190proof 1.0.115 → 1.0.117

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
@@ -280,7 +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.
+- `payload.reasoningEffort`: `string` - OpenAI and OpenRouter: forwarded as `reasoning_effort` on the request. Valid values are model-dependent (`none`/`minimal`/`low`/`medium`/`high`/`xhigh`/`max`). Direct OpenAI: 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`. Via OpenRouter the same models accept tools at any effort (OpenRouter fronts `/v1/responses`), so omitting this runs them at their native default. 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

@@ -296,11 +296,15 @@ interface GenericPayload {
      */
     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.
+     * OpenAI and OpenRouter: forwarded as `reasoning_effort` on the request.
+     * Valid values are model-dependent
+     * (`none`/`minimal`/`low`/`medium`/`high`/`xhigh`/`max`).
+     * Direct OpenAI: 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`. Via OpenRouter
+     * the same models accept tools at any effort (OpenRouter fronts
+     * /v1/responses), so omitting this runs them at their native default.
+     * Ignored by all other adapters.
      */
     reasoningEffort?: string;
     /**
@@ -352,6 +356,15 @@ 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;
 /**
@@ -366,4 +379,4 @@ 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

@@ -296,11 +296,15 @@ interface GenericPayload {
      */
     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.
+     * OpenAI and OpenRouter: forwarded as `reasoning_effort` on the request.
+     * Valid values are model-dependent
+     * (`none`/`minimal`/`low`/`medium`/`high`/`xhigh`/`max`).
+     * Direct OpenAI: 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`. Via OpenRouter
+     * the same models accept tools at any effort (OpenRouter fronts
+     * /v1/responses), so omitting this runs them at their native default.
+     * Ignored by all other adapters.
      */
     reasoningEffort?: string;
     /**
@@ -352,6 +356,15 @@ 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;
 /**
@@ -366,4 +379,4 @@ 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 };