| name | rule-trace |
| description | Make AI-agent rule application visible and auditable using stable rule IDs, a catalog, trace blocks, usage counters, and a deterministic validator. Use this whenever someone wants to add rule tracing to a repo, turn existing agent rules (CLAUDE.md, AGENTS.md, .cursorrules, .agents/, scattered docs) into a traceable ID-based format, track or count how often each rule is considered vs applied, audit which rules are noise or dead, validate that a rule catalog and its importers haven't drifted, or build a report/dashboard of rule usage. Triggers on phrases like "rule tracing", "trace which rules were applied", "stable rule IDs", "rules catalog", "collect my traces", "which rules never fire", "rule usage metrics", or wiring rules across multiple agent tools. |
| version | 1.4.0 |
| license | MIT |
Rule Trace
Rules loaded into an agent are loaded context, not applied context — a rule that was followed looks identical to one that was ignored. This skill closes that gap. Every rule gets a stable ID anchored at a markdown heading; an agent that does real work appends a trace block with human-readable prose plus fenced rule-trace JSON disclosing which in-scope rules it considered (candidates) versus which it actually let constrain the work (applied); a catalog indexes every ID; deterministic scripts validate the system and count candidate-vs-applied usage across sessions so dead and miscoped rules become visible.
The most valuable signal is the diff between candidates and applied — a rule that is always a candidate but never applied is noise, miscoped, or being ignored.
Pick a mode
Read the user's intent and route to one mode. Each links the reference you should read before acting.
| Intent | Mode | Read first |
|---|
| Add rule tracing to a repo that has none | init | references/convention.md, references/importer-wiring.md |
| Convert existing/scattered rules into the traceable format | migrate | references/migration-guide.md, references/rule-anatomy.md |
| Maintain/clean an existing rule set, using usage data as evidence | audit | references/catalog-format.md, plus the report (below) |
| Count usage and build the dashboard | report | this file's "Counters" section |
The deterministic scripts live in scripts/ and are portable (Node ≥ 18, no dependencies). They resolve repo layout from an optional .agents/rule-trace.config.json (see references/catalog-format.md), falling back to the conventional layout. Run them from the target repo root, or pass --root <dir>.
init — scaffold a fresh system
Goal: drop the convention in, create an empty catalog, seed one example rule, and wire the importers.
- Inspect the repo for existing agent entry points:
CLAUDE.md, AGENTS.md, .opencode/opencode.json, .cursorrules, .github/copilot-instructions.md. If real rules already exist there, switch to migrate instead — don't scaffold over content.
- Copy
templates/rule-trace.md.tmpl → .agents/rule-trace.md, templates/rules-catalog.md.tmpl → .agents/rules-catalog.md, and templates/rule-file.md.tmpl → .agents/rules/root.md (keep the one example rule so the layout is concrete).
- Wire every supported agent entry point present to reference the rule files in lockstep, or configure generated importers for tools that cannot expand references — see
references/importer-wiring.md for which tools actually expand references. The non-negotiable invariant for configured importers: reference importers load the identical set of files, while generated importers are fresh materializations of the same canonical content. The validator enforces both. If the repo uses only one agent tool, set importers in .agents/rule-trace.config.json to just that entry so the validator doesn't warn about the absent ones.
- Run
node <skill>/scripts/validate-rules.mjs --root <repo> and fix anything it reports.
- Optionally scaffold the operational wiring (a CI job that runs the validator, a metrics
.gitignore, the Claude Code Stop hook) with node <skill>/scripts/scaffold-wiring.mjs --root <repo> — it's non-destructive. Offer this; don't force it.
migrate — turn existing rules into traceable form
This is the high-value mode and the part that needs judgment — extraction is not mechanical. Follow references/migration-guide.md in full. In short:
- Gather every source of existing rules (the entry points above, plus READMEs/docs the user points to).
- Split prose into discrete, individually-citable rules. One rule = one enforceable idea.
- Assign IDs by layer: repo-wide (
ROOT-), topic/area (e.g. TEST-, STYLE-), package-local (PKG-<PKG>-<AREA>-). Number sequentially per prefix. IDs are immutable once published.
- Rewrite each rule into the anatomy in
references/rule-anatomy.md (Scope / Applies when / Severity / Rule).
- Build the catalog — run
node <skill>/scripts/generate-catalog.mjs --root <repo> --write to derive it from the headings (it preserves any summaries you've already written), then wire the importers (references/importer-wiring.md). If you configure generated importers for Cursor, Copilot, or other reference-blind tools, run node <skill>/scripts/sync-importers.mjs --root <repo>. See references/catalog-format.md.
- Validate:
node <skill>/scripts/validate-rules.mjs --root <repo> must pass.
Extraction is agent judgment; the validator is the deterministic check on the output. Don't invent rules the sources don't support.
audit — maintain the rule set with evidence
Maintain the rules by combining repo state, quantitative usage evidence, and the current session into concrete edits. Follow references/audit.md in full — it has the method, parameters (scope/paths/session/usage/apply), the report.json flag interpretation, the output template, and the rules of thumb. In short:
- Build the latest report:
node <skill>/scripts/report.mjs --root <repo> (writes report.json + dashboard.html under the metrics dir).
- Read its flags (dead / stale / always-candidate-never-applied / low-rate / un-waived MUST / unknown IDs) and combine with session evidence.
- Classify each rule Keep / Revise / Remove / Consolidate / Add, in the narrowest correct file; if applying fixes, keep to low-risk ones.
- Re-validate and regenerate the catalog after edits.
report — counters and dashboard
See "Counters" below.
Counters
Trace blocks emit candidate + applied IDs in prose and fenced rule-trace JSON in every relevant response, so the data exists in transcripts — it just needs collecting. Two collectors share one append-only event log (<metricsDir>/traces.jsonl), deduped by transcript message UUID so they never double-count:
- Offline collection/backfill (tool-agnostic):
node <skill>/scripts/parse-traces.mjs --root <repo> walks saved Claude Code transcripts (default ~/.claude/projects/<encoded-cwd>/) and appends any trace blocks it finds. Re-runnable; retroactive over history. Point --transcripts <dir> at another tool's transcript store if its records expose uuid + an assistant message.content.
- Live Stop-hook (Claude Code only): wire
scripts/record-trace.mjs as a Stop hook (see references/importer-wiring.md). It records each finished main-agent response automatically, including untraced responses so reports can state trace coverage. Ignores SubagentStop; never blocks the agent. The plugin install already wires this hook, so the manual hook is for standalone installs only — never both, or the recorder runs twice per turn.
Then node <skill>/scripts/report.mjs --root <repo> aggregates the log into per-rule candidate/applied/rate plus the flag lists, writing report.json and a self-contained dashboard.html. To publish the dashboard as a shareable Artifact, render the generated HTML with the Artifact tool. Pass --now <ISO-8601 date> to pin report time (staleness and generatedAt) for a reproducible run — e.g. when regenerating committed example output; it does not filter events (that's --since).
Honest limits to state when you present counts: they are self-reported (what the model claimed, not proof of compliance); a rule never surfaced as a candidate is invisible (false absence); and counts need volume before rates mean much. The live hook is Claude-Code-specific; the convention and offline collector are tool-agnostic.
Validate (CI)
node <skill>/scripts/validate-rules.mjs --root <repo> checks: every catalog ID resolves to a heading; every heading is catalogued; no duplicate IDs; reference importers load identical file sets; generated importers are fresh; required fields present (incl. Severity — pass --no-severity while migrating before severities are added); and warns on numbering gaps. Trace-lint mode --lint-file <path> flags cited IDs missing from the catalog; use it on real trace output, not documentation quoting the trace format, because four-backtick outer examples are ignored while three-backtick rule-trace examples are linted as real traces. Exit 1 on errors. For copy-paste GitHub Actions / GitLab / package-script / Stop-hook snippets, see references/ci-wiring.md.
