Summary

When using opencode-go/deepseek-v4-flash as the primary model for isolated cron sessions (agentTurn in isolated session), the model call enters stream_progress but then stalls indefinitely — the last stream event arrives, then 128+ seconds pass with no progress. The stuck_session_recovery mechanism aborts the run after ~370s, and the error surfaces as LLM request failed..

The same API endpoint works fine when called directly via curl (200, ~2s response time) and works in normal interactive sessions. The issue is specific to isolated cron sessions using opencode-go streaming.

Evidence

All times in Asia/Shanghai:

Diagnostic Logs (gateway.log / structured log)

stalled session: ... state=processing age=130s queueDepth=1 
reason=active_work_without_progress classification=stalled_agent_run 
activeWorkKind=model_call lastProgress=model_call:stream_progress 
lastProgressAge=128s ...

stuck session recovery: ... age=370s action=abort_embedded_run 
aborted=true drained=true ...

Direct API Test (working)

curl -s https://opencode.ai/zen/go/v1/chat/completions \
  -H "Authorization: Bearer sk-..." \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-v4-flash","messages":[{"role":"user","content":"Say hello"}],"max_tokens":3}'
# → HTTP 200, 1.78s

Root Cause (Hypothesis)

The opencode-go provider plugin does not properly establish or maintain streaming connections (model_call:stream_progress) within isolated cron sessions. This is likely connected to the known issue where applyOpencodeGoProviderConfig is an empty function (referenced in PR #92913, which tries to fix context window registration for opencode-go models). When isolated sessions attempt to initialize streaming, the provider configuration (context window, max tokens, etc.) may not be correctly propagated, causing the stream to hang after initialization.

Workaround

Adding fallbacks: ["Minimax-M2.7"] (or any non-opencode-go model) to the cron job payload resolved the issue:

{
  "kind": "agentTurn",
  "message": "...",
  "fallbacks": ["Minimax-M2.7"],
  "timeoutSeconds": 600
}

After adding this fallback, the cron job completes in ~32s (vs 370s stall) on fallback to Minimax-M2.7.

Steps to Reproduce

1. Create an isolated cron job using agentTurn with model: opencode-go/deepseek-v4-flash (or via agent default primary model)
2. Let it run at the scheduled time
3. Observe: the job takes >360s and fails with LLM request failed.
4. Diagnostic shows: activeWorkKind=model_call, lastProgress=model_call:stream_progress, lastProgressAge > 120s

Expected Behavior

opencode-go streaming should work reliably in isolated cron sessions, just as it works in interactive sessions and direct API calls.

Environment

• OpenClaw version: 2026.6.6
• OS: macOS (arm64, Darwin 25.4.0)
• opencode-go API base: https://opencode.ai/zen/go/v1
• Model: deepseek-v4-flash