| name | investigation-report |
| description | Standardized "investigate → written report" pipeline for SRE client work. Fixes the deliverable contract up front (format + done-criteria), delegates repo/infra scanning to subagents, verifies every claim, and outputs an issue-ready Markdown report. Use when the user says "調査して", "調べてレポート", "investigation report", "報告書にまとめて", or hands over an investigation-shaped task. |
| user-invocable | true |
| allowed-tools | Read, Glob, Grep, Bash, Agent, Write, WebSearch, WebFetch |
| argument-hint | <question or target> [--out <file>] [--depth quick|thorough] |
investigation-report
Investigation requests are the most frequent work type (docs/debugging/PR-analysis ≈ 60% of sessions) and the main source of misunderstood_request friction. This skill removes the ambiguity by fixing the contract BEFORE any work, then running hands-off.
Phase 0 — Contract (one round max)
From $ARGUMENTS and context, fill this contract. Ask the user ONE batched question only for fields you genuinely cannot infer; otherwise state your assumption and proceed:
- Question: the exact question the report must answer
- Audience: customer-facing / internal / self
- Deliverable: Markdown report (default), destination file (default
./reports/YYYY-MM-DD-<slug>.md, or --out)
- Done-criteria: what makes this answerable ("root cause identified or 3 hypotheses ranked", "migration steps listed with risks", ...)
- Depth:
quick (single-pass, ~1 subagent) / thorough (default; parallel subagents + verification pass)
Echo the contract in 3 lines, then start. Do not wait for approval unless the user is clearly present and the task is client-facing.
Phase 1 — Collect (delegate, don't read inline)
- Spawn
Explore/general-purpose subagents for repo scanning, gh history, cloud CLI reads. Parallelize independent angles (code, config, CI history, docs, external references).
- The main context receives conclusions only, not file dumps.
Phase 2 — Verify (mandatory)
Per the user's global policy: no technical claim without a citation or a check.
- Every finding gets a source:
file:line, command output, or doc URL.
- Claims that could not be verified are labeled unverified hypothesis — never silently promoted to fact.
- For
thorough: spawn one adversarial subagent to refute the top conclusion before writing.
Phase 3 — Report
Structure (Japanese output for the user, this order, no extras):
# <Question>
## 結論 — 1-3 sentences, answers the question directly
## 根拠 — findings with citations
## 推奨アクション — ranked, with effort estimates
## 未検証・残課題 — hypotheses and follow-ups
## 参考 — commands run, sources
Write to the destination file. Final message: 結論 + file path + (if client repo) suggest /sreaas:task report.
