[Jammy2211]

Closes nothing on its own — phase 1 of the autonomous-orchestrators epic, tracked on PyAutoLabs/PyAutoBrain#551. Read together with PyAutoLabs/PyAutoBrain#552.

What changes

• policy/shift_watcher.md (new) — the deterministic watcher is the one exemption to "sessions end at their deliverable": a non-LLM script, calls no Claude tool, arms nothing in a session, dispatches nothing without an open brief, STOP file checked first. Cites the 2026-08-31 and 2026-09-03 drains.
• batches/AGENTS.md — "Shift brief (2026-11-03)": brief: block (ordered goals, named tasks, backlog: none | fill-idle | scoped + exclude:, review-minutes, review-at, lanes, merge-cap, token-budget, optional ral-allowance), issue:, packet: yes; the calibration rule; a "Digest" subsection with counters; the one-screen "Proof packet"; batches/examples/ is not a record.
• REFERENCE.md — documents Lane: any | local-dev | ral | gpu | wsl-data (last three imply local-dev) and the planned RAL test packet (phase 5).

Approved deviation

policy/end_at_deliverable.md and its hook are byte-identical to main: the policy text is stamped into 50 repos by repos_sync.py, so editing it would redden --check until ~50 stanza commits land. The exemption lives in the new page instead; a one-line pointer in end_at_deliverable.md rides with the next routine stanza refresh.

Tier: judge — merge mode: human /prm. Never auto-merge. Merge this one first, then the PyAutoBrain PR.

Witness (run on both branches, laptop, 2026-11-03)

## (a) end_at_deliverable hook payloads (env -u PYAUTO_ALLOW_TIMERS)
{"tool_name":"ScheduleWakeup","tool_input":{"delaySeconds":600}} -> exit 2 | policy end_at_deliverable: sessions end at their deliverable — ScheduleWakeup would outlive the turn.
{"tool_name":"CronCreate","tool_input":{}} -> exit 2 | policy end_at_deliverable: sessions end at their deliverable — CronCreate would outlive the turn.
{"tool_name":"RemoteTrigger","tool_input":{"action":"run"}} -> exit 2 | policy end_at_deliverable: sessions end at their deliverable — RemoteTrigger would outlive the turn.
{"tool_name":"send_later","tool_input":{}} -> exit 2 | policy end_at_deliverable: sessions end at their deliverable — send_later would outlive the turn.
{"tool_name":"RemoteTrigger","tool_input":{"action":"list"}} -> exit 0 | 

## (a') policy/end_at_deliverable.md byte-identical to main
identical (no diff vs origin/main)

## (b) Revert condition count
main: 6  branch: 11
529:## Shift briefs and unattended `glance` — 2026-11-03
539:### Shift brief = launch — 2026-11-03
554:### Unattended `glance` merges — 2026-11-03
575:### One Mind writer per shift — 2026-11-03
590:### Goal decomposition files drafts only — 2026-11-03
601:### The push-notification test — 2026-11-03

## (c1) lifecycle.py check  →  DRIFT (pre-existing, not this task)
  - active/ prompt no registry entry claims: active/chapter2-explanations.md
  ! warning: batches/2026-08-31-pm.md: review-minutes-actual: empty (pre-existing)

## (c2) repos_sync.py --check (branch script, --root ~/Code/PyAutoLabs)
check generated smoke bootstraps: OK
check PyAutoHeart/config/repos.yaml: OK
check PyAutoHands/pre_build.sh: OK
check PyAutoHands/autohands/config/workspaces.yaml: OK
check ensure_workspace_labels.sh: OK
check hygiene conductor coverage: OK
check local checkout origins: OK
check workspace checkouts (manifest ↔ disk): OK
check tenant firewall (organ code): OK
check root AGENTS.md routing table (generated): OK
check WORKFLOW.md owner map (generated): OK
check organism-map blocks (generated): OK
check never-rewrite-history blocks (generated): OK
check remote-session blocks (generated): OK
check end-at-deliverable blocks (generated): OK
check shared-standards blocks (generated): OK
check where-to-file blocks (generated): OK
check public front-door organ tables (generated): OK
check hub organism blurb (organs present): OK
check CLAUDE.md → AGENTS.md pointers: OK
check generated hooks (session-start + end-at-deliverable): OK (50 of 50 checked out, 1 excluded)
check generated Codex hooks: OK
check target-repo layout lints: OK
exit 0
(from the worktree root the only ✗ is COWLS_COSMOS_Web_Lens_Survey, which worktree_create flattens from lens/ to the top level — a layout artifact; this diff touches no scripts/ or repos.yaml)

## (c3) pytest -q (PYAUTO_ROOT/MIND/BRAIN unset)
PyAutoMind:  711 passed in 37.66s
PyAutoBrain: 1409 passed, 1834 warnings in 86.91s (0:01:26)

(c3) note: with the per-task activate.sh env set, one test per repo fails
(test_repos_sync_root_default::test_a_root_that_does_not_hold_this_checkout_is_refused,
test_grouped_organ_consumers::test_hygiene_map_and_heart_board_find_grouped_checkouts).
Both fail identically on canonical main under that env and pass on this branch without it — they read the worktree's
PYAUTO_ROOT, they are not touched by this diff.

(d) PyAutoMind/batches/AGENTS.md now carries a fenced example shift brief, a fenced example digest and the proof-packet block — the phase-2 validator's target.

🤖 Generated with Claude Code