gate-attribution

Gate Attribution

The rule

The gate tracker attributes reviewer/test_engineer dispatches PER TASK. A single-task prompt still attributes by task_id / taskId / unambiguous prompt task ID. A set-dispatch can also count per-task when the reviewer/test_engineer output includes parseable per-task rows:

[REVIEWED] | task-2.1 | APPROVED | ...
[TESTED] | 2.1 | PASS | ...

[REVIEWED] verdicts are APPROVED | REJECTED | CONCERNS; [TESTED] verdicts are PASS | FAIL | SKIPPED. Rows with task-X.Y are normalized to X.Y; unsafe or non-plan IDs are ignored. Each parseable per-task verdict row creates gate evidence (regardless of verdict value); the gate's pass/fail decision is made elsewhere from the accumulated evidence. If no rows are parseable, attribution falls back to the single-task rule.

Protocol

1. For unrelated or high-risk tasks: Dispatch separate reviewer and/or test_engineer lanes with exactly ONE taskId.
2. For a true set-dispatch: Require one [REVIEWED] | task-id | verdict | ... (or [TESTED] | ...) row per task in the returned output. Each parseable row creates gate evidence; the gate decision is made from the accumulated evidence.
3. Minimize overhead via parallel dispatch when set-dispatch is not appropriate:

   dispatch_lanes_async with:
   - common_prompt: shared verification context
   - lanes: one lane per task, each with a single taskId
   - max_concurrent: up to 3
   

   This protocol applies only in sessions whose actual tool list includes the swarm controller's dispatch tools; on hosts without the controller it is not applicable — do not fabricate set-dispatch rows.
4. Collect + attribute: Single-task lanes auto-attribute to their taskId; set-dispatch rows auto-attribute per parsed row.
5. Do NOT rely on prose summaries: A batched dispatch without parseable rows is ambiguous and does not count per-task.

Gate evidence is persisted independently as .swarm/evidence/{taskId}.json for each task. Each parseable set-dispatch row causes the hook to write one task-scoped evidence file for that task (regardless of verdict value); a single multi-task evidence file cannot satisfy any task.