STEP 1579 (b) — Plan Runner Msg (arc 3/3 完)

STEP 1569 Ophyd 読解 top 3 の (b):

Msg indirection layer: rei-aios MCP tool 群を 「plan は tool call を yield → SafetyGate dispatch」 の 2 層化 — introspection / dry-run / 安全停止を得る

Bluesky の RunEngine は 10 年運用実績のある pattern:

# Bluesky
def my_plan():
    yield Msg('set', motor, 5)
    yield Msg('trigger', detector)
    yield Msg('read', detector)
RE(my_plan())   # RunEngine が dispatch

これを Rei stack (TypeScript / MCP tool 群 / D-FUMT₈ verdict) 向けに portable な形で minimum viable 実装。

const runner = new PlanRunner({
  dryRun: false,                          // true = tool 呼出 skip、 introspection のみ
  toolHandler: (name, args) => ...,       // 実 dispatch (MCP integration 点)
  safetyGate: (checkMsg) => ...,          // STEP 1345 SafetyGate 統合点
  waitFn: (ms) => new Promise(...),       // 差替可 (test 時 mock 化)
});

const trace: TraceEntry[] = await runner.run(myPlan());
// trace = 各 Msg の timestamp / result / noDataDetected を記録

const summary = summarizeTrace(trace);
// { totalMsgs, byType, noDataHits, safetyChecks, approvedChecks }

disk_health_verdict (STEP 1567、 Layer 1 sensor tool) を Msg indirection 越しに呼出:

function* diskDiagnosticPlan(input): Plan {
  yield dryRunHintMsg('disk-diagnostic plan v0.1');
  yield checkMsg('disk_health_verdict', input, 'caution',
                 'SMART attribute read requires privileged access');
  const r = yield callMsg('disk_health_verdict', input, 'primary-disk-scan');
  yield checkpointMsg('post-primary-scan', {
    verdict: r.verdict, reason: r.reason
  });
  // 条件分岐 follow-up
  if (r.verdict === 'ZERO' && r.reason === 'no_data') {
    yield callMsg('log_and_retry', {...}, 'no-data-retry');
  } else if (r.verdict === 'FALSE') {
    yield callMsg('escalate_backup_verify', {...}, 'critical-escalation');
  } else {
    yield callMsg('log_healthy', {...}, 'log-only');
  }
}

実行 3 パターン確認済:

1. Healthy disk (reallocatedSectors=0, temperature=40): verdict=TRUE → log_healthy 分岐
2. NoData (empty input): verdict=ZERO, reason=no_data、 noDataDetected=true flag → log_and_retry 分岐
3. Dry-run: 全 tool 呼出 skip、 introspection のみ (verdict undefined → else branch)

1. Msg constructors (5 types)
2. PlanRunner basic run — 1 checkpoint
3. toolHandler dispatch
4. SafetyGate rejection (block → throw)
5. Dry-run mode — 0 tool invocations
6. NoDataResult 自動検出 (noDataDetected flag)
7. Trace summarization (byType / noDataHits / safetyChecks)
8. Pilot plan full 5-Msg sequence
9. Pilot plan dry-run introspection
10. Pilot plan no-data path routing