190proof 1.0.114 → 1.0.116

package/README.md

@@ -136,6 +136,16 @@ const payload: GenericPayload = {
 const response = await callWithRetries("image-example", payload);
 ```
 
+How images reach the model depends on the provider. OpenAI, Anthropic, and
+Google get native image blocks. Groq is text-only: images degrade to an inline
+`Image (url)` text reference. OpenRouter sends OpenAI-style `image_url` content
+parts (the remote URL when present, else a `data:` URI) — but if the model has
+no vision-capable endpoints, OpenRouter rejects the request with a
+routing-layer 404, so the retry loop resends the payload with images degraded
+to the same inline text references Groq gets, and remembers the model
+(in-process, until restart) so later calls degrade up front. Messages without
+image attachments serialize identically either way.
+
 ### With System Messages
 
 ```typescript
@@ -280,6 +290,7 @@ Optional per-request knobs live on `payload` (`GenericPayload`):
 - `payload.streamTimeoutMs`: `number` - OpenRouter-only: total wall-clock budget per streaming attempt (default: 600000).
 - `payload.streamDeadlineAt`: `number` - OpenRouter-only: absolute deadline (epoch ms) for the whole call **including retries** — the caller's turn budget. Each attempt gets `min(streamTimeoutMs, deadline - now)`, and once under 10s remain the call fails fast instead of starting a generation that cannot be delivered. Use it whenever the caller has its own timeout: a per-attempt budget alone is re-granted on every retry and can outlive that timeout.
 - `payload.thinkingConfig`: `Record<string, unknown>` - Google-only: forwarded verbatim as `generationConfig.thinkingConfig` on the Gemini request — e.g. `{ thinkingBudget: 0 }` to disable thinking, `{ thinkingLevel: "HIGH" }` on models that take a level. Ignored by all other adapters; shapes are model-specific and validated by Google, not the SDK.
+- `payload.reasoningEffort`: `string` - OpenAI-only: forwarded as `reasoning_effort` on the request. Valid values are model-dependent (`none`/`minimal`/`low`/`medium`/`high`/`xhigh`/`max`). Reasoning-by-default models (the gpt-5.6 family) reject function tools on `/chat/completions` with a 400 unless this is explicitly `"none"` — their implicit default is `medium`. Ignored by all other adapters.
 
 When a streaming attempt is cut at its **total deadline** and prose has already arrived, the partial answer is returned with `truncated: true` on the response rather than discarded — those tokens were generated and billed, so throwing them away costs money and gives the user nothing. Surface such a reply as incomplete. Salvage never applies to tool-call turns (half-streamed arguments are unparseable JSON), to stalls (the provider died mid-thought), or to caller aborts. When nothing is salvageable, the discard is logged with an approximate token count — aborted attempts never receive OpenRouter's `usage` chunk, so that log line is the only record of the wasted spend.
 

package/dist/index.d.mts

@@ -295,6 +295,14 @@ interface GenericPayload {
      * adapters. Forwarded as the request body's `provider` field.
      */
     provider?: OpenRouterProviderPreferences;
+    /**
+     * OpenAI-only: forwarded as `reasoning_effort` on the request. Valid values
+     * are model-dependent (`none`/`minimal`/`low`/`medium`/`high`/`xhigh`/`max`).
+     * Reasoning-by-default models (gpt-5.6 family) 400 on /chat/completions when
+     * function tools are present unless this is explicitly `"none"` — their
+     * implicit default is `medium`. Ignored by all other adapters.
+     */
+    reasoningEffort?: string;
     /**
      * Per-request HTTP timeout in ms for the underlying provider call (applied
      * per attempt, not across retries). Honored by all adapters (Anthropic,
@@ -344,6 +352,15 @@ interface GenericPayload {
     signal?: AbortSignal;
 }
 
+/**
+ * In-process memory of OpenRouter models that rejected image input (the
+ * routing-layer 404 "No endpoints found that support image input"). Payloads
+ * for these models degrade image attachments to inline `Image (url)` text
+ * references up front — the exact pre-vision serialization — instead of
+ * burning a doomed attempt per call. Populated by the retry loop on first
+ * rejection; cleared only by process restart (exported so tests can reset it).
+ */
+declare const openRouterImageRejectedModels: Set<string>;
 declare const OPENROUTER_STREAM_TIMEOUT_MS = 600000;
 declare const OPENROUTER_NONSTREAM_TIMEOUT_MS = 180000;
 /**
@@ -358,4 +375,4 @@ declare function parseModelString(model: string): {
 };
 declare function callWithRetries(id: string | string[], aiPayload: GenericPayload, aiConfig?: OpenAIConfig | AnthropicAIConfig, retries?: number, chunkTimeoutMs?: number): Promise<ParsedResponseMessage>;
 
-export { type AnyModel, ClaudeModel, type FunctionCall, type FunctionDefinition, GPTModel, GeminiModel, type GenericMessage, type GenericPayload, GroqModel, MIN_STREAM_ATTEMPT_MS, OPENROUTER_NONSTREAM_TIMEOUT_MS, OPENROUTER_STREAM_TIMEOUT_MS, type OpenAIConfig, OpenRouterModel, type OpenRouterProviderPreferences, type ParsedResponseMessage, type Provider, type ToolResult, callWithRetries, parseModelString };
+export { type AnyModel, ClaudeModel, type FunctionCall, type FunctionDefinition, GPTModel, GeminiModel, type GenericMessage, type GenericPayload, GroqModel, MIN_STREAM_ATTEMPT_MS, OPENROUTER_NONSTREAM_TIMEOUT_MS, OPENROUTER_STREAM_TIMEOUT_MS, type OpenAIConfig, OpenRouterModel, type OpenRouterProviderPreferences, type ParsedResponseMessage, type Provider, type ToolResult, callWithRetries, openRouterImageRejectedModels, parseModelString };

package/dist/index.d.ts

@@ -295,6 +295,14 @@ interface GenericPayload {
      * adapters. Forwarded as the request body's `provider` field.
      */
     provider?: OpenRouterProviderPreferences;
+    /**
+     * OpenAI-only: forwarded as `reasoning_effort` on the request. Valid values
+     * are model-dependent (`none`/`minimal`/`low`/`medium`/`high`/`xhigh`/`max`).
+     * Reasoning-by-default models (gpt-5.6 family) 400 on /chat/completions when
+     * function tools are present unless this is explicitly `"none"` — their
+     * implicit default is `medium`. Ignored by all other adapters.
+     */
+    reasoningEffort?: string;
     /**
      * Per-request HTTP timeout in ms for the underlying provider call (applied
      * per attempt, not across retries). Honored by all adapters (Anthropic,
@@ -344,6 +352,15 @@ interface GenericPayload {
     signal?: AbortSignal;
 }
 
+/**
+ * In-process memory of OpenRouter models that rejected image input (the
+ * routing-layer 404 "No endpoints found that support image input"). Payloads
+ * for these models degrade image attachments to inline `Image (url)` text
+ * references up front — the exact pre-vision serialization — instead of
+ * burning a doomed attempt per call. Populated by the retry loop on first
+ * rejection; cleared only by process restart (exported so tests can reset it).
+ */
+declare const openRouterImageRejectedModels: Set<string>;
 declare const OPENROUTER_STREAM_TIMEOUT_MS = 600000;
 declare const OPENROUTER_NONSTREAM_TIMEOUT_MS = 180000;
 /**
@@ -358,4 +375,4 @@ declare function parseModelString(model: string): {
 };
 declare function callWithRetries(id: string | string[], aiPayload: GenericPayload, aiConfig?: OpenAIConfig | AnthropicAIConfig, retries?: number, chunkTimeoutMs?: number): Promise<ParsedResponseMessage>;
 
-export { type AnyModel, ClaudeModel, type FunctionCall, type FunctionDefinition, GPTModel, GeminiModel, type GenericMessage, type GenericPayload, GroqModel, MIN_STREAM_ATTEMPT_MS, OPENROUTER_NONSTREAM_TIMEOUT_MS, OPENROUTER_STREAM_TIMEOUT_MS, type OpenAIConfig, OpenRouterModel, type OpenRouterProviderPreferences, type ParsedResponseMessage, type Provider, type ToolResult, callWithRetries, parseModelString };
+export { type AnyModel, ClaudeModel, type FunctionCall, type FunctionDefinition, GPTModel, GeminiModel, type GenericMessage, type GenericPayload, GroqModel, MIN_STREAM_ATTEMPT_MS, OPENROUTER_NONSTREAM_TIMEOUT_MS, OPENROUTER_STREAM_TIMEOUT_MS, type OpenAIConfig, OpenRouterModel, type OpenRouterProviderPreferences, type ParsedResponseMessage, type Provider, type ToolResult, callWithRetries, openRouterImageRejectedModels, parseModelString };

package/dist/index.js

@@ -39,6 +39,7 @@ __export(index_exports, {
   OPENROUTER_STREAM_TIMEOUT_MS: () => OPENROUTER_STREAM_TIMEOUT_MS,
   OpenRouterModel: () => OpenRouterModel,
   callWithRetries: () => callWithRetries,
+  openRouterImageRejectedModels: () => openRouterImageRejectedModels,
   parseModelString: () => parseModelString
 });
 module.exports = __toCommonJS(index_exports);
@@ -346,6 +347,7 @@ async function prepareOpenAIPayload(identifier, payload) {
   const preparedPayload = {
     model: payload.model,
     messages: [],
+    reasoning_effort: payload.reasoningEffort,
     tools: (_a = payload.functions) == null ? void 0 : _a.map((fn) => ({
       type: "function",
       function: fn
@@ -416,7 +418,7 @@ async function prepareOpenAIPayload(identifier, payload) {
   return preparedPayload;
 }
 async function callOpenAIStream(id, openAiPayload, openAiConfig, chunkTimeoutMs, requestTimeoutMs = 12e4, signal) {
-  var _a, _b, _c, _d, _e, _f, _g, _h, _i, _j;
+  var _a, _b, _c, _d, _e, _f, _g, _h, _i, _j, _k, _l;
   const functionNames = openAiPayload.tools ? new Set(openAiPayload.tools.map((fn) => fn.function.name)) : null;
   const { endpoint, headers } = buildOpenAIRequestConfig(
     id,
@@ -434,7 +436,15 @@ async function callOpenAIStream(id, openAiPayload, openAiConfig, chunkTimeoutMs,
   const response = await fetch(endpoint, {
     method: "POST",
     headers,
-    body: JSON.stringify({ ...openAiPayload, stream: true }),
+    // include_usage: OpenAI only reports token usage on streams when asked,
+    // via a final choices-less chunk — without it every streamed call is
+    // invisible to usage accounting. Skipped for Azure, where older
+    // api-versions reject stream_options.
+    body: JSON.stringify({
+      ...openAiPayload,
+      stream: true,
+      ...(openAiConfig == null ? void 0 : openAiConfig.service) !== "azure" ? { stream_options: { include_usage: true } } : {}
+    }),
     // Merge (don't overwrite) the internal timeout controller with the caller's
     // cancellation signal so both an internal timeout and an external abort stop
     // the stream.
@@ -445,6 +455,7 @@ async function callOpenAIStream(id, openAiPayload, openAiConfig, chunkTimeoutMs,
   }
   let paragraph = "";
   let reasoning = "";
+  let streamUsage = null;
   const toolCallAccumulators = [];
   const reader = response.body.getReader();
   let partialChunk = "";
@@ -475,13 +486,15 @@ async function callOpenAIStream(id, openAiPayload, openAiConfig, chunkTimeoutMs,
       if (!jsonString) continue;
       if (jsonString.includes("[DONE]")) {
         clearTimeout(overallTimeout);
-        return parseStreamedResponse(
+        const parsed = parseStreamedResponse(
           id,
           paragraph,
           toolCallAccumulators,
           functionNames,
           reasoning
         );
+        parsed.usage = streamUsage;
+        return parsed;
       }
       let json;
       try {
@@ -491,6 +504,15 @@ async function callOpenAIStream(id, openAiPayload, openAiConfig, chunkTimeoutMs,
         continue;
       }
       if (!((_a = json.choices) == null ? void 0 : _a.length)) {
+        if (json.usage) {
+          streamUsage = {
+            prompt_tokens: json.usage.prompt_tokens,
+            completion_tokens: json.usage.completion_tokens,
+            total_tokens: json.usage.total_tokens,
+            cached_tokens: (_c = (_b = json.usage.prompt_tokens_details) == null ? void 0 : _b.cached_tokens) != null ? _c : 0
+          };
+          continue;
+        }
         if (json.error) {
           logger_default.error(id, "Stream error from OpenAI:", json.error);
           const error2 = new Error("Stream error: OpenAI error");
@@ -503,23 +525,23 @@ async function callOpenAIStream(id, openAiPayload, openAiConfig, chunkTimeoutMs,
         }
         continue;
       }
-      const toolCalls = (_c = (_b = json.choices[0]) == null ? void 0 : _b.delta) == null ? void 0 : _c.tool_calls;
+      const toolCalls = (_e = (_d = json.choices[0]) == null ? void 0 : _d.delta) == null ? void 0 : _e.tool_calls;
       if (toolCalls) {
         for (const toolCall of toolCalls) {
-          const idx = (_d = toolCall.index) != null ? _d : 0;
+          const idx = (_f = toolCall.index) != null ? _f : 0;
           while (toolCallAccumulators.length <= idx) {
             toolCallAccumulators.push({ name: "", arguments: "" });
           }
           if (toolCall.id) toolCallAccumulators[idx].id = toolCall.id;
-          if ((_e = toolCall.function) == null ? void 0 : _e.name)
+          if ((_g = toolCall.function) == null ? void 0 : _g.name)
             toolCallAccumulators[idx].name += toolCall.function.name;
-          if ((_f = toolCall.function) == null ? void 0 : _f.arguments)
+          if ((_h = toolCall.function) == null ? void 0 : _h.arguments)
             toolCallAccumulators[idx].arguments += toolCall.function.arguments;
         }
       }
-      const text = (_h = (_g = json.choices[0]) == null ? void 0 : _g.delta) == null ? void 0 : _h.content;
+      const text = (_j = (_i = json.choices[0]) == null ? void 0 : _i.delta) == null ? void 0 : _j.content;
       if (text) paragraph += text;
-      const reasoningDelta = (_j = (_i = json.choices[0]) == null ? void 0 : _i.delta) == null ? void 0 : _j.reasoning;
+      const reasoningDelta = (_l = (_k = json.choices[0]) == null ? void 0 : _k.delta) == null ? void 0 : _l.reasoning;
       if (reasoningDelta) reasoning += reasoningDelta;
     }
   }
@@ -1232,7 +1254,7 @@ async function callGoogleAIWithRetries(id, payload, retries = 5, requestTimeoutM
 function normalizeMessageContent(content) {
   return Array.isArray(content) ? content.map((c) => c.type === "text" ? c.text : `[${c.type}]`).join("\n") : content;
 }
-function prepareOpenAICompatMessages(messages) {
+function prepareOpenAICompatMessages(messages, opts = {}) {
   var _a;
   const out = [];
   for (const message of messages) {
@@ -1254,6 +1276,22 @@ function prepareOpenAICompatMessages(messages) {
       role: message.role,
       content
     };
+    if (opts.imageParts) {
+      const imageBlocks = (message.files || []).filter(
+        (file) => ALLOWED_IMAGE_MIME_TYPES.includes(file.mimeType) && (file.url || file.data)
+      ).map((file) => ({
+        type: "image_url",
+        image_url: {
+          url: file.url || `data:${file.mimeType};base64,${file.data}`
+        }
+      }));
+      if (imageBlocks.length) {
+        outMessage.content = [
+          ...content ? [{ type: "text", text: content }] : [],
+          ...imageBlocks
+        ];
+      }
+    }
     if ((_a = message.functionCalls) == null ? void 0 : _a.length) {
       outMessage.tool_calls = message.functionCalls.map((fc, i) => {
         var _a2;
@@ -1266,7 +1304,8 @@ function prepareOpenAICompatMessages(messages) {
           }
         };
       });
-      if (!content) outMessage.content = null;
+      if (!content && !Array.isArray(outMessage.content))
+        outMessage.content = null;
     }
     if (message.reasoning) outMessage.reasoning = message.reasoning;
     const reasoningDetails = filterOpenAICompatReasoningDetails(
@@ -1360,11 +1399,14 @@ async function callGroqWithRetries(id, payload, retries = 5, requestTimeoutMs =
     signal
   });
 }
+var openRouterImageRejectedModels = /* @__PURE__ */ new Set();
 function prepareOpenRouterPayload(payload) {
   var _a;
   return {
     model: payload.model,
-    messages: prepareOpenAICompatMessages(payload.messages),
+    messages: prepareOpenAICompatMessages(payload.messages, {
+      imageParts: !openRouterImageRejectedModels.has(String(payload.model))
+    }),
     tools: (_a = payload.functions) == null ? void 0 : _a.map((fn) => ({
       type: "function",
       function: fn
@@ -1754,6 +1796,11 @@ function moderationEvictionSlug(error2, payload) {
 }
 var MIN_STREAM_ATTEMPT_MS = 1e4;
 var estimateTokens = (text) => Math.round(text.length / 4);
+function isImageInputRejection(error2) {
+  var _a, _b, _c, _d;
+  const body = (_c = error2 == null ? void 0 : error2.data) != null ? _c : (_b = (_a = error2 == null ? void 0 : error2.response) == null ? void 0 : _a.data) == null ? void 0 : _b.error;
+  return /no endpoints found.*image input/i.test(String((_d = body == null ? void 0 : body.message) != null ? _d : ""));
+}
 function streamAttemptBudgetMs(options) {
   if (options.streamDeadlineAt === void 0) return options.streamTimeoutMs;
   const remaining = options.streamDeadlineAt - Date.now();
@@ -1804,6 +1851,16 @@ async function callOpenRouterWithRetries(id, payload, retries = 5, options, sign
           `OpenRouter moderation eviction: ignoring provider "${slug}" for remaining attempts`
         );
       }
+      if (isImageInputRejection(error2) && options.genericMessages) {
+        openRouterImageRejectedModels.add(String(payload.model));
+        payload.messages = prepareOpenAICompatMessages(
+          options.genericMessages
+        );
+        logger_default.log(
+          id,
+          `OpenRouter: ${payload.model} has no image-capable endpoints; retrying with images as text refs`
+        );
+      }
       throw error2;
     }),
     { retries, signal }
@@ -1904,7 +1961,8 @@ async function callWithRetries(id, aiPayload, aiConfig, retries = 5, chunkTimeou
             streamTimeoutMs: (_c = aiPayload.streamTimeoutMs) != null ? _c : OPENROUTER_STREAM_TIMEOUT_MS,
             streamDeadlineAt: aiPayload.streamDeadlineAt,
             requestTimeoutMs: (_d = aiPayload.requestTimeoutMs) != null ? _d : OPENROUTER_NONSTREAM_TIMEOUT_MS,
-            chunkTimeoutMs
+            chunkTimeoutMs,
+            genericMessages: routingPayload.messages
           },
           signal
         );
@@ -1949,6 +2007,7 @@ async function callWithRetries(id, aiPayload, aiConfig, retries = 5, chunkTimeou
   OPENROUTER_STREAM_TIMEOUT_MS,
   OpenRouterModel,
   callWithRetries,
+  openRouterImageRejectedModels,
   parseModelString
 });
 //# sourceMappingURL=index.js.map