190proof 1.0.117 → 1.0.118

package/README.md

@@ -290,7 +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 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.
+- `payload.reasoningEffort`: `string` - OpenAI and OpenRouter reasoning effort. Valid values are model-dependent (`none`/`minimal`/`low`/`medium`/`high`/`xhigh`/`max`). OpenRouter: sent as the nested `reasoning: { effort }` object — the canonical form, and the only one that accepts `max` (the flat `reasoning_effort` field caps at `xhigh`). Direct OpenAI: sent flat as `reasoning_effort`; `max` is rejected there, and 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

@@ -296,14 +296,17 @@ interface GenericPayload {
      */
     provider?: OpenRouterProviderPreferences;
     /**
-     * OpenAI and OpenRouter: forwarded as `reasoning_effort` on the request.
-     * Valid values are model-dependent
+     * OpenAI and OpenRouter reasoning effort. 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.
+     * OpenRouter: sent as the nested `reasoning: { effort }` object (the
+     * canonical form, and the only one where `max` is accepted — flat
+     * `reasoning_effort` caps at `xhigh`, and on gpt-5.6-luna nested `max`
+     * measurably out-reasons `xhigh` despite docs calling them aliases).
+     * Direct OpenAI: sent flat as `reasoning_effort`; `max` is rejected there,
+     * and 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).
      * Ignored by all other adapters.
      */
     reasoningEffort?: string;

package/dist/index.d.ts

@@ -296,14 +296,17 @@ interface GenericPayload {
      */
     provider?: OpenRouterProviderPreferences;
     /**
-     * OpenAI and OpenRouter: forwarded as `reasoning_effort` on the request.
-     * Valid values are model-dependent
+     * OpenAI and OpenRouter reasoning effort. 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.
+     * OpenRouter: sent as the nested `reasoning: { effort }` object (the
+     * canonical form, and the only one where `max` is accepted — flat
+     * `reasoning_effort` caps at `xhigh`, and on gpt-5.6-luna nested `max`
+     * measurably out-reasons `xhigh` despite docs calling them aliases).
+     * Direct OpenAI: sent flat as `reasoning_effort`; `max` is rejected there,
+     * and 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).
      * Ignored by all other adapters.
      */
     reasoningEffort?: string;

package/dist/index.js

@@ -1414,7 +1414,11 @@ function prepareOpenRouterPayload(payload) {
     tool_choice: payload.function_call ? typeof payload.function_call === "string" ? payload.function_call : { type: "function", function: payload.function_call } : void 0,
     temperature: payload.temperature,
     provider: payload.provider,
-    reasoning_effort: payload.reasoningEffort
+    // Nested form is OpenRouter's canonical param and the only one that
+    // accepts "max" (flat reasoning_effort rejects it — OpenAI enum tops out
+    // at xhigh, and measured on gpt-5.6-luna nested max buys more reasoning
+    // tokens than xhigh despite the docs calling them equivalent).
+    reasoning: payload.reasoningEffort ? { effort: payload.reasoningEffort } : void 0
   };
 }
 var DSML_ENVELOPE_RE = /<｜+DSML｜+tool_calls>/;