Headless mode hangs indefinitely on single tool call, ignores timeout (gemini-pro, o3-mini)

[ellisjr]

Description

When using certain models in headless mode (gemini-pro via OpenRouter, openai/o3-mini via direct API), the sidecar hangs indefinitely on the first tool call. The --timeout flag is ignored — sessions run for 12-20+ minutes without progressing past message 1 before the polling loop finally gives up with no output.

Steps to Reproduce

sidecar start --model gemini-pro --agent Plan --no-ui --timeout 20 \
  --prompt "Read and review _bmad-output/planning-artifacts/prd.md"

Observed Logs

[sidecar] 15s | 1 messages | Executing tool call...
[sidecar] 30s | 1 messages | Executing tool call...
...
[sidecar] 12m0s | 1 messages | Executing tool call...
...
[sidecar] 20m0s | 1 messages | Executing tool call...
{"level":"error","msg":"Polling loop exited","completed":false,"pollCount":599,"elapsed":1201030}
## Sidecar Results: No Output

Expected Behavior

Session detects a stalled tool call and either retries, times out gracefully, or produces whatever partial output is available.

Actual Behavior

Session hangs for the full duration stuck on a single tool call (file read), never progresses past message 1, then exits with zero output.

Affected Models

• gemini-pro via OpenRouter (resolves to google/gemini-3-pro-preview) — hung for 20 minutes
• openai/o3-mini via direct API — hung for 12+ minutes before manual kill

Notes

• The same task with --model gemini (Flash variant via OpenRouter) completed successfully in ~15 seconds with a full, high-quality review
• This suggests Pro/reasoning models have latency on tool calls that the polling/timeout logic doesn't account for
• The --timeout 20 flag had no effect on either model