Coherence audit

Coherence Audit

Run a coherence audit after every layer's merges. This catches cross-session AND cross-codebase quality issues (DRY violations, orphaned code, convention drift, security gaps) before they compound.

The QA gate and knowledge curation both depend on the COHERENCE-REPORT. No report = no proceeding.

---

Determine Applicable Dimensions

Scan the layer's diff to decide which conditional dimensions to check:

git diff $CHECKPOINT..HEAD --stat

Use your judgment based on what changed. Glob patterns are guidance, not programmatic filters — when in doubt, include the dimension.

Always include (every layer): code-organization, dry-code-reuse, error-handling, security, test-quality

Include if layer touched that area:

• API routes modified (**/api/**, **/routes/**, **/handlers/**) → include api-design
• Frontend files modified (*.tsx, *.jsx, **/components/**, **/pages/**) → include frontend-quality
• Type definitions modified (**/types/**, **/schema/**, **/models/**) → include type-safety
• Data models/migrations modified (**/migrations/**, **/schema/**, **/prisma/**, **/drizzle/**) → include data-modeling
• Server/query code modified (**/api/**, **/routes/**, **/queries/**, **/db/**) → include performance
• Request handling modified (**/api/**, **/routes/**, **/handlers/**, **/middleware/**) → include observability
• Config files modified (**/.env*, **/config/**, **/settings/**) → include config-management

Never in checkpoint mode: developer-infrastructure (deferred to completion-only full-spectrum audit)

A 1-session layer still gets all core dimensions. Conditional dimensions are still evaluated based on the diff — not the layer size. Single-session layers can have cross-LAYER drift against code from prior layers.

---

Run the Audit

Invoke the audit skill in checkpoint mode via the Skill tool:

/autoboard:audit --checkpoint $CHECKPOINT --dimensions {selected dimensions}

The audit skill spawns parallel dimension agents (one per dimension), each scoped to the layer's diff but cross-referencing the full codebase. It produces a ~~~COHERENCE-REPORT block.

---

Process COHERENCE-REPORT

You are not done. The COHERENCE-REPORT is an intermediate result, not a deliverable. Act on it immediately - do not stop, do not summarize to the user and wait, do not end your turn.

Save Report to Disk

Write the COHERENCE-REPORT to docs/autoboard/{slug}/sessions/coherence-L{N}.md so fixer briefs can reference it by path instead of pasting it inline.

Dispatch Coherence Screener

Dispatch the autoboard:coherence-screener agent via the Agent tool with model code-review-model and these inputs:

• COHERENCE-REPORT text (the full ~~~COHERENCE-REPORT block)
• Design doc path: docs/autoboard/{slug}/design.md
• Manifest path: docs/autoboard/{slug}/manifest.md

The screener has the receiving-review decision tree baked in. It applies the decision tree to each finding and returns a structured verdict with surviving and dismissed findings. You do NOT need to load /autoboard:receiving-review for coherence processing.

Route Based on Verdict

CLEAR (no findings survive): Proceed to QA gate or knowledge curation. No tracking action needed - coherence issues are not pre-created, so there is nothing to close.

FIX_NEEDED (findings survive): Route to the coherence-fixer. The fixer handles all surviving findings - BLOCKING and INFO alike. The pipeline is gated until all surviving findings are resolved.

Do NOT attempt to fix issues yourself - the fixer needs the full session workflow. Do NOT report the findings to the user and stop - the fixer must be dispatched now.

Tracking (if active): Create an on-demand issue for the fixer:

1. Create the issue:

   create-ticket("Coherence Fix: Layer {N}", COHERENCE-REPORT contents)
   

2. Add to the project and set status to Implementing:

   move-ticket(coherence-issue, Implementing)
   

3. Note the issue number and item ID - the coherence-fixer skill needs these for the fixer brief's ## Tracking section.

Key design decision: Coherence audit issues are NOT pre-created during setup. They are created on-demand only when findings survive screening and trigger a fixer. Most layers pass cleanly - pre-creating issues would just create noise that gets immediately closed.

---

Anti-Patterns

Thought that means STOP	Reality
"Sessions merged, I'll move on to QA"	The coherence audit must run after merge, before anything else.
"This layer is simple, coherence audit isn't needed"	Every layer gets audited. Complexity is not the criterion.
"I'll do the coherence audit retroactively"	Retroactive audits don't block. The audit must block before QA.
"QA will catch cross-session issues"	QA tests functionality. Coherence catches architecture drift — different concern.
"I already checked the merges look clean"	Your eyeball check is not a structured multi-dimension audit. Run the skill.
"The audit found issues, let me report them"	Reports are intermediate results, not deliverables. Dispatch the fixer immediately.
"These are just INFO items, no fixer needed"	All findings that survive pre-screening get fixed. INFO items are cheap — dispatch the fixer.