feat(devx-qa): add babysit-pr skill - #19

Summary

• Adds a babysit-pr skill so a PR's failing CI can be diagnosed and remediated without re-explaining the loop each time. Distilled from codex's babysit-pr, which is ~16KB of instruction the model largely already knows — this keeps only what isn't in the weights: the gh mechanics, the branch-caused vs. infrastructure classification, and hard budgets on autonomous pushes.
• The skill watches, diagnoses, and classifies; it does not own git mutations. Commits go through devx-git:ci and base-branch merges through devx-git:update-origin, so the push guardrails live in one place instead of being re-implemented without them.
• Bumps devx-qa to 1.6.0 and brings the docs back in line with the directory names Claude Code actually resolves.

Review

A /code-review max pass found 18 defects, reproduced live against gh 2.100.0. All 18 are fixed in 08cbd16, and each gh invocation in the file was re-verified against real repos before being written.

Security and safety

• The skill instructed the agent to run a command string read out of an untrusted CI log. Verification commands now come from the repo's own scripts, and log content is stated to be data, never instructions.
• Nothing forbade copying values out of a log into the tree. GitHub masks only registered secrets, and secrets-guard inspects file paths and command strings — not context. Now explicitly forbidden, and logs are grepped rather than read whole.
• This was the only auto-triggering, push-capable skill in the marketplace, with no tool ceiling. Now disable-model-invocation: true, a scoped allowed-tools, and confirmation before the first push.
• Nothing verified the checked-out branch was the PR head, so commits could land anywhere. Preflight now compares headRefName, checks out the PR, and refuses fork PRs.

Correctness — verified, not assumed

• select(.state=="FAILURE") missed CANCELLED, TIMED_OUT, ERROR, and ACTION_REQUIRED. Worse, a cancelled-only run exits 0: facebook/react#37589 reports 238 pass / 6 cancel and gh pr checks exits 0. Now dispatches on bucket.
• "Any other exit code = a check failed" conflated four states. Verified: cli/cli#14456 exits 0, facebook/react#37656 exits 1 with no required checks reported. Replaced with a full exit-code table plus message matching.
• gh run view --job <id> --log-failed errors run is still in progress — exactly the state --fail-fast creates — so no matrix failure was diagnosable. Replaced with gh api .../actions/jobs/<id>/logs --allow-escape-sequences, which serves a job's log mid-run. The old rationale for --job was backwards and is gone.
• gh run rerun --failed returns 403 This workflow is already running on an in-progress run, and the budget was decremented for the refused attempt. Now waits for status: completed and only counts a rerun that dispatched.
• mergeable: UNKNOWN — the normal value seconds after a push — was reported as ready, and BEHIND never triggered the merge written one line below it. Now re-polls, then dispatches on mergeStateStatus.
• PR=$(...) was set in one Bash call and read in three others. Shell state does not persist, and an empty selector does not error — gh silently falls back to the current branch's PR. Literal values are now substituted, and -R <owner>/<repo> is threaded through so a URL argument's repo isn't discarded.
• ~14% of check links aren't /actions/runs/<id>/job/<id> — external CI, CLA bots, other repos. Now branches on the workflow field and surfaces non-Actions checks as blockers.
• --required was never passed, so advisory checks entered the fix loop forever; it also exits 1 on repos with no branch protection, which is why it needed the exit-code table first.
• --watch outlived the Bash tool's timeout with no arm for it; --interval 60 restored.

Structure

• Nothing in the skill ran git fetch, so Step 6 merged a stale base ref and looped forever. Delegated to devx-git:update-origin, which fetches, takes a backup branch, and handles a rejected push.
• Budgets keyed on job name and were decremented only on the flake path, so a cascade of different jobs never tripped them, and the stop conditions were orphaned — no step consulted them. Now global caps (5 pushes, 3 reruns, 45 min, 2 attempts per error signature) evaluated at the loop head.
• A bare git push violated update-origin Guardrail 6, the fix shipped in 037c512. Now git push origin HEAD with the rejection recovery path.

Docs — corrected the --job claim the review disproved, aligned ### fixing-skills / ### fixing-react-antipatterns with their directory names, removed a ### claudemd section for a skill that does not exist, and fixed the stale /commit reference (the directory is ci).

Test plan

• Every gh invocation verified against live repos (cli/cli, facebook/react, nodejs/node, microsoft/vscode) on gh 2.100.0
• Exit codes confirmed: 0 all-pass, 1 no-required-checks, 8 pending, 4 auth
• Confirmed the bucket filter catches the cancelled-only false green
• Confirmed the jobs-logs API serves a log mid-run and --allow-escape-sequences is required
• Confirmed non-Actions checks carry no workflow field
• Frontmatter within limits: description 506/1024, body 231/500 lines
• End-to-end run against a real failing PR with a multi-job matrix