Workflow Audit
Usage Template
Provide: objective, evidence source, mode (manual, session, or interview), size bound, candidate limit, operator/approval owner, and desired handoff.
Workflow
SKILL_CANDIDATE when the work has a reusable procedure or multiple steps; otherwise KEEP_MANUAL.AUTOMATION_CANDIDATE only when repetition or a clear trigger is evidenced; otherwise HOST_REQUIRED_REVIEW.LOOP_CANDIDATE only when retry, monitoring, polling, follow-up, or a repeat-until condition is evidenced; otherwise NO_LOOP.
Record the evidence snippet, approval owner, verifier, stop condition, and side-effect expectation. Promotion is a separate supervised action with a receipt and rollback path.
Failure Protocol
Return NEEDS_INPUT for missing evidence or a required authority decision, INSUFFICIENT_EVIDENCE when repetition cannot be supported, and VERIFY_FAILED when a recommendation cannot be traced to the supplied source. Preserve the audit receipt and do not promote.
Output Contract
Return status, result, evidence, unknowns, and next_action. The result includes mode, source hash, bounded candidates, skill/automation/loop decisions, approval, verifier, stop condition, and side-effect count.
Edge Cases
- A quiet prior session is not proof that a workflow is eligible for automation; record
INSUFFICIENT_EVIDENCE or request a frequency probe.
- A recurring task that writes to a shared Vault remains a supervised staged-write candidate even when the extraction itself is deterministic.
Success Metrics
- Source provenance and candidate bounds are preserved.
- Every decision is evidence-linked and independently checkable.
- No skill, schedule, external service, or Vault file is changed by the audit.
Quality Gates
1---2name: workflow-audit3description: Use when manual work, prior sessions, or an operator interview must be audited for repeatable skill, automation, or bounded-loop candidates.4---56# Workflow Audit78<skill_contract>9 <input>Bounded manual notes, prior-session evidence, or interview answers describing work, repetition, decisions, and constraints.</input>10 <output>A deterministic audit with mode, evidence-bounded candidates, skill/automation/loop decisions, permissions, and a typed handoff.</output>11 <done>Every recommendation points to supplied evidence, remains within the input bound, and has an explicit supervised promotion decision.</done>12 <non_goals>Editing skills, changing schedules, mutating Vault state, inventing metrics, or executing an unapproved external action.</non_goals>1314## Usage Template1516Provide: objective, evidence source, mode (`manual`, `session`, or `interview`), size bound, candidate limit, operator/approval owner, and desired handoff.1718## Workflow1920<intake>21Classify the source as manual notes, prior-session history, or an operator interview. Preserve its hash and provenance. Infer a mode only when the text contains a clear session or interview signal; otherwise use `manual`.22</intake>2324<unknowns_gate>25Separate observed repetition from a proposed improvement. Missing frequency, owner, verifier, permission, or stop condition makes a candidate review-only; it does not justify a schedule or autonomous loop.26</unknowns_gate>2728<execute>29Extract a bounded list of distinct work candidates. For each candidate, decide:3031- `SKILL_CANDIDATE` when the work has a reusable procedure or multiple steps; otherwise `KEEP_MANUAL`.32- `AUTOMATION_CANDIDATE` only when repetition or a clear trigger is evidenced; otherwise `HOST_REQUIRED_REVIEW`.33- `LOOP_CANDIDATE` only when retry, monitoring, polling, follow-up, or a repeat-until condition is evidenced; otherwise `NO_LOOP`.3435Record the evidence snippet, approval owner, verifier, stop condition, and side-effect expectation. Promotion is a separate supervised action with a receipt and rollback path.36</execute>3738<evaluate>39Reject recommendations that rely on an unbounded input, an inferred external permission, a fluent success claim, or a video/demo as authority. Check candidate count, source hash, evidence linkage, and `side_effect_count: 0`.40</evaluate>4142## Failure Protocol4344Return `NEEDS_INPUT` for missing evidence or a required authority decision, `INSUFFICIENT_EVIDENCE` when repetition cannot be supported, and `VERIFY_FAILED` when a recommendation cannot be traced to the supplied source. Preserve the audit receipt and do not promote.4546## Output Contract4748Return `status`, `result`, `evidence`, `unknowns`, and `next_action`. The result includes mode, source hash, bounded candidates, skill/automation/loop decisions, approval, verifier, stop condition, and side-effect count.4950## Edge Cases5152- A quiet prior session is not proof that a workflow is eligible for automation; record `INSUFFICIENT_EVIDENCE` or request a frequency probe.53- A recurring task that writes to a shared Vault remains a supervised staged-write candidate even when the extraction itself is deterministic.5455## Success Metrics5657- Source provenance and candidate bounds are preserved.58- Every decision is evidence-linked and independently checkable.59- No skill, schedule, external service, or Vault file is changed by the audit.6061## Quality Gates6263- [ ] Mode is explicit or conservatively inferred.64- [ ] Unknown permissions, verifiers, and stop conditions are surfaced.65- [ ] Promotion is supervised and receipt-backed.66- [ ] `side_effect_count` is zero.6768</skill_contract>
Mark393295827