fix(pm): fleet-write derives the seat session from the container; the workflow’s Actions state is the kill switch; size-routed direct for post-stamped

[objectstack-fleet]

Fixes #19755

What changed

scripts/pm/** only — the fleet's own write tools, not on the governed register.

1. The seat session is derived from the container. sessionSource(env) in fleet-write/dispatch.mjs reads OS_FLEET_SESSION first (explicit wins, validation unchanged; an explicit value that is malformed is refused, never replaced). When it is absent, the container's CLAUDE_CODE_REMOTE_SESSION_ID — cse_ID — is read and becomes session_ID; malformed or absent gives null as before. The route's reason names which source was read, and --route now prints the session and its source. The container variable's name is CONTAINER_SESSION_ENV in fleet-write/ops.mjs, beside SESSION_ENV, with its shape CONTAINER_SESSION_SHAPE beside SESSION_SHAPE.
2. Condition ③ reads the workflow's Actions state. Relay live = the file on main (today's read) AND GET /repos/{board}/actions/workflows/fleet-write.yml answering 200 with state === 'active'. Any other state, status or no answer is not live, the reason naming it — disabled_manually verbatim when that is the state, described as the maintainer's kill switch. Both reads are cached once per process and reached only after ① and ②. Explicit OS_FLEET_TRANSPORT=dispatch against a disabled workflow is a prerequisite refusal (exit 3), never a silent fall-back.
3. Unchanged: the two ceilings, the no-run fall-back under auto, the op table, the executor, the workflow file, every allow rule. Every --help line and header comment that told a seat to export OS_FLEET_SESSION (dispatch.mjs, label-write.mjs, issue-create.mjs, with-fleet.sh) now says a cloud seat's session is read from the container and OS_FLEET_SESSION overrides it (a local checkout, a test). No documented spelling carries a prefix.
4. Self-tests for the derivation (well-formed / malformed / absent / explicit wins / malformed explicit never replaced / blank explicit counts as absent / the reason names the source), the state read (active / disabled_manually / another state / a 200 without a state / 404 / 500 / no answer / cached once / reached only after the file read / strict refusal), and the CLI (--dry-run and --route with the container variable alone). The dispatch and with-fleet CLI cases clear the inherited container variable, so both batteries hold inside a cloud container.
5. Size-routed fall-back under auto (addendum, item 6). sizeRoute(route, body) in post-stamped.mjs measures the RENDERED body in UTF-8 bytes against MAX_BODY_BYTES (60,000; the platform's client_payload cap is 64 KB). Over it under auto, THAT write takes direct and the transport line becomes one line naming the bytes, the cap and the identity used (the seat's own user, not objectstack-fleet[bot]). Explicit OS_FLEET_TRANSPORT=dispatch over the cap is refused (exit 3) naming the bytes. A route already direct, or already carrying an error, is untouched; a refused op, a failed run and UNCONFIRMED keep their register; label-write, issue-create, close-cards are not size-routed; no compression codec (pinned by import shape).

Verification record

--route in this cloud container, with NO OS_FLEET_SESSION set (the seat's own id from the claude-code-remote get_session tool, field ccr.id, is session_01GnJon4xkRvphn4w28xx3An; the container carries CLAUDE_CODE_REMOTE_SESSION_ID=cse_01GnJon4xkRvphn4w28xx3An — same tail). Angle-bracket placeholders in the pasted reason are spelled ID here for the body sanitizer.

$ node scripts/pm/fleet-write/dispatch.mjs --route
{"requested":"auto","transport":"dispatch","failed":null,"session":"session_01GnJon4xkRvphn4w28xx3An","session_source":"CLAUDE_CODE_REMOTE_SESSION_ID","reason":"OS_FLEET_TRANSPORT is auto → dispatch: CCR_AGENT_PROXY_ENABLED=1 (a cloud seat container, whose proxy replaces the Authorization header), the session was derived from the container's CLAUDE_CODE_REMOTE_SESSION_ID (cse_ID → session_ID; OS_FLEET_SESSION is absent), and the relay workflow .github/workflows/fleet-write.yml is on objectstack-ai/objectstack@main (GET answered 200) and its Actions state is active (GET /repos/objectstack-ai/objectstack/actions/workflows/fleet-write.yml answered 200)","error":null}
exit 0

Before this change the same command in the same container answered "transport":"direct","failed":"session".

Size route, live dry runs in this container (node --use-env-proxy scripts/pm/post-stamped.mjs --comment=19755 --file=… --dry-run; nothing written). A 60,159-byte rendered body:

post-stamped: transport direct — OS_FLEET_TRANSPORT is auto → direct for THIS write: the body is 60,159 UTF-8 bytes, 159 over the relay's 60,000-byte body cap (a repository_dispatch client_payload is capped by the platform at 64 KB), so it cannot take the relay. Written DIRECT as the seat's own user (the token this process holds), not as objectstack-fleet[bot]. ⛔ Not compressed: a body this size is the thing to shrink.

A small body from the same tree: post-stamped: transport dispatch — … the session was derived from the container's CLAUDE_CODE_REMOTE_SESSION_ID … its Actions state is active …. The disabled-workflow case is proven in the dispatch self-test's fake platform; the live workflow was not touched.

Batteries, each run alone with its exit captured before any pipe (worktree at 6470e4e):

command	exit	verdict line
pnpm check:pm-fleet-write-validate	0	69 cases pass across 7 batteries
pnpm check:pm-fleet-write-execute	0	48 cases pass across 9 batteries
pnpm check:pm-fleet-write-dispatch	0	90 cases pass across 10 batteries (was 64 across 9 at c120dbd)
pnpm check:pm-with-fleet	0	32 cases pass (was 31; floor 28 → 29)
pnpm check:pm-post-stamped	0	609 cases pass across 21 batteries (was 596 across 20)
pnpm check:pm-label-write	0	88 cases pass across 10 batteries
pnpm check:pm-issue-create	0	42 cases pass across 6 batteries
pnpm check:pm-close-cards	0	111 cases pass across 12 batteries
pnpm check:pm-write-pace	0	109 cases pass across 12 batteries

npx eslint --no-inline-config on the five changed .mjs files: exit 0, no findings.

Gates — node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands on the edited tree derived 38 families (6 paths vs merge base c120dbd; 449 changed lines, under the human-merge threshold). Each was run alone with its exit code captured before any pipe; all 38 exited 0. --ran with the codes recorded: 38 derived famil(ies) accounted for — 38 run, 0 NOT-MEASURED (a DERIVED zero — all 38 recorded an exit code and none of them is 3). Repo-wide pnpm lint is CI's run; the changed files were linted directly (above).

Changeset

None: scripts/pm/** publishes nothing from any released package, so skip-changeset applies. This dispatch forbids label writes, so the label is the seat's to apply.

Acceptance notes

• post-stamped --dry-run skips the proxy re-exec by design ("that path makes no request"), yet the route resolution on a dry run does read the relay, so in a cloud container a dry run without --use-env-proxy reads condition ③ as HTTP 401 and reports direct. Pre-existing (the read existed before this change; it was simply never reached without a session). Observation only, not filed — a live run re-arms first and is unaffected.
• close-cards.mjs and the sibling tools' self-test fixtures still spell OS_FLEET_SESSION is absent as a fixture error string; those are fixtures, not documented spellings, and were left as they are.
• .claude/skills/pm-dispatch/references/rest-channel.md still describes the old opt-in sentence; that path is on the governed register and outside this card's scripts/pm/** scope.

---

Generated by Claude Code

_{Dispatching seat (director, summon #27): skip-changeset applied by the seat — scripts/pm/** publishes nothing; the card's Claim: naming this branch is on #19755. Written 2026-09-23T00:48Z.}