190proof 1.0.117 → 1.0.119

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

@@ -27,7 +27,7 @@ var __toESM = (mod, isNodeMode, target) => (target = mod != null ? __create(__ge
 ));
 var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: true }), mod);
 
-// index.ts
+// src/index.ts
 var index_exports = {};
 __export(index_exports, {
   ClaudeModel: () => ClaudeModel,
@@ -44,7 +44,7 @@ __export(index_exports, {
 });
 module.exports = __toCommonJS(index_exports);
 
-// interfaces.ts
+// src/interfaces.ts
 var ClaudeModel = /* @__PURE__ */ ((ClaudeModel2) => {
   ClaudeModel2["HAIKU_3"] = "claude-3-haiku-20240307";
   ClaudeModel2["SONNET_3"] = "claude-3-sonnet-20240229";
@@ -103,7 +103,7 @@ var GeminiModel = /* @__PURE__ */ ((GeminiModel2) => {
   return GeminiModel2;
 })(GeminiModel || {});
 
-// logger.ts
+// src/logger.ts
 function formatIdentifier(identifier) {
   if (Array.isArray(identifier)) {
     return identifier.map((id) => `[${id}]`).join(" ");
@@ -128,11 +128,11 @@ var logger_default = {
   error
 };
 
-// index.ts
+// src/index.ts
 var import_client_bedrock_runtime = require("@aws-sdk/client-bedrock-runtime");
 var import_axios = __toESM(require("axios"));
 
-// utils.ts
+// src/utils.ts
 function timeout(ms) {
   return new Promise((resolve) => setTimeout(resolve, ms));
 }
@@ -142,7 +142,7 @@ function isHeicImage(name, mime) {
   return ["heic", "heif", "heics"].includes(extension) || !!(mime && ["image/heic", "image/heif", "image/heic-sequence"].includes(mime));
 }
 
-// index.ts
+// src/index.ts
 var sharp = require("sharp");
 var decode = require("heic-decode");
 function anySignal(signals) {
@@ -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>/;
@@ -1561,6 +1565,14 @@ async function callOpenRouterStream(id, payload, streamTimeoutMs = OPENROUTER_ST
   let finishReason = null;
   let sawDone = false;
   let dataChunks = 0;
+  const RAW_HEAD_MAX = 3e3;
+  const RAW_TAIL_MAX = 1500;
+  let rawHead = "";
+  let rawTail = "";
+  let generationId;
+  const unreadKeys = /* @__PURE__ */ new Set();
+  const KNOWN_DELTA_KEYS = /* @__PURE__ */ new Set(["role", "content", "reasoning", "reasoning_details", "tool_calls"]);
+  const KNOWN_CHOICE_KEYS = /* @__PURE__ */ new Set(["index", "delta", "finish_reason", "native_finish_reason", "logprobs"]);
   try {
     armStallTimer();
     const response = await fetch(openRouterEndpoint(), {
@@ -1612,6 +1624,11 @@ async function callOpenRouterStream(id, payload, streamTimeoutMs = OPENROUTER_ST
           sawDone = true;
           break outer;
         }
+        if (rawHead.length < RAW_HEAD_MAX) {
+          rawHead += dataStr.slice(0, RAW_HEAD_MAX - rawHead.length) + "\n";
+        } else {
+          rawTail = (rawTail + dataStr + "\n").slice(-RAW_TAIL_MAX);
+        }
         let json;
         try {
           json = JSON.parse(dataStr);
@@ -1633,6 +1650,7 @@ async function callOpenRouterStream(id, payload, streamTimeoutMs = OPENROUTER_ST
           throw error2;
         }
         if (json.provider) provider = json.provider;
+        if (json.id && !generationId) generationId = json.id;
         let useful = false;
         if (json.usage) {
           usage = json.usage;
@@ -1641,6 +1659,12 @@ async function callOpenRouterStream(id, payload, streamTimeoutMs = OPENROUTER_ST
         const choice = (_c = json.choices) == null ? void 0 : _c[0];
         if (choice) {
           const delta = (_d = choice.delta) != null ? _d : {};
+          for (const k of Object.keys(choice)) {
+            if (!KNOWN_CHOICE_KEYS.has(k) && choice[k] != null) unreadKeys.add(`choice.${k}`);
+          }
+          for (const k of Object.keys(delta)) {
+            if (!KNOWN_DELTA_KEYS.has(k) && delta[k] != null && delta[k] !== "") unreadKeys.add(`delta.${k}`);
+          }
           if (delta.content) {
             paragraph += delta.content;
             useful = true;
@@ -1697,7 +1721,12 @@ async function callOpenRouterStream(id, payload, streamTimeoutMs = OPENROUTER_ST
         usage,
         paragraph: paragraph.slice(0, 500),
         reasoningChars: reasoning.length,
-        toolCalls
+        toolCalls,
+        generationId,
+        dataChunks,
+        unreadKeys: [...unreadKeys],
+        rawHead,
+        rawTail: rawTail || void 0
       })
     });
   } catch (error2) {