Agent Skills: Investigation Loop

Default output: return only the result, blockers, and required evidence. Omit preambles, process narration, repeated context, confidence scores, and follow-up offers. Use at most five bullets unless a required artifact or schema needs more.

Investigation Loop

Long-form investigation skill for tickets where the answer requires walking multiple linked tickets, code paths, and Confluence pages. Hand the agent a ticket and a marker file; the cooking Stop hook keeps it iterating until either (a) a written hypothesis with at least three evidence citations exists, or (b) the iteration counter hits 10.

When to Use

• A Jira (or similar) ticket whose root cause is non-obvious and requires reading code, related tickets, and historical decisions
• The investigation will take 30+ minutes and you don't want to babysit it
• You need a written hypothesis at the end, not just a fix — investigation is the deliverable

When NOT to Use

• The fix is obvious — just do it; don't loop
• The work is mechanical (test fixing, refactor) — use test-fix-loop or cross-repo-refactor
• You don't have access to the systems involved (Jira, repo, etc.) — the loop will spin without progress

How to Run

1. Decide on the target. Have the ticket key (e.g. RF-9421) and any obvious entry point (a stack trace, a failing PR, a recent regression).
2. Create the marker file with the agent's PID:

   mkdir -p ~/.claude/cooking
   PID=$$
   START_EPOCH=$(ps -p $PID -o lstart= | sed 's/^ *//;s/ *$//')
   printf '{"start_epoch": "%s"}' "$START_EPOCH" > ~/.claude/cooking/active-$PID
   

3. Tell Claude: "Investigate JIRA <TICKET-KEY> in cook mode. Goal: written hypothesis with at least 3 evidence citations."
4. The agent reads the ticket, walks linked tickets and the relevant code, and either:
   • Finds enough — writes the hypothesis, removes the marker, exits
   • Needs another cycle — leaves the marker; the Stop hook returns exit 2 and the agent continues from where it left off
5. The skill must remove the marker file in a finally-equivalent path.

TERMINATION

Stop when ALL of these are true:

• A written hypothesis exists in the conversation answering the ticket's central question
• At least 3 evidence citations support the hypothesis (each citation must be a concrete reference: code at path:line, ticket key, doc URL, commit SHA, or test result)
• The agent has explicitly stated "I have enough to write up findings" (verbatim trigger string), OR the iteration counter has reached 10
• The marker file is removed

When termination triggers:

rm -f ~/.claude/cooking/active-$PID ~/.claude/cooking/iter-$PID

If iteration 10 hits without a hypothesis, write a brief note describing what's missing (which file/ticket couldn't be accessed, what evidence is incomplete) and exit anyway — never run past 10. The hook hard-caps regardless.

Anti-patterns

• Don't reach a hypothesis and immediately try to fix the bug — this skill's deliverable is the hypothesis. Hand off to a normal coding session for the fix.
• Don't count "I think it's probably X" as an evidence citation — citations must reference something concrete the user can verify.
• Don't silently skip a citation requirement to terminate early — if you can't find 3 sources, hit iteration 10 honestly and exit with a "what's missing" note.
• Don't write the hypothesis to a file in the repo — the deliverable lives in the conversation transcript. The user reads it directly.

Verification

/cook-status should show this loop's marker as active mid-run. After termination, the marker is gone and the conversation transcript contains the hypothesis with its citations. The iteration counter file should be ≤ 10 at any observation point.

Related

• Stop hook: hooks/cooking/stop-hook.sh
• Status command: /cook-status
• Sibling skills: test-fix-loop, cross-repo-refactor, dreaming-consolidator