feat: /debug sub-agent escalation from /qa + recommendations in /review and /ship (v0.6.5.0) (#192)

* feat: add browse access to /debug for visual verification

Debug skill can now use the browse binary to visually reproduce bugs,
take screenshots as evidence, and verify fixes. This makes /debug
effective for web app bugs when spawned as a sub-agent from /qa.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat: add /debug sub-agent escalation to /qa (Phase 8g)

When QA fix attempts fail twice on the same bug (reverted due to
regressions), /qa now spawns a /debug sub-agent with a structured
bug brief including symptoms, repro steps, failed fix details, and
file paths. Results are reported in Phase 10's debug escalation summary.

Sequential execution: one debug investigation at a time, working tree
cleaned between investigations. Graceful degradation on all failure
modes (BLOCKED, agent failure → deferred in report).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat: add /debug recommendation to /review (Step 5.7)

When /review finds what appears to be a pre-existing bug in the base
branch (not introduced by the PR's diff), it now classifies it as
INFORMATIONAL and recommends running /debug for systematic root-cause
investigation. No Agent spawning — /review's scope stays on the diff.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat: add reverted QA commit detection to /ship

During pre-landing review, /ship now checks for reverted fix(qa):
commits in the branch history and recommends /debug for systematic
investigation. Informational only — does not block shipping.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat: add debug escalation tests (validation + LLM judge + E2E)

Skill validation: 11 new assertions covering Phase 8g trigger, structured
handoff fields, agent result handlers, debug escalation summary, Step 5.7
recommendation, ship reverted QA detection, and debug browse setup.

LLM judge: evaluates Phase 8g template quality — structured brief format,
result handling, working tree cleanup, sequential processing.

E2E: prompt-level deterministic test (verifies escalation prompt has all
required fields) + full flow stub (fixture TODO for planted regression).

Touchfile entries for diff-based test selection.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* chore: add worktree parallel debug agents to TODOS.md (P2)

When /qa hits multiple stubborn bugs, parallel debug agents in
isolated git worktrees could investigate simultaneously. Deferred
from the sequential debug escalation PR as a follow-up.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* chore: bump version and changelog (v0.6.5.0)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* feat: add E2E evals for /review pre-existing bug + /ship reverted QA detection

Two new E2E tests:
- review-pre-existing-bug: plants SQL injection in base branch, verifies
  Step 5.7 classifies as INFORMATIONAL and recommends /debug
- ship-reverted-qa-commits: creates branch with reverted fix(qa): commits,
  verifies /ship detects them and recommends /debug

Also fixes qa-debug-prompt-logic to use correct workingDirectory, and
ensures test repo init uses -b main for portability.

All 4 debug-related evals pass: $0.34 total, 94s.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Garry Tan
2026-03-18 17:59:32 -05:00
committed by GitHub
parent 8914a0fdce
commit 94c1530efc
16 changed files with 579 additions and 2 deletions
+50
View File
@@ -17,6 +17,7 @@ allowed-tools:
- Grep
- AskUserQuestion
- WebSearch
- Agent
---
<!-- AUTO-GENERATED from SKILL.md.tmpl — do not edit directly -->
<!-- Regenerate: bun run gen:skill-docs -->
@@ -832,6 +833,41 @@ WTF-LIKELIHOOD:
**Hard cap: 50 fixes.** After 50 fixes, stop regardless of remaining issues.
### 8g. Debug Escalation
After the fix loop completes (all issues processed through 8a-8f), check for issues that were **reverted at least twice** (two or more separate fix attempts each caused regressions and were reverted).
If no issues match, skip this phase entirely.
For each matching issue, sequentially spawn a debug sub-agent using the Agent tool:
**Agent prompt — structured handoff:**
> Read `~/.claude/skills/gstack/debug/SKILL.md` (fallback: `.claude/skills/debug/SKILL.md`) and follow the systematic debugging methodology.
>
> **Bug Brief:**
> - Issue ID: ISSUE-NNN
> - Symptom: [what was observed during QA testing]
> - Severity: [critical/high/medium/low]
> - Affected URL: [page URL where bug was found]
> - Reproduction: [step-by-step repro from QA testing]
> - Screenshot evidence: [paths to before/after screenshots from QA]
> - Failed fix attempts:
> 1. [what was changed] → [what regression it caused] (reverted)
> 2. [what was changed] → [what regression it caused] (reverted)
> - Files investigated: [list of source files already examined during fix attempts]
> - Console errors: [relevant JS errors if any]
>
> Investigate this bug. You have full access to AskUserQuestion if you need user input. Output the DEBUG REPORT when done.
**When the debug agent returns:**
- **DONE status:** The fix is in the working tree. Re-test the affected page via browse. If verified, commit with `fix(qa/debug): ISSUE-NNN — [description]`. If regression detected, revert and mark as "deferred (debug-investigated, fix regressed)."
- **DONE_WITH_CONCERNS:** Apply the fix, note concerns in report. Re-test.
- **BLOCKED:** Run `git checkout .` to discard any uncommitted debug artifacts (temp logs, assertions). Mark as "deferred (debug-investigated, root cause unclear)." Include the debug report in Phase 10 output.
- **Agent call fails:** Mark as "deferred (debug unavailable)." Continue to Phase 9.
**Sequencing:** Process one debug issue at a time. Wait for each agent to complete before spawning the next. The working tree must be clean between debug investigations.
---
## Phase 9: Final QA
@@ -872,6 +908,20 @@ Write to `~/.gstack/projects/{slug}/{user}-{branch}-test-outcome-{datetime}.md`
**PR Summary:** Include a one-line summary suitable for PR descriptions:
> "QA found N issues, fixed M, health score X → Y."
**Debug Escalation Summary** (include only if any issues were escalated in Phase 8g):
```
DEBUG ESCALATION
Issues escalated: N
Resolved (DONE): X (commit SHAs)
Concerns: Y
Blocked: Z (root cause unclear)
Unavailable: W (agent call failed)
Per-issue details:
ISSUE-NNN: [symptom] → [root cause found / BLOCKED]
Debug report: [inline summary or full report]
```
---
## Phase 11: TODOS.md Update
+50
View File
@@ -17,6 +17,7 @@ allowed-tools:
- Grep
- AskUserQuestion
- WebSearch
- Agent
---
{{PREAMBLE}}
@@ -245,6 +246,41 @@ WTF-LIKELIHOOD:
**Hard cap: 50 fixes.** After 50 fixes, stop regardless of remaining issues.
### 8g. Debug Escalation
After the fix loop completes (all issues processed through 8a-8f), check for issues that were **reverted at least twice** (two or more separate fix attempts each caused regressions and were reverted).
If no issues match, skip this phase entirely.
For each matching issue, sequentially spawn a debug sub-agent using the Agent tool:
**Agent prompt — structured handoff:**
> Read `~/.claude/skills/gstack/debug/SKILL.md` (fallback: `.claude/skills/debug/SKILL.md`) and follow the systematic debugging methodology.
>
> **Bug Brief:**
> - Issue ID: ISSUE-NNN
> - Symptom: [what was observed during QA testing]
> - Severity: [critical/high/medium/low]
> - Affected URL: [page URL where bug was found]
> - Reproduction: [step-by-step repro from QA testing]
> - Screenshot evidence: [paths to before/after screenshots from QA]
> - Failed fix attempts:
> 1. [what was changed] → [what regression it caused] (reverted)
> 2. [what was changed] → [what regression it caused] (reverted)
> - Files investigated: [list of source files already examined during fix attempts]
> - Console errors: [relevant JS errors if any]
>
> Investigate this bug. You have full access to AskUserQuestion if you need user input. Output the DEBUG REPORT when done.
**When the debug agent returns:**
- **DONE status:** The fix is in the working tree. Re-test the affected page via browse. If verified, commit with `fix(qa/debug): ISSUE-NNN — [description]`. If regression detected, revert and mark as "deferred (debug-investigated, fix regressed)."
- **DONE_WITH_CONCERNS:** Apply the fix, note concerns in report. Re-test.
- **BLOCKED:** Run `git checkout .` to discard any uncommitted debug artifacts (temp logs, assertions). Mark as "deferred (debug-investigated, root cause unclear)." Include the debug report in Phase 10 output.
- **Agent call fails:** Mark as "deferred (debug unavailable)." Continue to Phase 9.
**Sequencing:** Process one debug issue at a time. Wait for each agent to complete before spawning the next. The working tree must be clean between debug investigations.
---
## Phase 9: Final QA
@@ -285,6 +321,20 @@ Write to `~/.gstack/projects/{slug}/{user}-{branch}-test-outcome-{datetime}.md`
**PR Summary:** Include a one-line summary suitable for PR descriptions:
> "QA found N issues, fixed M, health score X → Y."
**Debug Escalation Summary** (include only if any issues were escalated in Phase 8g):
```
DEBUG ESCALATION
Issues escalated: N
Resolved (DONE): X (commit SHAs)
Concerns: Y
Blocked: Z (root cause unclear)
Unavailable: W (agent call failed)
Per-issue details:
ISSUE-NNN: [symptom] → [root cause found / BLOCKED]
Debug report: [inline summary or full report]
```
---
## Phase 11: TODOS.md Update