woos-version-integration-audit

name	woos-version-integration-audit
description	Dedicated cross-feature integration audit. Extracts candidate conflicts by script, then performs evidence-backed semantic review across the full version scope.
version	1.0.0
author	Hermes Profile
license	MIT
metadata	{"hermes":{"tags":["integration","audit","cross-feature","traceability","gate","product-design"],"related_skills":["woos-product-design-flow","woos-prd-consistency-audit"]}}

Version Integration Audit

Output Language

• Author all user-facing prose (narratives, findings, summaries) in the user's most recent input language (e.g. if the user is writing in Chinese, write the PRD body, review findings, and summaries in Chinese).
• Keep these tokens verbatim in English regardless of the user's language, because they are contract identifiers consumed by other skills and gates:
  • Verdict tokens: PASS, REQUEST_CHANGES, BLOCKED, NOT_RUN
  • Severity tokens: critical, high, medium, low, warning
  • Tag tokens: [ASSUMPTION: ...], [NEEDS CLARIFICATION: ...], [NEEDS INPUT: ...]
  • Phase / dimension IDs: P0 … P11, P2a, Phase A, Phase B
  • Shape values: internal-tool, single-operator, consumer-product, multi-stakeholder, CLI
  • P0 rating values: strong, adequate, thin, broken
  • Section headings declared in templates (e.g. ## Background, ## Functional Requirements, ## Assumptions Index) — translations break downstream structural checks
  • Field labels inside templates (e.g. **Consequences (testable):**, **Out of Scope:**, **User value:**, Given … When … Then …)
• Code, file paths, IDs, and quoted PRD excerpts stay in their original form.

Purpose

Run Step 7 of woos-product-design-flow as a standalone audit skill with its own completion contract.

This skill exists because cross-feature review is where agents most often pretend they checked everything and then write a shallow PASS.

Incremental Execution Model

This skill runs incrementally — not just once at the end:

• After the 2nd feature completes Step 6.5 → first run (F1 + F2)
• After each subsequent feature → incremental run (all completed features so far)
• Final run after last feature → full integration report

Context management: To avoid context explosion, each incremental run loads:

• Full docs: ONLY the newest completed feature (requirements + PRD + UI brief when present)
• Interface summaries: ALL completed features (*-interface.md), including the newest when produced
• Script pre-filter: integration_gate.py extracts conflict candidates BEFORE semantic review
• Roadmap + architecture: always loaded (shared context)

This keeps context bounded at O(1 full feature + n interface summaries) rather than O(n full features).

Required Invocation (hard gate)

• MUST be invoked as a separate skill in fresh context
• MUST run scripts/integration_gate.py before semantic review
• MUST account for the full version scope (roadmap + architecture + all completed feature interfaces), not only the newest feature's full docs
• MUST produce a finding for every A1-A5, B1-B5, C1-C5 row
• MUST include row-level evidence in the final report
• If any row is missing evidence or judgment, return BLOCKED

Required Load Set (mandatory)

Before auditing, load and report:

• references/framework-implementation-readiness.md
• scripts/integration_gate.py
• docs/product/<project>-roadmap.md
• docs/product/<project>-architecture.md
• docs/prd/<version>/<newest-feature>-requirements.md (full doc for newest feature)
• docs/prd/<version>/<newest-feature>.md (full doc for newest feature)
• docs/design/<version>/<newest-feature>-ui-brief.md (if present, for newest feature)
• docs/prd/<version>/<feature-id>-interface.md for ALL completed features when produced

For the final run (after last feature), load full docs for all features instead of just the newest, plus all interface summaries when produced.

If any required file is not loaded, return BLOCKED.

Input Scope

Shared scope for every run:

• docs/product/<project>-roadmap.md
• docs/product/<project>-architecture.md
• docs/prd/<version>/<feature-id>-interface.md for all completed features when produced
• scripts/integration_gate.py

Incremental run additional full-doc scope:

• docs/prd/<version>/<newest-feature>-requirements.md
• docs/prd/<version>/<newest-feature>.md
• docs/design/<version>/<newest-feature>-ui-brief.md (if present)

Previously completed features may be represented by interface summaries only during incremental runs. They are still part of scope; they are not missing full-doc coverage unless this is the final run.

The orchestrator MUST invoke the script without --final-run for incremental checks, and WITH --final-run after the last feature completes. For every run it MUST pass --feature <feature-id> for each audited feature in the selected-version scope completed so far. For incremental checks it MUST also pass --newest-feature <feature-id>. For final checks it MUST add --final-run.

Final run additional full-doc scope:

• docs/prd/<version>/<feature-id>-requirements.md for all features
• docs/prd/<version>/<feature-id>.md for all features
• docs/design/<version>/<feature-id>-ui-brief.md for all features that have one

Output

• docs/reviews/<version>/integration-report.md (or integration-report-after-<feature-id>.md for incremental runs)

Two-Phase Protocol

Phase 1 — Script Extraction

Run scripts/integration_gate.py to extract candidate evidence, including:

• feature coverage matrix
• constants
• endpoints (including duplicate-endpoint signals — same METHOD /path defined by multiple sources)
• state signatures
• UI-to-PRD mappings
• deterministic coverage gaps
• A1-C5 script evidence rows (a row may explicitly say no deterministic signal was found)

The script emits SIGNALS_CLEAR or HOTSPOTS_FOUND as a script signal only (in a ## Script Verdict section). It does NOT emit PASS / CONFLICTS_FOUND. The gate verdict is set by Phase 2.

Phase 2 — Semantic Audit

Review the extracted evidence and decide whether candidate mismatches are:

• true conflicts
• semantic equivalents
• harmless naming differences
• intentional divergences that should be documented but not block

Required Checklist

Part A — Shared Concept Consistency

• A1 State machine unified
• A2 Constants consistent
• A3 Data model aligned
• A4 API contract consistent
• A5 Terminology unified

Part B — Completeness & Traceability

• B1 Roadmap → Requirements coverage
• B2 Requirements → PRD coverage
• B3 PRD → Interface Summary coverage
• B4 Architecture → PRD alignment
• B5 UI → PRD traceability

Part C — Cross-Feature Integration

• C1 No AC conflicts
• C2 User flows connectable
• C3 No duplicate effort
• C4 Dependency order clear
• C5 Error handling consistent

For incremental runs, C1/C2 are provisional checks based on the newest feature's full PRD plus previously completed features' interface summaries. Full PRD-to-PRD verification of C1/C2 is mandatory on the final run.

Output Contract (required)

The final report MUST include a row for every A1-C5 check with these columns:

Check	Script Evidence	Semantic Judgment	Result

The semantic-review report MUST carry forward the script's A1-C5 evidence rows. When the script has no deterministic signal for a row, preserve that fact explicitly rather than inventing evidence.

It MUST also include:

1. ## Summary
2. ## Part A — Shared Concepts
3. ## Part B — Traceability
4. ## Part C — Cross-Feature
5. ## Verdict — the gate verdict (PASS or CONFLICTS_FOUND). This is set by the audit skill in Phase 2 and is separate from the script's ## Script Verdict (which only emits SIGNALS_CLEAR / HOTSPOTS_FOUND).
6. ## Recommended Fix Order

A report that contains only the raw script output (no ## Verdict section, no per-row semantic judgments) is not a completed gate output and MUST be treated as BLOCKED.

Verdicts

• PASS — all 15 checks reviewed and no real conflicts remain
• CONFLICTS_FOUND — all 15 checks reviewed and one or more real conflicts remain
• BLOCKED — script not run, full input scope not read, row coverage incomplete, or evidence missing

Fail-Closed Rules

• PASS is forbidden if the report skips any A1-C5 row
• PASS is forbidden if a row has neither script-produced evidence nor an explicit "no deterministic signal" note
• PASS is forbidden if a row has no semantic judgment
• "Reviewed all docs, no major conflicts found" is not a valid completion state