190proof 1.0.108 → 1.0.110

package/README.md

@@ -273,9 +273,13 @@ Main function to make requests to any supported AI provider.
 - `retries`: `number` - Number of retry attempts (default: 5)
 - `chunkTimeoutMs`: `number` - Timeout for streaming chunks in ms (default: 15000)
 
-Two optional per-request knobs live on `payload` (`GenericPayload`):
+Optional per-request knobs live on `payload` (`GenericPayload`):
 
-- `payload.requestTimeoutMs`: `number` - Per-attempt HTTP timeout in ms (default: 120000), honored by every adapter.
+- `payload.requestTimeoutMs`: `number` - Per-attempt HTTP timeout in ms (default: 120000), honored by every adapter — except streaming OpenRouter attempts, which it deliberately does NOT bound (see below). For OpenRouter's non-streaming transport the default is 180000.
+- `payload.streaming`: `boolean` - OpenRouter-only (default: true). Streams the completion over SSE. A streaming attempt is bounded by two independent timers instead of `requestTimeoutMs`: `streamTimeoutMs` (total wall clock, default 600000) and the per-useful-chunk stall timeout (`chunkTimeoutMs` argument, default 15000). A chunk is "useful" only if it advances content, reasoning, tool-call fragments, finish_reason, or usage — SSE comment keep-alives (`: OPENROUTER PROCESSING`) and role-only deltas don't reset the stall timer, so a hung provider dies within one stall window while a healthy long generation can run to the total budget. Set `streaming: false` for the old single-JSON-body transport.
+- `payload.streamTimeoutMs`: `number` - OpenRouter-only: total wall-clock budget per streaming attempt (default: 600000).
+
+OpenRouter retries also perform **moderation eviction**: a provider content-moderation rejection (e.g. "Upstream error from Alibaba: Output data may contain inappropriate content.") is deterministic for a given payload, so on the first one the refusing provider is removed from the request's provider preferences (`ignore` += slug, `order` -= slug) and every remaining attempt reroutes to the next provider. Non-moderation errors retry with unchanged preferences, and `fallbackModel` still applies if the whole pool refuses.
 - `payload.signal`: `AbortSignal` - Caller-supplied cancellation. When it aborts, the in-flight provider request is cancelled and `callWithRetries` **rejects immediately — it does not retry or fall back** (both the retry loop and the fallback branch bail on `signal.aborted`). Threaded to the underlying fetch/axios/SDK call of each provider.
 
 #### Returns

package/dist/index.d.mts

@@ -279,6 +279,24 @@ interface GenericPayload {
      * valid response isn't cut short.
      */
     requestTimeoutMs?: number;
+    /**
+     * OpenRouter-only: stream the completion over SSE instead of waiting for a
+     * single JSON body. Defaults to true. Streaming attempts are bounded by
+     * `streamTimeoutMs` (total) plus a per-useful-chunk stall timeout — NOT by
+     * `requestTimeoutMs`, which only governs non-streaming attempts (default
+     * 180s for OpenRouter). Set to false to force the old non-streaming path.
+     */
+    streaming?: boolean;
+    /**
+     * OpenRouter-only: total wall-clock budget in ms for one streaming attempt
+     * (connect + full generation). Defaults to 600s. Independent of
+     * `requestTimeoutMs` by design: a healthy long generation keeps streaming
+     * useful chunks and may run far past any sane non-streaming deadline, while
+     * a hung one is killed much earlier by the per-useful-chunk stall timeout
+     * (`chunkTimeoutMs` argument of `callWithRetries`, default 15s — reset only
+     * by chunks that advance the output, never by keep-alive bytes/comments).
+     */
+    streamTimeoutMs?: number;
     /**
      * Optional caller-supplied cancellation signal. When it aborts, the in-flight
      * provider request is cancelled and `callWithRetries` rejects immediately —
@@ -289,10 +307,12 @@ interface GenericPayload {
     signal?: AbortSignal;
 }
 
+declare const OPENROUTER_STREAM_TIMEOUT_MS = 600000;
+declare const OPENROUTER_NONSTREAM_TIMEOUT_MS = 180000;
 declare function parseModelString(model: string): {
     provider: Provider;
     modelId: 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, 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, OPENROUTER_NONSTREAM_TIMEOUT_MS, OPENROUTER_STREAM_TIMEOUT_MS, type OpenAIConfig, OpenRouterModel, type OpenRouterProviderPreferences, type ParsedResponseMessage, type Provider, type ToolResult, callWithRetries, parseModelString };

package/dist/index.d.ts

@@ -279,6 +279,24 @@ interface GenericPayload {
      * valid response isn't cut short.
      */
     requestTimeoutMs?: number;
+    /**
+     * OpenRouter-only: stream the completion over SSE instead of waiting for a
+     * single JSON body. Defaults to true. Streaming attempts are bounded by
+     * `streamTimeoutMs` (total) plus a per-useful-chunk stall timeout — NOT by
+     * `requestTimeoutMs`, which only governs non-streaming attempts (default
+     * 180s for OpenRouter). Set to false to force the old non-streaming path.
+     */
+    streaming?: boolean;
+    /**
+     * OpenRouter-only: total wall-clock budget in ms for one streaming attempt
+     * (connect + full generation). Defaults to 600s. Independent of
+     * `requestTimeoutMs` by design: a healthy long generation keeps streaming
+     * useful chunks and may run far past any sane non-streaming deadline, while
+     * a hung one is killed much earlier by the per-useful-chunk stall timeout
+     * (`chunkTimeoutMs` argument of `callWithRetries`, default 15s — reset only
+     * by chunks that advance the output, never by keep-alive bytes/comments).
+     */
+    streamTimeoutMs?: number;
     /**
      * Optional caller-supplied cancellation signal. When it aborts, the in-flight
      * provider request is cancelled and `callWithRetries` rejects immediately —
@@ -289,10 +307,12 @@ interface GenericPayload {
     signal?: AbortSignal;
 }
 
+declare const OPENROUTER_STREAM_TIMEOUT_MS = 600000;
+declare const OPENROUTER_NONSTREAM_TIMEOUT_MS = 180000;
 declare function parseModelString(model: string): {
     provider: Provider;
     modelId: 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, 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, OPENROUTER_NONSTREAM_TIMEOUT_MS, OPENROUTER_STREAM_TIMEOUT_MS, type OpenAIConfig, OpenRouterModel, type OpenRouterProviderPreferences, type ParsedResponseMessage, type Provider, type ToolResult, callWithRetries, parseModelString };