Task: Consume server model reasoning capabilities

Identity

Task ID: 20260901-server-model-capabilities-9e31b6c4
Mode: Feature
Branch: codex/20260901-server-model-capabilities-9e31b6c4-server-model-capabilities
Worktree: D:\w\makelore-model-capabilities-9e31b6c4
Base commit: 850947c092
Owner: codex
Status: Ready for integration

Consume per-model reasoning capabilities from Works Square model config and persist them on the imported Main-owned Provider account.
Make server metadata override the local Pi reasoning map for server-provisioned models while retaining the local verified profile as an old-server/direct-provider fallback.
Add native max to the product-neutral thinking-level contract and composer menu so DeepSeek exposes off/low/high/max without relabeling provider values.
Keep one-api and Pi 0.84.2 unchanged.

Electron Main remains the sole Works/model-config, Provider, credential, and Pi-wire owner. Renderer receives only safe product-neutral levels.
Strictly normalize the trusted response shape; a missing capability field means an older server and falls back to the local profile.
Preserve persisted user choices and all non-Works Provider behavior.
Do not restore OpenCode, add a second runtime, or expose Provider credentials or raw response data.
Feature mode may update only product code, focused tests, and this task record.

Add failing tests for Works capability import, Provider metadata persistence, server-over-local Pi mapping, and native max menu/contract acceptance.
Implement the typed response parser, Provider metadata projection, capability-map precedence, local DeepSeek fallback, and minimal max propagation through existing product validators/UI.
Run focused unit tests, typecheck, scoped lint, Vite/Main build, and the project-docs completion gate.

Parsed and normalized Works Square model_capabilities, stored the safe result on the managed Provider account, removed stale overrides when the optional server field disappears, and included the metadata in Provider runtime-shape invalidation.
Server-provided efforts now override the local Pi reasoning/thinking-level map while preserving verified local modalities, token limits, and wire-format compatibility. A capability for a future model without a local profile enables standard reasoning_effort serialization automatically.
Updated the verified DeepSeek fallback to expose only off/low/high/max. Pi sends thinking.type=disabled with no reasoning_effort for off, and sends the exact low/high/max effort for enabled requests.
Added native max to the product contract, snapshot validation, project persistence, Host route, Pi runtime, and composer UI where it is labelled 最高.
Kept Pi at 0.84.2 and made no one-api, dependency, README, or second-runtime change.

Focused final Vitest combination: 8 files and 102/102 tests passed, including actual Pi streamSimple payload checks for off and each of low/high/max.
pnpm run typecheck: passed.
pnpm run lint:check: passed with the repository's existing five warnings and no errors; a separate scoped ESLint run over every changed TypeScript/TSX file passed.
pnpm run build:vite: passed; only existing Browserslist, dynamic-import, and chunk size warnings were emitted.
git diff --check: passed.

Integrate with the paired Works Square API change. Deployment order is tolerant: a client talking to an older server uses the corrected local DeepSeek profile, and older clients ignore the new server response member.
Verify future server model entries against their provider's exact effort vocabulary before publishing them; no one-api change is required while it remains transparent.

Promote the Main-owned capability normalization/precedence rule and the safe server-to-Provider metadata contract into canonical architecture documentation after the paired server and client branches are integrated.