Investigation Summary

• The model is a reasoning model: vLLM is launched with --reasoning-parser nemotron_v3, and a direct API call returns ~130 of 149 completion tokens in the reasoning field before any content (with max_tokens: 20, content is null and finish_reason: "length").
• OpenClaw is configured as if it is not: openclaw.json shows "reasoning": false on the model entry, and the container env shows NEMOCLAW_REASONING=false.
• The config generator supports the flag — scripts/generate-openclaw-config.mts contains const reasoning = (env.NEMOCLAW_REASONING || "false") === "true"; and Dockerfile has ARG NEMOCLAW_REASONING=false — but passing NEMOCLAW_REASONING=true to the installer, to rebuild --yes, and to the installer with NEMOCLAW_RECREATE_SANDBOX=1 leaves the baked value unchanged in all three cases.
• Resulting failure: agent runs abort with LLM idle timeout (120s): no response from model while logs show lastProgress=model_call:stream_progress lastProgressAge=117s — the stream is open and emitting reasoning tokens, but they are not credited as progress.
• Unknown: whether the intended path to set this on an existing sandbox is a different command/flag, and whether the idle timer is designed to ignore reasoning-token progress or this is unintended.

Description

On a local reasoning model (Nemotron 3 Super via vLLM with --reasoning-parser nemotron_v3), agent runs consistently fail with LLM idle timeout (120s): no response from model, even for simple single-tool requests such as "check my email".

The model is working correctly. A direct call to the vLLM endpoint returns a valid response quickly, but the majority of generated tokens are emitted in the reasoning field before any content appears. During an agent run the stream is open and progressing (lastProgress=model_call:stream_progress), yet the idle detector fires at 120s because no content has arrived.

The generated openclaw.json declares "reasoning": false for this model. scripts/generate-openclaw-config.mts reads NEMOCLAW_REASONING to set that value, and the Dockerfile exposes ARG NEMOCLAW_REASONING=false, but there appears to be no user-facing way to set it to true on an existing sandbox — the env var is not picked up by the installer's recovery path, by rebuild, or by recreate.

Expected: either a supported way to declare a local model as a reasoning model on an existing sandbox, or reasoning-token stream progress should reset the idle timer so runs are not aborted while the model is actively generating.

Reproduction Steps

1. Onboard a sandbox on DGX Spark using a local vLLM endpoint serving a reasoning model (Nemotron 3 Super, launched with --reasoning-parser nemotron_v3).
2. Confirm the model emits reasoning before content:
   curl -s http://localhost:8000/v1/chat/completions -H "Content-Type: application/json"
   -d '{"model":"nemotron-3-super","messages":[{"role":"user","content":"Say hello"}],"max_tokens":20}'
   → "content": null, finish_reason: "length", all tokens in reasoning.
3. Observe the generated config declares the model as non-reasoning:
   ssh openshell- "cat /sandbox/.openclaw/openclaw.json | python3 -m json.tool | grep -A2 -B2 reasoning"
   → "reasoning": false
4. Attempt to set the flag, in three ways:
   a. Installer: NEMOCLAW_REASONING=true NEMOCLAW_INSTALL_REF=v0.0.92 ... ./nemoclaw.sh
   b. Rebuild: NEMOCLAW_REASONING=true nemoclaw rebuild --yes
   c. Recreate: NEMOCLAW_REASONING=true NEMOCLAW_RECREATE_SANDBOX=1 ... ./nemoclaw.sh
5. After each attempt, verify it did not take:
   docker inspect $(docker ps -q --filter name=openshell-) --format '{{json .Config.Env}}' | python3 -m json.tool | grep -i reasoning
   → NEMOCLAW_REASONING=false
6. Send the agent any request requiring a tool call. The run aborts at 120s with LLM idle timeout.

Environment

• Platform: NVIDIA DGX Spark (GB10, aarch64), 128GB unified memory
• Host OS: Ubuntu 24.04 / DGX OS 7.5.0, kernel 6.17.0-1021-nvidia
• GPU driver: 580.159.03, CUDA 13.0
• NemoClaw: v0.0.92
• OpenShell: 0.0.85 (docker driver)
• OpenClaw: 2026.7.1
• Node.js: v22.22.2
• Inference: vLLM v0.18.1-cu130 (Docker), serving nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4
  launched with --reasoning-parser nemotron_v3 --enable-auto-tool-choice --tool-call-parser qwen3_coder
• Relevant config: agents.defaults.timeoutSeconds: 600, thinkingDefault: "off",
  container env NEMOCLAW_MAX_TOKENS=4096, NEMOCLAW_TOOL_DISCLOSURE=progressive
• Observed generation rate: ~12 tokens/sec

Debug Output

Full nemoclaw debug --quick output available; the relevant excerpts are in the Logs section above. Happy to attach the complete tarball if useful.

Logs

### 1. Gateway declares the model as non-reasoning

2026-07-24T05:37:46.869+00:00 [gateway] agent model: inference/nemotron-3-super (thinking=off, fast=off)


### 2. But vLLM is serving it WITH a reasoning parser

~/vllm/start-super.sh:
    --reasoning-parser nemotron_v3
    --enable-auto-tool-choice
    --tool-call-parser qwen3_coder


### 3. And the model demonstrably emits reasoning before content

$ curl -s http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" \
    -d '{"model":"nemotron-3-super","messages":[{"role":"user","content":"Say hello"}],"max_tokens":20}'

"content": null,
"finish_reason": "length",
"reasoning": "We are to say hello.\n The instruction is simple: \"Say hello\"\n We'll output the string"

Same request with max_tokens=500 completes normally in 12.4s:
  149 completion tokens total, ~130 emitted as `reasoning`,
  content = "Hello! 👋 How can I assist you today?"


### 4. Config and container env both show reasoning disabled

$ ssh openshell-<sandbox> "cat /sandbox/.openclaw/openclaw.json | python3 -m json.tool | grep -A2 -B2 reasoning"
    "id": "nemotron-3-super",
    "name": "inference/nemotron-3-super",
    "reasoning": false,

$ docker inspect $(docker ps -q --filter name=openshell-<sandbox>) --format '{{json .Config.Env}}' | grep -i reasoning
    "NEMOCLAW_REASONING=false"

(unchanged after passing NEMOCLAW_REASONING=true to the installer,
 to `rebuild --yes`, and to the installer with NEMOCLAW_RECREATE_SANDBOX=1)


### 5. Result: agent runs abort while the model is still streaming

2026-07-23T23:50:43.177 [provider-transport-fetch] [model-fetch] start provider=inference
  api=openai-completions model=nemotron-3-super method=POST
  url=https://inference.local/v1/chat/completions timeoutMs=undefined
2026-07-23T23:50:43.224 [provider-transport-fetch] [model-fetch] response
  status=200 elapsedMs=46 contentType=text/event-stream; charset=utf-8

2026-07-23T23:53:25.630 [diagnostic] long-running session: state=processing age=145s
  activeWorkKind=model_call lastProgress=model_call:stream_progress
  lastProgressAge=117s recovery=none
2026-07-23T23:53:30.536 [agent/embedded] embedded run failover decision: stage=assistant
  decision=surface_error reason=timeout from=inference/nemotron-3-super
  rawError=LLM idle timeout (120s): no response from model

The stream opens in 46ms and is actively producing reasoning tokens,
but the idle detector fires at 120s because no `content` has arrived.


### Related (possibly separate issue): Telegram webhook management fails at startup

2026-07-24T05:37:49.635 [telegram] deleteWebhook failed: Network request for 'deleteWebhook' failed!
2026-07-24T05:37:49.639 [telegram] deleteMyCommands failed: Network request for 'deleteMyCommands' failed!
2026-07-24T05:37:49.653 [telegram] setMyCommands failed: Network request for 'setMyCommands' failed!
[telegram] [default] bridge did not start within 15s; check channels.telegram.enabled, plugin entries, and gateway log

Checklist

• I confirmed this bug is reproducible
• I searched existing issues and this is not a duplicate