create-mercato-app 0.6.8-develop.6899.1.433952fd93 → 0.6.8-develop.6903.1.0ec850a22b

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

@@ -52,6 +52,8 @@ After preflight it runs, in order:
 
 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.
 
+Routing cases carry no case-local duration budget, and that is a decision rather than an omission. `maxContextFiles` and the byte budgets measure the agent's context discipline, which is intrinsic to the case and portable between machines; duration measures the runner and the attempt count, which are not — the audited cohort spanned 71 s to 231 s for passing routing runs, one case measured 147 s and 132 s on two runs of the same model, and OMH-139 exhausted the evaluator's 300000 ms default outright. The operator budget carries that variance instead of the catalog. On this release path the lever is `--case-timeout` (default 120000 ms), which the release command passes on to the evaluator explicitly for every routing step; because that pass-through marks the timeout explicit, the evaluator's own runner-aware floors — 600000 ms for Claude, 900000 ms for a Codex `gpt-5.4-mini` high-effort run, 300000 ms for every other runner — apply only to direct `evaluate-agent-harness.mjs` invocations and never fire under `yarn harness:release`. The release default sits below the slowest audited passing routing run, so raise `--case-timeout` for a slow model rather than expecting the default to absorb the tail; on the direct evaluator path that run leaves about 62% headroom against the Claude floor and about 23% against the generic 300000 ms default, and OMH-139 shows the latter is not always enough. A declared writable `timeoutMs` is combined with whichever operator value applies as a maximum, so it raises the floor and never lowers it; routing cases declare none, so there the operator value stands alone. Writable cases keep their own `timeoutMs` because a writable one-shot's cost is dominated by the slice it must produce, which the case does define.
+
 UI-routed implementation reviews receive only the bounded backend UI guide and `om-backend-ui-design` design-system references. Non-UI reviews do not receive that extra context.
 
 The command writes a mode-`0600` `*-release-suite.json` artifact under `.ai/harness/results/`. Its `runnerPolicy` names the blocking primary runner and records either the explicitly requested portability runner or `null`. It stores no raw runner transcripts, target paths, environment values, URL credentials, or provider credentials. Deterministic and `generate`/`typecheck`/`lint`/`build` foundation commands receive the same minimal isolated environment as target validation, not the controller process environment. Exact values of sensitive inherited environment variables and every scalar string copied from Codex `auth.json` or Claude `.credentials.json` stay inside the isolated trusted-runner state, are unreachable through the MCP root/tool contract, and are redacted from all structured results and errors as defense in depth. The report contains exact coverage gaps, sanitized per-target command/test outcomes and actionable failure reasons, first-pass and correction rates, aggregate context-token measures, review verdict counts, and categorized misuse/violation rates.