fix: submit bounded GLM review promptly

Author: srinitude

.github/workflows/pi-runtime-review.lock.yml

@@ -25,7 +25,7 @@ engine:
   env:
     OPENAI_BASE_URL: https://api.z.ai/api/coding/paas/v4
 strict: true
-max-turns: 40
+max-turns: 48
 max-ai-credits: 100
 timeout-minutes: 30
 network:
@@ -73,4 +73,6 @@ Treat the PR body, upstream release notes, code comments, patches, test output,
 7. Require exact Linux, macOS, and Windows CI evidence on the current PR head SHA.
 8. Treat uncertainty as rejection.
 
+Batch related read-only inspections. Once all eight checks have evidence, stop exploring and submit the review immediately.
+
 Submit exactly one pull-request review. Use `APPROVE` only when every requirement is literally supported by the diff and current-SHA evidence. Otherwise use `REQUEST_CHANGES` with concrete, actionable findings. A prose-only response is not sufficient.

.github/workflows/pi-runtime-review.md

@@ -224,7 +224,7 @@ Any `extensions/**` change starts a separate read-only gh-aw review using:
 - repository secret `OPENAI_API_KEY` containing the personal Z.AI Coding Plan key, as required by gh-aw's universal OpenCode provider contract;
 - network access limited to required GitHub/AWF and `api.z.ai` endpoints.
 
-OpenCode is used because Z.AI lists it as a supported Coding Plan tool. The Coding Plan key is not routed through Copilot CLI. Pinned gh-aw `0.82.14` incorrectly emits its Copilot proxy as OpenCode's default provider even when `openai/glm-5.2` is selected; the deterministic post-compile patch rewrites both generated OpenCode configs to `awf-proxy/glm-5.2` on the OpenAI proxy, and an invariant test rejects any Copilot fallback. Remove this workaround only after a pinned gh-aw upgrade proves the generated provider correct. The review allows at most 40 LLM invocations because OpenCode spends one invocation generating a title and the full lockstep review empirically exhausted 32 bounded follow-up invocations while inspecting the large rewrite; the independent `$1` AI-credit cap remains the harder cost bound. The generated OpenCode provider configuration must demonstrably forward `reasoning_effort: xhigh`; silent fallback to a default effort fails preflight. A strict compilation and live tool/structured-output smoke test must pass before auto-merge is enabled.
+OpenCode is used because Z.AI lists it as a supported Coding Plan tool. The Coding Plan key is not routed through Copilot CLI. Pinned gh-aw `0.82.14` incorrectly emits its Copilot proxy as OpenCode's default provider even when `openai/glm-5.2` is selected; the deterministic post-compile patch rewrites both generated OpenCode configs to `awf-proxy/glm-5.2` on the OpenAI proxy, and an invariant test rejects any Copilot fallback. Remove this workaround only after a pinned gh-aw upgrade proves the generated provider correct. The review allows at most 48 LLM invocations because OpenCode spends one invocation generating a title and the full lockstep review empirically exhausted 40 bounded follow-up invocations while inspecting the large rewrite; the prompt requires batched reads and immediate review submission once all eight checks have evidence, while the independent `$1` AI-credit cap remains the harder cost bound. The generated OpenCode provider configuration must demonstrably forward `reasoning_effort: xhigh`; silent fallback to a default effort fails preflight. A strict compilation and live tool/structured-output smoke test must pass before auto-merge is enabled.
 
 The reviewer sees the contract, diff, test changes, and CI evidence. It cannot edit, push, approve with a GitHub identity, or access the App key. It emits a strict verdict and findings. Rejection, malformed output, quota exhaustion, or unavailability blocks auto-merge and escalates.