@e2edev/e2e 0.10.0 → 0.11.0

package/dist/agent/model/adapter.js.map

@@ -1 +1 @@
-{"version":3,"file":"adapter.js","sourceRoot":"","sources":["../../../src/agent/model/adapter.ts"],"names":[],"mappings":"AAAA,8BAA8B;AAI9B,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAsEzC;;;;GAIG;AACH,MAAM,OAAO,uBAAwB,SAAQ,UAAU;IAC5C,OAAO,CAAqB;IAErC,YAAY,KAAa,EAAE,OAAO,GAA0C,EAAE;QAC5E,KAAK,CAAC,sBAAsB,EAAE,KAAK,EAAE,OAAO,CAAC,KAAK,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,OAAO,CAAC,KAAK,EAAE,CAAC,CAAC;QAClG,IAAI,CAAC,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC;IACjC,CAAC;CACF;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,eAAe,CAAC,IAAY;IAC1C,OAAO,IAAI,WAAW,EAAE,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,UAAU,CAAC;AACnD,CAAC;AAED,+EAA+E;AAC/E,MAAM,aAAa,GAAG,EAAE,CAAC;AAEzB;;;;;;;GAOG;AACH,MAAM,UAAU,oBAAoB,CAAC,KAAwC;IAC3E,OAAO,CACL,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC,KAAK,CAAC,GAAG,aAAa,CAAC;QACnD,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC,MAAM,CAAC,GAAG,aAAa,CAAC,CACrD,CAAC;AACJ,CAAC"}
+{"version":3,"file":"adapter.js","sourceRoot":"","sources":["../../../src/agent/model/adapter.ts"],"names":[],"mappings":"AAAA,8BAA8B;AAI9B,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAyEzC;;;;GAIG;AACH,MAAM,OAAO,uBAAwB,SAAQ,UAAU;IAC5C,OAAO,CAAqB;IAErC,YAAY,KAAa,EAAE,OAAO,GAA0C,EAAE;QAC5E,KAAK,CAAC,sBAAsB,EAAE,KAAK,EAAE,OAAO,CAAC,KAAK,KAAK,SAAS,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,OAAO,CAAC,KAAK,EAAE,CAAC,CAAC;QAClG,IAAI,CAAC,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC;IACjC,CAAC;CACF;AAED;;;;;;;GAOG;AACH,MAAM,UAAU,eAAe,CAAC,IAAY;IAC1C,OAAO,IAAI,WAAW,EAAE,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,UAAU,CAAC;AACnD,CAAC;AAED,+EAA+E;AAC/E,MAAM,aAAa,GAAG,EAAE,CAAC;AAEzB;;;;;;;GAOG;AACH,MAAM,UAAU,oBAAoB,CAAC,KAAwC;IAC3E,OAAO,CACL,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC,KAAK,CAAC,GAAG,aAAa,CAAC;QACnD,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC,MAAM,CAAC,GAAG,aAAa,CAAC,CACrD,CAAC;AACJ,CAAC"}

package/dist/agent/model/overflow.d.ts

@@ -0,0 +1,20 @@
+/**
+ * Recognizes a context-window overflow in a provider failure. Providers say
+ * it in as many ways as there are providers, and a gateway forwards their
+ * words; the patterns below are the ones seen in the wild (ported from the
+ * pi agent harness, which collected them across twenty providers), plus the
+ * HTTP 413 some of them answer with instead of a message.
+ *
+ * The runner reacts to an overflow differently from any other provider
+ * failure: the request was too big, so a smaller request can still succeed,
+ * where a 5xx or a rate limit needs only time.
+ */
+/**
+ * Whether a provider failure says the request exceeded the model's context
+ * window. Reads the message, the HTTP status, and the raw response body of
+ * the error and of everything it wraps (`cause`, or the last attempt of a
+ * spent retry chain), so a gateway or retry wrapper does not hide the
+ * provider's words.
+ */
+export declare function isContextOverflow(error: unknown): boolean;
+//# sourceMappingURL=overflow.d.ts.map

package/dist/agent/model/overflow.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"overflow.d.ts","sourceRoot":"","sources":["../../../src/agent/model/overflow.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAwCH;;;;;;GAMG;AACH,wBAAgB,iBAAiB,CAAC,KAAK,EAAE,OAAO,GAAG,OAAO,CASzD"}

package/dist/agent/model/overflow.js

@@ -0,0 +1,82 @@
+/**
+ * Recognizes a context-window overflow in a provider failure. Providers say
+ * it in as many ways as there are providers, and a gateway forwards their
+ * words; the patterns below are the ones seen in the wild (ported from the
+ * pi agent harness, which collected them across twenty providers), plus the
+ * HTTP 413 some of them answer with instead of a message.
+ *
+ * The runner reacts to an overflow differently from any other provider
+ * failure: the request was too big, so a smaller request can still succeed,
+ * where a 5xx or a rate limit needs only time.
+ */
+const OVERFLOW_PATTERNS = [
+    /prompt is too long/i, // Anthropic
+    /request_too_large/i, // Anthropic (HTTP 413)
+    /input is too long for requested model/i, // Amazon Bedrock
+    /exceeds the context window/i, // OpenAI
+    /exceeds (?:the )?(?:model'?s )?maximum context length(?: of [\d,]+ tokens?|\s*\([\d,]+\))/i, // OpenAI-compatible proxies
+    /input token count.*exceeds the maximum/i, // Google
+    /maximum prompt length is \d+/i, // xAI
+    /reduce the length of the messages/i, // Groq
+    /maximum context length is \d+ tokens/i, // OpenRouter
+    /exceeds (?:the )?maximum allowed input length of [\d,]+ tokens?/i, // OpenRouter/Poolside
+    /input \(\d+ tokens\) is longer than the model'?s context length \(\d+ tokens\)/i, // Together AI
+    /exceeds the limit of \d+/i, // GitHub Copilot
+    /exceeds the available context size/i, // llama.cpp
+    /greater than the context length/i, // LM Studio
+    /context window exceeds limit/i, // MiniMax
+    /exceeded model token limit/i, // Kimi
+    /too large for model with \d+ maximum context length/i, // Mistral
+    /prompt has [\d,]+ tokens?, but the configured context size is [\d,]+ tokens?/i, // DS4
+    /model_context_window_exceeded/i, // z.ai
+    /prompt too long; exceeded (?:max )?context length/i, // Ollama
+    /range of input length should be/i, // DashScope / Qwen
+    /context[_ ]length[_ ]exceeded/i, // generic
+    /too many tokens/i, // generic
+    /token limit exceeded/i, // generic
+];
+/** Failures that mention tokens without being an overflow: throttling and its cousins. */
+const NON_OVERFLOW_PATTERNS = [
+    /throttl/i, // Amazon Bedrock: "ThrottlingException: Too many tokens, please wait…"
+    /service unavailable/i,
+    /rate limit/i,
+    /too many requests/i,
+];
+/** Longest cause chain walked; a wrapped error rarely nests deeper. */
+const MAX_CAUSE_DEPTH = 8;
+/**
+ * Whether a provider failure says the request exceeded the model's context
+ * window. Reads the message, the HTTP status, and the raw response body of
+ * the error and of everything it wraps (`cause`, or the last attempt of a
+ * spent retry chain), so a gateway or retry wrapper does not hide the
+ * provider's words.
+ */
+export function isContextOverflow(error) {
+    let current = error;
+    for (let depth = 0; depth < MAX_CAUSE_DEPTH && current !== undefined && current !== null; depth += 1) {
+        if (describesOverflow(current))
+            return true;
+        if (typeof current !== 'object')
+            return false;
+        const wrapper = current;
+        current = wrapper.cause ?? wrapper.lastError;
+    }
+    return false;
+}
+/**
+ * One error object, read on its own. HTTP 413 is an overflow whatever the
+ * text says. A text field counts when it matches an overflow pattern and no
+ * throttling pattern; the veto is per field, so a wrapper's "rate limit"
+ * message cannot hide the provider's "prompt is too long" in the body.
+ */
+function describesOverflow(error) {
+    if (typeof error !== 'object' || error === null)
+        return false;
+    const record = error;
+    if ((record.statusCode ?? record.status) === 413)
+        return true;
+    return [record.message, record.responseBody].some((text) => typeof text === 'string' &&
+        OVERFLOW_PATTERNS.some((pattern) => pattern.test(text)) &&
+        !NON_OVERFLOW_PATTERNS.some((pattern) => pattern.test(text)));
+}
+//# sourceMappingURL=overflow.js.map

package/dist/agent/model/overflow.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"overflow.js","sourceRoot":"","sources":["../../../src/agent/model/overflow.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,MAAM,iBAAiB,GAAsB;IAC3C,qBAAqB,EAAE,YAAY;IACnC,oBAAoB,EAAE,uBAAuB;IAC7C,wCAAwC,EAAE,iBAAiB;IAC3D,6BAA6B,EAAE,SAAS;IACxC,4FAA4F,EAAE,4BAA4B;IAC1H,yCAAyC,EAAE,SAAS;IACpD,+BAA+B,EAAE,MAAM;IACvC,oCAAoC,EAAE,OAAO;IAC7C,uCAAuC,EAAE,aAAa;IACtD,kEAAkE,EAAE,sBAAsB;IAC1F,iFAAiF,EAAE,cAAc;IACjG,2BAA2B,EAAE,iBAAiB;IAC9C,qCAAqC,EAAE,YAAY;IACnD,kCAAkC,EAAE,YAAY;IAChD,+BAA+B,EAAE,UAAU;IAC3C,6BAA6B,EAAE,OAAO;IACtC,sDAAsD,EAAE,UAAU;IAClE,+EAA+E,EAAE,MAAM;IACvF,gCAAgC,EAAE,OAAO;IACzC,oDAAoD,EAAE,SAAS;IAC/D,kCAAkC,EAAE,mBAAmB;IACvD,gCAAgC,EAAE,UAAU;IAC5C,kBAAkB,EAAE,UAAU;IAC9B,uBAAuB,EAAE,UAAU;CACpC,CAAC;AAEF,0FAA0F;AAC1F,MAAM,qBAAqB,GAAsB;IAC/C,UAAU,EAAE,uEAAuE;IACnF,sBAAsB;IACtB,aAAa;IACb,oBAAoB;CACrB,CAAC;AAEF,uEAAuE;AACvE,MAAM,eAAe,GAAG,CAAC,CAAC;AAE1B;;;;;;GAMG;AACH,MAAM,UAAU,iBAAiB,CAAC,KAAc;IAC9C,IAAI,OAAO,GAAY,KAAK,CAAC;IAC7B,KAAK,IAAI,KAAK,GAAG,CAAC,EAAE,KAAK,GAAG,eAAe,IAAI,OAAO,KAAK,SAAS,IAAI,OAAO,KAAK,IAAI,EAAE,KAAK,IAAI,CAAC,EAAE,CAAC;QACrG,IAAI,iBAAiB,CAAC,OAAO,CAAC;YAAE,OAAO,IAAI,CAAC;QAC5C,IAAI,OAAO,OAAO,KAAK,QAAQ;YAAE,OAAO,KAAK,CAAC;QAC9C,MAAM,OAAO,GAAG,OAAmD,CAAC;QACpE,OAAO,GAAG,OAAO,CAAC,KAAK,IAAI,OAAO,CAAC,SAAS,CAAC;IAC/C,CAAC;IACD,OAAO,KAAK,CAAC;AACf,CAAC;AAED;;;;;GAKG;AACH,SAAS,iBAAiB,CAAC,KAAc;IACvC,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI;QAAE,OAAO,KAAK,CAAC;IAC9D,MAAM,MAAM,GAAG,KAA8F,CAAC;IAC9G,IAAI,CAAC,MAAM,CAAC,UAAU,IAAI,MAAM,CAAC,MAAM,CAAC,KAAK,GAAG;QAAE,OAAO,IAAI,CAAC;IAC9D,OAAO,CAAC,MAAM,CAAC,OAAO,EAAE,MAAM,CAAC,YAAY,CAAC,CAAC,IAAI,CAC/C,CAAC,IAAI,EAAE,EAAE,CACP,OAAO,IAAI,KAAK,QAAQ;QACxB,iBAAiB,CAAC,IAAI,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACvD,CAAC,qBAAqB,CAAC,IAAI,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,CAC/D,CAAC;AACJ,CAAC"}

package/dist/agent/model/prompt-cache.d.ts

@@ -0,0 +1,47 @@
+/**
+ * Prompt-cache hints: what a request carries so the provider can serve the
+ * repeated prefix (system prompt, tool definitions, the conversation so far)
+ * from its cache instead of reading it again at full price.
+ *
+ * Anthropic caches only up to an explicit breakpoint, so the system prompt
+ * carries one (it covers the tool definitions ahead of it) and the newest
+ * message carries the other: the next turn's request then matches everything
+ * up to it. OpenAI caches prefixes on its own and takes a routing key; one
+ * key per system prompt sends every call of a run to the same cache. Other
+ * providers get nothing extra, and the request is exactly what it was.
+ */
+import type { ModelMessage, SystemModelMessage } from 'ai';
+import type { ProviderOptions } from '../../types.ts';
+export type CacheFamily = 'anthropic' | 'openai';
+/** What the hints need to know about the model: the provider and model id the SDK reports. */
+export interface CacheModelRef {
+    readonly provider?: string | undefined;
+    readonly modelId?: string | undefined;
+}
+/**
+ * The provider family a model routes to. A gateway model names its upstream
+ * in the model id (`anthropic/claude-…`); a direct provider names it in the
+ * provider (`anthropic.messages`, `openai.responses`).
+ */
+export declare function cacheFamily(model: CacheModelRef | undefined): CacheFamily | undefined;
+export interface PromptCacheHints {
+    readonly family: CacheFamily | undefined;
+    /** The system prompt as the request's instructions, carrying a breakpoint where the provider needs one. */
+    instructions(system: string): string | SystemModelMessage;
+    /**
+     * Request-level provider options: the caller's own, plus a prompt-cache
+     * routing key where the provider takes one. The caller's values win.
+     */
+    providerOptions(base: ProviderOptions | undefined, system: string): ProviderOptions | undefined;
+    /**
+     * Moves the conversation breakpoint to the newest message. Every other
+     * message loses its breakpoint, so a history that carries forward across
+     * turns never accumulates more than the provider allows.
+     */
+    markLatest(messages: ModelMessage[]): ModelMessage[];
+}
+/** The hints for one model; a no-op for providers without a cache the request can address. */
+export declare function promptCacheHints(model: CacheModelRef | undefined): PromptCacheHints;
+/** A stable key for one system prompt, so every call sharing the prefix routes alike. */
+export declare function promptCacheKey(system: string): string;
+//# sourceMappingURL=prompt-cache.d.ts.map

package/dist/agent/model/prompt-cache.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"prompt-cache.d.ts","sourceRoot":"","sources":["../../../src/agent/model/prompt-cache.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAGH,OAAO,KAAK,EAAE,YAAY,EAAE,kBAAkB,EAAE,MAAM,IAAI,CAAC;AAC3D,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,gBAAgB,CAAC;AAEtD,MAAM,MAAM,WAAW,GAAG,WAAW,GAAG,QAAQ,CAAC;AAEjD,8FAA8F;AAC9F,MAAM,WAAW,aAAa;IAC5B,QAAQ,CAAC,QAAQ,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;IACvC,QAAQ,CAAC,OAAO,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;CACvC;AAED;;;;GAIG;AACH,wBAAgB,WAAW,CAAC,KAAK,EAAE,aAAa,GAAG,SAAS,GAAG,WAAW,GAAG,SAAS,CAMrF;AAID,MAAM,WAAW,gBAAgB;IAC/B,QAAQ,CAAC,MAAM,EAAE,WAAW,GAAG,SAAS,CAAC;IACzC,2GAA2G;IAC3G,YAAY,CAAC,MAAM,EAAE,MAAM,GAAG,MAAM,GAAG,kBAAkB,CAAC;IAC1D;;;OAGG;IACH,eAAe,CAAC,IAAI,EAAE,eAAe,GAAG,SAAS,EAAE,MAAM,EAAE,MAAM,GAAG,eAAe,GAAG,SAAS,CAAC;IAChG;;;;OAIG;IACH,UAAU,CAAC,QAAQ,EAAE,YAAY,EAAE,GAAG,YAAY,EAAE,CAAC;CACtD;AAED,8FAA8F;AAC9F,wBAAgB,gBAAgB,CAAC,KAAK,EAAE,aAAa,GAAG,SAAS,GAAG,gBAAgB,CAiBnF;AAED,yFAAyF;AACzF,wBAAgB,cAAc,CAAC,MAAM,EAAE,MAAM,GAAG,MAAM,CAErD"}

package/dist/agent/model/prompt-cache.js

@@ -0,0 +1,78 @@
+/**
+ * Prompt-cache hints: what a request carries so the provider can serve the
+ * repeated prefix (system prompt, tool definitions, the conversation so far)
+ * from its cache instead of reading it again at full price.
+ *
+ * Anthropic caches only up to an explicit breakpoint, so the system prompt
+ * carries one (it covers the tool definitions ahead of it) and the newest
+ * message carries the other: the next turn's request then matches everything
+ * up to it. OpenAI caches prefixes on its own and takes a routing key; one
+ * key per system prompt sends every call of a run to the same cache. Other
+ * providers get nothing extra, and the request is exactly what it was.
+ */
+import { createHash } from 'node:crypto';
+/**
+ * The provider family a model routes to. A gateway model names its upstream
+ * in the model id (`anthropic/claude-…`); a direct provider names it in the
+ * provider (`anthropic.messages`, `openai.responses`).
+ */
+export function cacheFamily(model) {
+    const provider = (model?.provider ?? '').toLowerCase();
+    const modelId = (model?.modelId ?? '').toLowerCase();
+    if (modelId.startsWith('anthropic/') || provider.startsWith('anthropic'))
+        return 'anthropic';
+    if (modelId.startsWith('openai/') || provider.startsWith('openai'))
+        return 'openai';
+    return undefined;
+}
+const ANTHROPIC_BREAKPOINT = { cacheControl: { type: 'ephemeral' } };
+/** The hints for one model; a no-op for providers without a cache the request can address. */
+export function promptCacheHints(model) {
+    const family = cacheFamily(model);
+    return {
+        family,
+        instructions: (system) => family === 'anthropic'
+            ? { role: 'system', content: system, providerOptions: { anthropic: ANTHROPIC_BREAKPOINT } }
+            : system,
+        providerOptions: (base, system) => {
+            if (family !== 'openai')
+                return base;
+            return {
+                ...base,
+                openai: { promptCacheKey: promptCacheKey(system), ...base?.['openai'] },
+            };
+        },
+        markLatest: (messages) => (family === 'anthropic' ? moveBreakpoint(messages) : messages),
+    };
+}
+/** A stable key for one system prompt, so every call sharing the prefix routes alike. */
+export function promptCacheKey(system) {
+    return `e2e-${createHash('sha256').update(system).digest('hex').slice(0, 16)}`;
+}
+function moveBreakpoint(messages) {
+    if (messages.length === 0)
+        return messages;
+    const last = messages.length - 1;
+    return messages.map((message, index) => (index === last ? withBreakpoint(message) : withoutBreakpoint(message)));
+}
+function withBreakpoint(message) {
+    const anthropic = message.providerOptions?.['anthropic'];
+    return {
+        ...message,
+        providerOptions: { ...message.providerOptions, anthropic: { ...anthropic, ...ANTHROPIC_BREAKPOINT } },
+    };
+}
+function withoutBreakpoint(message) {
+    const anthropic = message.providerOptions?.['anthropic'];
+    if (anthropic === undefined || !('cacheControl' in anthropic))
+        return message;
+    const { cacheControl: _dropped, ...rest } = anthropic;
+    const { anthropic: _anthropic, ...others } = message.providerOptions ?? {};
+    const providerOptions = Object.keys(rest).length === 0 ? others : { ...others, anthropic: rest };
+    if (Object.keys(providerOptions).length === 0) {
+        const { providerOptions: _none, ...bare } = message;
+        return bare;
+    }
+    return { ...message, providerOptions };
+}
+//# sourceMappingURL=prompt-cache.js.map

package/dist/agent/model/prompt-cache.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"prompt-cache.js","sourceRoot":"","sources":["../../../src/agent/model/prompt-cache.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAEH,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAYzC;;;;GAIG;AACH,MAAM,UAAU,WAAW,CAAC,KAAgC;IAC1D,MAAM,QAAQ,GAAG,CAAC,KAAK,EAAE,QAAQ,IAAI,EAAE,CAAC,CAAC,WAAW,EAAE,CAAC;IACvD,MAAM,OAAO,GAAG,CAAC,KAAK,EAAE,OAAO,IAAI,EAAE,CAAC,CAAC,WAAW,EAAE,CAAC;IACrD,IAAI,OAAO,CAAC,UAAU,CAAC,YAAY,CAAC,IAAI,QAAQ,CAAC,UAAU,CAAC,WAAW,CAAC;QAAE,OAAO,WAAW,CAAC;IAC7F,IAAI,OAAO,CAAC,UAAU,CAAC,SAAS,CAAC,IAAI,QAAQ,CAAC,UAAU,CAAC,QAAQ,CAAC;QAAE,OAAO,QAAQ,CAAC;IACpF,OAAO,SAAS,CAAC;AACnB,CAAC;AAED,MAAM,oBAAoB,GAAG,EAAE,YAAY,EAAE,EAAE,IAAI,EAAE,WAAW,EAAE,EAAW,CAAC;AAmB9E,8FAA8F;AAC9F,MAAM,UAAU,gBAAgB,CAAC,KAAgC;IAC/D,MAAM,MAAM,GAAG,WAAW,CAAC,KAAK,CAAC,CAAC;IAClC,OAAO;QACL,MAAM;QACN,YAAY,EAAE,CAAC,MAAM,EAAE,EAAE,CACvB,MAAM,KAAK,WAAW;YACpB,CAAC,CAAC,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,MAAM,EAAE,eAAe,EAAE,EAAE,SAAS,EAAE,oBAAoB,EAAE,EAAE;YAC3F,CAAC,CAAC,MAAM;QACZ,eAAe,EAAE,CAAC,IAAI,EAAE,MAAM,EAAE,EAAE;YAChC,IAAI,MAAM,KAAK,QAAQ;gBAAE,OAAO,IAAI,CAAC;YACrC,OAAO;gBACL,GAAG,IAAI;gBACP,MAAM,EAAE,EAAE,cAAc,EAAE,cAAc,CAAC,MAAM,CAAC,EAAE,GAAG,IAAI,EAAE,CAAC,QAAQ,CAAC,EAAE;aACxE,CAAC;QACJ,CAAC;QACD,UAAU,EAAE,CAAC,QAAQ,EAAE,EAAE,CAAC,CAAC,MAAM,KAAK,WAAW,CAAC,CAAC,CAAC,cAAc,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC;KACzF,CAAC;AACJ,CAAC;AAED,yFAAyF;AACzF,MAAM,UAAU,cAAc,CAAC,MAAc;IAC3C,OAAO,OAAO,UAAU,CAAC,QAAQ,CAAC,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,EAAE,CAAC,EAAE,CAAC;AACjF,CAAC;AAED,SAAS,cAAc,CAAC,QAAwB;IAC9C,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,QAAQ,CAAC;IAC3C,MAAM,IAAI,GAAG,QAAQ,CAAC,MAAM,GAAG,CAAC,CAAC;IACjC,OAAO,QAAQ,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,KAAK,EAAE,EAAE,CAAC,CAAC,KAAK,KAAK,IAAI,CAAC,CAAC,CAAC,cAAc,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,iBAAiB,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC;AACnH,CAAC;AAED,SAAS,cAAc,CAAC,OAAqB;IAC3C,MAAM,SAAS,GAAG,OAAO,CAAC,eAAe,EAAE,CAAC,WAAW,CAAC,CAAC;IACzD,OAAO;QACL,GAAG,OAAO;QACV,eAAe,EAAE,EAAE,GAAG,OAAO,CAAC,eAAe,EAAE,SAAS,EAAE,EAAE,GAAG,SAAS,EAAE,GAAG,oBAAoB,EAAE,EAAE;KACtG,CAAC;AACJ,CAAC;AAED,SAAS,iBAAiB,CAAC,OAAqB;IAC9C,MAAM,SAAS,GAAG,OAAO,CAAC,eAAe,EAAE,CAAC,WAAW,CAAC,CAAC;IACzD,IAAI,SAAS,KAAK,SAAS,IAAI,CAAC,CAAC,cAAc,IAAI,SAAS,CAAC;QAAE,OAAO,OAAO,CAAC;IAC9E,MAAM,EAAE,YAAY,EAAE,QAAQ,EAAE,GAAG,IAAI,EAAE,GAAG,SAAS,CAAC;IACtD,MAAM,EAAE,SAAS,EAAE,UAAU,EAAE,GAAG,MAAM,EAAE,GAAG,OAAO,CAAC,eAAe,IAAI,EAAE,CAAC;IAC3E,MAAM,eAAe,GAAG,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC,MAAM,KAAK,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,GAAG,MAAM,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC;IACjG,IAAI,MAAM,CAAC,IAAI,CAAC,eAAe,CAAC,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QAC9C,MAAM,EAAE,eAAe,EAAE,KAAK,EAAE,GAAG,IAAI,EAAE,GAAG,OAAO,CAAC;QACpD,OAAO,IAAoB,CAAC;IAC9B,CAAC;IACD,OAAO,EAAE,GAAG,OAAO,EAAE,eAAe,EAAE,CAAC;AACzC,CAAC"}

package/dist/agent/model/sdk.d.ts

@@ -1,12 +1,12 @@
 /**
- * AI SDK adapter. One implementation serves every
- * provider: a `provider/model-id` reference resolves through the AI Gateway,
- * and a caller-supplied AI SDK model instance (`openai('gpt-4o')`, a local
- * provider, a scripted test model) is used directly. Everything after model
- * construction — bounded requests, closed-grammar validation, usage and error
- * translation — is provider-independent.
+ * AI SDK adapter. One implementation serves every provider: the
+ * caller-supplied AI SDK model instance (`gateway('openai/gpt-5.4-mini')`,
+ * `openrouter(...)`, `openai('gpt-4o')`, a local provider, a scripted test
+ * model) is used as is. Everything after the model — bounded requests,
+ * closed-grammar validation, usage and error translation — is
+ * provider-independent.
  */
-import { type ResolvedModel, type SdkLanguageModel } from '../../config/agent.ts';
+import type { ResolvedModel, SdkLanguageModel } from '../../config/agent.ts';
 import { type ModelAdapter } from './adapter.ts';
 /**
  * Provider transport retries; distinct from runner-owned model-call budget.
@@ -23,30 +23,47 @@ export declare const TRANSPORT_RETRIES = 5;
 /** Creates the adapter for one resolved model, or fails with MODEL_UNAVAILABLE. */
 export declare function createModelAdapter(model: ResolvedModel | undefined): ModelAdapter;
 /**
- * Builds the bare AI SDK language model for one resolved model reference, for
- * callers that drive the SDK directly (the default step executor) rather than
- * through the adapter. Fails with MODEL_UNAVAILABLE exactly like the adapter.
- * Gateway references require the AI SDK to be loaded (loadAiSdk) first; model
- * instances never touch it.
+ * The bare AI SDK language model of one resolved model, for callers that
+ * drive the SDK directly (the default step executor) rather than through the
+ * adapter.
  */
 export declare function instantiateLanguageModel(model: ResolvedModel): SdkLanguageModel;
+/** The AI SDK usage fields the runner reads, structurally. */
+export interface SdkUsage {
+    readonly inputTokens?: number | undefined;
+    readonly outputTokens?: number | undefined;
+    readonly inputTokenDetails?: {
+        readonly cacheReadTokens?: number | undefined;
+        readonly cacheWriteTokens?: number | undefined;
+    } | undefined;
+}
 /**
- * The AI Gateway reports per-request cost in provider metadata when available;
- * other providers simply lack the key. `cost` is what the gateway bills; BYOK
- * routes bill the provider key directly and report `cost: "0"`, so fall back
- * to `marketCost`, the list-price estimate of the same request.
+ * The prompt-cache split of one call's input, when the provider reported it.
+ * Providers that cache nothing report zeros; providers that do not report the
+ * split leave the fields undefined, and so does the report.
+ */
+export declare function cacheTokenFields(usage: SdkUsage | undefined): {
+    cacheReadTokens?: number;
+    cacheWriteTokens?: number;
+};
+/**
+ * Per-request cost, where the provider reports one in its metadata. The
+ * Vercel AI Gateway's `cost` is what it bills; BYOK routes bill the provider
+ * key directly and report `cost: "0"`, so fall back to `marketCost`, the
+ * list-price estimate of the same request. OpenRouter reports `usage.cost`
+ * when its provider is asked to account usage. Other providers simply lack
+ * the key.
  */
 export declare function readCost(metadata: Readonly<Record<string, Readonly<Record<string, unknown>>>> | undefined): number | undefined;
 /**
- * A rejected credential names the variable the runner read it from, whether
- * that was `E2E_MODEL_API_KEY`, the gateway's own `AI_GATEWAY_API_KEY`, or
- * the one `apiKeyEnv` chose. The gateway's own text only ever says
- * `AI_GATEWAY_API_KEY`. Takes the resolved model (the adapter path) or the SDK
- * model object (the tool loop path); a model that brought its own credential
- * has no variable to name. Requires the SDK to be loaded, which every caller
- * has done by the time a provider call has failed.
+ * A rejected credential is named for what it is. The model instance owns its
+ * credential, read from the provider package's own variable
+ * (`AI_GATEWAY_API_KEY`, `OPENROUTER_API_KEY`, ...) or passed at
+ * construction, so the runner can only say that the provider refused it.
+ * Requires the SDK to be loaded, which every caller has done by the time a
+ * provider call has failed.
  */
-export declare function credentialHint(cause: unknown, model?: ResolvedModel | SdkLanguageModel | string): string;
+export declare function credentialHint(cause: unknown): string;
 /**
  * True when a provider call ended by abort or timeout rather than by a
  * provider answer, including a retry chain the deadline cut short. Requires

package/dist/agent/model/sdk.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"sdk.d.ts","sourceRoot":"","sources":["../../../src/agent/model/sdk.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAIH,OAAO,EAAuB,KAAK,aAAa,EAAE,KAAK,gBAAgB,EAAE,MAAM,uBAAuB,CAAC;AAIvG,OAAO,EAIL,KAAK,YAAY,EAKlB,MAAM,cAAc,CAAC;AAOtB;;;;;;;;;;GAUG;AACH,eAAO,MAAM,iBAAiB,IAAI,CAAC;AAEnC,mFAAmF;AACnF,wBAAgB,kBAAkB,CAAC,KAAK,EAAE,aAAa,GAAG,SAAS,GAAG,YAAY,CAuFjF;AAsBD;;;;;;GAMG;AACH,wBAAgB,wBAAwB,CAAC,KAAK,EAAE,aAAa,GAAG,gBAAgB,CAE/E;AAuFD;;;;;GAKG;AACH,wBAAgB,QAAQ,CACtB,QAAQ,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,GAAG,SAAS,GAChF,MAAM,GAAG,SAAS,CAIpB;AAqDD;;;;;;;;GAQG;AACH,wBAAgB,cAAc,CAAC,KAAK,EAAE,OAAO,EAAE,KAAK,CAAC,EAAE,aAAa,GAAG,gBAAgB,GAAG,MAAM,GAAG,MAAM,CAcxG;AAoBD;;;;GAIG;AACH,wBAAgB,OAAO,CAAC,KAAK,EAAE,OAAO,GAAG,OAAO,CAI/C"}
+{"version":3,"file":"sdk.d.ts","sourceRoot":"","sources":["../../../src/agent/model/sdk.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AAIH,OAAO,KAAK,EAAE,aAAa,EAAE,gBAAgB,EAAE,MAAM,uBAAuB,CAAC;AAM7E,OAAO,EAIL,KAAK,YAAY,EAKlB,MAAM,cAAc,CAAC;AAEtB;;;;;;;;;;GAUG;AACH,eAAO,MAAM,iBAAiB,IAAI,CAAC;AAEnC,mFAAmF;AACnF,wBAAgB,kBAAkB,CAAC,KAAK,EAAE,aAAa,GAAG,SAAS,GAAG,YAAY,CAuFjF;AAsBD;;;;GAIG;AACH,wBAAgB,wBAAwB,CAAC,KAAK,EAAE,aAAa,GAAG,gBAAgB,CAE/E;AAUD,8DAA8D;AAC9D,MAAM,WAAW,QAAQ;IACvB,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;IAC1C,QAAQ,CAAC,YAAY,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;IAC3C,QAAQ,CAAC,iBAAiB,CAAC,EACvB;QACE,QAAQ,CAAC,eAAe,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;QAC9C,QAAQ,CAAC,gBAAgB,CAAC,EAAE,MAAM,GAAG,SAAS,CAAC;KAChD,GACD,SAAS,CAAC;CACf;AAED;;;;GAIG;AACH,wBAAgB,gBAAgB,CAC9B,KAAK,EAAE,QAAQ,GAAG,SAAS,GAC1B;IAAE,eAAe,CAAC,EAAE,MAAM,CAAC;IAAC,gBAAgB,CAAC,EAAE,MAAM,CAAA;CAAE,CAMzD;AAqCD;;;;;;;GAOG;AACH,wBAAgB,QAAQ,CACtB,QAAQ,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,GAAG,SAAS,GAChF,MAAM,GAAG,SAAS,CAOpB;AA6DD;;;;;;;GAOG;AACH,wBAAgB,cAAc,CAAC,KAAK,EAAE,OAAO,GAAG,MAAM,CAYrD;AAaD;;;;GAIG;AACH,wBAAgB,OAAO,CAAC,KAAK,EAAE,OAAO,GAAG,OAAO,CAI/C"}

package/dist/agent/model/sdk.js

@@ -1,20 +1,18 @@
 /**
- * AI SDK adapter. One implementation serves every
- * provider: a `provider/model-id` reference resolves through the AI Gateway,
- * and a caller-supplied AI SDK model instance (`openai('gpt-4o')`, a local
- * provider, a scripted test model) is used directly. Everything after model
- * construction — bounded requests, closed-grammar validation, usage and error
- * translation — is provider-independent.
+ * AI SDK adapter. One implementation serves every provider: the
+ * caller-supplied AI SDK model instance (`gateway('openai/gpt-5.4-mini')`,
+ * `openrouter(...)`, `openai('gpt-4o')`, a local provider, a scripted test
+ * model) is used as is. Everything after the model — bounded requests,
+ * closed-grammar validation, usage and error translation — is
+ * provider-independent.
  */
 import { withHint } from '../../internal/errors.js';
-import { GATEWAY_API_KEY_ENV } from '../../config/agent.js';
 import { aiSdk, loadAiSdk } from '../ai-sdk.js';
 import { packageVersion } from '../../internal/package-version.js';
 import { AgentError } from '../error.js';
+import { isContextOverflow } from './overflow.js';
+import { promptCacheHints } from './prompt-cache.js';
 import { imageTokenUpperBound, ModelOutputInvalidError, tokenUpperBound, } from './adapter.js';
-/** Default AI Gateway base URL used when no endpoint override is configured. */
-const DEFAULT_GATEWAY_ENDPOINT = 'https://ai-gateway.vercel.sh/v4/ai';
-const PROVIDER_PATTERN = /^[a-z0-9][a-z0-9._-]*$/;
 /**
  * Provider transport retries; distinct from runner-owned model-call budget.
  *
@@ -30,23 +28,22 @@ export const TRANSPORT_RETRIES = 5;
 /** Creates the adapter for one resolved model, or fails with MODEL_UNAVAILABLE. */
 export function createModelAdapter(model) {
     if (model === undefined) {
-        throw new AgentError('MODEL_UNAVAILABLE', 'the agent fixture requires a model: pass one to createAgent({ model }), set agent.model, or set E2E_MODEL');
+        throw new AgentError('MODEL_UNAVAILABLE', 'the agent fixture requires a model: pass an AI SDK model instance to createAgent({ model }) or set agent.model');
     }
-    const { endpoint, flavor } = validateModel(model);
     const adapterVersion = packageVersion(import.meta.url, '../../../package.json', '0.0.0');
-    // The AI SDK is an optional peer: the language model is constructed on the
-    // first call, after the loader has resolved (or clearly refused) it.
-    let languageModel;
+    const languageModel = model.model;
     return {
         provenance: {
             provider: model.provider,
             model: model.id,
-            endpoint,
-            adapterVersion: `${flavor}/${adapterVersion}`,
+            // The instance owns its transport; the report records that the endpoint
+            // is whatever the provider package was configured with.
+            endpoint: 'provider-default',
+            adapterVersion: `ai-sdk/${adapterVersion}`,
         },
         async generate(call) {
             const { generateText, jsonSchema, Output } = await loadAiSdk();
-            languageModel ??= instantiate(model).languageModel;
+            const cache = promptCacheHints(languageModel);
             const images = call.images ?? [];
             const inputBound = tokenUpperBound(call.system) +
                 tokenUpperBound(call.prompt) +
@@ -56,9 +53,12 @@ export function createModelAdapter(model) {
             }
             let issue;
             const schema = call.schema;
+            // The policy prefix is what every judgment call of a run shares, so it
+            // is where the prompt cache is addressed; the prompt itself is one-off.
+            const providerOptions = cache.providerOptions(call.providerOptions, call.system);
             const settings = {
                 model: languageModel,
-                system: call.system,
+                instructions: cache.instructions(call.system),
                 // Text-only calls keep the plain prompt form; images require the
                 // multi-part message form, and both must carry the same instruction
                 // text in the same position relative to the system policy.
@@ -66,9 +66,7 @@ export function createModelAdapter(model) {
                     ? { prompt: call.prompt }
                     : { messages: [userMessage(call.prompt, images)] }),
                 maxOutputTokens: call.maxOutputTokens,
-                ...(call.providerOptions === undefined
-                    ? {}
-                    : { providerOptions: call.providerOptions }),
+                ...(providerOptions === undefined ? {} : { providerOptions: providerOptions }),
                 maxRetries: TRANSPORT_RETRIES,
                 abortSignal: call.signal,
                 timeout: call.timeoutMs,
@@ -105,7 +103,7 @@ export function createModelAdapter(model) {
                 return { value: output, usage: readUsage(result, inputBound) };
             }
             catch (cause) {
-                throw translateModelError(cause, issue, call.signal, model);
+                throw translateModelError(cause, issue, call.signal);
             }
         },
     };
@@ -130,61 +128,37 @@ function userMessage(prompt, images) {
     };
 }
 /**
- * Builds the bare AI SDK language model for one resolved model reference, for
- * callers that drive the SDK directly (the default step executor) rather than
- * through the adapter. Fails with MODEL_UNAVAILABLE exactly like the adapter.
- * Gateway references require the AI SDK to be loaded (loadAiSdk) first; model
- * instances never touch it.
+ * The bare AI SDK language model of one resolved model, for callers that
+ * drive the SDK directly (the default step executor) rather than through the
+ * adapter.
  */
 export function instantiateLanguageModel(model) {
-    return instantiate(model).languageModel;
+    return model.model;
 }
 /**
- * Validates one resolved model and derives its report provenance without
- * touching the AI SDK, so an adapter can be constructed — and refuse clearly —
- * before the optional peer dependency is ever loaded.
+ * The prompt-cache split of one call's input, when the provider reported it.
+ * Providers that cache nothing report zeros; providers that do not report the
+ * split leave the fields undefined, and so does the report.
  */
-function validateModel(model) {
-    if (model.kind === 'instance') {
-        // The instance owns its transport; the report records that the endpoint is
-        // whatever the provider package defaults to.
-        return { endpoint: 'provider-default', flavor: 'ai-sdk' };
-    }
-    if (!PROVIDER_PATTERN.test(model.provider)) {
-        throw new AgentError('MODEL_UNAVAILABLE', `unknown model provider "${model.provider}"; expected a gateway provider ID or an AI SDK model instance`);
-    }
-    if (model.apiKey === undefined) {
-        throw new AgentError('MODEL_UNAVAILABLE', `no model credential: set ${model.apiKeyEnv} or ${GATEWAY_API_KEY_ENV}`);
-    }
-    return { endpoint: model.endpoint ?? DEFAULT_GATEWAY_ENDPOINT, flavor: 'ai-gateway' };
-}
-/**
- * The variable each gateway language model's credential was read from, keyed
- * by the model object the SDK is handed. The tool loop sees only that object,
- * and this is how a rejected credential is still reported by its source. A
- * name, never a value.
- */
-const credentialSources = new WeakMap();
-/** Builds the AI SDK language model for one validated resolved model. */
-function instantiate(model) {
-    if (model.kind === 'instance')
-        return { languageModel: model.model };
-    const { endpoint } = validateModel(model);
-    const gateway = aiSdk().createGateway({ apiKey: model.apiKey, baseURL: endpoint });
-    const languageModel = gateway.languageModel(`${model.provider}/${model.id}`);
-    credentialSources.set(languageModel, model.apiKeySource ?? model.apiKeyEnv);
-    return { languageModel };
+export function cacheTokenFields(usage) {
+    const details = usage?.inputTokenDetails;
+    return {
+        ...(typeof details?.cacheReadTokens === 'number' ? { cacheReadTokens: details.cacheReadTokens } : {}),
+        ...(typeof details?.cacheWriteTokens === 'number' ? { cacheWriteTokens: details.cacheWriteTokens } : {}),
+    };
 }
 function readUsage(result, inputBound) {
     const inputTokens = result.usage?.inputTokens;
     const outputTokens = result.usage?.outputTokens;
     const estimatedCostUsd = readCost(result.providerMetadata);
+    const cache = cacheTokenFields(result.usage);
     if (typeof inputTokens === 'number' && typeof outputTokens === 'number') {
-        return { inputTokens, outputTokens, accounting: 'provider', estimatedCostUsd };
+        return { inputTokens, outputTokens, ...cache, accounting: 'provider', estimatedCostUsd };
     }
     return {
         inputTokens: inputBound,
         outputTokens: tokenUpperBound(result.text ?? ''),
+        ...cache,
         accounting: 'adapter-upper-bound',
         estimatedCostUsd,
     };
@@ -208,16 +182,22 @@ function parseJsonObject(text) {
     }
 }
 /**
- * The AI Gateway reports per-request cost in provider metadata when available;
- * other providers simply lack the key. `cost` is what the gateway bills; BYOK
- * routes bill the provider key directly and report `cost: "0"`, so fall back
- * to `marketCost`, the list-price estimate of the same request.
+ * Per-request cost, where the provider reports one in its metadata. The
+ * Vercel AI Gateway's `cost` is what it bills; BYOK routes bill the provider
+ * key directly and report `cost: "0"`, so fall back to `marketCost`, the
+ * list-price estimate of the same request. OpenRouter reports `usage.cost`
+ * when its provider is asked to account usage. Other providers simply lack
+ * the key.
  */
 export function readCost(metadata) {
     const billed = parseCost(metadata?.['gateway']?.['cost']);
     if (billed !== undefined && billed > 0)
         return billed;
-    return parseCost(metadata?.['gateway']?.['marketCost']) ?? billed;
+    const market = parseCost(metadata?.['gateway']?.['marketCost']) ?? billed;
+    if (market !== undefined)
+        return market;
+    const openrouter = metadata?.['openrouter']?.['usage'];
+    return typeof openrouter === 'object' && openrouter !== null ? parseCost(openrouter['cost']) : undefined;
 }
 function parseCost(raw) {
     const cost = typeof raw === 'string' ? Number(raw) : raw;
@@ -228,7 +208,7 @@ function parseCost(raw) {
  * only after `generate` loaded the SDK, so the cached module is available for
  * the error-class checks.
  */
-function translateModelError(rawCause, issue, signal, model) {
+function translateModelError(rawCause, issue, signal) {
     const { APICallError, NoObjectGeneratedError } = aiSdk();
     const cause = unwrapRetry(rawCause);
     if (cause instanceof AgentError)
@@ -244,25 +224,29 @@ function translateModelError(rawCause, issue, signal, model) {
     if (NoObjectGeneratedError.isInstance(cause)) {
         return new ModelOutputInvalidError(issue ?? 'provider response did not match the closed response grammar', { ...(cause.text !== undefined ? { rawText: cause.text } : {}), cause });
     }
+    // Too big a request is not a provider outage: the observation budget or
+    // the step is what needs to shrink, and the code says so.
+    if (isContextOverflow(cause)) {
+        return new AgentError('CONTEXT_OVERFLOW', `the model request exceeded the context window: ${cause instanceof Error ? cause.message : String(cause)}`, { cause });
+    }
     if (APICallError.isInstance(cause)) {
-        return new AgentError('MODEL_PROVIDER_FAILED', withHint(`model provider failed: ${cause.message}`, credentialHint(cause, model)), { cause });
+        return new AgentError('MODEL_PROVIDER_FAILED', withHint(`model provider failed: ${cause.message}`, credentialHint(cause)), { cause });
     }
     return new AgentError('MODEL_PROVIDER_FAILED', `model provider failed: ${cause instanceof Error ? cause.message : String(cause)}`, { cause });
 }
 /**
- * A rejected credential names the variable the runner read it from, whether
- * that was `E2E_MODEL_API_KEY`, the gateway's own `AI_GATEWAY_API_KEY`, or
- * the one `apiKeyEnv` chose. The gateway's own text only ever says
- * `AI_GATEWAY_API_KEY`. Takes the resolved model (the adapter path) or the SDK
- * model object (the tool loop path); a model that brought its own credential
- * has no variable to name. Requires the SDK to be loaded, which every caller
- * has done by the time a provider call has failed.
+ * A rejected credential is named for what it is. The model instance owns its
+ * credential, read from the provider package's own variable
+ * (`AI_GATEWAY_API_KEY`, `OPENROUTER_API_KEY`, ...) or passed at
+ * construction, so the runner can only say that the provider refused it.
+ * Requires the SDK to be loaded, which every caller has done by the time a
+ * provider call has failed.
  */
-export function credentialHint(cause, model) {
+export function credentialHint(cause) {
     const failure = unwrapRetry(cause);
     if (!(failure instanceof Error))
         return '';
-    // The gateway raises its own authentication error class and a raw provider
+    // Gateways raise their own authentication error classes and a raw provider
     // an APICallError; both carry the HTTP status, so that is what is read.
     const statusCode = failure.statusCode;
     const rejected = statusCode === 401 ||
@@ -270,18 +254,7 @@ export function credentialHint(cause, model) {
         /unauthenticated|unauthorized|authentication/i.test(`${failure.name} ${failure.message}`);
     if (!rejected)
         return '';
-    const source = credentialSource(model);
-    if (source === undefined)
-        return 'the provider rejected the credential the model instance was created with';
-    return `the gateway rejected the credential read from ${source}: check that the key is complete and belongs to the AI Gateway (https://vercel.com/docs/ai-gateway)`;
-}
-/** The variable a gateway model's credential came from; undefined for a model that owns its own. */
-function credentialSource(model) {
-    if (model === undefined || typeof model !== 'object')
-        return undefined;
-    if ('kind' in model)
-        return model.kind === 'gateway' ? (model.apiKeySource ?? model.apiKeyEnv) : undefined;
-    return credentialSources.get(model);
+    return 'the provider rejected the credential the model instance was created with: check the variable the provider package reads, or the key passed at construction';
 }
 /**
  * Unwraps a spent transport retry chain to the attempt that actually failed,