@chances-ai/engine 30.0.0 → 31.0.0

package/dist/ai/adapters/native-anthropic.js

@@ -0,0 +1,179 @@
+import { native } from "@chances-ai/native";
+import { detectCompat } from "../compat.js";
+import { effortToThinkingBudget } from "../thinking.js";
+import { extractRetryAfterSeconds } from "../retry.js";
+import { AiSdkAdapter, redactStream } from "./ai-sdk.js";
+/**
+ * (7.10 W6) `api="anthropic-messages"` adapter. Prefers the hand-rolled native
+ * client (chances-natives `anthropic` feature) which gives us cache_control
+ * 3-anchor placement, byte-faithful thinking signatures, and structured
+ * status/Retry-After the AI-SDK hides; falls back to `@ai-sdk/anthropic` when no
+ * native client is loaded (the per-fn JS-fallback invariant — the live path
+ * works today regardless of the addon build).
+ *
+ * Both paths emit the SAME `ProviderEvent` sequence (the native parser is the
+ * golden-equivalence target), so the engine is oblivious to which ran.
+ */
+export class NativeAnthropicAdapter {
+    opts;
+    id = "anthropic";
+    fallback;
+    constructor(opts) {
+        this.opts = opts;
+        this.fallback = new AiSdkAdapter({
+            provider: "anthropic",
+            apiKey: opts.apiKey,
+            models: opts.models,
+        });
+    }
+    listModels() {
+        return this.opts.models;
+    }
+    async *stream(req, signal) {
+        if (!native.hasAnthropicClient()) {
+            // No native build loaded → the AI-SDK anthropic path (the JS fallback).
+            yield* this.fallback.stream(req, signal);
+            return;
+        }
+        yield* redactStream(this.streamNative(req, signal), this.opts.apiKey);
+    }
+    /** Bridge the native callback stream to an async iterable. The Rust worker
+     *  pushes ProviderEvent JSON on a libuv-thread callback; we queue + hand them
+     *  out in order, ending on the terminal `done`/`error` or on abort. */
+    async *streamNative(req, signal) {
+        if (signal.aborted)
+            return; // pre-abort: never hit the network
+        const reqJson = JSON.stringify(this.buildNativeRequest(req));
+        const queue = [];
+        let wake = null;
+        let done = false;
+        const push = (ev) => {
+            queue.push(ev);
+            if (ev.type === "done" || ev.type === "error")
+                done = true;
+            wake?.();
+        };
+        const onEvent = (json) => {
+            try {
+                push(this.parseEvent(JSON.parse(json)));
+            }
+            catch {
+                push({ type: "error", message: "native anthropic: malformed event" });
+            }
+        };
+        const handle = native.anthropicStream(reqJson, this.buildNativeOptions(req), onEvent);
+        const onAbort = () => {
+            done = true;
+            handle.abort();
+            wake?.();
+        };
+        signal.addEventListener("abort", onAbort, { once: true });
+        try {
+            for (;;) {
+                while (queue.length > 0) {
+                    const ev = queue.shift();
+                    if (signal.aborted)
+                        return;
+                    yield ev;
+                    if (ev.type === "done" || ev.type === "error")
+                        return;
+                }
+                if (done || signal.aborted)
+                    return;
+                await new Promise((resolve) => {
+                    wake = resolve;
+                });
+                wake = null;
+            }
+        }
+        finally {
+            signal.removeEventListener("abort", onAbort);
+            handle.abort(); // stop the worker if we leave early (consumer break/throw)
+        }
+    }
+    parseEvent(e) {
+        return normalizeNativeError(e);
+    }
+    buildNativeRequest(req) {
+        return buildNativeAnthropicRequest(req, this.opts.models);
+    }
+    buildNativeOptions(req) {
+        const compat = detectCompat("anthropic-messages", {
+            provider: "anthropic",
+            modelId: req.model,
+        });
+        const beta = [];
+        // The extended (1h) cache TTL is a beta; only request it when this model
+        // supports it AND the caller asked for the long retention.
+        if (req.cacheRetention === "long" && compat.supportsLongCacheRetention) {
+            beta.push("extended-cache-ttl-2025-04-11");
+        }
+        return {
+            apiKey: this.opts.apiKey,
+            ...(this.opts.baseURL ? { baseUrl: this.opts.baseURL } : {}),
+            ...(beta.length > 0 ? { beta } : {}),
+            ...(req.sampling?.timeoutMs !== undefined ? { timeoutMs: req.sampling.timeoutMs } : {}),
+        };
+    }
+}
+/**
+ * (7.10 W6) The JSON the Rust `build_request_body` consumes. Pure + exported so
+ * the request shaping is unit-tested without a native build. effort→budget lives
+ * ONCE (here, via `thinking.ts`); the cache compat flags come from the model.
+ * Reasoning and sampling are mutually exclusive on anthropic — when a thinking
+ * budget is present the Rust builder drops temperature/top_p, so we omit them.
+ */
+export function buildNativeAnthropicRequest(req, models) {
+    const descriptor = models.find((m) => m.id === req.model);
+    const compat = detectCompat("anthropic-messages", {
+        provider: "anthropic",
+        modelId: req.model,
+    });
+    const out = {
+        model: req.model,
+        messages: req.messages,
+        tools: req.tools,
+        maxTokens: req.sampling?.maxTokens ?? descriptor?.maxTokens ?? 4096,
+        cacheRetention: req.cacheRetention ?? "short",
+        compat: {
+            longCache: compat.supportsLongCacheRetention ?? false,
+            cacheOnTools: compat.supportsCacheControlOnTools ?? false,
+            allowEmptySignature: compat.allowEmptySignature ?? false,
+        },
+    };
+    if (req.system)
+        out.system = req.system;
+    if (req.reasoning) {
+        out.thinkingBudget = effortToThinkingBudget(req.reasoning);
+    }
+    else {
+        if (req.sampling?.temperature !== undefined)
+            out.temperature = req.sampling.temperature;
+        if (req.sampling?.topP !== undefined)
+            out.topP = req.sampling.topP;
+    }
+    return out;
+}
+/**
+ * (7.10 W6) Map a raw native event object to a {@link ProviderEvent}. Pure +
+ * exported for unit tests. The native `error` event carries `status` +
+ * `retryAfter` (raw header); we normalize the Retry-After (incl. an HTTP-date)
+ * to seconds and fold it into the message so the engine's classifier
+ * (`classifyProviderError`) classifies by the HTTP code AND honors the
+ * server-directed delay — the precise path the AI-SDK hides.
+ */
+export function normalizeNativeError(e) {
+    if (e.type !== "error")
+        return e;
+    let message = typeof e.message === "string" ? e.message : "anthropic error";
+    const retryAfter = e.retryAfter;
+    if (retryAfter !== undefined && !/retry\s*after/i.test(message)) {
+        const secs = extractRetryAfterSeconds({
+            "retry-after": typeof retryAfter === "string" ? retryAfter : String(retryAfter),
+        });
+        if (secs !== undefined)
+            message += ` (retry after ${secs} seconds)`;
+    }
+    return { type: "error", message };
+}
+//# sourceMappingURL=native-anthropic.js.map

package/dist/ai/adapters/native-anthropic.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"native-anthropic.js","sourceRoot":"","sources":["../../../src/ai/adapters/native-anthropic.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,MAAM,EAAE,MAAM,oBAAoB,CAAC;AAQ5C,OAAO,EAAE,YAAY,EAAE,MAAM,cAAc,CAAC;AAC5C,OAAO,EAAE,sBAAsB,EAAE,MAAM,gBAAgB,CAAC;AACxD,OAAO,EAAE,wBAAwB,EAAE,MAAM,aAAa,CAAC;AACvD,OAAO,EAAE,YAAY,EAAE,YAAY,EAAE,MAAM,aAAa,CAAC;AASzD;;;;;;;;;;GAUG;AACH,MAAM,OAAO,sBAAsB;IAIJ;IAHpB,EAAE,GAAG,WAAW,CAAC;IACT,QAAQ,CAAe;IAExC,YAA6B,IAAmC;QAAnC,SAAI,GAAJ,IAAI,CAA+B;QAC9D,IAAI,CAAC,QAAQ,GAAG,IAAI,YAAY,CAAC;YAC/B,QAAQ,EAAE,WAAW;YACrB,MAAM,EAAE,IAAI,CAAC,MAAM;YACnB,MAAM,EAAE,IAAI,CAAC,MAAM;SACpB,CAAC,CAAC;IACL,CAAC;IAED,UAAU;QACR,OAAO,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC;IAC1B,CAAC;IAED,KAAK,CAAC,CAAC,MAAM,CAAC,GAAgB,EAAE,MAAmB;QACjD,IAAI,CAAC,MAAM,CAAC,kBAAkB,EAAE,EAAE,CAAC;YACjC,wEAAwE;YACxE,KAAK,CAAC,CAAC,IAAI,CAAC,QAAQ,CAAC,MAAM,CAAC,GAAG,EAAE,MAAM,CAAC,CAAC;YACzC,OAAO;QACT,CAAC;QACD,KAAK,CAAC,CAAC,YAAY,CAAC,IAAI,CAAC,YAAY,CAAC,GAAG,EAAE,MAAM,CAAC,EAAE,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;IACxE,CAAC;IAED;;2EAEuE;IAC/D,KAAK,CAAC,CAAC,YAAY,CAAC,GAAgB,EAAE,MAAmB;QAC/D,IAAI,MAAM,CAAC,OAAO;YAAE,OAAO,CAAC,mCAAmC;QAE/D,MAAM,OAAO,GAAG,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC,kBAAkB,CAAC,GAAG,CAAC,CAAC,CAAC;QAC7D,MAAM,KAAK,GAAoB,EAAE,CAAC;QAClC,IAAI,IAAI,GAAwB,IAAI,CAAC;QACrC,IAAI,IAAI,GAAG,KAAK,CAAC;QACjB,MAAM,IAAI,GAAG,CAAC,EAAiB,EAAE,EAAE;YACjC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YACf,IAAI,EAAE,CAAC,IAAI,KAAK,MAAM,IAAI,EAAE,CAAC,IAAI,KAAK,OAAO;gBAAE,IAAI,GAAG,IAAI,CAAC;YAC3D,IAAI,EAAE,EAAE,CAAC;QACX,CAAC,CAAC;QACF,MAAM,OAAO,GAAG,CAAC,IAAY,EAAE,EAAE;YAC/B,IAAI,CAAC;gBACH,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,IAAI,CAAC,KAAK,CAAC,IAAI,CAA4B,CAAC,CAAC,CAAC;YACrE,CAAC;YAAC,MAAM,CAAC;gBACP,IAAI,CAAC,EAAE,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,mCAAmC,EAAE,CAAC,CAAC;YACxE,CAAC;QACH,CAAC,CAAC;QAEF,MAAM,MAAM,GAAG,MAAM,CAAC,eAAe,CAAC,OAAO,EAAE,IAAI,CAAC,kBAAkB,CAAC,GAAG,CAAC,EAAE,OAAO,CAAC,CAAC;QACtF,MAAM,OAAO,GAAG,GAAG,EAAE;YACnB,IAAI,GAAG,IAAI,CAAC;YACZ,MAAM,CAAC,KAAK,EAAE,CAAC;YACf,IAAI,EAAE,EAAE,CAAC;QACX,CAAC,CAAC;QACF,MAAM,CAAC,gBAAgB,CAAC,OAAO,EAAE,OAAO,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,CAAC,CAAC;QAC1D,IAAI,CAAC;YACH,SAAS,CAAC;gBACR,OAAO,KAAK,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;oBACxB,MAAM,EAAE,GAAG,KAAK,CAAC,KAAK,EAAG,CAAC;oBAC1B,IAAI,MAAM,CAAC,OAAO;wBAAE,OAAO;oBAC3B,MAAM,EAAE,CAAC;oBACT,IAAI,EAAE,CAAC,IAAI,KAAK,MAAM,IAAI,EAAE,CAAC,IAAI,KAAK,OAAO;wBAAE,OAAO;gBACxD,CAAC;gBACD,IAAI,IAAI,IAAI,MAAM,CAAC,OAAO;oBAAE,OAAO;gBACnC,MAAM,IAAI,OAAO,CAAO,CAAC,OAAO,EAAE,EAAE;oBAClC,IAAI,GAAG,OAAO,CAAC;gBACjB,CAAC,CAAC,CAAC;gBACH,IAAI,GAAG,IAAI,CAAC;YACd,CAAC;QACH,CAAC;gBAAS,CAAC;YACT,MAAM,CAAC,mBAAmB,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC;YAC7C,MAAM,CAAC,KAAK,EAAE,CAAC,CAAC,2DAA2D;QAC7E,CAAC;IACH,CAAC;IAEO,UAAU,CAAC,CAA0B;QAC3C,OAAO,oBAAoB,CAAC,CAAC,CAAC,CAAC;IACjC,CAAC;IAEO,kBAAkB,CAAC,GAAgB;QACzC,OAAO,2BAA2B,CAAC,GAAG,EAAE,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;IAC5D,CAAC;IAEO,kBAAkB,CAAC,GAAgB;QAMzC,MAAM,MAAM,GAAG,YAAY,CAAC,oBAAoB,EAAE;YAChD,QAAQ,EAAE,WAAW;YACrB,OAAO,EAAE,GAAG,CAAC,KAAK;SACnB,CAA4B,CAAC;QAC9B,MAAM,IAAI,GAAa,EAAE,CAAC;QAC1B,yEAAyE;QACzE,2DAA2D;QAC3D,IAAI,GAAG,CAAC,cAAc,KAAK,MAAM,IAAI,MAAM,CAAC,0BAA0B,EAAE,CAAC;YACvE,IAAI,CAAC,IAAI,CAAC,+BAA+B,CAAC,CAAC;QAC7C,CAAC;QACD,OAAO;YACL,MAAM,EAAE,IAAI,CAAC,IAAI,CAAC,MAAM;YACxB,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,OAAO,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,IAAI,CAAC,IAAI,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YAC5D,GAAG,CAAC,IAAI,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;YACpC,GAAG,CAAC,GAAG,CAAC,QAAQ,EAAE,SAAS,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,SAAS,EAAE,GAAG,CAAC,QAAQ,CAAC,SAAS,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC;SACxF,CAAC;IACJ,CAAC;CACF;AAED;;;;;;GAMG;AACH,MAAM,UAAU,2BAA2B,CACzC,GAAgB,EAChB,MAAyB;IAEzB,MAAM,UAAU,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,GAAG,CAAC,KAAK,CAAC,CAAC;IAC1D,MAAM,MAAM,GAAG,YAAY,CAAC,oBAAoB,EAAE;QAChD,QAAQ,EAAE,WAAW;QACrB,OAAO,EAAE,GAAG,CAAC,KAAK;KACnB,CAA4B,CAAC;IAC9B,MAAM,GAAG,GAA4B;QACnC,KAAK,EAAE,GAAG,CAAC,KAAK;QAChB,QAAQ,EAAE,GAAG,CAAC,QAAQ;QACtB,KAAK,EAAE,GAAG,CAAC,KAAK;QAChB,SAAS,EAAE,GAAG,CAAC,QAAQ,EAAE,SAAS,IAAI,UAAU,EAAE,SAAS,IAAI,IAAI;QACnE,cAAc,EAAE,GAAG,CAAC,cAAc,IAAI,OAAO;QAC7C,MAAM,EAAE;YACN,SAAS,EAAE,MAAM,CAAC,0BAA0B,IAAI,KAAK;YACrD,YAAY,EAAE,MAAM,CAAC,2BAA2B,IAAI,KAAK;YACzD,mBAAmB,EAAE,MAAM,CAAC,mBAAmB,IAAI,KAAK;SACzD;KACF,CAAC;IACF,IAAI,GAAG,CAAC,MAAM;QAAE,GAAG,CAAC,MAAM,GAAG,GAAG,CAAC,MAAM,CAAC;IACxC,IAAI,GAAG,CAAC,SAAS,EAAE,CAAC;QAClB,GAAG,CAAC,cAAc,GAAG,sBAAsB,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;IAC7D,CAAC;SAAM,CAAC;QACN,IAAI,GAAG,CAAC,QAAQ,EAAE,WAAW,KAAK,SAAS;YAAE,GAAG,CAAC,WAAW,GAAG,GAAG,CAAC,QAAQ,CAAC,WAAW,CAAC;QACxF,IAAI,GAAG,CAAC,QAAQ,EAAE,IAAI,KAAK,SAAS;YAAE,GAAG,CAAC,IAAI,GAAG,GAAG,CAAC,QAAQ,CAAC,IAAI,CAAC;IACrE,CAAC;IACD,OAAO,GAAG,CAAC;AACb,CAAC;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,oBAAoB,CAAC,CAA0B;IAC7D,IAAI,CAAC,CAAC,IAAI,KAAK,OAAO;QAAE,OAAO,CAA6B,CAAC;IAC7D,IAAI,OAAO,GAAG,OAAO,CAAC,CAAC,OAAO,KAAK,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,iBAAiB,CAAC;IAC5E,MAAM,UAAU,GAAG,CAAC,CAAC,UAAU,CAAC;IAChC,IAAI,UAAU,KAAK,SAAS,IAAI,CAAC,gBAAgB,CAAC,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC;QAChE,MAAM,IAAI,GAAG,wBAAwB,CAAC;YACpC,aAAa,EAAE,OAAO,UAAU,KAAK,QAAQ,CAAC,CAAC,CAAC,UAAU,CAAC,CAAC,CAAC,MAAM,CAAC,UAAU,CAAC;SAChF,CAAC,CAAC;QACH,IAAI,IAAI,KAAK,SAAS;YAAE,OAAO,IAAI,iBAAiB,IAAI,WAAW,CAAC;IACtE,CAAC;IACD,OAAO,EAAE,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,CAAC;AACpC,CAAC"}

package/dist/ai/adapters/openai-compatible.d.ts

@@ -1,4 +1,4 @@
-import type { ChatRequest, ModelDescriptor, ProviderAdapter, ProviderEvent } from "../types.js";
+import type { ChatRequest, Effort, ModelDescriptor, ProviderAdapter, ProviderEvent, ThinkingFormat } from "../types.js";
 export interface OpenAICompatibleAdapterOptions {
     /** Provider id, e.g. "deepseek" or "xai". Becomes the adapter id. */
     provider: string;
@@ -20,4 +20,20 @@ export declare class OpenAICompatibleAdapter implements ProviderAdapter {
     listModels(): ModelDescriptor[];
     stream(req: ChatRequest, signal: AbortSignal): AsyncIterable<ProviderEvent>;
 }
+/**
+ * (7.10 §3.4, codex W3 R2-M1) The request-body fields that enable reasoning for
+ * an openai-completions vendor, keyed by its `thinkingFormat`. Returned object is
+ * spread into the body by the SDK's chat getArgs. `undefined` ⇒ no reasoning
+ * (off, or a dialect that needs no request field — string-thinking emits inline).
+ * Pure + exported so each dialect's wire shape is unit-tested.
+ *
+ * (7.10 TS re-review) `supportsReasoningEffort` GATES the plain-OpenAI dialect:
+ * kimi/moonshot, grok and groq detect as `thinkingFormat:"openai"` but
+ * `detectCompat` flags them `supportsReasoningEffort:false` (they have no
+ * `reasoning_effort` field — reasoning is automatic on their reasoning models).
+ * Without this gate the flag was dead and we sent `reasoning_effort` to vendors
+ * that reject it (6.9 work-vs-silently-dropped). The dedicated dialects
+ * (deepseek/openrouter/zai/qwen) carry their own enable field and are unaffected.
+ */
+export declare function reasoningBody(format: ThinkingFormat | undefined, reasoning: Effort | undefined, supportsReasoningEffort?: boolean): Record<string, unknown> | undefined;
 //# sourceMappingURL=openai-compatible.d.ts.map

package/dist/ai/adapters/openai-compatible.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"openai-compatible.d.ts","sourceRoot":"","sources":["../../../src/ai/adapters/openai-compatible.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EACV,WAAW,EACX,eAAe,EACf,eAAe,EACf,aAAa,EACd,MAAM,aAAa,CAAC;AAIrB,MAAM,WAAW,8BAA8B;IAC7C,qEAAqE;IACrE,QAAQ,EAAE,MAAM,CAAC;IACjB,mDAAmD;IACnD,OAAO,EAAE,MAAM,CAAC;IAChB,MAAM,EAAE,MAAM,CAAC;IACf,MAAM,EAAE,eAAe,EAAE,CAAC;CAC3B;AAED;;;;;GAKG;AACH,qBAAa,uBAAwB,YAAW,eAAe;IAGjD,OAAO,CAAC,QAAQ,CAAC,IAAI;IAFjC,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAC;gBAES,IAAI,EAAE,8BAA8B;IAIjE,UAAU,IAAI,eAAe,EAAE;IAIxB,MAAM,CAAC,GAAG,EAAE,WAAW,EAAE,MAAM,EAAE,WAAW,GAAG,aAAa,CAAC,aAAa,CAAC;CAgCnF"}
+{"version":3,"file":"openai-compatible.d.ts","sourceRoot":"","sources":["../../../src/ai/adapters/openai-compatible.ts"],"names":[],"mappings":"AAGA,OAAO,KAAK,EACV,WAAW,EACX,MAAM,EACN,eAAe,EAEf,eAAe,EACf,aAAa,EACb,cAAc,EACf,MAAM,aAAa,CAAC;AAIrB,MAAM,WAAW,8BAA8B;IAC7C,qEAAqE;IACrE,QAAQ,EAAE,MAAM,CAAC;IACjB,mDAAmD;IACnD,OAAO,EAAE,MAAM,CAAC;IAChB,MAAM,EAAE,MAAM,CAAC;IACf,MAAM,EAAE,eAAe,EAAE,CAAC;CAC3B;AAED;;;;;GAKG;AACH,qBAAa,uBAAwB,YAAW,eAAe;IAGjD,OAAO,CAAC,QAAQ,CAAC,IAAI;IAFjC,QAAQ,CAAC,EAAE,EAAE,MAAM,CAAC;gBAES,IAAI,EAAE,8BAA8B;IAIjE,UAAU,IAAI,eAAe,EAAE;IAIxB,MAAM,CAAC,GAAG,EAAE,WAAW,EAAE,MAAM,EAAE,WAAW,GAAG,aAAa,CAAC,aAAa,CAAC;CAmDnF;AAED;;;;;;;;;;;;;;GAcG;AACH,wBAAgB,aAAa,CAC3B,MAAM,EAAE,cAAc,GAAG,SAAS,EAClC,SAAS,EAAE,MAAM,GAAG,SAAS,EAC7B,uBAAuB,UAAO,GAC7B,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,SAAS,CAqBrC"}

package/dist/ai/adapters/openai-compatible.js

@@ -1,3 +1,6 @@
+import { detectCompat } from "../compat.js";
+import { ThinkFilter } from "../think-filter.js";
+import { mapEffortToOpenAIReasoningEffort } from "../thinking.js";
 import { streamFromAiSdk } from "./ai-sdk-stream.js";
 import { redact, redactStream } from "./ai-sdk.js";
 /**
@@ -36,10 +39,111 @@ export class OpenAICompatibleAdapter {
             yield { type: "error", message: redact(message, this.opts.apiKey) };
             return;
         }
+        // (7.10) Resolve this provider's quirks from the baseURL/id and translate the
+        // requested reasoning effort to THIS vendor's wire dialect (codex W3 R2-M1).
+        // The openai-compatible chat getArgs SPREADS providerOptions[providerName]
+        // raw into the request body, so per-dialect fields (deepseek `thinking`,
+        // zai/glm `enable_thinking`, openrouter `reasoning`) reach the endpoint.
+        const quirks = detectCompat("openai-completions", {
+            provider: this.opts.provider,
+            baseURL: this.opts.baseURL,
+            modelId: req.model,
+        });
+        const body = reasoningBody(quirks.thinkingFormat, req.reasoning, quirks.supportsReasoningEffort);
+        const providerOptions = body ? { [this.opts.provider]: body } : undefined;
         // Redact downstream stream errors too — same defense-in-depth posture as
         // AiSdkAdapter. We pass `apiKey` into a third-party constructor, so any
         // future SDK that echoes the key on throw shouldn't reach the bus.
-        yield* redactStream(streamFromAiSdk(provider(req.model), req, signal), this.opts.apiKey);
+        let stream = streamFromAiSdk(provider(req.model), req, signal, undefined, { providerOptions });
+        // Endpoints that fold reasoning INLINE as `<think>…</think>` in the content
+        // (rather than a separate reasoning field) get the streaming splitter so it
+        // surfaces as thinking, not answer text. Dormant for our declared providers
+        // (they use a reasoning field), available for proxies that don't.
+        if (quirks.thinkingFormat === "string-thinking")
+            stream = applyThinkFilter(stream);
+        yield* redactStream(stream, this.opts.apiKey);
+    }
+}
+/**
+ * (7.10 §3.4, codex W3 R2-M1) The request-body fields that enable reasoning for
+ * an openai-completions vendor, keyed by its `thinkingFormat`. Returned object is
+ * spread into the body by the SDK's chat getArgs. `undefined` ⇒ no reasoning
+ * (off, or a dialect that needs no request field — string-thinking emits inline).
+ * Pure + exported so each dialect's wire shape is unit-tested.
+ *
+ * (7.10 TS re-review) `supportsReasoningEffort` GATES the plain-OpenAI dialect:
+ * kimi/moonshot, grok and groq detect as `thinkingFormat:"openai"` but
+ * `detectCompat` flags them `supportsReasoningEffort:false` (they have no
+ * `reasoning_effort` field — reasoning is automatic on their reasoning models).
+ * Without this gate the flag was dead and we sent `reasoning_effort` to vendors
+ * that reject it (6.9 work-vs-silently-dropped). The dedicated dialects
+ * (deepseek/openrouter/zai/qwen) carry their own enable field and are unaffected.
+ */
+export function reasoningBody(format, reasoning, supportsReasoningEffort = true) {
+    if (!reasoning)
+        return undefined;
+    const effort = mapEffortToOpenAIReasoningEffort(reasoning);
+    switch (format) {
+        case "openai":
+            return supportsReasoningEffort ? { reasoning_effort: effort } : undefined;
+        case "deepseek":
+            // deepseek-reasoner: turn thinking on + the granular effort.
+            return { thinking: { type: "enabled" }, reasoning_effort: effort };
+        case "openrouter":
+            return { reasoning: { effort } };
+        case "qwen":
+        case "zai":
+            // Top-level boolean toggle — no granular effort tier.
+            return { enable_thinking: true };
+        case "string-thinking":
+        case undefined:
+            // Inline-<think> models emit reasoning in content (ThinkFilter handles the
+            // output); they take no request field. Unknown format → send nothing.
+            return undefined;
+    }
+}
+/**
+ * (7.10) Wrap a provider stream so inline `<think>…</think>` in `text-delta`
+ * events is split into the thinking lifecycle. Non-text events pass through.
+ * Opens the block on the first thinking text and closes it when answer content
+ * resumes (or at stream end).
+ */
+async function* applyThinkFilter(source) {
+    const filter = new ThinkFilter();
+    let thinkingOpen = false;
+    function* drain(content, thinking) {
+        if (thinking) {
+            if (!thinkingOpen) {
+                thinkingOpen = true;
+                yield { type: "thinking-start" };
+            }
+            yield { type: "thinking-delta", text: thinking };
+        }
+        if (content) {
+            if (thinkingOpen) {
+                thinkingOpen = false;
+                yield { type: "thinking-end" };
+            }
+            yield { type: "text-delta", text: content };
+        }
+    }
+    for await (const ev of source) {
+        if (ev.type === "text-delta") {
+            const out = filter.push(ev.text);
+            yield* drain(out.content, out.thinking);
+        }
+        else if (ev.type === "done") {
+            const out = filter.finish();
+            yield* drain(out.content, out.thinking);
+            if (thinkingOpen) {
+                thinkingOpen = false;
+                yield { type: "thinking-end" };
+            }
+            yield ev;
+        }
+        else {
+            yield ev;
+        }
     }
 }
 //# sourceMappingURL=openai-compatible.js.map

package/dist/ai/adapters/openai-compatible.js.map

@@ -1 +1 @@
-{"version":3,"file":"openai-compatible.js","sourceRoot":"","sources":["../../../src/ai/adapters/openai-compatible.ts"],"names":[],"mappings":"AAMA,OAAO,EAAE,eAAe,EAAE,MAAM,oBAAoB,CAAC;AACrD,OAAO,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,aAAa,CAAC;AAWnD;;;;;GAKG;AACH,MAAM,OAAO,uBAAuB;IAGL;IAFpB,EAAE,CAAS;IAEpB,YAA6B,IAAoC;QAApC,SAAI,GAAJ,IAAI,CAAgC;QAC/D,IAAI,CAAC,EAAE,GAAG,IAAI,CAAC,QAAQ,CAAC;IAC1B,CAAC;IAED,UAAU;QACR,OAAO,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC;IAC1B,CAAC;IAED,KAAK,CAAC,CAAC,MAAM,CAAC,GAAgB,EAAE,MAAmB;QASjD,MAAM,MAAM,GAAG,CAAC,MAAM,MAAM,CAAC,2BAA2B,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,CAAwB,CAAC;QACpG,MAAM,MAAM,GAAG,MAAM,EAAE,sBAAsB,CAAC;QAC9C,IAAI,CAAC,MAAM,EAAE,CAAC;YACZ,MAAM,EAAE,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,4CAA4C,EAAE,CAAC;YAC/E,OAAO;QACT,CAAC;QACD,IAAI,QAAsC,CAAC;QAC3C,IAAI,CAAC;YACH,QAAQ,GAAG,MAAM,CAAC;gBAChB,IAAI,EAAE,IAAI,CAAC,IAAI,CAAC,QAAQ;gBACxB,OAAO,EAAE,IAAI,CAAC,IAAI,CAAC,OAAO;gBAC1B,MAAM,EAAE,IAAI,CAAC,IAAI,CAAC,MAAM;aACzB,CAAC,CAAC;QACL,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,MAAM,OAAO,GAAG,CAAC,YAAY,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC;YAC3D,MAAM,EAAE,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,CAAC,OAAO,EAAE,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC;YACpE,OAAO;QACT,CAAC;QACD,yEAAyE;QACzE,wEAAwE;QACxE,mEAAmE;QACnE,KAAK,CAAC,CAAC,YAAY,CAAC,eAAe,CAAC,QAAQ,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,GAAG,EAAE,MAAM,CAAC,EAAE,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;IAC3F,CAAC;CACF"}
+{"version":3,"file":"openai-compatible.js","sourceRoot":"","sources":["../../../src/ai/adapters/openai-compatible.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,YAAY,EAAE,MAAM,cAAc,CAAC;AAC5C,OAAO,EAAE,WAAW,EAAE,MAAM,oBAAoB,CAAC;AACjD,OAAO,EAAE,gCAAgC,EAAE,MAAM,gBAAgB,CAAC;AAUlE,OAAO,EAAE,eAAe,EAAE,MAAM,oBAAoB,CAAC;AACrD,OAAO,EAAE,MAAM,EAAE,YAAY,EAAE,MAAM,aAAa,CAAC;AAWnD;;;;;GAKG;AACH,MAAM,OAAO,uBAAuB;IAGL;IAFpB,EAAE,CAAS;IAEpB,YAA6B,IAAoC;QAApC,SAAI,GAAJ,IAAI,CAAgC;QAC/D,IAAI,CAAC,EAAE,GAAG,IAAI,CAAC,QAAQ,CAAC;IAC1B,CAAC;IAED,UAAU;QACR,OAAO,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC;IAC1B,CAAC;IAED,KAAK,CAAC,CAAC,MAAM,CAAC,GAAgB,EAAE,MAAmB;QASjD,MAAM,MAAM,GAAG,CAAC,MAAM,MAAM,CAAC,2BAA2B,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,CAAwB,CAAC;QACpG,MAAM,MAAM,GAAG,MAAM,EAAE,sBAAsB,CAAC;QAC9C,IAAI,CAAC,MAAM,EAAE,CAAC;YACZ,MAAM,EAAE,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,4CAA4C,EAAE,CAAC;YAC/E,OAAO;QACT,CAAC;QACD,IAAI,QAAsC,CAAC;QAC3C,IAAI,CAAC;YACH,QAAQ,GAAG,MAAM,CAAC;gBAChB,IAAI,EAAE,IAAI,CAAC,IAAI,CAAC,QAAQ;gBACxB,OAAO,EAAE,IAAI,CAAC,IAAI,CAAC,OAAO;gBAC1B,MAAM,EAAE,IAAI,CAAC,IAAI,CAAC,MAAM;aACzB,CAAC,CAAC;QACL,CAAC;QAAC,OAAO,CAAC,EAAE,CAAC;YACX,MAAM,OAAO,GAAG,CAAC,YAAY,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC;YAC3D,MAAM,EAAE,IAAI,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,CAAC,OAAO,EAAE,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,EAAE,CAAC;YACpE,OAAO;QACT,CAAC;QACD,8EAA8E;QAC9E,6EAA6E;QAC7E,2EAA2E;QAC3E,yEAAyE;QACzE,yEAAyE;QACzE,MAAM,MAAM,GAAG,YAAY,CAAC,oBAAoB,EAAE;YAChD,QAAQ,EAAE,IAAI,CAAC,IAAI,CAAC,QAAQ;YAC5B,OAAO,EAAE,IAAI,CAAC,IAAI,CAAC,OAAO;YAC1B,OAAO,EAAE,GAAG,CAAC,KAAK;SACnB,CAA4B,CAAC;QAC9B,MAAM,IAAI,GAAG,aAAa,CAAC,MAAM,CAAC,cAAc,EAAE,GAAG,CAAC,SAAS,EAAE,MAAM,CAAC,uBAAuB,CAAC,CAAC;QACjG,MAAM,eAAe,GAAG,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC;QAE1E,yEAAyE;QACzE,wEAAwE;QACxE,mEAAmE;QACnE,IAAI,MAAM,GAAG,eAAe,CAAC,QAAQ,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,GAAG,EAAE,MAAM,EAAE,SAAS,EAAE,EAAE,eAAe,EAAE,CAAC,CAAC;QAC/F,4EAA4E;QAC5E,4EAA4E;QAC5E,4EAA4E;QAC5E,kEAAkE;QAClE,IAAI,MAAM,CAAC,cAAc,KAAK,iBAAiB;YAAE,MAAM,GAAG,gBAAgB,CAAC,MAAM,CAAC,CAAC;QACnF,KAAK,CAAC,CAAC,YAAY,CAAC,MAAM,EAAE,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,CAAC;IAChD,CAAC;CACF;AAED;;;;;;;;;;;;;;GAcG;AACH,MAAM,UAAU,aAAa,CAC3B,MAAkC,EAClC,SAA6B,EAC7B,uBAAuB,GAAG,IAAI;IAE9B,IAAI,CAAC,SAAS;QAAE,OAAO,SAAS,CAAC;IACjC,MAAM,MAAM,GAAG,gCAAgC,CAAC,SAAS,CAAC,CAAC;IAC3D,QAAQ,MAAM,EAAE,CAAC;QACf,KAAK,QAAQ;YACX,OAAO,uBAAuB,CAAC,CAAC,CAAC,EAAE,gBAAgB,EAAE,MAAM,EAAE,CAAC,CAAC,CAAC,SAAS,CAAC;QAC5E,KAAK,UAAU;YACb,6DAA6D;YAC7D,OAAO,EAAE,QAAQ,EAAE,EAAE,IAAI,EAAE,SAAS,EAAE,EAAE,gBAAgB,EAAE,MAAM,EAAE,CAAC;QACrE,KAAK,YAAY;YACf,OAAO,EAAE,SAAS,EAAE,EAAE,MAAM,EAAE,EAAE,CAAC;QACnC,KAAK,MAAM,CAAC;QACZ,KAAK,KAAK;YACR,sDAAsD;YACtD,OAAO,EAAE,eAAe,EAAE,IAAI,EAAE,CAAC;QACnC,KAAK,iBAAiB,CAAC;QACvB,KAAK,SAAS;YACZ,2EAA2E;YAC3E,sEAAsE;YACtE,OAAO,SAAS,CAAC;IACrB,CAAC;AACH,CAAC;AAED;;;;;GAKG;AACH,KAAK,SAAS,CAAC,CAAC,gBAAgB,CAC9B,MAAoC;IAEpC,MAAM,MAAM,GAAG,IAAI,WAAW,EAAE,CAAC;IACjC,IAAI,YAAY,GAAG,KAAK,CAAC;IACzB,QAAQ,CAAC,CAAC,KAAK,CAAC,OAAe,EAAE,QAAgB;QAC/C,IAAI,QAAQ,EAAE,CAAC;YACb,IAAI,CAAC,YAAY,EAAE,CAAC;gBAClB,YAAY,GAAG,IAAI,CAAC;gBACpB,MAAM,EAAE,IAAI,EAAE,gBAAgB,EAAE,CAAC;YACnC,CAAC;YACD,MAAM,EAAE,IAAI,EAAE,gBAAgB,EAAE,IAAI,EAAE,QAAQ,EAAE,CAAC;QACnD,CAAC;QACD,IAAI,OAAO,EAAE,CAAC;YACZ,IAAI,YAAY,EAAE,CAAC;gBACjB,YAAY,GAAG,KAAK,CAAC;gBACrB,MAAM,EAAE,IAAI,EAAE,cAAc,EAAE,CAAC;YACjC,CAAC;YACD,MAAM,EAAE,IAAI,EAAE,YAAY,EAAE,IAAI,EAAE,OAAO,EAAE,CAAC;QAC9C,CAAC;IACH,CAAC;IACD,IAAI,KAAK,EAAE,MAAM,EAAE,IAAI,MAAM,EAAE,CAAC;QAC9B,IAAI,EAAE,CAAC,IAAI,KAAK,YAAY,EAAE,CAAC;YAC7B,MAAM,GAAG,GAAG,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC,IAAI,CAAC,CAAC;YACjC,KAAK,CAAC,CAAC,KAAK,CAAC,GAAG,CAAC,OAAO,EAAE,GAAG,CAAC,QAAQ,CAAC,CAAC;QAC1C,CAAC;aAAM,IAAI,EAAE,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;YAC9B,MAAM,GAAG,GAAG,MAAM,CAAC,MAAM,EAAE,CAAC;YAC5B,KAAK,CAAC,CAAC,KAAK,CAAC,GAAG,CAAC,OAAO,EAAE,GAAG,CAAC,QAAQ,CAAC,CAAC;YACxC,IAAI,YAAY,EAAE,CAAC;gBACjB,YAAY,GAAG,KAAK,CAAC;gBACrB,MAAM,EAAE,IAAI,EAAE,cAAc,EAAE,CAAC;YACjC,CAAC;YACD,MAAM,EAAE,CAAC;QACX,CAAC;aAAM,CAAC;YACN,MAAM,EAAE,CAAC;QACX,CAAC;IACH,CAAC;AACH,CAAC"}

package/dist/ai/compat.d.ts

@@ -0,0 +1,30 @@
+import type { Api, ModelCompat } from "./types.js";
+/**
+ * (7.10 §3.1) Resolve the per-model quirk matrix from the wire protocol +
+ * baseURL + provider id — pi's `detectCompat` pattern (`openai-completions.ts:1075`).
+ * This is the mechanism that makes ONE `openai-completions` path serve every
+ * vendor correctly: the difference between reasoning working vs being silently
+ * dropped on deepseek/glm/kimi/qwen lives here, as data.
+ *
+ * The generated catalog pins `compat` per model; for any model that arrives
+ * WITHOUT one (custom baseURL, pre-catalog rows) the adapter falls back to this
+ * auto-detection. Explicit `model.compat?.X` always wins (`X ?? detected.X`,
+ * see {@link mergeCompat}).
+ *
+ * @param api      wire protocol — only `openai-completions` / `anthropic-messages`
+ *                 carry quirks; everything else returns `{}`.
+ * @param opts     `baseURL` + `provider` drive detection; `modelId` is needed
+ *                 only for the openrouter→anthropic cache-control corner case.
+ */
+export declare function detectCompat(api: Api, opts: {
+    baseURL?: string;
+    provider: string;
+    modelId?: string;
+}): ModelCompat;
+/**
+ * (7.10 §3.1) Field-level merge: an explicit per-model override beats the
+ * auto-detected default (`override ?? detected`), pi `getCompat` semantics. Only
+ * meaningful within one api family — callers pass same-shaped objects.
+ */
+export declare function mergeCompat<T extends ModelCompat>(detected: T, override?: Partial<T>): T;
+//# sourceMappingURL=compat.d.ts.map

package/dist/ai/compat.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"compat.d.ts","sourceRoot":"","sources":["../../src/ai/compat.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAEV,GAAG,EACH,WAAW,EAEZ,MAAM,YAAY,CAAC;AAEpB;;;;;;;;;;;;;;;;GAgBG;AACH,wBAAgB,YAAY,CAC1B,GAAG,EAAE,GAAG,EACR,IAAI,EAAE;IAAE,OAAO,CAAC,EAAE,MAAM,CAAC;IAAC,QAAQ,EAAE,MAAM,CAAC;IAAC,OAAO,CAAC,EAAE,MAAM,CAAA;CAAE,GAC7D,WAAW,CAIb;AAkED;;;;GAIG;AACH,wBAAgB,WAAW,CAAC,CAAC,SAAS,WAAW,EAAE,QAAQ,EAAE,CAAC,EAAE,QAAQ,CAAC,EAAE,OAAO,CAAC,CAAC,CAAC,GAAG,CAAC,CAQxF"}

package/dist/ai/compat.js

@@ -0,0 +1,93 @@
+/**
+ * (7.10 §3.1) Resolve the per-model quirk matrix from the wire protocol +
+ * baseURL + provider id — pi's `detectCompat` pattern (`openai-completions.ts:1075`).
+ * This is the mechanism that makes ONE `openai-completions` path serve every
+ * vendor correctly: the difference between reasoning working vs being silently
+ * dropped on deepseek/glm/kimi/qwen lives here, as data.
+ *
+ * The generated catalog pins `compat` per model; for any model that arrives
+ * WITHOUT one (custom baseURL, pre-catalog rows) the adapter falls back to this
+ * auto-detection. Explicit `model.compat?.X` always wins (`X ?? detected.X`,
+ * see {@link mergeCompat}).
+ *
+ * @param api      wire protocol — only `openai-completions` / `anthropic-messages`
+ *                 carry quirks; everything else returns `{}`.
+ * @param opts     `baseURL` + `provider` drive detection; `modelId` is needed
+ *                 only for the openrouter→anthropic cache-control corner case.
+ */
+export function detectCompat(api, opts) {
+    if (api === "anthropic-messages")
+        return detectAnthropicCompat(opts.provider);
+    if (api === "openai-completions")
+        return detectOpenAICompletionsCompat(opts);
+    return {};
+}
+function detectAnthropicCompat(provider) {
+    // First-party Anthropic supports 1h cache + cache_control on tools. Empty
+    // signatures are NOT allowed (downgrade interrupted thinking to text);
+    // Bedrock/Vertex anthropic relays (out of the declared set today) would set
+    // allowEmptySignature where their replay tolerates it.
+    const firstParty = provider === "anthropic";
+    return {
+        supportsLongCacheRetention: firstParty,
+        supportsCacheControlOnTools: firstParty,
+        allowEmptySignature: false,
+    };
+}
+function detectOpenAICompletionsCompat(opts) {
+    const url = (opts.baseURL ?? "").toLowerCase();
+    const p = opts.provider;
+    const isDeepSeek = p === "deepseek" || url.includes("deepseek.com");
+    // glm / z.ai share the "enable_thinking" (zai) dialect.
+    const isZai = p === "glm" || url.includes("open.bigmodel.cn") || url.includes("api.z.ai");
+    const isMoonshot = p === "kimi" || url.includes("moonshot");
+    const isOpenRouter = p === "openrouter" || url.includes("openrouter.ai");
+    const isGrok = p === "xai" || url.includes("x.ai");
+    const isGroq = p === "groq" || url.includes("groq.com");
+    const isDoubao = p === "doubao" || url.includes("volces.com");
+    const isMiniMax = p === "minimax" || url.includes("minimax");
+    // "Non-standard" OpenAI-compat endpoints: most Chinese vendors + grok/groq
+    // reject the OpenAI-native `store` field and want plain `max_tokens`.
+    const isNonStandard = isDeepSeek || isZai || isMoonshot || isGrok || isGroq || isDoubao || isMiniMax;
+    // enable_thinking vendors (zai/glm + qwen-family) don't take reasoning_effort;
+    // grok/groq/moonshot don't expose effort either.
+    const supportsReasoningEffort = !(isZai || isGrok || isGroq || isMoonshot);
+    const thinkingFormat = isDeepSeek
+        ? "deepseek"
+        : isZai
+            ? "zai"
+            : isOpenRouter
+                ? "openrouter"
+                : "openai";
+    return {
+        thinkingFormat,
+        maxTokensField: isNonStandard ? "max_tokens" : "max_completion_tokens",
+        supportsReasoningEffort,
+        supportsStore: !isNonStandard,
+        // Moonshot/Kimi reject `strict` in tool defs; everyone else accepts it.
+        supportsStrictMode: !isMoonshot,
+        supportsUsageInStreaming: true,
+        supportsDeveloperRole: false,
+        // DeepSeek requires replayed assistant messages to carry reasoning_content.
+        requiresReasoningContentOnAssistantMessages: isDeepSeek,
+        // OpenRouter routing an Anthropic model expects Anthropic cache markers.
+        cacheControlFormat: isOpenRouter && (opts.modelId ?? "").startsWith("anthropic/") ? "anthropic" : undefined,
+    };
+}
+/**
+ * (7.10 §3.1) Field-level merge: an explicit per-model override beats the
+ * auto-detected default (`override ?? detected`), pi `getCompat` semantics. Only
+ * meaningful within one api family — callers pass same-shaped objects.
+ */
+export function mergeCompat(detected, override) {
+    if (!override)
+        return detected;
+    const out = { ...detected };
+    for (const k of Object.keys(override)) {
+        const v = override[k];
+        if (v !== undefined)
+            out[k] = v;
+    }
+    return out;
+}
+//# sourceMappingURL=compat.js.map

package/dist/ai/compat.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"compat.js","sourceRoot":"","sources":["../../src/ai/compat.ts"],"names":[],"mappings":"AAOA;;;;;;;;;;;;;;;;GAgBG;AACH,MAAM,UAAU,YAAY,CAC1B,GAAQ,EACR,IAA8D;IAE9D,IAAI,GAAG,KAAK,oBAAoB;QAAE,OAAO,qBAAqB,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;IAC9E,IAAI,GAAG,KAAK,oBAAoB;QAAE,OAAO,6BAA6B,CAAC,IAAI,CAAC,CAAC;IAC7E,OAAO,EAAE,CAAC;AACZ,CAAC;AAED,SAAS,qBAAqB,CAAC,QAAgB;IAC7C,0EAA0E;IAC1E,uEAAuE;IACvE,4EAA4E;IAC5E,uDAAuD;IACvD,MAAM,UAAU,GAAG,QAAQ,KAAK,WAAW,CAAC;IAC5C,OAAO;QACL,0BAA0B,EAAE,UAAU;QACtC,2BAA2B,EAAE,UAAU;QACvC,mBAAmB,EAAE,KAAK;KAC3B,CAAC;AACJ,CAAC;AAED,SAAS,6BAA6B,CAAC,IAItC;IACC,MAAM,GAAG,GAAG,CAAC,IAAI,CAAC,OAAO,IAAI,EAAE,CAAC,CAAC,WAAW,EAAE,CAAC;IAC/C,MAAM,CAAC,GAAG,IAAI,CAAC,QAAQ,CAAC;IACxB,MAAM,UAAU,GAAG,CAAC,KAAK,UAAU,IAAI,GAAG,CAAC,QAAQ,CAAC,cAAc,CAAC,CAAC;IACpE,wDAAwD;IACxD,MAAM,KAAK,GAAG,CAAC,KAAK,KAAK,IAAI,GAAG,CAAC,QAAQ,CAAC,kBAAkB,CAAC,IAAI,GAAG,CAAC,QAAQ,CAAC,UAAU,CAAC,CAAC;IAC1F,MAAM,UAAU,GAAG,CAAC,KAAK,MAAM,IAAI,GAAG,CAAC,QAAQ,CAAC,UAAU,CAAC,CAAC;IAC5D,MAAM,YAAY,GAAG,CAAC,KAAK,YAAY,IAAI,GAAG,CAAC,QAAQ,CAAC,eAAe,CAAC,CAAC;IACzE,MAAM,MAAM,GAAG,CAAC,KAAK,KAAK,IAAI,GAAG,CAAC,QAAQ,CAAC,MAAM,CAAC,CAAC;IACnD,MAAM,MAAM,GAAG,CAAC,KAAK,MAAM,IAAI,GAAG,CAAC,QAAQ,CAAC,UAAU,CAAC,CAAC;IACxD,MAAM,QAAQ,GAAG,CAAC,KAAK,QAAQ,IAAI,GAAG,CAAC,QAAQ,CAAC,YAAY,CAAC,CAAC;IAC9D,MAAM,SAAS,GAAG,CAAC,KAAK,SAAS,IAAI,GAAG,CAAC,QAAQ,CAAC,SAAS,CAAC,CAAC;IAE7D,2EAA2E;IAC3E,sEAAsE;IACtE,MAAM,aAAa,GACjB,UAAU,IAAI,KAAK,IAAI,UAAU,IAAI,MAAM,IAAI,MAAM,IAAI,QAAQ,IAAI,SAAS,CAAC;IAEjF,+EAA+E;IAC/E,iDAAiD;IACjD,MAAM,uBAAuB,GAAG,CAAC,CAAC,KAAK,IAAI,MAAM,IAAI,MAAM,IAAI,UAAU,CAAC,CAAC;IAE3E,MAAM,cAAc,GAA8C,UAAU;QAC1E,CAAC,CAAC,UAAU;QACZ,CAAC,CAAC,KAAK;YACL,CAAC,CAAC,KAAK;YACP,CAAC,CAAC,YAAY;gBACZ,CAAC,CAAC,YAAY;gBACd,CAAC,CAAC,QAAQ,CAAC;IAEjB,OAAO;QACL,cAAc;QACd,cAAc,EAAE,aAAa,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC,CAAC,uBAAuB;QACtE,uBAAuB;QACvB,aAAa,EAAE,CAAC,aAAa;QAC7B,wEAAwE;QACxE,kBAAkB,EAAE,CAAC,UAAU;QAC/B,wBAAwB,EAAE,IAAI;QAC9B,qBAAqB,EAAE,KAAK;QAC5B,4EAA4E;QAC5E,2CAA2C,EAAE,UAAU;QACvD,yEAAyE;QACzE,kBAAkB,EAChB,YAAY,IAAI,CAAC,IAAI,CAAC,OAAO,IAAI,EAAE,CAAC,CAAC,UAAU,CAAC,YAAY,CAAC,CAAC,CAAC,CAAC,WAAW,CAAC,CAAC,CAAC,SAAS;KAC1F,CAAC;AACJ,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,WAAW,CAAwB,QAAW,EAAE,QAAqB;IACnF,IAAI,CAAC,QAAQ;QAAE,OAAO,QAAQ,CAAC;IAC/B,MAAM,GAAG,GAAG,EAAE,GAAG,QAAQ,EAAO,CAAC;IACjC,KAAK,MAAM,CAAC,IAAI,MAAM,CAAC,IAAI,CAAC,QAAQ,CAAgB,EAAE,CAAC;QACrD,MAAM,CAAC,GAAG,QAAQ,CAAC,CAAC,CAAC,CAAC;QACtB,IAAI,CAAC,KAAK,SAAS;YAAE,GAAG,CAAC,CAAC,CAAC,GAAG,CAAe,CAAC;IAChD,CAAC;IACD,OAAO,GAAG,CAAC;AACb,CAAC"}

package/dist/ai/cost.d.ts

@@ -1,3 +1,12 @@
 import type { ModelDescriptor, Usage } from "./types.js";
+/**
+ * (7.10 §3.3) Cache-aware cost over the 4-bucket {@link Usage}. Each bucket
+ * bills at its own $/1M rate; a cache hit is far cheaper than fresh input and a
+ * cache write slightly dearer. When the catalog omits a cache rate we fall back
+ * to the input rate (a conservative upper bound — never under-bills), so models
+ * without published cache pricing still compute a sane number. This replaces the
+ * old 2-rate formula that billed cached requests as if every token were fresh
+ * input (Axis 6.3 — a real over-count for cache-heavy coding turns).
+ */
 export declare function estimateCost(model: ModelDescriptor, usage: Usage): number;
 //# sourceMappingURL=cost.d.ts.map

package/dist/ai/cost.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"cost.d.ts","sourceRoot":"","sources":["../../src/ai/cost.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,eAAe,EAAE,KAAK,EAAE,MAAM,YAAY,CAAC;AAEzD,wBAAgB,YAAY,CAAC,KAAK,EAAE,eAAe,EAAE,KAAK,EAAE,KAAK,GAAG,MAAM,CAKzE"}
+{"version":3,"file":"cost.d.ts","sourceRoot":"","sources":["../../src/ai/cost.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,eAAe,EAAE,KAAK,EAAE,MAAM,YAAY,CAAC;AAEzD;;;;;;;;GAQG;AACH,wBAAgB,YAAY,CAAC,KAAK,EAAE,eAAe,EAAE,KAAK,EAAE,KAAK,GAAG,MAAM,CASzE"}

package/dist/ai/cost.js

@@ -1,5 +1,18 @@
+/**
+ * (7.10 §3.3) Cache-aware cost over the 4-bucket {@link Usage}. Each bucket
+ * bills at its own $/1M rate; a cache hit is far cheaper than fresh input and a
+ * cache write slightly dearer. When the catalog omits a cache rate we fall back
+ * to the input rate (a conservative upper bound — never under-bills), so models
+ * without published cache pricing still compute a sane number. This replaces the
+ * old 2-rate formula that billed cached requests as if every token were fresh
+ * input (Axis 6.3 — a real over-count for cache-heavy coding turns).
+ */
 export function estimateCost(model, usage) {
-    return ((usage.inputTokens / 1_000_000) * model.inputCostPerMTok +
-        (usage.outputTokens / 1_000_000) * model.outputCostPerMTok);
+    const cacheRead = model.cacheReadCostPerMTok ?? model.inputCostPerMTok;
+    const cacheWrite = model.cacheWriteCostPerMTok ?? model.inputCostPerMTok;
+    return ((usage.input / 1_000_000) * model.inputCostPerMTok +
+        (usage.output / 1_000_000) * model.outputCostPerMTok +
+        (usage.cacheRead / 1_000_000) * cacheRead +
+        (usage.cacheWrite / 1_000_000) * cacheWrite);
 }
 //# sourceMappingURL=cost.js.map

package/dist/ai/cost.js.map

@@ -1 +1 @@
-{"version":3,"file":"cost.js","sourceRoot":"","sources":["../../src/ai/cost.ts"],"names":[],"mappings":"AAEA,MAAM,UAAU,YAAY,CAAC,KAAsB,EAAE,KAAY;IAC/D,OAAO,CACL,CAAC,KAAK,CAAC,WAAW,GAAG,SAAS,CAAC,GAAG,KAAK,CAAC,gBAAgB;QACxD,CAAC,KAAK,CAAC,YAAY,GAAG,SAAS,CAAC,GAAG,KAAK,CAAC,iBAAiB,CAC3D,CAAC;AACJ,CAAC"}
+{"version":3,"file":"cost.js","sourceRoot":"","sources":["../../src/ai/cost.ts"],"names":[],"mappings":"AAEA;;;;;;;;GAQG;AACH,MAAM,UAAU,YAAY,CAAC,KAAsB,EAAE,KAAY;IAC/D,MAAM,SAAS,GAAG,KAAK,CAAC,oBAAoB,IAAI,KAAK,CAAC,gBAAgB,CAAC;IACvE,MAAM,UAAU,GAAG,KAAK,CAAC,qBAAqB,IAAI,KAAK,CAAC,gBAAgB,CAAC;IACzE,OAAO,CACL,CAAC,KAAK,CAAC,KAAK,GAAG,SAAS,CAAC,GAAG,KAAK,CAAC,gBAAgB;QAClD,CAAC,KAAK,CAAC,MAAM,GAAG,SAAS,CAAC,GAAG,KAAK,CAAC,iBAAiB;QACpD,CAAC,KAAK,CAAC,SAAS,GAAG,SAAS,CAAC,GAAG,SAAS;QACzC,CAAC,KAAK,CAAC,UAAU,GAAG,SAAS,CAAC,GAAG,UAAU,CAC5C,CAAC;AACJ,CAAC"}

package/dist/ai/index.d.ts

@@ -1,4 +1,8 @@
-export type { Role, ContentPart, Message, ToolDefinition, ToolCallRequest, Usage, ApiFamily, ModelDescriptor, ChatRequest, ProviderEvent, ProviderAdapter, } from "./types.js";
+export type { Role, ContentPart, Message, ToolDefinition, ToolCallRequest, Usage, Api, Effort, CacheRetention, SamplingOptions, ModelThinkingLevel, ThinkingLevelMap, ThinkingFormat, OpenAICompletionsCompat, AnthropicMessagesCompat, ModelCompat, ModelDescriptor, ChatRequest, ProviderEvent, ProviderAdapter, } from "./types.js";
+export { detectCompat, mergeCompat } from "./compat.js";
+export { clampSampling, SAMPLING_BOUNDS } from "./sampling.js";
+export { THINKING_LEVELS, getSupportedThinkingLevels, clampThinkingLevel, anthropicHasRealXHighEffort, mapEffortToAnthropicAdaptiveEffort, effortToThinkingBudget, mapEffortToOpenAIReasoningEffort, effortToGoogleThinkingBudget, thinkingBlockToProvider, type ThinkingBlock, } from "./thinking.js";
+export { ThinkFilter, splitThinkBlocks } from "./think-filter.js";
 export { ModelRegistry } from "./registry.js";
 export { ModelRouter, type RouteHints, type Route } from "./router.js";
 export { estimateCost } from "./cost.js";
@@ -8,6 +12,6 @@ export { AiSdkAdapter, AI_SDK_PACKAGES, type AiSdkAdapterOptions, type AiSdkProv
 export { OpenAICompatibleAdapter, type OpenAICompatibleAdapterOptions, } from "./adapters/openai-compatible.js";
 export { KNOWN_MODELS, type ProviderEntry } from "./known-models.js";
 export { buildRegistry, envByProviderFrom, type RegistryOptions } from "./setup.js";
-export { classifyProviderError, defaultRetryConfig, type RetryConfig, type RetryDecision, } from "./retry.js";
+export { classifyProviderError, classifyByStatus, extractRetryAfterSeconds, computeRetryDelayMs, defaultRetryConfig, withMaxRetries, type RetryConfig, type RetryDecision, type ProviderErrorKind, type HeadersLike, } from "./retry.js";
 export { summarizeConversation, COMPACT_SUMMARY_SYSTEM } from "./summarizer.js";
 //# sourceMappingURL=index.d.ts.map

package/dist/ai/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/ai/index.ts"],"names":[],"mappings":"AAAA,YAAY,EACV,IAAI,EACJ,WAAW,EACX,OAAO,EACP,cAAc,EACd,eAAe,EACf,KAAK,EACL,SAAS,EACT,eAAe,EACf,WAAW,EACX,aAAa,EACb,eAAe,GAChB,MAAM,YAAY,CAAC;AACpB,OAAO,EAAE,aAAa,EAAE,MAAM,eAAe,CAAC;AAC9C,OAAO,EAAE,WAAW,EAAE,KAAK,UAAU,EAAE,KAAK,KAAK,EAAE,MAAM,aAAa,CAAC;AACvE,OAAO,EAAE,YAAY,EAAE,MAAM,WAAW,CAAC;AACzC,OAAO,EAAE,KAAK,cAAc,EAAE,iBAAiB,EAAE,mBAAmB,EAAE,MAAM,eAAe,CAAC;AAC5F,OAAO,EAAE,WAAW,EAAE,MAAM,oBAAoB,CAAC;AACjD,OAAO,EACL,YAAY,EACZ,eAAe,EACf,KAAK,mBAAmB,EACxB,KAAK,aAAa,GACnB,MAAM,sBAAsB,CAAC;AAC9B,OAAO,EACL,uBAAuB,EACvB,KAAK,8BAA8B,GACpC,MAAM,iCAAiC,CAAC;AACzC,OAAO,EAAE,YAAY,EAAE,KAAK,aAAa,EAAE,MAAM,mBAAmB,CAAC;AACrE,OAAO,EAAE,aAAa,EAAE,iBAAiB,EAAE,KAAK,eAAe,EAAE,MAAM,YAAY,CAAC;AACpF,OAAO,EACL,qBAAqB,EACrB,kBAAkB,EAClB,KAAK,WAAW,EAChB,KAAK,aAAa,GACnB,MAAM,YAAY,CAAC;AACpB,OAAO,EAAE,qBAAqB,EAAE,sBAAsB,EAAE,MAAM,iBAAiB,CAAC"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/ai/index.ts"],"names":[],"mappings":"AAAA,YAAY,EACV,IAAI,EACJ,WAAW,EACX,OAAO,EACP,cAAc,EACd,eAAe,EACf,KAAK,EACL,GAAG,EACH,MAAM,EACN,cAAc,EACd,eAAe,EACf,kBAAkB,EAClB,gBAAgB,EAChB,cAAc,EACd,uBAAuB,EACvB,uBAAuB,EACvB,WAAW,EACX,eAAe,EACf,WAAW,EACX,aAAa,EACb,eAAe,GAChB,MAAM,YAAY,CAAC;AACpB,OAAO,EAAE,YAAY,EAAE,WAAW,EAAE,MAAM,aAAa,CAAC;AACxD,OAAO,EAAE,aAAa,EAAE,eAAe,EAAE,MAAM,eAAe,CAAC;AAC/D,OAAO,EACL,eAAe,EACf,0BAA0B,EAC1B,kBAAkB,EAClB,2BAA2B,EAC3B,kCAAkC,EAClC,sBAAsB,EACtB,gCAAgC,EAChC,4BAA4B,EAC5B,uBAAuB,EACvB,KAAK,aAAa,GACnB,MAAM,eAAe,CAAC;AACvB,OAAO,EAAE,WAAW,EAAE,gBAAgB,EAAE,MAAM,mBAAmB,CAAC;AAClE,OAAO,EAAE,aAAa,EAAE,MAAM,eAAe,CAAC;AAC9C,OAAO,EAAE,WAAW,EAAE,KAAK,UAAU,EAAE,KAAK,KAAK,EAAE,MAAM,aAAa,CAAC;AACvE,OAAO,EAAE,YAAY,EAAE,MAAM,WAAW,CAAC;AACzC,OAAO,EAAE,KAAK,cAAc,EAAE,iBAAiB,EAAE,mBAAmB,EAAE,MAAM,eAAe,CAAC;AAC5F,OAAO,EAAE,WAAW,EAAE,MAAM,oBAAoB,CAAC;AACjD,OAAO,EACL,YAAY,EACZ,eAAe,EACf,KAAK,mBAAmB,EACxB,KAAK,aAAa,GACnB,MAAM,sBAAsB,CAAC;AAC9B,OAAO,EACL,uBAAuB,EACvB,KAAK,8BAA8B,GACpC,MAAM,iCAAiC,CAAC;AACzC,OAAO,EAAE,YAAY,EAAE,KAAK,aAAa,EAAE,MAAM,mBAAmB,CAAC;AACrE,OAAO,EAAE,aAAa,EAAE,iBAAiB,EAAE,KAAK,eAAe,EAAE,MAAM,YAAY,CAAC;AACpF,OAAO,EACL,qBAAqB,EACrB,gBAAgB,EAChB,wBAAwB,EACxB,mBAAmB,EACnB,kBAAkB,EAClB,cAAc,EACd,KAAK,WAAW,EAChB,KAAK,aAAa,EAClB,KAAK,iBAAiB,EACtB,KAAK,WAAW,GACjB,MAAM,YAAY,CAAC;AACpB,OAAO,EAAE,qBAAqB,EAAE,sBAAsB,EAAE,MAAM,iBAAiB,CAAC"}

package/dist/ai/index.js

@@ -1,3 +1,7 @@
+export { detectCompat, mergeCompat } from "./compat.js";
+export { clampSampling, SAMPLING_BOUNDS } from "./sampling.js";
+export { THINKING_LEVELS, getSupportedThinkingLevels, clampThinkingLevel, anthropicHasRealXHighEffort, mapEffortToAnthropicAdaptiveEffort, effortToThinkingBudget, mapEffortToOpenAIReasoningEffort, effortToGoogleThinkingBudget, thinkingBlockToProvider, } from "./thinking.js";
+export { ThinkFilter, splitThinkBlocks } from "./think-filter.js";
 export { ModelRegistry } from "./registry.js";
 export { ModelRouter } from "./router.js";
 export { estimateCost } from "./cost.js";
@@ -7,6 +11,6 @@ export { AiSdkAdapter, AI_SDK_PACKAGES, } from "./adapters/ai-sdk.js";
 export { OpenAICompatibleAdapter, } from "./adapters/openai-compatible.js";
 export { KNOWN_MODELS } from "./known-models.js";
 export { buildRegistry, envByProviderFrom } from "./setup.js";
-export { classifyProviderError, defaultRetryConfig, } from "./retry.js";
+export { classifyProviderError, classifyByStatus, extractRetryAfterSeconds, computeRetryDelayMs, defaultRetryConfig, withMaxRetries, } from "./retry.js";
 export { summarizeConversation, COMPACT_SUMMARY_SYSTEM } from "./summarizer.js";
 //# sourceMappingURL=index.js.map

package/dist/ai/index.js.map

@@ -1 +1 @@
-{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/ai/index.ts"],"names":[],"mappings":"AAaA,OAAO,EAAE,aAAa,EAAE,MAAM,eAAe,CAAC;AAC9C,OAAO,EAAE,WAAW,EAA+B,MAAM,aAAa,CAAC;AACvE,OAAO,EAAE,YAAY,EAAE,MAAM,WAAW,CAAC;AACzC,OAAO,EAAuB,iBAAiB,EAAE,mBAAmB,EAAE,MAAM,eAAe,CAAC;AAC5F,OAAO,EAAE,WAAW,EAAE,MAAM,oBAAoB,CAAC;AACjD,OAAO,EACL,YAAY,EACZ,eAAe,GAGhB,MAAM,sBAAsB,CAAC;AAC9B,OAAO,EACL,uBAAuB,GAExB,MAAM,iCAAiC,CAAC;AACzC,OAAO,EAAE,YAAY,EAAsB,MAAM,mBAAmB,CAAC;AACrE,OAAO,EAAE,aAAa,EAAE,iBAAiB,EAAwB,MAAM,YAAY,CAAC;AACpF,OAAO,EACL,qBAAqB,EACrB,kBAAkB,GAGnB,MAAM,YAAY,CAAC;AACpB,OAAO,EAAE,qBAAqB,EAAE,sBAAsB,EAAE,MAAM,iBAAiB,CAAC"}
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/ai/index.ts"],"names":[],"mappings":"AAsBA,OAAO,EAAE,YAAY,EAAE,WAAW,EAAE,MAAM,aAAa,CAAC;AACxD,OAAO,EAAE,aAAa,EAAE,eAAe,EAAE,MAAM,eAAe,CAAC;AAC/D,OAAO,EACL,eAAe,EACf,0BAA0B,EAC1B,kBAAkB,EAClB,2BAA2B,EAC3B,kCAAkC,EAClC,sBAAsB,EACtB,gCAAgC,EAChC,4BAA4B,EAC5B,uBAAuB,GAExB,MAAM,eAAe,CAAC;AACvB,OAAO,EAAE,WAAW,EAAE,gBAAgB,EAAE,MAAM,mBAAmB,CAAC;AAClE,OAAO,EAAE,aAAa,EAAE,MAAM,eAAe,CAAC;AAC9C,OAAO,EAAE,WAAW,EAA+B,MAAM,aAAa,CAAC;AACvE,OAAO,EAAE,YAAY,EAAE,MAAM,WAAW,CAAC;AACzC,OAAO,EAAuB,iBAAiB,EAAE,mBAAmB,EAAE,MAAM,eAAe,CAAC;AAC5F,OAAO,EAAE,WAAW,EAAE,MAAM,oBAAoB,CAAC;AACjD,OAAO,EACL,YAAY,EACZ,eAAe,GAGhB,MAAM,sBAAsB,CAAC;AAC9B,OAAO,EACL,uBAAuB,GAExB,MAAM,iCAAiC,CAAC;AACzC,OAAO,EAAE,YAAY,EAAsB,MAAM,mBAAmB,CAAC;AACrE,OAAO,EAAE,aAAa,EAAE,iBAAiB,EAAwB,MAAM,YAAY,CAAC;AACpF,OAAO,EACL,qBAAqB,EACrB,gBAAgB,EAChB,wBAAwB,EACxB,mBAAmB,EACnB,kBAAkB,EAClB,cAAc,GAKf,MAAM,YAAY,CAAC;AACpB,OAAO,EAAE,qBAAqB,EAAE,sBAAsB,EAAE,MAAM,iBAAiB,CAAC"}

package/dist/ai/known-models.d.ts

@@ -1,20 +1,34 @@
-import type { ApiFamily, ModelDescriptor } from "./types.js";
+import type { Api, ModelDescriptor } from "./types.js";
 /**
  * Per-provider metadata used by `buildRegistry` to wire the right adapter.
  * `models` is the capability table the registry indexes; `envKey` resolves
- * the API key; `baseURL` is required for the `openai-compatible` family
+ * the API key; `baseURL` is required for the `openai-completions` wire protocol
  * (Chinese providers + xai/groq/openrouter) and ignored for the native ones.
  *
- * Prices are USD per 1M tokens and should be reviewed per release.
- * `supportsTools` is per-model — uneven across the OpenAI-compatible set
- * (especially older MiniMax / Kimi). Verify before flipping the flag.
+ * (7.10) Assembled from {@link PROVIDER_TABLE} (hand-maintained connection
+ * config) + {@link GENERATED_MODELS} (the models.dev catalog, regenerated by
+ * `scripts/generate-models.ts`). Prices are USD per 1M tokens.
  */
 export interface ProviderEntry {
     id: string;
-    api: ApiFamily;
+    api: Api;
     envKey: string;
     baseURL?: string;
     models: ModelDescriptor[];
 }
+/**
+ * (7.10 §3.2) Merge each provider's generated catalog rows with its
+ * hand-maintained `models` (doubao isn't on models.dev) and drop any `exclude`d
+ * ids. `provider` + `api` on every row already match the entry (the generator
+ * stamps them from the table), which the registry consistency guard verifies.
+ *
+ * Model ids must be globally unique (the registry indexes by id, so a collision
+ * is ambiguous routing). A few open-weight ids appear under BOTH a direct host
+ * AND the openrouter aggregator (e.g. groq + openrouter both list
+ * `openai/gpt-oss-120b`). We dedup with **earlier-table-entry-wins**: direct
+ * providers precede openrouter in {@link PROVIDER_TABLE}, so the direct (faster,
+ * cheaper) host keeps the id and the aggregator's duplicate is dropped — and any
+ * future collision resolves deterministically without manual `exclude` upkeep.
+ */
 export declare const KNOWN_MODELS: ProviderEntry[];
 //# sourceMappingURL=known-models.d.ts.map