audit

/agentflow:audit

Make it visible: the moment you start, say so in one line (skill + run-id) so it's clear an Agent Flow run is happening; /agentflow:board then lists every run on disk — the audit trail.

You are the orchestrator of a /agentflow:audit recipe. Your job is to construct a 6-stage pipeline.json, hand it to /agentflow:pipe, and let the framework drive execution. The recipe itself adds no new primitives.

Step 0 — Resolve config + compute manifest hash for the run-id

Read ${CLAUDE_PLUGIN_ROOT}/skills/audit/defaults.md. Apply override priority CLI > defaults.

Required: --target <path>. If missing → ask via AskUserQuestion or abort with a clear message.

Manifest hash (forces a new run when source files change):

• List all files under <target> matching <file_glob> and not <file_exclude>
• For each: take (rel_path, mtime_ns) — fast, no byte reads
• Sort by rel_path
• manifest_hash = sha256(joined entries)[:16]

Run-id derivation (when --run-id is NOT provided):

• audit-<8 char hash> where the hash is sha256(target + glob + review_model + manifest_hash)[:8]
• This way: same content → same run-id (resume works), changed content → new run-id (fresh run).

Echo the resolved config + run-id in one line. If target has > 200 files, ask for confirmation (cost/duration warning).

Step 1 — Working dirs

Create .agentflow/audit/<run-id>/ (recipe scratch) and .agentflow/pipe/<run-id>/ (pipeline state).

Step 2 — Resolve the discover params

The 6-stage structure is shipped as a self-contained, human-readable workflow-file at ${CLAUDE_PLUGIN_ROOT}/workflows/audit/WORKFLOW.md (this is the canonical, reusable artifact — you do NOT hand-build a stages.json). Its discover stage runs the sibling discover.mjs via {{workflow.dir}} (so the whole workflows/audit/ folder is movable); the script emits a /agentflow:foreach-compatible items array with a per-file content_hash (for the review --cache).

Per-invocation inputs are declared params, passed to init with --param (no env exports):

• target (required) — the directory to audit
• glob (default **/*) — comma-separated include globs, e.g. **/*.cs
• exclude (default empty) — comma-separated exclude globs

To tune the review/digest models or group depth, copy the whole workflows/audit/ folder into your project's workflows/ and edit the --model / --method-config values in its WORKFLOW.md; pass that copy's WORKFLOW.md to --workflow.

Step 3 — (the pipeline is the workflow-file)

No stages.json to build. The workflow-file already wires the 6 stages with declarative templates ({{stages.<name>.run_id}}, {{stages.<name>.result_pointer}}, {{run.dir}}, {{params.*}}), resolved by /agentflow:pipe at tick time. The stages are: discover (bash) → review (/agentflow:foreach, --kind code-review --cache) → build-group-input (json) → partition (/agentflow:group path-prefix) → build-digest-inputs (json) → digest (/agentflow:reduce, markdown).

Step 4 — Validate and init the pipe

node "${CLAUDE_PLUGIN_ROOT}/dist/state/pipe.js" init <run-id> \
  --workflow "${CLAUDE_PLUGIN_ROOT}/workflows/audit/WORKFLOW.md" \
  --param target="<resolved-target-path>" \
  [--param glob="**/*.cs"] [--param exclude="<globs>"] \
  [--no-stop-on-failure if user passed --keep-going] \
  [--force if --run-id was explicitly provided and overrides existing]

/agentflow:pipe init automatically runs schema validation on every primitive stage's init_args (catches typos in flags / bad kind values / missing required configs BEFORE the pipeline starts). A clear error listing surfaces here, not mid-run.

Step 5 — Drive the pipe

Use pipe drive (NOT manual tick + execute loops) — it auto-runs every bash stage and every deterministic primitive stage (in this recipe: discover, build-group-input, partition, build-digest-inputs) without your involvement, stopping ONLY when an Agent dispatch is needed (review and digest stages).

node "${CLAUDE_PLUGIN_ROOT}/dist/state/pipe.js" drive <run-id>

Output is JSON:

• {"action": "needs_agent", "cmd": "foreach", "suggested_child_run_id": "...", "init_args": [...]} — you must init the child + run the dispatch loop + advance, then call drive again.
• {"action": "done", "result_pointer": "..."} — pipeline complete, surface the digest.
• {"action": "failed", ...} — surface the error.

When needs_agent for /agentflow:foreach review:

1. Run the child init: node "${CLAUDE_PLUGIN_ROOT}/dist/state/foreach.js" init <suggested_child_run_id> <init_args...> --force
2. Record in pipe: node "${CLAUDE_PLUGIN_ROOT}/dist/state/pipe.js" start-primitive-child <run-id> --child-cmd foreach --child-run-id <suggested>
3. Follow /agentflow:foreach SKILL.md Step 4 dispatch loop: status → claim → split into chunks → fan-out parallel Agents in ONE message → complete-batch each result file.
4. After every Agent return, record token usage: node "${CLAUDE_PLUGIN_ROOT}/dist/state/foreach.js" budget-add <enum-run-id> --tokens <total_tokens> --model <review_model> so the budget aggregates correctly.
5. When complete-batch reports run_status: done, call drive again — it advances past the review stage, auto-runs the bash bridges + group, and stops at digest.

When needs_agent for /agentflow:reduce digest:

1. Init the reduce child + record + materialize (the reduce SKILL covers Steps 2-4).
2. Dispatch ONE digest agent (no fan-out for /agentflow:reduce).
3. After Agent returns: state/reduce.js budget-add for the tokens, then state/reduce.js complete <child-id> --output-path ./<run-id>-audit.md — write the digest to a visible file in the workspace root (e.g. audit-3f2a-audit.md), not buried under .agentflow/.
4. Call drive once more → returns done.

Step 6 — Final report

When drive returns done:

• Read the first 30 lines of result_pointer (the digest markdown) and surface inline.
• Run node "${CLAUDE_PLUGIN_ROOT}/dist/inspect.js" budget <run-id> and surface the total cost.
• Run node "${CLAUDE_PLUGIN_ROOT}/dist/inspect.js" tree <run-id> for the full child tree breakdown.
• Suggest follow-ups (e.g. "to refine the auth component, run /agentflow:foreach --kind audit --model opus --items <groups.json filtered>").

Important rules

• Idempotence: re-running with the same --run-id (or the same --target if run-id is auto-derived) resumes. With the manifest hash in run-id derivation: a file content change → new run-id → fresh run. Pure resume of a half-completed run uses the SAME run-id.
• No file modification: this recipe is read-only on the target. The digest lands as a visible ./<run-id>-audit.md in the workspace; internal state lives under .agentflow/audit/<run-id>/, .agentflow/pipe/<run-id>/, .agentflow/foreach/<run-id>-s1-foreach/, .agentflow/group/<run-id>-s3-partition/, .agentflow/reduce/<run-id>-s5-digest/.
• Incremental re-runs: with --cache on the review stage, files whose content_hash matches a prior cached result are skipped (no agent dispatch). Hits saved under .agentflow/cache/foreach-code-review/.
• Validate then drive: trust the dry-run validation /agentflow:pipe init performs. Catches recipe typos before any agent dispatch.

Quick example

/agentflow:audit --target examples/fake-repo

Expected on the bundled fake-repo (8 files, 4 components, 6 with intentional bugs):

• stage discover finds 8 .cs files (each with sha256 content_hash)
• stage review dispatches /agentflow:foreach (sonnet, conc=4, --cache) → 6 bug reports + 2 clean
• stage build-group-input materializes the run reference
• stage partition runs /agentflow:group path-prefix depth=1 → 4 groups (auth, billing, api, data)
• stage build-digest-inputs materializes the reduce inputs
• stage digest dispatches /agentflow:reduce (opus, markdown) → executive report
• Final digest written to a visible ./<run-id>-audit.md in the workspace

Re-running the same command without changing fake-repo: the same run-id is regenerated, all 8 review items become cache hits → no agent dispatch for review → only the digest runs again (still costs ~$0.10 for the opus call). If you change one file, only that file dispatches a new review agent.