agentfootprint/dist/types/thinking/AnthropicThinkingHandler.d.ts

The explainable agent framework — backtrack a wrong answer to the exact context that caused it (evidence, not guesses). Built on footprintjs.

/**
 * AnthropicThinkingHandler — normalizes Anthropic's extended-thinking
 * response shape into the framework's `ThinkingBlock[]` contract.
 *
 * Anthropic emits thinking via blocks in `response.content`:
 *
 *   ```ts
 *   { type: 'thinking',          thinking: 'reasoning text', signature: 'opaque-base64' }
 *   { type: 'redacted_thinking', data: 'opaque-encrypted' }
 *   { type: 'text',              text: 'visible content' }
 *   { type: 'tool_use',          id, name, input }
 *   ```
 *
 * The handler filters for `'thinking'` + `'redacted_thinking'` blocks,
 * preserves the `signature` field BYTE-EXACT (Anthropic validates
 * signatures server-side on the next turn — any modification = HTTP 400),
 * and ignores other block types (visible text + tool calls flow through
 * the existing `LLMResponse.content` / `LLMResponse.toolCalls` paths).
 *
 * The same shapes in both thinking modes (budget and adaptive). Two
 * adaptive-mode facts the echo already honours: a turn may carry NO
 * thinking block (the model chose not to think), and a block's `thinking`
 * text may be EMPTY — `display: 'omitted'`, the default on Claude 4.7 and
 * later when the request does not ask to think (`.thinking()` asks for
 * `'summarized'`), and the progress notes some models write between tool
 * calls. An empty block is still complete — its `signature` carries the
 * reasoning — so it is kept and echoed like any other, never dropped.
 *
 * **Critical invariant:** signature pass-through is byte-exact. The
 * handler MUST NOT trim, normalize encoding, JSON-roundtrip, or
 * otherwise touch the signature string. Tests verify this explicitly.
 *
 * **Three input shapes** Anthropic produces (per Phase 4a panel review):
 *   1. Non-streaming response: full `response.content` array
 *   2. Streaming aggregated:   AnthropicProvider accumulates chunks +
 *                              calls handler with the same array shape
 *   3. Bedrock-via-Anthropic:  deferred to Phase 5+ (separate handler
 *                              or extension of this one)
 *
 * Streaming + non-streaming converge on shape #1 because the provider
 * handles the distinction — handler only sees the assembled content
 * array.
 *
 * **`parseChunk` is OPTIONAL** — Phase 3's framework path populates
 * `LLMChunk.thinkingDelta` from inside AnthropicProvider's `stream()`
 * directly, bypassing handler.parseChunk. We still implement it for
 * consumer integrations that want to use the handler on raw Anthropic
 * chunks directly (e.g., custom transports).
 */
import type { ThinkingHandler } from './types.js';
export declare const anthropicThinkingHandler: ThinkingHandler;
//# sourceMappingURL=AnthropicThinkingHandler.d.ts.map