feat: add anthropic web search support

[darknessest]

What

Adds Anthropic as a fourth websearch_cited provider, backed by Anthropic's server-side web_search_20260318 tool, pinned to allowed_callers: ["direct"].

{
  "provider": {
    "anthropic": {
      "options": {
        "websearch_cited": { "model": "claude-sonnet-4-6" }
      }
    }
  }
}

How

src/anthropic.ts follows the existing provider shape (createAnthropicWebsearchClient → WebsearchClient), calls POST /v1/messages directly with fetch, and adds no dependencies.

Auth, both supported by opencode:

• API key — x-api-key + anthropic-version.
• Claude Pro/Max OAuth — Authorization: Bearer, anthropic-beta: oauth-2025-04-20, and the Claude Code system block that these credentials require as the first system entry.

The tool is pinned to allowed_callers: ["direct"] rather than the version default. On web_search_20260209 and later the default is ["code_execution_20260120"] (dynamic filtering), and a caller without code execution access gets HTTP 200 with code_execution_tool_result_error: too_many_requests, web_search_requests: 0, and an uncited answer from model knowledge — a silent failure for a tool whose contract is cited search. Measured on Claude Pro/Max OAuth credentials: two dynamic-filtering attempts failed that way, bracketing a successful direct call in between. Dynamic filtering is therefore not exposed as a config option.

Formatting mirrors the Google provider's output contract:

• Citation markers are derived from web_search_result_location citations, deduplicated by URL, and rendered as [n] before any trailing whitespace so line breaks survive.
• A Sources: list is appended only when at least one citation exists.
• Text blocks emitted before the first web_search_tool_result are the model narrating its intent to search, so they are dropped from the answer.
• A web_search_tool_result_error throws with its error_code, unless usable text was already produced — a later failed search should not discard an answer that is already grounded.
• stop_reason of max_tokens / pause_turn appends a truncation notice rather than silently returning a cut-off answer.

Testing

• bun check clean, bun test:agent 44 pass / 0 fail.
• Formatting tests are driven by a payload captured live from POST https://api.anthropic.com/v1/messages with web_search_20260318 + allowed_callers: ["direct"] (trimmed, provenance recorded at the fixture), plus cases for narration, citations: null, non-web-search citation types, truncation, and search errors. A test asserts the tool definition keeps allowed_callers: ["direct"], so dropping it fails the suite.
• Verified end-to-end against the live API with both claude-sonnet-4-6 and claude-haiku-4-5-20251001; "What is the latest stable version of Bun?" returns a cited answer with a Sources: list.

Note

Independent of, and does not depend on, the OpenAI streaming fix PR; both branch from main. They touch websearch.test.ts in different places, so whichever merges second may need a trivial rebase.