feat(anthropic): native Claude Pro/Max subscription support

[shuv1337]

Adds Claude Pro/Max as a first-class Anthropic integration method, so a subscription works without an API key and without an external plugin.

What's included

• OAuth method claude-pro-max — PKCE, code exchange, refresh.
• CLAUDE_CODE_OAUTH_TOKEN added to the integration's env method, so a headless host can use claude setup-token. A setup token arrives in the key slot but is an OAuth token — sent as x-api-key it 401s — so isSubscription recognizes the sk-ant-oat prefix and routes it accordingly.
• No token plumbing added. Integration.connection.resolve already refreshes and persists credentials, and ModelResolver already injects the resolved value.

Why the seam is a route transport

A subscription only draws on the plan when the request presents as genuine Claude Code. Otherwise it is accepted and silently billed as pay-as-you-go "extra usage". That needs Bearer auth, Claude Code's headers, a rewritten request body (identity, canonical <env> block, Claude Code tool casing), and the inverse tool-name mapping on the response.

aisdk.hook("sdk") cannot do this. ModelResolver short-circuits the @ai-sdk/anthropic package to the native AnthropicMessages route, so that hook is never invoked for Anthropic — the existing one in AnthropicPlugin was already dead code on this path. This was confirmed by instrumentation: zero hook invocations across every process while the request still went out and 401'd.

So the seam is a wrapped route transport, selected by a guarded branch in ModelResolver — the same place the ChatGPT-plan branch is selected. Shaping the decoded body leaves encoding with the protocol, and framing splits SSE into whole events before the reverse mapping runs, so it needs no boundary buffering.

The branch is guarded on isSubscription. Without a credential of that shape, behaviour is unchanged.

Billing liability (please read)

normalizeEnv reproduces the <env> block Anthropic fuzzy-matches to decide whether a request is really Claude Code. If opencode's prompt format drifts, requests still succeed and merely cost money — a silent failure. isCanonical is the canary: it warns when normalization stops producing the expected shape, and a golden fixture in the tests pins the format. Please don't silence that warning without re-verifying against a real subscription.

Verification

Against a live Claude Pro/Max subscription, on a from-source server with an isolated XDG_CONFIG_HOME so no external plugin could load:

• models catalogued; completion returned from claude-sonnet-5
• tool turn round-tripped: opencode read → Read on the wire → read restored from response frames → correct answer
• CLAUDE_CODE_OAUTH_TOKEN picked up natively as an env connection
• 27 new tests; 1485 pass / 0 fail across packages/core; typecheck clean; oxlint 0 errors on touched files

[blacksmith-sh]

Found 1 test failure on Blacksmith runners:

Failure

Test	View Logs
smoke: session timeline/renders seeded timeline in order while paging through history	View Logs

^{Need help on this PR? Tag @codesmith-bot with what you need.}