…ance

Enable a cheap executor model (Haiku/Sonnet) to consult Opus for strategic
guidance via the Anthropic Advisor Tool API (`advisor-tool-2026-03-01`),
achieving near-Opus quality at ~80% lower cost.

- Patch `@ai-sdk/[email protected]` to support advisor tool (7 patch points
  across 4 dist files: `prepareTools`, streaming/non-streaming handlers,
  multi-turn message reconstruction)
- Add `experimental.advisor` config schema with `enabled`, `model`,
  `max_uses` (capped at 10), and `caching` options
- Inject advisor as provider-defined tool for Anthropic models (skipped
  for non-Anthropic providers with warning, skipped for small queries)
- Add advisor system prompt guiding when to call `advisor()`
- Track advisor iteration costs from `usage.iterations[]` at Opus rates
- Add `advisor_call` telemetry event type
- Add 27 unit tests covering tool injection, config validation, result
  parsing (3 subtypes), multi-turn reconstruction, cost tracking, and
  `prepareTools` SDK logic

Co-Authored-By: Claude Opus 4.6 <[email protected]>

- Extract hardcoded server-side tool list to shared `ANTHROPIC_SERVER_SIDE_TOOLS` constant
- Fix `JSON.parse` crash on compacted advisor results in multi-turn reconstruction
- Add defense-in-depth `mcp_` prefix exclusion for server-side tools in message history
- Add `providerExecuted: true` flag for server-side tools in `message-v2.ts`
- Handle `isProviderExecuted` heuristic for server-executed tool results in `processor.ts`
- Migrate `AnthropicAuthPlugin` from npm to internal plugin registration
- SDK patch: full advisor tool support (Zod schemas, `prepareTools`, streaming/non-streaming
  handlers, multi-turn reconstruction)
- Add 41 unit tests covering tool injection, config schema, result parsing, multi-turn
  reconstruction, cost tracking, compaction, `mcp_` prefix exclusion, and `providerExecuted` flag

Co-Authored-By: Claude Opus 4.6 <[email protected]>