fix(session): don't clobber user's model swap on model-less prompts

[Qiiks]

Issue for this PR

Closes #42893

Type of change

• Bug fix

What does this PR do?

A model-less prompt (no input.model — e.g. plugin completion-reminders via promptAsync) resolves from the agent fallback (ag.model) for the turn. The existing design (PR #26765) intentionally uses the agent's model when no explicit model is given — this is correct for the turn's model and is preserved.

However, setAgentModel then overwrites the session row to the resolved model, permanently destroying a model the user explicitly stored via a prior swap. The TUI is unaffected (always sends draft.model explicitly), but opencode serve / mobile clients lose the swap on the next plugin reminder.

The fix: guard the setAgentModel write in createUserMessage — skip it when:

1. The prompt carried no explicit model (input.model == null), AND
2. The session already has a stored model that differs from the agent's resolved model (storedModelDiffers)

Fresh sessions (no stored model) still persist the agent's model on the first model-less prompt (hasNoStoredModel guard), and explicit-model prompts always persist normally. The model-resolution precedence is unchanged — model-less prompts still use the agent's model for the turn.

How did you verify your code works?

• New regression test: "model-less prompt does not clobber user's stored model swap" — creates a session, swaps to kimi-k2.5-free, sends a model-less prompt with an agent carrying test/test-model, asserts the turn uses test-model (design preserved) but the session row stays kimi-k2.5-free (the fix). Verified RED→GREEN.
• Fresh-session test: "model-less first prompt on fresh session persists agent model to row" — confirms the hasNoStoredModel guard fires setAgentModel on a fresh session so the row is initialized.
• Existing test preserved: "applies agent variant only when using agent model" (the test that pins the agent-model design from Persist session model switches outside event flag #26765) still passes — model-less prompts still use the agent's model for the turn.
• Full suite: 46 pass / 14 skip / 0 fail across test/session/prompt.test.ts.
• Typecheck: 30/30 packages clean (turbo typecheck).

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[HermanGrondahl]

This seems to fix the original model-clobber case, but I found a related state-update edge case: the guard can also suppress an explicit agent change because setAgentModel() persists the agent and model together.

Consider this sequence:

• stored session: agent = plan, model = A
• A was explicitly selected by the user
• next prompt explicitly selects agent = build
• the prompt has no explicit input.model
• build resolves fallback model B

The PR correctly prevents fallback model B from overwriting the explicitly stored model A.

However, because wouldClobberStoredModel skips the entire setAgentModel() call, the explicit agent change is skipped as well. The persisted session remains:

plan + A

rather than:

build + A

In this case the two fields need different treatment: preserve the stored model, but still persist the explicitly selected agent.

I think a regression covering this transition would be useful:

• before: plan + A
• input: explicit build, no explicit model
• build fallback: B
• after: build + A

That would preserve the model-swap fix without leaving the persisted agent stale.

[Enough1122]

AI code review — automated review for reference; please use your judgment.

• packages/opencode/src/session/prompt.ts:679 — logic checks out for the four model/agent combinations (explicit model always persists; model-less + stored-model-differs skips; fresh session initializes). Good minimal fix that preserves the Persist session model switches outside event flag #26765 turn-resolution design.
• prompt.ts:682 — edge case worth a decision: when input.model == null and the agent also changed (current.agent !== info.agent), wouldClobberStoredModel blocks the whole setAgentModel write, so the agent switch isn't persisted either. Since setAgentModel writes both fields together, you can't skip just the model today — but silently dropping an agent change may surprise callers. Consider persisting agent while keeping the stored model, or documenting the tradeoff.
• prompt.ts:24 — comment clearly explains intent; nice.
• Tests: both new instance tests assert the important invariant pair (turn uses agent model, row keeps user swap) plus fresh-session initialization. Thorough.
• Nit: test/session/prompt.test.ts:2443 — double blank line introduced at top of the new block; prettier will likely flag. No blocking issues found.

[github-actions]

Automated PR Cleanup

Thank you for contributing to opencode.

Due to the high volume of PRs from users and AI agents, we periodically close older PRs using automated criteria so maintainers can focus review time on the most active and community-supported contributions.

This PR was closed because it matched the following cleanup criteria:

• The PR was created more than 1 month ago
• The PR had fewer than 2 positive reactions
• Positive reactions are counted as thumbs-up, heart, celebration, or rocket reactions on the PR

PRs created within the last month are not affected by this cleanup.

If you believe this PR was closed incorrectly, or if you are still actively working on it, please leave a comment explaining why it should be reopened. A maintainer can review and reopen it if appropriate.

Thanks again for taking the time to contribute.