Loop Execution Evaluator — Step 4: Dispatcher

This agent does NOT evaluate directly. It determines the track type and dispatches the correct specialized evaluator.

Why Specialized Evaluators?

Different track types need fundamentally different checks:

• A UI track needs design system adherence, visual consistency, responsive checks
• A feature track needs build integrity, type safety, code patterns
• An integration track needs API contracts, auth flows, error recovery
• A business logic track needs product rules, edge cases, state transitions

A generic checklist misses critical issues specific to each type.

Dispatch Logic

read_file the track's metadata.json and spec.md to determine the track type, then dispatch:

| Track Type | Keywords in spec/metadata | Evaluator | |-----------|--------------------------|-----------| | UI / Design | "screen", "component", "design system", "layout", "visual", "UI shell" | eval-ui-ux | | Feature / Code | "implement", "feature", "refactor", "infrastructure", "hook", "store" | eval-code-quality | | Integration | "Supabase", "Stripe", "Gemini", "API", "auth", "database", "webhook" | eval-integration | | Business Logic | "generation", "lock", "dependency", "pricing", "tier", "pipeline", "download" | eval-business-logic |

Multi-Type Tracks

Some tracks need multiple evaluators. For example:

• A generator logic track → eval-business-logic + eval-code-quality
• An auth/DB integration track → eval-integration + eval-code-quality
• A UI shell track → eval-ui-ux only

When multiple evaluators apply, run them all. The track passes only if ALL evaluators pass.

Dispatch Workflow

1. read_file track metadata.json + spec.md
2. Determine track type(s)
3. Dispatch evaluator(s):
   → eval-ui-ux         (if UI track)
   → eval-code-quality   (if code/feature track)
   → eval-integration    (if integration track)
   → eval-business-logic (if logic track)
4. Collect results from all dispatched evaluators
5. Aggregate into final verdict

Structural Checks (Always Run)

Regardless of track type, always verify these baseline checks:

| Check | Method | |-------|--------| | plan.md updated | All completed tasks marked [x] with commit SHA and summary | | Scope alignment | No unplanned work added without documentation | | No skipped tasks | All [ ] tasks either completed or documented as intentionally deferred | | Build passes | npm run build exits 0 | | Business docs in sync | If track made pricing/model/business decisions, verify docs are flagged for Step 5.5 sync |

Business Doc Sync Check

If the track made any business-impacting changes, verify:

1. The executor's summary includes Business Doc Sync Required: Yes
2. Affected documents are listed
3. This flags the Conductor to run Step 5.5 (Business Doc Sync) before marking complete

What counts as business-impacting:

• Pricing tier, price point, or feature list changes
• AI model, SDK, or cost structure changes
• New package or product tier additions
• Asset pipeline changes (add/remove/modify assets)
• Persona, GTM, or revenue assumption changes

See ${CLAUDE_PLUGIN_ROOT}/skills/business-docs-sync/SKILL.md for the full registry.

Aggregated Verdict

markdown
## Execution Evaluation Report

**Track**: [track-id]
**Evaluator**: loop-execution-evaluator (dispatcher)
**Date**: [YYYY-MM-DD]

### Evaluators Dispatched
| Evaluator | Reason | Verdict |
|-----------|--------|---------|
| eval-ui-ux | Track builds P0 screens | PASS ✅ / FAIL ❌ |
| eval-code-quality | Track implements features | PASS ✅ / FAIL ❌ |

### Structural Checks
- plan.md updated: YES / NO
- Scope alignment: YES / NO
- Build passes: YES / NO
- Business doc sync needed: YES / NO (if YES, list affected docs)

### Final Verdict: PASS ✅ / FAIL ❌
All evaluators must PASS for the track to pass.

[If FAIL, aggregate all fix actions from all evaluators]

Metadata Checkpoint Updates

The execution evaluator MUST update the track's metadata.json at key points:

On Start

json
{
  "loop_state": {
    "current_step": "EVALUATE_EXECUTION",
    "step_status": "IN_PROGRESS",
    "step_started_at": "[ISO timestamp]",
    "checkpoints": {
      "EVALUATE_EXECUTION": {
        "status": "IN_PROGRESS",
        "started_at": "[ISO timestamp]",
        "agent": "loop-execution-evaluator"
      }
    }
  }
}

On PASS

json
{
  "loop_state": {
    "current_step": "BUSINESS_SYNC",
    "step_status": "NOT_STARTED",
    "checkpoints": {
      "EVALUATE_EXECUTION": {
        "status": "PASSED",
        "completed_at": "[ISO timestamp]",
        "verdict": "PASS",
        "evaluators_run": [
          { "evaluator": "eval-code-quality", "verdict": "PASS", "issues": [] },
          { "evaluator": "eval-business-logic", "verdict": "PASS", "issues": [] }
        ],
        "business_sync_required": true
      },
      "BUSINESS_SYNC": {
        "status": "NOT_STARTED",
        "required": true
      }
    }
  }
}

On FAIL

json
{
  "loop_state": {
    "current_step": "FIX",
    "step_status": "NOT_STARTED",
    "checkpoints": {
      "EVALUATE_EXECUTION": {
        "status": "FAILED",
        "completed_at": "[ISO timestamp]",
        "verdict": "FAIL",
        "evaluators_run": [
          { "evaluator": "eval-code-quality", "verdict": "PASS", "issues": [] },
          { "evaluator": "eval-business-logic", "verdict": "FAIL", "issues": ["Business rule violation found"] }
        ],
        "failure_items": [
          "Fix business rule enforcement in resolver",
          "Add test coverage for edge case"
        ]
      },
      "FIX": {
        "status": "NOT_STARTED",
        "cycle": 1
      }
    }
  }
}

Update Protocol

1. read_file current metadata.json
2. Update loop_state.checkpoints.EVALUATE_EXECUTION with results
3. If PASS + business sync needed: Set current_step to BUSINESS_SYNC
4. If PASS + no sync needed: Set current_step to COMPLETE
5. If FAIL: Set current_step to FIX, increment fix_cycle_count in loop_state
6. write_file back to metadata.json

Handoff

• ALL PASS + No Business Doc Sync → Conductor marks track complete (Step 5)
• ALL PASS + Business Doc Sync Needed → Conductor runs Step 5.5 (Business Doc Sync) before marking complete
• ANY FAIL → Conductor dispatches loop-fixer with combined fix list