omk-ai 1.3.2 → 1.4.1

package/CHANGELOG.md

@@ -1,5 +1,23 @@
 # Changelog
 
+## [Unreleased]
+
+## [1.4.1] - 2026-10-13
+
+### Changed
+
+- Lockstep version alignment with OMK 1.4.1. Provider behavior and external dependency versions are unchanged.
+
+### Fixed
+
+- Proxy-routing regression tests inject a connection failure instead of assuming a host immediately rejects a fixed local port. They still verify proxy-agent dispatch, failure propagation and absence of a direct-fetch fallback.
+
+## [1.4.0] - 2026-10-11
+
+### Fixed
+
+- xAI reasoning tokens are included in output usage without double-counting when total tokens are present. Provider-reported billed USD is available separately from the catalog cost estimate.
+
 ## [1.3.2] - 2026-10-10
 
 ### Added

package/dist/prompt-hash.d.ts

@@ -0,0 +1,16 @@
+import type { Context, Message } from "./types.ts";
+/** sha256 hex of the system prompt string as UTF-8; the empty string when there is none. */
+export declare function systemHash(context: Pick<Context, "systemPrompt">): string;
+/** The tool list as hashed by {@link toolsHash}: `[{name, description, parameters}]` in the order given. */
+export declare function serializePromptTools(context: Pick<Context, "tools">): string;
+/** sha256 hex of {@link serializePromptTools}: names, descriptions and schemas, in sent order, keys as given. */
+export declare function toolsHash(context: Pick<Context, "tools">): string;
+/**
+ * One message as it reaches a provider: role, content and the tool-result
+ * fields. Local bookkeeping (timestamp, usage, stopReason, model ids) is left
+ * out because providers do not send it.
+ */
+export declare function serializePromptMessage(message: Message): string;
+/** sha256 hex over the first `count` messages (default all), each serialized by {@link serializePromptMessage}. */
+export declare function messagesPrefixHash(context: Pick<Context, "messages">, count?: number): string;
+//# sourceMappingURL=prompt-hash.d.ts.map

package/dist/prompt-hash.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"prompt-hash.d.ts","sourceRoot":"","sources":["../src/prompt-hash.ts"],"names":[],"mappings":"AAgBA,OAAO,KAAK,EAAE,OAAO,EAAE,OAAO,EAAE,MAAM,YAAY,CAAC;AAMnD,4FAA4F;AAC5F,wBAAgB,UAAU,CAAC,OAAO,EAAE,IAAI,CAAC,OAAO,EAAE,cAAc,CAAC,GAAG,MAAM,CAEzE;AAED,4GAA4G;AAC5G,wBAAgB,oBAAoB,CAAC,OAAO,EAAE,IAAI,CAAC,OAAO,EAAE,OAAO,CAAC,GAAG,MAAM,CAQ5E;AAED,iHAAiH;AACjH,wBAAgB,SAAS,CAAC,OAAO,EAAE,IAAI,CAAC,OAAO,EAAE,OAAO,CAAC,GAAG,MAAM,CAEjE;AAED;;;;GAIG;AACH,wBAAgB,sBAAsB,CAAC,OAAO,EAAE,OAAO,GAAG,MAAM,CAW/D;AAED,mHAAmH;AACnH,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,IAAI,CAAC,OAAO,EAAE,UAAU,CAAC,EAAE,KAAK,SAA0B,GAAG,MAAM,CAO9G","sourcesContent":["/**\n * Hashes of the cacheable request prefix (spec 051).\n *\n * Provider prompt caches (Anthropic, xAI, OpenAI) reuse work only while the\n * system prompt, the tool list and the earlier messages of a request are\n * byte-identical to the previous request. These helpers hash exactly those\n * parts of a `Context`, so a test or a run log can tell which part changed.\n *\n * Node-only (`node:crypto`), so it is a separate entry (`omk-ai/prompt-hash`)\n * and not part of the browser-safe package index.\n *\n * Nothing is normalized: tool parameters are serialized with their own key\n * order, so a schema whose key order changes between calls shows up as a\n * changed hash instead of being hidden.\n */\nimport { createHash } from \"node:crypto\";\nimport type { Context, Message } from \"./types.ts\";\n\nfunction sha256(text: string): string {\n\treturn createHash(\"sha256\").update(text, \"utf8\").digest(\"hex\");\n}\n\n/** sha256 hex of the system prompt string as UTF-8; the empty string when there is none. */\nexport function systemHash(context: Pick<Context, \"systemPrompt\">): string {\n\treturn sha256(context.systemPrompt ?? \"\");\n}\n\n/** The tool list as hashed by {@link toolsHash}: `[{name, description, parameters}]` in the order given. */\nexport function serializePromptTools(context: Pick<Context, \"tools\">): string {\n\treturn JSON.stringify(\n\t\t(context.tools ?? []).map((tool) => ({\n\t\t\tname: tool.name,\n\t\t\tdescription: tool.description,\n\t\t\tparameters: tool.parameters,\n\t\t})),\n\t);\n}\n\n/** sha256 hex of {@link serializePromptTools}: names, descriptions and schemas, in sent order, keys as given. */\nexport function toolsHash(context: Pick<Context, \"tools\">): string {\n\treturn sha256(serializePromptTools(context));\n}\n\n/**\n * One message as it reaches a provider: role, content and the tool-result\n * fields. Local bookkeeping (timestamp, usage, stopReason, model ids) is left\n * out because providers do not send it.\n */\nexport function serializePromptMessage(message: Message): string {\n\tif (message.role === \"toolResult\") {\n\t\treturn JSON.stringify({\n\t\t\trole: message.role,\n\t\t\ttoolCallId: message.toolCallId,\n\t\t\ttoolName: message.toolName,\n\t\t\tisError: message.isError,\n\t\t\tcontent: message.content,\n\t\t});\n\t}\n\treturn JSON.stringify({ role: message.role, content: message.content });\n}\n\n/** sha256 hex over the first `count` messages (default all), each serialized by {@link serializePromptMessage}. */\nexport function messagesPrefixHash(context: Pick<Context, \"messages\">, count = context.messages.length): string {\n\tconst hash = createHash(\"sha256\");\n\tfor (const message of context.messages.slice(0, count)) {\n\t\thash.update(serializePromptMessage(message), \"utf8\");\n\t\thash.update(\"\\n\", \"utf8\");\n\t}\n\treturn hash.digest(\"hex\");\n}\n"]}

package/dist/prompt-hash.js

@@ -0,0 +1,62 @@
+/**
+ * Hashes of the cacheable request prefix (spec 051).
+ *
+ * Provider prompt caches (Anthropic, xAI, OpenAI) reuse work only while the
+ * system prompt, the tool list and the earlier messages of a request are
+ * byte-identical to the previous request. These helpers hash exactly those
+ * parts of a `Context`, so a test or a run log can tell which part changed.
+ *
+ * Node-only (`node:crypto`), so it is a separate entry (`omk-ai/prompt-hash`)
+ * and not part of the browser-safe package index.
+ *
+ * Nothing is normalized: tool parameters are serialized with their own key
+ * order, so a schema whose key order changes between calls shows up as a
+ * changed hash instead of being hidden.
+ */
+import { createHash } from "node:crypto";
+function sha256(text) {
+    return createHash("sha256").update(text, "utf8").digest("hex");
+}
+/** sha256 hex of the system prompt string as UTF-8; the empty string when there is none. */
+export function systemHash(context) {
+    return sha256(context.systemPrompt ?? "");
+}
+/** The tool list as hashed by {@link toolsHash}: `[{name, description, parameters}]` in the order given. */
+export function serializePromptTools(context) {
+    return JSON.stringify((context.tools ?? []).map((tool) => ({
+        name: tool.name,
+        description: tool.description,
+        parameters: tool.parameters,
+    })));
+}
+/** sha256 hex of {@link serializePromptTools}: names, descriptions and schemas, in sent order, keys as given. */
+export function toolsHash(context) {
+    return sha256(serializePromptTools(context));
+}
+/**
+ * One message as it reaches a provider: role, content and the tool-result
+ * fields. Local bookkeeping (timestamp, usage, stopReason, model ids) is left
+ * out because providers do not send it.
+ */
+export function serializePromptMessage(message) {
+    if (message.role === "toolResult") {
+        return JSON.stringify({
+            role: message.role,
+            toolCallId: message.toolCallId,
+            toolName: message.toolName,
+            isError: message.isError,
+            content: message.content,
+        });
+    }
+    return JSON.stringify({ role: message.role, content: message.content });
+}
+/** sha256 hex over the first `count` messages (default all), each serialized by {@link serializePromptMessage}. */
+export function messagesPrefixHash(context, count = context.messages.length) {
+    const hash = createHash("sha256");
+    for (const message of context.messages.slice(0, count)) {
+        hash.update(serializePromptMessage(message), "utf8");
+        hash.update("\n", "utf8");
+    }
+    return hash.digest("hex");
+}
+//# sourceMappingURL=prompt-hash.js.map

package/dist/prompt-hash.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"prompt-hash.js","sourceRoot":"","sources":["../src/prompt-hash.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;GAcG;AACH,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AAGzC,SAAS,MAAM,CAAC,IAAY,EAAU;IACrC,OAAO,UAAU,CAAC,QAAQ,CAAC,CAAC,MAAM,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;AAAA,CAC/D;AAED,4FAA4F;AAC5F,MAAM,UAAU,UAAU,CAAC,OAAsC,EAAU;IAC1E,OAAO,MAAM,CAAC,OAAO,CAAC,YAAY,IAAI,EAAE,CAAC,CAAC;AAAA,CAC1C;AAED,4GAA4G;AAC5G,MAAM,UAAU,oBAAoB,CAAC,OAA+B,EAAU;IAC7E,OAAO,IAAI,CAAC,SAAS,CACpB,CAAC,OAAO,CAAC,KAAK,IAAI,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,CAAC;QACpC,IAAI,EAAE,IAAI,CAAC,IAAI;QACf,WAAW,EAAE,IAAI,CAAC,WAAW;QAC7B,UAAU,EAAE,IAAI,CAAC,UAAU;KAC3B,CAAC,CAAC,CACH,CAAC;AAAA,CACF;AAED,iHAAiH;AACjH,MAAM,UAAU,SAAS,CAAC,OAA+B,EAAU;IAClE,OAAO,MAAM,CAAC,oBAAoB,CAAC,OAAO,CAAC,CAAC,CAAC;AAAA,CAC7C;AAED;;;;GAIG;AACH,MAAM,UAAU,sBAAsB,CAAC,OAAgB,EAAU;IAChE,IAAI,OAAO,CAAC,IAAI,KAAK,YAAY,EAAE,CAAC;QACnC,OAAO,IAAI,CAAC,SAAS,CAAC;YACrB,IAAI,EAAE,OAAO,CAAC,IAAI;YAClB,UAAU,EAAE,OAAO,CAAC,UAAU;YAC9B,QAAQ,EAAE,OAAO,CAAC,QAAQ;YAC1B,OAAO,EAAE,OAAO,CAAC,OAAO;YACxB,OAAO,EAAE,OAAO,CAAC,OAAO;SACxB,CAAC,CAAC;IACJ,CAAC;IACD,OAAO,IAAI,CAAC,SAAS,CAAC,EAAE,IAAI,EAAE,OAAO,CAAC,IAAI,EAAE,OAAO,EAAE,OAAO,CAAC,OAAO,EAAE,CAAC,CAAC;AAAA,CACxE;AAED,mHAAmH;AACnH,MAAM,UAAU,kBAAkB,CAAC,OAAkC,EAAE,KAAK,GAAG,OAAO,CAAC,QAAQ,CAAC,MAAM,EAAU;IAC/G,MAAM,IAAI,GAAG,UAAU,CAAC,QAAQ,CAAC,CAAC;IAClC,KAAK,MAAM,OAAO,IAAI,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC,EAAE,KAAK,CAAC,EAAE,CAAC;QACxD,IAAI,CAAC,MAAM,CAAC,sBAAsB,CAAC,OAAO,CAAC,EAAE,MAAM,CAAC,CAAC;QACrD,IAAI,CAAC,MAAM,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC;IAC3B,CAAC;IACD,OAAO,IAAI,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC;AAAA,CAC1B","sourcesContent":["/**\n * Hashes of the cacheable request prefix (spec 051).\n *\n * Provider prompt caches (Anthropic, xAI, OpenAI) reuse work only while the\n * system prompt, the tool list and the earlier messages of a request are\n * byte-identical to the previous request. These helpers hash exactly those\n * parts of a `Context`, so a test or a run log can tell which part changed.\n *\n * Node-only (`node:crypto`), so it is a separate entry (`omk-ai/prompt-hash`)\n * and not part of the browser-safe package index.\n *\n * Nothing is normalized: tool parameters are serialized with their own key\n * order, so a schema whose key order changes between calls shows up as a\n * changed hash instead of being hidden.\n */\nimport { createHash } from \"node:crypto\";\nimport type { Context, Message } from \"./types.ts\";\n\nfunction sha256(text: string): string {\n\treturn createHash(\"sha256\").update(text, \"utf8\").digest(\"hex\");\n}\n\n/** sha256 hex of the system prompt string as UTF-8; the empty string when there is none. */\nexport function systemHash(context: Pick<Context, \"systemPrompt\">): string {\n\treturn sha256(context.systemPrompt ?? \"\");\n}\n\n/** The tool list as hashed by {@link toolsHash}: `[{name, description, parameters}]` in the order given. */\nexport function serializePromptTools(context: Pick<Context, \"tools\">): string {\n\treturn JSON.stringify(\n\t\t(context.tools ?? []).map((tool) => ({\n\t\t\tname: tool.name,\n\t\t\tdescription: tool.description,\n\t\t\tparameters: tool.parameters,\n\t\t})),\n\t);\n}\n\n/** sha256 hex of {@link serializePromptTools}: names, descriptions and schemas, in sent order, keys as given. */\nexport function toolsHash(context: Pick<Context, \"tools\">): string {\n\treturn sha256(serializePromptTools(context));\n}\n\n/**\n * One message as it reaches a provider: role, content and the tool-result\n * fields. Local bookkeeping (timestamp, usage, stopReason, model ids) is left\n * out because providers do not send it.\n */\nexport function serializePromptMessage(message: Message): string {\n\tif (message.role === \"toolResult\") {\n\t\treturn JSON.stringify({\n\t\t\trole: message.role,\n\t\t\ttoolCallId: message.toolCallId,\n\t\t\ttoolName: message.toolName,\n\t\t\tisError: message.isError,\n\t\t\tcontent: message.content,\n\t\t});\n\t}\n\treturn JSON.stringify({ role: message.role, content: message.content });\n}\n\n/** sha256 hex over the first `count` messages (default all), each serialized by {@link serializePromptMessage}. */\nexport function messagesPrefixHash(context: Pick<Context, \"messages\">, count = context.messages.length): string {\n\tconst hash = createHash(\"sha256\");\n\tfor (const message of context.messages.slice(0, count)) {\n\t\thash.update(serializePromptMessage(message), \"utf8\");\n\t\thash.update(\"\\n\", \"utf8\");\n\t}\n\treturn hash.digest(\"hex\");\n}\n"]}

package/dist/providers/openai-completions-usage.d.ts

@@ -0,0 +1,28 @@
+import type { AssistantMessage, Model } from "../types.ts";
+export interface OpenAICompletionsRawUsage {
+    prompt_tokens?: number;
+    completion_tokens?: number;
+    total_tokens?: number;
+    prompt_cache_hit_tokens?: number;
+    prompt_tokens_details?: {
+        cached_tokens?: number;
+        cache_write_tokens?: number;
+    };
+    completion_tokens_details?: {
+        reasoning_tokens?: number;
+    };
+    /** xAI only: the amount actually billed for this request, in 1e-10 USD. */
+    cost_in_usd_ticks?: number;
+}
+/**
+ * Map a Chat Completions `usage` object to our usage record.
+ *
+ * OpenAI counts reasoning tokens inside `completion_tokens`. xAI does not: it
+ * reports them only in `completion_tokens_details.reasoning_tokens` (its
+ * `total_tokens` is prompt + completion + reasoning) and bills them at the
+ * output rate. Without adding them, Grok reasoning runs were under-counted and
+ * under-costed, by a lot at high effort. xAI also returns the billed amount as
+ * `cost_in_usd_ticks`, which is surfaced as `cost.billed`.
+ */
+export declare function parseChunkUsage(rawUsage: OpenAICompletionsRawUsage, model: Model<"openai-completions">): AssistantMessage["usage"];
+//# sourceMappingURL=openai-completions-usage.d.ts.map

package/dist/providers/openai-completions-usage.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"openai-completions-usage.d.ts","sourceRoot":"","sources":["../../src/providers/openai-completions-usage.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAAE,gBAAgB,EAAE,KAAK,EAAE,MAAM,aAAa,CAAC;AAE3D,MAAM,WAAW,yBAAyB;IACzC,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,uBAAuB,CAAC,EAAE,MAAM,CAAC;IACjC,qBAAqB,CAAC,EAAE;QAAE,aAAa,CAAC,EAAE,MAAM,CAAC;QAAC,kBAAkB,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;IAChF,yBAAyB,CAAC,EAAE;QAAE,gBAAgB,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;IAC1D,2EAA2E;IAC3E,iBAAiB,CAAC,EAAE,MAAM,CAAC;CAC3B;AA6BD;;;;;;;;;GASG;AACH,wBAAgB,eAAe,CAC9B,QAAQ,EAAE,yBAAyB,EACnC,KAAK,EAAE,KAAK,CAAC,oBAAoB,CAAC,GAChC,gBAAgB,CAAC,OAAO,CAAC,CA8B3B","sourcesContent":["import { calculateCost } from \"../models.ts\";\nimport type { AssistantMessage, Model } from \"../types.ts\";\n\nexport interface OpenAICompletionsRawUsage {\n\tprompt_tokens?: number;\n\tcompletion_tokens?: number;\n\ttotal_tokens?: number;\n\tprompt_cache_hit_tokens?: number;\n\tprompt_tokens_details?: { cached_tokens?: number; cache_write_tokens?: number };\n\tcompletion_tokens_details?: { reasoning_tokens?: number };\n\t/** xAI only: the amount actually billed for this request, in 1e-10 USD. */\n\tcost_in_usd_ticks?: number;\n}\n\nconst USD_TICKS = 10_000_000_000;\n\nfunction isXaiModel(model: Model<\"openai-completions\">): boolean {\n\treturn model.provider === \"xai\" || model.baseUrl.includes(\"api.x.ai\");\n}\n\nfunction count(value: unknown): number {\n\treturn typeof value === \"number\" && Number.isFinite(value) && value > 0 ? value : 0;\n}\n\n/**\n * Reasoning tokens xAI reported outside `completion_tokens`. OpenAI counts them\n * inside it; xAI's documented responses do not. Rather than trust one rule for\n * every model, use `total_tokens` when present: only the part of the reasoning\n * count that `total_tokens` shows beyond prompt + completion is added, so a\n * model that nests reasoning inside `completion_tokens` is never double-counted.\n */\nfunction separateReasoningTokens(\n\trawUsage: OpenAICompletionsRawUsage,\n\tpromptTokens: number,\n\tcompletionTokens: number,\n): number {\n\tconst reasoning = count(rawUsage.completion_tokens_details?.reasoning_tokens);\n\tif (typeof rawUsage.total_tokens !== \"number\") return reasoning;\n\treturn Math.min(reasoning, Math.max(0, count(rawUsage.total_tokens) - promptTokens - completionTokens));\n}\n\n/**\n * Map a Chat Completions `usage` object to our usage record.\n *\n * OpenAI counts reasoning tokens inside `completion_tokens`. xAI does not: it\n * reports them only in `completion_tokens_details.reasoning_tokens` (its\n * `total_tokens` is prompt + completion + reasoning) and bills them at the\n * output rate. Without adding them, Grok reasoning runs were under-counted and\n * under-costed, by a lot at high effort. xAI also returns the billed amount as\n * `cost_in_usd_ticks`, which is surfaced as `cost.billed`.\n */\nexport function parseChunkUsage(\n\trawUsage: OpenAICompletionsRawUsage,\n\tmodel: Model<\"openai-completions\">,\n): AssistantMessage[\"usage\"] {\n\tconst promptTokens = count(rawUsage.prompt_tokens);\n\tconst cacheReadTokens = count(rawUsage.prompt_tokens_details?.cached_tokens ?? rawUsage.prompt_cache_hit_tokens);\n\tconst cacheWriteTokens = count(rawUsage.prompt_tokens_details?.cache_write_tokens);\n\n\t// Follow documented OpenAI/OpenRouter semantics: cached_tokens is cache-read\n\t// tokens (hits). OpenAI does not document or emit cache_write_tokens, but\n\t// OpenRouter-compatible providers can include it as a separate write count.\n\t// OpenRouter's own provider/tests affirm the separate mapping:\n\t// https://github.com/OpenRouterTeam/ai-sdk-provider/pull/409\n\t// Do not subtract writes from cached_tokens, otherwise spec-compliant\n\t// providers are under-reported. DS4 mirrors this contract too:\n\t// https://github.com/antirez/ds4/pull/29\n\tconst input = Math.max(0, promptTokens - cacheReadTokens - cacheWriteTokens);\n\tconst xai = isXaiModel(model);\n\tconst completionTokens = count(rawUsage.completion_tokens);\n\tconst reasoningTokens = xai ? separateReasoningTokens(rawUsage, promptTokens, completionTokens) : 0;\n\tconst outputTokens = completionTokens + reasoningTokens;\n\tconst usage: AssistantMessage[\"usage\"] = {\n\t\tinput,\n\t\toutput: outputTokens,\n\t\tcacheRead: cacheReadTokens,\n\t\tcacheWrite: cacheWriteTokens,\n\t\ttotalTokens: input + outputTokens + cacheReadTokens + cacheWriteTokens,\n\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t};\n\tcalculateCost(model, usage);\n\tconst ticks = rawUsage.cost_in_usd_ticks;\n\tif (xai && typeof ticks === \"number\" && Number.isFinite(ticks) && ticks >= 0) usage.cost.billed = ticks / USD_TICKS;\n\treturn usage;\n}\n"]}

package/dist/providers/openai-completions-usage.js

@@ -0,0 +1,63 @@
+import { calculateCost } from "../models.js";
+const USD_TICKS = 10_000_000_000;
+function isXaiModel(model) {
+    return model.provider === "xai" || model.baseUrl.includes("api.x.ai");
+}
+function count(value) {
+    return typeof value === "number" && Number.isFinite(value) && value > 0 ? value : 0;
+}
+/**
+ * Reasoning tokens xAI reported outside `completion_tokens`. OpenAI counts them
+ * inside it; xAI's documented responses do not. Rather than trust one rule for
+ * every model, use `total_tokens` when present: only the part of the reasoning
+ * count that `total_tokens` shows beyond prompt + completion is added, so a
+ * model that nests reasoning inside `completion_tokens` is never double-counted.
+ */
+function separateReasoningTokens(rawUsage, promptTokens, completionTokens) {
+    const reasoning = count(rawUsage.completion_tokens_details?.reasoning_tokens);
+    if (typeof rawUsage.total_tokens !== "number")
+        return reasoning;
+    return Math.min(reasoning, Math.max(0, count(rawUsage.total_tokens) - promptTokens - completionTokens));
+}
+/**
+ * Map a Chat Completions `usage` object to our usage record.
+ *
+ * OpenAI counts reasoning tokens inside `completion_tokens`. xAI does not: it
+ * reports them only in `completion_tokens_details.reasoning_tokens` (its
+ * `total_tokens` is prompt + completion + reasoning) and bills them at the
+ * output rate. Without adding them, Grok reasoning runs were under-counted and
+ * under-costed, by a lot at high effort. xAI also returns the billed amount as
+ * `cost_in_usd_ticks`, which is surfaced as `cost.billed`.
+ */
+export function parseChunkUsage(rawUsage, model) {
+    const promptTokens = count(rawUsage.prompt_tokens);
+    const cacheReadTokens = count(rawUsage.prompt_tokens_details?.cached_tokens ?? rawUsage.prompt_cache_hit_tokens);
+    const cacheWriteTokens = count(rawUsage.prompt_tokens_details?.cache_write_tokens);
+    // Follow documented OpenAI/OpenRouter semantics: cached_tokens is cache-read
+    // tokens (hits). OpenAI does not document or emit cache_write_tokens, but
+    // OpenRouter-compatible providers can include it as a separate write count.
+    // OpenRouter's own provider/tests affirm the separate mapping:
+    // https://github.com/OpenRouterTeam/ai-sdk-provider/pull/409
+    // Do not subtract writes from cached_tokens, otherwise spec-compliant
+    // providers are under-reported. DS4 mirrors this contract too:
+    // https://github.com/antirez/ds4/pull/29
+    const input = Math.max(0, promptTokens - cacheReadTokens - cacheWriteTokens);
+    const xai = isXaiModel(model);
+    const completionTokens = count(rawUsage.completion_tokens);
+    const reasoningTokens = xai ? separateReasoningTokens(rawUsage, promptTokens, completionTokens) : 0;
+    const outputTokens = completionTokens + reasoningTokens;
+    const usage = {
+        input,
+        output: outputTokens,
+        cacheRead: cacheReadTokens,
+        cacheWrite: cacheWriteTokens,
+        totalTokens: input + outputTokens + cacheReadTokens + cacheWriteTokens,
+        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+    };
+    calculateCost(model, usage);
+    const ticks = rawUsage.cost_in_usd_ticks;
+    if (xai && typeof ticks === "number" && Number.isFinite(ticks) && ticks >= 0)
+        usage.cost.billed = ticks / USD_TICKS;
+    return usage;
+}
+//# sourceMappingURL=openai-completions-usage.js.map

package/dist/providers/openai-completions-usage.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"openai-completions-usage.js","sourceRoot":"","sources":["../../src/providers/openai-completions-usage.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,aAAa,EAAE,MAAM,cAAc,CAAC;AAc7C,MAAM,SAAS,GAAG,cAAc,CAAC;AAEjC,SAAS,UAAU,CAAC,KAAkC,EAAW;IAChE,OAAO,KAAK,CAAC,QAAQ,KAAK,KAAK,IAAI,KAAK,CAAC,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAC,CAAC;AAAA,CACtE;AAED,SAAS,KAAK,CAAC,KAAc,EAAU;IACtC,OAAO,OAAO,KAAK,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC;AAAA,CACpF;AAED;;;;;;GAMG;AACH,SAAS,uBAAuB,CAC/B,QAAmC,EACnC,YAAoB,EACpB,gBAAwB,EACf;IACT,MAAM,SAAS,GAAG,KAAK,CAAC,QAAQ,CAAC,yBAAyB,EAAE,gBAAgB,CAAC,CAAC;IAC9E,IAAI,OAAO,QAAQ,CAAC,YAAY,KAAK,QAAQ;QAAE,OAAO,SAAS,CAAC;IAChE,OAAO,IAAI,CAAC,GAAG,CAAC,SAAS,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,KAAK,CAAC,QAAQ,CAAC,YAAY,CAAC,GAAG,YAAY,GAAG,gBAAgB,CAAC,CAAC,CAAC;AAAA,CACxG;AAED;;;;;;;;;GASG;AACH,MAAM,UAAU,eAAe,CAC9B,QAAmC,EACnC,KAAkC,EACN;IAC5B,MAAM,YAAY,GAAG,KAAK,CAAC,QAAQ,CAAC,aAAa,CAAC,CAAC;IACnD,MAAM,eAAe,GAAG,KAAK,CAAC,QAAQ,CAAC,qBAAqB,EAAE,aAAa,IAAI,QAAQ,CAAC,uBAAuB,CAAC,CAAC;IACjH,MAAM,gBAAgB,GAAG,KAAK,CAAC,QAAQ,CAAC,qBAAqB,EAAE,kBAAkB,CAAC,CAAC;IAEnF,6EAA6E;IAC7E,0EAA0E;IAC1E,4EAA4E;IAC5E,+DAA+D;IAC/D,6DAA6D;IAC7D,sEAAsE;IACtE,+DAA+D;IAC/D,yCAAyC;IACzC,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,YAAY,GAAG,eAAe,GAAG,gBAAgB,CAAC,CAAC;IAC7E,MAAM,GAAG,GAAG,UAAU,CAAC,KAAK,CAAC,CAAC;IAC9B,MAAM,gBAAgB,GAAG,KAAK,CAAC,QAAQ,CAAC,iBAAiB,CAAC,CAAC;IAC3D,MAAM,eAAe,GAAG,GAAG,CAAC,CAAC,CAAC,uBAAuB,CAAC,QAAQ,EAAE,YAAY,EAAE,gBAAgB,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;IACpG,MAAM,YAAY,GAAG,gBAAgB,GAAG,eAAe,CAAC;IACxD,MAAM,KAAK,GAA8B;QACxC,KAAK;QACL,MAAM,EAAE,YAAY;QACpB,SAAS,EAAE,eAAe;QAC1B,UAAU,EAAE,gBAAgB;QAC5B,WAAW,EAAE,KAAK,GAAG,YAAY,GAAG,eAAe,GAAG,gBAAgB;QACtE,IAAI,EAAE,EAAE,KAAK,EAAE,CAAC,EAAE,MAAM,EAAE,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE,KAAK,EAAE,CAAC,EAAE;KACpE,CAAC;IACF,aAAa,CAAC,KAAK,EAAE,KAAK,CAAC,CAAC;IAC5B,MAAM,KAAK,GAAG,QAAQ,CAAC,iBAAiB,CAAC;IACzC,IAAI,GAAG,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,MAAM,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,KAAK,IAAI,CAAC;QAAE,KAAK,CAAC,IAAI,CAAC,MAAM,GAAG,KAAK,GAAG,SAAS,CAAC;IACpH,OAAO,KAAK,CAAC;AAAA,CACb","sourcesContent":["import { calculateCost } from \"../models.ts\";\nimport type { AssistantMessage, Model } from \"../types.ts\";\n\nexport interface OpenAICompletionsRawUsage {\n\tprompt_tokens?: number;\n\tcompletion_tokens?: number;\n\ttotal_tokens?: number;\n\tprompt_cache_hit_tokens?: number;\n\tprompt_tokens_details?: { cached_tokens?: number; cache_write_tokens?: number };\n\tcompletion_tokens_details?: { reasoning_tokens?: number };\n\t/** xAI only: the amount actually billed for this request, in 1e-10 USD. */\n\tcost_in_usd_ticks?: number;\n}\n\nconst USD_TICKS = 10_000_000_000;\n\nfunction isXaiModel(model: Model<\"openai-completions\">): boolean {\n\treturn model.provider === \"xai\" || model.baseUrl.includes(\"api.x.ai\");\n}\n\nfunction count(value: unknown): number {\n\treturn typeof value === \"number\" && Number.isFinite(value) && value > 0 ? value : 0;\n}\n\n/**\n * Reasoning tokens xAI reported outside `completion_tokens`. OpenAI counts them\n * inside it; xAI's documented responses do not. Rather than trust one rule for\n * every model, use `total_tokens` when present: only the part of the reasoning\n * count that `total_tokens` shows beyond prompt + completion is added, so a\n * model that nests reasoning inside `completion_tokens` is never double-counted.\n */\nfunction separateReasoningTokens(\n\trawUsage: OpenAICompletionsRawUsage,\n\tpromptTokens: number,\n\tcompletionTokens: number,\n): number {\n\tconst reasoning = count(rawUsage.completion_tokens_details?.reasoning_tokens);\n\tif (typeof rawUsage.total_tokens !== \"number\") return reasoning;\n\treturn Math.min(reasoning, Math.max(0, count(rawUsage.total_tokens) - promptTokens - completionTokens));\n}\n\n/**\n * Map a Chat Completions `usage` object to our usage record.\n *\n * OpenAI counts reasoning tokens inside `completion_tokens`. xAI does not: it\n * reports them only in `completion_tokens_details.reasoning_tokens` (its\n * `total_tokens` is prompt + completion + reasoning) and bills them at the\n * output rate. Without adding them, Grok reasoning runs were under-counted and\n * under-costed, by a lot at high effort. xAI also returns the billed amount as\n * `cost_in_usd_ticks`, which is surfaced as `cost.billed`.\n */\nexport function parseChunkUsage(\n\trawUsage: OpenAICompletionsRawUsage,\n\tmodel: Model<\"openai-completions\">,\n): AssistantMessage[\"usage\"] {\n\tconst promptTokens = count(rawUsage.prompt_tokens);\n\tconst cacheReadTokens = count(rawUsage.prompt_tokens_details?.cached_tokens ?? rawUsage.prompt_cache_hit_tokens);\n\tconst cacheWriteTokens = count(rawUsage.prompt_tokens_details?.cache_write_tokens);\n\n\t// Follow documented OpenAI/OpenRouter semantics: cached_tokens is cache-read\n\t// tokens (hits). OpenAI does not document or emit cache_write_tokens, but\n\t// OpenRouter-compatible providers can include it as a separate write count.\n\t// OpenRouter's own provider/tests affirm the separate mapping:\n\t// https://github.com/OpenRouterTeam/ai-sdk-provider/pull/409\n\t// Do not subtract writes from cached_tokens, otherwise spec-compliant\n\t// providers are under-reported. DS4 mirrors this contract too:\n\t// https://github.com/antirez/ds4/pull/29\n\tconst input = Math.max(0, promptTokens - cacheReadTokens - cacheWriteTokens);\n\tconst xai = isXaiModel(model);\n\tconst completionTokens = count(rawUsage.completion_tokens);\n\tconst reasoningTokens = xai ? separateReasoningTokens(rawUsage, promptTokens, completionTokens) : 0;\n\tconst outputTokens = completionTokens + reasoningTokens;\n\tconst usage: AssistantMessage[\"usage\"] = {\n\t\tinput,\n\t\toutput: outputTokens,\n\t\tcacheRead: cacheReadTokens,\n\t\tcacheWrite: cacheWriteTokens,\n\t\ttotalTokens: input + outputTokens + cacheReadTokens + cacheWriteTokens,\n\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t};\n\tcalculateCost(model, usage);\n\tconst ticks = rawUsage.cost_in_usd_ticks;\n\tif (xai && typeof ticks === \"number\" && Number.isFinite(ticks) && ticks >= 0) usage.cost.billed = ticks / USD_TICKS;\n\treturn usage;\n}\n"]}

package/dist/providers/openai-completions.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"openai-completions.d.ts","sourceRoot":"","sources":["../../src/providers/openai-completions.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAOX,0BAA0B,EAG1B,MAAM,sCAAsC,CAAC;AAE9C,OAAO,KAAK,EAGX,OAAO,EAGP,KAAK,EACL,mBAAmB,EACnB,cAAc,EACd,aAAa,EAMb,MAAM,aAAa,CAAC;AASrB,OAAO,EAAiC,KAAK,+BAA+B,EAAE,MAAM,gCAAgC,CAAC;AA0CrH,MAAM,WAAW,wBAAyB,SAAQ,aAAa;IAC9D,UAAU,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,UAAU,GAAG;QAAE,IAAI,EAAE,UAAU,CAAC;QAAC,QAAQ,EAAE;YAAE,IAAI,EAAE,MAAM,CAAA;SAAE,CAAA;KAAE,CAAC;IAC7F,eAAe,CAAC,EAAE,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,GAAG,KAAK,GAAG,OAAO,CAAC;CACpF;AA2BD,eAAO,MAAM,uBAAuB,EAAE,cAAc,CAAC,oBAAoB,EAAE,wBAAwB,CAgUlG,CAAC;AAEF,eAAO,MAAM,6BAA6B,EAAE,cAAc,CAAC,oBAAoB,EAAE,mBAAmB,CAoBnG,CAAC;AAwTF,wBAAgB,eAAe,CAC9B,KAAK,EAAE,KAAK,CAAC,oBAAoB,CAAC,EAClC,OAAO,EAAE,OAAO,EAChB,MAAM,EAAE,+BAA+B,GACrC,0BAA0B,EAAE,CAyQ9B","sourcesContent":["import OpenAI from \"openai\";\nimport type {\n\tChatCompletionAssistantMessageParam,\n\tChatCompletionChunk,\n\tChatCompletionContentPart,\n\tChatCompletionContentPartImage,\n\tChatCompletionContentPartText,\n\tChatCompletionDeveloperMessageParam,\n\tChatCompletionMessageParam,\n\tChatCompletionSystemMessageParam,\n\tChatCompletionToolMessageParam,\n} from \"openai/resources/chat/completions.js\";\nimport { calculateCost, clampThinkingLevel } from \"../models.ts\";\nimport type {\n\tAssistantMessage,\n\tCacheRetention,\n\tContext,\n\tImageContent,\n\tMessage,\n\tModel,\n\tSimpleStreamOptions,\n\tStreamFunction,\n\tStreamOptions,\n\tTextContent,\n\tThinkingContent,\n\tTool,\n\tToolCall,\n\tToolResultMessage,\n} from \"../types.ts\";\nimport { AssistantMessageEventStream } from \"../utils/event-stream.ts\";\nimport { shortHash } from \"../utils/hash.ts\";\nimport { headersToRecord } from \"../utils/headers.ts\";\nimport { parseStreamingJson } from \"../utils/json-parse.ts\";\nimport { retryProviderRequest } from \"../utils/provider-retry.ts\";\nimport { sanitizeSurrogates } from \"../utils/sanitize-unicode.ts\";\nimport { isCloudflareProvider, resolveCloudflareBaseUrl } from \"./cloudflare.ts\";\nimport { buildCopilotDynamicHeaders, hasCopilotVisionInput } from \"./github-copilot-headers.ts\";\nimport { getCompat, isModelStudioModel, type ResolvedOpenAICompletionsCompat } from \"./openai-completions-compat.ts\";\nimport { resolveOpenAIPromptCacheKey } from \"./openai-prompt-cache.ts\";\nimport { openAICompletionStopReason as mapStopReason } from \"./provider-stop-reasons.ts\";\nimport { buildBaseOptions } from \"./simple-options.ts\";\nimport { stableTools } from \"./tool-schema.ts\";\nimport { transformMessages } from \"./transform-messages.ts\";\n\n/**\n * Check if conversation messages contain tool calls or tool results.\n * This is needed because Anthropic (via proxy) requires the tools param\n * to be present when messages include tool_calls or tool role messages.\n */\nfunction hasToolHistory(messages: Message[]): boolean {\n\tfor (const msg of messages) {\n\t\tif (msg.role === \"toolResult\") {\n\t\t\treturn true;\n\t\t}\n\t\tif (msg.role === \"assistant\") {\n\t\t\tif (msg.content.some((block) => block.type === \"toolCall\")) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t}\n\t}\n\treturn false;\n}\n\nfunction isTextContentBlock(block: { type: string }): block is TextContent {\n\treturn block.type === \"text\";\n}\n\nfunction isThinkingContentBlock(block: { type: string }): block is ThinkingContent {\n\treturn block.type === \"thinking\";\n}\n\nfunction isToolCallBlock(block: { type: string }): block is ToolCall {\n\treturn block.type === \"toolCall\";\n}\n\nfunction isImageContentBlock(block: { type: string }): block is ImageContent {\n\treturn block.type === \"image\";\n}\n\nexport interface OpenAICompletionsOptions extends StreamOptions {\n\ttoolChoice?: \"auto\" | \"none\" | \"required\" | { type: \"function\"; function: { name: string } };\n\treasoningEffort?: \"minimal\" | \"low\" | \"medium\" | \"high\" | \"xhigh\" | \"max\" | \"ultra\";\n}\n\ninterface OpenAICompatCacheControl {\n\ttype: \"ephemeral\";\n\tttl?: string;\n}\n\ntype ChatCompletionInstructionMessageParam = ChatCompletionDeveloperMessageParam | ChatCompletionSystemMessageParam;\n\ntype ChatCompletionTextPartWithCacheControl = ChatCompletionContentPartText & {\n\tcache_control?: OpenAICompatCacheControl;\n};\n\ntype ChatCompletionToolWithCacheControl = OpenAI.Chat.Completions.ChatCompletionTool & {\n\tcache_control?: OpenAICompatCacheControl;\n};\n\nfunction resolveCacheRetention(cacheRetention?: CacheRetention): CacheRetention {\n\tif (cacheRetention) {\n\t\treturn cacheRetention;\n\t}\n\tif (typeof process !== \"undefined\" && process.env.OMK_CACHE_RETENTION === \"long\") {\n\t\treturn \"long\";\n\t}\n\treturn \"short\";\n}\n\nexport const streamOpenAICompletions: StreamFunction<\"openai-completions\", OpenAICompletionsOptions> = (\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\toptions?: OpenAICompletionsOptions,\n): AssistantMessageEventStream => {\n\tconst stream = new AssistantMessageEventStream();\n\n\t(async () => {\n\t\tconst output: AssistantMessage = {\n\t\t\trole: \"assistant\",\n\t\t\tcontent: [],\n\t\t\tapi: model.api,\n\t\t\tprovider: model.provider,\n\t\t\tmodel: model.id,\n\t\t\tusage: {\n\t\t\t\tinput: 0,\n\t\t\t\toutput: 0,\n\t\t\t\tcacheRead: 0,\n\t\t\t\tcacheWrite: 0,\n\t\t\t\ttotalTokens: 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t},\n\t\t\tstopReason: \"stop\",\n\t\t\ttimestamp: Date.now(),\n\t\t};\n\n\t\ttry {\n\t\t\tconst apiKey = options?.apiKey;\n\t\t\tif (!apiKey) {\n\t\t\t\tthrow new Error(`No API key for provider: ${model.provider}`);\n\t\t\t}\n\t\t\tconst compat = getCompat(model);\n\t\t\tconst cacheRetention = resolveCacheRetention(options?.cacheRetention);\n\t\t\tconst cacheSessionId = cacheRetention === \"none\" ? undefined : options?.sessionId;\n\t\t\tconst client = createClient(model, context, apiKey, options?.headers, cacheSessionId, compat);\n\t\t\tlet params = buildParams(model, context, options, compat, cacheRetention);\n\t\t\tconst nextParams = await options?.onPayload?.(params, model);\n\t\t\tif (nextParams !== undefined) {\n\t\t\t\tparams = nextParams as OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming;\n\t\t\t}\n\t\t\tconst requestOptions = {\n\t\t\t\t...(options?.signal ? { signal: options.signal } : {}),\n\t\t\t\t...(options?.timeoutMs !== undefined ? { timeout: options.timeoutMs } : {}),\n\t\t\t\tmaxRetries: 0,\n\t\t\t};\n\t\t\tconst { data: openaiStream, response } = await retryProviderRequest(\n\t\t\t\t() => client.chat.completions.create(params, requestOptions).withResponse(),\n\t\t\t\t{\n\t\t\t\t\tmaxRetries: options?.maxRetries,\n\t\t\t\t\tmaxRetryDelayMs: options?.maxRetryDelayMs,\n\t\t\t\t\tsignal: options?.signal,\n\t\t\t\t},\n\t\t\t);\n\t\t\tawait options?.onResponse?.({ status: response.status, headers: headersToRecord(response.headers) }, model);\n\t\t\tstream.push({ type: \"start\", partial: output });\n\n\t\t\tinterface StreamingToolCallBlock extends ToolCall {\n\t\t\t\tpartialArgs?: string;\n\t\t\t\tstreamIndex?: number;\n\t\t\t}\n\t\t\ttype StreamingBlock = TextContent | ThinkingContent | StreamingToolCallBlock;\n\t\t\ttype StreamingToolCallDelta = NonNullable<ChatCompletionChunk.Choice.Delta[\"tool_calls\"]>[number];\n\n\t\t\tlet textBlock: TextContent | null = null;\n\t\t\tlet thinkingBlock: ThinkingContent | null = null;\n\t\t\tlet hasFinishReason = false;\n\t\t\tconst toolCallBlocksByIndex = new Map<number, StreamingToolCallBlock>();\n\t\t\tconst toolCallBlocksById = new Map<string, StreamingToolCallBlock>();\n\t\t\tconst blocks = output.content as StreamingBlock[];\n\t\t\tconst getContentIndex = (block: StreamingBlock) => blocks.indexOf(block);\n\t\t\tconst finishBlock = (block: StreamingBlock) => {\n\t\t\t\tconst contentIndex = getContentIndex(block);\n\t\t\t\tif (contentIndex === -1) {\n\t\t\t\t\treturn;\n\t\t\t\t}\n\t\t\t\tif (block.type === \"text\") {\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"text_end\",\n\t\t\t\t\t\tcontentIndex,\n\t\t\t\t\t\tcontent: block.text,\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t} else if (block.type === \"thinking\") {\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"thinking_end\",\n\t\t\t\t\t\tcontentIndex,\n\t\t\t\t\t\tcontent: block.thinking,\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\tblock.arguments = parseStreamingJson(block.partialArgs);\n\t\t\t\t\t// Finalize in-place and strip the scratch buffers so replay only\n\t\t\t\t\t// carries parsed arguments.\n\t\t\t\t\tdelete block.partialArgs;\n\t\t\t\t\tdelete block.streamIndex;\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"toolcall_end\",\n\t\t\t\t\t\tcontentIndex,\n\t\t\t\t\t\ttoolCall: block,\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t};\n\t\t\tconst ensureTextBlock = () => {\n\t\t\t\tif (!textBlock) {\n\t\t\t\t\ttextBlock = { type: \"text\", text: \"\" };\n\t\t\t\t\tblocks.push(textBlock);\n\t\t\t\t\tstream.push({ type: \"text_start\", contentIndex: getContentIndex(textBlock), partial: output });\n\t\t\t\t}\n\t\t\t\treturn textBlock;\n\t\t\t};\n\t\t\tconst ensureThinkingBlock = (thinkingSignature: string) => {\n\t\t\t\tif (!thinkingBlock) {\n\t\t\t\t\tthinkingBlock = {\n\t\t\t\t\t\ttype: \"thinking\",\n\t\t\t\t\t\tthinking: \"\",\n\t\t\t\t\t\tthinkingSignature,\n\t\t\t\t\t};\n\t\t\t\t\tblocks.push(thinkingBlock);\n\t\t\t\t\tstream.push({ type: \"thinking_start\", contentIndex: getContentIndex(thinkingBlock), partial: output });\n\t\t\t\t}\n\t\t\t\treturn thinkingBlock;\n\t\t\t};\n\t\t\tconst ensureToolCallBlock = (toolCall: StreamingToolCallDelta) => {\n\t\t\t\tconst streamIndex = typeof toolCall.index === \"number\" ? toolCall.index : undefined;\n\t\t\t\tlet block = streamIndex !== undefined ? toolCallBlocksByIndex.get(streamIndex) : undefined;\n\t\t\t\tif (!block && toolCall.id) {\n\t\t\t\t\tblock = toolCallBlocksById.get(toolCall.id);\n\t\t\t\t}\n\t\t\t\tif (!block) {\n\t\t\t\t\tblock = {\n\t\t\t\t\t\ttype: \"toolCall\",\n\t\t\t\t\t\tid: toolCall.id || \"\",\n\t\t\t\t\t\tname: toolCall.function?.name || \"\",\n\t\t\t\t\t\targuments: {},\n\t\t\t\t\t\tpartialArgs: \"\",\n\t\t\t\t\t\tstreamIndex,\n\t\t\t\t\t};\n\t\t\t\t\tif (streamIndex !== undefined) {\n\t\t\t\t\t\ttoolCallBlocksByIndex.set(streamIndex, block);\n\t\t\t\t\t}\n\t\t\t\t\tif (toolCall.id) {\n\t\t\t\t\t\ttoolCallBlocksById.set(toolCall.id, block);\n\t\t\t\t\t}\n\t\t\t\t\tblocks.push(block);\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"toolcall_start\",\n\t\t\t\t\t\tcontentIndex: getContentIndex(block),\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t\tif (streamIndex !== undefined && block.streamIndex === undefined) {\n\t\t\t\t\tblock.streamIndex = streamIndex;\n\t\t\t\t\ttoolCallBlocksByIndex.set(streamIndex, block);\n\t\t\t\t}\n\t\t\t\tif (toolCall.id) {\n\t\t\t\t\ttoolCallBlocksById.set(toolCall.id, block);\n\t\t\t\t}\n\t\t\t\treturn block;\n\t\t\t};\n\n\t\t\tfor await (const chunk of openaiStream) {\n\t\t\t\tif (!chunk || typeof chunk !== \"object\") continue;\n\n\t\t\t\t// OpenAI documents ChatCompletionChunk.id as the unique chat completion identifier,\n\t\t\t\t// and each chunk in a streamed completion carries the same id.\n\t\t\t\toutput.responseId ||= chunk.id;\n\t\t\t\tif (typeof chunk.model === \"string\" && chunk.model.length > 0 && chunk.model !== model.id) {\n\t\t\t\t\toutput.responseModel ||= chunk.model;\n\t\t\t\t}\n\t\t\t\tif (chunk.usage) {\n\t\t\t\t\toutput.usage = parseChunkUsage(chunk.usage, model);\n\t\t\t\t}\n\n\t\t\t\tconst choice = Array.isArray(chunk.choices) ? chunk.choices[0] : undefined;\n\t\t\t\tif (!choice) continue;\n\n\t\t\t\t// Fallback: some providers (e.g., Moonshot) return usage\n\t\t\t\t// in choice.usage instead of the standard chunk.usage\n\t\t\t\tif (!chunk.usage && (choice as any).usage) {\n\t\t\t\t\toutput.usage = parseChunkUsage((choice as any).usage, model);\n\t\t\t\t}\n\n\t\t\t\tif (choice.finish_reason) {\n\t\t\t\t\tconst finishReasonResult = mapStopReason(choice.finish_reason);\n\t\t\t\t\toutput.stopReason = finishReasonResult.stopReason;\n\t\t\t\t\tif (finishReasonResult.errorMessage) {\n\t\t\t\t\t\toutput.errorMessage = finishReasonResult.errorMessage;\n\t\t\t\t\t}\n\t\t\t\t\thasFinishReason = true;\n\t\t\t\t}\n\n\t\t\t\tif (choice.delta) {\n\t\t\t\t\tif (\n\t\t\t\t\t\tchoice.delta.content !== null &&\n\t\t\t\t\t\tchoice.delta.content !== undefined &&\n\t\t\t\t\t\tchoice.delta.content.length > 0\n\t\t\t\t\t) {\n\t\t\t\t\t\tconst block = ensureTextBlock();\n\t\t\t\t\t\tblock.text += choice.delta.content;\n\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\ttype: \"text_delta\",\n\t\t\t\t\t\t\tcontentIndex: getContentIndex(block),\n\t\t\t\t\t\t\tdelta: choice.delta.content,\n\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t});\n\t\t\t\t\t}\n\n\t\t\t\t\t// Some endpoints return reasoning in reasoning_content (llama.cpp),\n\t\t\t\t\t// or reasoning (other openai compatible endpoints)\n\t\t\t\t\t// Use the first non-empty reasoning field to avoid duplication\n\t\t\t\t\t// (e.g., chutes.ai returns both reasoning_content and reasoning with same content)\n\t\t\t\t\tconst reasoningFields = [\"reasoning_content\", \"reasoning\", \"reasoning_text\"];\n\t\t\t\t\tconst deltaFields = choice.delta as Record<string, unknown>;\n\t\t\t\t\tlet foundReasoningField: string | null = null;\n\t\t\t\t\tfor (const field of reasoningFields) {\n\t\t\t\t\t\tconst value = deltaFields[field];\n\t\t\t\t\t\tif (typeof value === \"string\" && value.length > 0) {\n\t\t\t\t\t\t\tfoundReasoningField = field;\n\t\t\t\t\t\t\tbreak;\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\tif (foundReasoningField) {\n\t\t\t\t\t\tconst delta = deltaFields[foundReasoningField];\n\t\t\t\t\t\tif (typeof delta === \"string\" && delta.length > 0) {\n\t\t\t\t\t\t\tconst thinkingSignature =\n\t\t\t\t\t\t\t\tmodel.provider === \"opencode-go\" && foundReasoningField === \"reasoning\"\n\t\t\t\t\t\t\t\t\t? \"reasoning_content\"\n\t\t\t\t\t\t\t\t\t: foundReasoningField;\n\t\t\t\t\t\t\tconst block = ensureThinkingBlock(thinkingSignature);\n\t\t\t\t\t\t\tblock.thinking += delta;\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\t\t\t\t\tcontentIndex: getContentIndex(block),\n\t\t\t\t\t\t\t\tdelta,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\tif (choice?.delta?.tool_calls) {\n\t\t\t\t\t\tfor (const toolCall of choice.delta.tool_calls) {\n\t\t\t\t\t\t\tconst block = ensureToolCallBlock(toolCall);\n\t\t\t\t\t\t\tif (!block.id && toolCall.id) {\n\t\t\t\t\t\t\t\tblock.id = toolCall.id;\n\t\t\t\t\t\t\t\ttoolCallBlocksById.set(toolCall.id, block);\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\tif (!block.name && toolCall.function?.name) {\n\t\t\t\t\t\t\t\tblock.name = toolCall.function.name;\n\t\t\t\t\t\t\t}\n\n\t\t\t\t\t\t\tlet delta = \"\";\n\t\t\t\t\t\t\tif (toolCall.function?.arguments) {\n\t\t\t\t\t\t\t\tdelta = toolCall.function.arguments;\n\t\t\t\t\t\t\t\tblock.partialArgs = (block.partialArgs ?? \"\") + toolCall.function.arguments;\n\t\t\t\t\t\t\t\tblock.arguments = parseStreamingJson(block.partialArgs);\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"toolcall_delta\",\n\t\t\t\t\t\t\t\tcontentIndex: getContentIndex(block),\n\t\t\t\t\t\t\t\tdelta,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\tconst reasoningDetails = (choice.delta as any).reasoning_details;\n\t\t\t\t\tif (reasoningDetails && Array.isArray(reasoningDetails)) {\n\t\t\t\t\t\tfor (const detail of reasoningDetails) {\n\t\t\t\t\t\t\tif (detail.type === \"reasoning.encrypted\" && detail.id && detail.data) {\n\t\t\t\t\t\t\t\tconst matchingToolCall = output.content.find(\n\t\t\t\t\t\t\t\t\t(b) => b.type === \"toolCall\" && b.id === detail.id,\n\t\t\t\t\t\t\t\t) as ToolCall | undefined;\n\t\t\t\t\t\t\t\tif (matchingToolCall) {\n\t\t\t\t\t\t\t\t\tmatchingToolCall.thoughtSignature = JSON.stringify(detail);\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tfor (const block of blocks) {\n\t\t\t\tfinishBlock(block);\n\t\t\t}\n\t\t\tif (options?.signal?.aborted) {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\n\t\t\tif (output.stopReason === \"aborted\") {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\t\t\tif (output.stopReason === \"error\") {\n\t\t\t\tthrow new Error(output.errorMessage || \"Provider returned an error stop reason\");\n\t\t\t}\n\t\t\tif (!hasFinishReason) {\n\t\t\t\tthrow new Error(\"Stream ended without finish_reason\");\n\t\t\t}\n\n\t\t\tstream.push({ type: \"done\", reason: output.stopReason, message: output });\n\t\t\tstream.end();\n\t\t} catch (error) {\n\t\t\tfor (const block of output.content) {\n\t\t\t\tdelete (block as { index?: number }).index;\n\t\t\t\t// Streaming scratch buffers are only used during parsing; never persist them.\n\t\t\t\tdelete (block as { partialArgs?: string }).partialArgs;\n\t\t\t\tdelete (block as { streamIndex?: number }).streamIndex;\n\t\t\t}\n\t\t\toutput.stopReason = options?.signal?.aborted ? \"aborted\" : \"error\";\n\t\t\toutput.errorMessage = error instanceof Error ? error.message : JSON.stringify(error);\n\t\t\t// Some providers via OpenRouter give additional information in this field.\n\t\t\tconst rawMetadata = (error as any)?.error?.metadata?.raw;\n\t\t\tif (rawMetadata) output.errorMessage += `\\n${rawMetadata}`;\n\t\t\tstream.push({ type: \"error\", reason: output.stopReason, error: output });\n\t\t\tstream.end();\n\t\t}\n\t})();\n\n\treturn stream;\n};\n\nexport const streamSimpleOpenAICompletions: StreamFunction<\"openai-completions\", SimpleStreamOptions> = (\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): AssistantMessageEventStream => {\n\tconst apiKey = options?.apiKey;\n\tif (!apiKey) {\n\t\tthrow new Error(`No API key for provider: ${model.provider}`);\n\t}\n\n\tconst base = buildBaseOptions(model, options, apiKey);\n\tconst clampedReasoning = options?.reasoning ? clampThinkingLevel(model, options.reasoning) : undefined;\n\tconst reasoningEffort = clampedReasoning === \"off\" ? undefined : clampedReasoning;\n\tconst toolChoice = (options as OpenAICompletionsOptions | undefined)?.toolChoice;\n\n\treturn streamOpenAICompletions(model, context, {\n\t\t...base,\n\t\treasoningEffort,\n\t\ttoolChoice,\n\t} satisfies OpenAICompletionsOptions);\n};\n\nfunction createClient(\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\tapiKey: string,\n\toptionsHeaders?: Record<string, string>,\n\tsessionId?: string,\n\tcompat: ResolvedOpenAICompletionsCompat = getCompat(model),\n) {\n\tconst headers = { ...model.headers };\n\tif (model.provider === \"github-copilot\") {\n\t\tconst hasImages = hasCopilotVisionInput(context.messages);\n\t\tconst copilotHeaders = buildCopilotDynamicHeaders({\n\t\t\tmessages: context.messages,\n\t\t\thasImages,\n\t\t});\n\t\tObject.assign(headers, copilotHeaders);\n\t}\n\n\tif (sessionId && compat.sendSessionAffinityHeaders) {\n\t\theaders.session_id = sessionId;\n\t\theaders[\"x-client-request-id\"] = sessionId;\n\t\theaders[\"x-session-affinity\"] = sessionId;\n\t}\n\n\t// Merge options headers last so they can override defaults\n\tif (optionsHeaders) {\n\t\tObject.assign(headers, optionsHeaders);\n\t}\n\n\tconst defaultHeaders =\n\t\tmodel.provider === \"cloudflare-ai-gateway\"\n\t\t\t? {\n\t\t\t\t\t...headers,\n\t\t\t\t\tAuthorization: headers.Authorization ?? null,\n\t\t\t\t\t\"cf-aig-authorization\": `Bearer ${apiKey}`,\n\t\t\t\t}\n\t\t\t: headers;\n\n\treturn new OpenAI({\n\t\tapiKey,\n\t\tbaseURL: isCloudflareProvider(model.provider) ? resolveCloudflareBaseUrl(model) : model.baseUrl,\n\t\tdangerouslyAllowBrowser: true,\n\t\tdefaultHeaders,\n\t});\n}\n\nfunction buildParams(\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\toptions?: OpenAICompletionsOptions,\n\tcompat: ResolvedOpenAICompletionsCompat = getCompat(model),\n\tcacheRetention: CacheRetention = resolveCacheRetention(options?.cacheRetention),\n) {\n\tconst messages = convertMessages(model, context, compat);\n\tconst cacheControl = getCompatCacheControl(compat, cacheRetention);\n\n\tconst params: OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming = {\n\t\tmodel: model.provider === \"zyloo\" && !model.id.startsWith(\"zyloo/\") ? `zyloo/${model.id}` : model.id,\n\t\tmessages,\n\t\tstream: true,\n\t\tprompt_cache_key:\n\t\t\t(model.baseUrl.includes(\"api.openai.com\") && cacheRetention !== \"none\") ||\n\t\t\t(cacheRetention === \"long\" && compat.supportsLongCacheRetention)\n\t\t\t\t? resolveOpenAIPromptCacheKey(context, options?.sessionId, `${model.provider}/${model.id}`)\n\t\t\t\t: undefined,\n\t\tprompt_cache_retention: cacheRetention === \"long\" && compat.supportsLongCacheRetention ? \"24h\" : undefined,\n\t};\n\n\tif (compat.supportsUsageInStreaming !== false) {\n\t\t(params as any).stream_options = { include_usage: true };\n\t}\n\n\tif (compat.supportsStore) {\n\t\tparams.store = false;\n\t}\n\n\tif (options?.maxTokens) {\n\t\tif (compat.maxTokensField === \"max_tokens\") {\n\t\t\t(params as any).max_tokens = options.maxTokens;\n\t\t} else {\n\t\t\tparams.max_completion_tokens = options.maxTokens;\n\t\t}\n\t}\n\n\tif (options?.temperature !== undefined) {\n\t\tparams.temperature = options.temperature;\n\t}\n\n\tif (context.tools && context.tools.length > 0) {\n\t\tparams.tools = convertTools(context.tools, compat);\n\t\tif (compat.zaiToolStream) {\n\t\t\t(params as any).tool_stream = true;\n\t\t}\n\t} else if (hasToolHistory(context.messages)) {\n\t\t// Anthropic (via LiteLLM/proxy) requires tools param when conversation has tool_calls/tool_results\n\t\tparams.tools = [];\n\t}\n\n\tif (cacheControl) {\n\t\tapplyAnthropicCacheControl(messages, params.tools, cacheControl);\n\t}\n\n\tif (options?.toolChoice) {\n\t\tparams.tool_choice = options.toolChoice;\n\t}\n\n\tif (compat.thinkingFormat === \"zai\" && model.reasoning) {\n\t\t(params as any).enable_thinking = !!options?.reasoningEffort;\n\t\tif (options?.reasoningEffort && compat.supportsReasoningEffort) {\n\t\t\t(params as any).reasoning_effort =\n\t\t\t\tmodel.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;\n\t\t}\n\t} else if (compat.thinkingFormat === \"qwen\" && model.reasoning) {\n\t\tObject.assign(params, { enable_thinking: !!options?.reasoningEffort });\n\t\tif (options?.reasoningEffort && compat.supportsReasoningEffort && isModelStudioModel(model)) {\n\t\t\tconst effort =\n\t\t\t\toptions.reasoningEffort === \"minimal\" && model.id.startsWith(\"deepseek-v4-\")\n\t\t\t\t\t? \"low\"\n\t\t\t\t\t: options.reasoningEffort;\n\t\t\tObject.assign(params, {\n\t\t\t\treasoning_effort: model.thinkingLevelMap?.[options.reasoningEffort] ?? effort,\n\t\t\t});\n\t\t}\n\t} else if (compat.thinkingFormat === \"qwen-chat-template\" && model.reasoning) {\n\t\t(params as any).chat_template_kwargs = {\n\t\t\tenable_thinking: !!options?.reasoningEffort,\n\t\t\tpreserve_thinking: true,\n\t\t};\n\t} else if (compat.thinkingFormat === \"deepseek\" && model.reasoning) {\n\t\t(params as any).thinking = { type: options?.reasoningEffort ? \"enabled\" : \"disabled\" };\n\t\tif (options?.reasoningEffort && compat.supportsReasoningEffort) {\n\t\t\t(params as any).reasoning_effort =\n\t\t\t\tmodel.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;\n\t\t}\n\t} else if (compat.thinkingFormat === \"openrouter\" && model.reasoning) {\n\t\t// OpenRouter normalizes reasoning across providers via a nested reasoning object.\n\t\tconst openRouterParams = params as typeof params & { reasoning?: { effort?: string; enabled?: boolean } };\n\t\tif (options?.reasoningEffort) {\n\t\t\topenRouterParams.reasoning = {\n\t\t\t\teffort: model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort,\n\t\t\t};\n\t\t} else if (model.thinkingLevelMap?.off !== null) {\n\t\t\tconst offValue = model.thinkingLevelMap?.off ?? \"none\";\n\t\t\topenRouterParams.reasoning = offValue === \"none\" ? { enabled: false } : { effort: offValue };\n\t\t}\n\t} else if (compat.thinkingFormat === \"ant-ling\" && model.reasoning && options?.reasoningEffort) {\n\t\tconst effort = model.thinkingLevelMap?.[options.reasoningEffort];\n\t\tif (typeof effort === \"string\") {\n\t\t\t(params as typeof params & { reasoning?: { effort: string } }).reasoning = { effort };\n\t\t}\n\t} else if (compat.thinkingFormat === \"together\" && model.reasoning) {\n\t\tconst togetherParams = params as Omit<typeof params, \"reasoning_effort\"> & {\n\t\t\treasoning?: { enabled: boolean };\n\t\t\treasoning_effort?: string;\n\t\t};\n\t\ttogetherParams.reasoning = { enabled: !!options?.reasoningEffort };\n\t\tif (options?.reasoningEffort && compat.supportsReasoningEffort) {\n\t\t\ttogetherParams.reasoning_effort = model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;\n\t\t}\n\t} else if (compat.thinkingFormat === \"string-thinking\" && model.reasoning) {\n\t\tconst stringThinkingParams = params as typeof params & { thinking?: string };\n\t\tif (options?.reasoningEffort) {\n\t\t\tstringThinkingParams.thinking = model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;\n\t\t} else if (model.thinkingLevelMap?.off !== null) {\n\t\t\tstringThinkingParams.thinking = model.thinkingLevelMap?.off ?? \"none\";\n\t\t}\n\t} else if (options?.reasoningEffort && model.reasoning && compat.supportsReasoningEffort) {\n\t\t// OpenAI-style reasoning_effort\n\t\t(params as any).reasoning_effort = model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;\n\t} else if (!options?.reasoningEffort && model.reasoning && compat.supportsReasoningEffort) {\n\t\tconst offValue = model.thinkingLevelMap?.off;\n\t\tif (typeof offValue === \"string\") {\n\t\t\t(params as any).reasoning_effort = offValue;\n\t\t}\n\t}\n\n\t// OpenRouter provider routing preferences\n\tif (model.baseUrl.includes(\"openrouter.ai\") && model.compat?.openRouterRouting) {\n\t\t(params as any).provider = model.compat.openRouterRouting;\n\t}\n\n\t// Vercel AI Gateway provider routing preferences\n\tif (model.baseUrl.includes(\"ai-gateway.vercel.sh\") && model.compat?.vercelGatewayRouting) {\n\t\tconst routing = model.compat.vercelGatewayRouting;\n\t\tif (routing.only || routing.order) {\n\t\t\tconst gatewayOptions: Record<string, string[]> = {};\n\t\t\tif (routing.only) gatewayOptions.only = routing.only;\n\t\t\tif (routing.order) gatewayOptions.order = routing.order;\n\t\t\t(params as any).providerOptions = { gateway: gatewayOptions };\n\t\t}\n\t}\n\n\treturn params;\n}\n\nfunction getCompatCacheControl(\n\tcompat: ResolvedOpenAICompletionsCompat,\n\tcacheRetention: CacheRetention,\n): OpenAICompatCacheControl | undefined {\n\tif (compat.cacheControlFormat !== \"anthropic\" || cacheRetention === \"none\") {\n\t\treturn undefined;\n\t}\n\n\tconst ttl = cacheRetention === \"long\" && compat.supportsLongCacheRetention ? \"1h\" : undefined;\n\treturn { type: \"ephemeral\", ...(ttl ? { ttl } : {}) };\n}\n\nfunction applyAnthropicCacheControl(\n\tmessages: ChatCompletionMessageParam[],\n\ttools: OpenAI.Chat.Completions.ChatCompletionTool[] | undefined,\n\tcacheControl: OpenAICompatCacheControl,\n): void {\n\taddCacheControlToSystemPrompt(messages, cacheControl);\n\taddCacheControlToLastTool(tools, cacheControl);\n\taddCacheControlToLastConversationMessage(messages, cacheControl);\n}\n\nfunction addCacheControlToSystemPrompt(\n\tmessages: ChatCompletionMessageParam[],\n\tcacheControl: OpenAICompatCacheControl,\n): void {\n\tfor (const message of messages) {\n\t\tif (message.role === \"system\" || message.role === \"developer\") {\n\t\t\taddCacheControlToInstructionMessage(message, cacheControl);\n\t\t\treturn;\n\t\t}\n\t}\n}\n\nfunction addCacheControlToLastConversationMessage(\n\tmessages: ChatCompletionMessageParam[],\n\tcacheControl: OpenAICompatCacheControl,\n): void {\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst message = messages[i];\n\t\tif (message.role === \"user\" || message.role === \"assistant\") {\n\t\t\tif (addCacheControlToMessage(message, cacheControl)) {\n\t\t\t\treturn;\n\t\t\t}\n\t\t}\n\t}\n}\n\nfunction addCacheControlToLastTool(\n\ttools: OpenAI.Chat.Completions.ChatCompletionTool[] | undefined,\n\tcacheControl: OpenAICompatCacheControl,\n): void {\n\tif (!tools || tools.length === 0) {\n\t\treturn;\n\t}\n\n\tconst lastTool = tools[tools.length - 1] as ChatCompletionToolWithCacheControl;\n\tlastTool.cache_control = cacheControl;\n}\n\nfunction addCacheControlToInstructionMessage(\n\tmessage: ChatCompletionInstructionMessageParam,\n\tcacheControl: OpenAICompatCacheControl,\n): boolean {\n\treturn addCacheControlToTextContent(message, cacheControl);\n}\n\nfunction addCacheControlToMessage(\n\tmessage: ChatCompletionMessageParam,\n\tcacheControl: OpenAICompatCacheControl,\n): boolean {\n\tif (message.role === \"user\" || message.role === \"assistant\") {\n\t\treturn addCacheControlToTextContent(message, cacheControl);\n\t}\n\treturn false;\n}\n\nfunction addCacheControlToTextContent(\n\tmessage:\n\t\t| ChatCompletionInstructionMessageParam\n\t\t| ChatCompletionAssistantMessageParam\n\t\t| Extract<ChatCompletionMessageParam, { role: \"user\" }>,\n\tcacheControl: OpenAICompatCacheControl,\n): boolean {\n\tconst content = message.content;\n\tif (typeof content === \"string\") {\n\t\tif (content.length === 0) {\n\t\t\treturn false;\n\t\t}\n\t\tmessage.content = [\n\t\t\t{\n\t\t\t\ttype: \"text\",\n\t\t\t\ttext: content,\n\t\t\t\tcache_control: cacheControl,\n\t\t\t},\n\t\t] as ChatCompletionTextPartWithCacheControl[];\n\t\treturn true;\n\t}\n\n\tif (!Array.isArray(content)) {\n\t\treturn false;\n\t}\n\n\tfor (let i = content.length - 1; i >= 0; i--) {\n\t\tconst part = content[i];\n\t\tif (part?.type === \"text\") {\n\t\t\tconst textPart = part as ChatCompletionTextPartWithCacheControl;\n\t\t\ttextPart.cache_control = cacheControl;\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\nexport function convertMessages(\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\tcompat: ResolvedOpenAICompletionsCompat,\n): ChatCompletionMessageParam[] {\n\tconst params: ChatCompletionMessageParam[] = [];\n\n\tconst normalizeToolCallId = (id: string): string => {\n\t\t// Handle pipe-separated IDs from OpenAI Responses API\n\t\t// Format: {call_id}|{id} where {id} can be 400+ chars with special chars (+, /, =)\n\t\t// These come from providers like github-copilot, openai-codex, opencode\n\t\t// Extract just the call_id part and normalize it\n\t\t// Multiple tool calls in the same turn can share call_id but differ by item_id.\n\t\t// Preserve item-level uniqueness when replaying into Chat Completions, which\n\t\t// requires distinct tool call ids.\n\t\tif (id.includes(\"|\")) {\n\t\t\t// Sanitize to allowed chars and truncate to 40 chars (OpenAI limit)\n\t\t\tconst separatorIndex = id.indexOf(\"|\");\n\t\t\tconst callId = id.slice(0, separatorIndex).replace(/[^a-zA-Z0-9_-]/g, \"_\");\n\t\t\tconst itemId = id.slice(separatorIndex + 1).replace(/[^a-zA-Z0-9_-]/g, \"_\");\n\t\t\tconst combinedId = itemId.length > 0 ? `${callId}_${itemId}` : callId;\n\t\t\tif (combinedId.length <= 40) {\n\t\t\t\treturn combinedId;\n\t\t\t}\n\t\t\tconst hash = shortHash(id).slice(0, 8);\n\t\t\tconst prefix = callId.slice(0, Math.max(1, 40 - hash.length - 1));\n\t\t\treturn `${prefix}_${hash}`;\n\t\t}\n\n\t\tif (model.provider === \"openai\") return id.length > 40 ? id.slice(0, 40) : id;\n\t\treturn id;\n\t};\n\n\tconst transformedMessages = transformMessages(context.messages, model, (id) => normalizeToolCallId(id));\n\n\t// WorkBuddy rejects a leading user message (`400 11128 \"first message is not\n\t// system prompt\"`) but accepts an empty system message, so a caller with no\n\t// system prompt still opens the channel instead of being rejected upstream.\n\tconst systemPrompt = context.systemPrompt || (compat.requiresSystemMessageFirst ? \"\" : undefined);\n\tif (systemPrompt !== undefined) {\n\t\tconst role = model.reasoning && compat.supportsDeveloperRole ? \"developer\" : \"system\";\n\t\tparams.push({ role, content: sanitizeSurrogates(systemPrompt) });\n\t}\n\n\tlet lastRole: string | null = null;\n\n\tfor (let i = 0; i < transformedMessages.length; i++) {\n\t\tconst msg = transformedMessages[i];\n\t\t// Some providers don't allow user messages directly after tool results\n\t\t// Insert a synthetic assistant message to bridge the gap\n\t\tif (compat.requiresAssistantAfterToolResult && lastRole === \"toolResult\" && msg.role === \"user\") {\n\t\t\tparams.push({\n\t\t\t\trole: \"assistant\",\n\t\t\t\tcontent: \"I have processed the tool results.\",\n\t\t\t});\n\t\t}\n\n\t\tif (msg.role === \"user\") {\n\t\t\tif (typeof msg.content === \"string\") {\n\t\t\t\tparams.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent: sanitizeSurrogates(msg.content),\n\t\t\t\t});\n\t\t\t} else {\n\t\t\t\tconst supportsImages = model.input.includes(\"image\");\n\t\t\t\tconst content: ChatCompletionContentPart[] = [];\n\t\t\t\tlet omittedImages = false;\n\t\t\t\tfor (const item of msg.content) {\n\t\t\t\t\tif (item.type === \"text\") {\n\t\t\t\t\t\tcontent.push({\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: sanitizeSurrogates(item.text),\n\t\t\t\t\t\t} satisfies ChatCompletionContentPartText);\n\t\t\t\t\t} else if (supportsImages) {\n\t\t\t\t\t\tcontent.push({\n\t\t\t\t\t\t\ttype: \"image_url\",\n\t\t\t\t\t\t\timage_url: {\n\t\t\t\t\t\t\t\turl: `data:${item.mimeType};base64,${item.data}`,\n\t\t\t\t\t\t\t},\n\t\t\t\t\t\t} satisfies ChatCompletionContentPartImage);\n\t\t\t\t\t} else {\n\t\t\t\t\t\tomittedImages = true;\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t\tif (omittedImages) {\n\t\t\t\t\tcontent.push({\n\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\ttext: \"[Image]\",\n\t\t\t\t\t} satisfies ChatCompletionContentPartText);\n\t\t\t\t}\n\t\t\t\tif (content.length === 0) continue;\n\t\t\t\tparams.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent,\n\t\t\t\t});\n\t\t\t}\n\t\t} else if (msg.role === \"assistant\") {\n\t\t\t// Some providers don't accept null content, use empty string instead\n\t\t\tconst assistantMsg: ChatCompletionAssistantMessageParam = {\n\t\t\t\trole: \"assistant\",\n\t\t\t\tcontent: compat.requiresAssistantAfterToolResult ? \"\" : null,\n\t\t\t};\n\n\t\t\tconst assistantTextParts = msg.content\n\t\t\t\t.filter(isTextContentBlock)\n\t\t\t\t.filter((block) => block.text.trim().length > 0)\n\t\t\t\t.map(\n\t\t\t\t\t(block) =>\n\t\t\t\t\t\t({\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: sanitizeSurrogates(block.text),\n\t\t\t\t\t\t}) satisfies ChatCompletionContentPartText,\n\t\t\t\t);\n\t\t\tconst assistantText = assistantTextParts.map((part) => part.text).join(\"\");\n\n\t\t\tconst nonEmptyThinkingBlocks = msg.content\n\t\t\t\t.filter(isThinkingContentBlock)\n\t\t\t\t.filter((block) => block.thinking.trim().length > 0);\n\t\t\tif (nonEmptyThinkingBlocks.length > 0) {\n\t\t\t\tif (compat.requiresThinkingAsText) {\n\t\t\t\t\t// Convert thinking blocks to plain text (no tags to avoid model mimicking them)\n\t\t\t\t\tconst thinkingText = nonEmptyThinkingBlocks\n\t\t\t\t\t\t.map((block) => sanitizeSurrogates(block.thinking))\n\t\t\t\t\t\t.join(\"\\n\\n\");\n\t\t\t\t\tassistantMsg.content = [{ type: \"text\", text: thinkingText }, ...assistantTextParts];\n\t\t\t\t} else {\n\t\t\t\t\t// Always send assistant content as a plain string (OpenAI Chat Completions\n\t\t\t\t\t// API standard format). Sending as an array of {type:\"text\", text:\"...\"}\n\t\t\t\t\t// objects is non-standard and causes some models (e.g. DeepSeek V3.2 via\n\t\t\t\t\t// NVIDIA NIM) to mirror the content-block structure literally in their\n\t\t\t\t\t// output, producing recursive nesting like [{'type':'text','text':'[{...}]'}].\n\t\t\t\t\tif (assistantText.length > 0) {\n\t\t\t\t\t\tassistantMsg.content = assistantText;\n\t\t\t\t\t}\n\n\t\t\t\t\t// Use the signature from the first thinking block if available (for llama.cpp server + gpt-oss)\n\t\t\t\t\tlet signature = nonEmptyThinkingBlocks[0].thinkingSignature;\n\t\t\t\t\tif (model.provider === \"opencode-go\" && signature === \"reasoning\") {\n\t\t\t\t\t\tsignature = \"reasoning_content\";\n\t\t\t\t\t}\n\t\t\t\t\tif (signature && signature.length > 0) {\n\t\t\t\t\t\t(assistantMsg as any)[signature] = nonEmptyThinkingBlocks.map((block) => block.thinking).join(\"\\n\");\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t} else if (assistantText.length > 0) {\n\t\t\t\t// Always send assistant content as a plain string (OpenAI Chat Completions\n\t\t\t\t// API standard format). Sending as an array of {type:\"text\", text:\"...\"}\n\t\t\t\t// objects is non-standard and causes some models (e.g. DeepSeek V3.2 via\n\t\t\t\t// NVIDIA NIM) to mirror the content-block structure literally in their\n\t\t\t\t// output, producing recursive nesting like [{'type':'text','text':'[{...}]'}].\n\t\t\t\tassistantMsg.content = assistantText;\n\t\t\t}\n\n\t\t\tconst toolCalls = msg.content.filter(isToolCallBlock);\n\t\t\tif (toolCalls.length > 0) {\n\t\t\t\tassistantMsg.tool_calls = toolCalls.map((tc) => ({\n\t\t\t\t\tid: tc.id,\n\t\t\t\t\ttype: \"function\" as const,\n\t\t\t\t\tfunction: {\n\t\t\t\t\t\tname: tc.name,\n\t\t\t\t\t\targuments: JSON.stringify(tc.arguments),\n\t\t\t\t\t},\n\t\t\t\t}));\n\t\t\t\tconst reasoningDetails = toolCalls\n\t\t\t\t\t.filter((tc) => tc.thoughtSignature)\n\t\t\t\t\t.map((tc) => {\n\t\t\t\t\t\ttry {\n\t\t\t\t\t\t\treturn JSON.parse(tc.thoughtSignature!);\n\t\t\t\t\t\t} catch {\n\t\t\t\t\t\t\treturn null;\n\t\t\t\t\t\t}\n\t\t\t\t\t})\n\t\t\t\t\t.filter(Boolean);\n\t\t\t\tif (reasoningDetails.length > 0) {\n\t\t\t\t\t(assistantMsg as any).reasoning_details = reasoningDetails;\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (\n\t\t\t\tcompat.requiresReasoningContentOnAssistantMessages &&\n\t\t\t\tmodel.reasoning &&\n\t\t\t\t(assistantMsg as { reasoning_content?: string }).reasoning_content === undefined\n\t\t\t) {\n\t\t\t\t(assistantMsg as { reasoning_content?: string }).reasoning_content = \"\";\n\t\t\t}\n\t\t\t// Skip assistant messages that have no content and no tool calls.\n\t\t\t// Some providers require \"either content or tool_calls, but not none\".\n\t\t\t// Other providers also don't accept empty assistant messages.\n\t\t\t// This handles aborted assistant responses that got no content.\n\t\t\tconst content = assistantMsg.content;\n\t\t\tconst hasContent =\n\t\t\t\tcontent !== null &&\n\t\t\t\tcontent !== undefined &&\n\t\t\t\t(typeof content === \"string\" ? content.length > 0 : content.length > 0);\n\t\t\tif (!hasContent && !assistantMsg.tool_calls) {\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tparams.push(assistantMsg);\n\t\t} else if (msg.role === \"toolResult\") {\n\t\t\tconst imageBlocks: Array<{ type: \"image_url\"; image_url: { url: string } }> = [];\n\t\t\tlet j = i;\n\n\t\t\tfor (; j < transformedMessages.length && transformedMessages[j].role === \"toolResult\"; j++) {\n\t\t\t\tconst toolMsg = transformedMessages[j] as ToolResultMessage;\n\n\t\t\t\t// Extract text and image content\n\t\t\t\tconst textResult = toolMsg.content\n\t\t\t\t\t.filter(isTextContentBlock)\n\t\t\t\t\t.map((block) => block.text)\n\t\t\t\t\t.join(\"\\n\");\n\t\t\t\tconst hasImages = toolMsg.content.some((c) => c.type === \"image\");\n\n\t\t\t\t// Always send tool result with text (or placeholder if only images)\n\t\t\t\tconst hasText = textResult.length > 0;\n\t\t\t\t// Some providers require the 'name' field in tool results\n\t\t\t\tconst toolResultMsg: ChatCompletionToolMessageParam = {\n\t\t\t\t\trole: \"tool\",\n\t\t\t\t\tcontent: sanitizeSurrogates(hasText ? textResult : \"(see attached image)\"),\n\t\t\t\t\ttool_call_id: toolMsg.toolCallId,\n\t\t\t\t};\n\t\t\t\tif (compat.requiresToolResultName && toolMsg.toolName) {\n\t\t\t\t\t(toolResultMsg as any).name = toolMsg.toolName;\n\t\t\t\t}\n\t\t\t\tparams.push(toolResultMsg);\n\n\t\t\t\tif (hasImages && model.input.includes(\"image\")) {\n\t\t\t\t\tfor (const block of toolMsg.content) {\n\t\t\t\t\t\tif (isImageContentBlock(block)) {\n\t\t\t\t\t\t\timageBlocks.push({\n\t\t\t\t\t\t\t\ttype: \"image_url\",\n\t\t\t\t\t\t\t\timage_url: {\n\t\t\t\t\t\t\t\t\turl: `data:${block.mimeType};base64,${block.data}`,\n\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\ti = j - 1;\n\n\t\t\tif (imageBlocks.length > 0) {\n\t\t\t\tif (compat.requiresAssistantAfterToolResult) {\n\t\t\t\t\tparams.push({\n\t\t\t\t\t\trole: \"assistant\",\n\t\t\t\t\t\tcontent: \"I have processed the tool results.\",\n\t\t\t\t\t});\n\t\t\t\t}\n\n\t\t\t\tparams.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent: [\n\t\t\t\t\t\t{\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: \"Attached image(s) from tool result:\",\n\t\t\t\t\t\t},\n\t\t\t\t\t\t...imageBlocks,\n\t\t\t\t\t],\n\t\t\t\t});\n\t\t\t\tlastRole = \"user\";\n\t\t\t} else {\n\t\t\t\tlastRole = \"toolResult\";\n\t\t\t}\n\t\t\tcontinue;\n\t\t}\n\n\t\tlastRole = msg.role;\n\t}\n\n\treturn params;\n}\n\nfunction convertTools(\n\ttools: Tool[],\n\tcompat: ResolvedOpenAICompletionsCompat,\n): OpenAI.Chat.Completions.ChatCompletionTool[] {\n\treturn stableTools(tools).map((tool) => ({\n\t\ttype: \"function\",\n\t\tfunction: {\n\t\t\tname: tool.name,\n\t\t\tdescription: tool.description,\n\t\t\tparameters: tool.parameters as any, // TypeBox already generates JSON Schema\n\t\t\t// Only include strict if provider supports it. Some reject unknown fields.\n\t\t\t...(compat.supportsStrictMode !== false && { strict: false }),\n\t\t},\n\t}));\n}\n\nfunction parseChunkUsage(\n\trawUsage: {\n\t\tprompt_tokens?: number;\n\t\tcompletion_tokens?: number;\n\t\tprompt_cache_hit_tokens?: number;\n\t\tprompt_tokens_details?: { cached_tokens?: number; cache_write_tokens?: number };\n\t},\n\tmodel: Model<\"openai-completions\">,\n): AssistantMessage[\"usage\"] {\n\tconst promptTokens = rawUsage.prompt_tokens || 0;\n\tconst cacheReadTokens = rawUsage.prompt_tokens_details?.cached_tokens ?? rawUsage.prompt_cache_hit_tokens ?? 0;\n\tconst cacheWriteTokens = rawUsage.prompt_tokens_details?.cache_write_tokens || 0;\n\n\t// Follow documented OpenAI/OpenRouter semantics: cached_tokens is cache-read\n\t// tokens (hits). OpenAI does not document or emit cache_write_tokens, but\n\t// OpenRouter-compatible providers can include it as a separate write count.\n\t// OpenRouter's own provider/tests affirm the separate mapping:\n\t// https://github.com/OpenRouterTeam/ai-sdk-provider/pull/409\n\t// Do not subtract writes from cached_tokens, otherwise spec-compliant\n\t// providers are under-reported. DS4 mirrors this contract too:\n\t// https://github.com/antirez/ds4/pull/29\n\tconst input = Math.max(0, promptTokens - cacheReadTokens - cacheWriteTokens);\n\t// OpenAI completion_tokens already includes reasoning_tokens.\n\tconst outputTokens = rawUsage.completion_tokens || 0;\n\tconst usage: AssistantMessage[\"usage\"] = {\n\t\tinput,\n\t\toutput: outputTokens,\n\t\tcacheRead: cacheReadTokens,\n\t\tcacheWrite: cacheWriteTokens,\n\t\ttotalTokens: input + outputTokens + cacheReadTokens + cacheWriteTokens,\n\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t};\n\tcalculateCost(model, usage);\n\treturn usage;\n}\n"]}
+{"version":3,"file":"openai-completions.d.ts","sourceRoot":"","sources":["../../src/providers/openai-completions.ts"],"names":[],"mappings":"AACA,OAAO,KAAK,EAOX,0BAA0B,EAG1B,MAAM,sCAAsC,CAAC;AAE9C,OAAO,KAAK,EAGX,OAAO,EAGP,KAAK,EACL,mBAAmB,EACnB,cAAc,EACd,aAAa,EAMb,MAAM,aAAa,CAAC;AASrB,OAAO,EAAiC,KAAK,+BAA+B,EAAE,MAAM,gCAAgC,CAAC;AA2CrH,MAAM,WAAW,wBAAyB,SAAQ,aAAa;IAC9D,UAAU,CAAC,EAAE,MAAM,GAAG,MAAM,GAAG,UAAU,GAAG;QAAE,IAAI,EAAE,UAAU,CAAC;QAAC,QAAQ,EAAE;YAAE,IAAI,EAAE,MAAM,CAAA;SAAE,CAAA;KAAE,CAAC;IAC7F,eAAe,CAAC,EAAE,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,GAAG,KAAK,GAAG,OAAO,CAAC;CACpF;AA2BD,eAAO,MAAM,uBAAuB,EAAE,cAAc,CAAC,oBAAoB,EAAE,wBAAwB,CAgUlG,CAAC;AAEF,eAAO,MAAM,6BAA6B,EAAE,cAAc,CAAC,oBAAoB,EAAE,mBAAmB,CAoBnG,CAAC;AAwTF,wBAAgB,eAAe,CAC9B,KAAK,EAAE,KAAK,CAAC,oBAAoB,CAAC,EAClC,OAAO,EAAE,OAAO,EAChB,MAAM,EAAE,+BAA+B,GACrC,0BAA0B,EAAE,CAyQ9B","sourcesContent":["import OpenAI from \"openai\";\nimport type {\n\tChatCompletionAssistantMessageParam,\n\tChatCompletionChunk,\n\tChatCompletionContentPart,\n\tChatCompletionContentPartImage,\n\tChatCompletionContentPartText,\n\tChatCompletionDeveloperMessageParam,\n\tChatCompletionMessageParam,\n\tChatCompletionSystemMessageParam,\n\tChatCompletionToolMessageParam,\n} from \"openai/resources/chat/completions.js\";\nimport { clampThinkingLevel } from \"../models.ts\";\nimport type {\n\tAssistantMessage,\n\tCacheRetention,\n\tContext,\n\tImageContent,\n\tMessage,\n\tModel,\n\tSimpleStreamOptions,\n\tStreamFunction,\n\tStreamOptions,\n\tTextContent,\n\tThinkingContent,\n\tTool,\n\tToolCall,\n\tToolResultMessage,\n} from \"../types.ts\";\nimport { AssistantMessageEventStream } from \"../utils/event-stream.ts\";\nimport { shortHash } from \"../utils/hash.ts\";\nimport { headersToRecord } from \"../utils/headers.ts\";\nimport { parseStreamingJson } from \"../utils/json-parse.ts\";\nimport { retryProviderRequest } from \"../utils/provider-retry.ts\";\nimport { sanitizeSurrogates } from \"../utils/sanitize-unicode.ts\";\nimport { isCloudflareProvider, resolveCloudflareBaseUrl } from \"./cloudflare.ts\";\nimport { buildCopilotDynamicHeaders, hasCopilotVisionInput } from \"./github-copilot-headers.ts\";\nimport { getCompat, isModelStudioModel, type ResolvedOpenAICompletionsCompat } from \"./openai-completions-compat.ts\";\nimport { parseChunkUsage } from \"./openai-completions-usage.ts\";\nimport { resolveOpenAIPromptCacheKey } from \"./openai-prompt-cache.ts\";\nimport { openAICompletionStopReason as mapStopReason } from \"./provider-stop-reasons.ts\";\nimport { buildBaseOptions } from \"./simple-options.ts\";\nimport { stableTools } from \"./tool-schema.ts\";\nimport { transformMessages } from \"./transform-messages.ts\";\n\n/**\n * Check if conversation messages contain tool calls or tool results.\n * This is needed because Anthropic (via proxy) requires the tools param\n * to be present when messages include tool_calls or tool role messages.\n */\nfunction hasToolHistory(messages: Message[]): boolean {\n\tfor (const msg of messages) {\n\t\tif (msg.role === \"toolResult\") {\n\t\t\treturn true;\n\t\t}\n\t\tif (msg.role === \"assistant\") {\n\t\t\tif (msg.content.some((block) => block.type === \"toolCall\")) {\n\t\t\t\treturn true;\n\t\t\t}\n\t\t}\n\t}\n\treturn false;\n}\n\nfunction isTextContentBlock(block: { type: string }): block is TextContent {\n\treturn block.type === \"text\";\n}\n\nfunction isThinkingContentBlock(block: { type: string }): block is ThinkingContent {\n\treturn block.type === \"thinking\";\n}\n\nfunction isToolCallBlock(block: { type: string }): block is ToolCall {\n\treturn block.type === \"toolCall\";\n}\n\nfunction isImageContentBlock(block: { type: string }): block is ImageContent {\n\treturn block.type === \"image\";\n}\n\nexport interface OpenAICompletionsOptions extends StreamOptions {\n\ttoolChoice?: \"auto\" | \"none\" | \"required\" | { type: \"function\"; function: { name: string } };\n\treasoningEffort?: \"minimal\" | \"low\" | \"medium\" | \"high\" | \"xhigh\" | \"max\" | \"ultra\";\n}\n\ninterface OpenAICompatCacheControl {\n\ttype: \"ephemeral\";\n\tttl?: string;\n}\n\ntype ChatCompletionInstructionMessageParam = ChatCompletionDeveloperMessageParam | ChatCompletionSystemMessageParam;\n\ntype ChatCompletionTextPartWithCacheControl = ChatCompletionContentPartText & {\n\tcache_control?: OpenAICompatCacheControl;\n};\n\ntype ChatCompletionToolWithCacheControl = OpenAI.Chat.Completions.ChatCompletionTool & {\n\tcache_control?: OpenAICompatCacheControl;\n};\n\nfunction resolveCacheRetention(cacheRetention?: CacheRetention): CacheRetention {\n\tif (cacheRetention) {\n\t\treturn cacheRetention;\n\t}\n\tif (typeof process !== \"undefined\" && process.env.OMK_CACHE_RETENTION === \"long\") {\n\t\treturn \"long\";\n\t}\n\treturn \"short\";\n}\n\nexport const streamOpenAICompletions: StreamFunction<\"openai-completions\", OpenAICompletionsOptions> = (\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\toptions?: OpenAICompletionsOptions,\n): AssistantMessageEventStream => {\n\tconst stream = new AssistantMessageEventStream();\n\n\t(async () => {\n\t\tconst output: AssistantMessage = {\n\t\t\trole: \"assistant\",\n\t\t\tcontent: [],\n\t\t\tapi: model.api,\n\t\t\tprovider: model.provider,\n\t\t\tmodel: model.id,\n\t\t\tusage: {\n\t\t\t\tinput: 0,\n\t\t\t\toutput: 0,\n\t\t\t\tcacheRead: 0,\n\t\t\t\tcacheWrite: 0,\n\t\t\t\ttotalTokens: 0,\n\t\t\t\tcost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },\n\t\t\t},\n\t\t\tstopReason: \"stop\",\n\t\t\ttimestamp: Date.now(),\n\t\t};\n\n\t\ttry {\n\t\t\tconst apiKey = options?.apiKey;\n\t\t\tif (!apiKey) {\n\t\t\t\tthrow new Error(`No API key for provider: ${model.provider}`);\n\t\t\t}\n\t\t\tconst compat = getCompat(model);\n\t\t\tconst cacheRetention = resolveCacheRetention(options?.cacheRetention);\n\t\t\tconst cacheSessionId = cacheRetention === \"none\" ? undefined : options?.sessionId;\n\t\t\tconst client = createClient(model, context, apiKey, options?.headers, cacheSessionId, compat);\n\t\t\tlet params = buildParams(model, context, options, compat, cacheRetention);\n\t\t\tconst nextParams = await options?.onPayload?.(params, model);\n\t\t\tif (nextParams !== undefined) {\n\t\t\t\tparams = nextParams as OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming;\n\t\t\t}\n\t\t\tconst requestOptions = {\n\t\t\t\t...(options?.signal ? { signal: options.signal } : {}),\n\t\t\t\t...(options?.timeoutMs !== undefined ? { timeout: options.timeoutMs } : {}),\n\t\t\t\tmaxRetries: 0,\n\t\t\t};\n\t\t\tconst { data: openaiStream, response } = await retryProviderRequest(\n\t\t\t\t() => client.chat.completions.create(params, requestOptions).withResponse(),\n\t\t\t\t{\n\t\t\t\t\tmaxRetries: options?.maxRetries,\n\t\t\t\t\tmaxRetryDelayMs: options?.maxRetryDelayMs,\n\t\t\t\t\tsignal: options?.signal,\n\t\t\t\t},\n\t\t\t);\n\t\t\tawait options?.onResponse?.({ status: response.status, headers: headersToRecord(response.headers) }, model);\n\t\t\tstream.push({ type: \"start\", partial: output });\n\n\t\t\tinterface StreamingToolCallBlock extends ToolCall {\n\t\t\t\tpartialArgs?: string;\n\t\t\t\tstreamIndex?: number;\n\t\t\t}\n\t\t\ttype StreamingBlock = TextContent | ThinkingContent | StreamingToolCallBlock;\n\t\t\ttype StreamingToolCallDelta = NonNullable<ChatCompletionChunk.Choice.Delta[\"tool_calls\"]>[number];\n\n\t\t\tlet textBlock: TextContent | null = null;\n\t\t\tlet thinkingBlock: ThinkingContent | null = null;\n\t\t\tlet hasFinishReason = false;\n\t\t\tconst toolCallBlocksByIndex = new Map<number, StreamingToolCallBlock>();\n\t\t\tconst toolCallBlocksById = new Map<string, StreamingToolCallBlock>();\n\t\t\tconst blocks = output.content as StreamingBlock[];\n\t\t\tconst getContentIndex = (block: StreamingBlock) => blocks.indexOf(block);\n\t\t\tconst finishBlock = (block: StreamingBlock) => {\n\t\t\t\tconst contentIndex = getContentIndex(block);\n\t\t\t\tif (contentIndex === -1) {\n\t\t\t\t\treturn;\n\t\t\t\t}\n\t\t\t\tif (block.type === \"text\") {\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"text_end\",\n\t\t\t\t\t\tcontentIndex,\n\t\t\t\t\t\tcontent: block.text,\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t} else if (block.type === \"thinking\") {\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"thinking_end\",\n\t\t\t\t\t\tcontentIndex,\n\t\t\t\t\t\tcontent: block.thinking,\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\tblock.arguments = parseStreamingJson(block.partialArgs);\n\t\t\t\t\t// Finalize in-place and strip the scratch buffers so replay only\n\t\t\t\t\t// carries parsed arguments.\n\t\t\t\t\tdelete block.partialArgs;\n\t\t\t\t\tdelete block.streamIndex;\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"toolcall_end\",\n\t\t\t\t\t\tcontentIndex,\n\t\t\t\t\t\ttoolCall: block,\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t};\n\t\t\tconst ensureTextBlock = () => {\n\t\t\t\tif (!textBlock) {\n\t\t\t\t\ttextBlock = { type: \"text\", text: \"\" };\n\t\t\t\t\tblocks.push(textBlock);\n\t\t\t\t\tstream.push({ type: \"text_start\", contentIndex: getContentIndex(textBlock), partial: output });\n\t\t\t\t}\n\t\t\t\treturn textBlock;\n\t\t\t};\n\t\t\tconst ensureThinkingBlock = (thinkingSignature: string) => {\n\t\t\t\tif (!thinkingBlock) {\n\t\t\t\t\tthinkingBlock = {\n\t\t\t\t\t\ttype: \"thinking\",\n\t\t\t\t\t\tthinking: \"\",\n\t\t\t\t\t\tthinkingSignature,\n\t\t\t\t\t};\n\t\t\t\t\tblocks.push(thinkingBlock);\n\t\t\t\t\tstream.push({ type: \"thinking_start\", contentIndex: getContentIndex(thinkingBlock), partial: output });\n\t\t\t\t}\n\t\t\t\treturn thinkingBlock;\n\t\t\t};\n\t\t\tconst ensureToolCallBlock = (toolCall: StreamingToolCallDelta) => {\n\t\t\t\tconst streamIndex = typeof toolCall.index === \"number\" ? toolCall.index : undefined;\n\t\t\t\tlet block = streamIndex !== undefined ? toolCallBlocksByIndex.get(streamIndex) : undefined;\n\t\t\t\tif (!block && toolCall.id) {\n\t\t\t\t\tblock = toolCallBlocksById.get(toolCall.id);\n\t\t\t\t}\n\t\t\t\tif (!block) {\n\t\t\t\t\tblock = {\n\t\t\t\t\t\ttype: \"toolCall\",\n\t\t\t\t\t\tid: toolCall.id || \"\",\n\t\t\t\t\t\tname: toolCall.function?.name || \"\",\n\t\t\t\t\t\targuments: {},\n\t\t\t\t\t\tpartialArgs: \"\",\n\t\t\t\t\t\tstreamIndex,\n\t\t\t\t\t};\n\t\t\t\t\tif (streamIndex !== undefined) {\n\t\t\t\t\t\ttoolCallBlocksByIndex.set(streamIndex, block);\n\t\t\t\t\t}\n\t\t\t\t\tif (toolCall.id) {\n\t\t\t\t\t\ttoolCallBlocksById.set(toolCall.id, block);\n\t\t\t\t\t}\n\t\t\t\t\tblocks.push(block);\n\t\t\t\t\tstream.push({\n\t\t\t\t\t\ttype: \"toolcall_start\",\n\t\t\t\t\t\tcontentIndex: getContentIndex(block),\n\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t});\n\t\t\t\t}\n\t\t\t\tif (streamIndex !== undefined && block.streamIndex === undefined) {\n\t\t\t\t\tblock.streamIndex = streamIndex;\n\t\t\t\t\ttoolCallBlocksByIndex.set(streamIndex, block);\n\t\t\t\t}\n\t\t\t\tif (toolCall.id) {\n\t\t\t\t\ttoolCallBlocksById.set(toolCall.id, block);\n\t\t\t\t}\n\t\t\t\treturn block;\n\t\t\t};\n\n\t\t\tfor await (const chunk of openaiStream) {\n\t\t\t\tif (!chunk || typeof chunk !== \"object\") continue;\n\n\t\t\t\t// OpenAI documents ChatCompletionChunk.id as the unique chat completion identifier,\n\t\t\t\t// and each chunk in a streamed completion carries the same id.\n\t\t\t\toutput.responseId ||= chunk.id;\n\t\t\t\tif (typeof chunk.model === \"string\" && chunk.model.length > 0 && chunk.model !== model.id) {\n\t\t\t\t\toutput.responseModel ||= chunk.model;\n\t\t\t\t}\n\t\t\t\tif (chunk.usage) {\n\t\t\t\t\toutput.usage = parseChunkUsage(chunk.usage, model);\n\t\t\t\t}\n\n\t\t\t\tconst choice = Array.isArray(chunk.choices) ? chunk.choices[0] : undefined;\n\t\t\t\tif (!choice) continue;\n\n\t\t\t\t// Fallback: some providers (e.g., Moonshot) return usage\n\t\t\t\t// in choice.usage instead of the standard chunk.usage\n\t\t\t\tif (!chunk.usage && (choice as any).usage) {\n\t\t\t\t\toutput.usage = parseChunkUsage((choice as any).usage, model);\n\t\t\t\t}\n\n\t\t\t\tif (choice.finish_reason) {\n\t\t\t\t\tconst finishReasonResult = mapStopReason(choice.finish_reason);\n\t\t\t\t\toutput.stopReason = finishReasonResult.stopReason;\n\t\t\t\t\tif (finishReasonResult.errorMessage) {\n\t\t\t\t\t\toutput.errorMessage = finishReasonResult.errorMessage;\n\t\t\t\t\t}\n\t\t\t\t\thasFinishReason = true;\n\t\t\t\t}\n\n\t\t\t\tif (choice.delta) {\n\t\t\t\t\tif (\n\t\t\t\t\t\tchoice.delta.content !== null &&\n\t\t\t\t\t\tchoice.delta.content !== undefined &&\n\t\t\t\t\t\tchoice.delta.content.length > 0\n\t\t\t\t\t) {\n\t\t\t\t\t\tconst block = ensureTextBlock();\n\t\t\t\t\t\tblock.text += choice.delta.content;\n\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\ttype: \"text_delta\",\n\t\t\t\t\t\t\tcontentIndex: getContentIndex(block),\n\t\t\t\t\t\t\tdelta: choice.delta.content,\n\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t});\n\t\t\t\t\t}\n\n\t\t\t\t\t// Some endpoints return reasoning in reasoning_content (llama.cpp),\n\t\t\t\t\t// or reasoning (other openai compatible endpoints)\n\t\t\t\t\t// Use the first non-empty reasoning field to avoid duplication\n\t\t\t\t\t// (e.g., chutes.ai returns both reasoning_content and reasoning with same content)\n\t\t\t\t\tconst reasoningFields = [\"reasoning_content\", \"reasoning\", \"reasoning_text\"];\n\t\t\t\t\tconst deltaFields = choice.delta as Record<string, unknown>;\n\t\t\t\t\tlet foundReasoningField: string | null = null;\n\t\t\t\t\tfor (const field of reasoningFields) {\n\t\t\t\t\t\tconst value = deltaFields[field];\n\t\t\t\t\t\tif (typeof value === \"string\" && value.length > 0) {\n\t\t\t\t\t\t\tfoundReasoningField = field;\n\t\t\t\t\t\t\tbreak;\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\tif (foundReasoningField) {\n\t\t\t\t\t\tconst delta = deltaFields[foundReasoningField];\n\t\t\t\t\t\tif (typeof delta === \"string\" && delta.length > 0) {\n\t\t\t\t\t\t\tconst thinkingSignature =\n\t\t\t\t\t\t\t\tmodel.provider === \"opencode-go\" && foundReasoningField === \"reasoning\"\n\t\t\t\t\t\t\t\t\t? \"reasoning_content\"\n\t\t\t\t\t\t\t\t\t: foundReasoningField;\n\t\t\t\t\t\t\tconst block = ensureThinkingBlock(thinkingSignature);\n\t\t\t\t\t\t\tblock.thinking += delta;\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"thinking_delta\",\n\t\t\t\t\t\t\t\tcontentIndex: getContentIndex(block),\n\t\t\t\t\t\t\t\tdelta,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\tif (choice?.delta?.tool_calls) {\n\t\t\t\t\t\tfor (const toolCall of choice.delta.tool_calls) {\n\t\t\t\t\t\t\tconst block = ensureToolCallBlock(toolCall);\n\t\t\t\t\t\t\tif (!block.id && toolCall.id) {\n\t\t\t\t\t\t\t\tblock.id = toolCall.id;\n\t\t\t\t\t\t\t\ttoolCallBlocksById.set(toolCall.id, block);\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\tif (!block.name && toolCall.function?.name) {\n\t\t\t\t\t\t\t\tblock.name = toolCall.function.name;\n\t\t\t\t\t\t\t}\n\n\t\t\t\t\t\t\tlet delta = \"\";\n\t\t\t\t\t\t\tif (toolCall.function?.arguments) {\n\t\t\t\t\t\t\t\tdelta = toolCall.function.arguments;\n\t\t\t\t\t\t\t\tblock.partialArgs = (block.partialArgs ?? \"\") + toolCall.function.arguments;\n\t\t\t\t\t\t\t\tblock.arguments = parseStreamingJson(block.partialArgs);\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\tstream.push({\n\t\t\t\t\t\t\t\ttype: \"toolcall_delta\",\n\t\t\t\t\t\t\t\tcontentIndex: getContentIndex(block),\n\t\t\t\t\t\t\t\tdelta,\n\t\t\t\t\t\t\t\tpartial: output,\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\n\t\t\t\t\tconst reasoningDetails = (choice.delta as any).reasoning_details;\n\t\t\t\t\tif (reasoningDetails && Array.isArray(reasoningDetails)) {\n\t\t\t\t\t\tfor (const detail of reasoningDetails) {\n\t\t\t\t\t\t\tif (detail.type === \"reasoning.encrypted\" && detail.id && detail.data) {\n\t\t\t\t\t\t\t\tconst matchingToolCall = output.content.find(\n\t\t\t\t\t\t\t\t\t(b) => b.type === \"toolCall\" && b.id === detail.id,\n\t\t\t\t\t\t\t\t) as ToolCall | undefined;\n\t\t\t\t\t\t\t\tif (matchingToolCall) {\n\t\t\t\t\t\t\t\t\tmatchingToolCall.thoughtSignature = JSON.stringify(detail);\n\t\t\t\t\t\t\t\t}\n\t\t\t\t\t\t\t}\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\tfor (const block of blocks) {\n\t\t\t\tfinishBlock(block);\n\t\t\t}\n\t\t\tif (options?.signal?.aborted) {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\n\t\t\tif (output.stopReason === \"aborted\") {\n\t\t\t\tthrow new Error(\"Request was aborted\");\n\t\t\t}\n\t\t\tif (output.stopReason === \"error\") {\n\t\t\t\tthrow new Error(output.errorMessage || \"Provider returned an error stop reason\");\n\t\t\t}\n\t\t\tif (!hasFinishReason) {\n\t\t\t\tthrow new Error(\"Stream ended without finish_reason\");\n\t\t\t}\n\n\t\t\tstream.push({ type: \"done\", reason: output.stopReason, message: output });\n\t\t\tstream.end();\n\t\t} catch (error) {\n\t\t\tfor (const block of output.content) {\n\t\t\t\tdelete (block as { index?: number }).index;\n\t\t\t\t// Streaming scratch buffers are only used during parsing; never persist them.\n\t\t\t\tdelete (block as { partialArgs?: string }).partialArgs;\n\t\t\t\tdelete (block as { streamIndex?: number }).streamIndex;\n\t\t\t}\n\t\t\toutput.stopReason = options?.signal?.aborted ? \"aborted\" : \"error\";\n\t\t\toutput.errorMessage = error instanceof Error ? error.message : JSON.stringify(error);\n\t\t\t// Some providers via OpenRouter give additional information in this field.\n\t\t\tconst rawMetadata = (error as any)?.error?.metadata?.raw;\n\t\t\tif (rawMetadata) output.errorMessage += `\\n${rawMetadata}`;\n\t\t\tstream.push({ type: \"error\", reason: output.stopReason, error: output });\n\t\t\tstream.end();\n\t\t}\n\t})();\n\n\treturn stream;\n};\n\nexport const streamSimpleOpenAICompletions: StreamFunction<\"openai-completions\", SimpleStreamOptions> = (\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\toptions?: SimpleStreamOptions,\n): AssistantMessageEventStream => {\n\tconst apiKey = options?.apiKey;\n\tif (!apiKey) {\n\t\tthrow new Error(`No API key for provider: ${model.provider}`);\n\t}\n\n\tconst base = buildBaseOptions(model, options, apiKey);\n\tconst clampedReasoning = options?.reasoning ? clampThinkingLevel(model, options.reasoning) : undefined;\n\tconst reasoningEffort = clampedReasoning === \"off\" ? undefined : clampedReasoning;\n\tconst toolChoice = (options as OpenAICompletionsOptions | undefined)?.toolChoice;\n\n\treturn streamOpenAICompletions(model, context, {\n\t\t...base,\n\t\treasoningEffort,\n\t\ttoolChoice,\n\t} satisfies OpenAICompletionsOptions);\n};\n\nfunction createClient(\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\tapiKey: string,\n\toptionsHeaders?: Record<string, string>,\n\tsessionId?: string,\n\tcompat: ResolvedOpenAICompletionsCompat = getCompat(model),\n) {\n\tconst headers = { ...model.headers };\n\tif (model.provider === \"github-copilot\") {\n\t\tconst hasImages = hasCopilotVisionInput(context.messages);\n\t\tconst copilotHeaders = buildCopilotDynamicHeaders({\n\t\t\tmessages: context.messages,\n\t\t\thasImages,\n\t\t});\n\t\tObject.assign(headers, copilotHeaders);\n\t}\n\n\tif (sessionId && compat.sendSessionAffinityHeaders) {\n\t\theaders.session_id = sessionId;\n\t\theaders[\"x-client-request-id\"] = sessionId;\n\t\theaders[\"x-session-affinity\"] = sessionId;\n\t}\n\n\t// Merge options headers last so they can override defaults\n\tif (optionsHeaders) {\n\t\tObject.assign(headers, optionsHeaders);\n\t}\n\n\tconst defaultHeaders =\n\t\tmodel.provider === \"cloudflare-ai-gateway\"\n\t\t\t? {\n\t\t\t\t\t...headers,\n\t\t\t\t\tAuthorization: headers.Authorization ?? null,\n\t\t\t\t\t\"cf-aig-authorization\": `Bearer ${apiKey}`,\n\t\t\t\t}\n\t\t\t: headers;\n\n\treturn new OpenAI({\n\t\tapiKey,\n\t\tbaseURL: isCloudflareProvider(model.provider) ? resolveCloudflareBaseUrl(model) : model.baseUrl,\n\t\tdangerouslyAllowBrowser: true,\n\t\tdefaultHeaders,\n\t});\n}\n\nfunction buildParams(\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\toptions?: OpenAICompletionsOptions,\n\tcompat: ResolvedOpenAICompletionsCompat = getCompat(model),\n\tcacheRetention: CacheRetention = resolveCacheRetention(options?.cacheRetention),\n) {\n\tconst messages = convertMessages(model, context, compat);\n\tconst cacheControl = getCompatCacheControl(compat, cacheRetention);\n\n\tconst params: OpenAI.Chat.Completions.ChatCompletionCreateParamsStreaming = {\n\t\tmodel: model.provider === \"zyloo\" && !model.id.startsWith(\"zyloo/\") ? `zyloo/${model.id}` : model.id,\n\t\tmessages,\n\t\tstream: true,\n\t\tprompt_cache_key:\n\t\t\t(model.baseUrl.includes(\"api.openai.com\") && cacheRetention !== \"none\") ||\n\t\t\t(cacheRetention === \"long\" && compat.supportsLongCacheRetention)\n\t\t\t\t? resolveOpenAIPromptCacheKey(context, options?.sessionId, `${model.provider}/${model.id}`)\n\t\t\t\t: undefined,\n\t\tprompt_cache_retention: cacheRetention === \"long\" && compat.supportsLongCacheRetention ? \"24h\" : undefined,\n\t};\n\n\tif (compat.supportsUsageInStreaming !== false) {\n\t\t(params as any).stream_options = { include_usage: true };\n\t}\n\n\tif (compat.supportsStore) {\n\t\tparams.store = false;\n\t}\n\n\tif (options?.maxTokens) {\n\t\tif (compat.maxTokensField === \"max_tokens\") {\n\t\t\t(params as any).max_tokens = options.maxTokens;\n\t\t} else {\n\t\t\tparams.max_completion_tokens = options.maxTokens;\n\t\t}\n\t}\n\n\tif (options?.temperature !== undefined) {\n\t\tparams.temperature = options.temperature;\n\t}\n\n\tif (context.tools && context.tools.length > 0) {\n\t\tparams.tools = convertTools(context.tools, compat);\n\t\tif (compat.zaiToolStream) {\n\t\t\t(params as any).tool_stream = true;\n\t\t}\n\t} else if (hasToolHistory(context.messages)) {\n\t\t// Anthropic (via LiteLLM/proxy) requires tools param when conversation has tool_calls/tool_results\n\t\tparams.tools = [];\n\t}\n\n\tif (cacheControl) {\n\t\tapplyAnthropicCacheControl(messages, params.tools, cacheControl);\n\t}\n\n\tif (options?.toolChoice) {\n\t\tparams.tool_choice = options.toolChoice;\n\t}\n\n\tif (compat.thinkingFormat === \"zai\" && model.reasoning) {\n\t\t(params as any).enable_thinking = !!options?.reasoningEffort;\n\t\tif (options?.reasoningEffort && compat.supportsReasoningEffort) {\n\t\t\t(params as any).reasoning_effort =\n\t\t\t\tmodel.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;\n\t\t}\n\t} else if (compat.thinkingFormat === \"qwen\" && model.reasoning) {\n\t\tObject.assign(params, { enable_thinking: !!options?.reasoningEffort });\n\t\tif (options?.reasoningEffort && compat.supportsReasoningEffort && isModelStudioModel(model)) {\n\t\t\tconst effort =\n\t\t\t\toptions.reasoningEffort === \"minimal\" && model.id.startsWith(\"deepseek-v4-\")\n\t\t\t\t\t? \"low\"\n\t\t\t\t\t: options.reasoningEffort;\n\t\t\tObject.assign(params, {\n\t\t\t\treasoning_effort: model.thinkingLevelMap?.[options.reasoningEffort] ?? effort,\n\t\t\t});\n\t\t}\n\t} else if (compat.thinkingFormat === \"qwen-chat-template\" && model.reasoning) {\n\t\t(params as any).chat_template_kwargs = {\n\t\t\tenable_thinking: !!options?.reasoningEffort,\n\t\t\tpreserve_thinking: true,\n\t\t};\n\t} else if (compat.thinkingFormat === \"deepseek\" && model.reasoning) {\n\t\t(params as any).thinking = { type: options?.reasoningEffort ? \"enabled\" : \"disabled\" };\n\t\tif (options?.reasoningEffort && compat.supportsReasoningEffort) {\n\t\t\t(params as any).reasoning_effort =\n\t\t\t\tmodel.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;\n\t\t}\n\t} else if (compat.thinkingFormat === \"openrouter\" && model.reasoning) {\n\t\t// OpenRouter normalizes reasoning across providers via a nested reasoning object.\n\t\tconst openRouterParams = params as typeof params & { reasoning?: { effort?: string; enabled?: boolean } };\n\t\tif (options?.reasoningEffort) {\n\t\t\topenRouterParams.reasoning = {\n\t\t\t\teffort: model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort,\n\t\t\t};\n\t\t} else if (model.thinkingLevelMap?.off !== null) {\n\t\t\tconst offValue = model.thinkingLevelMap?.off ?? \"none\";\n\t\t\topenRouterParams.reasoning = offValue === \"none\" ? { enabled: false } : { effort: offValue };\n\t\t}\n\t} else if (compat.thinkingFormat === \"ant-ling\" && model.reasoning && options?.reasoningEffort) {\n\t\tconst effort = model.thinkingLevelMap?.[options.reasoningEffort];\n\t\tif (typeof effort === \"string\") {\n\t\t\t(params as typeof params & { reasoning?: { effort: string } }).reasoning = { effort };\n\t\t}\n\t} else if (compat.thinkingFormat === \"together\" && model.reasoning) {\n\t\tconst togetherParams = params as Omit<typeof params, \"reasoning_effort\"> & {\n\t\t\treasoning?: { enabled: boolean };\n\t\t\treasoning_effort?: string;\n\t\t};\n\t\ttogetherParams.reasoning = { enabled: !!options?.reasoningEffort };\n\t\tif (options?.reasoningEffort && compat.supportsReasoningEffort) {\n\t\t\ttogetherParams.reasoning_effort = model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;\n\t\t}\n\t} else if (compat.thinkingFormat === \"string-thinking\" && model.reasoning) {\n\t\tconst stringThinkingParams = params as typeof params & { thinking?: string };\n\t\tif (options?.reasoningEffort) {\n\t\t\tstringThinkingParams.thinking = model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;\n\t\t} else if (model.thinkingLevelMap?.off !== null) {\n\t\t\tstringThinkingParams.thinking = model.thinkingLevelMap?.off ?? \"none\";\n\t\t}\n\t} else if (options?.reasoningEffort && model.reasoning && compat.supportsReasoningEffort) {\n\t\t// OpenAI-style reasoning_effort\n\t\t(params as any).reasoning_effort = model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;\n\t} else if (!options?.reasoningEffort && model.reasoning && compat.supportsReasoningEffort) {\n\t\tconst offValue = model.thinkingLevelMap?.off;\n\t\tif (typeof offValue === \"string\") {\n\t\t\t(params as any).reasoning_effort = offValue;\n\t\t}\n\t}\n\n\t// OpenRouter provider routing preferences\n\tif (model.baseUrl.includes(\"openrouter.ai\") && model.compat?.openRouterRouting) {\n\t\t(params as any).provider = model.compat.openRouterRouting;\n\t}\n\n\t// Vercel AI Gateway provider routing preferences\n\tif (model.baseUrl.includes(\"ai-gateway.vercel.sh\") && model.compat?.vercelGatewayRouting) {\n\t\tconst routing = model.compat.vercelGatewayRouting;\n\t\tif (routing.only || routing.order) {\n\t\t\tconst gatewayOptions: Record<string, string[]> = {};\n\t\t\tif (routing.only) gatewayOptions.only = routing.only;\n\t\t\tif (routing.order) gatewayOptions.order = routing.order;\n\t\t\t(params as any).providerOptions = { gateway: gatewayOptions };\n\t\t}\n\t}\n\n\treturn params;\n}\n\nfunction getCompatCacheControl(\n\tcompat: ResolvedOpenAICompletionsCompat,\n\tcacheRetention: CacheRetention,\n): OpenAICompatCacheControl | undefined {\n\tif (compat.cacheControlFormat !== \"anthropic\" || cacheRetention === \"none\") {\n\t\treturn undefined;\n\t}\n\n\tconst ttl = cacheRetention === \"long\" && compat.supportsLongCacheRetention ? \"1h\" : undefined;\n\treturn { type: \"ephemeral\", ...(ttl ? { ttl } : {}) };\n}\n\nfunction applyAnthropicCacheControl(\n\tmessages: ChatCompletionMessageParam[],\n\ttools: OpenAI.Chat.Completions.ChatCompletionTool[] | undefined,\n\tcacheControl: OpenAICompatCacheControl,\n): void {\n\taddCacheControlToSystemPrompt(messages, cacheControl);\n\taddCacheControlToLastTool(tools, cacheControl);\n\taddCacheControlToLastConversationMessage(messages, cacheControl);\n}\n\nfunction addCacheControlToSystemPrompt(\n\tmessages: ChatCompletionMessageParam[],\n\tcacheControl: OpenAICompatCacheControl,\n): void {\n\tfor (const message of messages) {\n\t\tif (message.role === \"system\" || message.role === \"developer\") {\n\t\t\taddCacheControlToInstructionMessage(message, cacheControl);\n\t\t\treturn;\n\t\t}\n\t}\n}\n\nfunction addCacheControlToLastConversationMessage(\n\tmessages: ChatCompletionMessageParam[],\n\tcacheControl: OpenAICompatCacheControl,\n): void {\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst message = messages[i];\n\t\tif (message.role === \"user\" || message.role === \"assistant\") {\n\t\t\tif (addCacheControlToMessage(message, cacheControl)) {\n\t\t\t\treturn;\n\t\t\t}\n\t\t}\n\t}\n}\n\nfunction addCacheControlToLastTool(\n\ttools: OpenAI.Chat.Completions.ChatCompletionTool[] | undefined,\n\tcacheControl: OpenAICompatCacheControl,\n): void {\n\tif (!tools || tools.length === 0) {\n\t\treturn;\n\t}\n\n\tconst lastTool = tools[tools.length - 1] as ChatCompletionToolWithCacheControl;\n\tlastTool.cache_control = cacheControl;\n}\n\nfunction addCacheControlToInstructionMessage(\n\tmessage: ChatCompletionInstructionMessageParam,\n\tcacheControl: OpenAICompatCacheControl,\n): boolean {\n\treturn addCacheControlToTextContent(message, cacheControl);\n}\n\nfunction addCacheControlToMessage(\n\tmessage: ChatCompletionMessageParam,\n\tcacheControl: OpenAICompatCacheControl,\n): boolean {\n\tif (message.role === \"user\" || message.role === \"assistant\") {\n\t\treturn addCacheControlToTextContent(message, cacheControl);\n\t}\n\treturn false;\n}\n\nfunction addCacheControlToTextContent(\n\tmessage:\n\t\t| ChatCompletionInstructionMessageParam\n\t\t| ChatCompletionAssistantMessageParam\n\t\t| Extract<ChatCompletionMessageParam, { role: \"user\" }>,\n\tcacheControl: OpenAICompatCacheControl,\n): boolean {\n\tconst content = message.content;\n\tif (typeof content === \"string\") {\n\t\tif (content.length === 0) {\n\t\t\treturn false;\n\t\t}\n\t\tmessage.content = [\n\t\t\t{\n\t\t\t\ttype: \"text\",\n\t\t\t\ttext: content,\n\t\t\t\tcache_control: cacheControl,\n\t\t\t},\n\t\t] as ChatCompletionTextPartWithCacheControl[];\n\t\treturn true;\n\t}\n\n\tif (!Array.isArray(content)) {\n\t\treturn false;\n\t}\n\n\tfor (let i = content.length - 1; i >= 0; i--) {\n\t\tconst part = content[i];\n\t\tif (part?.type === \"text\") {\n\t\t\tconst textPart = part as ChatCompletionTextPartWithCacheControl;\n\t\t\ttextPart.cache_control = cacheControl;\n\t\t\treturn true;\n\t\t}\n\t}\n\n\treturn false;\n}\n\nexport function convertMessages(\n\tmodel: Model<\"openai-completions\">,\n\tcontext: Context,\n\tcompat: ResolvedOpenAICompletionsCompat,\n): ChatCompletionMessageParam[] {\n\tconst params: ChatCompletionMessageParam[] = [];\n\n\tconst normalizeToolCallId = (id: string): string => {\n\t\t// Handle pipe-separated IDs from OpenAI Responses API\n\t\t// Format: {call_id}|{id} where {id} can be 400+ chars with special chars (+, /, =)\n\t\t// These come from providers like github-copilot, openai-codex, opencode\n\t\t// Extract just the call_id part and normalize it\n\t\t// Multiple tool calls in the same turn can share call_id but differ by item_id.\n\t\t// Preserve item-level uniqueness when replaying into Chat Completions, which\n\t\t// requires distinct tool call ids.\n\t\tif (id.includes(\"|\")) {\n\t\t\t// Sanitize to allowed chars and truncate to 40 chars (OpenAI limit)\n\t\t\tconst separatorIndex = id.indexOf(\"|\");\n\t\t\tconst callId = id.slice(0, separatorIndex).replace(/[^a-zA-Z0-9_-]/g, \"_\");\n\t\t\tconst itemId = id.slice(separatorIndex + 1).replace(/[^a-zA-Z0-9_-]/g, \"_\");\n\t\t\tconst combinedId = itemId.length > 0 ? `${callId}_${itemId}` : callId;\n\t\t\tif (combinedId.length <= 40) {\n\t\t\t\treturn combinedId;\n\t\t\t}\n\t\t\tconst hash = shortHash(id).slice(0, 8);\n\t\t\tconst prefix = callId.slice(0, Math.max(1, 40 - hash.length - 1));\n\t\t\treturn `${prefix}_${hash}`;\n\t\t}\n\n\t\tif (model.provider === \"openai\") return id.length > 40 ? id.slice(0, 40) : id;\n\t\treturn id;\n\t};\n\n\tconst transformedMessages = transformMessages(context.messages, model, (id) => normalizeToolCallId(id));\n\n\t// WorkBuddy rejects a leading user message (`400 11128 \"first message is not\n\t// system prompt\"`) but accepts an empty system message, so a caller with no\n\t// system prompt still opens the channel instead of being rejected upstream.\n\tconst systemPrompt = context.systemPrompt || (compat.requiresSystemMessageFirst ? \"\" : undefined);\n\tif (systemPrompt !== undefined) {\n\t\tconst role = model.reasoning && compat.supportsDeveloperRole ? \"developer\" : \"system\";\n\t\tparams.push({ role, content: sanitizeSurrogates(systemPrompt) });\n\t}\n\n\tlet lastRole: string | null = null;\n\n\tfor (let i = 0; i < transformedMessages.length; i++) {\n\t\tconst msg = transformedMessages[i];\n\t\t// Some providers don't allow user messages directly after tool results\n\t\t// Insert a synthetic assistant message to bridge the gap\n\t\tif (compat.requiresAssistantAfterToolResult && lastRole === \"toolResult\" && msg.role === \"user\") {\n\t\t\tparams.push({\n\t\t\t\trole: \"assistant\",\n\t\t\t\tcontent: \"I have processed the tool results.\",\n\t\t\t});\n\t\t}\n\n\t\tif (msg.role === \"user\") {\n\t\t\tif (typeof msg.content === \"string\") {\n\t\t\t\tparams.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent: sanitizeSurrogates(msg.content),\n\t\t\t\t});\n\t\t\t} else {\n\t\t\t\tconst supportsImages = model.input.includes(\"image\");\n\t\t\t\tconst content: ChatCompletionContentPart[] = [];\n\t\t\t\tlet omittedImages = false;\n\t\t\t\tfor (const item of msg.content) {\n\t\t\t\t\tif (item.type === \"text\") {\n\t\t\t\t\t\tcontent.push({\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: sanitizeSurrogates(item.text),\n\t\t\t\t\t\t} satisfies ChatCompletionContentPartText);\n\t\t\t\t\t} else if (supportsImages) {\n\t\t\t\t\t\tcontent.push({\n\t\t\t\t\t\t\ttype: \"image_url\",\n\t\t\t\t\t\t\timage_url: {\n\t\t\t\t\t\t\t\turl: `data:${item.mimeType};base64,${item.data}`,\n\t\t\t\t\t\t\t},\n\t\t\t\t\t\t} satisfies ChatCompletionContentPartImage);\n\t\t\t\t\t} else {\n\t\t\t\t\t\tomittedImages = true;\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t\tif (omittedImages) {\n\t\t\t\t\tcontent.push({\n\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\ttext: \"[Image]\",\n\t\t\t\t\t} satisfies ChatCompletionContentPartText);\n\t\t\t\t}\n\t\t\t\tif (content.length === 0) continue;\n\t\t\t\tparams.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent,\n\t\t\t\t});\n\t\t\t}\n\t\t} else if (msg.role === \"assistant\") {\n\t\t\t// Some providers don't accept null content, use empty string instead\n\t\t\tconst assistantMsg: ChatCompletionAssistantMessageParam = {\n\t\t\t\trole: \"assistant\",\n\t\t\t\tcontent: compat.requiresAssistantAfterToolResult ? \"\" : null,\n\t\t\t};\n\n\t\t\tconst assistantTextParts = msg.content\n\t\t\t\t.filter(isTextContentBlock)\n\t\t\t\t.filter((block) => block.text.trim().length > 0)\n\t\t\t\t.map(\n\t\t\t\t\t(block) =>\n\t\t\t\t\t\t({\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: sanitizeSurrogates(block.text),\n\t\t\t\t\t\t}) satisfies ChatCompletionContentPartText,\n\t\t\t\t);\n\t\t\tconst assistantText = assistantTextParts.map((part) => part.text).join(\"\");\n\n\t\t\tconst nonEmptyThinkingBlocks = msg.content\n\t\t\t\t.filter(isThinkingContentBlock)\n\t\t\t\t.filter((block) => block.thinking.trim().length > 0);\n\t\t\tif (nonEmptyThinkingBlocks.length > 0) {\n\t\t\t\tif (compat.requiresThinkingAsText) {\n\t\t\t\t\t// Convert thinking blocks to plain text (no tags to avoid model mimicking them)\n\t\t\t\t\tconst thinkingText = nonEmptyThinkingBlocks\n\t\t\t\t\t\t.map((block) => sanitizeSurrogates(block.thinking))\n\t\t\t\t\t\t.join(\"\\n\\n\");\n\t\t\t\t\tassistantMsg.content = [{ type: \"text\", text: thinkingText }, ...assistantTextParts];\n\t\t\t\t} else {\n\t\t\t\t\t// Always send assistant content as a plain string (OpenAI Chat Completions\n\t\t\t\t\t// API standard format). Sending as an array of {type:\"text\", text:\"...\"}\n\t\t\t\t\t// objects is non-standard and causes some models (e.g. DeepSeek V3.2 via\n\t\t\t\t\t// NVIDIA NIM) to mirror the content-block structure literally in their\n\t\t\t\t\t// output, producing recursive nesting like [{'type':'text','text':'[{...}]'}].\n\t\t\t\t\tif (assistantText.length > 0) {\n\t\t\t\t\t\tassistantMsg.content = assistantText;\n\t\t\t\t\t}\n\n\t\t\t\t\t// Use the signature from the first thinking block if available (for llama.cpp server + gpt-oss)\n\t\t\t\t\tlet signature = nonEmptyThinkingBlocks[0].thinkingSignature;\n\t\t\t\t\tif (model.provider === \"opencode-go\" && signature === \"reasoning\") {\n\t\t\t\t\t\tsignature = \"reasoning_content\";\n\t\t\t\t\t}\n\t\t\t\t\tif (signature && signature.length > 0) {\n\t\t\t\t\t\t(assistantMsg as any)[signature] = nonEmptyThinkingBlocks.map((block) => block.thinking).join(\"\\n\");\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t} else if (assistantText.length > 0) {\n\t\t\t\t// Always send assistant content as a plain string (OpenAI Chat Completions\n\t\t\t\t// API standard format). Sending as an array of {type:\"text\", text:\"...\"}\n\t\t\t\t// objects is non-standard and causes some models (e.g. DeepSeek V3.2 via\n\t\t\t\t// NVIDIA NIM) to mirror the content-block structure literally in their\n\t\t\t\t// output, producing recursive nesting like [{'type':'text','text':'[{...}]'}].\n\t\t\t\tassistantMsg.content = assistantText;\n\t\t\t}\n\n\t\t\tconst toolCalls = msg.content.filter(isToolCallBlock);\n\t\t\tif (toolCalls.length > 0) {\n\t\t\t\tassistantMsg.tool_calls = toolCalls.map((tc) => ({\n\t\t\t\t\tid: tc.id,\n\t\t\t\t\ttype: \"function\" as const,\n\t\t\t\t\tfunction: {\n\t\t\t\t\t\tname: tc.name,\n\t\t\t\t\t\targuments: JSON.stringify(tc.arguments),\n\t\t\t\t\t},\n\t\t\t\t}));\n\t\t\t\tconst reasoningDetails = toolCalls\n\t\t\t\t\t.filter((tc) => tc.thoughtSignature)\n\t\t\t\t\t.map((tc) => {\n\t\t\t\t\t\ttry {\n\t\t\t\t\t\t\treturn JSON.parse(tc.thoughtSignature!);\n\t\t\t\t\t\t} catch {\n\t\t\t\t\t\t\treturn null;\n\t\t\t\t\t\t}\n\t\t\t\t\t})\n\t\t\t\t\t.filter(Boolean);\n\t\t\t\tif (reasoningDetails.length > 0) {\n\t\t\t\t\t(assistantMsg as any).reasoning_details = reasoningDetails;\n\t\t\t\t}\n\t\t\t}\n\t\t\tif (\n\t\t\t\tcompat.requiresReasoningContentOnAssistantMessages &&\n\t\t\t\tmodel.reasoning &&\n\t\t\t\t(assistantMsg as { reasoning_content?: string }).reasoning_content === undefined\n\t\t\t) {\n\t\t\t\t(assistantMsg as { reasoning_content?: string }).reasoning_content = \"\";\n\t\t\t}\n\t\t\t// Skip assistant messages that have no content and no tool calls.\n\t\t\t// Some providers require \"either content or tool_calls, but not none\".\n\t\t\t// Other providers also don't accept empty assistant messages.\n\t\t\t// This handles aborted assistant responses that got no content.\n\t\t\tconst content = assistantMsg.content;\n\t\t\tconst hasContent =\n\t\t\t\tcontent !== null &&\n\t\t\t\tcontent !== undefined &&\n\t\t\t\t(typeof content === \"string\" ? content.length > 0 : content.length > 0);\n\t\t\tif (!hasContent && !assistantMsg.tool_calls) {\n\t\t\t\tcontinue;\n\t\t\t}\n\t\t\tparams.push(assistantMsg);\n\t\t} else if (msg.role === \"toolResult\") {\n\t\t\tconst imageBlocks: Array<{ type: \"image_url\"; image_url: { url: string } }> = [];\n\t\t\tlet j = i;\n\n\t\t\tfor (; j < transformedMessages.length && transformedMessages[j].role === \"toolResult\"; j++) {\n\t\t\t\tconst toolMsg = transformedMessages[j] as ToolResultMessage;\n\n\t\t\t\t// Extract text and image content\n\t\t\t\tconst textResult = toolMsg.content\n\t\t\t\t\t.filter(isTextContentBlock)\n\t\t\t\t\t.map((block) => block.text)\n\t\t\t\t\t.join(\"\\n\");\n\t\t\t\tconst hasImages = toolMsg.content.some((c) => c.type === \"image\");\n\n\t\t\t\t// Always send tool result with text (or placeholder if only images)\n\t\t\t\tconst hasText = textResult.length > 0;\n\t\t\t\t// Some providers require the 'name' field in tool results\n\t\t\t\tconst toolResultMsg: ChatCompletionToolMessageParam = {\n\t\t\t\t\trole: \"tool\",\n\t\t\t\t\tcontent: sanitizeSurrogates(hasText ? textResult : \"(see attached image)\"),\n\t\t\t\t\ttool_call_id: toolMsg.toolCallId,\n\t\t\t\t};\n\t\t\t\tif (compat.requiresToolResultName && toolMsg.toolName) {\n\t\t\t\t\t(toolResultMsg as any).name = toolMsg.toolName;\n\t\t\t\t}\n\t\t\t\tparams.push(toolResultMsg);\n\n\t\t\t\tif (hasImages && model.input.includes(\"image\")) {\n\t\t\t\t\tfor (const block of toolMsg.content) {\n\t\t\t\t\t\tif (isImageContentBlock(block)) {\n\t\t\t\t\t\t\timageBlocks.push({\n\t\t\t\t\t\t\t\ttype: \"image_url\",\n\t\t\t\t\t\t\t\timage_url: {\n\t\t\t\t\t\t\t\t\turl: `data:${block.mimeType};base64,${block.data}`,\n\t\t\t\t\t\t\t\t},\n\t\t\t\t\t\t\t});\n\t\t\t\t\t\t}\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t}\n\n\t\t\ti = j - 1;\n\n\t\t\tif (imageBlocks.length > 0) {\n\t\t\t\tif (compat.requiresAssistantAfterToolResult) {\n\t\t\t\t\tparams.push({\n\t\t\t\t\t\trole: \"assistant\",\n\t\t\t\t\t\tcontent: \"I have processed the tool results.\",\n\t\t\t\t\t});\n\t\t\t\t}\n\n\t\t\t\tparams.push({\n\t\t\t\t\trole: \"user\",\n\t\t\t\t\tcontent: [\n\t\t\t\t\t\t{\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: \"Attached image(s) from tool result:\",\n\t\t\t\t\t\t},\n\t\t\t\t\t\t...imageBlocks,\n\t\t\t\t\t],\n\t\t\t\t});\n\t\t\t\tlastRole = \"user\";\n\t\t\t} else {\n\t\t\t\tlastRole = \"toolResult\";\n\t\t\t}\n\t\t\tcontinue;\n\t\t}\n\n\t\tlastRole = msg.role;\n\t}\n\n\treturn params;\n}\n\nfunction convertTools(\n\ttools: Tool[],\n\tcompat: ResolvedOpenAICompletionsCompat,\n): OpenAI.Chat.Completions.ChatCompletionTool[] {\n\treturn stableTools(tools).map((tool) => ({\n\t\ttype: \"function\",\n\t\tfunction: {\n\t\t\tname: tool.name,\n\t\t\tdescription: tool.description,\n\t\t\tparameters: tool.parameters as any, // TypeBox already generates JSON Schema\n\t\t\t// Only include strict if provider supports it. Some reject unknown fields.\n\t\t\t...(compat.supportsStrictMode !== false && { strict: false }),\n\t\t},\n\t}));\n}\n"]}

package/dist/providers/openai-completions.js

@@ -1,5 +1,5 @@
 import OpenAI from "openai";
-import { calculateCost, clampThinkingLevel } from "../models.js";
+import { clampThinkingLevel } from "../models.js";
 import { AssistantMessageEventStream } from "../utils/event-stream.js";
 import { shortHash } from "../utils/hash.js";
 import { headersToRecord } from "../utils/headers.js";
@@ -9,6 +9,7 @@ import { sanitizeSurrogates } from "../utils/sanitize-unicode.js";
 import { isCloudflareProvider, resolveCloudflareBaseUrl } from "./cloudflare.js";
 import { buildCopilotDynamicHeaders, hasCopilotVisionInput } from "./github-copilot-headers.js";
 import { getCompat, isModelStudioModel } from "./openai-completions-compat.js";
+import { parseChunkUsage } from "./openai-completions-usage.js";
 import { resolveOpenAIPromptCacheKey } from "./openai-prompt-cache.js";
 import { openAICompletionStopReason as mapStopReason } from "./provider-stop-reasons.js";
 import { buildBaseOptions } from "./simple-options.js";
@@ -868,30 +869,4 @@ function convertTools(tools, compat) {
         },
     }));
 }
-function parseChunkUsage(rawUsage, model) {
-    const promptTokens = rawUsage.prompt_tokens || 0;
-    const cacheReadTokens = rawUsage.prompt_tokens_details?.cached_tokens ?? rawUsage.prompt_cache_hit_tokens ?? 0;
-    const cacheWriteTokens = rawUsage.prompt_tokens_details?.cache_write_tokens || 0;
-    // Follow documented OpenAI/OpenRouter semantics: cached_tokens is cache-read
-    // tokens (hits). OpenAI does not document or emit cache_write_tokens, but
-    // OpenRouter-compatible providers can include it as a separate write count.
-    // OpenRouter's own provider/tests affirm the separate mapping:
-    // https://github.com/OpenRouterTeam/ai-sdk-provider/pull/409
-    // Do not subtract writes from cached_tokens, otherwise spec-compliant
-    // providers are under-reported. DS4 mirrors this contract too:
-    // https://github.com/antirez/ds4/pull/29
-    const input = Math.max(0, promptTokens - cacheReadTokens - cacheWriteTokens);
-    // OpenAI completion_tokens already includes reasoning_tokens.
-    const outputTokens = rawUsage.completion_tokens || 0;
-    const usage = {
-        input,
-        output: outputTokens,
-        cacheRead: cacheReadTokens,
-        cacheWrite: cacheWriteTokens,
-        totalTokens: input + outputTokens + cacheReadTokens + cacheWriteTokens,
-        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
-    };
-    calculateCost(model, usage);
-    return usage;
-}
 //# sourceMappingURL=openai-completions.js.map