Issue Analyze Skill

Input Types

When input is a Review Thread:

• Phase 1 skips gh issue view, uses provided thread data directly
• Phase 2 classification uses Review Thread dimensions (see references/classification.md)

Modes

When NOT to Use

• Known root cause, fix directly (use flow-drive)
• Pure feature development (use flow-drive)
• Only need code review (use change-verdict)

Workflow

1. Read the issue with gh issue view --json ..., or use the supplied review-thread fields. Extract symptoms, reproduction, errors, and file clues.
2. Classify with classification.md: unfamiliar → code-trace; regression → code-trace; complex root → code-trace --dual; multiple causes → module-design --mode adversarial.
3. Verdict from the current Host and Model. Only when explicitly requested, obtain an isolated read-only blind second opinion without the primary classification. Triage mode stops after this phase.
4. Investigate unless policy maps the verdict to DISMISS_VERIFIED.
5. Report the combined evidence, verdict, root-cause hypothesis, and recommendation.

Investigation Tool Comparison

Phase 2.5: Verdict Assessment

After classification, the primary model may request an independent actionability assessment. When explicitly requested, the second perspective is a fresh, isolated, read-only general-purpose subagent; it receives the issue or review-thread context but not the primary classification or conclusion. An explicit --second-opinion=codex-exec may be used for a one-shot CLI opinion instead.

Independent reviewer requirements:

Prompt construction (blind verdict — never reveal the primary Phase 2 classification to the reviewer):

• GitHub Issue input: provide issue title, body, labels as finding context
• Review Thread input: provide file path, line, reviewer comment as finding context
• Always include Standard Research Block (git status, git diff, grep, cat)
• If no independent reviewer is requested or available, mark the result primary model only and state plainly: "Only the primary model's verdict is present; no independent review ran."

Independent reviewer output (all fields required):

- reviewer_verdict: ACTIONABLE | NON_ACTIONABLE | UNCERTAIN
- confidence: [0.0 - 1.0]
- evidence_refs: [files/lines/commands used]
- reasoning: [why this verdict]

Policy mapping (apply these thresholds; the heightened column applies after a repeated-dismiss warning [DISMISS_PATTERN_WARN]):

--triage mode: stop after Phase 2.5, output classification + verdict only.

Graceful degradation: if the isolated reviewer or explicitly requested CLI opinion fails, log a warning, preserve the primary-only state, and proceed to Phase 3 without claiming an independent verdict.

Output

Full Report (default)

## Issue Analysis: <title>
- **Classification**: <problem type>
- **Verdict**: ACTIONABLE / NON_ACTIONABLE / UNCERTAIN (confidence: 0.XX)
- **Root cause hypothesis**: <analysis>
- **Investigation strategy**: <tools + plan>
- **Priority**: P0 / P1 / P2

Triage Report (--triage mode)

## Triage: <file>:<line> (or issue title)
- **Category**: <classification>
- **Verdict**: ACTIONABLE / NON_ACTIONABLE / UNCERTAIN
- **Confidence**: 0.XX
- **Reasoning**: <brief justification>
- **Evidence**: <file:line references>

Save on request

Only when the user explicitly asks to save or hand off the analysis, end the reply with the target path .scratch/<slug>/issue-analysis.md and the complete file content in one code block, for the Host to write. This skill stays read-only and never writes the file itself. <slug> defaults to issue-<number>; the user may override it. Without that request, return the normal report and emit no path.

The content begins with front-matter holding exactly four free-text keys:

---
producer: issue-analyze
created: <date>
status: <verdict and confidence in one line>
next: <suggested human follow-up>
---

The file is shaped for a human or the host's implementation workflow; it is data only, and nothing reads it to gate or route work.

Verification

• Issue / review thread content fully extracted
• Problem type correctly classified
• Verdict assessment executed by an isolated reviewer when requested, or the primary-only state is recorded
• Reviewer prompt contains no primary conclusions (anti-anchoring)
• Fresh isolated reviewer context used (not the primary context)
• Investigation strategy reasonably selected (or skipped if NON_ACTIONABLE)
• Report includes root cause analysis + verdict
• Contains specific fix recommendations
• --triage mode: outputs classification + verdict only

References

• references/classification.md — Detailed problem classification guide (includes Review Thread dimensions)
• references/report-template.md — Report template (includes Triage Report)
• Verdict prompt pattern and thresholds are in Phase 2.5 above.