ln-310-multi-agent-validator

What this skill does

> **Paths:** File paths (`references/`, `../ln-*`) are relative to this skill directory.

**Type:** L2 Coordinator
**Category:** 3XX Planning

# Multi-Agent Validator

Evaluation-platform coordinator for:
- `mode=story`
- `mode=plan_review`

This skill uses the evaluation platform for:
- mandatory official-doc, MCP Ref, Context7, and current-web research
- parallel read-only evidence lanes
- sequential documentation, repair, merge, refinement, and approval
- runtime-backed worker plans, worker summaries, agent sync, and cleanup verification

## Inputs

| Input | Required | Source | Description |
|-------|----------|--------|-------------|
| `storyId` | `mode=story` | args, git branch, kanban, user | Story to validate |
| `plan {file}` | `mode=plan_review` | args or auto | Plan file to validate |

Mode detection:
- `plan` or `plan {file}` -> `mode=plan_review`
- otherwise -> `mode=story`

## Mandatory Read

**MANDATORY READ:** Load `references/environment_state_contract.md`, `references/storage_mode_detection.md`, `references/input_resolution_pattern.md`
**MANDATORY READ:** Load `references/evaluation_coordinator_runtime_contract.md`, `references/evaluation_summary_contract.md`, `references/evaluation_parallelism_policy.md`, `references/evaluation_research_contract.md`
**MANDATORY READ:** Load `references/agent_delegation_pattern.md`
**MANDATORY READ:** Load `references/penalty_points.md`
**MANDATORY READ:** Load `references/researchgraph_mcp_usage.md` when researchgraph files changed or the target claims hypothesis, goal, benchmark, or proposal readiness.
Conditional read: load `references/phase2_research_audit.md` only when the coordinator performs inline criteria mapping instead of consuming ln-312 findings summaries.

Agent review policy: run health check, record skipped reason when no advisor is available, verify every advisor claim before merge, and treat transport/auth/tool failures as operator evidence rather than domain findings. Load `references/agent_review_workflow.md` only when debugging lifecycle/liveness details outside the evaluation runtime.

## Worker Set

The coordinator uses these evaluation workers:
- `ln-311-review-research-worker`
- `ln-312-review-findings-worker`
- `ln-313-review-docs-worker`
- `ln-314-review-repair-worker`
- `ln-315-review-merge-worker`
- `ln-316-review-refinement-worker`

## Worker Invocation (MANDATORY)

**Host Skill Invocation:** `Skill(skill: "...", args: "...")` is mandatory delegation.
- Claude: call the Skill tool exactly as shown.
- Codex: if no Skill tool exists, locate the named skill in available skills, read its `SKILL.md`, treat `args` as `$ARGUMENTS`, execute that skill workflow, then return here with its result/artifact.
- Do not inline worker logic or mark the worker complete without executing the target skill.

Use the Skill tool for delegated workers. Do not inline worker logic inside the coordinator.

TodoWrite format (mandatory):
- `Resolve target and build runtime manifest`
- `Load target artifacts and metadata`
- `Launch external agents and verify health`
- `Run research and findings workers in parallel`
- `Generate documentation updates`
- `Apply accepted low-risk repairs`
- `Sync agents and merge all evidence`
- `Run refinement (MANDATORY in ALL modes when advisor available — do NOT skip)`
- `Compute verdict and write review output`
- `Verify runtime cleanup and self-check`

Representative invocations:

```text
Skill(skill: "ln-311-review-research-worker", args: "{identifier} research")
Skill(skill: "ln-312-review-findings-worker", args: "{identifier} findings")
Skill(skill: "ln-313-review-docs-worker", args: "{identifier} docs")
Skill(skill: "ln-314-review-repair-worker", args: "{identifier} repair")
Skill(skill: "ln-315-review-merge-worker", args: "{identifier} merge")
Skill(skill: "ln-316-review-refinement-worker", args: "{identifier} refinement")
```

## Runtime Contract

**MANDATORY READ:** Load `references/loop_health_contract.md`

Runtime family:
- `evaluation-runtime`

Identifier:
- `story-{storyId}` for story mode
- `plan-{slug}` for plan review

Phase order:
1. `PHASE_0_CONFIG`
2. `PHASE_1_DISCOVERY`
3. `PHASE_2_AGENT_LAUNCH`
4. `PHASE_3_EVIDENCE_LANES`
5. `PHASE_4_DOCS`
6. `PHASE_5_REPAIR`
7. `PHASE_6_MERGE`
8. `PHASE_7_REFINEMENT`
9. `PHASE_8_APPROVAL`
10. `PHASE_9_SELF_CHECK`

Phase policy:
- `delegate_phases = [PHASE_3_EVIDENCE_LANES, PHASE_4_DOCS, PHASE_5_REPAIR, PHASE_6_MERGE, PHASE_7_REFINEMENT]`
- `aggregate_phase = PHASE_6_MERGE`
- `report_phase = PHASE_8_APPROVAL`
- `cleanup_phase = PHASE_9_SELF_CHECK`
- `self_check_phase = PHASE_9_SELF_CHECK`
- `agent_resolve_before = [PHASE_6_MERGE]`
- `required_phases_when_advisor_available = [PHASE_7_REFINEMENT]`

## Parallelism Rules

Allowed overlap:
- external agents
- `ln-311`
- `ln-312`
- local repo inspection and evidence gathering

Sequential only:
- `ln-313`
- `ln-314`
- `ln-315`
- `ln-316`
- approval and status mutation

## Workflow

### Phase 0: Config

1. Resolve `mode`, identifier, and storage mode.
2. Resolve story or plan target.
3. Build evaluation runtime manifest with:
   - `expected_agents`
   - `required_research=true`
   - exact `phase_order`
   - `phase_policy`
   - report path
4. Start runtime:

```bash
node references/scripts/evaluation-runtime/cli.mjs start \
  --skill ln-310 \
  --identifier {identifier} \
  --manifest-file .hex-skills/evaluation/{identifier}_manifest.json
```

5. Checkpoint Phase 0.

### Phase 1: Discovery

1. Materialize the exact target artifact.
2. Load only the metadata needed for the current mode.
3. In `mode=story`, resolve Story and child tasks.
4. In `mode=plan_review`, resolve the plan file.
5. If researchgraph files changed or the target cites `H##`, `G##`, run IDs, benchmark manifests, or readiness claims, run read-only researchgraph verification/audits and attach the result as validation evidence.
6. Checkpoint Phase 1 with resolved refs.

### Phase 2: Agent Launch

1. Run agent health check.
2. Exclude disabled agents from `.hex-skills/environment_state.json`.
3. If no agents are available:
   - record `agents_skipped_reason`
   - checkpoint Phase 2
   - continue
4. Otherwise:
   - build per-agent prompts
   - launch each available agent
   - register each launched agent:

```bash
node references/scripts/evaluation-runtime/cli.mjs register-agent \
  --skill ln-310 \
  --identifier {identifier} \
  --agent {name} \
  --prompt-file {promptPath} \
  --result-file {resultPath} \
  --metadata-file {metadataPath}
```

5. Checkpoint Phase 2 with `health_check_done`, `agents_available`, `agents_required`, and optional `agents_skipped_reason`.
6. Classify each external agent result before domain verdict:
   - `rate_limited`, `tool_missing`, `auth_missing`, `permission_denial`, and `asked_question` are transport/operator states.
   - Do not convert them into `NO-GO` without domain evidence from artifacts or findings.
   - Record loop health for repeated advisor/session failures and pause when retry usefulness is exhausted.

### Phase 3: Evidence Lanes

This phase is the mandatory parallel evidence barrier.

1. Build `worker_plan` with:
   - `ln-311` lane `research` (mandatory)
   - `ln-312` lane `findings` (mandatory)
2. Launch all planned workers in parallel.
3. While those workers run, continue local repo inspection and collect additional evidence.
4. Sync agents opportunistically, but do not block on them until merge.
5. Record each worker summary with:

```bash
node references/scripts/evaluation-runtime/cli.mjs record-worker-result \
  --skill ln-310 \
  --identifier {identifier} \
  --payload-file {childSummaryArtifactPath}
```

Research is mandatory in every mode:
- official documentation or standards
- MCP Ref
- Context7 when a library or framework is involved
- current web best-practice research

For `mode=story`, findings must still produce penalty-point evidence and coverage analysis.

### Phase 4: Docs

1. In `mode=story`, run `ln-313-review-docs-worker` when documentati