Repository navigation
feat(fleet): add /task-manager skill + CI skills-install gate (v0.10.0) - #9
Merged
Merged
Conversation
Adds a fourth dedicated-seat skill — the peer of /coordinator (git) and /brain (knowledge) — for GOALS + TASKS + AGENT STATUS coordination. - /task-manager onboards a window as the fleet's coordination seat: it keeps every agent lane's active goal current, surfaces stalls + reachability gaps + stale claims + blockers from `fleet.sh roster` / `monitors`, routes work with `fleet.sh msg`, keeps the board tidy, and gives the operator a "who's on what / what's next / what's stuck" readout. - Generic + leak-free for this PUBLIC repo: Fleet's own roster / board / msg / goal / retire are the substrate; any richer project-specific task board is an OPTIONAL integration, referenced by capability only (no project paths, no dashboard port, no SQLite/UI specifics). The seat works with Fleet alone. - No new script or install.sh change needed — install.sh already copies every dir under .fleet/skills/ into the target's .claude/skills/. Forcing function: a new `skills-install` CI job runs install.sh into a temp project and asserts EVERY skill under .fleet/skills/ propagates into .claude/skills/ (dynamic over the source, so future skills are auto-covered; explicit asserts for the known skills guard a silent glob no-op). This closes a real gap — the smoke job only ran `fleet.sh init`, which does not copy skills, so skill propagation was previously untested. Negative-control-verified: the job goes red when a skill fails to land. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
convenientlymike
added a commit
that referenced
this pull request
Sep 11, 2026
…ing (#9) Per coordinator decision, finding #9 (upgrade-transition collision) is propagation-gating and folds into this PR so the pre-propagation head is complete and #9 is closed BY CONSTRUCTION, not by rollout timing. On a rollout onto an ACTIVE fleet, existing windows carry an agent-FILE label but have no reservation yet (created on their next register.sh). Without a guard, a NEW registration during that window is handed an agent-N a live window's file already shows — the duplicate-agent-N collision, reintroduced purely by rollout order. reserve_label (step 1b preferred + step 2) and next_label now consult a new _label_held_by_live_file(label, self): a label carried by a LIVE agent-file owned by a different sid OCCUPIES its slot, so a new registration skips it. Self-excluded (a session re-registering its own label is never blocked by its own file); live-only (a stale/dead file-label stays reusable). The check runs only on an otherwise-free slot (mkdir won / dir absent), so it is ~free in steady state and only bites during the transition. Forcing test L8: a LIVE agent-file labeled agent-1 with NO reservation → a new window's reserve_label returns agent-2, NOT agent-1; control: a DEAD/stale agent-1 file-label IS reusable (new window takes agent-1 — no over-block). Watched to BITE with the guard reverted. Gate: shellcheck clean; selftest-labels 19/19 (L1-L8); collateral 16/16 + brain-retire 15/15 (no regression); smoke green. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
convenientlymike
added a commit
that referenced
this pull request
Sep 11, 2026
…ardening (#11) * fix(fleet): close reap-vs-reserve race + complete label-reservation hardening Pre-merge red-team follow-up to #10 (atomic agent-N reservation). A 6-lens adversarial review (each finding refuted before trust) surfaced one critical residual and two completions. CRITICAL — reap-vs-reserve race (reintroduced the collision #10 fixed): reap()'s label-GC removed a reservation whose owner had no agent file, but register.sh creates the reservation BEFORE writing the agent file, so a concurrent reap (runs on every fleet command in every session) could delete a just-created in-flight reservation → a 3rd session re-mkdir's the same agent-N → two live agent-N. The reap-GC path now carries the same reservation-age grace guard _label_reclaimable uses: GC only when owner departed AND the reservation dir aged past FLEET_LABEL_GRACE. Protects the fresh in-flight slot and the empty-sid sub-window (mkdir→sid-write); a genuinely-departed owner's minutes-old reservation still frees promptly. COMPLETIONS: - sid_for_target pass-2 (kept-stale files) now fails loud (exit 3 + candidate shorts) on a duplicate label, symmetric with pass-1 — a handoff DM to a departed agent-N is no longer silently delivered to one of two same-label inboxes. - FLEET_LABEL_GRACE is config-readable (label_grace_s) like stale_after_s, and its default is raised 3→10s to comfortably exceed the reserve→agent-file write gap on a cold host spawning many windows. FORCING COVERAGE (selftest-labels.sh, each watched to BITE with its guard reverted): - L7: a concurrent reap during an in-flight reserve → the fresh reservation SURVIVES; control: an aged departed reservation IS still GC-removed. - L7b: the empty-sid sub-window (mkdir'd, sid not yet written) is likewise protected. - L4b: pass-2 kept-stale duplicate fails loud; control: a unique stale label still delivers. Gate: shellcheck clean; selftest-labels 17/17; selftest-collateral 16/16 + selftest-brain-retire 15/15 (no regression); smoke green. Follow-ups tracked in the PR (confirmed but out of this PR's tight scope): reclaim-vs-reclaim atomicity, and file-label↔reservation reconciliation on the upgrade transition (existing windows carry file-labels but no reservation). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(fleet): fold upgrade-transition guard into the reservation hardening (#9) Per coordinator decision, finding #9 (upgrade-transition collision) is propagation-gating and folds into this PR so the pre-propagation head is complete and #9 is closed BY CONSTRUCTION, not by rollout timing. On a rollout onto an ACTIVE fleet, existing windows carry an agent-FILE label but have no reservation yet (created on their next register.sh). Without a guard, a NEW registration during that window is handed an agent-N a live window's file already shows — the duplicate-agent-N collision, reintroduced purely by rollout order. reserve_label (step 1b preferred + step 2) and next_label now consult a new _label_held_by_live_file(label, self): a label carried by a LIVE agent-file owned by a different sid OCCUPIES its slot, so a new registration skips it. Self-excluded (a session re-registering its own label is never blocked by its own file); live-only (a stale/dead file-label stays reusable). The check runs only on an otherwise-free slot (mkdir won / dir absent), so it is ~free in steady state and only bites during the transition. Forcing test L8: a LIVE agent-file labeled agent-1 with NO reservation → a new window's reserve_label returns agent-2, NOT agent-1; control: a DEAD/stale agent-1 file-label IS reusable (new window takes agent-1 — no over-block). Watched to BITE with the guard reverted. Gate: shellcheck clean; selftest-labels 19/19 (L1-L8); collateral 16/16 + brain-retire 15/15 (no regression); smoke green. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What & why
Adds a fourth dedicated-seat skill — the peer of
/coordinator(git) and/brain(knowledge) — for goals + tasks + agent-status coordination. A multi-window fleet benefits from one window owning "who's doing what, what's next, what's stalled, who won't get their wakes."Changes
/task-manageronboards a window as the fleet's coordination seat: keeps every agent lane's active goal current, surfaces stalls + reachability gaps + stale claims + blockers fromfleet.sh roster/monitors, routes work withfleet.sh msg, keeps the board tidy, and gives the operator a who's on what / what's next / what's stuck readout.roster/board/msg/goal/retireare the substrate; any richer project-specific task board is an optional integration referenced by capability only (no project paths, dashboard port, or store/UI specifics). The seat works with Fleet alone.install.shchange —install.shalready copies every dir under.fleet/skills/into the target's.claude/skills/.Forcing function (new)
A
skills-installCI job runsinstall.shinto a temp project and asserts every skill under.fleet/skills/propagates into.claude/skills/— dynamic over the source (future skills auto-covered), with explicit asserts for the known skills to guard a silent glob no-op. Closes a real gap: the smoke job only ranfleet.sh init(which does not copy skills), so skill propagation was previously untested. Negative-control-verified locally: the job goes red when a skill fails to land.Verification (byproduct)
:9477/taskboard.py/ private paths)install.shmirror: all 4 skills (arm, coordinator, brain, task-manager) propagateci.ymlvalid YAML · shellcheck clean · retire+brain selftest still 15/15VERSION 0.9.0 → 0.10.0; CHANGELOG + README updated.
🤖 Generated with Claude Code