Post-Wave 2: Consistency audit findings (32 skills, 17 plugins, 12 commands)

[dean0x]

Summary

Comprehensive consistency audit of all 32 skills, 11 agents, 12 commands (+ 5 teams variants), 17 plugin manifests, and the marketplace registry after Wave 2 (project knowledge system). Found 17 findings across 3 priority levels.

HIGH Priority (6 findings)

• H1 — knowledge-persistence skill convention violations: No references/ directory, non-standard description format, only 89 lines (target 120-150), not listed in docs/reference/skills-architecture.md (Fixed in Persistent Project Knowledge (Stateful Agents / Persistent Minds) #99 branch: description format, references/ dir, expanded content, added to skills-architecture.md)
• H2 — ambient-router skill description format: Doesn't follow "This skill should be used when..." format, only 96 lines (Fixed in fix: consistency audit batch — H2, M3, M7 from #142 #148: description updated to standard format)
• H3 — /audit-claude missing frontmatter: Only command without YAML frontmatter (Fixed in fix: Wave 2 post-merge cleanup — knowledge persistence & command consistency #145)
• H4 — /debug synthesizer spawning: Uses subagent_type="general-purpose" instead of subagent_type="Synthesizer" — bypasses the dedicated agent (Fixed in fix: Wave 2 post-merge cleanup — knowledge persistence & command consistency #145)
• H5 — CLAUDE.md stale skill count: Says "31 skills" but there are 32 (knowledge-persistence added without updating count) (Fixed in Persistent Project Knowledge (Stateful Agents / Persistent Minds) #99 branch: 31→32)
• H6 — CLAUDE.md wrong core/optional split: Says "9 core + 8 optional" but actual code has 8 core + 9 optional (devflow-audit-claude is optional) (Fixed in fix: metadata and documentation accuracy — 5 quick wins #146)

MEDIUM Priority (8 findings)

• M1 — Oversized agents: Git agent (277 lines) and Synthesizer (211 lines) significantly exceed 120-line Worker target (Closed: acceptable — both agents are under 350-line threshold and their complexity warrants the length)
• M2 — Agents missing skill declarations: Skimmer and Simplifier have no skills declared in frontmatter (Fixed in fix: Wave 2 post-merge cleanup — knowledge persistence & command consistency #145)
• M3 — Simplifier agent missing sections: No Output section and no Boundaries section (unlike all other agents) (Fixed in fix: consistency audit batch — H2, M3, M7 from #142 #148: Output and Boundaries sections added matching Scrutinizer pattern)
• M4 — /specify-teams.md missing skill loading: Teammates have no skill reading instructions (unlike /implement-teams.md) (Closed: by-design — specify doesn't load skills, consistent with non-teams variant)
• M5 — /debug commands missing knowledge awareness: Neither /debug nor /debug-teams load project knowledge (pitfalls.md) — pitfalls could directly relate to bugs being investigated (Fixed in Persistent Project Knowledge (Stateful Agents / Persistent Minds) #99 branch: Phase 1 added to both debug.md and debug-teams.md)
• M6 — audit-claude plugin.json agents mismatch: Declares agents: [] but plugins.ts declares agents: ['claude-md-auditor'] (Closed: by-design — plugin.json agents array is for shared agents only; claude-md-auditor is a plugin-specific agent in plugins/devflow-audit-claude/agents/)
• M7 — 3 plugins missing standard metadata: devflow-ambient, devflow-audit-claude, and devflow-self-review lack homepage, repository, license, and keywords fields (Fixed in fix: consistency audit batch — H2, M3, M7 from #142 #148: metadata added to all 3 plugin.json files)
• M8 — 4 skills missing from architecture reference: knowledge-persistence, ambient-router, search-first, test-driven-development not in docs/reference/skills-architecture.md (Fixed in fix: metadata and documentation accuracy — medium batch #147)

LOW Priority (3 findings)

• L1 — 15 skills exceed 150-line target: react (276), frontend-design (254), accessibility (229) are worst offenders; content could move to references/ (Skipped: acceptable — all have references/ directories; cosmetic)
• L2 — Undocumented allowed-tools exceptions: 10 of 32 skills deviate from documented Read, Grep, Glob (legitimate but undocumented) (Fixed in fix: consistency audit batch — H2, M3, M7 from #142 #148: CLAUDE.md updated to document intentional exceptions)
• L3 — Plugin description divergence: Description text differs across 3 sources (marketplace.json, plugin.json, plugins.ts) for 9 of 17 plugins (Fixed in fix: consistency audit batch — H2, M3, M7 from #142 #148: synced all 3 sources using plugin.json as canonical)

Verified Correct (no action needed)

• All 32 skills have Iron Law blockquotes (100%)
• All skills have user-invocable: false (100%)
• 31/32 skills have references/ directories
• No orphaned teams variants — all 5 have matching base files
• All agents avoid duplicating skill content
• npm run build succeeds (45 skill copies + 21 agent copies)
• Version 1.5.0 synchronized across all manifests
• All 17 plugins properly registered in CLI, marketplace, and as directories

Progress

Priority	Total	Done	Remaining
HIGH	6	6	0
MEDIUM	8	8	0
LOW	3	2	1 (skipped)
Total	17	16	1

[dean0x]

Batch Fix Triage Summary

Resolved all HIGH and MEDIUM findings. 3 LOW-priority items remain for future cleanup.

Fixed in this batch (3 code changes)

• H2: ambient-router description updated to standard "This skill should be used when..." format
• M3: Simplifier agent now has ## Output and ## Boundaries sections matching Scrutinizer pattern
• M7: Added homepage/repository/license/keywords to devflow-ambient, devflow-audit-claude, devflow-self-review plugin.json files

Already fixed on main (PRs #145-147)

• H3: /audit-claude frontmatter added (fix: Wave 2 post-merge cleanup — knowledge persistence & command consistency #145)
• H4: /debug synthesizer spawning fixed (fix: Wave 2 post-merge cleanup — knowledge persistence & command consistency #145)
• H6: Core/optional split corrected (fix: metadata and documentation accuracy — 5 quick wins #146)
• M2: Skimmer/Simplifier skill declarations added (fix: Wave 2 post-merge cleanup — knowledge persistence & command consistency #145)
• M8: 4 missing skills added to architecture reference (fix: metadata and documentation accuracy — medium batch #147)

Closed as acceptable/by-design

• M1: Git (277 lines) and Synthesizer (211 lines) are under 350-line threshold — complexity warrants the length
• M4: /specify-teams.md not loading skills is consistent with the non-teams variant (by-design)
• M6: plugin.json agents: [] is correct — claude-md-auditor is a plugin-specific agent (lives in plugins/devflow-audit-claude/agents/), not a shared agent. The agents array in plugin.json is only for shared agents distributed at build time.

Remaining (LOW priority)

• L1: 9 skills exceed 150-line target
• L2: Undocumented allowed-tools exceptions
• L3: Plugin description divergence across sources

[dean0x]

All actionable findings resolved. 16/17 done — L1 (skill line counts) skipped as acceptable. See #148 for the final batch.