Skip to content

[Bug] Agent stops without leaving artifacts #3050

Closed as not planned
Closed as not planned

Description

@awaescher

What's wrong? Maybe some steps to reproduce.

Pretty often, the agent just stops before producing any results, neither in the chat nor as file output.
I had this today with a new stealth model ox-alpha but also with Qwen 2.6/2.8 27b running locally.

Where does it happen?

Desktop (macOS)

Version (if known)

OpenChamber 1.19.0 , OpenCode 1.18.18

Screenshots / recordings (optional)

Image

Ignore the continuation after "und?" which was then aborted by me.

Logs (optional)

Activity

  1. added
    area:agentsAgents, subagents, multi-run, agent manager
    platform:macosmacOS desktop (Electron/Tauri)
    priority:highMaintainer-set only: blocks core workflows, data loss, or many users
    needs-infoNeeds more info from reporter to reproduce
    on Aug 21, 2026
  2. openchamber-bot commented on Aug 21, 2026

    @openchamber-bot
    Contributor

    Thanks for the report. An agent run that ends without any chat or file output is a core workflow breaker, and seeing it on both ox-alpha and local Qwen suggests it's not tied to one model.

    To help us pin this down, could you share a bit more?

    • The logs section came through empty. Any log output from around the time the run stops would help a lot.
    • Does the run stop with an error, or does it just go quiet? Does the session still show up in the sidebar afterward?
    • Does it happen at a particular point (start of a run, mid-response), or does it feel random?

    The screenshot is helpful too. If anything on screen looks unusual at the moment it stops, that's worth mentioning.

  3. awaescher commented on Aug 21, 2026

    @awaescher
    Author

    I can't find any thinking traces that got cut off, which could show that the model came to a limit and had to stop, forced by the provider. But it looks okay to me in OpenChamber.

    Image

    However, I did the same in the OpenCode UI and it did not finish either. Here however it looks like it got force-stopped:

    Image
  4. yulia-ivashko commented on Sep 15, 2026

    @yulia-ivashko
    Contributor

    Your comparison with the OpenCode UI shows the turn was stopped there too. 563911ad8, included in v1.23.0, improves OpenChamber's handling by displaying failed-turn errors and diagnostic history instead of leaving the failure silent. It does not repair the underlying OpenCode/provider turn failure, so this is closing as not planned here. Thanks for checking both clients and sharing the results.

  5. awaescher commented on Sep 16, 2026

    @awaescher
    Author

    Related to #2809, setting OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX actually works. Defining max_tokens in the OpenCode provider config does not work (yet).

    Discusstion over at: anomalyco/opencode#1735

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    area:agentsAgents, subagents, multi-run, agent managerbugSomething isn't workingneeds-infoNeeds more info from reporter to reproduceplatform:macosmacOS desktop (Electron/Tauri)priority:highMaintainer-set only: blocks core workflows, data loss, or many users

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions