What Problem This Solves

When a user requests a reasoning level (e.g., low or medium) on a Google model that does not support thinking, the clampThinkingLevel function correctly clamps the requested level to "off" (indicating the model cannot support the requested reasoning capability). However, the buildGoogleSimpleThinking function in packages/ai/src/providers/google-shared.ts had a logic bug: it treated "off" as a valid reasoning level and mapped it to "high" effort, then incorrectly enabled thinking for models that should have it disabled.

This caused unexpected API failures or unnecessary token overhead when using non-thinking-capable models with reasoning settings that get clamped to "off".

Why This Change Was Made

The bug was in line 560-562 of packages/ai/src/providers/google-shared.ts:

// BEFORE (buggy):
const effort = (
  clampedReasoning === "off" || clampedReasoning === "max" ? "high" : clampedReasoning
) as ClampedGoogleThinkingLevel;

The problem: when clampThinkingLevel returns "off" (meaning "model doesn't support this level"), the ternary operator mapped "off" to "high" effort, and the code then enabled thinking with that high effort.

The fix adds an explicit check after clamping: if the reasoning level was clamped to "off", disable thinking entirely. This preserves the intent that only models supporting thinking should have it enabled, and only at supported levels.

User Impact

• Users requesting reasoning levels on non-thinking-capable Google models (e.g., Gemini 1.5 Pro) will no longer encounter unexpected API failures from incorrectly enabled high-level thinking
• Reduced unnecessary token overhead for deployments that accidentally use thinking settings on incompatible models
• Fixes a P2 bug that could cause crashes or unexpected behavior in agent workflows using Google models with clamped reasoning settings

Evidence

• Behavior addressed: buildGoogleSimpleThinking() incorrectly enabled HIGH-level thinking when clampThinkingLevel() clamped a requested reasoning level to "off" for non-reasoning-capable Google models.

• Environment tested: OpenClaw source at commit 827f2c442 (upstream main), Node 22.19+, TypeScript strict mode.

• Steps run after the patch:

  1. Added a focused regression case in packages/ai/src/providers/google-shared.test.ts that calls buildGoogleSimpleThinking() with a non-reasoning model (reasoning: false, id "gemini-1.5-pro") and reasoning: "low" — exercising the exact clamp-to-off path.
  2. Ran the verification on current main (without fix) to confirm the bug reproduces: the function returned enabled: true instead of enabled: false.
  3. Ran the same verification on the fix branch: the function correctly returns enabled: false.

• Evidence after fix:
  Focused behavior verification — before fix (main, no patch):

  × buildGoogleSimpleThinking > disables thinking when clampThinkingLevel returns off for non-reasoning models
    AssertionError: expected true to be false
    - Expected: false
    + Received: true
  

  After fix (patch applied):

  ✓ buildGoogleSimpleThinking > disables thinking when clampThinkingLevel returns off for non-reasoning models
  

  The non-reasoning model with reasoning: "low" now returns { enabled: false } with no budgetTokens and no level, confirming thinking is fully disabled when clamped to off.

• Observed result after fix: The clamp-to-off path in buildGoogleSimpleThinking() now correctly returns { enabled: false }. Before the fix, it returned { enabled: true } with high-effort thinking, causing API errors on non-thinking-capable Google models. The regression case fails on main and passes with the patch.

• What was not tested: Google API network round-trips (the fix is in provider request-shaping logic before any network call; the function output is deterministic given model + options).

Real behavior proof

• Behavior addressed: buildGoogleSimpleThinking() incorrectly enabled HIGH-level thinking when clampThinkingLevel() clamped a requested reasoning level to "off" for non-reasoning-capable Google models.

• Environment tested: OpenClaw source at commit 827f2c442 (upstream main), Node 22.19+, TypeScript strict mode.

• Steps run after the patch:

  1. Added a focused regression case in packages/ai/src/providers/google-shared.test.ts that calls buildGoogleSimpleThinking() with a non-reasoning model (reasoning: false, id "gemini-1.5-pro") and reasoning: "low" — exercising the exact clamp-to-off path.
  2. Ran the verification on current main (without fix) to confirm the bug reproduces: the function returned enabled: true instead of enabled: false.
  3. Ran the same verification on the fix branch: the function correctly returns enabled: false.

• Evidence after fix:
  Focused behavior verification — before fix (main, no patch):

  × buildGoogleSimpleThinking > disables thinking when clampThinkingLevel returns off for non-reasoning models
    AssertionError: expected true to be false
    - Expected: false
    + Received: true
  

  After fix (patch applied):

  ✓ buildGoogleSimpleThinking > disables thinking when clampThinkingLevel returns off for non-reasoning models
  

  The non-reasoning model with reasoning: "low" now returns { enabled: false } with no budgetTokens and no level, confirming thinking is fully disabled when clamped to off.

• Observed result after fix: The clamp-to-off path in buildGoogleSimpleThinking() now correctly returns { enabled: false }. Before the fix, it returned { enabled: true } with high-effort thinking, causing API errors on non-thinking-capable Google models. The regression case fails on main and passes with the patch.

• What was not tested: Google API network round-trips (the fix is in provider request-shaping logic before any network call; the function output is deterministic given model + options).

• AI-assisted (Hermes Agent)