Repository navigation
subagent: spawns fail with "The provider response ended with an unknown finish reason" #44695
Closed as not planned
Description
Summary
Subagents (e.g. explore) intermittently fail at spawn with The provider response ended with an unknown finish reason. The session dies immediately with no further detail; a retry of the same prompt often succeeds. In one session, 3 of 5 concurrently-spawned explore agents failed this way; a second attempt got 2 of the remaining ones through, and one never succeeded after repeated tries. This makes parallel research via subagents unreliable.
Environment
- opencode version: opencode2 v0.0.0-beta-18050 (beta)
- OS: Linux 6.8.0-111-generic x86_64
- Terminal: tmux (TERM=tmux-256color)
- Shell: /bin/bash
- Install/channel: beta
- Active plugins: /home/cunabu/.config/opencode/plugins/cunabu-agent-git.mjs
- Model in use: opencode-go/deepseek-v4-pro (per opencode.json)
Reproduction
- Spawn 3-5 subagents in parallel (e.g.
explorewith distinct prompts). - Observe that one or more fail immediately with
session.stream: The provider response ended with an unknown finish reason. - Re-run the identical prompt; it often completes.
Expected Behavior
Subagents run, or fail with a normal refusal/stream error that carries a reason and retries cleanly.
Actual Behavior
Immediate failure at stream start with no finish-reason detail; outcome is effectively a coin flip per spawn.
Additional Context
- Fails at session start (no long-lived context involved); different prompts fail, so content is not the trigger.
- Intermittent across runs; more frequent under parallel spawns, so worth checking whether concurrency/connection reuse plays a part.
- No server/log snippet available in this environment beyond the quoted error.
Activity
github-actions commentedon Aug 24, 2026
This issue might be a duplicate of existing issues. Please check:
core(v2): "unknown finish reason" error omits raw provider finish_reason— same error string, V2 beta, discusses the root cause and where the raw finish value is droppedAborted provider stream recorded as clean stop (finish=unknown, zero usage, no text) — subagent returns empty with no error— same symptom (subagent silently fails on unknown finish), parallel spawns angleSession silently stops on empty LLM response (finish: unknown, 0 tokens)— same loop-exit behavior on unknown finish reasonSession exits silently (exit 0, no error) when a turn returns finish "unknown"— directly tracks the loop-exit-on-unknown-finish bugIf your issue is specifically about the intermittent/concurrency aspect (more failures under parallel spawns), it may still be worth keeping open as a separate signal, but please review the above and comment if one of them already covers your case.
neriousy commentedon Oct 22, 2026
Closing this: on current V2, an unknown finish reason from the provider is treated as an incomplete stream and retried automatically when no output has started yet (packages/core/src/session/runner/step.ts, #51635). That should cover subagent spawns failing this way. If parallel spawns still fail with this error on a current build, comment here with the version and provider and we'll reopen it.