Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

This has happened for at least a week now, now sure what version. But basically I see all my cron jobs fail, whereas if I run the task manually it works fine.

The opencode-go provider plugin intermittently fails to forward stream completion events from the provider API to the gateway runtime. The provider-side HTTP stream delivers tokens successfully, but the final [DONE] / stream-close signal is never propagated. The gateway sits stuck on model_call:stream_progress until stuckSessionAbortMs (600ms) kills it, resulting in "LLM request failed." and "Request was aborted" errors.

This affects any model routed through the opencode-go provider (observed on deepseek-v4-flash).

Symptoms

• Cron jobs using opencode-go/{model} fail with "LLM request failed."
  

• Session trajectory shows prompt.submitted → 600+ seconds of silence → model.completed (aborted=true, externalAbort=true)
  

• Provider-side dashboard shows the API call completed normally with full token counts and no errors
  

• Failover is never attempted despite fallbackConfigured: true (always surface_error)
  

• Occurs intermittently — same jobs succeed on other days with identical configs
  

Steps to reproduce

1. Configure a cron job using any model through the opencode-go provider (e.g., opencode-go/deepseek-v4-flash)
2. Let it run during a routine schedule
3. After ~10 minutes — observe "LLM request failed." error
4. Check the provider dashboard — the API call shows as completed with partial or full output tokens

Expected behavior

Cron jobs to run completely.

Actual behavior

• Cron jobs using opencode-go/{model} fail with "LLM request failed."

OpenClaw version

2026.6.6

Operating system

docker

Install method

docker

Model

opencode-go/deepseek-v4-flash

Provider / routing chain

openclaw->opencode-go

Additional provider/model setup details

No response

Logs, screenshots, and evidence

Session Trajectory (from one affected run)

seq=1  session.started             05:00:01 EDT
seq=2  trace.metadata              05:00:01
seq=3  context.compiled            05:00:01
seq=4  prompt.submitted            05:00:01    18k total tokens
                          ⬐ 622 seconds of silence ⬏
seq=5  model.completed             05:10:23    aborted=true externalAbort=true
seq=6  trace.artifacts             05:10:23
seq=7  session.ended               05:10:23    promptError="This operation was aborted"

Key detail: model.completed at seq=5 shows:

The model produced 226 output tokens — visible on the provider side as a completed call — but model.completed only fired because the stuck-session recovery aborted it. The stream never closed naturally.
Gateway Log Entries (diagnostic)

stalled session: age=502s activeWorkKind=model_call
  lastProgress=model_call:stream_progress lastProgressAge=497s
  recovery=none

stuck session recovery: age=622s action=abort_embedded_run aborted=true

embedded_run_failover_decision:
  decision: surface_error
  provider: "opencode-go"
  fallbackConfigured: true
  aborted: true
  rawErrorPreview: "Request was aborted"

Impact and severity

Affected: opencode-go users
Severity: blocks cron jobs
Frequency: 90% of cron jobs are failing for me.
Consequence: Missed tasks

Additional information

No response