crash-analysis

name	crash_analysis
description	Analyze reproduced crash evidence and produce a structured verdict for workflow routing.
compatibility	opencode
metadata	{"stage":"crash-analysis","owner":"tianheng"}

What this skill does

This skill performs analysis-only crash verdicting after repro, without patching source code.

When to use this skill

Use this skill in the crash-analysis stage after re-run has produced crash evidence.

Required inputs

• crash_info.md
• re_run_report.md
• crash_triage.json (if present)
• coordinator hint/context

Required outputs

• crash_analysis.json
• crash_analysis.md

crash_analysis.json must follow this minimal shape:

{
  "verdict": "false_positive|real_bug|unknown",
  "reason": "short explanation",
  "confidence": 0.0,
  "signals": ["concrete log lines or findings"]
}

Workflow

1. Read crash and repro artifacts first.
2. Correlate triage label with concrete log evidence.
3. Write crash_analysis.json and crash_analysis.md.
4. Do not modify source files in this stage.

Constraints

• Analysis-only stage; no code edits.
• Prefer explicit evidence lines over generic conclusions.
• Use conservative verdict unknown when evidence is weak.
• Do not classify real_bug from sanitizer keywords alone.
• If evidence is weak or missing, output unknown and explain missing evidence explicitly.

Command policy

• Allowed: read-only commands (find, grep, rg, cat, ls, sed, awk, head, tail).
• Forbidden: build or execution commands.

Acceptance checklist

• verdict is exactly one of false_positive, real_bug, unknown.
• reason is non-empty and grounded in observed evidence.
• signals is a non-empty string array with concrete findings.
• Output remains analysis-only.

Done contract

• Create ./done.
• Write the path string crash_analysis.json as the sole text of ./done (run echo 'crash_analysis.json' > ./done; do not copy the file's contents).