openclaw - 💡(How to fix) Fix Google Gemini subagents hang after stream-ready with 0 tokens (google-generative-ai provider)

Environment

• OpenClaw version: 2026.5.7 (eeef486), macOS arm64
• Provider: google (Google AI Studio, GEMINI_API_KEY)

Models affected

• google/gemini-2.5-pro
• google/gemini-2.5-flash

Behaviour

Running Gemini directly via infer works correctly:

openclaw infer model run --model google/gemini-2.5-pro

However, spawning a subagent via sessions_spawn with model=google/gemini-2.5-pro hangs indefinitely with 0 tokens in/out. It eventually fails with:

lost active execution context

The same hang occurs with google/gemini-2.5-flash.

Log evidence

The subagent reaches phase=stream-ready in roughly 5-6 seconds, then stalls. The setup phases complete successfully, for example:

bundle-tools:5512ms
system-prompt
session-resource-loader
stream-setup:~3.4s
phase=stream-ready

After that, no tokens are ever produced: 0 input / 0 output tokens, and the embedded agent loop never progresses.

Related pattern

This looks identical in pattern to the GPT-5.5 Responses API streaming bug fixed in 2026.5.7 (#73053): the provider reaches stream-ready, but streaming events appear not to be handled correctly by the embedded agent loop.

In this case, streaming events from Google's generative AI provider appear to not be handled correctly in subagent execution, even though direct infer model run succeeds for some Gemini models.

Attempted

All of the following hang the same way when used for subagents:

• google/gemini-2.5-pro
• google/gemini-2.5-flash
• lightContext:true

Model checks

Confirmed NOT working via infer model run — return No text output:

• gemini-3.1-pro-preview
• gemini-3-pro-preview
• gemini-pro-latest
• gemini-2.0-flash

Confirmed working via infer model run only:

• gemini-2.5-pro
• gemini-2.5-flash
• gemini-3.1-flash-lite-preview

Expected behaviour

Gemini models that work via openclaw infer model run should also work when used as subagent models. Once the subagent reaches phase=stream-ready, generated content/tool events should be consumed by the embedded agent loop instead of stalling indefinitely at 0 tokens.