Campaign

Use $campaign for the durable autonomous research loop.

Controller Contract

1. Read harness_core/campaign.yaml and the accepted docs/context/campaign.md Campaign Contract.
2. If the Contract is missing or draft, route to Grill and enter waiting_human; only explicit Human Approval accepts it.
3. Let the Agent propose the next Operation, then let the Controller validate policy, budget, prerequisites, JobHandle, and mandatory rules.
4. Reuse grill, prepare, build, run, analyze, write, and change; do not duplicate their behavior inside this wrapper.
5. Queue mutating steering while a training job is active. Read-only status and code questions remain available immediately.
6. Give each worker a bounded Operation Packet; persist cross-session facts through typed artifacts rather than conversation history.
7. Record Light Conclusion Evidence per Iteration and create full Evidence Chains only at review, Claim Delta, release, or audit.
8. After durable changes under docs/, run python tooling/harness.py docs sync --json; this maintains the live project link and the project name selects the external subfolder.

Start with python tooling/harness.py campaign start --json; the default is the current interactive Agent session. Set runtime.campaign_execution_mode: worker_process only for unattended background execution. Inspect with python tooling/harness.py campaign status --json. Use --standalone on a focused Operation only when isolation is intended.