feat(models): support GPT-6 Astra reasoning variants

[Chloe-JY]

Problem

GPT-6 Astra is omitted by OpenAI API discovery and the Codex fallback catalog. Even when added manually, it falls through the generic GPT context metadata and the GPT-5-only endpoint/reasoning checks, so direct tool requests can start on the wrong API and aggregator requests can lose the selected effort. The model picker also lacks an Astra family label.

Solution

• Admit GPT-6 in OpenAI discovery and add gpt-6-astra to the Codex setup/refresh fallback catalog. Keep existing default-enabled models and live account metadata precedence.
• Expose Low, Medium, High, Extra High, Max, and the existing Codex Ultra mode, with standard/fast variants. Ultra sends max plus the existing request-local delegation instructions; it is not a new API effort.
• Route direct public OpenAI Astra requests to Responses, classify Astra reasoning for compatible providers, and select max_completion_tokens for compatible Chat Completions. Custom endpoints retain their existing endpoint policy.
• Add the GPT 6 Astra group and matching frontend/backend metadata: 1,050,000-token context, 128,000-token maximum output in the frontend descriptor, vision, and reasoning.
• Cover discovery payload projection, fallback/saved-account parity, live metadata precedence, serialized provider requests, streaming/non-streaming compatible requests, grouping, parsing, and model metadata.

Official references: Astra model specification and Astra migration guidance. Existing accounts use Refresh Models and enable Astra; this change does not refresh accounts in the background.

Potential risks

• Account access and Fast availability remain provider-controlled. Astra Fast is unavailable with EU data residency; standard variants remain available. Live Astra API/OAuth calls and regional Fast behavior were not exercised.
• Ultra retains the existing Codex delegation policy and permission checks. Live discovery can expose a narrower effort set, which remains authoritative.
• The API request change is confined to Astra: direct OpenAI uses Responses; compatible Chat Completions uses its completion-token field. Custom gateway protocol requirements still use the existing policy and error negotiation.
• No dependency, schema, IPC, or persistence-format changes. Revert this commit to roll back; any stored Astra selection should be switched to an available model. No historical data cleanup is required.

Verification

• cd src-tauri && cargo test -p key_vault --lib --no-fail-fast — 382 passed.
• cd src-tauri && cargo test -p agent_core --lib providers:: --no-fail-fast — 451 passed; 2,749 unrelated tests filtered out. No compiler warnings in either Rust test run.
• pnpm test src/util/__tests__/modelGrouping.test.ts src/util/__tests__/modelGrouping.thresholds.test.ts src/util/__tests__/modelVariants.test.ts src/types/model/info.openai.test.ts — 54 passed across four files.
• pnpm exec tsgo --noEmit --pretty false — passed.
• Commit hooks ran normally: lint-staged (Oxlint, ESLint, Prettier), TypeScript, commitlint, and cd src-tauri && cargo clippy --lib --message-format=short -p agent_core -p key_vault all passed.
• pnpm exec eslint src/util/modelGrouping.ts src/util/__tests__/modelGrouping.test.ts src/util/__tests__/modelVariants.test.ts src/types/model/info.openai.ts src/types/model/info.openai.test.ts --max-warnings 0 — passed.
• pnpm exec eslint src/ --ext .ts,.tsx,.js,.jsx --format json — 6,431 files, zero errors or warnings. Used with the verified zero cycle count to retain the required audit trailer despite the worktree-incompatible trailer helper.
• pnpm run check:circular — no cycles across 6,527 modules.
• pnpm run check:test-placement — consistent across 495 directories.
• git diff --cached --check — passed. Reviewed the full diff for scope, secrets, private paths, generated artifacts, and unrelated formatting.
• No desktop GUI/E2E or live provider inference run. This changes catalog/metadata values without changing component layout; automated grouping/variant tests cover those values, so screenshots would add little evidence. No computer control was used.

Audit

Architecture checklist: covered compilation, production call-chain wiring, naming, effort-vs-Ultra semantics, fallback branches, provider boundaries, readability, serialized payloads, setup/saved-account parity, and live metadata/default precedence (layers 1–10 within this change). Unrelated domains and session initialization were outside scope.

Area	Verdict	Evidence	Change or reason kept	Verification
Background work	keep	Discovery remains explicit setup/refresh; inference remains request-driven	No timers, subscriptions, workers, or retries added	Production call-chain/diff review
Memory	keep	One bounded catalog entry and 12 fallback variants; request-local values	No new retained structures or cache lifecycle	Catalog cardinality and request construction tests
Scope/isolation	keep	Existing account/endpoint resolution and live metadata precedence	Custom endpoint policy and user selections preserved	Endpoint policy and live catalog tests
Rendering/hot path	keep	Static tier/metadata entries; no delta-loop changes	No added work while idle, hidden, closed, signed out, or in a second instance	Parser/grouping and transport tests

Performance verdict: pass for the scoped catalog/request changes, supported by source review and passing catalog/transport tests. No CPU/RAM improvement is claimed; no resource ownership or lifecycle changed.