openclaw - 💡(How to fix) Fix openai/gpt-5.4: Command lane cleared on every message in channel sessions (WS session interrupted mid-turn) [1 participants]

The model fallback chain fires (gpt-5.4 → gemini-3.1-pro-preview) but both fail with the same cleared-lane error, leaving the channel unable to respond. 4. Agent fails before reply; channel shows the ⚠️ Agent failed before reply error 2026-03-29T07:53:55.717Z warn model-fallback/decision candidate_failed requestedModel=gpt-5.4 errorPreview="Command lane 'session:agent:main:discord:channel:1485311563157213258' cleared" 2026-03-29T07:55:07.021Z error Embedded agent failed before reply: Command lane "session:agent:main:discord:channel:1485311563157213258" cleared The error occurs consistently across multiple message attempts. The channel recovers only after a context compaction event resets the session. {"error": {"message": "Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead."}}

Fix Plan

To resolve the issue, we need to update the model probe and onboard paths to use max_completion_tokens instead of max_tokens for models that require it.

Here are the steps:

• Extract the requiresFoundryMaxCompletionTokens function from shared-x9RbOMks.js and create a shared requiresMaxCompletionTokens(modelId) helper function.
• Update onboard-custom-DY0XuRoF.js and onboard-iE_s6O7s.js to use the new requiresMaxCompletionTokens helper function to determine whether to use max_tokens or max_completion_tokens in probe calls.
• Update any other Chat Completions request builders that hardcode max_tokens to use the requiresMaxCompletionTokens helper function.

Example code for the requiresMaxCompletionTokens helper function:

function requiresMaxCompletionTokens(modelId) {
  const normalized = normalizeModelName(modelId);
  if (!normalized) return false;
  return normalized.startsWith("gpt-5") || normalized.startsWith("o1") || 
         normalized.startsWith("o3") || normalized.startsWith("o4");
}

Example code for updating onboard-custom-DY0XuRoF.js:

// ...

const params = {
  // ...
  max_tokens: requiresMaxCompletionTokens(modelId) ? 'max_completion_tokens' : 'max_tokens',
  // ...
};

// ...

Verification

To verify that the fix worked, test the Discord channel session with the openai/gpt-5.4 model and ensure that the agent responds correctly without showing the ⚠️ Agent failed before reply error.

Extra Tips

• Make sure to update all relevant code paths that use max_tokens to use the requiresMaxCompletionTokens helper function.
• Test the fix with different models to ensure that it works correctly for all cases.
• Consider adding logging or monitoring to detect and report any future issues with model compatibility.