ZCL Orchestrator (Codex Skill)

This skill is for the orchestrator (you), not the spawned mission agent.

Native Runtime Boundary (Current)

• ZCL has first-class native runtime execution in zcl suite run and campaign flows via runner.type=codex_app_server.
• Runtime selection is strategy-based and deterministic via ordered fallback chains (--runtime-strategies, ZCL_RUNTIME_STRATEGIES).
• Process runner mode remains supported as explicit fallback (--session-isolation process).
• Native runtime strategy health/failure classes are typed and surfaced in artifacts/JSON outputs.

Capability Matrix

Primary evidence:

• .zcl/.../tool.calls.jsonl
• .zcl/.../feedback.json

Operator Invocation Story

When asked to "run this through ZCL", use this order:

1. zcl init
2. Optional preflight:
   • zcl update status --json
   • zcl doctor --json
3. Optional single-attempt allocation path:
   • zcl attempt start --suite <suiteId> --mission <missionId> --prompt <text> --isolation-model native_spawn --json
4. Preferred native suite path:
   • zcl suite run --file <suite.(yaml|yml|json)> --session-isolation native --runtime-strategies codex_app_server --feedback-policy auto_fail --finalization-mode auto_from_result_json --result-channel file_json --result-min-turn 3 --campaign-id <campaignId> --progress-jsonl <path|-> --json
5. Process fallback path:
   • zcl suite run --file <suite.(yaml|yml|json)> --session-isolation process --feedback-policy auto_fail --finalization-mode auto_from_result_json --result-channel file_json --result-min-turn 3 --campaign-id <campaignId> --progress-jsonl <path|-> --shim tool-cli --json -- <runner-cmd> [args...]
6. First-class campaign path:
   • zcl campaign lint --spec <campaign.(yaml|yml|json)> --json
   • zcl campaign canary --spec <campaign.(yaml|yml|json)> --missions 3 --json
   • zcl campaign run --spec <campaign.(yaml|yml|json)> --json
   • zcl campaign resume --campaign-id <id> --json
   • zcl campaign status --campaign-id <id> --json
   • zcl campaign report --campaign-id <id> --allow-invalid --json
   • zcl campaign publish-check --campaign-id <id> --json
7. Validate/report from artifacts:
   • zcl report --strict <attemptDir|runDir>
   • zcl validate --strict <attemptDir|runDir>
   • zcl attempt explain --json <attemptDir>

Explicit finalization path for harness-aware prompts:

• zcl feedback --ok --result <text> or zcl feedback --fail --result-json <json>

Operator Guardrails

• Native suite attempts now anchor timeout windows before native turn execution; stalled native attempts classify as ZCL_E_RUNTIME_STALL.
• Native suite orchestration now guarantees terminal artifacts on failure paths (feedback.json + best-effort attempt.report.json) for post-mortem continuity.
• Campaign read/report/resume commands fail fast with ZCL_E_CAMPAIGN_STATE_DRIFT when persisted campaign.run.state.json disagrees with current spec mission selection (for example totalMissions=0 on a non-empty selection).
• For CI/automation, prefer zcl campaign report --allow-invalid --json to collect report payloads without branching on process exit codes; keep zcl campaign publish-check as the strict publish gate.

Campaign Guidance

Recommended native Codex flow:

flows:
  - flowId: codex-native
    runner:
      type: codex_app_server
      sessionIsolation: native
      runtimeStrategies: ["codex_app_server"]
      feedbackPolicy: auto_fail
      freshAgentPerAttempt: true

Mission-only recommended variant:

promptMode: mission_only
flows:
  - flowId: codex-native
    runner:
      type: codex_app_server
      sessionIsolation: native
      runtimeStrategies: ["codex_app_server"]
      finalization:
        mode: auto_from_result_json
        minResultTurn: 3
        resultChannel:
          kind: file_json
          path: mission.result.json

Prompt Policy

Use exactly one mode per campaign:

• promptMode: mission_only (preferred): mission intent + output contract only; no harness terms.
• promptMode: default: harness-aware prompts permitted.

Native Recommendation Criteria (Measured)

Codex native runtime is recommended when both hold in CI/nightly checks:

1. Reliability: native 20-attempt parallel smoke run success rate >= 95%.
2. Throughput: same run completes in <= 30 seconds on CI worker baseline.

Guard test path:

• internal/interfaces/cli/suite_run_integration_test.go (TestSuiteRun_NativeParallelUniqueSessions, TestSuiteRun_NativeSchedulerRateLimitIsDeterministic).

Templates

• examples/campaign.canonical.yaml
• examples/campaign.no-context.comparison.yaml
• examples/campaign.no-context.codex-exec.yaml
• examples/campaign.no-context.codex-subagent.yaml
• examples/campaign.no-context.claude-subagent.yaml
• examples/campaign.native.codex.minimal.yaml
• examples/campaign.native.codex.advanced.yaml
• docs/migration/shell-adapter-to-native-codex.md