[Bug]: nvidia model: minimaxai/minimax-m2.7, Model generated invalid tool call:

[HarmonClaw]

Bug Description

1. curl nvidia api call works
   google@openclaw:$ curl -X POST https://integrate.api.nvidia.com/v1/chat/completions -H "Authorization: Bearer nvapi-xxx" -H "Content-Type: application/json" -d '{"model": "minimaxai/minimax-m2.7", "messages": [{"role": "user", "content": "what model are you running?"}]}'
   {"id":"chatcmpl-a16c3f9a34714361","object":"chat.completion","created":1775982026,"model":"minimaxai/minimax-m2.7","choices":[{"index":0,"message":{"role":"assistant","content":"\n\nI’m a large‑language‑model AI assistant built by MiniMax. The specific version you’re interacting with is the MiniMax‑M2.7 model.","refusal":null,"annotations":null,"audio":null,"function_call":null,"tool_calls":[],"reasoning":null,"reasoning_content":null},"logprobs":null,"finish_reason":"stop","stop_reason":null,"token_ids":null}],"service_tier":null,"system_fingerprint":null,"usage":{"prompt_tokens":44,"total_tokens":930,"completion_tokens":886,"prompt_tokens_details":null},"prompt_logprobs":null,"prompt_token_ids":null,"kv_transfer_params":null}google-fulll-harmonrix@openclaw:$

2. hermes bot generating invalid tool calls with nvidia: minimaxai/minimax-m2.7
   [2026/4/12 10:36] agentHermes_bot: ⚠️ Model generated invalid tool call: read_fileread_file
   [2026/4/12 10:38] Chang: isn't it allowed by the config?
   [2026/4/12 10:38] agentHermes_bot: 📖 read_file: "~/.hermes/workspace/crypto-backtester..."
   [2026/4/12 10:38] agentHermes_bot: ⚠️ Model generated invalid tool call: read_fileread_fileread_fileread_file
   [2026/4/12 10:39] Chang: what model are you running?
   [2026/4/12 10:39] agentHermes_bot: ⚠️ Model generated invalid tool call: terminalterminal
   [2026/4/12 10:43] Chang: what model are you running?
   [2026/4/12 10:44] agentHermes_bot: ⚠️ Connection to provider dropped (ReadTimeout). Reconnecting… (attempt 2/3)
   [2026/4/12 10:44] agentHermes_bot: ⚠️ Model generated invalid tool call: search_filessearch_files
   [2026/4/12 10:56] Chang: what model are you running?
   [2026/4/12 10:57] agentHermes_bot: ⚠️ Model generated invalid tool call: search_filessearch_filessearch_filessearch_files

Steps to Reproduce

1. endpoint: https://integrate.api.nvidia.com/v1
   model: minimaxai/minimax-m2.7
2. [2026/4/12 10:36] agentHermes_bot: ⚠️ Model generated invalid tool call: read_fileread_file
   [2026/4/12 10:38] Chang: isn't it allowed by the config?
   [2026/4/12 10:38] agentHermes_bot: 📖 read_file: "~/.hermes/workspace/crypto-backtester..."
   [2026/4/12 10:38] agentHermes_bot: ⚠️ Model generated invalid tool call: read_fileread_fileread_fileread_file

Expected Behavior

[2026/4/12 11:38] Chang: what model are you running?
[2026/4/12 11:38] agentHermes_bot: ❓ clarify: "You've asked me what model I'm runnin..."
[2026/4/12 11:38] agentHermes_bot: I'm running MiniMax M2.5 (minimax/minimax-m2.5:free) via OpenRouter.
You've asked this question several times now. Is there something specific about the model you'd like to know, or are you testing something?
[2026/4/12 11:39] Chang: i've just changed the config.yaml, i thought you were running on qwen3-codex, just to confirm, nothing eles
[2026/4/12 11:39] agentHermes_bot: 💻 terminal: "cat ~/.hermes/config.yaml"

Actual Behavior

[2026/4/12 10:36] agentHermes_bot: ⚠️ Model generated invalid tool call: read_fileread_file
[2026/4/12 10:38] Chang: isn't it allowed by the config?
[2026/4/12 10:38] agentHermes_bot: 📖 read_file: "~/.hermes/workspace/crypto-backtester..."
[2026/4/12 10:38] agentHermes_bot: ⚠️ Model generated invalid tool call: read_fileread_fileread_fileread_file
[2026/4/12 10:39] Chang: what model are you running?
[2026/4/12 10:39] agentHermes_bot: ⚠️ Model generated invalid tool call: terminalterminal
[2026/4/12 10:43] Chang: what model are you running?
[2026/4/12 10:44] agentHermes_bot: ⚠️ Connection to provider dropped (ReadTimeout). Reconnecting… (attempt 2/3)
[2026/4/12 10:44] agentHermes_bot: ⚠️ Model generated invalid tool call: search_filessearch_files
[2026/4/12 10:56] Chang: what model are you running?
[2026/4/12 10:57] agentHermes_bot: ⚠️ Model generated invalid tool call: search_filessearch_filessearch_filessearch_files

Affected Component

Tools (terminal, file ops, web, code execution, etc.)

Messaging Platform (if gateway-related)

Telegram

Operating System

Ubuntu 22.04.5 LTS

Python Version

Python 3.10.12

Hermes Version

v0.8.0

Relevant Logs / Traceback

Root Cause Analysis (optional)

No response

Proposed Fix (optional)

No response

Are you willing to submit a PR for this?

• I'd like to fix this myself and submit a PR

[jackshouse26]

Root cause + fix for the terminalterminal / doubled tool name bug

Environment: Hermes v0.9.x, model minimaxai/minimax-m2.7 via https://integrate.api.nvidia.com/v1

The issue is in the streaming tool call accumulator in run_agent.py. Tool call names are built by appending each streaming delta:

entry["function"]["name"] += tc_delta.function.name

Most models send the full name in the first chunk only and an empty string in subsequent chunks, so this works fine. MiniMax M2.7 via NVIDIA NIM sends the full name in every chunk, causing it to concatenate: "terminal" + "terminal" = "terminalterminal".

Fix — replace the name append block (~line 5163) with a repeat-aware version:

if tc_delta.function.name:
    current_name = entry["function"]["name"]
    delta_name = tc_delta.function.name
    # Some models (e.g. MiniMax via NVIDIA NIM) resend the full name in
    # every chunk instead of incrementally. Detect and skip duplicates.
    if not current_name:
        entry["function"]["name"] = delta_name
    elif not current_name.endswith(delta_name):
        entry["function"]["name"] += delta_name

This handles both behaviors:

• Incremental deltas (normal models): appended as before since partial chunks won't match the suffix
• Full-name-per-chunk (MiniMax/NIM): only the first chunk sets the name, repeats are skipped

[bozoweed]

i think that you forget most important part like some ids in chunck ;) reade the OpenAI documentation nvidia is working like them for chunking function streamed

[LVT382009]

#9372 (comment)

Check this