| name | awh-gate |
| description | Out-of-band verification gate for the testing to done transition. Independently reruns the Task's section-1 cited tests plus its module suite via `awh eval` against a sealed, read-only baseline, and blocks the transition on any regression. Emits an awh-eval@1 artefact: per-task pass@1, the weighted aggregate, the sealed-baseline hash, and a GREEN or RED verdict. Used by chief-technology-officer/ship-tasks at step 28, after the post-implementation Task-audit (step 27) and before the done flip (step 30). GREEN is required to reach done; RED routes the task back to ready_to_implement per STATUS-REFERENCE section 1.3. Use when the user asks to "run the awh gate", "verify this task out of band", or "gate testing to done". Do NOT use for spec correctness (that is task-audit, during draft to ready_to_implement) or for in-context coverage (that is coverage-gate-author); this skill is the independent rerun those two are not, because an agent grading its own work is not a check. |
| license | Apache-2.0 |
| metadata | {"version":"1.0.0","module":"skill","stage":"e","cyberos-template":"awh-eval@1","cyberos-rubric-target":"awh_gate_rubric@1.0"} |
| allowed_memory_scopes | {"read":["project:*","module:*"],"write":["project:task/{task_id}/awh-gate"]} |
| audit | {"row_kind":"awh_gate_result","required_fields":["task_id","module","outcome","weighted_pass","harness_version","sealed_acceptance_hash"]} |
| inputs | [{"name":"task","format":"task@1","required":true},{"name":"module","format":"string","required":true},{"name":"goldenset","format":"path","required":true}] |
| outputs | [{"name":"report","format":"awh-eval@1"}] |
| triggers | ["run the awh gate","verify this task out of band","gate testing to done","independent rerun before done"] |
awh-gate
The out-of-band verification gate. CyberOS already has author and audit skill pairs, a coverage gate, and a task audit, but each of those runs in the authoring context, so the model is grading its own homework. This skill is the independent layer absorbed from auto-work-harness: it reruns the real tests outside that context and blocks the testing to done transition on regression.
When it runs
Step 28 of chief-technology-officer/ship-tasks, between the post-implementation task-audit (step 27) and the backlog-state-update-author done flip (step 30). The done flip is conditional on this skill returning GREEN.
What it does
- Resolve the task's module and its golden set at
modules/<module>/.awh/goldenset.yaml.
- Run
awh eval <goldenset> --base-dir . --seeds 1 --baseline modules/<module>/.awh/eval-baseline.json --max-regression 0.0. The golden set reruns the module's real build and test plus the held-out acceptance test, which is sealed read-only via awh lock so the agent cannot edit the bar it is graded against.
- Read the verdict. GREEN (no task regressed against the sealed baseline) is required to proceed to the done flip. RED routes the task back to
ready_to_implement per STATUS-REFERENCE section 1.3 with routed_back_count += 1.
- Emit one
awh_gate_result row into the memory audit chain carrying {task_id, module, outcome, weighted_pass, harness_version, sealed_acceptance_hash}. The row kind is gated on protocol change P23 section 6; until that lands, the verdict is written to a side log (.awh/gate-results.jsonl).
What it is not
This skill does not rewrite code or tests. It gates and measures. It does not replace the coverage gate or the spec audit; it is the independent rerun that confirms their result. See website/docs/architecture/verification-gate.html and tools/awh/ for the vendored tool, and tools/awh/RETIREMENT.md for how the standalone auto-work-harness is retired once every module is green under this gate.
Provenance
Vendored from auto-work-harness (source sha c1f2c77). Maturity ledger at .awh/evolution-log.jsonl (read with awh maturity report --log .awh/evolution-log.jsonl).
