Skip to content

UI hangs indefinitely when LLM stream fails mid-flight (fetch failed / network drop) #46583

Description

@Rajeshwaran-R

Description

The TUI freezes and becomes unresponsive when the LLM provider stream errors out mid-flight with a "fetch failed" error. The UI enters a waiting state with no recovery path — no timeout fires, and the interface stays frozen indefinitely while the process continues to consume ~31% CPU in a render loop.

Steps to Reproduce

  1. Start opencode with an LLM provider that requires a network connection
  2. Send a message to the agent
  3. Have the network connection drop or the provider return an error mid-stream
  4. Observe: the UI freezes and never recovers

Expected Behavior

The UI should detect the stream failure and display an error, allowing the user to retry or exit cleanly.

Actual Behavior

  • TUI hangs indefinitely — no error displayed, no way to recover without killing the process
  • Process stays alive consuming ~31% CPU (stat: Rl+)
  • Main thread blocked on ep_poll (epoll waiting for I/O that never arrives)
  • Log shows the stream error was recorded, but the UI never unblocks

Log Evidence

level=ERROR message="stream error" providerID=zcode modelID=claude-sonnet-4-6
  agent=title error.error.type=api_error error.error.message="fetch failed"

level=ERROR message="stream error" providerID=zcode modelID=claude-sonnet-4-6
  agent=build error.error.type=api_error error.error.message="fetch failed"

The title agent failed at T+0:05, the build agent at T+2:21. After that, the UI was frozen for 5+ hours until manually killed.

Sequence of Events

10:13:10  models.dev fetch → TimeoutError
10:16:40  New session created, user message sent
10:16:45  Title agent stream → "fetch failed"
10:19:06  Build agent stream → "fetch failed"
          ← UI freezes here, no recovery

Process Snapshot (at hang)

PID   STAT  %CPU  THREADS  ELAPSED  WCHAN
58867  Rl+   31.1    33     4:08:31  ep_poll
  • 33 threads, most in futex_wait (normal)
  • Main thread: ep_poll — blocked waiting for I/O event that never fires
  • No subprocess leak, no DB lock, SQLite WAL clean

Environment

  • opencode version: 1.18.25
  • OS: Ubuntu 24.04.4 LTS (Noble Numbat)
  • Kernel: 6.8.0-138-generic
  • Platform: Linux x86_64
  • Node: v24 (Bun runtime)
  • Provider: zcode (Claude Sonnet 4.6 proxy)

Suggested Fix

Add a stream-level timeout or watchdog: if a stream has errored and the promise has not resolved within N seconds, surface the error in the UI and unblock the render loop. A Promise.race(stream, timeout) pattern around the provider stream would prevent the indefinite wait.

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions