mirror of https://github.com/garrytan/gstack.git
Same root cause as land-and-deploy's Step 7a: canary's alert path already routes through AskUserQuestion (Phase 5), but the most common outcome — nothing broke across the whole monitoring window — never puts the live page in front of a human. Phase 6 prints "VERDICT: DEPLOY IS HEALTHY" straight from automated checks, and Phase 7 offers to promote that unconfirmed state into the new baseline other runs compare against. Phase 6 now opens the live URL and asks for an explicit pass/fail after monitoring completes, distinguishing `Human smoke: CONFIRMED` from `Human smoke: AWAITING` in the report and JSONL log — matching the verdict vocabulary land-and-deploy's Step 9 already uses. Phase 7's baseline-update offer is gated on CONFIRMED so an unconfirmed all-clear can't quietly become the new reference point. Part of the same fix wave as #2580 (land-and-deploy). Scoped this one to canary only after auditing the other ~50 skills for the same pattern (automated verdict of HEALTHY/VERIFIED with no human-confirmation gate on the quiet all-green path) — qa, design-review, ios-design-review, ios-qa, and benchmark all already report per-item evidence (verified/best-effort/ reverted/deferred, or objective metrics) rather than a global "it works in prod" claim, so they don't share this specific gap. Tested: bun run gen:skill-docs --dry-run reports FRESH, exit 0. Targeted bun test (skill-parser, skill-validation, gen-skill-docs, skill-size-budget): 757 pass / 1 fail, the fail is the same pre-existing package.json/VERSION mismatch already on main, unrelated to this diff. |
||
|---|---|---|
| .. | ||
| SKILL.md | ||
| SKILL.md.tmpl | ||