omnigateway 0.13.3 → 0.13.4

package/README.md

@@ -397,6 +397,11 @@ Worth knowing before you deploy it:
 - **The gateway does not know which model accepts which request shape.** An
   unsupported combination surfaces as the provider's own 400 rather than being
   caught earlier.
+- **Thinking controls retain their wire intent.** Anthropic `between_tools` passes
+  through; other providers treat it as off and report a degradation. Unparsed
+  Anthropic thinking fields (including beta controls) pass through only to
+  Anthropic; supply any required beta header yourself. Hidden adaptive display
+  omits generated Responses reasoning summaries, without inventing a default.
 - Not in scope: semantic caching, billing.
 
 What is missing on purpose, what is missing for now, and what is designed but not

package/bin/omni.js

@@ -13963,6 +13963,9 @@ function toWire(req, model, opts) {
     addAutoCacheBreakpoints(body, req, note);
   if (req.reasoning !== undefined) {
     switch (req.reasoning.mode) {
+      case "betweenTools":
+        body.thinking = { type: "between_tools" };
+        break;
       case "adaptive":
         if (anthropicReasoningForm(model) === "budget") {
           body.thinking = { type: "disabled" };
@@ -14001,7 +14004,10 @@ function toWire(req, model, opts) {
   } else if (!nativeFormat && read.kind === "tooDeep") {
     note("anthropic:response-schema-too-deep");
   }
-  const vendorRaw = req.vendor?.anthropic ?? {};
+  const { thinking: thinkingExtras, ...vendorRaw } = req.vendor?.anthropic ?? {};
+  if (body.thinking !== undefined && isRecord2(thinkingExtras)) {
+    body.thinking = { ...thinkingExtras, ...body.thinking };
+  }
   const vendorConfig = vendorRaw.output_config;
   const encoderConfig = isRecord2(body.output_config) ? body.output_config : undefined;
   let unreadableFormat = false;
@@ -15178,6 +15184,10 @@ function toAntigravityWire(req, model, identity) {
   }
   if (req.reasoning !== undefined) {
     switch (req.reasoning.mode) {
+      case "betweenTools":
+        note("antigravity:reasoning-between-tools-as-off");
+        generationConfig.thinkingConfig = { thinkingBudget: 0, includeThoughts: false };
+        break;
       case "off":
         generationConfig.thinkingConfig = { thinkingBudget: 0, includeThoughts: false };
         break;
@@ -15195,7 +15205,7 @@ function toAntigravityWire(req, model, identity) {
         break;
       case "adaptive":
         generationConfig.thinkingConfig = {
-          includeThoughts: req.reasoning.display !== "omitted"
+          includeThoughts: req.reasoning.display !== "omitted" && req.reasoning.display !== "updates"
         };
         if (req.reasoning.effort !== undefined)
           note("antigravity:reasoning-effort-dropped");
@@ -17070,6 +17080,8 @@ function toCustomChatWire(req, model) {
   }
   if (req.toolChoice !== undefined)
     body.tool_choice = encodeChatToolChoice(req.toolChoice);
+  if (req.reasoning?.mode === "betweenTools")
+    note("custom:reasoning-between-tools-as-off");
   const effort = customEffort(req.reasoning);
   if (effort !== undefined)
     body.reasoning_effort = effort;
@@ -17200,9 +17212,14 @@ function toCustomResponsesWire(req, model) {
   }
   if (req.toolChoice !== undefined)
     body.tool_choice = encodeResponsesToolChoice(req.toolChoice);
+  if (req.reasoning?.mode === "betweenTools")
+    note("custom:reasoning-between-tools-as-off");
   const effort = customEffort(req.reasoning);
   if (effort !== undefined)
-    body.reasoning = { effort, summary: "auto" };
+    body.reasoning = {
+      effort,
+      ...req.reasoning?.mode === "adaptive" && (req.reasoning.display === "omitted" || req.reasoning.display === "updates") ? {} : { summary: "auto" }
+    };
   else if (req.reasoning?.mode === "budget")
     note("custom:reasoning-budget-dropped");
   Object.assign(body, req.vendor?.openai ?? {});
@@ -17595,12 +17612,17 @@ function toGrokWire(req, model) {
   }
   if (req.toolChoice !== undefined)
     body.tool_choice = encodeToolChoice3(req.toolChoice);
-  if (req.reasoning !== undefined && req.reasoning.mode !== "off") {
+  if (req.reasoning?.mode === "betweenTools")
+    note("grok:reasoning-between-tools-as-off");
+  if (req.reasoning !== undefined && req.reasoning.mode !== "off" && req.reasoning.mode !== "betweenTools") {
     if (req.reasoning.mode === "budget") {
       note("grok:reasoning-budget-dropped");
     } else {
       const effort = req.reasoning.effort ?? "medium";
-      body.reasoning = { effort, summary: "concise" };
+      body.reasoning = {
+        effort,
+        ...req.reasoning.display === "omitted" || req.reasoning.display === "updates" ? {} : { summary: "concise" }
+      };
     }
   }
   Object.assign(body, req.vendor?.grok ?? {});
@@ -18060,7 +18082,9 @@ function toKiloWire(req, model) {
   }
   if (req.toolChoice !== undefined)
     body.tool_choice = encodeToolChoice4(req.toolChoice);
-  if (req.reasoning !== undefined && req.reasoning.mode !== "off") {
+  if (req.reasoning?.mode === "betweenTools")
+    note("kilo:reasoning-between-tools-as-off");
+  if (req.reasoning !== undefined && req.reasoning.mode !== "off" && req.reasoning.mode !== "betweenTools") {
     if (req.reasoning.mode === "budget") {
       body.reasoning = { max_tokens: req.reasoning.budgetTokens };
     } else {
@@ -18338,6 +18362,8 @@ function toChatWire(req, model, vendor = "kimi") {
   }
   if (req.toolChoice !== undefined)
     body.tool_choice = encodeToolChoice5(req.toolChoice);
+  if (req.reasoning?.mode === "betweenTools")
+    note("kimi:reasoning-between-tools-as-off");
   if (req.reasoning !== undefined)
     note("kimi:reasoning-dropped");
   Object.assign(body, req.vendor?.[vendor] ?? {});
@@ -18760,12 +18786,17 @@ function toMuseWire(req, model) {
   }
   if (req.toolChoice !== undefined)
     body.tool_choice = encodeToolChoice6(req.toolChoice);
-  if (req.reasoning !== undefined && req.reasoning.mode !== "off") {
+  if (req.reasoning?.mode === "betweenTools")
+    note("muse:reasoning-between-tools-as-off");
+  if (req.reasoning !== undefined && req.reasoning.mode !== "off" && req.reasoning.mode !== "betweenTools") {
     if (req.reasoning.mode === "budget") {
       note("muse:reasoning-budget-dropped");
     } else {
       const effort = req.reasoning.effort ?? "medium";
-      body.reasoning = { effort, summary: "auto" };
+      body.reasoning = {
+        effort,
+        ...req.reasoning.display === "omitted" || req.reasoning.display === "updates" ? {} : { summary: "auto" }
+      };
     }
   }
   const key = suppliedKey(req) ?? cacheKey(req, instructions ?? "", input[0]);
@@ -19196,12 +19227,17 @@ function toResponsesWire(req, model, opts = { oauth: false }) {
   }
   if (req.toolChoice !== undefined)
     body.tool_choice = encodeToolChoice7(req.toolChoice);
-  if (req.reasoning !== undefined && req.reasoning.mode !== "off") {
+  if (req.reasoning?.mode === "betweenTools")
+    degradations.push("openai:reasoning-between-tools-as-off");
+  if (req.reasoning !== undefined && req.reasoning.mode !== "off" && req.reasoning.mode !== "betweenTools") {
     if (req.reasoning.mode === "budget") {
       degradations.push("openai:reasoning-budget-dropped");
     } else {
       const effort = req.reasoning.effort ?? "medium";
-      body.reasoning = { effort, summary: "auto" };
+      body.reasoning = {
+        effort,
+        ...req.reasoning.display === "omitted" || req.reasoning.display === "updates" ? {} : { summary: "auto" }
+      };
     }
   }
   const key = suppliedKey2(req) ?? cacheKey2(req, instructions ?? "", input[0]);
@@ -19794,7 +19830,7 @@ function requiredCapabilities(request) {
   return {
     tools: (request.tools?.length ?? 0) > 0,
     images,
-    reasoning: request.reasoning !== undefined && request.reasoning.mode !== "off"
+    reasoning: request.reasoning !== undefined && request.reasoning.mode !== "off" && request.reasoning.mode !== "betweenTools"
   };
 }
 function cooldownMs(failures, threshold, base) {
@@ -25073,7 +25109,7 @@ function createServiceDeps(input) {
 }
 
 // apps/cli/src/version.ts
-var VERSION = "0.13.3";
+var VERSION = "0.13.4";
 
 // apps/cli/src/commands/service.ts
 function gatewayArgv(root) {

package/gateway.js

@@ -19628,6 +19628,9 @@ function toWire(req, model, opts) {
     addAutoCacheBreakpoints(body, req, note);
   if (req.reasoning !== undefined) {
     switch (req.reasoning.mode) {
+      case "betweenTools":
+        body.thinking = { type: "between_tools" };
+        break;
       case "adaptive":
         if (anthropicReasoningForm(model) === "budget") {
           body.thinking = { type: "disabled" };
@@ -19666,7 +19669,10 @@ function toWire(req, model, opts) {
   } else if (!nativeFormat && read.kind === "tooDeep") {
     note("anthropic:response-schema-too-deep");
   }
-  const vendorRaw = req.vendor?.anthropic ?? {};
+  const { thinking: thinkingExtras, ...vendorRaw } = req.vendor?.anthropic ?? {};
+  if (body.thinking !== undefined && isRecord2(thinkingExtras)) {
+    body.thinking = { ...thinkingExtras, ...body.thinking };
+  }
   const vendorConfig = vendorRaw.output_config;
   const encoderConfig = isRecord2(body.output_config) ? body.output_config : undefined;
   let unreadableFormat = false;
@@ -20843,6 +20849,10 @@ function toAntigravityWire(req, model, identity) {
   }
   if (req.reasoning !== undefined) {
     switch (req.reasoning.mode) {
+      case "betweenTools":
+        note("antigravity:reasoning-between-tools-as-off");
+        generationConfig.thinkingConfig = { thinkingBudget: 0, includeThoughts: false };
+        break;
       case "off":
         generationConfig.thinkingConfig = { thinkingBudget: 0, includeThoughts: false };
         break;
@@ -20860,7 +20870,7 @@ function toAntigravityWire(req, model, identity) {
         break;
       case "adaptive":
         generationConfig.thinkingConfig = {
-          includeThoughts: req.reasoning.display !== "omitted"
+          includeThoughts: req.reasoning.display !== "omitted" && req.reasoning.display !== "updates"
         };
         if (req.reasoning.effort !== undefined)
           note("antigravity:reasoning-effort-dropped");
@@ -22735,6 +22745,8 @@ function toCustomChatWire(req, model) {
   }
   if (req.toolChoice !== undefined)
     body.tool_choice = encodeChatToolChoice(req.toolChoice);
+  if (req.reasoning?.mode === "betweenTools")
+    note("custom:reasoning-between-tools-as-off");
   const effort = customEffort(req.reasoning);
   if (effort !== undefined)
     body.reasoning_effort = effort;
@@ -22865,9 +22877,14 @@ function toCustomResponsesWire(req, model) {
   }
   if (req.toolChoice !== undefined)
     body.tool_choice = encodeResponsesToolChoice(req.toolChoice);
+  if (req.reasoning?.mode === "betweenTools")
+    note("custom:reasoning-between-tools-as-off");
   const effort = customEffort(req.reasoning);
   if (effort !== undefined)
-    body.reasoning = { effort, summary: "auto" };
+    body.reasoning = {
+      effort,
+      ...req.reasoning?.mode === "adaptive" && (req.reasoning.display === "omitted" || req.reasoning.display === "updates") ? {} : { summary: "auto" }
+    };
   else if (req.reasoning?.mode === "budget")
     note("custom:reasoning-budget-dropped");
   Object.assign(body, req.vendor?.openai ?? {});
@@ -23260,12 +23277,17 @@ function toGrokWire(req, model) {
   }
   if (req.toolChoice !== undefined)
     body.tool_choice = encodeToolChoice3(req.toolChoice);
-  if (req.reasoning !== undefined && req.reasoning.mode !== "off") {
+  if (req.reasoning?.mode === "betweenTools")
+    note("grok:reasoning-between-tools-as-off");
+  if (req.reasoning !== undefined && req.reasoning.mode !== "off" && req.reasoning.mode !== "betweenTools") {
     if (req.reasoning.mode === "budget") {
       note("grok:reasoning-budget-dropped");
     } else {
       const effort = req.reasoning.effort ?? "medium";
-      body.reasoning = { effort, summary: "concise" };
+      body.reasoning = {
+        effort,
+        ...req.reasoning.display === "omitted" || req.reasoning.display === "updates" ? {} : { summary: "concise" }
+      };
     }
   }
   Object.assign(body, req.vendor?.grok ?? {});
@@ -23725,7 +23747,9 @@ function toKiloWire(req, model) {
   }
   if (req.toolChoice !== undefined)
     body.tool_choice = encodeToolChoice4(req.toolChoice);
-  if (req.reasoning !== undefined && req.reasoning.mode !== "off") {
+  if (req.reasoning?.mode === "betweenTools")
+    note("kilo:reasoning-between-tools-as-off");
+  if (req.reasoning !== undefined && req.reasoning.mode !== "off" && req.reasoning.mode !== "betweenTools") {
     if (req.reasoning.mode === "budget") {
       body.reasoning = { max_tokens: req.reasoning.budgetTokens };
     } else {
@@ -24003,6 +24027,8 @@ function toChatWire(req, model, vendor = "kimi") {
   }
   if (req.toolChoice !== undefined)
     body.tool_choice = encodeToolChoice5(req.toolChoice);
+  if (req.reasoning?.mode === "betweenTools")
+    note("kimi:reasoning-between-tools-as-off");
   if (req.reasoning !== undefined)
     note("kimi:reasoning-dropped");
   Object.assign(body, req.vendor?.[vendor] ?? {});
@@ -24425,12 +24451,17 @@ function toMuseWire(req, model) {
   }
   if (req.toolChoice !== undefined)
     body.tool_choice = encodeToolChoice6(req.toolChoice);
-  if (req.reasoning !== undefined && req.reasoning.mode !== "off") {
+  if (req.reasoning?.mode === "betweenTools")
+    note("muse:reasoning-between-tools-as-off");
+  if (req.reasoning !== undefined && req.reasoning.mode !== "off" && req.reasoning.mode !== "betweenTools") {
     if (req.reasoning.mode === "budget") {
       note("muse:reasoning-budget-dropped");
     } else {
       const effort = req.reasoning.effort ?? "medium";
-      body.reasoning = { effort, summary: "auto" };
+      body.reasoning = {
+        effort,
+        ...req.reasoning.display === "omitted" || req.reasoning.display === "updates" ? {} : { summary: "auto" }
+      };
     }
   }
   const key = suppliedKey(req) ?? cacheKey(req, instructions ?? "", input[0]);
@@ -24861,12 +24892,17 @@ function toResponsesWire(req, model, opts = { oauth: false }) {
   }
   if (req.toolChoice !== undefined)
     body.tool_choice = encodeToolChoice7(req.toolChoice);
-  if (req.reasoning !== undefined && req.reasoning.mode !== "off") {
+  if (req.reasoning?.mode === "betweenTools")
+    degradations.push("openai:reasoning-between-tools-as-off");
+  if (req.reasoning !== undefined && req.reasoning.mode !== "off" && req.reasoning.mode !== "betweenTools") {
     if (req.reasoning.mode === "budget") {
       degradations.push("openai:reasoning-budget-dropped");
     } else {
       const effort = req.reasoning.effort ?? "medium";
-      body.reasoning = { effort, summary: "auto" };
+      body.reasoning = {
+        effort,
+        ...req.reasoning.display === "omitted" || req.reasoning.display === "updates" ? {} : { summary: "auto" }
+      };
     }
   }
   const key = suppliedKey2(req) ?? cacheKey2(req, instructions ?? "", input[0]);
@@ -25383,7 +25419,7 @@ function requiredCapabilities(request) {
   return {
     tools: (request.tools?.length ?? 0) > 0,
     images,
-    reasoning: request.reasoning !== undefined && request.reasoning.mode !== "off"
+    reasoning: request.reasoning !== undefined && request.reasoning.mode !== "off" && request.reasoning.mode !== "betweenTools"
   };
 }
 function cooldownMs(failures, threshold, base) {
@@ -46949,10 +46985,11 @@ var schema = object({
   thinking: union([
     object({
       type: literal("adaptive"),
-      display: _enum(["summarized", "omitted"]).optional()
+      display: _enum(["summarized", "omitted", "updates"]).optional()
     }),
     object({ type: literal("enabled"), budget_tokens: number2().int().positive() }),
-    object({ type: literal("disabled") })
+    object({ type: literal("disabled") }),
+    object({ type: literal("between_tools") })
   ]).optional(),
   metadata: object({ user_id: string2().max(512).nullish() }).nullish()
 });
@@ -47178,6 +47215,8 @@ function toIrReasoning(thinking, effort) {
       return { mode: "budget", budgetTokens: thinking.budget_tokens };
     case "disabled":
       return { mode: "off" };
+    case "between_tools":
+      return { mode: "betweenTools" };
   }
 }
 function readBetas(headers) {
@@ -47237,8 +47276,13 @@ function parseAnthropicRequest(body, headers) {
   const conversation = typeof userId === "string" && userId.length > 0 ? userId : readConversationHeader(headers ?? undefined);
   if (conversation !== undefined)
     request.conversationId = conversation;
-  const extras = extraFields(body, KNOWN);
-  if (extras !== undefined)
+  const rawThinking = body.thinking;
+  const thinkingExtras = parsed.thinking !== undefined && typeof rawThinking === "object" && rawThinking !== null ? extraFields(rawThinking, ["type", "display", "budget_tokens"]) : undefined;
+  const extras = {
+    ...extraFields(body, KNOWN),
+    ...thinkingExtras === undefined ? {} : { thinking: thinkingExtras }
+  };
+  if (Object.keys(extras).length > 0)
     request.vendor = { anthropic: extras };
   if (named.betas.length > 0)
     request.betas = named.betas;
@@ -49035,7 +49079,7 @@ function createRing(limits = DEFAULT_RING_LIMITS) {
 }
 
 // apps/gateway/src/version.ts
-var VERSION = "0.13.3";
+var VERSION = "0.13.4";
 
 // apps/gateway/src/app.ts
 var ADMIN_SESSION_TTL_MS = 12 * 60 * 60 * 1000;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnigateway",
-  "version": "0.13.3",
+  "version": "0.13.4",
   "description": "Self-hosted AI gateway with Anthropic- and OpenAI-compatible APIs, an admin console, and a CLI",
   "license": "MIT",
   "author": "Harismawan <mail@harismawan.com>",