feat(workflow): /next-issue and /wrap-up, with priority labels and an issue claim protocol

[MarkMichaelis]

Summary

This replaces hand-maintained tracking issues with a queue that lives on the issues themselves, and adds the two commands that work it.

• /next-issue: a stateless dispatcher. It selects open issues that are not held, not claimed and not blocked, ordered by lowest priority-N and then oldest, and claims each one before dispatching it. Each issue goes to a new named session (through the existing Start-IssueAgent.ps1), to a background subagent when the work is small and settled, or to this session (here).
  • Claim: push the branch, add in-progress, and post a marker comment.
  • Races: the earliest claim wins.
  • Stale claims: a claim that has been silent for 3 days is taken over, with a comment saying so.
• /wrap-up: for each of this session's PRs whose exit criteria all hold, it checks the repository's shared instructions for an explicit merge-without-asking grant.
  • With the grant, it merges and cleans up.
  • Without it, or when it can't tell, it does not merge. It lists the PR under Needs you as ready to merge, with the evidence and the exact merge command.
  • This makes it correct whether or not fix(dev-loop): merging a finished PR is pre-authorized -- merge, close, clean up, then report #507 (merge pre-authorization) ever merges.
  • Everything else is parked with a hand-off comment and a claim release. Loose ends are filed as issues, then it reports.
• Delivery: both are Claude Code skills under .claude/skills/, which Claude Code turns into slash commands. The sync manifest names exactly those two directories, so the rest of a consumer's .claude/ (settings, hooks, its own skills) is never touched.
• Instructions:
  • New section: Issue Queue -- Priority Labels and Claims, covering the label contract and the claim, race, stale and release rules.
  • Generic standing rules moved out of trackers and one user's local memory: judge Pester by FailedContainersCount; re-scan stale issue inventories; a hand-applied rebase is new authorship.
  • Repo-specific traps now live in project.instructions.md.

Owner decisions (recorded on #498)

Decision	Choice
Command names	/next-issue, /wrap-up
Stale-claim timeout	3 days
Default fan-out	1
Delivery	.claude/skills/next-issue/ and .claude/skills/wrap-up/, by exact directory
Copilot prompt file	not in this PR, filed as #504

Independent review

The review ran on Sonnet 5, a different model from the Opus 5 author, under an explicit model override. It passed selection (verified live, with #117 on top), the claim race, sync delivery, leak hygiene, and the no-new-command-line-options check. It returned changes requested, with two findings, both accepted and fixed test-first in d0cf8a7:

1. Important. /wrap-up asserted that merging "is pre-authorized by the Merge Step in the shared instructions". That isn't true on main, because the grant lives in the held PR fix(dev-loop): merging a finished PR is pre-authorized -- merge, close, clean up, then report #507. Step 2 now checks for the grant and, without it, lists the PR under Needs you instead of merging. Shipped skill text no longer cites any issue or PR number, and a scoped assertion pins that.
2. Important, weak oracle. The /next-issue selection tests matched qualifiers anywhere in the body. They now extract the fenced gh issue list block under Step 1 and assert the exact --search argument. The ranking and launcher assertions are scoped to their own sections.

While fixing (2), the first version of the section helpers used break/continue, which Pester 6 reports as an escaped loop label and fails the whole block. The helpers are now regex-based.

Verification

GitHub Actions is billing-blocked, so every result below is a local run with its scope named. All were re-run at d0cf8a7:

Scope	Command	Result
.github suite	Invoke-Pester -Path .\.github	Result=Passed, 850 passed, 0 failed, 5 skipped, FailedContainersCount=0, 71 containers
Root sync suite (not run by CI, #309)	Invoke-Pester -Path (Resolve-Path .\Pull-SDLC.ai.Tests.ps1).Path	Result=Passed, 394 passed, 0 failed, 0 skipped, FailedContainersCount=0
Skill-contract suite	.github/agents/tests/issue-queue-skills.Tests.ps1 (inside the 71 above)	Result=Passed, 22 passed, FailedContainersCount=0
Leak scan + internal cross-references	the validate-instructions.yml patterns and link check, run locally with bash	no leaks, 0 broken references
Frontmatter	parsed with a real YAML parser	both skills: name matches the directory; description present

Falsification. Every ablation ran in a scratch copy, never in the worktree:

• Widening the manifest entry to a .claude/ prefix fails 9 of 11 sync tests. The 2 that still pass are the "treats as upstream-managed" cases, which a wider prefix also satisfies.
• Moving -is:blocked out of the Step 1 command and into the surrounding prose fails the exact-search assertion. A qualifier mentioned only in prose no longer satisfies it.
• Putting back /wrap-up's unconditional "merging is pre-authorized" sentence fails all four of the new /wrap-up assertions.
• Before the implementation, the skill tests and 4 of 11 sync tests failed.

Functional dry run. I ran the skill's selection query against this repository: 72 candidates. #500, which has open blockers, was correctly excluded.

Assumptions

• Adding the two skill directories to the manifest is not a new command-line option. No script gained a parameter.
• The repo-specific reason for "never -Sweep here" (concurrent sessions' branches) is inferred. feat(workflow): replace tracking issues with priority labels, issue claims, and a /next-issue dispatcher #498 states the rule without a reason.

Closes #498
Closes #500

🤖 Generated with Claude Code

https://claude.ai/code/session_01Fda3nBeB9rJBPdJSmD7sRN