[BUG]:session.processor crashes sidecar on unhandled AbortError #26667

Activity

1. github-actions

   assigned

   kitlangton

   on May 10, 2026
2. github-actions commented on May 10, 2026

   

   github-actionsBot

   on May 10, 2026 – with GitHub Actions

   Contributor

   More actions

   This issue might be related to existing issues. Please check if any of these are duplicates or related:

   • Sessions permanently stuck after server restart or stream interruption — no startup recovery for orphaned messages/tool parts #19023: Sessions permanently stuck after server restart or stream interruption — covers similar scenario of unhandled stream interruption in session.processor and orphaned state, though focused on stuck sessions rather than sidecar crash
   • Windows desktop local sessions can hang/terminate on exact-file tasks (glob NotFoundError / undici terminated) #26180: Windows desktop local sessions can hang/terminate — also surfaces AbortError: Aborted from stream interruption on Windows, though triggered by a different code path (glob/file tasks)
   • Streaming ECONNRESET / "socket connection was closed unexpectedly" with SGLang OpenAI-compatible provider during tool-using agent runs (Qwen3.5 + reasoning content) #22803: Streaming ECONNRESET / socket connection closed unexpectedly — similar symptom of stream connection drops causing session.processor errors during long agent runs

   If your issue is distinct from these (e.g., the sidecar process fully exits rather than just erroring), it is worth keeping open as a separate report. It may help to note whether the sidecar process fully terminates or just logs the error and continues.
3. lb714-bit commented on May 10, 2026

   

   lb714-bit

   on May 10, 2026

   Last edited by lb714-bit

   Author

   More actions

   I've reviewed all three and my issue is distinct:

   • Sessions permanently stuck after server restart or stream interruption — no startup recovery for orphaned messages/tool parts #19023 — sessions getting stuck (process survives). Mine: sidecar process fully crashes and exits.
   • Windows desktop local sessions can hang/terminate on exact-file tasks (glob NotFoundError / undici terminated) #26180 — sessions hang/terminate on Windows with AbortError, but triggered by glob/file tasks. Mine: triggered by LLM stream interruption during continuous agent loops (repeated API requests). The error path is the same (session.processor → AbortError: Aborted via Effect.js fiber), but the trigger and severity differ.
   • Streaming ECONNRESET / "socket connection was closed unexpectedly" with SGLang OpenAI-compatible provider during tool-using agent runs (Qwen3.5 + reasoning content) #22803 — ECONNRESET during streaming, process continues. Mine: process does not continue — it fully terminates.

   the sidecar process fully terminates, not just logs an error. After each crash, the only logs produced are MCP initialization errors from the fresh restart, confirming the process fully died.

   The key distinction from all three: in my case the unhandled AbortError in session.processor crashes the process outright. The other issues deal with the process surviving but sessions becoming stuck or errored.

   Model info: zai-coding-plan / glm-5.1 (ZhipuAI GLM-5.1 via OpenAI-compatible API).
4. lb714-bit

   changed the title [-]session.processor crashes sidecar on unhandled AbortError[/-] [+][BUG]:session.processor crashes sidecar on unhandled AbortError[/+] on May 10, 2026
5. lb714-bit commented on May 11, 2026

   

   lb714-bit

   on May 11, 2026

   Author

   More actions

   Reproduction with Evidence — Confirmed on v1.14.48, Two Different Providers

   I'm experiencing the exact same issue on OpenCode Desktop v1.14.48 (Windows 11). I cleared all logs and captured two separate crashes in a clean environment. Both crashes share the identical AbortError stack trace in session.processor, but were triggered via two completely different model providers — confirming this is an OpenCode-side parsing/handling bug, not a provider-specific issue.

   ---

   Crash #1 — Zhipu GLM-5.1 (empty stream response)

   Server log (~/.local/share/opencode/log/2026-05-11T105738.log):
   ERROR 2026-05-11T11:07:55 +611226ms service=session.processor session.id=ses_1e947a4e3ffeRYMSOvWO4I1giK messageID=msg_e16b8aa9d0014wgPqlh47p7dE8 error=Aborted stack=AbortError: Aborted

   main.log (%APPDATA%\@opencode-ai\desktop\logs\main.log):
   [2026-05-11 19:07:55.851] [warn] sidecar exited { code: 1 }

   Message data from OpenCode DB — the API returned completely empty:

   {
     "message": {
       "id": "msg_e16b8aa9d0014wgPqlh47p7dE8",
       "role": "assistant",
       "modelID": "glm-5.1",
       "providerID": "zhipuai-coding-plan",
       "cost": 0,
       "tokens": { "input": 0, "output": 0, "reasoning": 0 }
     },
     "parts": []
   }

   Zero tokens. Empty parts. The stream opened and closed with no content.

   Crash #2 — big-pickle via opencode provider (interrupted mid-stream)

   Same session, 16 minutes later, after manual restart and switching to a completely different model and provider.
   main.log:
   [2026-05-11 19:23:44.555] [warn] sidecar exited { code: 1 }
   Message data from OpenCode DB — the stream was interrupted mid-generation:

   {
     "message": {
       "id": "msg_e16c749eb0017Zx5MIEF92HRfi",
       "role": "assistant",
       "modelID": "big-pickle",
       "providerID": "opencode",
       "cost": 0,
       "tokens": { "input": 0, "output": 0, "reasoning": 0 }
     },
     "parts": [
       { "type": "reasoning", "text": "Let me gather the context directly..." },
       { "type": "text", "text": "" },
       { "type": "tool", "tool": "read",  "state": { "status": "completed" } },
       { "type": "tool", "tool": "glob",  "state": { "status": "running" } },
       { "type": "tool", "tool": "glob",  "state": { "status": "running" } },
       { "type": "tool", "tool": "glob",  "state": { "status": "pending", "input": {} } }
     ]
   }

   The stream was interrupted mid-generation — two glob tool calls stuck in "running", one stuck in "pending" with empty input. The response never completed.

   wo entirely unrelated LLM providers trigger the identical crash. The common factor is OpenCode's session.processor not gracefully handling stream interruptions — whether the stream returns empty or is cut off mid-generation.

   The session.processor Effect.js fiber lacks an error boundary for AbortError, so any stream interruption propagates up and kills the entire sidecar process.

   Environment
   OpenCode Desktop: v1.14.48
   OS: Windows 11 (10.0.26200)
   Plugin: oh-my-openagent@latest

   oh-my-opencode.log

   desktop-logs-main.log

   opencode-log-2026-05-11T105738.log
6. fibbery commented on May 13, 2026

   

   fibbery

   on May 13, 2026

   More actions

   When will this problem be solved? I also encountered the same issue, as soon as the AI reports an error, OpenCode Desktop will crash, but there's no problem using TUI.
7. UncleXJ commented on May 15, 2026

   

   UncleXJ

   on May 15, 2026

   More actions

   Confirmed on v1.15.0 + Windows 11 (with full crash trace)

   Still reproducible on OpenCode Desktop v1.15.0 (released 2026-05-15), Windows 11 x64.

   Crash Trace

   From sidecar log (share/log/2026-05-15T045357.log):

   \
   ERROR service=session.processor session.id=ses_xxxx messageID=msg_xxxx error=Aborted stack=AbortError: Aborted
   at new DOMException (node:internal/per_context/domexception:76:18)
   at Array. (node-CDG1DJG7.js:293605:27)
   at Generator.next ()
   at ~effect/Effect/successCont (node-CDG1DJG7.js:7907:29)
   at ~effect/Effect/evaluate (node-CDG1DJG7.js:7921:25)
   at FiberImpl.runLoop (node-CDG1DJG7.js:7733:112)
   at FiberImpl.evaluate (node-CDG1DJG7.js:7703:26)
   \\

   Crash Pattern

   • Sidecar starts normally, serves requests
   • After completing a conversation turn (LLM response + tool calls), the \session.processor\ throws an unhandled \AbortError\
   • Sidecar exits with code 1
   • OpenCode auto-restarts sidecar, but all in-progress context is lost

   Timeline from main.log

   \
   [12:53:55] sidecar started
   [12:55:46] sidecar exited { code: 1 } // ~2 min after first request
   [12:56:51] sidecar exited { code: 1 } // second crash after auto-restart
   \\

   Frequency

   Crash is not 100% deterministic — some turns complete fine, others trigger the AbortError. Estimated ~30-50% of turns crash. No obvious trigger pattern (not limited to tool calls, long responses, or specific providers).

   OMO Interaction

   We have \oh-my-openagent\ installed. OMO registers an \uncaughtException\ handler that calls \process.exit(1), which converts the unhandled AbortError into a hard sidecar crash. However, the root cause is OpenCode's \session.processor\ not catching AbortError, not OMO.

   Evidence: disabling OMO changes the crash behavior (sidecar may survive but session becomes corrupted), confirming the underlying AbortError is from OpenCode code.

   Workaround

   None found. The crash happens regardless of:

   • OpenCode version (tested v1.14.50, v1.14.51, v1.15.0)
   • OMO version (tested v4.0.0, v4.1.2)
   • Provider or model

   Suggested Fix

   \session.processor\ should catch \AbortError\ gracefully — log the error, clean up the streaming state, and allow the session to continue rather than propagating the exception as unhandled.
8. WhiteBite commented on May 17, 2026

   

   WhiteBite

   on May 17, 2026

   More actions

   Confirming connection leak issue on latest Windows 11 + adding logs
   Environment:

   • Windows 11 24H2 (10.0.26200)
   • OpenCode Desktop 1.14.41
   • Using Fireworks AI API (minimax-m2p7, kimi-k2p6)
   • 15 MCP servers configured (mix of local stdio + remote)
   • No proxy (direct connection)
     Reproduction:
   • Happens multiple times per day (5+ times today)
   • Triggers on: file glob tasks, MCP tool calls, LLM streaming responses
   • Not only after git commit — also on regular file operations
     Observed errors from logs:
     ERROR service=llm providerID=fireworks-ai error={"name":"AI_APICallError","cause":{"errno":-4079,"code":"ECONNABORTED","syscall":"write"}}
     ERROR service=llm providerID=fireworks-ai error={"name":"AI_APICallError","cause":{"errno":-4077,"code":"ECONNRESET"}}
     ERROR service=session.processor error=Aborted
     Additional symptom:
   • Desktop UI freezes completely
   • Lost connection to local server (Sisyphus can't reach CLI)
   • Need to fully kill all opencode processes and restart
   • Empty log files (0 bytes) appear during crash
9. Brendonovich commented on May 18, 2026

   

   Brendonovich

   on May 18, 2026

   Member

   More actions

   Thanks for the error logs @lb714-bit, looking into this now
10. Brendonovich commented on May 18, 2026

    

    Brendonovich

    on May 18, 2026

    Member

    More actions

    The AbortError being referenced is never thrown, it's only created via new DOMExpection.
    This seems to be a windows-only problem since I've seen a similar thing happen on there where the server goes offline, likely related to invalid tool calls instead of this AbortError.

    
11. lb714-bit commented on May 19, 2026

    

    lb714-bit

    on May 19, 2026

    Author

    More actions

    Fair points. The invalid tool call theory sounds plausible — can't confirm or deny that from my end.
    That said, OpenCode's web and TUI modes on Windows work fine, so it doesn't seem to be purely a Windows thing.
    Ultimately though, regardless of what triggers the interrupt, the real issue is that halt() produces an error that nothing catches — and that's what crashes the sidecar. The trigger and the crash are separate problems. Even if the root cause gets fixed, the fiber still needs proper error handling to prevent this class of crashes.
12. lb714-bit

    closed this as completedon Jun 1, 2026