feat(models): enrich provider model list from models.dev

[hshinosa]

What

Bundle a snapshot of the models.dev community catalogue (930 models, 19 providers, ~400 KB JSON) and use it to enrich llm.ModelInfo when the provider's own /models endpoint leaves fields blank.

Live provider values always win; the snapshot is metadata-only (context window, pricing, capability flags) — never a source for "what models exist".

The snapshot ships as a JSON asset consumed via //go:embed, not as a generated Go source file full of struct literals. Data as data, not code: review diff stays small, GitHub auto-collapses via linguist-generated, and the runtime cost is one json.Unmarshal at init (~ms for 900 entries).

Regenerate with make sync-models.

Why

Provider /models endpoints rarely carry the metadata a picker needs:

• Anthropic hard-coded ContextWindow: 200000, MaxOutput: 64000, Vision: true, Tools: true, Reasoning: true for every model. That lied for Claude Sonnet 4.6 (real: 1M) and every future model with different capabilities.
• Gemini exposed real InputTokenLimit/OutputTokenLimit but tagged reasoning with strings.Contains(id, "2.5") — wrong for every non-Gemini id a proxy happens to re-serve.
• Most providers report zero pricing.

The dashboard's Models page, the composer gauge, and the TUI's picker therefore either showed wrong numbers or none at all. Users who picked Claude Sonnet 4.6 saw a 200k gauge and Antares would trigger compaction at ~160k when the real window was 800k unused.

Cascade

Context window and enrichment walk the same order:

1. Per-provider user override in cfg.Providers[…].ModelMeta (kept from upstream).
2. Strict providers.MetaByProvider — exact provider/id, with slug variant fallback (Anthropic ships claude-opus-4.7 while models.dev catalogues claude-opus-4-7 — same model, different spelling).
3. Proxy fallback via MetaByAnyProvider — only for openai-compatible and custom providers (EnxAPI, LiteLLM, self-hosted endpoints) that re-serve official ids. First-party providers that genuinely do not carry a model never inherit some other provider's entry.
4. cfg.Model.ContextWindow global fallback.
5. 128000 safety net.

The earlier cascade iterated every provider in the config and returned the first hit, which caused MetaByProvider's loose fallback to attribute the wrong entry to unrelated providers — e.g. reading Anthropic's row for a Kimi request. The new candidateProviders walks only providers that actually claim the id in their Models list or ModelMeta keys.

Adapter cleanup

• internal/llm/anthropic.go: drop the hard-coded metadata block; emit id+name only.
• internal/llm/gemini.go: drop the reasoning heuristic. Keep the real token limits the API returns.

Composer gauge

• ChatPage refetches /context-window whenever activeModel changes, passing the qualified provider/model id so the server honours the caller's provider hint.
• ModelPicker emits an antares:model-changed event on successful /model/set; StatusPill listens so the "Connected · " label refreshes at once instead of waiting up to 10 s for the next poll.

TUI

• New providers.FetchModelInfos preserves the full ModelInfo from the live probe. FetchModels still returns []string for cmd/antares/main.go's model-listing subcommand.
• Model picker's right column now shows 1000K ctx · anthropic when the cascade knows the window, falling back to the provider name.
• Header adds 145K/1000K (14%) alongside the ↑↓ counters; a Context section joins the Tokens section in the sidebar. /status prints the same ratio. Both fall back to the metadata cascade for the active model so a fresh session shows the model's budget before the first turn fires the live usage event.
• The agent-side contextWindowFor already picks up the new cascade, so compaction now triggers at the model's real window (Claude Sonnet 4.6 compacts at 800k instead of 160k) even for TUI-only users.

Files

Handwritten changes total ~1.2K lines. The bundled snapshot (models_generated.json, ~400 KB) is linguist-generated so GitHub collapses it in reviews.

File	Purpose
internal/providers/metadata.go (+264)	ModelMeta, cascade helpers, slug variants
internal/providers/metadata_test.go (+157)	11 subtests covering cascade layers
internal/providers/generated_models.go (+59)	Tiny //go:embed loader — 60 lines, not 12k
internal/providers/models_generated.json (+15,399)	Bundled snapshot, collapsed in review
scripts/sync-models-dev.go (+349)	Regenerator
internal/providers/catalog.go, models.go (+44/-8)	ContextWindow() cascade, FetchModelInfos helper
internal/agent/compact.go (+77/-8)	contextWindowFor uses new cascade
internal/llm/anthropic.go, gemini.go (+24/-5)	Drop hard-coded metadata + heuristic
internal/server/handlers_config.go, handlers_providers.go (+215/-14)	User-override plumbing + enrichment endpoint
internal/tui/* (+117/-13)	Context header, sidebar section, picker badge, /status line
web/src/pages/ChatPage.tsx, components/chat/ModelPicker.tsx, components/layout/StatusPill.tsx (+25/-5)	Composer gauge refetch + model-changed event
Makefile, .gitattributes (+7/-0)	sync-models target + linguist tag

Not breaking

• Existing configs untouched; ModelMeta struct gains json: tags (additive) with no runtime behaviour change.
• Live provider values still win when non-zero — snapshot only fills blanks.
• The 3-entry hand-curated contextWindows map stays as an escape hatch.
• No new runtime dependencies. Sync script uses stdlib net/http + encoding/json.
• FetchModels API unchanged; FetchModelInfos is additive.

Validation

• GOTOOLCHAIN=go1.26.3 go test ./internal/providers/... ./internal/agent/... ./internal/server/... ./internal/config/... ./internal/llm/... ./internal/tui/... — all pass, 11 new subtests cover cascade layers.
• GOTOOLCHAIN=go1.26.3 go vet ./... clean.
• bun run typecheck in web/ clean.
• Smoke: bundled snapshot resolves Claude Sonnet 4.6 → 1M (was 200k), GPT-5.4 → 1.05M, Kimi K2.5 via enx proxy → 262k (from opencode entry), unknown model → 128k safety net (never fabricates).
• Live TUI smoke: go run ./cmd/antares tui shows the sidebar Context section and header ratio for the active model on first launch.

[enowdev]

Fixed and merged with corrective commit 7b3b5d8. Review found four reproducible issues: runtime refresh omitted seven providers present in the bundled snapshot; empty refresh data erased the catalogue; Attachment incorrectly enabled Vision for non-image models; and equal-ranked proxy candidates produced different context windows across repeated lookups. The fix shares the provider allowlist between runtime and generator, retains last-known-good data on empty refresh/cache input, derives Vision only from image capability, and applies deterministic provider ranking and tie-breaking. Added regression coverage for these cases. Local provider/server/agent/LLM/TUI tests, go vet, integrated frontend typecheck/tests/build, and GitHub go/web/smoke checks passed. Closed by merging the corrected PR, not by discarding it.