agent-worker-flow

name	agent-worker-flow
description	Standard claim/branch/verify/publish workflow for pod agent sessions. Read this skill at the start of any feature, review, summarize, or meditate session.
allowed-tools	Bash, Read, Glob, Grep

Standard Worker Flow for Pod Agent Sessions

This skill covers the shared workflow used by all pod worker agents. Session-specific commands reference this skill rather than duplicating it.

Coordination Reference

The coordination script handles all GitHub-based multi-agent coordination. Session UUID is available as $POD_SESSION_ID (exported by pod). The gh CLI defaults to the current repo, so --repo is not needed.

Command	What it does
coordination orient	List unclaimed/claimed issues, open PRs, PRs needing attention
coordination plan [--label L] "title"	Create GitHub issue with agent-plan + optional label; body from stdin
coordination create-pr N [--partial] ["title"]	Push branch, create PR closing issue #N, enable auto-merge, swap claimed → has-pr. With --partial: adds replan label.
coordination claim-fix N	Comment on failing PR #N claiming fix (30min cooldown)
coordination close-pr N "reason"	Comment reason and close PR #N
coordination list-unclaimed [--label L]	List unclaimed agent-plan issues (FIFO order); optional label filter
coordination queue-depth [L]	Count of unclaimed issues; optional label for per-type count
coordination claim N	Claim issue #N — adds claimed label + comment, detects races
coordination skip N "reason"	Mark claimed issue as needing replan — removes claimed, adds replan label
coordination add-dep N M	Add depends-on: #M to issue #N's body; adds blocked label if #M is open
coordination check-blocked	Unblock issues whose depends-on dependencies are all closed; remove orphan blocked from issues whose body has no depends-on: lines
coordination check-has-pr	Remove orphan has-pr from open issues that have no currently-open PR closing them (post audit comment)
coordination release-stale-claims [SECS]	Release claimed issues with no PR after SECS seconds (default 4h); manual use only
coordination release-orphan-claims	Release claims whose owning session UUID is no longer in .pod/agents/ (liveness-based, no age threshold); manual use only
coordination lock-planner	Acquire advisory planner lock (20min TTL)
coordination unlock-planner	Release planner lock early
coordination critical-path-depth [L]	Count unclaimed critical-path issues; optional label filter
coordination set-target N	Planner sets recommended target agent count (wind-down use only)

Issue lifecycle: planner creates issue (label: agent-plan) → worker claims it (adds label: claimed) → worker creates PR closing it (label swaps to has-pr) → auto-merge squash-merges. Issues marked replan (by skip, partial completion, or worker-led decomposition) are handled by the next planner. Issues with has-pr are excluded from list-unclaimed and queue-depth.

Never apply has-pr manually. The label is set automatically by coordination create-pr (full path) and cleared by GitHub's auto-close on merge of a Closes #N PR. Hand-applying it desynchronises the label from any actual PR — when the supposed PR closes or merges without Closes #N, the issue stays has-pr forever and is silently excluded from the work queue. If you want to "park" an issue while sub-issues do the work, use coordination add-dep <parent> <sub> for each sub-issue: the parent becomes blocked, and check-blocked auto-clears the label when all subs close. The orphan-label housekeeping cycle (check-has-pr, check-blocked) auto-removes mis-applied labels and posts an audit comment on the issue.

Partial completion: worker uses --partial → label swaps to replan. A planner creates a new issue for remaining work, then closes the replan issue with a link to the new one.

Worker-led decomposition: if the claimed issue is too large for one session, worker creates sub-issues and coordination skips the parent with a Decomposed into #X, #Y breadcrumb comment. The next planner either closes the parent (sub-issues fully cover it) or narrows it to the residual scope. See "Assess Scope" (Step 4b) for the full procedure.

Dependencies: Issues can declare depends-on: #N in their body. coordination plan auto-adds the blocked label if any dependency is open. check-blocked (run by pod each loop) removes blocked when all dependencies close. Blocked issues are excluded from list-unclaimed and queue-depth.

Branch naming: agent/<first-8-chars-of-UUID> Plan files: plans/<UUID-prefix>.md Progress files: progress/<UTC-timestamp>_<UUID-prefix>.md

Step 1: Claim a Work Item

coordination orient

Priority order: 0. Directives first: Check for open directive issues before anything else. These are direct instructions from the project owner — work flowing down from the human, not work awaiting human attention — and take absolute precedence over all other work:

coordination list-unclaimed --label directive

If any are open and unclaimed, claim the oldest one immediately. Directives cannot be skipped or refused because you disagree with the approach. The valid exits from a directive are (a) completing the deliverables and closing the issue yourself, (b) opening a partial PR and noting the remaining scope so a successor can pick it up, or (c) posting a comment explaining a genuine technical blocker (e.g. a missing dependency) and coordination skip with that reason. Do not leave a directive open with a "for owner closure" note — if the deliverables are in, close it. Do not skip because you think a different approach is better — that is the owner's call, not yours.

1. Oldest unclaimed issue of your type:

   coordination list-unclaimed --label <your-label>
   

Don't repair PRs from a worker session. PR health (merge conflicts, failed CI, stuck CI) is the repair agent's responsibility; pod dispatches /repair automatically when coordination list-pr-repair reports candidates, ahead of /plan / /work. Focus on fresh issue work.

If the queue is empty, write a brief progress note and exit.

coordination claim <issue-number>

You MUST check the output. If it says CLAIM FAILED, you MUST NOT work on that issue — pick a different one. Only proceed if the output says Claimed issue #N. Read the full issue body:

coordination read-issue <N> --json body --jq .body

Step 2: Set Up

git checkout -b agent/<first-8-chars-of-session-UUID>
git rev-parse HEAD      # record starting commit

If the branch already exists (common in reused worktrees): check for an open PR on it first (gh pr list --head agent/<id>). If a PR exists, create a new branch with a suffix (agent/<id>-v2). If no PR exists, reset it to master: git checkout agent/<id> && git reset --hard origin/master.

Before that reset --hard, run git status --short first. A reused worktree may carry uncommitted working-tree edits left by a prior session (e.g. half-finished skill/command tweaks) — reset --hard discards these unrecoverably. If you see any, git diff them: they are almost always disposable cruft the reset is meant to clear, but confirm they are not unpushed work before discarding. If they look worth keeping, commit them to a throwaway branch (git stash is banned in the home repo but fine here) so the next session can recover them.

Record any project-specific quality metrics (e.g. sorry count, test coverage) as described in the project's CLAUDE.md.

Editing progress/items.json (any session, especially review/fidelity audits that bulk-update fidelity/status fields): edit surgically — grep -n the item id, Read those ~15 lines, Edit just the field value, and drop any now-stale fidelity_note. Never json.load+json.dump the whole file: the reserializer reflows indentation/key-order/unicode into a multi-thousand-line diff against the shared 13k-line file (it only stays clean by luck if your dump params happen to match). When flipping a gap back to verified, also remove its fidelity_issue and confirm the linked repair issue actually merged. (Full rationale in the lean-formalization skill.)

Step 3: Codebase Orientation

Read the specific files mentioned in the plan/issue. Understand the current state of code you'll be modifying. Don't read progress history — the issue body provides that context.

Step 4: Verify Assumptions

Check that the plan's assumptions still hold:

• Quality metrics match what the issue says
• Files mentioned in the issue still exist and haven't been restructured
• No recently merged PR invalidates the plan
• For "infra"/"helper" issues, check the deliverable does not already exist — possibly under a different name or in a differently-named file. Planners write these without full codebase knowledge, and this repo has many overlapping per-chapter files. Before writing any code, grep for the target lemma name and its mathematical content (key symbols in the statement, e.g. character.*Nonempty.*≅). An equivalent (or more general) sorry-free lemma already present means the issue is redundant: point to it in a comment and skip for the planner to close. Do not re-prove it under a new name. (Seen with #6425, whose deliverable was already Etingof.charEq_iso from an earlier PR.)
• For "final assembly" issues that consume prerequisites: verify the infrastructure it depends on actually exists sorry-free. check-blocked unblocks an issue when its depends-on deps close — but a dep closed as replan (decomposed) means its real work moved to still-open sub-issues, so the assembly is not actually ready. Confirm the named lemmas/defs exist in the Lean files before working; if not, re-add depends-on on the real open sub-issues and skip.
• Sanity-check the target statement is actually provable before attempting the proof. A planner-written signature can be false as stated (hypotheses too weak, wrong quantifier, missing finiteness). Spend a moment trying to break it: do the hypotheses actually pin down the conclusion, or can you build a counterexample? (E.g. "M has one composition factor" does not imply finite length — R=ℤ, M=ℤ is indecomposable with unlinked factors ℤ/2, ℤ/3.) If the statement is unprovable, do not attempt it or silently strengthen the signature: post a comment with the counterexample and suggested fix, then skip for replan (Escalation ladder: "Ordering mistake in the plan → report, request replan").

If stale:

coordination skip <issue-number> "reason: <what changed>"

Go back to Step 1 and try the next issue.

PR fix plans: If the plan asks you to fix a broken PR, use judgement. If the PR is low quality or not worth salvaging:

coordination close-pr <pr-number> "reason: <why not worth fixing>"

Step 4b: Assess Scope

After orienting but before writing code, check whether the task fits in a single session. Warning signs it doesn't:

• Target file is 500+ lines and you need to understand most of it
• The work naturally splits into independent sub-lemmas or sub-tasks
• Difficulty feels higher than the issue says

If the issue is too large, decomposing it into smaller sub-issues is a normal success path, not a failure mode. You have the freshest codebase context and can usually scope sub-tasks more accurately than a planner could in advance. A good decomposition is more valuable than a failed heroic attempt — and far better than overrunning the session trying to salvage it.

You may decompose when any of these is true:

• the claimed issue is too large for one session,
• the work naturally splits into independent sub-tasks,
• you can write self-contained successor issues without further investigation.

# 0. FIRST search for an existing continuation issue before creating one. A
#    planner may have already pre-split your claimed issue into a follow-up,
#    and that follow-up will NOT appear in `coordination list-unclaimed` if it
#    carries `replan`/`blocked`. Search ALL open issues by item ID / theorem
#    name, not just the unclaimed queue:
#      gh issue list --state open --search "<ItemID or theorem> in:title" \
#        --json number,title,labels
#    If a suitable continuation already exists, reuse it (point your breadcrumb
#    at it) instead of creating a duplicate.
#
# 1. Otherwise create self-contained sub-issues. Use `coordination plan`
#    exactly as a planner would — same body template (Current state /
#    Deliverables / Context / Verification), same label. Note: `coordination
#    plan` does only best-effort title-keyword overlap warnings; it does not
#    hold the planner lock and cannot atomically dedupe against concurrent
#    creators. If you see open issues that look related, link or coordinate
#    explicitly in the sub-issue body rather than relying on the warning.
#    Do NOT write a literal `depends-on: #N` token in prose (even to say it is
#    NOT needed): `coordination plan` parses that string and auto-applies the
#    `blocked` label. Phrase such notes without the literal token.
echo "body..." | coordination plan --label feature "Sub-task 1: ..."
echo "body..." | coordination plan --label feature "Sub-task 2: ..."

# 2. Link ordering dependencies if any sub-task must precede another.
#    Do NOT add `depends-on: #<parent>` — the parent is about to be
#    superseded; depend on real predecessor sub-issues instead.
coordination add-dep <sub2> <sub1>

# 3. Leave a machine-readable breadcrumb on the parent. The planner's
#    replan-triage step keys off this exact `Decomposed into #X, #Y`
#    phrasing — keep it on a single line at the start of the comment.
gh issue comment <parent> --body "Decomposed into #<sub1>, #<sub2>

(reason: <one-line scope assessment>)"

# 4. Release the claim by marking the parent for planner triage. This
#    routes through pod's claim-state machinery and clears the in-process
#    `state.claimed_issue` correctly. Do NOT use `gh issue close` directly:
#    it leaves pod's session-end cleanup thinking the issue is still
#    claimed and will attempt a stray `coordination skip` on a closed
#    issue at exit.
coordination skip <parent> "Decomposed into #<sub1>, #<sub2> — see comment"

The planner's next replan-triage cycle picks the parent up and either closes it (if the sub-issues fully cover it) or narrows the body to the residual scope (if not).

After decomposing, you have two options:

1. Continue on one of the sub-issues: claim it via coordination claim, then return to Step 2 with the sub-issue. Common case when the parent was just two work items glued together.
2. Stop and exit: if you've used most of your session orienting, write a brief progress entry and exit. The next worker will claim a sub-issue. The next agent onboards from progress/ on main, so a progress entry committed only to your unmerged branch is lost. If your handoff carries design analysis worth preserving (exact reusable lemmas found, route chosen, subtleties), land it as a small progress-handoff PR: commit just the progress/ file, push, gh pr create, gh pr merge <N> --auto --squash. The durable design detail should also live in the sub-issue bodies you created — those persist regardless of any PR.

If you've already done a coherent subset of the parent's work before deciding to decompose, prefer the partial-PR path:

# Steps 1-3 above (create sub-issues for the remaining work, leave the
# `Decomposed into #X, #Y` breadcrumb on the parent).

# Then land your coherent subset. `--partial` marks the parent `replan`.
coordination create-pr <parent> --partial "feat: <what landed>"

The next planner sees the breadcrumb on the parent and closes it with a forward link to the sub-issues.

Step 5: Execute

After each coherent chunk of changes:

• Build and test using the project's build commands (see project CLAUDE.md)
• Commit with conventional prefixes: feat:, fix:, refactor:, test:, doc:, chore:

Each commit must compile. One logical change per commit.

Commit early, create PRs early. Sessions can terminate at any time. Pushed-but-not-PR'd work is effectively lost — nobody will find it.

• Commit after every compiling milestone. Don't wait for the full feature.
• WIP commits are fine: feat: WIP prove helper_lemma (2/4 sorries remain)
• If 20+ minutes have passed without a commit, stop and commit now.
• Use coordination create-pr N --partial as soon as you have useful progress, even if incomplete. This saves the work as a visible PR.

Failure handling:

• Build fails on pre-existing issue → log and work around
• Stuck after 3 fundamentally different attempts → decompose into sub-issues (Step 4b)
• 3 consecutive iterations with no commits → end session, document blockers (does not apply to review or self-improvement sessions)
• If /second-opinion or /reflect is unavailable, skip and note in progress entry

Step 5b: Context Health

If conversation compaction occurs:

1. Finish your current sub-task (get to compiling state)
2. Commit what you have
3. Skip remaining deliverables — do NOT start new work
4. Go directly to Step 6 then Step 7 with --partial

Commit early and often. Each commit is a checkpoint.

Step 6: Verify

Build and test the project. Compare quality metrics with the starting values. Review your diff: git diff <starting-commit>..HEAD. Use /second-opinion if available.

Step 7: Publish

Write a progress entry to progress/<UTC-timestamp>_<UUID-prefix>.md:

• Date/time (UTC), session type, what was accomplished
• Decisions made, key patterns discovered
• What remains, quality metric deltas

Full completion:

git push -u origin <branch>
coordination create-pr <issue-number>

For a long session, origin/main may have advanced (other agents merged PRs) since you branched. Before create-pr, sync so your branch builds against and diffs cleanly against current main:

git fetch origin main
git merge origin/main --no-edit   # resolve any conflicts, then rebuild
git diff --stat origin/main..HEAD # sanity check: only YOUR files should appear

If unrelated files show up as deleted, your branch is stale — merge first.

Once the PR is created, exit. Do not poll CI, wait for the merge, or otherwise spin on the PR. Another session will pick up any follow-up work (e.g. a "fix PR #N" issue if CI fails). Polling burns context and tokens for no benefit.

Partial completion (did NOT complete all deliverables):

• Progress entry lists: completed deliverables, NOT-completed deliverables and why, whether unfinished work needs a new issue
• Use --partial:

  coordination create-pr <N> --partial "feat: what was actually done"
  

If you only closed a bad PR (no code changes):

gh issue close <N> --comment "Closed PR #M as not worth salvaging. See progress entry."

Step 8: Reflect

Run /reflect. If it suggests improvements to skills or commands, make those changes and commit before finishing. Do NOT modify the project's top-level CLAUDE.md or roadmap files — those are off-limits to agents.