AGENT SKILL

audit-progress

Use when an iteration has just finished and you need to verify behavior evidence quality in three tiers — deep evidence for current stories, impacted behavior for touched scenarios, and sentinel corpus for high-value regression detection.

License 许可
Apache-2.0
所属仓库
zak-keown/moe
最近更新
2026-09-05
Stars
★ 0
Forks
1

01 前言 Front Matter

name
audit-progress
description
Use when an iteration has just finished and you need to verify behavior evidence quality in three tiers — deep evidence for current stories, impacted behavior for touched scenarios, and sentinel corpus for high-value regression detection.

02 资源文件 1

03 SKILL.md 正文

Auditing Progress

Overview

Runs after every iteration as part of the planning cycle. Verifies behavior evidence quality in three tiers using parallel adversarial review (PAR) — two paired auditor subagents evaluate the same work in parallel with competitive framing.

The audit answers: "Does durable, reusable evidence exist at the correct seam for every externally observable behavior this iteration touched?"

When to Use

Invoked by iterate after every run-iteration call, before picking the next iteration.

Audit Process

1. Partition the audit into three tiers

Read the per-epic requirement files in docs/moe/iterations/requirements/, docs/moe/iterations/behavior-scenarios.md, and docs/moe/iterations/behavior-corpus.md:

  • Tier 1 — Deep evidence: stories marked done:ITER-<current> and scenarios added or updated in this iteration. Audit every AC and its proof obligation thoroughly.
  • Tier 2 — Impacted behavior: all existing scenarios whose owning stories had code changes in this iteration (even if those stories were completed in earlier iterations). Verify the scenarios still pass.
  • Tier 3 — Sentinel corpus: all scenarios with run cadence sentinel in the behavior corpus. Compare against the pre-iteration baseline from run-iteration step 3.

2. Dispatch paired auditor subagents (PAR)

Following the PAR methodology in skills/_shared/parallel-adversarial-review.md, resolved relative to this loaded document:

  1. Build the auditor prompt using auditor-subagent-prompt.md. Include ALL THREE tiers:
    • Tier 1: full story cards with proof obligations + new/changed scenario cards
    • Tier 2: impacted scenario cards + their current test results
    • Tier 3: sentinel scenario IDs + baseline results + current results
  2. Resolve skills/_shared/par-reviewer-wrapper.md relative to this loaded document and wrap in its competitive framing
  3. Dispatch TWO auditor subagents in parallel
  4. Wait for both to return

3. Aggregate findings

Following PAR aggregation rules:

  • Same finding from both auditors → one finding, high confidence
  • Finding from only one auditor → separate finding, still actionable
  • Severity disagreement → take the more severe assessment, always fix it

4. Process results

  • If gaps found (any AC fails, evidence is too weak, sentinel regression detected):
    • For AC failures: append gap stories to requirements/ (status pending) or flip existing stories back from done to pending
    • For weak evidence: create evidence-improvement stories (add scenario, strengthen seam)
    • For sentinel regressions: create regression-fix stories with CRITICAL priority
    • Revise roadmap.md to add a follow-up iteration for the gaps
  • If clean (all tiers pass, evidence is adequate):
    • The iteration is confirmed done
    • Return clean signal to the orchestrator

5. Return control

Return the audit result (clean or gaps) to the orchestrator. The orchestrator decides whether to loop or terminate.

Quick Reference

Tier What it checks Failure means
Deep evidence Every AC + proof obligation for current iteration Story not done, evidence too weak
Impacted behavior Scenarios whose surfaces were touched Stale or broken scenario
Sentinel corpus High-value journey scenarios Regression in previously-working behavior
Reads Writes Dispatches
requirements/, behavior-scenarios.md, behavior-corpus.md, product code/tests requirements/ (gaps), roadmap.md (new iteration) if gaps, behavior-scenarios.md (stale flags) Two auditor subagents in parallel (PAR)

References

04 同仓库更多 Skills