work-items:work — autonomous orchestrator-dispatch lifecycle underspecified (branch/worktree provisioning, PR-creation ownership, fix-subagent re-dispatch)

[kyle-sexton]

Deferred from PR #563 review (Codex P1 + claude[bot] findings). Three facets of one root cause: /work-items:work Step 5 documents an autonomous orchestrator-dispatch flow whose git/worktree/PR lifecycle seams with /implementation:implement-dispatch, /source-control:worktree, and /source-control:pull-request are not fully specified. A truthful fix is not confined to work/SKILL.md — it requires a design decision and likely edits to the sibling skills — so it was deferred rather than papered over in the PR.

Facet 1 — worker branch/worktree not provisioned before dispatch (Codex P1, SKILL.md:143)

• work Step 5 "Suggest branch name" (SKILL.md:141) only emits git checkout -b ... for a user and "never executes itself" — in autonomous/loop mode from a default-branch checkout, nothing creates the branch.
• /implementation:implement-dispatch preflight (implement-dispatch SKILL.md:29) requires "branch correct (never the default branch)" before composing the first brief (orchestrator-side).
• /source-control:worktree create calls EnterWorktree as its final action and transitions the whole session (worktree SKILL.md:66) — so the orchestrator cannot "create a worktree and then keep orchestrating."
• Net: no specified mechanism puts the autonomous orchestrator (or the dispatch entry) on a non-default branch/worktree before the preflight runs. Needs a decision on who provisions the per-pick worktree and when.

Facet 2 — PR-creation ownership vs the diff gate (claude[bot], SKILL.md:149)

• The high-blast-radius diff gate is "pre-PR" and orchestrator-owned, but no bullet states who opens the PR or when. The execute bullet requires the worker to "follow every step of the consuming project's development workflow," which may include PR creation — a worker could open the PR before the orchestrator's gate runs, making the gate a no-op.
• The worker scope-fence should explicitly list PR-creation as FORBIDDEN and the orchestrator should own PR timing — but "orchestrator never edits source" + out-of-tree worker leaves it unclear how the orchestrator reaches the worker's branch to open the PR. Ties back to Facet 1's provisioning decision.

Facet 3 — "owning subagent" for CI-found fixes is ambiguous (claude[bot], SKILL.md:151)

• "fix branch-owned findings via the owning subagent" does not say whether that is the original worker (whose out-of-tree worktree may already be cleaned up by /source-control:worktree) or a newly dispatched scope-fenced fix subagent, nor which worktree it uses. Depends on whether the per-pick worktree persists from worker-return through CI-green.

Acceptance

Specify (in work/SKILL.md plus whatever sibling-skill contract changes the decision requires) the end-to-end lifecycle: branch/worktree provisioning before dispatch, PR-creation ownership + worker scope-fence exclusion, and the fix-subagent re-dispatch + worktree-persistence contract — such that an autonomous run from a default-branch checkout can actually execute.

Filed by PR #563 review-response (babysit). Source threads: Codex P1 inline + claude[bot] PR-review comments on #563.

[kyle-sexton]

Facet-2 framing addition (control tower, 2026-07-20, from PR #563's held P1 review thread). The absorbed work-skill execute step's gap is sharper than "who/when opens the PR": as absorbed, NO step invokes /source-control:pull-request at all, so in the no-project-PR-stage fallback a completed worker change has no path to become a PR. Interim reality: lane prompts still carry the canonical PR contract (rule 6), so nothing breaks operationally today, and PR #563's docs now disclose the deferred handoff explicitly (no-false-guarantees narrowing). Design here should decide: orchestrator-owned invocation of /source-control:pull-request vs worker-owned, and where the project-PR-stage detection seam lives.

[kyle-sexton]

This was generated by AI during triage.

Triage — T3, needs-human, wayfind: design.

Genuinely open, multi-skill architecture gap, not confined to one file: the autonomous orchestrator-dispatch lifecycle (who provisions the worker branch/worktree before dispatch, who owns PR-creation timing relative to the pre-PR diff gate, which subagent re-touches a branch-owned CI fix) spans /work-items:work, /implementation:implement-dispatch, /source-control:worktree, and /source-control:pull-request. Each sibling skill's own preflight/contract assumes the others already solved the seam (worktree's create transitions the whole session as its final action; implement-dispatch's preflight requires an already-correct branch). Related to #573 (same source PR #563 review, same "work dispatch lifecycle underspecified" root, distinct specific gap) — filed and routed separately since resolving one doesn't resolve the other.