Chat Still Broken After API Parameters Fix

Date: December 31, 2025
Status: Chat endpoint still returning errors
Context: Episodes, sessions, and voice are working ✅, but chat is broken ❌

Current Status

  • ✅ Episodes: Working
  • ✅ Sessions: Working
  • ✅ Voice: Working
  • ❌ Chat: Still broken

This indicates:

  • Authentication is working (episodes/sessions/voice use same auth)
  • Issue is specific to chat endpoint or LLM API call

Recent Fix Applied

We fixed the GPT-5.1-chat API parameters:

  • Changed max_tokens → max_completion_tokens
  • Removed temperature parameter for gpt-5.1-chat models
  • Code changes committed and pushed

Fix Location: backend/agents/base.py → FoundryChatClient.ainvoke()

Possible Issues

1. Deployment Not Complete

Symptom: Code fix is deployed but deployment hasn’t completed yet

Check:

gh run list --workflow=deploy.yml --limit 1

Solution: Wait for deployment to complete (~14 minutes)

2. Configuration Not Updated in Azure

Symptom: Code fix is correct, but Azure Container Apps environment variables are wrong

Required Configuration:

  • AZURE_AI_ENDPOINT = https://zimax-gw.azure-api.net/zimax (base URL, no /openai/v1/)
  • AZURE_AI_DEPLOYMENT = gpt-5.1-chat
  • AZURE_AI_API_VERSION = 2024-12-01-preview ✅ (required for gpt-5.1-chat)
  • AZURE_AI_MODEL_ROUTER = (empty or not set) ✅
  • AZURE_AI_KEY = (API key from Key Vault)

Check:

az containerapp show \
  --name staging-env-api \
  --resource-group zimax-ai \
  --query "properties.template.containers[0].env" \
  --output table

Solution: Update environment variables in Azure Container Apps if incorrect

3. Code Not Deployed

Symptom: Code changes committed but not yet deployed

Check:

  • Verify code is in main branch
  • Check if deployment workflow is running
  • Review deployment logs

4. Different Error Than Expected

Symptom: Fix addressed one issue, but there’s another problem

Check Backend Logs:

az containerapp logs show \
  --name staging-env-api \
  --resource-group zimax-ai \
  --tail 200 \
  --follow

Look for:

  • "Agent execution failed: ..."
  • "FoundryChatClient: Error calling LLM"
  • "FoundryChatClient: Error response body: ..."
  • "Full traceback: ..."

5. Model Router Still Enabled

Symptom: AZURE_AI_MODEL_ROUTER is still set in Azure, causing wrong endpoint format

Check:

  • Verify AZURE_AI_MODEL_ROUTER is empty/not set in Azure Container Apps
  • Code checks: if self.settings.azure_ai_model_router and self.settings.azure_ai_model_router.strip()

Solution: Delete or clear AZURE_AI_MODEL_ROUTER environment variable

Diagnostic Steps

Step 1: Test Chat Endpoint 

# Get JWT token from browser
python3 scripts/test-chat-debug.py \
  --token "YOUR_JWT_TOKEN" \
  --message "hi"

Expected: HTTP 200 with actual response (not error message)
If error: Check response body for details

Step 2: Check Deployment Status 

gh run list --workflow=deploy.yml --limit 1 --json conclusion,status,createdAt

Expected: conclusion: "success", status: "completed"
If still running: Wait for completion

Step 3: Verify Azure Configuration 

az containerapp show \
  --name staging-env-api \
  --resource-group zimax-ai \
  --query "properties.template.containers[0].env[?name=='AZURE_AI_API_VERSION' || name=='AZURE_AI_DEPLOYMENT' || name=='AZURE_AI_ENDPOINT' || name=='AZURE_AI_MODEL_ROUTER']" \
  --output table

Expected:

  • AZURE_AI_API_VERSION = 2024-12-01-preview
  • AZURE_AI_DEPLOYMENT = gpt-5.1-chat
  • AZURE_AI_ENDPOINT = https://zimax-gw.azure-api.net/zimax
  • AZURE_AI_MODEL_ROUTER = (not present or empty)

Step 4: Check Backend Logs 

az containerapp logs show \
  --name staging-env-api \
  --resource-group zimax-ai \
  --tail 100 \
  --follow

Look for chat-related errors when testing

Quick Fixes

Fix 1: Wait for Deployment

If deployment is still running, wait for it to complete.

Fix 2: Verify Code is Deployed

Check that the latest commit with the fix is in the deployed version:

# Check latest commit
git log --oneline -5

# Verify the fix is in the code
grep -A 10 "max_completion_tokens" backend/agents/base.py

Fix 3: Update Azure Configuration

If configuration is wrong, update it:

az containerapp update \
  --name staging-env-api \
  --resource-group zimax-ai \
  --set-env-vars \
    AZURE_AI_API_VERSION=2024-12-01-preview \
    AZURE_AI_DEPLOYMENT=gpt-5.1-chat \
    AZURE_AI_ENDPOINT=https://zimax-gw.azure-api.net/zimax

# Remove Model Router if set
az containerapp update \
  --name staging-env-api \
  --resource-group zimax-ai \
  --remove-env-vars AZURE_AI_MODEL_ROUTER

Expected Behavior After Fix

Once the fix is deployed and configuration is correct:

  1. Chat request succeeds:
    • HTTP 200 OK
    • Response contains actual agent response (not error message)
    • tokens_used > 0 (indicates LLM call succeeded)
  2. Backend logs show: 
    INFO: Calling agent_chat for user ...
INFO: FoundryChatClient: Calling https://zimax-gw.azure-api.net/zimax/openai/deployments/gpt-5.1-chat/chat/completions?api-version=2024-12-01-preview
INFO: FoundryChatClient: Response status=200
INFO: Agent chat succeeded: agent=elena, response_length=50
  3. Request payload includes:
    • max_completion_tokens (not max_tokens)
    • No temperature parameter
  • docs/troubleshooting/gpt-5.1-chat-api-parameters-fix.md - Original fix documentation
  • docs/troubleshooting/chat-error-diagnosis.md - General chat troubleshooting
  • docs/configuration/config-alignment.md - Configuration reference
  • docs/troubleshooting/bypass-model-router.md - Disable Model Router