findings-ledger

Use when recording, tallying, or triaging stochastic review findings — the append-only JSONL ledger where Tier 2 (advisory, unevidenced) reviewer output goes instead of blocking language. Triggers on

SKILL.md

GitHub

Findings Ledger

Stochastic judgment proposes; deterministic verification disposes (tier doctrine: review-tiers — .claude/rules/review-tiers.md). This skill is the proposes side's inbox: a single append-only JSONL at .claude/ledger/findings.jsonl where Tier 2 findings accumulate fingerprinted, so the same defect reported twice across runs collides to one entry — deterministically when the reviewer passes a --category, heuristically on claim wording when it doesn't — and recurrence across independent runs — not rhetoric — decides what gets investigated.

The ledger file

  • Location: .claude/ledger/findings.jsonl, one JSON event per line.
  • Append-only. Status transitions are new events for the same fingerprint; prior lines are never mutated. Current status = the fingerprint's most recent event.
  • Gitignored by default (machine-local noise). Team-shared mode, two options: (a) commit ledger.py export --out <path> output — one line per fingerprint with category, file, count and first/last seen, no claim text — to any tracked path outside .claude/ledger/; no .gitignore change needed. file and symbol are still reviewer-supplied free text and get committed, so review the export before pushing. (b) Commit the raw ledger: in .gitignore, replace the .claude/ledger/ line with .claude/ledger/*.lock (the lock sidecar stays local). Append-only + stable ordering keeps either diff reviewable.
  • CI runners don't persist it. Appends made during a CI job vanish with the runner — and CI must never push the ledger back to the repo. Upload the run's findings as a build artifact and harvest locally instead; patterns in references/ledger-format.md.

Entry fields: fingerprint, fingerprint_scheme (1 = text, 2 = structured; absent on pre-#279 entries, read as 1), category and symbol (or null), file, claim, tier, source, run_id, date, evidence (path or null), status (NEW | RECURRING | INVESTIGATING | PROMOTED | RETIRED-NOISE). Full schema, lifecycle, and fingerprint-normalization limits: references/ledger-format.md.

Commands

All via scripts/ledger.py (python3, stdlib only; exit 0 = ok, 2 = setup/usage error; output ordering is deterministic). Resolve the script path first — references/install-paths.md:

PROJ="${CLAUDE_PROJECT_DIR:-.}"
# 1. Repo checkout
LEDGER="$PROJ/.claude/skills/findings-ledger/scripts/ledger.py"
# 2. Project Codex install
[ -f "$LEDGER" ] || LEDGER="$PROJ/.agents/skills/findings-ledger/scripts/ledger.py"
# 3. Global Claude Code install
[ -f "$LEDGER" ] || LEDGER="$HOME/.claude/skills/findings-ledger/scripts/ledger.py"
# 4. Global Codex install
[ -f "$LEDGER" ] || LEDGER="$HOME/.agents/skills/findings-ledger/scripts/ledger.py"

python3 "$LEDGER" add \
  --file <path> --claim "<one sentence>" --tier 2 \
  --source <agent-name> --run-id <id> [--evidence <path>] \
  [--category <[ns/]kebab-slug> [--symbol <function|heading|key>]]

python3 "$LEDGER" tally
python3 "$LEDGER" triage [--threshold 2] [--age-days 14]
python3 "$LEDGER" promote <fingerprint> [--status INVESTIGATING] [--evidence <path>]
python3 "$LEDGER" retire <fingerprint>
python3 "$LEDGER" export --out <path>
  • add computes the fingerprint (see Categories below) and labels a sighting RECURRING only when an earlier sighting came from a different --run-id; a repeat within the same run stays NEW. A --tier 1 finding without --evidence is demoted to tier 2 automatically — the tier rule, enforced. Control characters (newline, tab, …) in --file, --category or --symbol are rejected.
  • Recurrence counts distinct run ids, not raw sightings — an agent repeating itself within one run cannot cross the threshold alone.
  • triage lists fingerprints at/over the recurrence threshold as promotion candidates, and still-NEW single-run findings older than the age cutoff as retire-as-noise candidates. The /triage-findings command (repo-local maintainer tooling; not shipped by the installer) wraps this and proposes a ratchet target per candidate; the human disposes.
  • promote refuses to record PROMOTED without --evidence — a promotion is its encoded check.
  • export writes the deduplicated per-fingerprint summary (sorted, byte-deterministic JSONL) to --out, and refuses any --out inside the ledger's directory.

Categories (structured fingerprints)

Two reviewers rarely phrase one defect the same way, so text fingerprints rarely recur. Pass --category and the fingerprint becomes sha256("v2" + normalized file path + category + symbol) — claim wording no longer matters, and paraphrases collide.

  • --category — a short kebab-case slug naming the defect class, optionally under one namespace (slop/unused-dep). Lowercased; anything that isn't [ns/]slug with each part [a-z0-9]+(-[a-z0-9]+)* is rejected. Seed vocabulary — prefer these, coin a new slug when none fits:

    Slug Use for
    fail-open error/missing input path that grants or proceeds instead of denying
    missing-validation user input reaches a sink unchecked
    missing-ci-wiring a test/check exists but CI never runs it
    missing-test behavior change with no test exercising it
    stale-doc docs contradict the current code
    broken-reference link, path or cross-reference that does not resolve
    vague-routing skill/agent description too unspecific to route on
    dead-code unreachable or unused code/config
    secret-exposure credential or token logged, committed or leaked

    Reviewer vocabulary (slop/*, security/*, used by code-reviewer and security-reviewer): references/ledger-format.md.

  • --symbol (optional, requires --category) — the function, heading, or config key the finding is about. Backticks and a trailing () are stripped, so `check_token()`, `check_token`(), check_token () and check_token match. Omit it for file-level findings.

  • Without --category, add uses the legacy text fingerprint (scheme 1), unchanged — existing entries keep their identity. A scheme 1 and a scheme 2 sighting of the same defect get different fingerprints, so stick to one scheme per defect.

Reviewers should pass both whenever the finding names a defect class and a location, reusing the same slug for the same class across runs:

python3 "$LEDGER" add --file src/auth.py --tier 2 --source code-reviewer \
  --run-id pr-281 --category fail-open --symbol check_token \
  --claim "check_token returns True when the verifier raises"

The ratchet

A recurring fingerprint is a candidate to leave the stochastic layer forever: investigate → encode as a scripts/validate.sh rule (Tier 0) or an exit-nonzero evidence script (Tier 1) → promote with the encoded check as --evidence → no LLM ever re-litigates it. Singles nobody re-reports get retired. The ratchet only turns toward lower variance.

Related skills