All contributors have signed the CLA. Thank you! ✅
_{Posted by the CLA Assistant Lite bot.}

I have read the CLA Document and I hereby sign the CLA

recheck

I have read the CLA Document and I hereby sign the CLA

[sisyphus-bot]

PR sweep first-pass triage on dev.

• Author: afafara
• Title: Fix primary agent model precedence
• Closing issues: none detected
• Review decision: none yet
• Mergeable state: MERGEABLE
• CI status: green / no failing checks

Awaiting a first-pass review.

Assigning code-yeongyu so the maintainer can prioritize a focused review of this PR. If the linked issue above has been triaged, the verdict there is the authoritative direction for this change.

[sisyphus-bot] Hi afafara. 🙏 Thanks for putting this together; I read through the entire patch and I want to be transparent about why I'm not landing it without a maintainer pass.

What the code does, accurately

• atlas-agent.ts and sisyphus-agent.ts now always pass uiSelectedModel into applyModelResolution, capture the resolution result, and then re-apply model + variant onto orchestratorConfig / sisyphusConfig after applyOverrides, but only when normalizeModel(uiSelectedModel) is defined. The mechanical effect is "if the user actively selected a model in the UI, the UI model wins; otherwise the configured override wins".
• The new test "uiSelectedModel takes priority over sisyphus category model override" pins the category-override branch in addition to the direct model-override branch.

Where I'd like maintainer input before this lands

This PR inverts the precedence that the existing tests on dev explicitly named:

• dev currently has test("user config model takes priority over uiSelectedModel for sisyphus", ...) and the matching for atlas test, asserting that agents.sisyphus.model === "google/antigravity-claude-opus-4-5-thinking" when the user has both a UI selection and overrides.sisyphus.model. Those test names are themselves a deliberate contract.
• This PR keeps the same test cases but flips the assertions to "openai/gpt-5.4", i.e. "UI wins". It also renames the tests to "uiSelectedModel takes priority over <agent> config model override".

That's a semantic change, not a bug fix. The existing behavior on dev was an intentional design decision (the test names spell it out), and reversing it changes behavior for every user who relied on agents.sisyphus.model / agents.atlas.model overrides in oh-my-opencode.jsonc. Until code-yeongyu confirms the intended precedence has changed, I want to leave this as a discussion rather than a merge.

Two things that would unblock

1. A short confirmation from a maintainer that the new precedence ("UI selection wins over agent/category model overrides for primary agents") is the desired contract going forward.
2. If yes, a doc / changelog blurb noting the behavior change so existing users can migrate or remove their override entries.

Smaller observations

• The model string updates in the test (gpt-5.5 → gpt-5.4) probably need rebasing against current dev, where gpt-5.5 is the in-use model string.
• The normalizeModel(uiSelectedModel) !== undefined check is a clean way to discriminate "UI explicitly set" from "no UI selection". I'd lean toward extracting that into a small named helper (hasExplicitUiSelection) since both atlas and sisyphus agents repeat the same pattern, but that's a follow-up, not a blocker.
• The { variant: _ignoredVariant, ...configWithoutVariant } pattern reads cleanly; nice job avoiding a mutation.

Once the precedence direction is confirmed by a maintainer, I'm happy to come back and finish the review.