| name | decision-audit |
| description | Governance audit of an engagement decision log. Runs a dynamic multi-agent
workflow: one verifier per rule (structurally-earned coverage of every entry)
followed by an adversarial verification pass that removes false positives.
Use when asked to audit / QA / lint the decision log, check governance hygiene,
or before a sprint close-out or client handover.
|
| version | 1.0.0 |
| kind | workflow-harness |
| workflow | decision-audit.workflow.js |
| triggers | ["audit the decision log","decision log hygiene","check the decision log","governance audit","lint decisions","decision-log QA"] |
decision-audit
A dynamic-workflow harness that audits a decision log against a set of governance
rules and returns only adversarially-verified findings — not a raw pass that
dumps unverified flags.
When to use it
| Use the harness when | Use a quick single-pass read when |
|---|
| The log is large (dozens+ of entries) | The log has only a handful of entries |
| Before a sprint close-out / client handover | A casual spot-check |
| You need trustworthy, low-false-positive findings | You just want a rough scan |
In testing, a single-pass audit produced many findings of which the large majority
were false positives; the harness's adversarial pass cut that to a small set of
verified, actionable items. See docs/workflows.md.
How to run
This is a template, not a fixed script. Read decision-audit.workflow.js,
adjust RULES/paths/rounds to the engagement if needed, then invoke the
Workflow tool:
Workflow({
scriptPath: ".claude/skills/decision-audit/decision-audit.workflow.js",
args: {
repoPath: "<absolute path to this team-ops repo>",
decisionLog: "03-status/decision-log.md", // optional; this is the default
clientSurfaceLog: "03-status/client-surface-log.md",
runDate: "YYYY-MM-DD", // for the stale-decision rule
rounds: 1 // loop-until-dry cap; raise for very large logs
}
})
Gotcha: the Workflow tool passes args as a JSON string. The script already
calls JSON.parse(args) — keep that if you fork it.
Returns { total_entries, coverage, confirmed, rejected, findings, report_markdown }.
The workflow is read-only against the log; the calling session writes the report
wherever the engagement keeps audit output.
Design notes
- One finder per rule → coverage of every entry is structurally earned, not
self-asserted.
- Adversarial verification is the value-add — it removes the false-positive noise
a raw pass emits. See
../README.md for the shared conventions.
- Read once: finders return each offending row's
entry_excerpt; verifiers receive
it inline and do not re-read the file.
- Batch verify by rule: one verifier judges all of a rule's candidates. Agent
count is the dominant cost driver — batching keeps a large-log run affordable.
Tuning levers
| Lever | Effect |
|---|
rounds (default 1) | loop-until-dry; raise only for very large logs |
| verify batch granularity | per-rule (default) = cheaper, more conservative; per-finding = higher recall, more agents |
ruleSet | swap/extend the governance rules per engagement |
