@yanlinglabs/winter-provider-catalog 0.0.24 → 0.0.28

package/dist/types.d.ts

@@ -54,6 +54,12 @@ export type ModelStatus = "candidate" | "experimental" | "supported" | "deprecat
  * continuation domain (Anthropic strips thinking across Claude models), so `continuationDomain`
  * lists the models PROVEN to accept this endpoint's native continuation object.
  */
+/** WS-23: how a row takes a per-message effort change -- see `ReasoningCapabilities.perMessageEffort`. */
+export type PerMessageEffortMechanism = {
+    beta: "mid-conversation-output-config-2026-07-01";
+} | {
+    item: "configuration_update";
+};
 export interface ReasoningCapabilities {
     supported: CapabilityEvidence<boolean>;
     /** The model's own effort vocabulary, verbatim. Vocabularies are NOT interchangeable across providers (WS-13 §8.2). */
@@ -95,6 +101,27 @@ export interface ReasoningCapabilities {
     blockBinding?: CapabilityEvidence<{
         beta: "thinking-binding-controls-2026-08-01";
     }>;
+    /**
+     * The model takes a PER-MESSAGE effort change (WS-23, 2026-09-25): a `role: "system"` message with
+     * empty `content` and `output_config.effort` inside `messages`, applied "from the next `user` turn"
+     * while the top-level `output_config.effort` stays fixed, so the cached prefix survives the switch
+     * (https://platform.claude.com/docs/en/build-with-claude/effort#change-effort-mid-conversation-beta).
+     * Documented for Claude Fable 5.1, Mythos 5.1, Opus 5.5 and Opus 5; Claude Fable 5 returns a 400.
+     * ABSENT means a change is a new top-level value, which restarts the cache. `beta` is the header
+     * value, a closed vocabulary (`CATALOG_VOCABULARIES.perMessageEffortBetas`).
+     *
+     * WS-23 (midconv lane): the value names the row's MECHANISM, because two vendors document one:
+     *   - `{ beta }` -- Anthropic's effort-only `system` message above, behind that beta header;
+     *   - `{ item: "configuration_update" }` -- OpenAI's Responses input item
+     *     `{"type":"configuration_update","reasoning":{"effort":…}}`, placed "before the next user
+     *     message", with the top-level `reasoning.effort` kept fixed ("Configuration updates are
+     *     supported by the GPT-6 model family in standard, single-agent mode. They change only reasoning
+     *     effort.", https://developers.openai.com/api/docs/guides/reasoning). A closed vocabulary
+     *     (`CATALOG_VOCABULARIES.perMessageEffortItems`).
+     * The engine keys only on PRESENCE (either mechanism lays out the same markers); each adapter
+     * refuses, typed, a row whose mechanism is not its own.
+     */
+    perMessageEffort?: CapabilityEvidence<PerMessageEffortMechanism>;
 }
 /** List prices, USD per million tokens. R6-H: the ONLY price source Winter has — an unpriced model reports `0` / `costBasis: "unknown"`, never an invented number. */
 export interface ModelPricing {
@@ -203,6 +230,92 @@ export interface WinterModelDescriptor {
     parallelTools?: CapabilityEvidence<boolean>;
     structuredOutput?: CapabilityEvidence<boolean>;
     promptCaching?: CapabilityEvidence<boolean>;
+    /**
+     * WS-23: the model takes tools declared with `defer_loading: true` and expands the `tool_reference`
+     * blocks a client's own tool-search tool returns inside a `tool_result` -- Anthropic's "custom tool
+     * search implementation" (https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool).
+     * A deferred tool stays out of the rendered prompt prefix until it is referenced, so loading one no
+     * longer changes `tools` and no longer invalidates the prompt cache. `true` is the only meaningful
+     * value; absent keeps today's shape (a loaded deferred tool is appended to `tools`).
+     */
+    deferredToolLoading?: CapabilityEvidence<boolean>;
+    /**
+     * WS-23: the model takes a MID-CONVERSATION `role: "system"` message carrying text -- an operator
+     * instruction appended at the point it becomes relevant instead of editing the top-level `system`
+     * field, so the cached prefix survives it
+     * (https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages; GA, no
+     * beta header). `true` is the only meaningful value; absent keeps reminders as user-turn text.
+     */
+    midConversationSystem?: CapabilityEvidence<boolean>;
+    /**
+     * WS-23: the endpoint takes a `prompt_cache_key` that groups a caller's requests for cache routing
+     * (OpenAI Responses: "Use a stable `prompt_cache_key` to optimize cache routing for requests that
+     * share a reusable prefix", https://developers.openai.com/api/docs/guides/prompt-caching). `true` is
+     * the only meaningful value; absent sends no key.
+     */
+    promptCacheKey?: CapabilityEvidence<boolean>;
+    /**
+     * WS-23 (midconv): the model takes MID-CONVERSATION TOOL CHANGES BY REFERENCE -- a `role: "system"`
+     * message carrying `tool_addition` / `tool_removal` blocks that name a tool `tools` declares, so what
+     * the model may call changes without editing `tools` and the cached prefix survives
+     * (https://platform.claude.com/docs/en/build-with-claude/mid-conversation-system-messages). `beta`
+     * is the header value, a closed vocabulary (`CATALOG_VOCABULARIES.midConversationToolChangeBetas`).
+     * Documented for Claude Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5.5, Opus 4.8 and Opus 5 on
+     * the Claude API, Amazon Bedrock and Google Cloud; "Not available on Claude Sonnet 5".
+     */
+    midConversationToolChanges?: CapabilityEvidence<{
+        beta: "mid-conversation-tool-changes-2026-07-01";
+    }>;
+    /**
+     * WS-23 (midconv): the model also takes a tool DEFINED BY VALUE mid-conversation -- a `tool_addition`
+     * whose `tool` is a `tool_definition` -- for a tool unknown at the first request, or a new definition
+     * under an existing name ("The new definition replaces the earlier one from that position onward").
+     * The header covers the reference changes too. Claude API only (same page). `beta` is a closed
+     * vocabulary (`CATALOG_VOCABULARIES.inlineToolDefinitionBetas`).
+     */
+    inlineToolDefinitions?: CapabilityEvidence<{
+        beta: "inline-tools-2026-09-15";
+    }>;
+    /**
+     * WS-23 (midconv): the model takes OpenAI's CLIENT-EXECUTED tool search -- a
+     * `{"type": "tool_search", "execution": "client", description, parameters}` tool, `tool_search_call`
+     * output items the client answers with `tool_search_output` (the loaded definitions, `defer_loading`
+     * kept, optionally grouped in `namespace` entries), so deferred tools never sit in `tools`
+     * (https://developers.openai.com/api/docs/guides/tools-tool-search: "Only gpt-5.4 and later models
+     * support tool_search"). `true` is the only meaningful value.
+     */
+    clientToolSearch?: CapabilityEvidence<boolean>;
+    /**
+     * WS-23 (midconv): the model takes an `additional_tools` developer input item -- tools that "become
+     * available only after that item appears in the input", replayed at their position -- so a tool can be
+     * added mid-conversation without editing `tools` (same page). `true` is the only meaningful value.
+     */
+    additionalToolsItem?: CapabilityEvidence<boolean>;
+    /**
+     * WS-23 (midconv): the model takes `tool_choice: {"type": "allowed_tools", "mode": "auto" | "required",
+     * "tools": [...]}` -- a callable subset of `tools` "but not modify the list of tools you pass in, so you
+     * can maximize savings from prompt caching" (https://developers.openai.com/api/docs/guides/function-calling).
+     * `true` is the only meaningful value.
+     */
+    allowedToolsChoice?: CapabilityEvidence<boolean>;
+    /**
+     * WS-24: the endpoint takes a call to a tool that is NOT in the request's `tools` -- the model called
+     * it from a definition it was shown in a tool result -- and takes that call and its result in the
+     * history of later requests. A fork of a session (whose `tools` is its parent's, frozen so the cached
+     * prefix is shared) relies on it to run a tool it loads itself where the row documents no mechanism
+     * that declares the tool (`deferredToolLoading`, `clientToolSearch`). No vendor documents this, so it is
+     * LIVE-PROBE-PROVEN only (`scripts/probe-fork-undeclared-tool.ts`), never set from a page. `true` is the
+     * only meaningful value; absent keeps such a call refused ("No such tool available").
+     */
+    undeclaredToolCalls?: CapabilityEvidence<boolean>;
+    /**
+     * WS-23 (midconv, live gate): whether the model takes an ASSISTANT PREFILL -- a request whose
+     * `messages` ends with an assistant turn. `false` is the meaningful value: "Prefilling assistant
+     * messages returns a 400 error on Claude Opus 4.6 and later Opus models, including Claude Opus 5.5"
+     * and on Claude Sonnet 5 (https://platform.claude.com/docs/en/models/opus-5-5/migration-guide), and on
+     * Claude Fable 5 / 5.1. An adapter refuses such a request typed instead of sending it. Absent = unknown.
+     */
+    assistantPrefill?: CapabilityEvidence<boolean>;
     reasoning?: ReasoningCapabilities;
     pricing?: CapabilityEvidence<ModelPricing>;
     /** R6-14: set only after the safety corpus passes live. A worker with no configured classifier route serves only when this is true AND `structuredOutput.confidence === "verified"`. */

package/dist/validate.d.ts

@@ -64,6 +64,10 @@ export declare const CATALOG_VOCABULARIES: {
     readonly toolLoopRequirements: readonly ["hard-error", "silent-degradation", "not-required"];
     readonly effortRequestFields: readonly ["output_config.effort"];
     readonly blockBindingBetas: readonly ["thinking-binding-controls-2026-08-01"];
+    readonly perMessageEffortBetas: readonly ["mid-conversation-output-config-2026-07-01"];
+    readonly perMessageEffortItems: readonly ["configuration_update"];
+    readonly midConversationToolChangeBetas: readonly ["mid-conversation-tool-changes-2026-07-01"];
+    readonly inlineToolDefinitionBetas: readonly ["inline-tools-2026-09-15"];
 };
 /**
  * Recursively scans any JSON value for credential material. Exported because Lane X's generator