Summary

When the API returns a 400 error with message "invalid tool call arguments", Hermes cascades through ALL fallback providers before giving up. This wastes 20-60 seconds per occurrence because the error is a model-output problem (malformed JSON in tool call arguments), not an infrastructure problem that a different provider could solve.

Observed Behavior

From agent logs, 7 occurrences of this pattern:

glm-5.1:cloud → qwen3.5:397b-cloud → mistral-large-3:675b-cloud → glm-5:cloud → kimi-k2.5:cloud → ERROR

Each cascade = 4-5 sequential API calls, each adding ~5-15s of latency. The final model often generates its own malformed JSON (unterminated strings, invalid escapes in execute_code/terminal tool arguments), making the cascade end in an even worse state.

LLM Provider: All models in this chain are served via Ollama (acting as an OpenAI-compatible proxy on 127.0.0.1:11434). The proxy routes cloud-hosted models (glm-5.1:cloud, qwen3.5:397b-cloud, mistral-large-3:675b-cloud, glm-5:cloud, kimi-k2.5:cloud) through the Ollama API endpoint. The "invalid tool call arguments" 400 error is returned by Ollama's proxy layer when it rejects malformed tool call JSON before forwarding to the upstream cloud LLM.

Root Cause

1. agent/error_classifier.py lines ~623-628

_classify_400() handles all 400 errors. When none match specific patterns (context overflow, model not found, rate limit, billing), it falls through to:

return result_fn(
    FailoverReason.format_error,
    retryable=False,
    should_fallback=True,   # ← THIS IS THE PROBLEM
)

A 400 with "invalid tool call arguments" doesn't match any special case, so it gets should_fallback=True.

2. run_agent.py lines ~10395-10428

The is_client_error check does NOT exclude FailoverReason.format_error from triggering fallback:

classified.reason not in (
    FailoverReason.rate_limit,
    FailoverReason.billing,
    FailoverReason.overloaded,
    FailoverReason.context_overflow,
    ...
)

Since format_error is not excluded, it triggers _try_activate_fallback().

Contrast with correct handling

When the API call succeeds (200 OK) but the response contains malformed JSON, run_agent.py ~L10865-10952 handles this correctly with a retry-with-feedback loop (up to 3 retries), NOT cascading fallbacks.

Proposed Fix

In _classify_400(), detect "invalid tool call" / "tool call arguments" patterns and return a new reason with should_fallback=False:

if any(p in error_msg for p in ["invalid tool call", "tool_call", "function_call"]):
    return result_fn(
        FailoverReason.model_output_error,
        retryable=False,
        should_fallback=False,
    )

Then add FailoverReason.model_output_error to the is_client_error exclusion list.

Environment

• Hermes Agent v0.10.0, 170 commits behind latest
• LLM Provider: Ollama (openai-compatible proxy) on 127.0.0.1:11434
• Primary: glm-5.1:cloud via Ollama proxy
• Fallback chain: 4 cloud models through same Ollama proxy, 1 local model on 127.0.0.1:8081
• OS: Ubuntu 6.17.0-22-generic, kernel 6.17.0-22, i7-12700H (20 threads), RTX 3060 Laptop GPU 6GB (driver 590.48.01, CUDA 13.1), 30GB RAM
• Node: v24.14.0