triage
A Claude Code skill that triages a batch of raw security-scanner findings:
verifies each is real, collapses duplicates, re-ranks by derived
exploitability, and tags each survivor with a component owner. Turns a raw
dump into a short, ranked, owned list. Pairs with /vuln-scan (which
generates findings) and this repo's vuln-pipeline (which generates crash
reports), but also ingests loosely-structured JSON or markdown from other
scanners.
Requirements
- Claude Code CLI installed and authenticated
- A read-only checkout of the target codebase (verification reads source; it does not build or run anything)
- A file or directory of findings to triage
Installation
Project-scoped (ships with this repo — nothing to do if you cloned it):
ls .claude/skills/triage/SKILL.mdOr user-scoped:
mkdir -p ~/.claude/skills
cp -r .claude/skills/triage ~/.claude/skills/Usage
From a Claude Code session in the target repo:
/triage path/to/findings.jsonInteractive mode (the default) opens with a short interview: trust boundary, threat model, scoring standard (HIGH/MED/LOW vs. CVSS vs. your org bug-bar), and whether to bias toward precision or recall on split votes. These answers shape how reachability is judged and how severity is labeled. If you pick CVSS, have the model emit vectors and compute the scores with a tool — see docs/best-practices.md#cvss. To skip the interview and use precision-biased defaults:
/triage path/to/findings.json --autoCommon invocations:
/triage VULN-FINDINGS.json # vuln-scan output, repo = cwd
/triage results/mytarget/2026-04-14/ --repo . # vuln-pipeline output
/triage scanner_export/ --votes 5 --repo ~/src/app # high-stakes batch, 5-vote verify
/triage backlog.md --auto --votes 1 # quick first pass on a markdown reportOutput
./TRIAGE.json— every input finding, annotated withverdict,verify_verdict, recomputedseverity,severity_alignmentvs. the scanner's claim,preconditions,vote_breakdown,rationaleciting file:line evidence,owner_hint, andduplicate_ofwhere applicable. Sorted by what to act on first../TRIAGE.md— reviewer-facing report: an "Act on these" section with one entry per confirmed finding, then a "Dropped" table explaining every rejection.
A needs_manual_test verdict means static reasoning hit its limit on that
finding — treat it as a recommendation for a human to build a controlled
proof-of-concept, not as a failure.
Checkpointing and resume
Per-phase state is written to ./.triage-state/. If a run is interrupted
(rate limit, context exhaustion, Ctrl-C), re-invoking /triage with the same
arguments resumes from the last completed phase — the interview is not
re-asked and verifiers already tallied are not re-spawned. Pass --fresh to
start over. ./.triage-state/ is scratch; add it to .gitignore.
What it does and doesn't do
- Does: read source, grep for callers, reason about reachability and protections, vote, rank, and route.
- Does not: build, run, or test the target; install dependencies; reach the network; write proof-of-concept exploits. All conclusions are static. This is deliberate — the skill is meant to run in a review box alongside a read-only checkout.
Questions
Reach out to your Anthropic contact.