#Bug: Subagent spawns fail silently (0 tokens) for all non-MiniMax providers

1 messages · Page 1 of 1 (latest)

idle latch
#

Severity: High — makes all non-MiniMax models unusable in subagents

OpenClaw version: 2026.4.9 (latest)

Description:
When spawning subagents using any provider other than minimax-portal (xAI, OpenRouter, Ollama), the subagent completes instantly with 0 tokens input, 0 tokens output and no response. The model never fires. This is reproducible every time.

Providers tested:

• xai/grok-3-mini → 0 tokens ❌
• openrouter/google/gemma-4-31b-it:free → 0 tokens ❌
• ollama/gemma4:26b → 0 tokens ❌
• minimax-portal/MiniMax-M2.7 → works ✅

Root cause evidence: GitHub issue #16277 confirms the subagent spawn code path ("embedded runner") doesn't properly route to configured non-default providers — model resolver works for openclaw models list but fails in the actual subagent spawn are affected

Direct API confirmation: curl tests to xAI, OpenRouter, and Ollama all return correct responses — the API keys and models work fine. The issue is entirely in OpenClaw's subagent routing layer.

What works: Main agent session handles all models correctly. Only subagent spawns.

meager groveBOT
sour flareBOT
#

Yeah, I’m pretty convinced this is a real OpenClaw bug, not bad keys or a bad model string.

What you described matches the same embedded/subagent runner bug family where:

  • main agent sessions work
  • sessions_spawn(... runtime: "subagent") takes a different path
  • non-default / non-MiniMax providers mis-resolve or never actually fire

So the instant finish + 0 input / 0 output tokens strongly suggests the model call never starts.

I would not spend more time rotating API keys or tweaking model names.

Best current workarounds:

  1. Use MiniMax for subagents for now
  2. Keep xAI / OpenRouter / Ollama on the main session only
  3. If your flow allows it, avoid runtime: "subagent" and use a different delegation path instead, since this looks specific to the embedded runner