Onboard probe selects `openai-responses` for LiteLLM proxy, agent returns empty text on subsequent calls

[Kennyezzz]

Agent Diagnostic

• Loaded nemoclaw-reference skill (architecture.md, commands.md)
• Loaded nemoclaw-configure-inference skill (inference-options.md)
• Ran python3 -c "import json; print(json.load(open('/sandbox/.openclaw/openclaw.json'))['models']['providers']['inference']['api'])" inside sandbox → confirmed openai-responses
• Ran curl -s https://inference.local/v1/responses ... inside sandbox → endpoint returns valid text every time
• Ran openclaw agent --agent main -m "hello" --session-id test --verbose on multiple times → intermittent empty replies
• Parsed session transcript (jsonl) — tokens always consumed, text intermittently empty (~6/9 calls):

  17:15 text=''      output_tokens=108
  17:22 text='Hey...' output_tokens=597  ← works
  17:23 text=''      output_tokens=531
  17:24 text=''      output_tokens=199
  17:27 text=''      output_tokens=252
  17:29 text='Hey...' output_tokens=249  ← works
  17:29 text='Hey...' output_tokens=363  ← works
  17:30 text=''      output_tokens=253
  

• Traced probeOpenAiLikeEndpoint in bin/lib/onboard.js:966-1016 — probe tries /responses first, selects on HTTP 200 only, no response parsing validation
• Overriding ENV NEMOCLAW_INFERENCE_API=openai-completions in a --from Dockerfile resolves the issue — all calls return text via /v1/chat/completions
• Agent could not resolve this — the inference API type is baked into openclaw.json at build time (Landlock read-only), cannot be changed at runtime

Description

When onboarding with a LiteLLM proxy as "Other OpenAI-compatible endpoint", the onboard probe selects openai-responses because LiteLLM returns HTTP 200 on /responses. After sandbox creation, agent calls intermittently return empty text — tokens are consumed (stopReason: "stop", isError: false) but the session transcript records "text":"".

The inference endpoint itself works — curling /v1/responses directly inside the sandbox always returns valid text. The issue is in OpenClaw's response parsing when using openai-responses API with a LiteLLM proxy.

Expected: agent returns text on every call.
Actual: agent returns empty text on most calls (~6/9 in testing).

Reproduction Steps

1. nemoclaw onboard
2. Choose 3 (Other OpenAI-compatible endpoint)
3. Enter LiteLLM proxy URL (e.g. http://<host>:4000/v1) + API key + model (e.g. gemini-3.1-flash-lite-preview)
4. Onboard prints: Responses API available — OpenClaw will use openai-responses.
5. Build output shows: Step 24/40 : ARG NEMOCLAW_INFERENCE_API=openai-responses
6. nemoclaw <name> connect
7. openclaw agent --agent main -m "hello" --session-id test --verbose on
8. Observe: "completed" with no text reply (intermittent — may work on first call, fails on most subsequent calls)

Workaround: Use --from with a custom Dockerfile containing ENV NEMOCLAW_INFERENCE_API=openai-completions. No workaround without --from.

Environment

• OS: macOS 15 (Apple M3)
• Docker: Docker Desktop 4.x
• NemoClaw: v0.0.7
• OpenShell: 0.0.23
• OpenClaw: 2026.3.11
• LiteLLM proxy with gemini-3.1-flash-lite-preview

Logs

Session transcript showing empty text with consumed tokens:


{"role":"assistant","content":[{"type":"text","text":""}],"usage":{"input":12020,"output":540},"stopReason":"stop"}


Direct curl inside sandbox returns valid text:


{"output":[{"content":[{"type":"output_text","text":"Hello! How can I help you today?"}]}],"usage":{"input_tokens":1,"output_tokens":135}}


Related code:
- `bin/lib/onboard.js:966-1016` — `probeOpenAiLikeEndpoint` tries `/responses` first
- `bin/lib/onboard.js:1081` — `"Responses API available — OpenClaw will use openai-responses."`
- `bin/lib/onboard.js:926-928` — `patchStagedDockerfile` writes probe result into `ARG NEMOCLAW_INFERENCE_API`

Agent-First Checklist

• I pointed my agent at the repo and had it investigate this issue
• I loaded relevant skills (e.g., debug-openshell-cluster, debug-inference, openshell-cli)
• My agent could not resolve this — the diagnostic above explains why

[johntmyers]

This isn't a bug in OpenShell if you're able to run inference directly on the sandbox and it works. Your workaround includes using OpenShell's --from but the fix is setting an env var specific to NemoClaw. I'd suggest checking with NemoClaw or OpenClaw is this is a valid bug.