Fix reasoning_effort error for GPT-5.1/5.2 in local mode

[shrey150]

Summary

• GPT-5.1 and GPT-5.2 reject reasoning_effort: "minimal" — valid values are none, low, medium, high, xhigh
• Sets reasoning_effort: "low" for GPT-5.1/5.2 models in the local mode LLM client (litellm path)
• Other GPT-5 variants (e.g. gpt-5-nano) are unaffected — they don't get reasoning_effort set

Context

Users on v0 SDK hitting this error when using GPT-5.2:

Unsupported value: 'reasoning_effort' does not support 'minimal' with this model.
Supported values are: 'none', 'low', 'medium', 'high', and 'xhigh'.

This matches the fix already shipped in the v3 TypeScript core (aisdk.ts) which differentiates GPT-5.1/5.2 from other GPT-5 models.

Note: For API mode (BROWSERBASE) users, the reasoning_effort is set server-side. Users hitting this in API mode should upgrade to stagehand>=3.6.0 which connects to a server with the fix.

Test plan

• Unit logic verified for GPT-5.2 ("low"), GPT-5.1 ("low"), GPT-5-nano (no override), GPT-4o (no override)
• E2E tested locally: extract + observe with openai/gpt-5.2 in LOCAL mode passes
• 57 unit tests pass (1 pre-existing failure unrelated)

🤖 Generated with Claude Code

---

Summary by cubic

Fixes OpenAI errors for gpt-5.1 and gpt-5.2 in local mode by setting reasoning_effort to "low" instead of the unsupported "minimal". Prevents failed requests when using these models via the local litellm client.

• Bug Fixes
  • Apply reasoning_effort: "low" only for gpt-5.1 and gpt-5.2 in local (litellm) mode.
  • Other GPT-5 variants (e.g., gpt-5-nano) are unchanged.
  • For API mode, this is handled server-side; upgrade to stagehand>=3.6.0 if you still see the error.

^{Written for commit 6f348f2. Summary will update on new commits. Review in cubic}

[cubic-dev-ai]

1 issue found across 1 file

Confidence score: 3/5

• There is a concrete behavior risk in stagehand/llm/client.py: reasoning_effort is always overridden for GPT-5.1/5.2, even when callers pass valid explicit values like "high" or "medium".
• Given the issue’s high severity/confidence (7/10, 9/10), this could cause user-visible regressions in model configuration rather than being a purely internal refactor concern.
• Pay close attention to stagehand/llm/client.py - ensure the GPT-5.1/5.2 fallback only handles the rejected "minimal" case and preserves valid user-provided reasoning_effort values.

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="stagehand/llm/client.py">

<violation number="1" location="stagehand/llm/client.py:120">
P1: This unconditionally overrides `reasoning_effort` for GPT-5.1/5.2, even when the user explicitly passes a valid value (e.g., `"high"`, `"medium"`). Since the bug is specifically about `"minimal"` being rejected, only that value should be replaced.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

P1: This unconditionally overrides reasoning_effort for GPT-5.1/5.2, even when the user explicitly passes a valid value (e.g., "high", "medium"). Since the bug is specifically about "minimal" being rejected, only that value should be replaced.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At stagehand/llm/client.py, line 120:

<comment>This unconditionally overrides `reasoning_effort` for GPT-5.1/5.2, even when the user explicitly passes a valid value (e.g., `"high"`, `"medium"`). Since the bug is specifically about `"minimal"` being rejected, only that value should be replaced.</comment>

<file context>
@@ -114,6 +114,11 @@ async def create_response(
+            # GPT-5.1 and GPT-5.2 don't support "minimal" reasoning_effort.
+            # Set "low" for these models to avoid OpenAI API errors.
+            if "gpt-5.1" in completion_model or "gpt-5.2" in completion_model:
+                filtered_params["reasoning_effort"] = "low"
+
         self.logger.debug(
</file context>

Suggested change

	filtered_params["reasoning_effort"] = "low"
	if filtered_params.get("reasoning_effort") == "minimal":
	filtered_params["reasoning_effort"] = "low"

[miguelg719]

@shrey150 can users pass specific reasoning effort?

[shrey150]

no, not possible on v3 nor in v2