work-items(work): absorb v4 loop-prompt execution rules into the skill

[kyle-sexton]

Category: behavioral/operational · Source: 2026-07-19 deep-dive audit (external PR audit: 7/7 GOOD; gaps below are process, not code quality). Session-scoped prompt rules that are skill gaps.

Absorb into the skill

• Orchestrator-dispatch posture as the documented default (= work-items(work): loop orchestrator implements issues inline instead of dispatching per-issue subagents #451): pick → spawn implementation subagent in out-of-tree worktree → collect → bookkeeping. Orchestrator never edits source.
• Canonical PR contract shipped in EVERY subagent brief: body skeleton (Summary/Fix/Verification), Closes #N, ## Related, per-plugin version bump + CHANGELOG, attribution trailer + session link. Audit: 3 red PRs + fixer round-trips were pure rework from learning the contract via CI; 2/7 PRs dropped the trailer (template drift = no shared template).
• Post-green review pass: after CI green, fetch review threads once; fix branch-owned findings via the owning subagent; classification reply; RESOLVE addressed threads; VALID-but-deferred findings REQUIRE a filed follow-up issue cited in reply + ## Related; then hand to babysit. Self-audit confirmed the lane never looked back at its PRs.
• High-blast-radius diff gate: PRs touching skill frontmatter descriptions/trigger keywords, cross-plugin contracts, or hooks get an orchestrator full-diff read before opening (self-audit: least-confident PR was exactly this class, passed through on CI alone).
• Concurrency + batch caps as userConfig (defaults, not hardcoded): audit showed 2-concurrent and 10-batch both binding for zero quality reason. Same-plugin serialization awareness until repo: per-change version bump + top-insert CHANGELOG serialize concurrent same-plugin PRs (conflict by construction) #464 lands.
• Skip ready-with-open-PR (interim in-progress detection; retire when work-items(work): status: ready issues with an open PR stay pickable — no in-review state (re-pick risk) #463 lands).
• Never-merge boundary stated explicitly (merging = babysit lane / human).

Related

#451 dispatch · #462 PR template · #463 in-progress marker · #464 version serialization

Stays prompt-level

Week priority directive + pick-order weighting; batch numbers chosen for the week.

[kyle-sexton]

Additional absorption item from live failure analysis: overnight the lane STOPPED after one 55-min cycle — the prompt's 'at most 10 issues per session-window' read as a full-session quota, and in self-paced /loop mode a quota-exhausted lane schedules no next wakeup. Skill must own the semantics: batch caps bound a CYCLE, never the loop; cap-reached/queue-empty ends the cycle and schedules the next (delay scaled to queue depth); the loop ends only on human stop or repeated-failure lane STOP. v4 prompt wording fixed both work and triage lanes; skill should make hand-prompt wording irrelevant.

[kyle-sexton]

This was generated by AI during triage. · Category: behavioral/operational · Priority: priority: high (week directive — pipeline improving itself). Routed status: ready.

Absorbing the v4 loop-prompt execution rules into /work-items:work

These seven rules currently live in a session prompt but are really work-skill gaps: today the skill's execute step (Step 5) defers generically to "the project's development workflow" and says nothing about inline-vs-dispatch, the PR contract, post-green review, blast-radius, concurrency, open-PR skipping, or the merge boundary. Plan below absorbs each as a behavioral change, delegating anything another skill already owns rather than re-stating it.

Seams this must NOT duplicate

• /implementation:implement-dispatch already owns the dispatch mechanics and already makes orchestrated cadence the default for autonomous runs (scope-fenced worker briefs, orchestrator-never-edits, verify-returns-against-evidence, concurrency cap). work chains to it — it does not re-describe dispatch.
• /source-control:pull-request owns the PR body shape, the branch→Closes #N closing-keyword injection, and the post-green review-thread loop (fetch → validate → classify → threaded reply → react → resolve bot-authored threads).
• /source-control:babysit-prs owns the fleet loop and never merges.
• /work-items:track owns filing follow-up issues.
• /source-control:worktree owns worktree lifecycle (the out-of-tree isolation per pick).

Per-bullet plan

• 1 — Orchestrator-dispatch as the documented default → absorb (chain to /implementation:implement-dispatch). Make dispatch the documented default execution posture for autonomous work-item execution: pick → spawn an implementation subagent in an out-of-tree worktree (via /source-control:worktree) → collect → bookkeeping, orchestrator never edits source. Delegate all dispatch mechanics to /implementation:implement-dispatch; work only states the default and the per-pick worktree isolation. This is where work-items(work): loop orchestrator implements issues inline instead of dispatching per-issue subagents #451 lands. Contestable fork — see the defaulted decision below.
• 2 — Canonical PR contract in every subagent brief → absorb into the brief template; contract owned elsewhere. The PR body skeleton (Summary/Fix/Verification), Closes #N, and merge style are owned by /source-control:pull-request — do not redefine them here. The gap is that the dispatch brief must carry the contract forward so a worker knows it up front instead of learning it via red CI: reference /source-control:pull-request's body contract, and enumerate the consuming-project conventions the brief must relay — per-plugin version bump + CHANGELOG entry, attribution trailer + session link, ## Related. (Confirmed: /source-control:pull-request owns the body shape; this bullet is brief-carriage, not a second definition.)
• 3 — Post-green review pass → mostly already-covered-by /source-control:pull-request; absorb only the work-item linkage + sequencing. The fetch-once → validate → classify → reply → resolve-bot-threads loop is already the pull-request monitor phase; babysit owns the fleet form. Absorb into work: (a) after CI green, run one review pass, fixing branch-owned findings via the owning subagent; (b) a VALID-but-deferred finding requires a filed follow-up issue (via /work-items:track add) cited in the classification reply and in ## Related; (c) then hand off to /source-control:babysit-prs. Only (b) and the sequencing are new to work.
• 4 — High-blast-radius diff gate → absorb (new pre-PR gate). Before opening a PR, the orchestrator does a full-diff read when the diff touches skill frontmatter descriptions / trigger keywords, cross-plugin contracts, or hooks. Complements /implementation:implement-dispatch's scope-fence discipline; not currently expressed anywhere.
• 5 — Concurrency + batch caps as userConfig → absorb as config defaults (not hardcoded). Introduce userConfig keys for max concurrent dispatch waves and batch size, read as defaults; recommended concurrency default mirrors /implementation:implement-dispatch's existing 3–5, following the plugin's existing userConfig precedent. Same-plugin serialization defer-to-repo: per-change version bump + top-insert CHANGELOG serialize concurrent same-plugin PRs (conflict by construction) #464 — until it lands, work carries an interim same-plugin serialization awareness note. Second contestable surface, but the default is grounded (existing 3–5) and maintainer-vetoable.
• 6 — Skip ready-with-open-PR → absorb (interim), defer cleanup to work-items(work): status: ready issues with an open PR stay pickable — no in-review state (re-pick risk) #463. work's staleness pre-check gains an interim check: skip a frontier item that already has an open PR. Retire this interim detection when the durable in-progress marker (work-items(work): status: ready issues with an open PR stay pickable — no in-review state (re-pick risk) #463) lands.
• 7 — Never-merge boundary → absorb (explicit statement). State that work's lane ends at PR creation and handoff (review pass, then babysit); merging is the babysit lane / human, never work. Consistent with /source-control:babysit-prs never merging and /source-control:pull-request merge being human-gated.

Stays prompt-level (out of scope)

Week-priority directive, pick-order weighting, and concrete batch numbers remain session-prompt concerns, not skill behavior.

Acceptance criteria

• work's execute step documents orchestrator-dispatch (out-of-tree worktree per pick, orchestrator never edits source) as the default for autonomous execution and chains to /implementation:implement-dispatch without re-describing its mechanics.
• The subagent-brief surface carries the PR contract by referencing /source-control:pull-request and enumerates version-bump / CHANGELOG / attribution-trailer / ## Related / Closes #N obligations — with no second definition of the PR body shape.
• A VALID-but-deferred review finding cannot be resolved without a filed follow-up issue cited in the reply and ## Related.
• A PR whose diff hits skill descriptions/trigger keywords, cross-plugin contracts, or hooks triggers an orchestrator full-diff read before the PR is opened.
• Concurrency and batch caps resolve from userConfig defaults (no hardcoded literal in work), with the recommended concurrency default matching the existing 3–5.
• A frontier item with an open PR is skipped by selection (interim), with a documented retirement trigger of work-items(work): status: ready issues with an open PR stay pickable — no in-review state (re-pick risk) #463.
• work states the never-merge boundary explicitly.
• #451 (dispatch), #463 (in-progress marker), #464 (same-plugin serialization) referenced at their respective bullets; source-control/repo: PRs fail pr-issue-linkage (missing ## Related) — no template scaffolds the required sections #462 template acknowledged via the fix: correct plugin-system inaccuracies and harden references #2 pull-request seam.

Decision defaulted: bullet 1 — orchestrator-dispatch (spawn a scope-fenced implementation subagent in an out-of-tree worktree; orchestrator never edits source) is made the DOCUMENTED DEFAULT for autonomous /work-items:work execution, replacing the skill's current provider-neutral "follow the project's development workflow" deference. — veto before merge if dispatch should stay opt-in rather than default.