Summary

Agent Zero v1.x silently routes utility_model calls to api.openai.com when
utility_model.api_base is an empty string (""). For users running NVIDIA NIM models (key prefix
nvapi-), this causes a repeated AuthenticationError flood on every background memory call.

Agent Zero correctly diagnosed the problem and attempted autonomous self-repair — adding try/except AuthenticationError guards to _50_recall_memories.py and _51_memorize_solutions.py. The logic
was correct, but the patch was written with 1-space Python indentation instead of 8-space,
causing an IndentationError on the next extension load. This blocked all agent responses
entirely.

Root Causes

1. utility_model.api_base: "" in _model_config/config.json → LiteLLM silent fallback to
   api.openai.com
2. Extension loader raises IndentationError/SyntaxError globally — one broken file blocks ALL
   extensions
3. Self-repair has no py_compile validation step before activating a patched file

Impact

• NVIDIA NIM users with provider: openai + NVIDIA key + empty utility_model.api_base will see
  repeated AuthenticationError in logs after every response
• If Agent Zero attempts autonomous self-repair, the indentation bug in the output causes complete
  agent failure
• Memory recall and memorization silently fail or crash background threads throughout

Proposed Fixes

Full white paper, root cause analysis, corrected extension files, and A0's self-generated diagnostic
documents:
https://github.com/MrTrenchTrucker/agent-zero-nvidia-routing-fix

Short version:

1. Config validator at startup — warn/error if utility_model.api_base is empty or key prefix
   mismatches endpoint
2. Graceful extension loader — catch IndentationError/SyntaxError per-extension, skip broken
   file, continue loading others
3. Self-repair safety — run py_compile on patched files before activating, restore original on
   failure
4. Core AuthenticationError guard — the nvapi- detection logic should live in core, not
   patched per-file

Environment

• Agent Zero v1.x (agent0ai/agent-zero:latest)
• LiteLLM with NVIDIA NIM (provider: openai, api_base: https://integrate.api.nvidia.com/v1)
• utility_model.api_base unset / empty string in _model_config/config.json

Verified Fix

Full recovery performed and confirmed. Post-recovery functional test: memory save/recall, sub-agent
delegation, and Docker socket access all passed. Corrected .py files with proper indentation +
AuthenticationError guards are in the repo above.

A fork with the code changes ready for review is at:
https://github.com/MrTrenchTrucker/agent-zero