create-mercato-app 0.6.7-develop.6726.1.983ae8a07e → 0.6.7-develop.6749.1.6b54c56dfe

package/agentic/shared/ai/harness/README.md

@@ -16,7 +16,7 @@ yarn harness:release --runner codex --prepare-targets /absolute/empty-release-ta
 yarn harness:release --runner codex --portability-runner claude --prepare-targets /absolute/empty-release-targets --acknowledge-writes
 ```
 
-The primary runner owns all 202 routing cases, all 46 writable cases, and all generated-code reviews. No per-case fallback or mixed primary ownership is allowed. Omitting `--portability-runner` is valid and the sanitized report records `portabilityRunner: null`; explicitly requesting an unavailable or failing secondary runner fails that extended run.
+The primary runner owns all 202 routing cases, all 46 writable cases, and all generative-judge runs. No per-case fallback or mixed primary ownership is allowed. Omitting `--portability-runner` is valid and the sanitized report records `portabilityRunner: null`; explicitly requesting an unavailable or failing secondary runner fails that extended run.
 
 Writable evaluation is intentionally opt-in. The expanded catalog has a 46-case writable release target, but only cases registered in `release-matrix.json` and backed by controller-owned fixtures and oracles are executable. Copy or create a fresh standalone app for one registered case, then seed only that case and mark the target disposable:
 
@@ -31,13 +31,13 @@ The preparer refuses the controller app, non-standalone targets, reused targets,
 
 The four generated-test cases use direct controller-resolved Jest/Playwright CLIs, never target package scripts. The target and dependencies stay read-only; Jest has no network, Playwright has loopback only, and the browser lane exposes only the exact installed headless-shell runtime. JSON reporters must attest at least one passed test and zero skipped, todo, focused, flaky, or expected-failure tests before review evidence can be created.
 
-Generated-code review is a separate post-oracle lane, never a nested second-model call. The individual command is opt-in while developing one case; the once-per-release suite requires it for every writable case:
+The generative judge is a separate post-oracle lane, never a nested second-model call. It applies the reusable `om-judge-agent-session` workflow, the pinned `om-code-review` skill, and UI design-system references when applicable. The individual command is opt-in while developing one case; the once-per-release suite requires it for every writable case:
 
 ```text
-yarn harness:validate --runner codex --review-writable-result /absolute/controller/.ai/harness/results/<writable-result>.json --writable-root /absolute/disposable/app
+yarn harness:validate --runner codex --judge-writable-result /absolute/controller/.ai/harness/results/<writable-result>.json --writable-root /absolute/disposable/app
 ```
 
-The source must be a passing one-shot implementation result from the current harness and its final whole-target fingerprint must still match. The controller copies changed regular text files as line-numbered inert snapshots, plus its pinned installed `om-code-review` skill and bounded trusted evidence, into a temporary read-only bundle. Target package scripts, dependencies, Git data, tracker state, original source files, and the target's absolute path are not copied or supplied to the reviewer; trace-verified out-of-bundle, environment, or process inspection and any bundle or target mutation fail closed. A separate sanitized artifact records source-result, target, policy, and installed-skill hashes plus the strict verdict, findings, and report. This supplemental review uses prior controller oracle evidence and does not claim the full repository validation gate or CI passed.
+The source must be a passing one-shot implementation result from the current harness and its final whole-target fingerprint must still match. The controller copies changed regular text files as line-numbered inert snapshots, the local judge skill, its pinned installed `om-code-review` skill, and bounded trusted evidence into a temporary read-only bundle. Target package scripts, dependencies, Git data, tracker state, original source files, and the target's absolute path are not copied or supplied to the judge; trace-verified out-of-bundle, environment, or process inspection and any bundle or target mutation fail closed. A separate sanitized artifact records source-result, target, policy, and skill hashes plus fixed evidence, artifact findings, design-system assessment, the smallest harness-owner findings, strict verdicts, and both reports. The legacy `--review-writable-result` and `--review-validation-result` flags remain read-only compatibility aliases. This supplemental judge uses prior controller evidence and does not claim validation or CI that did not run.
 
 ## Live-run security boundary
 

package/agentic/shared/ai/harness/RELEASE.md

@@ -34,7 +34,7 @@ For externally prepared apps, `--writable-targets /absolute/release-targets.json
 }
 ```
 
-The current catalog contains 202 cases, including 46 writable implementation/regression cases (22.8%). The command still derives all counts and case IDs from `cases.json`, `validators.json`, and `release-matrix.json`; those figures are documented release facts, not runner constants. The matrix keeps both supported runner model selectors, an exact all-case primary profile, an exact 46-case portability profile, and runner-neutral writable assignments. Run `yarn install-skills` first so the pinned external `om-code-review` skill and ownership evidence are present. Before running a model or writing a fixture, the release command requires complete deterministic, primary live-routing, writable, trusted-oracle, target, generated-test, and generated-code-review coverage. Every one of the 46 writable cases must have an `om-code-review` assignment. Missing business fixtures or release-matrix entries fail preflight and are listed by exact case ID in the report.
+The current catalog contains 202 cases, including 46 writable implementation/regression cases (22.8%). The command still derives all counts and case IDs from `cases.json`, `validators.json`, and `release-matrix.json`; those figures are documented release facts, not runner constants. The matrix keeps both supported runner model selectors, an exact all-case primary profile, an exact 46-case portability profile, and runner-neutral writable assignments. Run `yarn install-skills` first so the pinned external `om-code-review` skill and ownership evidence are present; the reusable local `om-judge-agent-session` skill ships with the scaffold. Before running a model or writing a fixture, the release command requires complete deterministic, primary live-routing, writable, trusted-oracle, target, generated-test, and generative-judge coverage. Every one of the 46 writable cases must have a judge assignment composing both skills. Missing business fixtures or release-matrix entries fail preflight and are listed by exact case ID in the report.
 
 ## PR #4529 remediation evidence
 
@@ -48,7 +48,7 @@ After preflight it runs, in order:
 4. fixture preparation and the selected primary runner for every writable case, including the controller-owned AST/behavior oracles and target typecheck;
 5. `yarn generate`, `yarn typecheck`, `yarn lint`, and `yarn build` in every writable target, after its trusted oracles;
 6. real generated-code execution for OMH-163 and OMH-192 through fixed Jest, OMH-164 through API-only Playwright, and OMH-165 through real-browser Playwright; and
-7. explicit isolated `om-code-review` for every writable result, bound to its passing command attestation, any required generated-test result and artifact hash, and the final target fingerprint.
+7. explicit isolated `om-judge-agent-session` for every writable result, composing `om-code-review` and applicable design-system guidance, bound to its passing command attestation, any required generated-test result and artifact hash, and the final target fingerprint.
 
 Each writable target is single-use because fixture preparation marks it disposable. Externally supplied target realpaths must be pairwise disjoint and neither equal to, contain, nor be contained by the controller. A failed deterministic or foundation-validation step prevents model execution. Once fixture preparation succeeds, all four target commands run even when the writable gate itself fails, so every generated target has exact diagnostics. A writable case may declare `timeoutMs` only to raise the release `--case-timeout` floor (never lower it); OMH-185 and its business-language parity case OMH-193 use 600000 ms because the complete module slice exceeded the generic five-minute evaluator default while actively producing source. Generated tests run only after the trusted writable oracle and all four target commands pass; review then requires all applicable gates. A target command or generated-test failure is recorded with its sanitized diagnostic and review is skipped. Other matrix entries continue so the report remains useful.