@warlock.js/ai-anthropic 5.30.0 → 5.31.0

package/CHANGELOG.md

@@ -4,6 +4,44 @@ All notable changes to `@warlock.js/ai-anthropic` are documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/). `@warlock.js/*` packages are released in lockstep — every package shares the same version number, so a version below may list only the changes that affected this package.
 
+## 5.31.0 - 2026-10-18
+
+### Upgrading
+
+- Claude models are now priced by default, so `Usage.cost` appears and USD budgets enforce on them without extra configuration. Pass `pricing` to keep contract rates.
+- Responses can be longer, and cost more, when `maxTokens` was never set. Set `maxTokens` on the model or per call to keep a lower cap. For a gateway or proxy id, set `maxOutputTokens` so streaming uses the real limit.
+
+### Added
+
+- Built-in Claude API list prices (`ANTHROPIC_MODEL_PRICING`, `anthropicPricingFor`, `withAnthropicCacheWriteRates`), from Anthropic's pricing page as of 2026-10-17, used when neither the model nor the SDK registry is priced. Dated snapshot ids resolve to their alias. Claude Haiku 5.5 (priced by prompt length) is left unpriced.
+- `complete()` and the streaming `done` chunk attach `usage.cost` when the model is priced.
+- 529 / `overloaded_error` responses and mid-stream overload events throw `ProviderOverloadedError`; every wrapped error carries `status`, `billed` and `retryable`.
+- An adapter contract test verifies that aborting `signal` rejects `complete()` and `stream()` on the model and cancels the HTTP request.
+- `AnthropicModelConfig.maxOutputTokens` overrides the inferred model output limit that sizes the default `max_tokens`.
+- With `streamChunks`, `stream()` yields `reasoning` for `thinking_delta`, `tool-args` for the `input_json_delta` of client `tool_use` blocks, and `usage` after `message_start` and each `message_delta`. `send()` takes the same `streamChunks` transport option.
+- `ai.providerTool({ provider: "anthropic", definition })` entries (`web_search`, `web_fetch`, `code_execution`, `bash`, `text_editor`, `memory`, …) are sent verbatim ahead of the function tools; the prompt-cache breakpoint stays on the last function tool. Other providers' tools are skipped with a warning. The reserved `providerOptions.anthropic.body` keys are unchanged.
+- Typed context management on `providerOptions.anthropic`: `contextManagement.clearToolUses` / `clearThinking` (`clear_tool_uses_20250919` / `clear_thinking_20251015`, beta `context-management-2025-06-27`) and `serverCompaction` (`compact_20260112`, beta `compact-2026-01-12`, trigger of at least 50,000 input tokens). The adapter builds the edits in the order the API requires and adds the beta headers; combining them with `body.context_management` fails with `InvalidRequestError`.
+- `thinkingBlockBinding: "error" | "drop_block"` sends `thinking.block_binding.prefix_mismatch_behavior` with the `thinking-binding-controls-2026-08-01` beta.
+- Preserved thinking after client-side compaction: on Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 an assistant turn marked `compaction: "retained"` is replayed without its thinking blocks, which were produced against the replaced history. `AnthropicModelConfig.preservedThinking` overrides the inference; `inferPreservedThinking(name)` and `ToAnthropicMessagesOptions` are exported.
+- `AnthropicModel.buildRequest()` creates a detached request snapshot for preflight auditing; `send(request)` sends that snapshot as a completion or stream without rebuilding its body.
+- Message, tool and provider-block converters and request types are exported from the package root for hosts using their own transport.
+
+### Changed
+
+- `reasoning: { effort: "none" }` now turns thinking off on Claude Sonnet 5 and Claude Haiku 5.5 with `thinking: { type: "disabled" }` (sent alone, so Haiku 5.5 stays at its default effort, inside the range where `disabled` is accepted). Previously no `thinking` field was sent and both models ran adaptive thinking by default.
+- `reasoning: { effort: "none" }` on Claude Opus 5 now sends adaptive thinking at `output_config.effort: "low"` instead of omitting `thinking` (which ran at the default `high`). Opus 5 is deliberately not sent `disabled`: with thinking off it can write a tool call as plain text that never runs, and leak thinking tags. Claude Opus 5.5, Sonnet 5.5 and Fable 5 / 5.1 are unchanged.
+- **BREAKING** The default `max_tokens` (when no `maxTokens` is set) rose from 4096. Streaming requests use the model's documented max output — 128K for the Claude 5 family and Opus 4.6–4.8 / Sonnet 4.6, 64K for Opus / Sonnet / Haiku 4.5 — and non-streaming requests use 16384, capped by the model's max output (the Anthropic SDK requires streaming for larger non-streaming requests unless the client sets a `timeout`). Unknown ids default to 16384. With thinking on, the default still keeps room for the thinking budget plus 4096 answer tokens, within the model's limit.
+
+### Fixed
+
+- Usage is marked `inputIncludesCache: false` (Anthropic's `input_tokens` excludes cache reads and writes), and 1-hour cache writes are reported in `cacheWrite1hTokens` from `cache_creation.ephemeral_1h_input_tokens` — or, when the breakdown is missing, from a request whose every cache marker asked for `ttl: "1h"`. Cost now bills reads and writes once each, at their own rates.
+- Caller pricing that omits cache-write rates gets Anthropic's documented multipliers (5-minute writes 1.25x input, 1-hour writes 2x) instead of billing writes at nothing.
+- Usage includes server-side compaction: the `compaction` entries of `usage.iterations`, which the top-level token counts leave out, are added to `input`, `output` and the cache counts (completions and streams).
+- `complete()` always builds a non-streaming request, including when an indexed `stream` option is supplied.
+- A request that fails to build in `complete()`, `stream()` or `countTokens()` (a reserved `providerOptions.anthropic.body` key, or a `body` value `structuredClone` cannot copy, such as a function) is logged like a send failure and rejects with an `AIError`; errors that are not already one become an `InvalidRequestError` carrying the original as `cause`. Before, the raw error escaped unlogged. `buildRequest()` itself still throws the raw error.
+- The debug log line for a streaming request reads "Starting streaming call to messages.create" again; non-streaming requests keep "Starting call to messages.create".
+- The provider docs and `setup-anthropic` skill say that `send()` passes `betas` only to the `"beta.messages"` endpoint.
+
 ## 5.30.0 - 2026-10-16
 
 ### Upgrading
@@ -13,16 +51,16 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ### Added
 
-- Responses with thinking, redacted-thinking, compaction, server-tool or cited-text blocks return the whole ordered turn as `ModelResponse.blocks` (and on the streaming `done` chunk); an assistant message carrying Anthropic `blocks` is replayed verbatim (WR-022).
-- `providerOptions.anthropic` (`AnthropicCallOptions`): `betas` routes the call through `client.beta.messages`, `body` merges extra request fields last (`context_management`, `fallbacks`, `inference_geo`, …), `thinkingDisplay` sets `thinking.display` (WR-025).
-- `AnthropicModel.countTokens(messages, options?)` counts input tokens through `messages.countTokens`; `AnthropicModel` is exported and `AnthropicSDK.model()` returns it (WR-025).
+- Responses with thinking, redacted-thinking, compaction, server-tool or cited-text blocks return the whole ordered turn as `ModelResponse.blocks` (and on the streaming `done` chunk); an assistant message carrying Anthropic `blocks` is replayed verbatim.
+- `providerOptions.anthropic` (`AnthropicCallOptions`): `betas` routes the call through `client.beta.messages`, `body` merges extra request fields last (`context_management`, `fallbacks`, `inference_geo`, …), `thinkingDisplay` sets `thinking.display`.
+- `AnthropicModel.countTokens(messages, options?)` counts input tokens through `messages.countTokens`; `AnthropicModel` is exported and `AnthropicSDK.model()` returns it.
 - Per-call `cacheControl.ttl` overrides the model's `promptCacheTtl`.
 
 ### Changed
 
-- Consecutive tool results are sent as one `user` message with several `tool_result` blocks, as Anthropic expects for parallel tool use (WR-024).
-- Claude 5-family thinking rules: Opus 5.5, Sonnet 5.5 and Fable 5 / 5.1 never receive `budget_tokens` or `disabled` (WR-026).
-- `effort: "none"` on Claude Sonnet 5.5 sends `thinking: { type: "between_tools" }` (its only thinking-off form, sent alone with no `output_config.effort` or `display`); it used to omit `thinking`, so the model still thought at its default effort. Opus 5.5 and Fable 5 / 5.1 have no thinking-off form and keep omitting `thinking` (WR-026).
+- Consecutive tool results are sent as one `user` message with several `tool_result` blocks, as Anthropic expects for parallel tool use.
+- Claude 5-family thinking rules: Opus 5.5, Sonnet 5.5 and Fable 5 / 5.1 never receive `budget_tokens` or `disabled`.
+- `effort: "none"` on Claude Sonnet 5.5 sends `thinking: { type: "between_tools" }` (its only thinking-off form, sent alone with no `output_config.effort` or `display`); it used to omit `thinking`, so the model still thought at its default effort. Opus 5.5 and Fable 5 / 5.1 have no thinking-off form and keep omitting `thinking`.
 
 ### Fixed