[BUG] Auto review fails with ChatGPT/Codex: Unsupported parameter: temperature

[pnpwin]

Environment

• DeepSeek Harness Desktop 0.2.2-rc.1, macOS.
• Package: @deepseek-ai/dsh-experimental-auto-review.
• Provider: openai-codex, authenticated through a ChatGPT subscription.

Steps to reproduce

1. Select a GPT model through the Codex provider.
2. Enable Auto review for the current session.
3. Ask the agent: “Run only pwd. Do not modify anything.”

Actual behavior

The review request fails before the tool executes:

Auto review of tool "skill" failed; its body was not executed:
auto-review: reviewer ended with error PI_AI_ERROR:
Codex error: Unsupported parameter: temperature

The structured error code is AUTO_REVIEW_FAILED.

Expected behavior

Auto review should evaluate the tool call and either allow it or request user approval, without failing because of an unsupported request parameter.

Root cause

The classifyRisk function unconditionally sends temperature: 0. The ChatGPT/Codex endpoint rejects this parameter, including zero.

Proposed minimal patch

Omit temperature for openai-codex, preserving the existing setting for other providers.

Based on commit fcc0ad462172522aaf5c93a7b2e01f6b9f0b0159:

diff --git a/packages/experimental/auto-review/src/index.ts b/packages/experimental/auto-review/src/index.ts
--- a/packages/experimental/auto-review/src/index.ts
+++ b/packages/experimental/auto-review/src/index.ts
@@ -684,7 +684,9 @@
     model: snapshot.model,
     system: REVIEW_POLICY,
     messages: reviewMessages(snapshot),
-    temperature: 0,
+    // The ChatGPT/Codex endpoint rejects temperature, including zero.
+    // Keep the existing sampling setting for other provider routes.
+    ...snapshot.provider === 'openai-codex' ? {} : { temperature: 0 },
     sessionId: agent.session.id,
     purpose: 'auto-review',
     signal,

This leaves the authorization policy, review history, strict response validation, and fail-closed behavior unchanged.