maplab-durable-job-orchestrator

Imported from page1010/maplab-ai-handbook (.agents/skills/maplab-durable-job-orchestrator/SKILL.md). Install upstream with npx skills add page1010/maplab-ai-handbook --skill maplab-durable-job-orchestrator. Copyright stays with the author.

MAPLAB durable job orchestrator

The Owner states the outcome; the agent chooses DeerFlow, Codex/Hermes continuation, and domain workers. Never require the Owner to remember /research-*, a model name, or a worker name.

Route automatically

Create or continue a durable job when one or more apply:

• the user asks to keep going, run several rounds, finish end to end, or return only when a result is ready;
• the work spans research plus generation, QA, upload, browser work, or more than one role;
• a provider job or render may outlive the current turn;
• progress has stopped before the stated acceptance condition;
• A8 is asked to generate music/video or upload a reviewable YouTube draft;
• Hermes is asked to train, evaluate, or repair LINE replies over several rounds.

A small one-turn lookup or edit stays synchronous. An explicit /research-public remains a diagnostic override, not the normal user experience.

Build the control loop

1. Read AGENT_CORE.md when present, CURRENT_STATUS.md, pitfalls.md, and the active Task Card.
2. Create or load the durable job under workbook/reviews/MAPLAB-DURABLE-JOBS/. Follow references/job-contract.md.
3. Record the user's outcome, acceptance evidence, authorization already granted, budgets, domain owner, next bounded action, and Resume Prompt.
4. Choose execution surfaces without exposing that choice to the Owner:
   • DeerFlow: public/synthetic research, decomposition, comparison, and next-action reasoning.
   • Domain worker: private assets, customer data, LINE corpus, A8 rendering, platform adapters, and fixed local scripts.
   • Codex thread heartbeat: wake and continue cross-tool, authenticated-browser, or repo work that a one-shot process cannot finish.
5. After every bounded action, verify an artifact or live surface, update the job receipt, and either continue automatically or enter a named gate.
6. Finish only when the Owner can inspect the promised artifact/link and the receipt proves it. A plan, worker chat, process exit, or API 200 alone is not completion.

Break objective-level plateaus

Method novelty is not outcome progress. Before every bounded action, compare the last three receipts against the job's Owner-facing acceptance and primary objective metrics, not only method fingerprints or test counts. Read references/objective-plateau-sop.md when two consecutive actions changed code, infrastructure, prompts, or synthetic proof without advancing an Owner-facing acceptance item.

• Record objective_metrics_before, objective_metrics_after, owner_acceptance_delta, unlocked_next_action, and attempt_consumed.
• Two consecutive zero objective deltas trigger the first-principles five questions before any further execution, even when fingerprints differ.
• Do not spend a domain attempt on the review/re-route itself. State attempt_consumed=false and preserve the prior attempt count. This exemption may be used once for the detected plateau; the next executed domain experiment consumes one attempt. A pure poll/readback of an already-running external action does not.
• Defer or split infrastructure/security work that is not a current blocker. A synthetic gate may continue only when the receipt names the exact immediately executable Owner-facing action it unlocks, proves every other prerequisite is satisfied, and gives a fixed stop condition. A split supporting job gets an explicit lower priority and attempt/spend cap and may not displace the main job unless an urgent verified safety issue or the Owner changes priority.
• After the circuit breaker, the next action must be the smallest falsifiable experiment against the real objective, not a broader version of the same supporting system.

Use authorization already present

Do not ask again for an action the Owner explicitly included in the goal. Examples:

• “生歌＋影片” authorizes the necessary bounded generation calls for that job.
• “上傳到 YouTube 給我看” authorizes a private/unlisted draft upload and readback; it does not authorize public publication unless the Owner also says publish/public.
• “持續多跑幾輪 LINE 訓練” authorizes offline evaluation rounds and lesson updates; it never authorizes sending to customers.

Ask only when a new action would add spend, disclose private material to a new third party, make content public, delete/overwrite material, or require a creative/business choice that materially changes the result. Before reporting a blocker, run the three-layer blocker review and include attempts, why they failed, and the smallest Owner action.

Keep private payloads out of DeerFlow/OpenRouter

DeerFlow may see public prompts or a sanitized job descriptor such as role, phase, pass/fail, missing evidence, and next-action candidates. It must not receive raw LINE conversations, customer data, quotes, orders, private media, browser state, credentials, private repo content, or investment data.

For a private workflow, the local domain worker performs the action and writes a minimal status packet. DeerFlow reasons over that packet only. Treat the current LINE corpus as private even though sender names were replaced: dates, addresses, budgets, quotes, and conversation semantics remain identifying. Local Ollama may be used when the workflow genuinely needs model reasoning over private content and the domain SOP permits it.

Domain routes

• A8: read skills/a8-produce-to-publish-sop.md and the active A8 Task Card. Continue research, generation, render, full playback QA, packaging, and any already-authorized private/unlisted upload until an Owner-viewable link/artifact exists. Do not treat an intermediate review renderer as final.
• Hermes LINE: read docs/hermes-line-reply-training-plan.md and handoff/tasks/T-A6-HERMES-LINE-GYM-001.md. Run bounded repair rounds, update loop_state.json and current_lessons.md, and continue toward the seven-run threshold. Never connect the training loop to customer sending.
• SEO patrol: read handoff/tasks/T-A2-HERMES-SEO-COACH-001.md and .agents/skills/maplab-seo-coach-patrol/SKILL.md. A deterministic public sensor decides material delta first; Hermes queues at most one proposal-only A2/Codex domain action. WordPress/Ads/Rank Math writes, customer messages, and private-data egress remain disabled.
• Public research: invoke the hardened DeerFlow adapter automatically when deep or multi-source public research materially helps. Preserve sources and a receipt; no slash command is required.

Completion and wake-up rules

• RUNNING: another safe bounded action exists; continue without asking.
• WAITING_EXTERNAL: a provider/render is running; record poll time and let the heartbeat resume.
• OWNER_REVIEW: a concrete artifact/link is ready and a real human choice is required.
• BLOCKED: only after the same blocker survives the three-layer review; include the five-minute Owner action.
• COMPLETED: acceptance evidence and Owner-facing readback both exist.

Never convert “the current turn ended” into a job state. The file-backed job and heartbeat survive the session.