the acting turn's result and send a synthetic follow-up message telling the agent to re-arm. Technically correct, but costs a real extra turn every single cycle, and the owner pushed back on that cost before it was built: an actual multi-hour, real-world watcher (the mobile side's own, on this repo, this session) had already been running with the agent re-arming ITSELF, same turn, for hours — zero failures. That is hard evidence to override with a hypothetical reliability argument.
3. What is actually built: the agent re-arms itself, as the LAST action of the SAME turn that handles the notification — no separate turn, no platform-side spawn. The platform's role is narrower than 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 boolean). handoff — "Exit and hand back to the agent, then re-arm" — carries the contract (handoffContract in background-job-prompt.ts) verbatim: "When the condition fires you will be handed back, with the change. Do the work — and before you end this turn, start the SAME watch again yourself, exactly as you did the first time: same condition, same command. That is the whole loop, every time: act, then re-arm, in the one turn." The closing "IF IT EXITS" section is rewritten for handoff too, so it does not contradict rule 7 by telling the model an exit means stay-stopped.

• isCommander, gating the promise to where it can be kept. The proposal frames this entirely in terms of commanders — "a commander already has a row," arming state's only home — and outside one there is nowhere to persist "this should keep re-arming itself," so nothing would fulfil the promise. Rather than make it anyway, BackgroundJobDialog only offers handoff when isCommander is true, defaulting to it there and to the pre-existing stop everywhere else. Not yet wired to true anywhere — Command HQ's own composer, which would set it, does not exist yet (web) or does not use this dialog (mobile has its own). Every existing /chat caller is unaffected: two options, stop default, exactly as before tonight.

• Stopped no longer reads as failed. watchers-what-broke.md §5: a job Officer had just SIGTERM'd came back status: 'failed' (exit code 144) from the harness's own classification, which this codebase does not control. Fixed client-side, at the one place it can be fixed cheaply: stopTask in BackgroundTaskTray.tsx now also calls markStopRequested(taskId) (background-job-origin.ts), and useChat.ts's task:notification handler overrides the displayed status to stopped when a stop was just requested for that id, regardless of what the harness reported. 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, 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 && onFire === 'handoff') → sendWithHiddenPrefix → QueuedPrompt.armWatcher → sendPrompt's new last parameter → the chat websocket message → ClientMessage's 'chat' variant → startClaudeTurn's msg.armWatcher, which sets session.pendingArmWatcher = true — on the SESSION object, not captured by createMessageHandler's closure, because the dialog is routinely submitted turns after that closure was built (the session-scoped event subscription wires ONCE, on the session's first-ever turn) and a value captured then would never see a later arm.

• The noticing, in createMessageHandler (websocket.ts). On task:started: if pendingArmWatcher is set, consume it (one-shot, the same arm/claim shape background-job-origin.ts already uses client-side for the identical problem) and call armCommandHqWatcher — this is what turns "some task started" into "this is the watcher" in the first place. If the commander is ALREADY armed instead, a new 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 — 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 than the common case, and not attempted to be closed tonight.

Suggested order from the proposal, all four items now done: (1) prompt + dialog, (2) stopped-vs-died, (3) the re-arm mechanism itself — landed as agent-driven, same-turn, after two rejected designs recorded above — (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 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.

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

Raised during spec and set aside for a second version, in the owner's words: "that's not... that's out of scope for tonight."

• Silent death is invisible. A background job is a child of the agent sidecar process, not of Officer. pm2 restart officer-claude-code kills every running watcher with no task:notification ever emitted, so collectRunningTasks (started-minus-notified) reports it running until the 7-day event prune. For a watcher meant to run for months, this is the actual failure mode — a card that says "running" and has been dead since the last deploy. Deferred anyway, because tonight is development against a live server that gets restarted constantly; the target environment (a server nobody restarts) makes this rare in practice. When it is picked up: task-restart.ts's pidsHoldingFile (fuser against the task's output file, run as the owning user) is the existing liveness probe used by the restart watcher — comparing "log 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 — 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 /chat/tasks/:id (log tail or subagent transcript) already cover the raw material.
• Filtering and search on the category list. Named explicitly as v2 in the first message of the spec.
• Tag UI. Schema is in scope for v1 (see above); a picker, filter chips, anything user-facing is not.
• Category management UI details — delete behaviour (default: uncategorise the affected commanders rather than cascade-delete their rows, not explicitly confirmed), reordering, anything beyond create/list.
• Multi-account visibility for the owner — "sometime for the next version we will want the super user to have visibility of who is running what." user_id on every table already makes this a query change, not a schema one.

Resolved, previously open

• Creation flow — resolved, see "How a commander is created".
• Naming / title storage — resolved, see "Naming".
• Category cardinality — resolved: one per commander, a column not a join table. Tags cover the many-per-commander case.
• /clear chains — the owner's answer, plainly: nobody sane runs /clear on a commander, and this is not being built for the case where they do. A commander's session_id is one transcript; chain-splicing is not replicated.
• Working directory — resolved, see "Working directory" above. Not something the owner had originally planned for; raised and settled within this same conversation.

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. 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.