chore(ai): refresh the generated model catalog from live provider sources

[snimu]

What

First explicit, reviewed refresh of the generated model catalog (npm run generate-models), per the RES-1269 discussion: catalog regeneration is a deliberate, audited step rather than a CI side effect. This PR stands alone against current main: CI's build regenerates the catalog, and the committed file here is byte-identical (md5 af81f536fceafbdeac473ee4f6f5b0d8) to a fresh regeneration at validation time, so committed == live. If upstream shifts before merge and CI fails, the catalog should be regenerated and this PR updated.

Totals: 1238 -> 1262 models (41 added, 17 removed). No provider dropped wholesale or lost >30% of its rows (the silent-[] fetch-failure signature from #2008 did not occur; generator logs showed all four sources loading: models.dev 564, OpenRouter 294, Vercel AI Gateway 233, Prime Inference 107).

Per-provider changes

provider	before	after	delta
amazon-bedrock	119	122	+3 / -0
anthropic	13	14	+1 / -0
cloudflare-ai-gateway	31	32	+1 / -0
cloudflare-workers-ai	17	18	+1 / -0
fireworks	17	19	+3 / -1
github-copilot	33	27	+2 / -8
google	17	17	+1 / -1
groq	6	7	+1 / -0
huggingface	69	71	+2 / -0
opencode	60	63	+4 / -1
opencode-go	24	27	+3 / -0
openrouter	292	295	+8 / -5
prime-inference	106	107	+1 / -0
vercel-ai-gateway	224	233	+10 / -1

Providers not listed are unchanged.

Added

• amazon-bedrock: anthropic.claude-fable-5-1, global.anthropic.claude-fable-5-1, us.anthropic.claude-fable-5-1
• anthropic: claude-fable-5-1
• cloudflare-ai-gateway: claude-fable-5.1
• cloudflare-workers-ai: @cf/zai-org/glm-5.3
• fireworks: accounts/fireworks/models/deepseek-v4-flash-vision-exp, accounts/fireworks/models/glm-5p3, accounts/fireworks/models/glm-5p3-flash
• github-copilot: claude-fable-5.1, gemini-3.8-flash
• google: gemini-3.8-flash
• groq: qwen/qwen3.8-27b
• huggingface: deepseek-ai/DeepSeek-V4-Flash-Vision-Exp, zai-org/GLM-5.3
• opencode: claude-fable-5-1, gemini-3.8-flash, ling-3.0-flash-fin-free, muse-spark-1.3-contributor-free
• opencode-go: hy4-preview, muse-spark-1.3-contributor, omen-alpha
• openrouter: anthropic/claude-fable-5.1, google/gemini-3.8-flash, ibm-granite/granite-4.2-8b, inception/mercury-2.5-preview, inclusionai/ling-3.0-flash-fin, meta/muse-spark-1.3, meta/muse-spark-1.3-contributor, ~z-ai/glm-flash-latest
• prime-inference: x-ai/grok-4.6
• vercel-ai-gateway: alibaba/qwen3.8-flash-next, alibaba/qwen3.8-max-0902, anthropic/claude-fable-5.1, google/gemini-3.8-flash, meta/muse-spark-1.3, meta/muse-spark-1.3-contributor, tencent/hy4-preview, xiaomi/mimo-v2.5-pro-ultraspeed, zai/glm-5.3-fast, zai/glm-5.3-promo-50

Removed

• fireworks: accounts/fireworks/models/deepseek-v4-flash
• github-copilot: claude-opus-4.5, claude-opus-4.6, claude-sonnet-4, claude-sonnet-4.5, gemini-3.1-pro-preview, gpt-4.1, gpt-5.2, gpt-5.2-codex
• google: gemini-robotics-er-1.6-preview
• opencode: hy3-free
• openrouter: anthropic/claude-opus-4.7-fast, anthropic/claude-opus-4.8-fast, anthropic/claude-opus-5-fast, arcee-ai/virtuoso-large, kwaipilot/kat-coder-air-v2.5
• vercel-ai-gateway: deepseek/deepseek-v3

Reference fixes (substitution table)

GitHub Copilot dropped 8 models; two of them were referenced in typed getModel("github-copilot", ...) calls (the exact breakage that motivated RES-1269 — the failing union in CI was Copilot's model list). Substitutions, applied only to (provider, id) pairs that were actually removed:

removed (github-copilot)	replacement	where
claude-sonnet-4.5	claude-sonnet-4.6	18 sites across 11 ai test files + 1 expect + 1 handoff pair
claude-sonnet-4.5	claude-haiku-4.5	1 site: the interleaved-thinking beta test in github-copilot-anthropic.test.ts; the beta header is only sent for non-adaptive-thinking models and haiku is the only remaining non-adaptive Claude in the Copilot catalog
gpt-5.2-codex	gpt-5.3-codex	2 sites in tool-call-id-normalization.test.ts + 1 handoff pair

The same ids under other providers (anthropic claude-sonnet-4-5, openai-codex gpt-5.2-codex, openai gpt-5.2, google gemini-3.1-pro-preview, etc.) still exist and their references are untouched. No test added or removed; counts unchanged.

defaultModelPerProvider check

Every entry resolves in the refreshed catalog except zai: glm-5.1, which is missing from the OLD catalog too — a pre-existing gap, already fixed (with a catalog-existence pin) in open #2032. Not duplicated here.

Data-quality notes (report only, for the upcoming validation work)

• huggingface / thinkingmachines/Inkling-Small: maxTokens 1048576 > contextWindow 524288 (likely swapped upstream). Pre-existing — identical values in the old catalog. Only violation of maxTokens <= contextWindow in the 1262 rows; no zero or missing contextWindow/maxTokens fields.

Verification (clean checkout, Prime sandbox, source-only sync + npm ci)

• npm run build (which on current main regenerates the catalog — exact CI mirror): exit 0; regenerated file byte-identical to the committed one.
• npm run check: exit 0.
• packages/ai npm test: 50 files passed / 22 skipped, 0 failures (includes the cross-provider-handoff catalog-existence guard).
• coding-agent model-resolver, model-registry, tree-selector, settings-manager test files: 4/4 passed.

Linear: RES-1269

---

Note

Medium Risk
Large generated catalog diff can break compile-time model unions and runtime routing for users pinned to removed ids; Copilot catalog shrink is the main behavioral change, mitigated by targeted test substitutions.

Overview
Refreshes the committed models.generated.ts catalog (RES-1269) so it matches a live generate-models run: 1262 models net (+41 / −17), with metadata and pricing updates across Bedrock, Anthropic, Google, Fireworks, OpenRouter, Vercel AI Gateway, opencode, and others.

Notable catalog deltas: adds Claude Fable 5.1, Gemini 3.8 Flash, and GLM-5.3 (and related variants) on multiple routes; GitHub Copilot drops eight ids (e.g. claude-sonnet-4.5, gpt-5.2-codex, older Opus/Sonnet/GPT entries) and gains claude-fable-5.1 and gemini-3.8-flash. Several rows are renamed or retuned (Fireworks DeepSeek ids, OpenRouter “fast” Opus removals, Google gemini-robotics-er removal, etc.).

Tests that typed getModel("github-copilot", …) against removed ids now use claude-sonnet-4.6 and gpt-5.3-codex; the Copilot interleaved-thinking mock test switches to claude-haiku-4.5 because it is the remaining non-adaptive Claude in that catalog. A one-line .changes note documents the refresh.

^{Reviewed by Cursor Bugbot for commit 8a7bc2a. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Refresh generated model catalog from live provider sources

• Regenerates models.generated.ts with updated metadata: 41 models added, 17 removed, and existing entries updated for provider, API, capability, pricing, context-window, and cache-cost fields.
• Migrates GitHub Copilot test fixtures from Claude Sonnet 4.5 to Claude Sonnet 4.6 and from GPT-5.2 Codex to GPT-5.3 Codex across context-overflow.test.ts, stream.test.ts, and other test suites.
• Switches the interleaved-thinking beta-header test to Claude Haiku 4.5.
• Risk: removed and renamed model identifiers in models.generated.ts may break callers referencing deprecated IDs.

^{Macroscope summarized 8a7bc2a.}