investigate
Investigation flow. Takes a question about the system (why does X happen, where does Y live, is Z feasible, what broke) and returns a verified written answer with evidence. Read-only end to end, no br
SKILL.mdskills/investigate/SKILL.md
GitHubinvestigate: question in, verified answer out
Read-only, end to end. If the fix becomes obvious mid-flight, the finding becomes the spec for the task funnel; you do not edit files here. An investigation that quietly turns into a change skips triage, tests and review, which is every gate the funnel has.
Input: the question after the command. Empty, ask what to investigate and stop.
Stage 1 · Sharpen the question
Restate the question and state what evidence would settle it. If you cannot say what would settle it, ask and stop. Fanning out on a vague question buys file dumps, not an answer.
Stage 2 · Fan out, read-only, in parallel, in one message
Each subagent gets the full question, because it inherits nothing, plus one angle:
- code path, where it lives, who calls it, what it returns
- data, schema, migrations, what the rows actually allow
- history,
git logandgit blame, the tracker, why it is the way it is - runtime, config, environment, scheduled jobs, webhooks, external services
Skip the angles the question obviously does not touch. Two is a normal number,
and the concurrency ceiling is kit config get .maxParallelAgents (with no kit on
PATH, ${CLAUDE_PLUGIN_ROOT}/scripts/kit).
Every claim a subagent returns is tagged MEASURED, with the command and its output, or INFERRED.
Stage 3 · Verify
An INFERRED claim that bears on the conclusion is measured before the report: run, or dispatch, the one command that would prove it. Two angles disagreeing is resolved by measuring, never by picking the more plausible one. An inherited claim is not a fact.
Stage 4 · Deliver
- The answer in short prose, conclusion first, evidence after it.
- Tied to an existing issue, post it as a comment on that issue, in the language the rest of the tracker uses.
- Out-of-scope findings discovered along the way become new issues, the way the repo's pipeline document says to open them. Never code.
- If the fix is then asked for, hand the finding to the task funnel as its spec.
Token rules
- No worktree and no dependency install. Nothing here writes, so nothing here installs.
- You hold conclusions, not file dumps. Subagents read, you consolidate.
- One MEASURED claim beats a paragraph of maybes. The command's output is cheaper than a wrong conclusion re-litigated across the session.