n9router 0.4.43 → 0.4.45

package/app/open-sse/translator/concerns/modality.js

@@ -0,0 +1,155 @@
+// Strip multimodal content blocks a model cannot read, BEFORE translation.
+// Driven by getCapabilitiesForModel: vision/audioInput/pdf. Replaces removed
+// media with a short text placeholder so messages never become empty.
+import { FORMATS } from "../formats.js";
+
+// Placeholder text inserted where a media block was removed.
+// Current turn: explain the active model can't read what the user just sent.
+const PLACEHOLDER_CURRENT = {
+  vision: "[image omitted: model has no vision support]",
+  audioInput: "[audio omitted: model has no audio support]",
+  pdf: "[file omitted: model has no document support]",
+};
+// Earlier turns: neutral (a combo may route to a different model each turn).
+const PLACEHOLDER_PREV = {
+  vision: "[Previous image omitted from context.]",
+  audioInput: "[Previous audio omitted from context.]",
+  pdf: "[Previous file omitted from context.]",
+};
+const ph = (cap, isLast) => (isLast ? PLACEHOLDER_CURRENT : PLACEHOLDER_PREV)[cap];
+
+// Map gemini inlineData/fileData mime prefix -> capability it requires.
+function capForMime(mime) {
+  if (typeof mime !== "string") return null;
+  if (mime.startsWith("image/")) return "vision";
+  if (mime.startsWith("audio/")) return "audioInput";
+  if (mime === "application/pdf") return "pdf";
+  return null;
+}
+
+// OpenAI chat content block -> required capability (null = plain text/other, keep).
+function capForOpenAIBlock(block) {
+  const t = block?.type;
+  if (t === "image_url" || t === "image") return "vision";
+  if (t === "input_audio" || t === "audio_url") return "audioInput";
+  if (t === "file") return "pdf";
+  return null;
+}
+
+// Claude content block -> required capability.
+function capForClaudeBlock(block) {
+  const t = block?.type;
+  if (t === "image") return "vision";
+  if (t === "document") return "pdf";
+  return null;
+}
+
+// Filter an array of content blocks; drop unsupported, inject one placeholder per kind.
+// isLast = block belongs to the current user turn (picks the explanatory placeholder).
+function filterBlocks(blocks, capOf, caps, removed, isLast) {
+  const out = [];
+  for (const block of blocks) {
+    const cap = capOf(block);
+    if (cap && caps[cap] === false) { removed.add(cap); continue; }
+    out.push(block);
+  }
+  for (const cap of removed) out.push({ type: "text", text: ph(cap, isLast) });
+  return out;
+}
+
+// OpenAI / OpenAI-compatible chat messages[].content[].
+function stripOpenAI(body, caps) {
+  if (!Array.isArray(body.messages)) return;
+  const last = body.messages.length - 1;
+  body.messages.forEach((msg, i) => {
+    if (!Array.isArray(msg.content)) return;
+    const removed = new Set();
+    msg.content = filterBlocks(msg.content, capForOpenAIBlock, caps, removed, i === last);
+  });
+}
+
+// Claude messages[].content[].
+function stripClaude(body, caps) {
+  if (!Array.isArray(body.messages)) return;
+  const last = body.messages.length - 1;
+  body.messages.forEach((msg, i) => {
+    if (!Array.isArray(msg.content)) return;
+    const removed = new Set();
+    msg.content = filterBlocks(msg.content, capForClaudeBlock, caps, removed, i === last);
+  });
+}
+
+// OpenAI Responses input[].content[] (input_image / input_file).
+function stripResponses(body, caps) {
+  if (!Array.isArray(body.input)) return;
+  const last = body.input.length - 1;
+  body.input.forEach((item, i) => {
+    if (!Array.isArray(item.content)) return;
+    const removed = new Set();
+    item.content = item.content.filter((b) => {
+      const cap = b?.type === "input_image" ? "vision" : b?.type === "input_file" ? "pdf" : null;
+      if (cap && caps[cap] === false) { removed.add(cap); return false; }
+      return true;
+    });
+    for (const cap of removed) item.content.push({ type: "input_text", text: ph(cap, i === last) });
+  });
+}
+
+// Gemini / gemini-cli contents[].parts[] (inlineData / fileData by mime).
+function stripGeminiParts(contents, caps) {
+  if (!Array.isArray(contents)) return;
+  const last = contents.length - 1;
+  contents.forEach((c, i) => {
+    if (!Array.isArray(c.parts)) return;
+    const removed = new Set();
+    c.parts = c.parts.filter((p) => {
+      const mime = p?.inlineData?.mimeType || p?.fileData?.mimeType;
+      const cap = capForMime(mime);
+      if (cap && caps[cap] === false) { removed.add(cap); return false; }
+      return true;
+    });
+    for (const cap of removed) c.parts.push({ text: ph(cap, i === last) });
+  });
+}
+
+/**
+ * Remove media blocks the model can't read, in-place on the source-format body.
+ * @param {object} body - request body (source format)
+ * @param {string} sourceFormat - one of FORMATS
+ * @param {object} caps - capabilities from getCapabilitiesForModel
+ * @returns {boolean} true if anything was stripped-eligible (cap false for some modality)
+ */
+export function stripUnsupportedModalities(body, sourceFormat, caps) {
+  if (!body || !caps) return false;
+  // Fast exit: model supports everything we'd strip.
+  if (caps.vision !== false && caps.audioInput !== false && caps.pdf !== false) return false;
+
+  switch (sourceFormat) {
+    case FORMATS.OPENAI:
+    case FORMATS.OLLAMA:
+    case FORMATS.KIRO:
+    case FORMATS.CURSOR:
+    case FORMATS.COMMANDCODE:
+      stripOpenAI(body, caps);
+      break;
+    case FORMATS.CLAUDE:
+      stripClaude(body, caps);
+      break;
+    case FORMATS.OPENAI_RESPONSES:
+    case FORMATS.OPENAI_RESPONSE:
+    case FORMATS.CODEX:
+      stripResponses(body, caps);
+      break;
+    case FORMATS.GEMINI:
+    case FORMATS.GEMINI_CLI:
+    case FORMATS.VERTEX:
+      stripGeminiParts(body.contents, caps);
+      break;
+    case FORMATS.ANTIGRAVITY:
+      stripGeminiParts(body?.request?.contents, caps);
+      break;
+    default:
+      stripOpenAI(body, caps);
+  }
+  return true;
+}

package/app/open-sse/translator/concerns/paramSupport.js

@@ -0,0 +1,44 @@
+// Strip request params a given provider/model rejects upstream (e.g. HTTP 400).
+// Config-driven: add a rule instead of scattering `delete body.x` across executors.
+
+// Each rule: optional provider, regex match on model, list of params to drop.
+// A param is removed only when it is present (!== undefined).
+const STRIP_RULES = [
+  // claude-opus-4 series: temperature is deprecated (Anthropic 400). #1748
+  { match: /claude-opus-4/i, drop: ["temperature"] },
+  // GitHub Copilot gpt-5.4: temperature unsupported.
+  { provider: "github", match: /gpt-5\.4/i, drop: ["temperature"] },
+  // GitHub Copilot Claude (except opus/sonnet 4.6): thinking + reasoning_effort rejected. #713
+  { provider: "github", match: (m) => /claude/i.test(m) && !/claude.*(opus|sonnet).*4\.6/i.test(m), drop: ["thinking", "reasoning_effort"] },
+  // Cloudflare Workers AI: content must be plain string, rejects OpenAI content-part array (#1926)
+  { provider: "cloudflare-ai", flattenContent: true },
+];
+
+// Test a rule's match (regex or predicate) against the model id.
+function matches(rule, model) {
+  if (!rule.match) return true;
+  return typeof rule.match === "function" ? rule.match(model) : rule.match.test(model);
+}
+
+// Remove unsupported params from body in place; returns body.
+export function stripUnsupportedParams(provider, model, body) {
+  if (!model || !body || typeof body !== "object") return body;
+  for (const rule of STRIP_RULES) {
+    if (rule.provider && rule.provider !== provider) continue;
+    if (!matches(rule, model)) continue;
+    for (const key of rule.drop || []) {
+      if (body[key] !== undefined) delete body[key];
+    }
+    // CF Workers AI oneOf root schema only accepts content as plain string (#1926)
+    if (rule.flattenContent && Array.isArray(body.messages)) {
+      for (const msg of body.messages) {
+        if (msg && Array.isArray(msg.content)) {
+          msg.content = msg.content
+            .map(b => (b?.type === "text" && typeof b.text === "string") ? b.text : "")
+            .join("");
+        }
+      }
+    }
+  }
+  return body;
+}

package/app/open-sse/translator/concerns/prefetch.js

@@ -0,0 +1,96 @@
+// Pre-fetch remote image URLs into base64 BEFORE translation, for target
+// formats whose upstream providers cannot fetch remote URLs themselves
+// (they require inline base64). Runs on the source-format body.
+import { FORMATS } from "../formats.js";
+import { fetchImageAsBase64, parseDataUri } from "./image.js";
+
+// Targets that require inline base64 images (cannot accept remote URLs).
+const TARGETS_NEED_BASE64 = new Set([
+  FORMATS.GEMINI, FORMATS.GEMINI_CLI, FORMATS.VERTEX,
+  FORMATS.ANTIGRAVITY, FORMATS.OLLAMA, FORMATS.KIRO,
+]);
+
+function isRemoteUrl(url) {
+  return typeof url === "string" && (url.startsWith("http://") || url.startsWith("https://"));
+}
+
+// Collect {get,set} accessors for every remote image URL in a source body.
+function collectImageRefs(body, sourceFormat) {
+  const refs = [];
+  const pushOpenAI = (messages) => {
+    for (const msg of messages || []) {
+      if (!Array.isArray(msg.content)) continue;
+      for (const block of msg.content) {
+        if (block?.type === "image_url") {
+          const url = typeof block.image_url === "string" ? block.image_url : block.image_url?.url;
+          if (isRemoteUrl(url)) refs.push({ get: () => url, set: (v) => {
+            if (typeof block.image_url === "string") block.image_url = v; else block.image_url.url = v;
+          } });
+        }
+      }
+    }
+  };
+  const pushGemini = (contents) => {
+    for (const c of contents || []) {
+      for (const p of c.parts || []) {
+        const uri = p?.fileData?.fileUri;
+        if (isRemoteUrl(uri)) refs.push({ get: () => uri, part: p });
+      }
+    }
+  };
+
+  switch (sourceFormat) {
+    case FORMATS.OPENAI:
+    case FORMATS.OLLAMA:
+    case FORMATS.KIRO:
+    case FORMATS.CURSOR:
+    case FORMATS.COMMANDCODE:
+      pushOpenAI(body.messages);
+      break;
+    case FORMATS.CLAUDE:
+      for (const msg of body.messages || []) {
+        if (!Array.isArray(msg.content)) continue;
+        for (const block of msg.content) {
+          if (block?.type === "image" && block.source?.type === "url" && isRemoteUrl(block.source.url)) {
+            refs.push({ get: () => block.source.url, claudeBlock: block });
+          }
+        }
+      }
+      break;
+    case FORMATS.GEMINI:
+    case FORMATS.GEMINI_CLI:
+    case FORMATS.VERTEX:
+      pushGemini(body.contents);
+      break;
+    case FORMATS.ANTIGRAVITY:
+      pushGemini(body?.request?.contents);
+      break;
+    default:
+      pushOpenAI(body.messages);
+  }
+  return refs;
+}
+
+/**
+ * Replace remote image URLs with base64 data when the target needs inline data.
+ * No-op when target accepts remote URLs (e.g. openai, claude) or body has none.
+ * @returns {Promise<number>} count of images converted
+ */
+export async function prefetchRemoteImages(body, sourceFormat, targetFormat, options = {}) {
+  if (!body || !TARGETS_NEED_BASE64.has(targetFormat)) return 0;
+  const refs = collectImageRefs(body, sourceFormat);
+  if (!refs.length) return 0;
+
+  let converted = 0;
+  for (const ref of refs) {
+    const url = ref.get();
+    if (parseDataUri(url)) continue; // already inline
+    const fetched = await fetchImageAsBase64(url, options);
+    if (!fetched) continue;
+    if (ref.set) ref.set(fetched.url);
+    else if (ref.part) { delete ref.part.fileData; ref.part.inlineData = { mimeType: fetched.mimeType, data: fetched.url.split(",")[1] }; }
+    else if (ref.claudeBlock) ref.claudeBlock.source = { type: "base64", media_type: fetched.mimeType, data: fetched.url.split(",")[1] };
+    converted++;
+  }
+  return converted;
+}

package/app/open-sse/translator/concerns/reasoning.js

@@ -0,0 +1,24 @@
+import { ROLE } from "../schema/index.js";
+
+// Build OpenAI delta carrying reasoning_content (optional leading assistant role)
+export function reasoningDelta(text, withRole = false) {
+  return withRole
+    ? { role: ROLE.ASSISTANT, reasoning_content: text }
+    : { reasoning_content: text };
+}
+
+// Extract reasoning text from a streamed OpenAI-compatible delta across vendor shapes:
+//   - reasoning_content (GLM, Qwen, DeepSeek, Kimi, Step, Hunyuan)
+//   - reasoning (some compat layers)
+//   - reasoning_details[] (MiniMax reasoning_split=true): [{ text|content }]
+// Returns concatenated reasoning string, or "" when none.
+export function extractReasoningText(delta) {
+  if (!delta || typeof delta !== "object") return "";
+  if (typeof delta.reasoning_content === "string" && delta.reasoning_content) return delta.reasoning_content;
+  if (typeof delta.reasoning === "string" && delta.reasoning) return delta.reasoning;
+  const details = delta.reasoning_details;
+  if (Array.isArray(details)) {
+    return details.map((d) => (typeof d === "string" ? d : d?.text || d?.content || "")).join("");
+  }
+  return "";
+}

package/app/open-sse/translator/concerns/thinking.js

@@ -0,0 +1,53 @@
+// Concern: reasoning_effort ↔ provider-native thinking config.
+// Central source of truth for level↔budget maps (web-standard values).
+// Provider-specific application lives in thinkingUnified.js; this file is maps-only.
+
+// Discrete effort levels, ordered low→high.
+export const EFFORT_LEVELS = ["minimal", "low", "medium", "high", "xhigh", "max"];
+
+// Web-standard level → budget_tokens (Anthropic/Gemini docs).
+export const LEVEL_TO_BUDGET = {
+  none: 0,
+  minimal: 512,
+  low: 1024,
+  medium: 8192,
+  high: 24576,
+  xhigh: 32768,
+  max: 128000,
+};
+
+// Returns budget_tokens for an effort level, or undefined if unknown.
+// 0 means "no thinking"; undefined means "effort not recognized".
+export function effortToBudget(effort) {
+  if (!effort) return undefined;
+  return LEVEL_TO_BUDGET[String(effort).toLowerCase()];
+}
+
+// OpenAI reasoning_effort → Gemini thinkingLevel (gemini-3 enum: minimal|low|medium|high).
+// Gemini 3 cannot fully disable thinking; "none"/"off" map to "minimal".
+export function effortToThinkingLevel(effort) {
+  const e = String(effort).toLowerCase().trim();
+  if (e === "none" || e === "off") return "minimal";
+  if (e === "xhigh" || e === "max") return "high";
+  return e;
+}
+
+// Numeric budget → nearest discrete level (reverse map via thresholds).
+// Returns null when budget <= 0 (no reasoning).
+export function budgetToLevel(budget) {
+  const b = Number(budget);
+  if (!b || b <= 0) return null;
+  if (b <= 768) return "minimal";
+  if (b <= 4096) return "low";
+  if (b <= 16384) return "medium";
+  if (b <= 28672) return "high";
+  return "xhigh";
+}
+
+// Gemini thinkingBudget (numeric) → OpenAI reasoning_effort (antigravity reverse map).
+export function budgetToEffort(budget) {
+  if (!budget || budget <= 0) return null;
+  if (budget <= 2048) return "low";
+  if (budget <= 16384) return "medium";
+  return "high";
+}

package/app/open-sse/translator/concerns/thinkingUnified.js

@@ -0,0 +1,271 @@
+// Unified thinking normalization: extract client intent → apply provider-native format.
+// Config-driven: thinking format/limits come from capabilities.js + registry transport,
+// never hardcoded per-model here. See .docs/thinking/plan.md MATRIX VI-A.
+
+import { getCapabilitiesForModel } from "../../providers/capabilities.js";
+import { PROVIDERS } from "../../providers/index.js";
+import { LEVEL_TO_BUDGET, budgetToLevel, effortToBudget, effortToThinkingLevel } from "./thinking.js";
+
+// Map a target wire-format to its native thinking format (when capability has none).
+const FORMAT_TO_NATIVE = {
+  openai: "openai",
+  "openai-responses": "openai",
+  "openai-response": "openai",
+  codex: "openai",
+  claude: "claude-budget",
+  gemini: "gemini-budget",
+  "gemini-cli": "gemini-budget",
+  vertex: "gemini-budget",
+  antigravity: "gemini-budget",
+  kiro: "kiro",
+};
+
+// Parse model-name suffix "model(value)" → { cleanModel, override }.
+// value: level name (high) | number (8192) | auto | none. null override when absent.
+export function parseSuffix(model) {
+  if (typeof model !== "string") return { cleanModel: model, override: null };
+  const m = model.match(/^(.*)\(([^()]+)\)\s*$/);
+  if (!m) return { cleanModel: model, override: null };
+  const cleanModel = m[1].trim();
+  const raw = m[2].trim().toLowerCase();
+  if (raw === "none" || raw === "off") return { cleanModel, override: { mode: "none" } };
+  if (raw === "auto") return { cleanModel, override: { mode: "auto" } };
+  if (/^\d+$/.test(raw)) return { cleanModel, override: { mode: "budget", budget: Number(raw) } };
+  if (LEVEL_TO_BUDGET[raw] !== undefined) return { cleanModel, override: { mode: "level", level: raw } };
+  return { cleanModel, override: null };
+}
+
+// Extract unified thinking intent from a request body (post-translation, mixed shapes).
+// Returns { mode, budget?, level? } or null when no thinking intent present.
+export function extractThinking(body) {
+  if (!body || typeof body !== "object") return null;
+
+  // Claude output_config.effort (explicit) — priority over adaptive thinking
+  const oc = body.output_config?.effort;
+  if (typeof oc === "string" && oc) {
+    const e = oc.toLowerCase();
+    if (e === "none" || e === "off") return { mode: "none" };
+    if (e === "auto") return { mode: "auto" };
+    return { mode: "level", level: e };
+  }
+
+  // Claude shape
+  const t = body.thinking;
+  if (t && typeof t === "object") {
+    if (t.type === "disabled") return { mode: "none" };
+    if (t.type === "adaptive" || t.type === "enabled") {
+      const budget = Number(t.budget_tokens);
+      if (Number.isFinite(budget) && budget > 0) return { mode: "budget", budget };
+      return { mode: "auto" };
+    }
+  }
+
+  // OpenAI chat / Responses shape
+  const effort = body.reasoning_effort ?? (typeof body.reasoning === "object" ? body.reasoning?.effort : null);
+  if (typeof effort === "string" && effort) {
+    const e = effort.toLowerCase();
+    if (e === "none" || e === "off") return { mode: "none" };
+    if (e === "auto") return { mode: "auto" };
+    return { mode: "level", level: e };
+  }
+
+  // Gemini shape (top-level, generationConfig, or request envelope)
+  const tc = body.thinkingConfig || body.generationConfig?.thinkingConfig || body.request?.generationConfig?.thinkingConfig;
+  if (tc && typeof tc === "object") {
+    if (typeof tc.thinkingLevel === "string") return { mode: "level", level: tc.thinkingLevel.toLowerCase() };
+    const tb = Number(tc.thinkingBudget);
+    if (Number.isFinite(tb)) {
+      if (tb === 0) return { mode: "none" };
+      if (tb < 0) return { mode: "auto" };
+      return { mode: "budget", budget: tb };
+    }
+  }
+
+  // Qwen shape
+  if (body.enable_thinking === false) return { mode: "none" };
+  if (body.enable_thinking === true) {
+    const tb = Number(body.thinking_budget);
+    if (Number.isFinite(tb) && tb > 0) return { mode: "budget", budget: tb };
+    return { mode: "auto" };
+  }
+
+  return null;
+}
+
+// Capture thinking intent from a body. Alias of extractThinking, named for clarity
+// at the call-site where intent is snapshotted before format translation.
+export const captureThinking = extractThinking;
+
+// Resolve thinking format: provider override > capability > derive(targetFormat).
+function resolveFormat(targetFormat, model, provider) {
+  const providerFmt = provider ? PROVIDERS[provider]?.thinkingFormat : null;
+  if (providerFmt) return providerFmt;
+  const caps = getCapabilitiesForModel(provider, model);
+  if (caps.thinkingFormat) return caps.thinkingFormat;
+  return FORMAT_TO_NATIVE[targetFormat] || "openai";
+}
+
+// Convert unified config to a budget number (for budget-based formats).
+function toBudget(cfg, range) {
+  let budget;
+  if (cfg.mode === "budget") budget = cfg.budget;
+  else if (cfg.mode === "level") budget = effortToBudget(cfg.level);
+  else if (cfg.mode === "auto") return -1;
+  if (!Number.isFinite(budget)) return undefined;
+  if (range) {
+    if (range.min != null && budget < range.min) budget = range.min;
+    if (range.max != null && budget > range.max) budget = range.max;
+  }
+  return budget;
+}
+
+// Convert unified config to a discrete level string.
+function toLevel(cfg) {
+  if (cfg.mode === "level") return cfg.level;
+  if (cfg.mode === "budget") return budgetToLevel(cfg.budget) || "medium";
+  if (cfg.mode === "auto") return "auto";
+  return null;
+}
+
+function toGeminiThinkingLevel(cfg) {
+  const raw = cfg.mode === "auto" ? "high" : (toLevel(cfg) || "high");
+  return effortToThinkingLevel(raw);
+}
+
+// Gemini nests thinkingConfig under generationConfig. gemini-cli / antigravity wrap
+// the whole request in a { request: { generationConfig } } envelope — target the
+// envelope's generationConfig when present, else the top-level one.
+function setGeminiThinking(body, tc) {
+  const gc = body.request?.generationConfig
+    ? body.request.generationConfig
+    : (body.generationConfig && typeof body.generationConfig === "object"
+        ? body.generationConfig
+        : (body.generationConfig = {}));
+  gc.thinkingConfig = tc;
+}
+
+// Strip every known thinking field from a body (used before re-applying / when unsupported).
+function stripAll(body) {
+  delete body.thinking;
+  delete body.reasoning_effort;
+  delete body.reasoning;
+  delete body.thinkingConfig;
+  delete body.enable_thinking;
+  delete body.thinking_budget;
+  delete body.output_config;
+  if (body.generationConfig) delete body.generationConfig.thinkingConfig;
+  if (body.request?.generationConfig) delete body.request.generationConfig.thinkingConfig;
+}
+
+// Apply unified thinking config to body in the resolved provider-native format.
+function applyFormat(fmt, body, cfg, caps) {
+  const none = cfg.mode === "none";
+  const canDisable = caps.thinkingCanDisable !== false;
+  // Model cannot disable thinking → clamp "none" to minimal effort instead.
+  const eff = none && !canDisable ? { mode: "level", level: "minimal" } : cfg;
+
+  switch (fmt) {
+    case "openai": {
+      if (none && canDisable) { body.reasoning_effort = "none"; break; }
+      const level = toLevel(eff);
+      if (level) body.reasoning_effort = level;
+      break;
+    }
+    case "claude-adaptive": {
+      if (none && canDisable) { body.thinking = { type: "disabled" }; break; }
+      const level = toLevel(eff);
+      body.output_config = { effort: level === "xhigh" ? "high" : level };
+      break;
+    }
+    case "claude-budget": {
+      if (none && canDisable) { body.thinking = { type: "disabled" }; break; }
+      const budget = toBudget(eff, caps.thinkingRange);
+      body.thinking = budget === -1 ? { type: "enabled" } : { type: "enabled", budget_tokens: budget || 8192 };
+      break;
+    }
+    case "gemini-level": {
+      const level = none ? "minimal" : toGeminiThinkingLevel(eff);
+      setGeminiThinking(body, { thinkingLevel: level, includeThoughts: level !== "minimal" });
+      break;
+    }
+    case "gemini-budget": {
+      if (none && canDisable) { setGeminiThinking(body, { thinkingBudget: 0, includeThoughts: false }); break; }
+      const budget = toBudget(eff, caps.thinkingRange);
+      setGeminiThinking(body, { thinkingBudget: budget ?? -1, includeThoughts: true });
+      break;
+    }
+    case "zai": {
+      // Z.ai ignores thinking.disabled → must use enable_thinking:false to turn off.
+      if (none && canDisable) { body.enable_thinking = false; delete body.thinking; break; }
+      body.thinking = { type: "enabled" };
+      break;
+    }
+    case "qwen": {
+      if (none && canDisable) { body.enable_thinking = false; break; }
+      body.enable_thinking = true;
+      const budget = toBudget(eff, caps.thinkingRange);
+      if (Number.isFinite(budget) && budget > 0) body.thinking_budget = budget;
+      break;
+    }
+    case "deepseek": {
+      if (none && canDisable) { body.thinking = { type: "disabled" }; break; }
+      body.thinking = { type: "enabled" };
+      // DeepSeek: low/medium→high, xhigh/max→max.
+      const level = toLevel(eff);
+      body.reasoning_effort = level === "xhigh" || level === "max" ? "max" : "high";
+      break;
+    }
+    case "kimi": {
+      if (none && canDisable) { body.thinking = { type: "disabled" }; break; }
+      const level = toLevel(eff);
+      if (level) body.reasoning_effort = level === "max" ? "high" : level;
+      break;
+    }
+    case "minimax": {
+      // M3 adaptive; M2.x cannot disable (handled via canDisable clamp).
+      body.thinking = { type: none && canDisable ? "disabled" : "adaptive" };
+      break;
+    }
+    case "hunyuan": {
+      if (none && canDisable) { body.thinking = { type: "disabled" }; break; }
+      const budget = toBudget(eff, caps.thinkingRange);
+      body.thinking = budget === -1 ? { type: "enabled" } : { type: "enabled", budget_tokens: budget || 8192 };
+      break;
+    }
+    case "step": {
+      if (none && canDisable) break;
+      const level = toLevel(eff);
+      if (level) body.reasoning_effort = level === "xhigh" || level === "max" ? "high" : level;
+      break;
+    }
+    case "kiro":
+      // Kiro thinking handled via system-tag injection in openai-to-kiro.js; no body field here.
+      break;
+    default:
+      break;
+  }
+}
+
+// Public entry: normalize thinking for the resolved target format.
+// Mutates and returns body. No-op when model has no reasoning capability.
+// `intent` is a pre-captured config (from captureThinking on the original body);
+// falls back to extracting from the current body when omitted.
+export function applyThinking(targetFormat, model, body, provider = null, intent = undefined) {
+  if (!body || typeof body !== "object") return body;
+
+  const { cleanModel, override } = parseSuffix(model);
+  const cfg = override || intent || extractThinking(body);
+  const caps = getCapabilitiesForModel(provider, cleanModel);
+
+  // Model cannot reason → strip any stray thinking fields.
+  if (!caps.reasoning) {
+    stripAll(body);
+    return body;
+  }
+  if (!cfg) return body;
+
+  const fmt = resolveFormat(targetFormat, cleanModel, provider);
+  stripAll(body);
+  applyFormat(fmt, body, cfg, caps);
+  return body;
+}

package/app/open-sse/translator/{helpers/toolCallHelper.js → concerns/toolCall.js}

@@ -3,6 +3,11 @@
 // Anthropic tool_use.id must match: ^[a-zA-Z0-9_-]+$
 const TOOL_ID_PATTERN = /^[a-zA-Z0-9_-]+$/;
 
+// Fallback streaming tool_call id when provider omits one (index optional)
+export function fallbackToolCallId(index) {
+  return index === undefined ? `call_${Date.now()}` : `call_${index}_${Date.now()}`;
+}
+
 // Generate deterministic tool call ID from position + tool name (cache-friendly)
 export function generateToolCallId(msgIndex = 0, tcIndex = 0, toolName = "") {
   const name = toolName ? `_${toolName.replace(/[^a-zA-Z0-9_-]/g, "")}` : "";