SIGN IN SIGN UP

feat(capabilities): snapshot per-model capabilities via /v1/models/{id}

Adds scripts/capabilities/snapshot.ts — a zero-inference-tokens tool that
fetches structured capability info for Claude models (thinking modes, effort
levels, context_management strategies, code_execution, batch, citations,
image/pdf input) from the Anthropic Beta API.

Purpose: replace speculation in src/models.ts with authoritative per-model
ground truth. Drift caught here (e.g. the 2 Phase-1-discovered bugs fixed in
cc0616e) would otherwise only surface as silent plugin misbehavior.

Snapshot policy: re-run on every CC version bump and diff against the previous
snapshot. Output lives at manifests/capabilities/<model>-<date>.json. Default
matrix: Haiku 4.5, Sonnet 4.6, Opus 4.7 (overridable via CAPABILITY_MODELS).

Also commits the 2026-04-17 baseline snapshots for the 3 models currently
available on the OAuth subscription. Expiring items (`effort.xhigh` on
Opus 4.7) are intentionally absent from response bodies — per docs, trust
user/operator + adaptive-thinking docs over API response lag for newly
shipped features.

npm script: `bun run capabilities:snapshot`
V
Vacbo committed
771667221879e7047f1632d94ce5c19cd2c06dcc
Parent: cc0616e