fix(kiro): honor thinking effort budgets

Resolve the Kiro thinking budget from client intent (OpenAI reasoning_effort,
OpenAI Responses reasoning.effort, Claude output_config.effort, and Claude
thinking.budget_tokens) by reusing the shared thinkingUnified extractThinking
parser, then inject the resolved budget into the Kiro thinking system prefix.
Explicit none/off/disabled stops the prefix injection; synthetic -thinking
aliases keep the default budget.

Co-authored-by: Cursor <cursoragent@cursor.com>

open-sse/config/kiroConstants.js

@@ -15,6 +15,9 @@
  * fiction. The suffix is stripped before the request leaves this process.
  */
 
+import { extractThinking } from "../translator/concerns/thinkingUnified.js";
+import { effortToBudget } from "../translator/concerns/thinking.js";
+
 export const KIRO_AGENTIC_SUFFIX = "-agentic";
 export const KIRO_THINKING_SUFFIX = "-thinking";
 
@@ -89,16 +92,48 @@ REMEMBER: When in doubt, write LESS per operation. Multiple small operations > o
 `.trim();
 
 /**
- * Detect whether an inbound request is asking for reasoning / thinking output.
+ * Resolve the Kiro thinking budget requested by a client.
  *
- * Sources of intent (any one is enough):
- *   - HTTP header `Anthropic-Beta: ...interleaved-thinking...`
- *   - JSON `thinking.type === "enabled"` (Claude Messages API)
- *   - JSON `reasoning_effort` in {low, medium, high, auto} (OpenAI o1/o3)
- *   - JSON `reasoning.effort` in {low, medium, high, auto} (OpenAI Responses)
- *   - System prompt contains `<thinking_mode>enabled</thinking_mode>` or
- *     `<thinking_mode>interleaved</thinking_mode>` (AMP / Cursor)
- *   - Model name contains `thinking` or `-reason`
+ * Reuses the shared thinkingUnified parser (extractThinking) so every client
+ * shape (Claude output_config.effort / thinking.budget_tokens, OpenAI
+ * reasoning_effort / reasoning.effort, Gemini, Qwen) maps consistently. Explicit
+ * `none`/`off`/disabled wins and returns null (no prefix injected).
+ * buildThinkingSystemPrefix performs Kiro's final 1..32000 clamp.
+ *
+ * @param {object} body OpenAI/Claude-shaped request body
+ * @param {object} [headers] Original inbound HTTP headers (case-insensitive)
+ * @param {string} [model] Model id the caller asked for
+ * @returns {number|null} budget to inject, or null when thinking is disabled
+ */
+export function resolveKiroThinkingBudget(body, headers, model) {
+  const cfg = extractThinking(body);
+  if (cfg) {
+    if (cfg.mode === "none") return null;
+    if (cfg.mode === "budget") return cfg.budget;
+    if (cfg.mode === "level") return effortToBudget(cfg.level) ?? KIRO_THINKING_BUDGET_DEFAULT;
+    return KIRO_THINKING_BUDGET_DEFAULT;
+  }
+
+  if (headers) {
+    const beta = pickHeader(headers, "anthropic-beta");
+    if (typeof beta === "string" && beta.toLowerCase().includes("interleaved-thinking")) {
+      return KIRO_THINKING_BUDGET_DEFAULT;
+    }
+  }
+
+  if (containsThinkingModeTag(body)) return KIRO_THINKING_BUDGET_DEFAULT;
+
+  if (typeof model === "string" && model) {
+    const m = model.toLowerCase();
+    if (m.includes("thinking") || m.includes("-reason")) return KIRO_THINKING_BUDGET_DEFAULT;
+  }
+
+  return null;
+}
+
+/**
+ * Detect whether an inbound request is asking for reasoning / thinking output.
+ * Thin wrapper over resolveKiroThinkingBudget (single source of truth).
  *
  * @param {object} body OpenAI-shaped request body (post-translation)
  * @param {object} [headers] Original inbound HTTP headers (case-insensitive)
@@ -106,44 +141,7 @@ REMEMBER: When in doubt, write LESS per operation. Multiple small operations > o
  * @returns {boolean}
  */
 export function isThinkingEnabled(body, headers, model) {
-  if (headers) {
-    const beta = pickHeader(headers, "anthropic-beta");
-    if (typeof beta === "string" && beta.toLowerCase().includes("interleaved-thinking")) {
-      return true;
-    }
-  }
-
-  if (body && typeof body === "object") {
-    const thinking = body.thinking;
-    if (thinking && typeof thinking === "object" && thinking.type === "enabled") {
-      const budget = Number(thinking.budget_tokens);
-      if (!Number.isFinite(budget) || budget > 0) {
-        return true;
-      }
-    }
-
-    const effort = body.reasoning_effort
-      ?? (body.reasoning && typeof body.reasoning === "object" ? body.reasoning.effort : null);
-    if (typeof effort === "string") {
-      const v = effort.toLowerCase();
-      if (v && v !== "none" && (v === "low" || v === "medium" || v === "high" || v === "auto")) {
-        return true;
-      }
-    }
-
-    if (containsThinkingModeTag(body)) {
-      return true;
-    }
-  }
-
-  if (typeof model === "string" && model) {
-    const m = model.toLowerCase();
-    if (m.includes("thinking") || m.includes("-reason")) {
-      return true;
-    }
-  }
-
-  return false;
+  return resolveKiroThinkingBudget(body, headers, model) !== null;
 }
 
 /**

open-sse/translator/request/claude-to-kiro.js

@@ -27,7 +27,7 @@ import { FORMATS } from "../formats.js";
 import { v4 as uuidv4 } from "uuid";
 import {
   resolveKiroModel,
-  isThinkingEnabled,
+  resolveKiroThinkingBudget,
   buildThinkingSystemPrefix,
   KIRO_AGENTIC_SYSTEM_PROMPT,
   resolveDefaultProfileArn,
@@ -374,13 +374,8 @@ export function claudeToKiroRequest(model, body, stream, credentials) {
   const temperature = body.temperature;
   const topP = body.top_p;
 
-  const {
-    upstream: upstreamModel,
-    agentic,
-    thinking: modelImpliesThinking,
-  } = resolveKiroModel(model);
-  const thinkingEnabled =
-    modelImpliesThinking || isThinkingEnabled(body, null, model);
+  const { upstream: upstreamModel, agentic } = resolveKiroModel(model);
+  const thinkingBudget = resolveKiroThinkingBudget(body, credentials?.rawHeaders, model);
 
   // Guard 1: no client tools → flatten all tool interactions to text.
   if (!clientProvidedTools) {
@@ -420,7 +415,7 @@ export function claudeToKiroRequest(model, body, stream, credentials) {
   // Prefix order: thinking_mode tag, timestamp marker, then agentic prompt.
   const timestamp = new Date().toISOString();
   const prefixParts = [];
-  if (thinkingEnabled) prefixParts.push(buildThinkingSystemPrefix());
+  if (thinkingBudget !== null) prefixParts.push(buildThinkingSystemPrefix(thinkingBudget));
   prefixParts.push(`[Context: Current time is ${timestamp}]`);
   if (agentic) prefixParts.push(KIRO_AGENTIC_SYSTEM_PROMPT);
   finalContent = `${prefixParts.join("\n\n")}\n\n${finalContent}`;

open-sse/translator/request/openai-to-kiro.js

@@ -8,7 +8,7 @@ import { v4 as uuidv4 } from "uuid";
 import { resolveSessionId } from "../../utils/sessionManager.js";
 import {
   resolveKiroModel,
-  isThinkingEnabled,
+  resolveKiroThinkingBudget,
   buildThinkingSystemPrefix,
   KIRO_AGENTIC_SYSTEM_PROMPT,
   resolveDefaultProfileArn
@@ -519,8 +519,8 @@ export function openaiToKiroRequest(model, body, stream, credentials) {
   const temperature = body.temperature;
   const topP = body.top_p;
 
-  const { upstream: upstreamModel, agentic, thinking: modelImpliesThinking } = resolveKiroModel(model);
-  const thinkingEnabled = modelImpliesThinking || isThinkingEnabled(body, null, model);
+  const { upstream: upstreamModel, agentic } = resolveKiroModel(model);
+  const thinkingBudget = resolveKiroThinkingBudget(body, credentials?.rawHeaders, model);
 
   const { history, currentMessage } = convertMessages(messages, tools, upstreamModel);
 
@@ -543,8 +543,8 @@ export function openaiToKiroRequest(model, body, stream, credentials) {
   // Order: thinking_mode tag first (so Kiro sees it before any user text),
   // then context/timestamp marker, then optional agentic chunked-write prompt.
   const prefixParts = [];
-  if (thinkingEnabled) {
-    prefixParts.push(buildThinkingSystemPrefix());
+  if (thinkingBudget !== null) {
+    prefixParts.push(buildThinkingSystemPrefix(thinkingBudget));
   }
   prefixParts.push(`[Context: Current time is ${timestamp}]`);
   if (agentic) {