[FEATURE]: v2 SDK: bind agent/model/variant options to queued prompts #48356

Activity

1. github-actions

   assigned

   neriousy

   on Sep 10, 2026
2. github-actions

   added

   needs:complianceThis means the issue will auto-close after 2 hours.This means the issue will auto-close after 2 hours.

   on Sep 10, 2026
3. remorses

   changed the title [-]v2 SDK: bind agent/model options to queued prompts[/-] [+][FEATURE]: v2 SDK: bind agent/model/variant options to queued prompts[/+] on Sep 10, 2026
4. github-actions

   removed

   needs:complianceThis means the issue will auto-close after 2 hours.This means the issue will auto-close after 2 hours.

   on Sep 10, 2026
5. github-actions commented on Sep 10, 2026

   

   github-actionsBot

   on Sep 10, 2026 – with GitHub Actions

   Contributor

   More actions

   Thanks for updating the issue! It now meets our contributing guidelines — the feature request is clear, uses the correct template, includes the verification checkbox, and explains the use case well. We'll take a look.
6. remorses

   added a commit that references this issue on Sep 13, 2026

   Store Kimaki system instructions through native OpenCode session entr…

   ...

   6a0846b
7. hir0xygen commented on Sep 23, 2026

   

   hir0xygen

   on Sep 23, 2026

   More actions

   +1, hitting the same thing in the V2 TUI. Reporting the client-side repro since this issue is framed as SDK-only:

   1. Start a turn in build (agent working).
   2. While it runs, switch agent to plan and queue the prompt with ctrl+x enter.
   3. The active agent switches to plan at the next steer opportunity, before the queued prompt is delivered, so the in-flight turn finishes under plan mode.

   Expected: session-wide selection stays put, and the queued prompt's agent is applied only when that prompt is actually delivered.
   Same "apply when delivered, not when queued" semantics requested here.
8. remorses

   added 2 commits that reference this issue on Sep 25, 2026

   Store Kimaki system instructions through native OpenCode session entr…

   ...

   4eadc97

   Store Kimaki system instructions through native OpenCode session entr…

   ...

   ddaa23c
9. pasogott

   added a commit that references this issue on Oct 9, 2026

   Store Kimaki system instructions through native OpenCode session entr…

   ...

   020b1ea
10. yee94 commented on Oct 28, 2026

    

    yee94

    on Oct 28, 2026

    More actions

    Another concrete integration use case from Yuri, a multi-client UI for OpenCode: users switch models while a turn is running, then queue a follow-up expecting it to use the model selected when they submitted it. Currently we can retain that selection in metadata, but execution still uses the session-wide model, so the queued task can run with the previous model instead.

    Client-side coordination is not a viable substitute for us: users may close the app or go offline immediately after submitting. We need OpenCode's durable queue to own execution independently of the submitting client. Switching the session model at admission also does not provide per-item isolation and can affect work already running.

    The behavior proposed here is exactly what we need: persist agent/model/variant with each admitted input, apply them at the delivery boundary, and preserve existing behavior when those options are omitted. It should continue to work after the client disconnects and when multiple queued inputs select different models.

    This is a real user-facing limitation for us, not just SDK ergonomics. Is there a planned implementation or an upstream PR we can follow and help validate? Happy to provide the integration scenario and test the eventual API.
11. opencodejr commented on Oct 28, 2026

    

    opencodejrBot

    on Oct 28, 2026

    More actions

    Thanks for the request, and for the clear use case from porting Kimaki.

    To restate it: you'd like session.prompt (especially with delivery: "queue") to accept optional agent and model (including variant). These would be stored with the queued input and applied when it is delivered, so a queued prompt can run with its own selection without changing the work already running. Today the prompt input doesn't carry these, and session.switchAgent() / session.switchModel() change the whole session right away. This needs a change in core, so it can't be done as a plugin.

    Maintainers will consider it. Related work you may want to follow:

    • fix(tui): defer queued agent and model selection until delivery #54779: an open PR that saves the agent, model and variant with each queued prompt, applies them when the prompt is delivered, and adds optional agent/model fields for API callers. It references this issue.
    • fix(core): apply prompt agent selection at the delivery boundary #59290: an open PR that carries the agent selection on the prompt and applies it at delivery, fixing Switching the composer agent while a turn is running reconfigures the running turn (plan turn gains write access) #59289. It covers agents only, not models.
    • Switching the composer agent while a turn is running reconfigures the running turn (plan turn gains write access) #59289: a bug report where switching the agent while a turn is running changes the turn that is already running.