[Bug]: Default model config for Commit & push doesn't work with opencode #10776

Activity

1. razzeee

   added

   bugSomething is broken or behaving incorrectly.Something is broken or behaving incorrectly.

   needs-triageIssue needs maintainer review and initial categorization.Issue needs maintainer review and initial categorization.

   on Sep 8, 2026
2. juliusmarminge commented on Sep 8, 2026

   

   juliusmarminge

   on Sep 8, 2026

   Member

   More actions

   Thanks for the report — this looks like a real OpenCode commit-message generation failure, not a web-button bug.

   What we know

   The toast text is produced only in OpenCodeTextGeneration when the SDK session.prompt call throws after a successful session.create:

   Text generation failed in generateCommitMessage: OpenCode session.prompt request failed.

   That is a different failure than:

   • [Bug]: OpenCode returns "no structured output" error during commit message generation #2199 (OpenCode returned no structured output) — fixed in fix(server): handle OpenCode text response format in commit message g… #2202
   • [Bug]: GIT Text generation failed in generateCommitMessage. When Codex not installed. #5182 (Codex spawn / default writer) — different error; setting Settings → Source Control → Source control writer model was the workaround there
   • [Bug]: Thread title and branch-name generation fail silently when the text generation provider is unhealthy #5359 (silent title / branch generation) — same text-gen family, different surface

   Commit & push (git.runStackedAction → generateCommitMessage) has no fallback. If generation throws, the whole commit/push is aborted.

   Likely cause

   Chat already talks to OpenCode with session.promptAsync. Commit/PR/title/branch generation still uses blocking session.prompt on a throw-on-error SDK client (@opencode-ai/sdk@1.15.13 in 0.0.40). Newer OpenCode (1.17+ / v2) changed that prompt endpoint so it no longer returns a completed message body. A parse/schema/HTTP throw then gets rewritten to this generic toast, hiding the real cause. Official OpenCode 2 support is still open (#8207).

   Workaround

   Type a commit message in the git dialog before Commit & push — a non-empty custom message skips generation.

   Ask

   Please add:

   1. opencode --version (and whether this is OpenCode 1 or 2)
   2. Whether OpenCode chat still works in the same setup
   3. The matching server log line around the failure (the SDK cause is stored but not shown in the toast)

   Next

   Treat as accepted. Fix belongs in apps/server/src/textGeneration/OpenCodeTextGeneration.ts: align with the chat adapter’s promptAsync + wait path, and surface the SDK/HTTP error in the toast.
3. juliusmarminge

   added

   via-triageFiled through npx t3 triageFiled through npx t3 triage

   acceptedfeature request acceptedfeature request accepted

   and removed

   needs-triageIssue needs maintainer review and initial categorization.Issue needs maintainer review and initial categorization.

   on Sep 8, 2026
4. razzeee commented on Sep 8, 2026

   

   razzeee

   on Sep 8, 2026

   ContributorAuthor

   More actions

   ❯ opencode --version
   1.18.29
   

   1. It's opencode 1
   2. Regular chat works with github-copilot/gpt-5.6-sol or similar

   created ... title="T3 Code generateCommitMessage"
   ERROR ... ProviderModelNotFoundError:
   Model not found: openai/gpt-5.
   Did you mean: gpt-5, gpt-5.4, gpt-5.6?
   

   I do not have an openai subscription on this system (or ever had)
   Settings show Haiku to be configured, but I don't remember doing that or it was ages ago. Guess this might be a different setting anyway.

   
5. Destreyf commented on Sep 8, 2026

   

   Destreyf

   on Sep 8, 2026

   Last edited by Destreyf

   More actions

   @razzeee do you use opencode go or zen, i have seen a similar error when the model is disabled?

   The model definition openai/gpt-5 would definitely point to an opanai key/sub of some sort typically.

   The "Text generation model" I think is what was used for generating thread titles and commit messages depending on how it's configured at the bottom of the "Source Control" settings page, which by default falls back to the global text generation model.

   
6. razzeee commented on Sep 8, 2026

   

   razzeee

   on Sep 8, 2026

   ContributorAuthor

   More actions

   I don't use go or zen.
7. Destreyf commented on Sep 8, 2026

   

   Destreyf

   on Sep 8, 2026

   More actions

   If you modify the source control writer model do you get the same error?
8. razzeee commented on Sep 8, 2026

   

   razzeee

   on Sep 8, 2026

   ContributorAuthor

   More actions

   Works with that enabled
9. ciphrnotfound commented on Sep 8, 2026

   

   ciphrnotfound

   on Sep 8, 2026

   More actions

   I’ll take this. I’ll keep it scoped to OpenCodeTextGeneration: align commit-message generation with the existing promptAsync + wait flow and improve error propagation, with focused regression coverage.
10. josephv123 commented on Oct 6, 2026

    

    josephv123

    on Oct 6, 2026

    Contributor

    More actions

    Some findings that may change the fix direction here.

    The failure in the follow-up is ProviderModelNotFoundError: openai/gpt-5, and it went away after picking a writer model. openai/gpt-5 is the OpenCode default text-generation model (DEFAULT_TEXT_GENERATION_MODEL_BY_PROVIDER in packages/contracts/src/model.ts). Commit & push fails for any OpenCode user who has no OpenAI provider connected and hasn't set a writer model.

    Surfacing the SDK error alone wouldn't name that cause. I ran OpenCode 1.18.29 (the reporter's version) and called session.prompt through @opencode-ai/sdk 1.15.13 with throwOnError: true, as OpenCodeTextGeneration does, using openai/gpt-5 with no OpenAI provider configured. The server answers 500 with:

    { "name": "UnknownError", "data": { "message": "Unexpected server error. Check server logs for details." } }

    The toast would only change from "session.prompt request failed" to "Unexpected server error". Both OpenCode text-generation paths also deliberately use fixed error text.

    Two small options, both product calls:

    • Name the model in the failure, e.g. "OpenCode couldn't generate with openai/gpt-5. Choose a writer model in Settings → Source Control." That's enough for the user to fix it.
    • Don't default OpenCode text generation to a model from a provider the user hasn't connected.
11. razzeee

    changed the title [-][Bug]: Commit & push doesn't work with opencode[/-] [+][Bug]: Default model config for Commit & push doesn't work with opencode[/+] on Oct 15, 2026
12. t3dotgg commented on Oct 21, 2026

    

    t3dotgg

    on Oct 21, 2026

    Member

    More actions

    Note

    🤖 Claude Opus 5.5 responding on behalf of Theo

    Thanks for the report. One part of this is fixed on main in #19386. Please test on the nightly built from main on 2026-10-21, or on the next stable release. One part is still open: Other OpenCode-only setups can still fall back to openai/gpt-5 and show a generic generation error. If that part still matters to you, please open a new issue for that part only.
13. t3dotgg

    closed this as completedon Oct 21, 2026