| name | clawpatch-bugs |
| description | Run clawpatch, triage findings, write a prioritized bug list. Trigger on "run clawpatch", "clawpatch bug hunt", "scan for bugs". |
| disable-model-invocation | true |
clawpatch bug hunt
Two-phase bug discovery flow; the deliverable is a curated markdown file under docs/bugs/:
- Discover — let
clawpatch produce a raw bug report.
- Triage — analyze each finding with parallel subagents, drop false positives / low-value noise, and save a curated, priority-sorted list.
The user already has clawpatch installed on PATH. Do not try to install it.
Instructions
Step 0 — Reset clawpatch workspace
Always start clean so a previous partial run does not contaminate this one:
rm -rf .clawpatch
Step 1 — Run clawpatch and capture the raw report
Compute today's date as YYYY-MM-DD (use date +%Y-%m-%d). Use it for the output filename. Ensure docs/bugs/ exists.
DATE=$(date +%Y-%m-%d)
mkdir -p docs/bugs
clawpatch init
clawpatch map
clawpatch review --limit 20
clawpatch report > docs/bugs/${DATE}-bugs.md
Notes:
clawpatch review --limit 20 may take a while; run it in the foreground and surface a single status line ("clawpatch reviewing…") rather than streaming its output.- If any of the four
clawpatch commands exits non-zero, stop and show the user the failing command's stderr. Do not proceed to step 2 on a partial report.
- After the report is written, tell the user the path and a one-line summary (count of findings).
Step 2 — Analyze findings to filter out non-bugs
Read docs/bugs/${DATE}-bugs.md and parse out each finding (title + file refs + clawpatch's rationale).
For each finding, spawn an Explore (or general-purpose) subagent in parallel with a self-contained prompt:
- Hand it the finding's title, the cited file(s)/line(s), and clawpatch's claim verbatim.
- Tell it the goal: "Read the cited code in this repo and decide whether this is a real, currently-shipping bug worth filing — or a false positive / stylistic nit / already-handled case."
- Ask for a structured short reply:
VERDICT: real-bug | not-a-bug | low-value followed by 2–4 sentences of evidence with path:line references.
Batch the subagent spawns: send them all in a single message (one Agent tool call per finding, all in parallel). Wait for all to return before continuing.
Aggregate the verdicts:
- Keep findings whose verdict is
real-bug.
- Drop
not-a-bug and low-value outright. Briefly note in the curated file how many were dropped and why (one sentence summary per dropped item is enough).
Sort the kept bugs by priority before writing the file. Use three buckets:
- P0 — ship-blocking: data loss, crashes on a common code path, runaway resource use, fundamental breakage of a documented flow, destructive TOCTOU.
- P1 — real user-facing bug, recoverable: bad UX, silent drops, broken automation, contract violations that affect real users.
- P2 — polish / defense-in-depth / edge case / low-impact perf.
Within each bucket, order by your own judgment of impact (highest first). Renumber the kept bugs 1..N in this final priority order (do not preserve clawpatch's emission order).
Write the curated list to:
docs/bugs/${DATE}-bugs-serious.md
Use this structure:
# Serious bugs — <DATE>
Source: `docs/bugs/<DATE>-bugs.md` (clawpatch review, limit 20, <RAW_COUNT> raw findings).
Filtered by parallel subagent triage on <DATE> (one Explore agent per finding).
Kept <N> real bugs, sorted by priority (P0 → P1 → P2). Dropped <M>.
Priority guide:
- **P0** — ship-blocking: data loss, crashes, runaway resource use, fundamental breakage.
- **P1** — real user-facing bug, recoverable: bad UX, silent drops, broken automation.
- **P2** — polish / defense-in-depth / edge case / low-impact perf.
## P0 (<count>)
### 1. <short title>
- **Where:** `path/to/file.ts:LINE`
- **What clawpatch said:** <one-line summary>
- **Why it is real:** <subagent evidence, 2–4 sentences>
- **Suggested fix sketch:** <one or two sentences>
(repeat per finding, continuing the numbering across buckets)
## P1 (<count>)
### <next-number>. <short title>
...
## P2 (<count>)
### <next-number>. <short title>
...
## Dropped (<M>)
- <title> — <one-sentence reason it was dropped>
Tell the user where the file lives, the keep/drop counts, and the P0/P1/P2 split. The curated markdown is the deliverable — stop here.
Guardrails
- Never edit application code in this flow. This skill only writes files under
docs/bugs/.
- Do not skip Step 0. A stale
.clawpatch/ directory silently changes review/report behavior.
- Do not run
clawpatch against unrelated directories. Always invoke from the repo root.
- Subagent prompts must be self-contained — they cannot see this conversation. Paste the finding text into each prompt verbatim.
- If
docs/bugs/${DATE}-bugs.md already exists from an earlier run today, overwrite it (the Step 0 rm -rf .clawpatch plus a fresh clawpatch report > redirect already implies this — don't preserve the old file).
