feat(sync): Cloudflare AI Gateway model catalog sync

[rekram1-node]

Summary

• Add central cloudflare-ai-gateway sync module (bun models:sync cloudflare-ai-gateway / cloudflare group).
• Reuse existing Workers AI sync secrets (CLOUDFLARE_WORKERS_AI_SYNC_ACCOUNT_ID / CLOUDFLARE_WORKERS_AI_SYNC_API_TOKEN) against GET /compat/models on gateway default.
• Hybrid sync: API is authoritative for membership + list prices; capabilities/limits/cache/reasoning_options come from base_model / existing TOMLs / source OpenAI·Anthropic·Workers AI files (endpoint is too thin for full authoring).
• Catalog refresh: +25 models (gpt-5 family, gpt-4.1, o1-pro, deep-research, gemma-4, kimi-k2.7-code, …), −1 (nemotron-3-120b-a12b no longer served), cost/format updates on a few existing entries. 82 → 106 models.

Evidence

• Live fetch with Workers AI sync secrets + gateway default returned 2361 models.
• Managed namespaces only: openai/*, anthropic/* (single-segment), workers-ai/@cf/* (skips whisper/aura-1).
• Docs: https://developers.cloudflare.com/ai-gateway/usage/chat-completion/ (/compat/models), https://developers.cloudflare.com/ai-gateway/usage/rest-api/

Test plan

• bun models:sync cloudflare-ai-gateway (live)
• Second dry-run: 0 created / 0 updated / 0 deleted
• bun validate
• bun models:sync --list-providers includes cloudflare-ai-gateway
• CI hourly matrix picks up the new provider (uses existing CF secrets)

[github-actions]

Now I have all the information I need. Let me compile my findings.

Action items

• [medium] [violation] providers/cloudflare-ai-gateway/models/openai/gpt-5.6.toml:1 - Check: New model files must resolve to a display name that matches their model ID when the source provider carries a name override. Why: The file uses base_model = "openai/gpt-5.6-sol" (whose metadata name = "GPT-5.6 Sol") with no name override, but the model ID from the filename is openai/gpt-5.6. The source providers/openai/models/gpt-5.6.toml declares name = "GPT-5.6" as an explicit override over the gpt-5.6-sol base, but buildNewModel only copies source?.base_model, cost, reasoning_options, interleaved, temperature, and status into the factored values — it never includes source?.name. The generated catalog entry will therefore display "GPT-5.6 Sol" for model ID openai/gpt-5.6, contradicting the OpenAI provider's own naming. Action: Add name: source?.name to the values object in buildNewModel (so source-level name overrides are preserved when source.base_model differs from the gateway model ID), or hand-add name = "GPT-5.6" to this TOML.

• [medium] [possible mistake] providers/cloudflare-ai-gateway/models/anthropic/claude-3.5-sonnet.toml:14-15 - Check: API-driven cost changes for pass-through models should be consistent with the gateway's other entries for the same upstream provider. Why: Input/output changed from 3/15 to 6/30 — exactly 2× Anthropic's known direct price for Claude Sonnet 3.5 v2. Every other Anthropic model in this gateway (claude-sonnet-4, claude-sonnet-4-5, claude-opus-4-6, claude-haiku-4-5, etc.) retained Anthropic's direct prices and was not touched by the sync, making this the sole outlier. The PR body cites a live fetch of 2361 models but does not map this specific doubling to a provider pricing reference. Action: Confirm the gateway /compat/models endpoint actually returns cost_in/cost_out corresponding to $6/$30 per million for anthropic/claude-3.5-sonnet, and cite the provider reference that supports it; otherwise restore 3/15.

• [low] [possible mistake] providers/cloudflare-ai-gateway/models/workers-ai/@cf/zai-org/glm-5.2.toml:8-9 - Check: API-driven cost changes for Workers AI models served through the gateway should be plausible relative to the Workers AI direct price. Why: Input/output dropped from 1.4/4.4 to 0.06/0.4 — a 23×/11× decrease from the Workers AI direct price for the same @cf/zai-org/glm-5.2 model. The newly added @cf/google/gemma-4-26b-a4b-it moves in the opposite direction (0.6/3 vs Workers AI's 0.1/0.3, a 6×/10× increase), so the two workers-ai cost deltas are inconsistent with a uniform gateway markup or discount. Action: Verify the gateway API's cost_in/cost_out for workers-ai/@cf/zai-org/glm-5.2 and confirm the per-token-to-per-million conversion is correct for this entry; cite the source if the drop is intentional.