Files
pixelheros/CCGS Skill Testing Framework/skills/review/architecture-review.md
2026-05-15 14:52:29 +08:00

7.4 KiB

Skill Test Spec: /architecture-review

Skill Summary

/architecture-review is an Opus-tier skill that validates a technical architecture document against the project's 8 required architecture sections and checks that it is internally consistent, non-contradictory with existing ADRs, and correctly targeting the pinned engine version. It produces a verdict of APPROVED / NEEDS REVISION / MAJOR REVISION NEEDED.

In full review mode, the skill spawns two director gate agents in parallel: TD-ARCHITECTURE (technical-director) and LP-FEASIBILITY (lead-programmer). In lean or solo mode, both gates are skipped and noted. The skill is read-only — no files are written.


Static Assertions (Structural)

Verified automatically by /skill-test static — no fixture needed.

  • Has required frontmatter fields: name, description, argument-hint, user-invocable, allowed-tools
  • Has ≥2 phase headings
  • Contains verdict keywords: APPROVED, NEEDS REVISION, MAJOR REVISION NEEDED
  • Does NOT require "May I write" language (read-only skill)
  • Has a next-step handoff at the end
  • Documents gate behavior: TD-ARCHITECTURE + LP-FEASIBILITY in full mode; skipped in lean/solo

Director Gate Checks

In full mode: TD-ARCHITECTURE (technical-director) and LP-FEASIBILITY (lead-programmer) are spawned in parallel after the skill reads the architecture doc.

In lean mode: both gates are skipped. Output notes: "TD-ARCHITECTURE skipped — lean mode" and "LP-FEASIBILITY skipped — lean mode".

In solo mode: both gates are skipped with equivalent notes.


Test Cases

Case 1: Happy Path — Complete architecture doc in full mode

Fixture:

  • docs/architecture/architecture.md exists with all 8 required sections populated
  • All sections reference the correct engine version from docs/engine-reference/
  • No contradictions with existing Accepted ADRs in docs/architecture/
  • production/session-state/review-mode.txt contains full

Input: /architecture-review docs/architecture/architecture.md

Expected behavior:

  1. Skill reads the architecture document
  2. Skill reads existing ADRs for cross-reference
  3. Skill reads engine version reference
  4. TD-ARCHITECTURE and LP-FEASIBILITY gate agents spawn in parallel
  5. Both gates return APPROVED
  6. Skill outputs section-by-section completeness check (8/8 sections present)
  7. Verdict: APPROVED

Assertions:

  • All 8 required sections are checked and reported
  • TD-ARCHITECTURE and LP-FEASIBILITY spawn in parallel (not sequentially)
  • Verdict is APPROVED when all sections are present and no conflicts exist
  • Skill does NOT write any files
  • Next-step handoff to /create-control-manifest or /create-epics is present

Case 2: Failure Path — Missing required sections

Fixture:

  • docs/architecture/architecture.md exists but is missing at least 2 required sections (e.g., no data model section, no error handling section)
  • production/session-state/review-mode.txt contains full

Input: /architecture-review docs/architecture/architecture.md

Expected behavior:

  1. Skill reads the document and identifies missing sections
  2. Section completeness shows fewer than 8/8 sections present
  3. Missing sections are listed by name with specific remediation guidance
  4. Verdict: MAJOR REVISION NEEDED (≥2 missing sections)

Assertions:

  • Verdict is MAJOR REVISION NEEDED (not APPROVED or NEEDS REVISION) for ≥2 missing sections
  • Each missing section is named explicitly in the output
  • Remediation guidance is specific (what to add, not just "add missing sections")
  • Skill does NOT pass a document missing required sections

Case 3: Partial Path — Architecture contradicts an existing ADR

Fixture:

  • docs/architecture/architecture.md exists with all 8 sections present
  • One Accepted ADR in docs/architecture/ establishes a constraint that the architecture doc contradicts (e.g., ADR-001 mandates ECS pattern; architecture.md describes a different pattern for the same system)

Input: /architecture-review docs/architecture/architecture.md

Expected behavior:

  1. Skill reads the architecture doc and all existing ADRs
  2. Conflict is detected between the architecture doc and the named ADR
  3. Conflict entry names: the ADR number/title, the contradicting sections, and impact
  4. Verdict: NEEDS REVISION (conflict exists but structure is otherwise sound)

Assertions:

  • Verdict is NEEDS REVISION (not MAJOR REVISION NEEDED for a single contradiction)
  • The specific ADR number and title are named in the conflict entry
  • The contradicting sections in both documents are identified
  • Skill does NOT auto-resolve the contradiction

Case 4: Edge Case — File not found

Fixture:

  • The path provided does not exist in the project

Input: /architecture-review docs/architecture/nonexistent.md

Expected behavior:

  1. Skill attempts to read the file
  2. File not found
  3. Skill outputs a clear error naming the missing file
  4. Skill suggests checking docs/architecture/ or running /create-architecture
  5. Skill does NOT produce a verdict

Assertions:

  • Skill outputs a clear error when the file is not found
  • No verdict is produced (APPROVED / NEEDS REVISION / MAJOR REVISION NEEDED)
  • Skill suggests a corrective action
  • Skill does NOT crash or produce a partial report

Case 5: Director Gate — Full mode spawns both gates; solo mode skips both

Fixture (full mode):

  • docs/architecture/architecture.md exists with all 8 sections
  • production/session-state/review-mode.txt contains full

Full mode expected behavior:

  1. TD-ARCHITECTURE gate spawns
  2. LP-FEASIBILITY gate spawns in parallel with TD-ARCHITECTURE
  3. Both gates complete before verdict is issued

Assertions (full mode):

  • TD-ARCHITECTURE and LP-FEASIBILITY both appear in the output as completed gates
  • Both gates spawn in parallel (not one after the other)
  • Verdict reflects gate feedback

Fixture (solo mode):

  • Same architecture doc
  • production/session-state/review-mode.txt contains solo

Solo mode expected behavior:

  1. Skill reads the architecture doc
  2. Gates are NOT spawned
  3. Output notes: "TD-ARCHITECTURE skipped — solo mode" and "LP-FEASIBILITY skipped — solo mode"
  4. Verdict is based on structural checks only

Assertions (solo mode):

  • Neither TD-ARCHITECTURE nor LP-FEASIBILITY appears as an active gate
  • Both skipped gates are noted in the output
  • Verdict is still produced based on the structural check alone

Protocol Compliance

  • Does NOT write any files (read-only skill)
  • Presents section completeness check before issuing verdict
  • TD-ARCHITECTURE and LP-FEASIBILITY spawn in parallel in full mode
  • Skipped gates are noted by name and mode in lean/solo output
  • Verdict is one of exactly: APPROVED, NEEDS REVISION, MAJOR REVISION NEEDED
  • Ends with next-step handoff appropriate to verdict

Coverage Notes

  • The 8 required architecture sections are project-specific; tests use the section list defined in the skill body — not re-enumerated here.
  • Engine version compatibility checking (cross-referencing docs/engine-reference/) is part of Case 1's happy path but not independently fixture-tested.
  • RTM (requirement traceability matrix) mode is a separate concern covered by the /architecture-review skill's own rtm argument mode, not tested here.