[yajo]

Problem

GET /claude/v1/models returned only the static model cards shipped with the proxy (DEFAULT_CLAUDE_MODEL_CARDS). New Claude models never showed up there, even though the proxy could serve them: unknown full model IDs pass through unchanged to Anthropic.

Solution

Query Anthropic's Models API with the OAuth token the adapter already manages, and merge the result with the static cards.

• ccproxy/plugins/claude_api/live_models.py (new): fetches https://api.anthropic.com/v1/models with Authorization: Bearer <token> and anthropic-version: 2023-06-01, converts every entry into the OpenAI-compatible model shape (id, created, object, owned_by, plus display_name and line when present) and caches the catalogue for one hour.
• ccproxy/plugins/claude_api/routes.py: /v1/models returns the live catalogue merged with the static cards, so legacy models that Anthropic no longer lists stay available. Live entries win on id collisions.
• ccproxy/plugins/claude_api/adapter.py: adds a public get_access_token() so other modules can reuse the token without reaching into the private resolver.
• ccproxy/plugins/claude_api/config.py: new live_models_enabled flag (default true) to turn the live lookup off.

Failure behaviour

The endpoint never breaks because of the upstream:

• No usable token (not logged in, missing scope) → static cards.
• Anthropic unreachable or returning an error → the last successful catalogue is kept, and the static cards are used if there is none yet.

Test plan

• uv run pytest tests/plugins/claude_api -q - 53 passed
• uv run ruff check and uv run ruff format --check on the touched files
• uv run mypy on the touched files - no issues
• Live check against Anthropic: 13 active models returned, merged with the static cards

Note for maintainers

/v1/models now depends on an upstream call, so it is slightly slower on a cold cache. The one hour cache keeps that cost negligible and the fallback keeps the endpoint usable offline.

Assisted-by: goose + MiniMax-M3.1-Flash