Triage

triage

Adversarial triage of raw security-scanner output. Does four jobs:

verify each finding is real, deduplicate across runs and scanners,

rank survivors by derived exploitability rather than the scanner's

claimed severity, and route each to a component owner. Output is a

short, ranked, owned list instead of a raw dump.

Invoke with /triage <findings-path> [--auto] [--votes N] [--repo PATH] [--fp-rules FILE].

Arguments (parse from $ARGUMENTS; positional $1/$2 expansion is

not stable across runtimes):

• findings path (first positional, required): a JSON file, a directory of

JSON files, a VULN-FINDINGS.json, an INCIDENTS.json, a pipeline

results/<target>/<ts>/ directory, or a markdown report.

• --auto: skip the interview and use defaults. Default mode is

interactive.

• --votes N: verifier votes per finding (default 3; use 1 for a quick

pass, 5 for high-stakes batches).

• --repo PATH: path to the target codebase, read-only (default cwd).

Verification needs source access; the skill stops with an error if the

cited files aren't reachable.

• --fp-rules FILE: append the contents of FILE to the verifier's

exclusion-rule list (Phase 3a). Use for org-specific precedents: "we use

Prisma ORM everywhere — raw-query SQLi only", "k8s resource limits cover

DoS", etc. Plain text, one rule per line or paragraph.

• --fresh: ignore any existing checkpoint in ./.triage-state/ and start

from Phase 0. Without this flag the skill resumes from the last completed

phase if a checkpoint is present.

Tools: Read, Glob, Grep, Write, Task, AskUserQuestion. Bash is

permitted only for git, find, wc, ls, jq, and

python3 .claude/skills/_lib/checkpoint.py (checkpoint I/O).

Do not execute target code. No building, running, installing

dependencies, or sending requests. A proof-of-concept that accidentally

works against something real is unacceptable, and "couldn't write a working

PoC" is weak evidence of non-exploitability. Every conclusion comes from

reading source. This applies to the orchestrator and every subagent;

include the constraint in every Task prompt. For high-confidence HIGH

findings, recommend a human-built PoC as a follow-up instead.

Do not reach the network. No package-registry lookups, CVE-database

queries, or upstream-commit fetches.

Checkpointing (runs before Phase 0 and after every phase)

On large finding batches a full run can exhaust context or hit rate limits

mid-way — particularly Phase 3, which spawns candidates × votes verifiers.

Phase state persists to ./.triage-state/ so a fresh /triage session can

resume without re-asking the interview or re-spawning verifiers.

All checkpoint I/O goes through python3 .claude/skills/_lib/checkpoint.py

(atomic writes, JSON-validated). Never use the Write tool for progress.json

directly. Never pass payload via heredoc or stdin; target-derived strings

could collide with the heredoc delimiter and break out to shell. The

Write→--from pattern keeps repo-derived bytes out of Bash argv.

State files in ./.triage-state/:

• progress.json — single source of truth for resume position:

{"status": "running"|"complete", "phase_done": N, "shards_done": [...]}.

Resume decisions read ONLY this file, never a glob of phase*.json or

shard files (stale files from a prior run must not be trusted).

• phaseN.json — data payload for phase N (schemas at the tail of each phase

section below).

• _chunk.tmp — transient payload buffer; overwritten before every

save/shard/append call.

Start of run — resume check. Bash:

python3 .claude/skills/_lib/checkpoint.py load ./.triage-state

• status == "absent" OR "complete", OR --fresh in $ARGUMENTS →

fresh start. Bash:

python3 .claude/skills/_lib/checkpoint.py reset ./.triage-state,

then proceed to Phase 0.

• status == "running" with phase_done == N → resume. Read

./.triage-state/phase0.json through phaseN.json in order (and any

shard_*.json files listed in shards_done), merging keys into working

state (later files override earlier — checkpoints may be deltas). Print

Resuming from checkpoint: Phase N complete (./.triage-state/phaseN.json),

and skip directly to Phase N+1.

End of every phase N. Two tool calls:

• Write tool → ./.triage-state/_chunk.tmp containing the phase's output

JSON (schema at the tail of each phase section).

• Bash → python3 .claude/skills/_lib/checkpoint.py save ./.triage-state <N> <name> --from ./.triage-state/_chunk.tmp

End of run. After writing TRIAGE.json and TRIAGE.md, Bash:

python3 .claude/skills/_lib/checkpoint.py done ./.triage-state 6

Phase 0: Mode select and interview

0a. Parse arguments

From $ARGUMENTS: extract the findings path (first positional), --auto

flag, --votes N (default 3), --repo PATH (default .), `--fp-rules

FILE` (default none). If no findings path was given, ask for one and stop.

If --fp-rules was given, Read the file now and carry its contents as

context.extra_fp_rules for injection into the Phase 3a verifier prompt.

0b. Interactive mode (default): interview the user

Unless --auto was passed, use AskUserQuestion to gather context that

shapes verification and ranking. Batch into one or two calls of up to four

questions. Expect free-text answers via "Other"; the multiple-choice options

are prompts, not constraints.

Round 1 (single AskUserQuestion call):

• Environment & trust boundary (header Environment, single-select)

`What kind of system are these findings from, and where does untrusted

input enter it?`

Options: Internet-facing web service (HTTP is untrusted),

Internal service (callers are authenticated peers),

Library / SDK (caller is the trust boundary),

CLI / batch tool (operator inputs trusted, file inputs not),

Embedded / firmware (physical access in scope).

Reachability is judged against this boundary; "command injection from env

var" is a true positive in a multi-tenant web service and a rule-8 false

positive in an operator CLI.

• Threat model (header Threat model, multi-select)

`What does a worst-case attacker look like for this system, and what

must never happen? Free text is best.`

Options: Unauthenticated remote code execution,

Tenant-to-tenant data leakage, Privilege escalation to admin,

Supply-chain compromise of downstream users,

Denial of service against a paid SLA,

Compliance-scoped data exposure (PII / PCI / PHI).

Phase 4 boosts findings that map onto a stated threat.

• Scoring standard (header Scoring, single-select)

How should severity be expressed in the output?

Options: Derived HIGH/MEDIUM/LOW from preconditions (default),

CVSS v3.1 vector + base score, CVSS v4.0 vector + base score,

OWASP Risk Rating (likelihood x impact),

Organization bug-bar (describe in Other).

The precondition rule is always computed; this controls what

severity_label additionally shows.

• Noise tolerance (header Noise tolerance, single-select)

When verifiers disagree, which way should ties break?

Options:

Precision: drop anything not majority-confirmed (fewer FPs, may miss real bugs),

Recall: keep split votes as needs_manual_test (more to review, fewer misses),

Ask me per-finding when it happens.

Round 2 (conditional): if the threat-model answer was empty or generic,

or the scoring answer was Organization bug-bar, ask one targeted follow-up.

Record the answers as a context dict carried through every phase and

echoed in the output under triage_context.

0c. Auto mode defaults

When --auto is set, do not call AskUserQuestion. Use:

• Environment: `Unknown. Treat any externally-reachable entry point as

untrusted; flag trust-boundary assumptions explicitly in rationale.`

• Threat model: empty (no boost).
• Scoring: derived HIGH/MEDIUM/LOW.
• Noise tolerance: precision.

Checkpoint: Write tool → ./.triage-state/_chunk.tmp:

{"phase": 0, "context": {mode, environment, threat_model, scoring, noise_tolerance, votes_per_finding, repo, findings_path}}

Then Bash:

python3 .claude/skills/_lib/checkpoint.py save ./.triage-state 0 interview --from ./.triage-state/_chunk.tmp

On resume past Phase 0, the interview is not re-asked; context is

restored from this file.

Phase 1: Ingest and normalize

Turn the input into a flat findings[] list with stable ids, regardless of

source format.

1a. Detect input shape

Inspect the findings path:

• Directory: Glob for /*.json and /*.jsonl. Recognized

containers, in priority order:

• VULN-FINDINGS.json (a {findings: [...]} container): read

.findings[].

• INCIDENTS.json (the /dnr-hunt / /dnr-respond output): read

.incidents[]; each incident is one finding. Map the nested fields —

vuln.file → file, vuln.line → line, vuln.type → category,

vuln.function → function, description ← vuln.mechanism plus

impact — and keep id, title, verdict, and poc; the 1b

aliases route poc → exploit_scenario, and a confirmed_exploited

verdict means the incident arrived execution-verified, not statically

guessed — verifiers should weight that. The top-level ruled_out

array is not findings — don't ingest it. Incidents with no

vuln.file follow the 1b missing-file rule like any other finding.

• reports/bug_*/report.json or reports/manifest.jsonl (this repo's

pipeline output): one finding per bug_NN. Map crash.crash_type →

category, verdict.severity_rating → severity, the prose report →

description, crash file from the ASAN top frame → file/line.

• found_bugs.jsonl: one finding per line.
• Any other *.json whose top level is a list of objects, or an object

with a findings/results/issues/vulnerabilities array: that

array.

• Single .json / .jsonl file: same recognition as above.
• Markdown / text: split on level-2/3 headings or --- rules; for each

section, extract file, line, category, severity, description by

pattern (File:, Line:, Severity: labels or path:NN spans).

Best-effort; mark source_format: "markdown_heuristic".

If nothing parseable is found, stop and report what was seen.

1b. Normalize fields

For each raw record, build a finding dict. **Pull what's present; never

guess what's absent.** Field map (source-key aliases → canonical):

| Canonical | Also accept |

|-----------------|----------------------------------------------------------|

| file | path, location.file, filename, ASAN top-frame file |

| line | line_number, location.line, lineno |

| category | type, cwe, rule_id, crash_type, vulnerability_class |

| severity | severity_rating, level, priority, risk |

| title | name, summary, message |

| description | details, report, body, evidence |

| exploit_scenario | attack_scenario, poc, reproduction |

| preconditions | requirements, assumptions |

| recommendation| fix, remediation, mitigation |

| scanner_confidence | confidence, score, certainty (normalize to 0.0-1.0) |

Attach to every finding:

• id: f001, f002, ... in ingest order. If scanner_confidence is

present on most findings, order ingest by it descending so high-signal

findings get verified (and surface in partial output) first; otherwise

keep source order. This is a scheduling prior only — it does not affect

verdicts.

• source: relative path of the file it came from, plus source format.
• missing_fields: list of canonical fields that were absent. If file is

missing or does not resolve under --repo, the finding is

unlocatable: it skips dedup and verification and is emitted directly

with verdict: false_positive, verify_verdict: needs_manual_test,

confidence: 0, refute_reasons: ["doesnt_exist"], `rationale: "no

source location in input; cannot verify statically; human review

required"`. Never emit a confident verdict on a finding you could not

locate, and never let it absorb or be absorbed by dedup.

1c. Locate the target codebase

Resolve --repo (default cwd). For the first 5 findings with a file,

check the path resolves under the repo. Try, in order: (a) repo/file

as-given; (b) file as an absolute or cwd-relative path; (c) repo/file

with common prefixes stripped from file (src/, app/, ./, or the

repo's own basename, e.g. harness/grade.py with --repo harness).

Record which resolution worked and apply it to every finding. If none

resolve, stop: tell the user verification needs source access and the

cited files aren't reachable, and suggest a --repo value based on the

longest common suffix you can see.

Checkpoint: Write tool → ./.triage-state/_chunk.tmp:

{"phase": 1, "context": {...}, "findings": [ {normalized finding dicts with id/source/file/line/category/...} ], "path_resolution": "<which of a/b/c worked>"}

Then Bash:

python3 .claude/skills/_lib/checkpoint.py save ./.triage-state 1 ingest --from ./.triage-state/_chunk.tmp

Phase 2: Deduplicate (before verification)

Collapse repeats so duplicate findings don't each burn N verifiers.

2a. Deterministic pass (inline, no subagent)

Cluster findings where all of:

• same file (after path normalization), AND
• same category (case-insensitive, punctuation stripped), AND
• line numbers within 10 of each other. Both-missing matches; one-side-

missing does NOT (a line-less record must not absorb a located one).

Within each cluster, the canonical is the record with the fewest

missing_fields; ties break to lowest id. Every other member gets

verdict: duplicate, duplicate_of: <canonical id>, and is removed from

the working set. Record duplicate ids on the canonical as absorbed: [...].

2b. Semantic pass (one subagent, only if >1 cluster survives)

Spawn ONE Task with subagent_type: "general-purpose" and this prompt:

You are deduplicating security findings before expensive verification. Two
findings are DUPLICATES if fixing one would also fix the other. Two findings
are DISTINCT if they have genuinely independent root causes, even if they
share a category or file.

Treat as DUPLICATE:
- Same root cause described with different wording or by different scanners
- A shared vulnerable helper function reported once per call site
- A missing global protection (auth check, output encoding) reported once
  per endpoint that lacks it
- A cause ("missing input validation on `name`") and its consequence
  ("SQL injection via `name`") in the same code path

Treat as DISTINCT:
- Different categories in the same file region (an "ssrf" near a
  "buffer_overflow" is not a duplicate just because the lines are close)
- Same file, same category, but different tainted variables reaching
  different sinks
- Same helper, but two independent bugs inside it
- Two endpoints missing the same check, where the fix is per-endpoint
  rather than a shared gate

Below are the candidate findings (one per line: id | file:line | category |
title). Group them. Respond with ONLY lines of the form:

  GROUP: <canonical_id> <- <dup_id>, <dup_id>, ...

One line per group that has duplicates. Omit singletons. Pick the most
specific / best-described finding as canonical. No prose.

CANDIDATES:
{one line per surviving finding: "f003 | src/auth.py:112 | sql_injection | User lookup concatenates name into query"}

Parse GROUP: lines. For each, mark the listed dup ids with

verdict: duplicate, duplicate_of: <canonical>, append them to the

canonical's absorbed, and drop them from the working set.

Carry forward candidates[] = the surviving canonicals.

Checkpoint: Write tool → ./.triage-state/_chunk.tmp:

{"phase": 2, "context": {...}, "findings": [ {all findings; duplicates carry verdict/duplicate_of} ], "candidates": ["f001", "f003", "..."]}

Then Bash:

python3 .claude/skills/_lib/checkpoint.py save ./.triage-state 2 dedup --from ./.triage-state/_chunk.tmp

Phase 3: Verify

For each candidate, N independent adversarial verifiers re-derive the claim

from the code and vote. Each verifier's stance is "find any reason this is

wrong." Each starts from the code at the cited location, not the scanner's

description, and never sees the other verifiers' reasoning (shared context

propagates blind spots).

3a. Verifier prompt (assemble once, reuse for every spawn)

You are a skeptical security engineer adversarially verifying ONE finding
from an automated scanner. Your default assumption is that the scanner is
WRONG. Your job is to re-derive the claim from the source code yourself and
decide TRUE_POSITIVE or FALSE_POSITIVE.

You have read-only access to the target codebase at: {REPO_PATH}
You may use Read, Glob, and Grep, but ONLY on paths inside {REPO_PATH}.
Do NOT read, grep, or glob outside that root: anything outside it (the
triage pipeline itself, scanner outputs, fixtures, other repos on disk) is
out of scope and citing it contaminates your verdict. If a finding's
`file` resolves outside {REPO_PATH}, return CANNOT_VERIFY with
REFUTE_REASON: doesnt_exist. You may NOT build, run, or test the target,
install dependencies, or reach the network. Every conclusion must come
from reading source under {REPO_PATH}.

ENVIRONMENT (from the operator; this defines the trust boundary):
{context.environment or "Unknown. Treat any externally-reachable entry point as untrusted."}

────────────────────────────────────────────────────────────────────────
PROCEDURE: follow all four steps. Each exists because skipping it lets a
specific false-positive class through.

1. READ THE CODE AT THE CITED LOCATION YOURSELF.
   Open {file} at line {line}. Understand what the code actually does. Do
   NOT trust the scanner's description: scanners misread code surprisingly
   often, and if you start from the summary you inherit the misreading.

2. TRACE REACHABILITY BACKWARDS FROM THE SINK.
   Grep for callers of this function/method. Follow imports. Establish
   whether attacker-controlled input (per the ENVIRONMENT above) can
   actually reach this line. A plausible-sounding chain is NOT enough: for
   at least the FIRST link in the chain, READ the actual call site and
   QUOTE the file:line in your rationale. Unreachable code is the single
   largest false-positive source.

3. HUNT FOR PROTECTIONS.
   Actively look for reasons the finding is WRONG:
   - Input validation / sanitization upstream of the sink
   - Framework auto-escaping, parameterized queries, prepared statements
   - Type constraints (the value is an int, an enum, a fixed-length token)
   - Authentication / authorization gates before this path
   - Configuration that limits exposure (feature flag off, debug-only)
   - Dead code, test-only code, example/fixture code

4. STRESS-TEST EACH PROTECTION.
   For each protection you found: is it applied on EVERY path to the sink,
   or only the one the scanner happened to trace? Are there encodings,
   edge cases, or alternate entry points that bypass it?

────────────────────────────────────────────────────────────────────────
EXCLUSION RULES: if the finding matches any of these, it is FALSE_POSITIVE
even if technically accurate. Cite the rule number in your verdict.

  1. Volumetric DoS or missing rate-limiting (handled at infrastructure
     layer). ReDoS, algorithmic complexity, and unbounded recursion ARE
     still valid findings.
  2. Test-only code, dead code, example/fixture code, or a crash with no
     security impact.
  3. Behavior that is the intended design (compression middleware, a
     backward-compatible weak algorithm offered alongside a strong one).
  4. Memory-safety concerns in memory-safe languages outside `unsafe` /
     FFI blocks.
  5. SSRF where the attacker controls only the path, not the host or
     protocol.
  6. User input flowing into an AI/LLM prompt (prompt injection is not a
     code vulnerability in the target).
  7. Path traversal in object storage (S3/GCS) where `../` does not escape
     a trust boundary.
  8. Trusted inputs used as the attack vector (env vars, CLI flags set by
     the operator), UNLESS the ENVIRONMENT above marks them untrusted.
  9. Client-side code flagged for server-side vulnerability classes.
 10. Outdated dependency versions (managed by a separate process).
 11. Weak random used for non-security purposes (jitter, shuffling,
     dev-only fallbacks).
 12. Low-impact nuisance issues (log spoofing, CSRF on logout, self-XSS,
     tabnabbing, open redirect, regex injection).
 13. Missing hardening or best-practice gap with no concrete exploit path
     (missing security headers, no audit logging, permissive config that
     isn't actually reached by untrusted input).
 14. XSS in a framework with default auto-escaping (React, Angular, Vue,
     Jinja2 autoescape=on) UNLESS the sink is a raw-HTML escape hatch
     (dangerouslySetInnerHTML, bypassSecurityTrustHtml, v-html, |safe).
 15. Identifiers that are unguessable by construction (UUIDv4, 128-bit+
     random tokens) flagged as "predictable" or "needs validation".
 16. Race conditions or TOCTOU that are theoretical only — no realistic
     window, or no security-relevant state changes between check and use.

{if context.extra_fp_rules: append here verbatim under an
 "ORG-SPECIFIC RULES:" heading}

────────────────────────────────────────────────────────────────────────
VERDICT: your response MUST end with EXACTLY this block:

  VERDICT: TRUE_POSITIVE | FALSE_POSITIVE | CANNOT_VERIFY
  CONFIDENCE: <0-10>
  REFUTE_REASON: <one of: doesnt_exist, already_handled,
    implausible_trigger, intentional_behavior, misread_code, duplicate,
    not_actionable, n/a>
  EXCLUSION_RULE: <1-16, org rule, or none>
  FIRST_LINK: <file:line of the first call site you read, or "none found">
  RATIONALE: <2-5 sentences citing specific file:line evidence for
    reachability, protections found/absent, and why each held or didn't>

TRUE_POSITIVE requires ALL of: path is reachable from untrusted input per
the ENVIRONMENT; protections are insufficient or bypassable; real-world
exploitation is feasible.

FALSE_POSITIVE requires ANY of: unreachable from untrusted input;
adequately protected on all paths; scanner misread the code; an exclusion
rule applies.

CANNOT_VERIFY: static reasoning genuinely hit its limit (e.g. behavior
depends on runtime configuration you cannot read, or the code path crosses
into a binary you cannot inspect). Use sparingly; it must not become the
default.

3b. Spawn N verifiers per candidate, all in one message

For each finding in candidates[], build N Task calls (N = --votes,

default 3) with subagent_type: "general-purpose" and `description:

"verify {id} vote {k}/{N}"`.

Always set subagent_type; never fork. Omitting subagent_type forks

the orchestrator, and a fork inherits the full conversation context: every

other finding's description, the scanner's prose, and any prior verifier

results. That defeats verifier independence and re-introduces the

inherited-framing failure mode this phase exists to prevent. Each verifier

must start with a fresh, empty context and receive only the 3a prompt

plus the single finding under review. The same applies to the ranking

subagents in 4a.

Each prompt is the verifier prompt from 3a with this block appended:

────────────────────────────────────────────────────────────────────────
FINDING UNDER REVIEW (from the scanner; treat as a CLAIM, not a fact):

  id:        {id}
  file:      {file}
  line:      {line}
  category:  {category}
  severity (claimed): {severity}
  title:     {title}

  description:
  {description}

  exploit_scenario:
  {exploit_scenario or "(not provided)"}

  preconditions (claimed):
  {preconditions as bullets or "(not provided)"}

You are vote {k} of {N}. You have NOT seen the other verifiers' reasoning
and you must NOT try to find it. Work independently from the code.

Put all verifier Task calls in a single assistant message so they run

concurrently. Do not set run_in_background; you need the final text, not

an async handle. If len(candidates) * N exceeds ~40, shard into

sequential batches of ~40, but keep each batch a single message.

Prompt size at scale. The 3a prompt is ~1200 words. When

candidates * votes > ~50, use this compact form instead (same procedure

and output contract, prose stripped):

Adversarially verify ONE scanner finding. Default: scanner is WRONG.
Read-only access scoped to {REPO_PATH} ONLY. No exec, no network.
ENVIRONMENT: {context.environment}

Steps: (1) Read {file}:{line} yourself; don't trust the description.
(2) Trace callers backwards; quote the first call-site file:line.
(3) Hunt for protections: validation, escaping, type bounds, auth gates,
dead/test code. (4) Stress-test each protection on every path.

Exclusion rules (FALSE_POSITIVE if matched): 1 volumetric DoS;
2 test/dead/fixture code; 3 intended design; 4 memory-safety in safe
lang outside unsafe/FFI; 5 SSRF path-only; 6 LLM prompt input;
7 object-storage traversal; 8 trusted operator env/CLI inputs;
9 client code, server vuln class; 10 outdated deps; 11 weak random
non-security; 12 low-impact nuisance (log spoof, open redirect, regex
inject); 13 missing-hardening-only, no concrete exploit; 14 XSS in
auto-escape framework w/o raw-HTML escape hatch; 15 unguessable
UUID/token flagged predictable; 16 theoretical-only race/TOCTOU.
{+ org rules from --fp-rules if any}

End with EXACTLY:
  VERDICT: TRUE_POSITIVE | FALSE_POSITIVE | CANNOT_VERIFY
  CONFIDENCE: <0-10>
  REFUTE_REASON: <doesnt_exist|already_handled|implausible_trigger|
    intentional_behavior|misread_code|duplicate|not_actionable|n/a>
  EXCLUSION_RULE: <1-16, org rule, or none>
  FIRST_LINK: <file:line or "none found">
  RATIONALE: <2-5 sentences, file:line cited>

FINDING: {id} {file}:{line} {category} (claimed {severity})
{title}
{description}
Vote {k}/{N}. Independent; do not seek other votes.

Findings with a file but no line get one verifier vote regardless

of --votes (a file-level sweep is expensive and doesn't benefit from

voting).

**If any Task call returns status: "async_launched" instead of the

verifier's text**, the runtime backgrounded it (some runtimes do this

automatically for large parallel batches). Pick one recovery and use it for

the whole batch:

• If completion notifications arrive in your conversation: parse each

verifier's VERDICT block from its notification result as it lands.

Do not end your turn until every vote is accounted for.

• If notifications do not arrive: do not poll transcript files. Re-spawn

the missing verifiers in a fresh Task batch (smaller shard size, e.g.

10) and use the synchronous results.

The same recovery applies to the dedupe subagent in 2b and the ranking

subagents in 4a.

3c. Tally votes

For each candidate, parse the trailing block from each of its N verifiers

(tolerate code fences and whitespace). If a verifier errored, timed out,

or produced no parseable VERDICT block, re-spawn it once. If the retry

also fails, count that vote as cannot_verify with confidence: 0 and

note "verifier_error" in refute_reasons. The remaining N-1 votes still

decide.

Build:

• vote_breakdown: `{"true_positive": x, "false_positive": y,

"cannot_verify": z}`

• confidence: mean CONFIDENCE across votes that agree with the majority,

rounded to one decimal.

• exclusion_rule: the modal EXCLUSION_RULE among FALSE_POSITIVE votes,

else null.

• refute_reasons: sorted unique REFUTE_REASON values from FALSE_POSITIVE

votes.

• first_links: unique FIRST_LINK values across all votes (reachability

audit trail).

• rationale: the RATIONALE from the highest-confidence vote on the

winning side, verbatim.

Decide verdict:

• Majority TRUE_POSITIVE → verdict: true_positive. Proceeds to Phase 4.
• Majority FALSE_POSITIVE → verdict: false_positive. Skips Phase 4.
• No majority (tie, or majority CANNOT_VERIFY):
• Noise tolerance precision → verdict: false_positive; append

"(split vote, dropped under precision policy)" to rationale.

• Noise tolerance recall → verdict: true_positive with

verify_verdict: needs_manual_test. Proceeds to Phase 4.

• Noise tolerance ask → collect all split findings and present them in

one AskUserQuestion call at the end of Phase 3 (header: id + title,

options: keep / drop), then apply the user's choices.

Build confirmed[] = candidates with verdict == true_positive.

Checkpoint: Write tool → ./.triage-state/_chunk.tmp:

{"phase": 3, "context": {...}, "findings": [ {all findings with verdict/vote_breakdown/confidence/refute_reasons/first_links/rationale/exclusion_rule} ], "confirmed": ["f001", "..."]}

Then Bash:

python3 .claude/skills/_lib/checkpoint.py save ./.triage-state 3 verify --from ./.triage-state/_chunk.tmp

This is the most expensive checkpoint. When len(candidates) * votes exceeds

~40 and verifier spawns are sharded into sequential batches, additionally

checkpoint per candidate as its votes are tallied:

• Write tool → ./.triage-state/_chunk.tmp = that finding's post-tally dict.
• Bash:

python3 .claude/skills/_lib/checkpoint.py shard ./.triage-state <id> --from ./.triage-state/_chunk.tmp

On resume at phase_done == 2, the Phase-3 entry point reads

progress.json:shards_done (default [] — do not glob shard files on

disk; stale shards from a prior run may exist), loads the corresponding

shard_{id}.json files, and spawns verifiers only for candidates[] ids

from phase2.json that are NOT in shards_done. Once every candidate is in

shards_done, write the consolidated phase3.json checkpoint as above.

Phase 4: Rank by exploitability (confirmed findings only)

Recompute severity from preconditions and reachability rather than category

name, and judge the scanner's claimed severity separately. Verification and

severity are independent judgments; "this is real" must not inflate into

"this is critical."

4a. Ranking prompt

Spawn one Task per confirmed finding (subagent_type: "general-purpose",

all in one message) with:

You are assigning severity to a CONFIRMED security finding. Verification
already happened; assume the finding is real. Your only job is to derive
how bad it is, independently of what the scanner claimed.

You may Read/Grep the codebase at {REPO_PATH} to check preconditions. Do
NOT execute code.

ENVIRONMENT: {context.environment}
THREAT MODEL (operator-stated, may be empty):
{context.threat_model as bullets, or "(none provided)"}
SCORING STANDARD: {context.scoring}

FINDING:
  id:        {id}
  file:      {file}:{line}
  category:  {category}
  claimed severity: {severity}
  reachability evidence: {first_links from Phase 3}
  verifier rationale: {rationale from Phase 3}

────────────────────────────────────────────────────────────────────────
STEP 1: Enumerate EVERY precondition that must hold for exploitation.
Be concrete: required auth state, configuration, prior request, race
window, attacker position. Then state the minimum ACCESS LEVEL required
(unauthenticated remote / authenticated / local / physical).

STEP 2: Derive severity from the precondition count and access level:

  | Preconditions | Access required          | Severity |
  |---------------|--------------------------|----------|
  | 0             | Unauthenticated remote   | HIGH     |
  | 1-2           | Authenticated            | MEDIUM   |
  | 3+            | Local-only / no demo path| LOW      |

  Evaluate each column independently and take the LOWER result. Example:
  0 preconditions but authenticated-only is MEDIUM, not HIGH; 1
  precondition but local-only is LOW. Cross-check: if your preconditions
  list has 3+ items, HIGH is almost certainly wrong.

STEP 3: Threat-model match. If the THREAT MODEL is non-empty and this
finding maps onto one of its entries, note which one. A match may raise
severity by ONE step (LOW to MEDIUM or MEDIUM to HIGH), never two. If the
threat model is empty, skip this step.

STEP 4: Judge the scanner's claimed severity. From the perspective of an
engineer who has reviewed two hundred scanner findings this week and is
allergic to inflation: would the CLAIMED severity contribute to alert
fatigue? Is it comparable to a real CVE at that level? Is the code in test
fixtures or dev-only config? Score in -5..+5:
  +3..+5  claimed severity is justified or understated
   0..+2  roughly right
  -1..-3  inflated by one level
  -4..-5  badly inflated (LOW dressed as HIGH)

STEP 5: verify_verdict. Exactly one of:
  exploitable        preconditions are realistically satisfiable
  mitigated          real, but a deployed control reduces it below the
                     derived severity (name the control)
  needs_manual_test  severity hinges on something only a runtime test can
                     settle; recommend a human build a PoC

STEP 6: If SCORING STANDARD is a CVSS or OWASP variant, emit a
`severity_label` in that format (vector string + base score for CVSS;
likelihood x impact for OWASP). Otherwise set it equal to the derived
HIGH/MEDIUM/LOW.

────────────────────────────────────────────────────────────────────────
Respond with ONLY this block:

  PRECONDITIONS:
  - <one per line>
  ACCESS_LEVEL: <unauthenticated_remote|authenticated|local|physical>
  SEVERITY: <HIGH|MEDIUM|LOW>
  SEVERITY_LABEL: <per scoring standard>
  THREAT_MATCH: <matched threat-model entry, or none>
  SEVERITY_ALIGNMENT: <-5..+5>
  VERIFY_VERDICT: <exploitable|mitigated|needs_manual_test>
  RANK_RATIONALE: <2-4 sentences>

4b. Merge

For each confirmed finding, parse the block and attach preconditions

(replacing any scanner-supplied list), access_level, severity

(recomputed), severity_label, threat_match, severity_alignment,

verify_verdict, and append RANK_RATIONALE to rationale (separated by a

blank line from the Phase-3 rationale).

For findings that did NOT reach Phase 4 (false_positive, duplicate,

unlocatable): set severity: null, verify_verdict: null,

severity_alignment: null, preconditions: [].

Checkpoint: Write tool → ./.triage-state/_chunk.tmp:

{"phase": 4, "context": {...}, "findings": [ {all findings with severity/severity_label/preconditions/access_level/threat_match/severity_alignment/verify_verdict} ]}

Then Bash:

python3 .claude/skills/_lib/checkpoint.py save ./.triage-state 4 rank --from ./.triage-state/_chunk.tmp

Phase 5: Route

Tag each confirmed true-positive with the most specific component or owner

inferable. For each finding in confirmed[], stop at the first hit:

• CODEOWNERS / OWNERS. Grep --repo for CODEOWNERS, OWNERS,

.github/CODEOWNERS, docs/CODEOWNERS. If found, match the finding's

file against its patterns (last match wins). Hint:

"CODEOWNERS: <pattern> -> <owner(s)>".

• git log. If --repo is a git checkout, run

git -C {REPO} log --format='%an' -n 50 -- "{file}" | sort | uniq -c | sort -rn | head -3.

Hint: `"top committer: <name> (<n>/<total> recent commits); no

CODEOWNERS entry"`.

• Module fallback. Hint: `"component: <top-level dir of file>/; no

CODEOWNERS or git history"`.

Attach as owner_hint. State the source so confidence is clear; a bare

username is less useful than `"component: auth/; no CODEOWNERS entry; top

committer jsmith (14/20 recent commits)"`. For non-true-positive findings,

set owner_hint: null.

Checkpoint: Write tool → ./.triage-state/_chunk.tmp:

{"phase": 5, "context": {...}, "findings": [ {all findings with owner_hint} ]}

Then Bash:

python3 .claude/skills/_lib/checkpoint.py save ./.triage-state 5 route --from ./.triage-state/_chunk.tmp

Phase 6: Output

6a. Sort

Order all findings by:

• verdict: true_positive, then duplicate, then false_positive.
• Within true positives: severity HIGH > MEDIUM > LOW, then confidence

descending, then severity_alignment descending.

• Within others: original id.

6b. Write ./TRIAGE.json

{
  "triage_completed": true,
  "triage_context": {
    "mode": "interactive|auto",
    "environment": "...",
    "threat_model": ["..."],
    "scoring": "...",
    "noise_tolerance": "...",
    "votes_per_finding": 3,
    "repo": "..."
  },
  "summary": {
    "input_count": 0,
    "duplicates": 0,
    "false_positives": 0,
    "true_positives": 0,
    "needs_manual_test": 0,
    "by_severity": {"HIGH": 0, "MEDIUM": 0, "LOW": 0}
  },
  "findings": [
    {
      "id": "f001",
      "source": "VULN-FINDINGS.json#0",
      "title": "...",
      "file": "...",
      "line": 0,
      "category": "...",
      "claimed_severity": "HIGH",
      "verdict": "true_positive|false_positive|duplicate",
      "verify_verdict": "exploitable|mitigated|needs_manual_test|null",
      "confidence": 0.0,
      "severity": "HIGH|MEDIUM|LOW|null",
      "severity_label": "...",
      "severity_alignment": 0,
      "preconditions": ["..."],
      "access_level": "...",
      "threat_match": "...|null",
      "rationale": "file:line-cited prose: reachability, protections, why each held or didn't; then ranking rationale",
      "vote_breakdown": {"true_positive": 0, "false_positive": 0, "cannot_verify": 0},
      "refute_reasons": ["..."],
      "exclusion_rule": null,
      "first_links": ["file:line", "..."],
      "duplicate_of": null,
      "absorbed": ["..."],
      "owner_hint": "...",
      "missing_fields": ["..."]
    }
  ]
}

Every input finding appears exactly once (duplicates reference their

canonical via duplicate_of). Do not silently drop anything. Do not print

this JSON to the terminal; write to file only.

6c. Write ./TRIAGE.md

Reviewer-facing report. Build it incrementally. Do NOT emit the whole

file in one Write. One chunk per finding; a stalled chunk loses that one

section, not the file.

Step 1 — header. Write tool → ./TRIAGE.md (clobbers any prior file)

containing only the title block, summary, and ## Act on these heading:

# Triage Report

{summary line: N in -> D duplicates, F false positives, T confirmed (H high / M med / L low), X need manual test}

Context: {mode}; environment = {environment}; scoring = {scoring}; {votes}-vote verification.

## Act on these

Step 2 — per finding. For each true_positive in severity order:

• Write tool → ./.triage-state/_chunk.tmp containing ONE finding's section:

### [{severity}] {title}  ({id})
`{file}:{line}` | {category} | claimed {claimed_severity} (alignment {severity_alignment:+d}) | confidence {confidence}/10
**Owner:** {owner_hint}
**Verdict:** {verify_verdict}, votes {vote_breakdown}
**Preconditions ({n}):** {bulleted}
**Threat-model match:** {threat_match or "none"}
**Why:** {rationale}
**Reachability evidence:** {first_links}
{if verify_verdict == needs_manual_test:}
> Recommend a human build a PoC; static reasoning hit its limit.

• Bash:

python3 .claude/skills/_lib/checkpoint.py append ./TRIAGE.md --from ./.triage-state/_chunk.tmp

Repeat for each true_positive.

Step 3 — footer. Write tool → ./.triage-state/_chunk.tmp containing the

Dropped table, then checkpoint.py append it the same way:

## Dropped

| id | title | file:line | why dropped |
{false_positives: refute_reasons + exclusion_rule}
{duplicates: "duplicate of {duplicate_of}"}
{unlocatable: "no source location in input"}

Checkpoint (final): Bash:

python3 .claude/skills/_lib/checkpoint.py done ./.triage-state 6

The next invocation's resume check sees status == "complete" and starts

fresh.

6d. Terminal summary

Under ~12 lines:

Triage complete: {N} findings -> {T} confirmed, {F} false positives, {D} duplicates.

  HIGH:   {n}   {title of top HIGH, owner_hint}
  MEDIUM: {n}
  LOW:    {n}
  Needs manual test: {n}

  Top refute reasons: {top 3 refute_reasons with counts}

Wrote ./TRIAGE.md and ./TRIAGE.json

Testing this skill

Smoke test (five-finding fixture: 2 real, 1 dup, 2 FP):

/triage .claude/skills/triage/fixtures/canary-findings.json --auto --repo targets/canary

Expected: f001 and f003 confirmed; f002 duplicate of f001; f004 dropped

(misread_code: it's a read buffer, not a randomness source); f005 dropped

(already_handled: there is a null check at line 68).

Or against pipeline output:

vuln-pipeline run drlibs --runs 3 --parallel --stream
/triage results/drlibs/<ts>/ --repo targets/drlibs

Hand-check a sample of TRUE_POSITIVE/HIGH results (the first_links should

point at real call sites) and a sample of FALSE_POSITIVE rejects (the

exclusion_rule or refute_reasons should be defensible).

Design notes

• Checkpoints are per-phase JSON, not conversation state. The pipeline's

--resume <session_id> (docs/pipeline.md) restores transcript history but

doesn't help when the orchestrator's context window itself fills;

file-backed checkpoints let a brand-new session pick up from the last

completed phase. ./.triage-state/ is scratch — add to .gitignore.

• Dedupe runs before verify to cut verifier spend by the duplication

factor (often 2-4x on multi-scanner input) at the cost of one cheap

subagent.

• Semantic dedupe is one agent, given only id/file/line/category/title:

enough to cluster, not enough to leak one scanner's reasoning into

another finding's verification.

• Bash is allowed narrowly for git log (owner hints), jq/find

(ingest), and python3 .claude/skills/_lib/checkpoint.py (state I/O).

The actual safety property is "no execution of target code," which is

preserved.

• CANNOT_VERIFY exists so verifiers aren't forced into a false

binary. It maps to needs_manual_test under recall policy and to a drop

under precision policy.

• Threat-model boost is capped at one step so a stated threat can't

re-inflate a LOW back to HIGH and defeat the precondition rule.

• severity_label is separate from severity. Sorting always uses the

precondition-derived HIGH/MEDIUM/LOW; the label is presentation-layer for

whatever standard the reviewer's tooling expects.

• Pipeline report.json ingest is best-effort. Those reports describe

ASAN crashes with prose exploitability analysis rather than the

file/line/category shape static verifiers expect. Expect more

needs_manual_test verdicts on that input than on static-scanner JSON.

• Sharding at ~40 parallel Tasks is a conservative ceiling for typical

agent-spawn limits; tune up if your runtime allows.

• No network, deliberately. CVE-database enrichment and upstream-fix

checks would help ranking but break the air-gapped-review property.