feat: add LongCat provider with LongCat-2.0 model

[mookechee]

Summary

Adds the LongCat provider (Meituan) with the LongCat-2.0 model.

• providers/longcat/provider.toml — OpenAI-compatible provider (@ai-sdk/openai-compatible, api https://api.longcat.chat/openai)
• providers/longcat/models/LongCat-2.0.toml — provider-specific fields (cost, reasoning options, interleaved), inheriting shared model facts from models/meituan/longcat-2.0.toml via base_model
• providers/longcat/logo.svg — official cat icon, monochrome currentColor, centered

Model facts (sourced from official docs)

Field	Value	Source
context	1,000,000	docs (1M context)
output	131,072	docs (max output 128K)
cost input / output / cache_read	0.75 / 2.95 / 0.015 USD per 1M	official USD pricing page (standard rate)
reasoning	toggle (thinking: enabled/disabled)	chat API docs
interleaved	reasoning_content	chat API docs
tool_call	true	changelog ("native tool calling")
modalities	text-only	docs ("text input only")
release_date	2026-06-30	changelog

Self-test

• Schema: bun validate → exit 0, provider generates correctly.
• Model id case: GET /v1/models returns exactly LongCat-2.0; a lowercase id is rejected with 400 Unsupported model — hence the exact-case filename (same convention as MiniMax-M2.toml).
• opencode end-to-end: loaded this branch's generated api.json into opencode (v1.16.2); live completion succeeded and the computed cost matched the listed prices exactly (reasoning billed at the output rate).

Evidence (raw API responses & cost records)

Model id is case-sensitive — GET https://api.longcat.chat/openai/v1/models returns exactly one id:

"LongCat-2.0"

Request with the lowercase id (as recorded by opencode):

HTTP 400 — https://api.longcat.chat/openai/chat/completions
{"error":{"code":"invalid_parameter","message":"Unsupported model (model=longcat-2.0)","type":"invalid_request_error"}}

Cost accounting — successful run after the rename (only change), recorded by opencode:

{"modelID": "LongCat-2.0", "cost": 0.00786295,
 "tokens": {"input": 10421, "output": 4, "reasoning": 12, "cache": {"read": 0, "write": 0}}}

Hand check against this PR's prices: (10421 × $0.75 + (4 + 12) × $2.95) / 1e6 = $0.00786295 — exact match.

Notes: the catalog was generated from this branch with this repo's generateCatalog() and loaded via OPENCODE_MODELS_PATH; cache_read pricing was not exercised (no cache hit); both …/openai and …/openai/v1 base URLs respond 200 — the documented form (…/openai) is used.

Official sources

• Provider endpoint, model id, context/output, text-only input: https://longcat.chat/platform/docs/
• Chat API thinking toggle and reasoning_content: https://longcat.chat/platform/docs/api/chat.html
• Pay-as-you-go pricing: https://longcat.chat/platform/docs/Pricing/LongCat-2.0.html
• Release date, native tool calling, and 1M context: https://longcat.chat/platform/docs/ChangeLog.html

[nathannli]

i think you should use base model syntax since longcat/meituan is the lab for this model

[mookechee]

Good call, thanks! Refactored to use base_model syntax: moved the provider-agnostic facts into models/meituan/longcat-2.0.toml (meituan being the lab) and providers/longcat/models/longcat-2.0.toml now inherits via base_model = "meituan/longcat-2.0", keeping only the provider-specific fields (cost, reasoning_options, interleaved). bun validate passes and the resolved output is identical. This also aligns with how meituan's other longcat models are namespaced.