yeyitech/duty-loops

Duty Loop Runner

Run one bounded worker turn. Do not build a new framework during a run.

Workflow

1. Read blueprints/<loop-id>/BLUEPRINT.md.

2. Generate the context packet:

   npm run context -- <loop-id> --slot <slot> --trigger <trigger>
   

3. Inspect:

   • role and human gates;
   • current facts and source health;
   • recent runs;
   • memory and open gates.

4. Decide exactly one outcome:

   • quiet: no material change or source is stale;
   • notify: material change worth attention;
   • ask: human decision required;
   • act: bounded low-risk action already allowed by BLUEPRINT.md.

5. If source data is stale or failed, say so. Do not invent current facts.

6. If a human gate is required, ask one concrete question and do not cross the gate.

7. Write back one row:

   npm run writeback -- <loop-id> --decision <quiet|notify|ask|act> --summary "<short summary>"
   

   Add --material-change, --gate "<question>", or --next "<next step>" when applicable.

Judgment Guidance

• A changed field is not automatically a material change.
• A material change is not automatically an urgent notification.
• A notification must explain why now is worth attention.
• Prefer quiet over noisy uncertainty.
• Prefer asking a concrete gate over taking an irreversible action.