[gering]

Summary

First slice of /work-system:manager: the status part. Dispatch/collect stay in add-manager-watch-loop.

• scripts/workers.sh / workers.py — list, screen, question, ready (all read-only) and a one-shot events.
  • Roles come from the shared classify_cwd (exported by the wrapper from herdr-agent.sh, like sessions.sh), so there is no second path classifier.
  • Flags: question (blocked dialog), asks (trailing question), verify (idle while the screen shows activity or could not be read; herdr #4376/#3090), draft (someone types).
  • screen clamps --lines to 20..200 (herdr #5462); question reads only the live dialog's options and never answers.
• events --state <file>, deduplicated:
  • blocked: one event per pane + session UUID and transition; an /exit successor starts fresh; herdr's unknown status is no observation.
  • ready-for-merge: one event per lane + head. Rule, shared with workers.sh ready: open, non-draft, same-repo PR at the lane's HEAD; merge state CLEAN/HAS_HOOKS; no changes requested; at least one check and all green; mandate.sh covered. Tri-state: an unreadable answer never re-arms an event.
  • Stage 1 pings the Manager through manager.sh prompt, resolved in the worker's workspace. Stage 2 runs the user's on-blocked command (argv, JSON on stdin, either event) only when the ping did not land or after escalate_after; a failing command is retried on at most 3 runs.
  • The command comes only from a CLI flag or ~/.config/work-system/events.toml (regular, own, not group/world-writable), never from a repo. The main checkout's .work-system.toml gets [events] enabled / escalate_after; a lane's own copy is ignored.
  • flock against overlapping timer ticks; the state file is validated before any delivery.
  • No loop ships: herdr 0.9.5 has no status-change hook, so the README documents a crontab example.
• manager.sh prompt never types into the caller's own pane (reason=manager-is-self).
• Skill /work-system:manager; marks pane text, titles, recaps and payloads as data, never instructions. The Manager reports "ready for merge", never merges or approves.
• work-system 1.32.0: CHANGELOG, READMEs, CLAUDE.md, schema (events.*), knowledge entry.

Decisions from the task (Robert, 2026-11-03)

Only this slice now; ready-for-merge from assess + PR/CI + covered until a run record exists; command only from user scope or a flag; events is one-shot for a user timer. Patch PR #100 (1.31.1) is separate — expect a CHANGELOG rebase if it lands first.

Test plan

• test_workers.py: 123 checks with fake herdr/gh, real git and real mandate.sh; argv-log proof that nothing calls agent send-keys|prompt|send-text or pane send-*
• test_manager.py (incl. the own-pane guard)
• scripts/check-structure.py: 0 errors, 0 warnings
• Live, read-only against herdr 0.9.5: list, list --all --json, screen <manager> --lines 5000 → lines=200, question <manager> → blocked=no
• events live only with a throwaway tab and test worker (not run in the lane)

Review

Three local /swarm:review rounds. Round 1 ran with claude lenses only (both cc-harness voices failed with "no gateway token" despite ready); round 2 ran with native codex (gpt-6.1-sol) + grok-4.7, 3 families, 7 consensus findings. Agreed fixes applied and recorded (mandate.sh covered → via=agreed-fixes). Declined: deduping the /continue gate ping against the event ping, folding assess + gh pr view into one call (assess is the agreed source), and dropping last_line from the payload (the spec requires it; the README warns it may hold secrets).

Round 3 (/swarm:review --fix, codex gpt-6.1-sol + grok-4.7 + lenses, 25 findings, 3 consensus): fixed ready-for-merge for lanes with several panes (all must be idle), dialog options with description lines and with a prompt below an old list, the hook's captured pipe (output now discarded), re-firing the same head after a red re-run, pruning on an empty agent list, pinging forever on reasons that cannot change, events without tomllib (now exit 3), the task name typed raw into the Manager (slugged), asks=unknown on a failed read, plus doc gaps. 13 new tests (136). Declined: SKIPPED/NEUTRAL-only rollups (GitHub counts them as passing), the on_blocked name (from the spec), assess + gh again, a single combined screen read (recent ≠ visible), and unifying the sanitizers with sessions.py (no canonical one; separate task).

Round 4 (/swarm:review --fix, same voices, 13 findings, 3 consensus, 8 refuted): ready now applies the lane-wide idle rule too (reason=lane-busy), so the skill and events agree; a record is pruned only when its pane, session or lane was seen gone (a role that fell to other on a main-repo-path timeout, or one poll without a UUID, re-fired the event); a retried ping prints retry= instead of a new event=; the ping bound covers manager.sh's four herdr calls; the screen and dialog blocks blank bidi controls; herdr-agent.sh's JSON checks run python3 -I; one rev-parse per lane; doc drift (final ping reasons, re-arm wording, question exit codes, composer consumers). 5 new tests (141). Declined: assess + gh again (Robert's decision).

Round 5 (/swarm:review --fix, same voices, 15 findings, 5 consensus, 6 refuted; review budget now used up): events judges a lane busy over every workspace (a --workspace run missed a busy pane elsewhere) and counts any non-self agent in the lane or a sub-directory of it, as ready does; ready reads only that lane's screens (was O(workers²) per Manager pass); screen/dialog blocks blank C1 and unterminated escapes; the lock file is opened with O_NOFOLLOW; a bad HA_CALL_TIMEOUT_SECS falls back instead of a traceback; manager.sh's task= attribution is a slug; doc fixes (skill and README synopses, settings README for linked worktrees, knowledge frontmatter). 5 new tests (146). Declined: no-main-repo stays final (a deterministic failure), assess + gh (Robert's decision), a lazy screen read for delivered blocked events, and unifying the sanitizers (separate task).

Round 6 (/swarm:review --fix at Robert's request beyond the budget, over the branch including Oskar's checkpoint commit; 20 findings, 5 consensus, 4 refuted): a typed ping carries task= and worktree= only as slugs (a worktree name with spaces read as a sentence outside the bracket), while the SendMessage body keeps the real names (the round-5 slug had changed it too); ping_text no longer repeats the task name, so the slug rule lives in one place; light reads also cover other agents, so an idle-but-working agent in a lane sub-directory keeps the lane busy; a ready record survives while its agent sits in a sub-directory; a scoped run whose workspace emptied prunes its vanished panes (the empty-list guard is over every workspace); the skill calls ready once per lane; doc fixes (lane-busy wording, --escalate-after synopsis, ready_for_merge docstring, CHANGELOG). 4 new tests (150) plus one in test_manager. Declined: replaying a still-current event after re-enabling, failing closed on an unreadable events.enabled (ws-settings never blocks, by design), a hyphenated name inside the attribution bracket (residual), the own-pane guard without HERDR_PANE_ID (herdr always sets it), assess + gh (Robert's decision), scope-limited screen reads for events, and sharing the composer glue (separate task).

Review by Oskar; no merge without Robert's go.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RtoWD1aAxWazpn1eWEyVmq

[oskar-portmann]

Review by Jonas Schmied (Dev bot), posted by Oskar because the jonas-schmied app is not installed on this repo. Reviewed head 6c876fa, verdict: approve. S1/S2 pushed as follow-up commit a4814f9.

Review of 6c876fa. Mainly workers.py (events), workers.sh, manager.sh, herdr-agent.sh, the tests, the schema and the version files.

Verdict: approve. No blockers.

Headless use (dev-1, no TTY)

• No read from the terminal, no tput or colours, and no terminal-size assumptions.
• External calls go through run() with timeouts, and the hook has a 30 s bound with stdin carrying JSON and stdout/stderr discarded.
• An unreadable herdr exits 3/4/5 before the state is touched, and gh/GitHub blips give unknown, which never re-arms an event.
• Concurrent runs are excluded by flock (summary: skipped=locked), and the lock is opened with O_NOFOLLOW.
• The hook is argv-only (shlex.split / TOML list, no shell). The ping text is slugged, and the own-pane guard is in manager.sh:629.

Idempotence and dispatch

• blocked: one event per pane+session transition (workers.py:918).
• ready-for-merge: one event per lane+head (workers.py:950-952).
• The hook runs only on an undelivered ping or after escalate_after (workers.py:797-802), with at most 3 tries.

Mutation probes (each one turns a test red)

• new = True for blocked
• hook_due ignoring a sent ping
• ready-for-merge head dedupe removed
• hook stdin payload dropped
• events.enabled=false ignored
• manager.sh own-pane guard removed

Tests and versions

• test_workers.py: 146 checks pass. test_manager.py: pass.
• Version 1.32.0 is in sync in plugin.json and marketplace.json, with a CHANGELOG entry as the repo's CLAUDE.md requires.
• The schema events.enabled / escalate_after defaults match the code (!= "false", 0).
• No .github/ changes, and no secrets.

CI on 6c876fa: scan pass, structure pass.

Should (non-blocking, will be addressed in a follow-up commit on this branch)

• S1 workers.py:805-822 / :982-985: the state is saved only at the end of the run. If the run is killed mid-loop (timer timeout, reboot), already-delivered pings and hooks are sent again on the next run. Save atomically after each deliver() that sent something.
• S2 workers.py:600-601 (and the matching README/docstring wording): "must not silence its own events" overstates it. A worker under the same user can still edit the main checkout's .work-system.toml, so ignoring the lane's copy is defense in depth, not a boundary.

Nit

• N1 workers.sh:16: shellcheck SC1007 on CDPATH= cd. It is the same idiom as manager.sh:82 and intentional, so leaving it is fine.

Not checkable here: real herdr/dev-1 behaviour and a live --ping into real Manager panes; the tests use fake herdr/gh.

[oskar-portmann]

Delta review 6c876fa → a4814f9.

Verdict: approve. No blockers.

• range-diff 6c876fa..f6d0a26 vs 6c876fa..a4814f9: 1: f6d0a26 = 1: a4814f9. The content is identical to the S1/S2 patch, and the tree diff f6d0a26 ↔ a4814f9 is empty. Only the committer differs; the author stays jonas-schmied[bot]. No .github/ changes, version unchanged at 1.32.0 (one CHANGELOG line under 1.32.0).
• S1: checkpoint() saves the state atomically after each delivery that ran a ping or hook. The kill-mid-loop test passes. Mutation probe: with checkpoint disabled, exactly that test fails (FAIL: 1 of 147).
• S2: wording only (README and workers.py): ignoring the lane's .work-system.toml is defense in depth, not a boundary.
• Local results: test_workers.py 147 checks pass, test_manager.py pass, check-structure.py OK (0 errors, 0 warnings).

CI on a4814f9: scan pass, structure pass.

Disclosure: I authored the S1/S2 commit (applied by Oskar via git am), so this delta review covers my own change; the rest of #101 was reviewed on 6c876fa.