feat(agent): show backend-selected thought level - #3734
Conversation
…alue The thinking-level pill and submenus now render ONLY what the backend reports as active. The single source of truth is the config option's current_value (which the Core seeds from the assistant default and keeps reconciled); the UI no longer invents a selection: - getCurrentThoughtLevelLabel/composeRuntimeSelectorLabel accept a localized default label and render 'model · Default' when the thought axis exists but no current is known — never options[0], which may not be what the backend actually runs. - GuidModelSelector drops its options[0] normalization, and useGuidAssistantSelection no longer pre-seeds the selection: an untouched picker sends NO thought_level override at create, so the backend resolves the real default (the old implicit override silently defeated assistant fixed defaults). - useAcpConfigOptions gains an anti-flicker merge: a whole-snapshot replace whose options carry no current information at all preserves known non-null currents (matched by category then id, only when still selectable); a frame with any non-null current stays authoritative, so the Core's reject re-push still clears a refused level. - Thought-level matching falls back to ids effort/thinking/ thinking_budget (category thought_level stays priority; 'thinking' keeps this compatible with upstream PR iOfficeAI#3597). Reuses the existing common.default i18n key (present in all locales).
Codecov Report❌ Patch coverage is 📢 Thoughts on this report? Let us know! |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 6be5085e10
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| const informed = next.some((option) => option.current_value != null); | ||
| if (informed) return next; |
There was a problem hiding this comment.
When a backend intentionally clears the only exposed config option—or clears every option at once—its snapshot contains only null current_value fields. This branch classifies that authoritative update as informationless and restores every still-selectable previous value, leaving the selector showing stale model or thought-level state. A non-null sibling is not a reliable indicator of whether nulls are authoritative; preservation needs to distinguish omitted/partial data from explicit nulls rather than treating every all-null snapshot as empty information.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
This is an intentional anti-flicker tradeoff, and the distinction you're describing isn't representable at the current DTO level.
mergeSnapshotPreservingKnownCurrents only preserves previous currents for a frame where every option's current_value is null. Telling an authoritative all-null snapshot (a backend deliberately clearing every exposed option) apart from an informationless one (an early catalog push before currents land, or an older Core that never stamps currents) requires signaling omitted vs explicit null — but AcpConfigOptionDto carries current_value: null as the single representation for both cases. There is no field today that marks a null as authoritative.
Given that, preferring preservation for all-null frames is the safer default: it avoids the far more common flash-to-Model-only regression, while any frame carrying at least one non-null current stays authoritative — so the Core's reject re-push still wipes a refused level (the corrected frame keeps the model current, so its effort null passes through). The deliberate clear-everything-at-once case needs a DTO change to distinguish explicit nulls, which is out of scope for this PR. Leaving this thread open to track that as a follow-up rather than expanding scope here.
…override The defaults effect fell back to `options[0]` when an assistant had no resolved thought-level default and the backend reported no current value. That pre-selected a concrete level which the send path then serialized as an explicit `thought_level`, silently defeating the assistant's backend default — the same implicit override useGuidAssistantSelection already stopped emitting. Drop the `options[0]` fallback so an untouched picker stays empty (`''`) unless the backend actually reports a current value; the send path then omits `thought_level` and the backend resolves the real default. Add an end-to-end send-path regression test asserting that an untouched (empty) selection serializes to `thought_level: undefined` in the conversation create request.
1 participant
Pull Request
Description
Make thought-level indicators and selectors reflect the backend's reported
current_valueinstead of inferring that the first available option is selected.This keeps conversation and guide selectors honest when the backend is still resolving a default, preserves known current values across empty snapshot refreshes, and avoids sending an implicit thought-level override from an untouched guide picker. The selector falls back to the existing localized “Default” label until the backend reports an active value.
Related Issues
thinkingoption identifierType of Change
fix— Bug fix (non-breaking change which fixes an issue)feat— New feature (non-breaking change which adds functionality)perf— Performance improvementrefactor— Code restructuring (no behavior change)docs— Documentation updateAtomic PR Checklist (Rule 1)
<type>(<scope>): <subject>(English)Local Checks (Rule 3)
bun run format:check— formatting passesbun run lint -- --quiet— no lint errorsbunx tsc --noEmit— no type errorsbunx vitest run— 2,593 tests passed; 5 skippedbun run i18n:types+node scripts/check-i18n.js)common.default)Runtime Verification
Screenshots
Not applicable; this changes selection-state accuracy without introducing new UI structure.
Additional Context
The full pre-push gate completed successfully from a clean worktree with dependencies installed from the lockfile. Existing translation-completeness warnings remain for
es-ES,fr-FR, andfa-IR; validation passes and this PR adds no translation keys.