feat: add gmd skill

[d-morrison]

Closes #2635

Adds the gmd skill: a composite directive that runs /gia (clear the PR queue, then the issue backlog) under a session-scoped away judgment grant and the mwc merge-confidence grant, so a single keyword launches an unattended clear-everything session.

• skills/gmd/SKILL.md — the Claude Code skill (delegates decision latitude to away, merges via mwc's confidence gate).
• codex-skills/gmd/SKILL.md — the Codex-side counterpart, per the generalize-to-every-agent rule.

Review history: four findings addressed (daytb→away scope, missing Grep/Glob tools, hardcoded personal path in hooks.json, unquoted ${extensionPath}), plus a semantic-line-breaks pass. Latest verdict: Ready for merge on 14c9e20b.

[d-morrison]

@claude review

[d-morrison]

@claude review

[github-actions]

👀 Picked up by workflow run #33292749646. Setup runs first; Claude itself responds after that.

[github-actions]

👀 Picked up by workflow run #33292759146. Setup runs first; Claude itself responds after that.

[d-morrison]

@claude review

[github-actions]

👀 Picked up by workflow run #33294963539. Setup runs first; Claude itself responds after that.

[d-morrison]

@claude review

[github-actions]

👀 Picked up by workflow run #33295168731. Setup runs first; Claude itself responds after that.

[d-morrison]

@claude review

[github-actions]

💰 Cost: $0.15895320000000002

[d-morrison]

@claude review

[github-actions]

👀 Picked up by workflow run #33295466606. Setup runs first; Claude itself responds after that.

[d-morrison]

@claude review

[github-actions]

👀 Picked up by workflow run #33298652744. Setup runs first; Claude itself responds after that.

[d-morrison]

Claude Code CLI (local session) is taking this PR the last mile per the user's direct instruction. It is fully clean per scripts/check-pr-fully-clean.py (clean verdict on HEAD 14c9e20b, CI green), and the clean verdict has stood for 10+ hours. I intend to merge it under the standing ai-config mwc grant in ~5 minutes unless the owning session posts a hold-off here.

Posted by Claude Code (AI agent) --- not written by a human.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.

_{2026-08-30 18:04 UTC}

[github-actions]

Claude finished review — View run

---

All checks that have completed are passing (build, validate ×2, new-line-breaks ×2, gather-context, preempt-previous); only this review run itself is pending — that's this run's own sibling job, not a content finding.

Code review: Morrison-Lab/ai-config#2638 — add gmd skill

What changed since the last review round, verified with git rather than assumed from the diffstat: the range 14c9e20b..0b43c673 given in the task consists entirely of merge commits (300e7bd "Merge branch 'main' into feat/gmd-skill" and 0b43c67, the PR's own merge-commit head) plus the commits they bring in from main. I confirmed directly (git diff 14c9e20b 0b43c673 -- skills/gmd/SKILL.md codex-skills/gmd/SKILL.md) that both of this PR's own files are byte-identical between the last-reviewed commit and the current head — the merge introduced no new content to this PR's actual diff. The saved PR diff (.claude-review-pr.diff, PR-vs-target-branch) confirms this independently: it contains only the two new SKILL.md files, unchanged from what was already reviewed.

Prior findings — status

All four previously-fixed items remain resolved and untouched by this round:

1. daytb→away scope mismatch — fixed, unchanged.
2. Missing Grep/Glob in allowed-tools — fixed, unchanged.
3. Hardcoded personal absolute path in hooks.json — fixed, unchanged (as of the PR's own commits).
4. Unquoted ${extensionPath} interpolation — fixed, unchanged (as of the PR's own commits).

A note on plugins/ai-config/hooks.json, checked and ruled out of scope

Diffing c626989..HEAD (i.e. including everything the merge pulled in) shows hooks.json now reads python3 ~/.gemini/config/plugins/ai-config/<script>.py with no ${extensionPath} token at all — at first glance this looks like a regression of finding #4. I traced it with git log/git show and confirmed it is not part of this PR's diff and not a regression to flag here: origin/main's hooks.json is byte-identical to this PR's current hooks.json, and the switch away from ${extensionPath} to the stable ~/.gemini/config/... path was made deliberately on main itself, in commit 50badca8 ("fix(plugin): use stable absolute path instead of extensionPath in hooks.json", part of PR #2664, already merged). Since the saved PR-vs-target-branch diff shows zero changes to this file, it is inherited main state, already reviewed and merged independently of this PR — not something to hold #2638 to.

This round's diff content

No new commits specific to this PR — only merge commits with no net effect on the PR's own files. No new findings.

This confirms the most recent prior review ("Ready for merge", reviewing 14c9e20b) still stands.

Verdict

Ready for merge

Stopping Point: Clean stopping point reached — review posted with a "Ready for merge" verdict; no work remains queued in this review-only run.

Confirmed identical at the true PR head too — git diff 14c9e20b 300e7bd8 -- skills/gmd/SKILL.md codex-skills/gmd/SKILL.md is empty.

I ran the required instrument rather than relying on gh pr checks:

python3 scripts/check-pr-fully-clean.py 2638 -R Morrison-Lab/ai-config

PR #2638 (feat/gmd-skill): state=OPEN, HEAD=300e7bd8 (committed 2026-08-30T17:56:26Z)
  verdict scan: examined 13 dated automated review item(s), 7 bore a verdict, latest = clean; per-reviewer: Claude=clean

❌ PR is NOT fully clean:
  - Check run 'review / claude-review' is still in status 'in_progress'
  - No review comment has been posted evaluating HEAD SHA 300e7bd8 yet

Exit code 1, with genuine - finding bullets (not an unhandled exception), so this is a real not-clean verdict from the instrument — not something to wave away by assumption. But both listed reasons are this review run's own in-flight state: the claude-review check that's currently executing is this run, and the "no review comment yet" is because this run's posting job submits my message only after I finish — there is nothing else pending. Per the task's own Verdict semantics section, a verdict must not be conditioned on the reviewing run's own in-flight sibling checks or deferred to this script's exit status, precisely to avoid the Needs more work self-referential loop (ai-config#2475) where every run reproduces the blocker it's waiting on. The true PR head is 300e7bd8 (the 0b43c67 commit I initially cited is GitHub's merge-preview wrapper around it, per this task's own stated caveat); I re-verified the content match against that correct head above, and no other check has failed.

Correcting my earlier citation: the reviewed commit is 300e7bd8, not 0b43c67.

Verdict

Ready for merge

Reviewed commit: 300e7bd

Stopping Point: Clean stopping point reached — review posted (content-clean, no failed checks, instrument output reported transparently); no work remains queued in this review-only run.

Reviewed commit: 300e7bd

[github-actions]

💰 Cost: $1.8344 (review) — run