Merge origin/master (v0.5.75) into gitea/new_feature

Resolve conflicts: - package.json / cli/package.json: take 0.5.75 - .gitignore: union both sides (upstream 9router-*/temp files + local state dirs) - CHANGELOG.md: keep both blocks, v0.5.75 above v0.5.70 - nonStreamingHandler.js: merge imports (unwrapClineEnvelope + tokensForDetail/shouldPersistRequestDetail); drop dead appendRequestLog - providers/[id]/page.js: union useState blocks (compatible-model states + importingClineModels) Co-authored-by: CommandCodeBot <noreply@commandcode.ai>
2026-09-17 14:14:43 +07:00
parent 81a47f36c6 17c4cc7687
commit 28e26f4295
113 changed files with 6634 additions and 482 deletions
--- a/open-sse/config/appConstants.js
+++ b/open-sse/config/appConstants.js
@@ -7,6 +7,9 @@ import { createRequire } from "module";
 export const GEMINI_CLI_VERSION = PROVIDERS["gemini-cli"]?.cliVersion;
 export const GEMINI_CLI_API_CLIENT = PROVIDERS["gemini-cli"]?.apiClient;

+// === Codex CLI === derive từ registry codex.transport
+export const CODEX_CLI_VERSION = PROVIDERS["codex"]?.cliVersion;
+
 // Map Node arch to Gemini CLI arch string (x64/x86/arm64/...)
 function geminiCLIArch() {
  const a = arch();
--- a/open-sse/executors/antigravity.js
+++ b/open-sse/executors/antigravity.js
@@ -5,7 +5,7 @@ import { OAUTH_ENDPOINTS, ANTIGRAVITY_HEADERS, AG_DEFAULT_TOOLS, AG_TOOL_SUFFIX,
 import { HTTP_STATUS } from "../config/runtimeConfig.js";
 import { resolveSessionId, toNumericSessionId } from "../utils/sessionManager.js";
 import { proxyAwareFetch } from "../utils/proxyFetch.js";
-import { cleanJSONSchemaForAntigravity } from "../translator/formats/gemini.js";
+import { cleanJSONSchemaForAntigravity, normalizeGeminiContents } from "../translator/formats/gemini.js";
 import { DEFAULT_THINKING_AG_SIGNATURE } from "../config/defaultThinkingSignature.js";
 import { getGeminiThoughtSignatureSync } from "../services/thoughtSignatureStore.js";

@@ -193,7 +193,7 @@ export class AntigravityExecutor extends BaseExecutor {

    // ─── Standard (non-image) request ───
    // Fix contents for Claude models via Antigravity
-    const contents = body.request?.contents?.map(c => {
+    const rawContents = (body.request?.contents || []).map(c => {
      let role = c.role;
      // functionResponse must be role "user" for Claude models
      if (c.parts?.some(p => p.functionResponse)) {
@@ -226,15 +226,13 @@ export class AntigravityExecutor extends BaseExecutor {
        return p;
      });

-      const partsChanged = parts?.length !== c.parts?.length || modifiedParts?.some((p, idx) => p !== c.parts[idx]);
-      if (role !== c.role || partsChanged) {
-        return {
-          ...c, role,
-          parts: modifiedParts || parts,
-        };
-      }
-      return c;
+      return {
+        ...c,
+        role,
+        parts: modifiedParts || parts || [],
+      };
    });
+    const contents = normalizeGeminiContents(rawContents);

    // Sanitize tool schemas and function names before sending to Antigravity.
    let tools = body.request?.tools;
--- a/open-sse/executors/codex.js
+++ b/open-sse/executors/codex.js
@@ -12,6 +12,7 @@ import { getThinkingLevels } from "../providers/thinkingLevels.js";
 import { DEFAULT_RETRY_CONFIG, HTTP_STATUS, resolveRetryEntry } from "../config/runtimeConfig.js";
 import { dbg } from "../utils/debugLog.js";
 import { resolveSessionId } from "../utils/sessionManager.js";
+import { stripCodexUnsupportedPatterns } from "../utils/codexToolSchema.js";

 // SSE error patterns inside 200-OK bodies. Some retry same account first; capacity rotates accounts.
 const CODEX_SSE_RETRY_PATTERNS = ["server_is_overloaded", "service_unavailable_error"];
@@ -72,6 +73,9 @@ function stripStoredItemReferences(body) {
 function normalizeCodexTools(body) {
  if (!Array.isArray(body.tools)) return;
  const validNames = new Set();
+  // Codex's schema validator has no Unicode property escapes; a `pattern`
+  // carrying `\p{...}` 400s the whole request on every account (#3922).
+  const patternStats = { removed: 0 };
  body.tools = body.tools.filter((tool) => {
    if (!tool || typeof tool !== "object" || Array.isArray(tool)) return false;
    const type = typeof tool.type === "string" ? tool.type : "";
@@ -80,6 +84,9 @@ function normalizeCodexTools(body) {
        for (const st of tool.tools) {
          const n = typeof st?.name === "string" ? st.name.trim().slice(0, 128) : "";
          if (n) validNames.add(n);
+          if (st?.parameters && typeof st.parameters === "object") {
+            st.parameters = stripCodexUnsupportedPatterns(st.parameters, patternStats);
+          }
        }
      }
      return true;
@@ -101,10 +108,13 @@ function normalizeCodexTools(body) {
    tool.type = "function";
    tool.name = name.slice(0, 128);
    if (description) tool.description = description;
-    tool.parameters = parameters;
+    tool.parameters = stripCodexUnsupportedPatterns(parameters, patternStats);
    validNames.add(name);
    return true;
  });
+  if (patternStats.removed > 0) {
+    dbg("CODEX", `stripped ${patternStats.removed} unsupported tool schema pattern(s)`);
+  }
  // Drop tool_choice if it references an unknown function name
  if (body.tool_choice && typeof body.tool_choice === "object" && !Array.isArray(body.tool_choice)) {
    if (body.tool_choice.type === "function") {
--- a/open-sse/executors/index.js
+++ b/open-sse/executors/index.js
@@ -17,6 +17,7 @@ import { PerplexityWebExecutor } from "./perplexity-web.js";
 import { OllamaLocalExecutor } from "./ollama-local.js";
 import { CommandCodeExecutor } from "./commandcode.js";
 import { XiaomiTokenplanExecutor } from "./xiaomi-tokenplan.js";
+import { XiaomiMimoExecutor } from "./xiaomi-mimo.js";
 import { MimoFreeExecutor } from "./mimo-free.js";
 import { CodeBuddyExecutor } from "./codebuddy-cn.js";
 import { CodeBuddyIntlExecutor } from "./codebuddy-intl.js";
@@ -50,6 +51,7 @@ const executors = {
  "ollama-local": new OllamaLocalExecutor(),
  commandcode: new CommandCodeExecutor(),
  "xiaomi-tokenplan": new XiaomiTokenplanExecutor(),
+  "xiaomi-mimo": new XiaomiMimoExecutor(),
  "mimo-free": new MimoFreeExecutor(),
  mmf: new MimoFreeExecutor(), // Alias for mimo-free
  "codebuddy-cn": new CodeBuddyExecutor(),
@@ -93,6 +95,7 @@ export { PerplexityWebExecutor } from "./perplexity-web.js";
 export { OllamaLocalExecutor } from "./ollama-local.js";
 export { CommandCodeExecutor } from "./commandcode.js";
 export { XiaomiTokenplanExecutor } from "./xiaomi-tokenplan.js";
+export { XiaomiMimoExecutor } from "./xiaomi-mimo.js";
 export { MimoFreeExecutor } from "./mimo-free.js";
 export { CodeBuddyExecutor } from "./codebuddy-cn.js";
 export { CodeBuddyIntlExecutor } from "./codebuddy-intl.js";
--- a/open-sse/executors/kiro.js
+++ b/open-sse/executors/kiro.js
@@ -127,12 +127,18 @@ async function readResponsePrefix(response, signal, maxBytes, timeoutMs) {
  return decoder.decode(concatChunks(chunks, totalBytes));
 }

+// The instruction goes into the current user turn, never into a top-level
+// `systemPrompt`: kiro.dev answers any body carrying that field with
+// 400 REQUEST_BODY_INVALID, so writing it here turned every repair retry into
+// a hard failure.
 function appendRepairInstruction(body, kind) {
  const repaired = structuredClone(body || {});
  const instruction = REPAIR_INSTRUCTIONS[kind] || "Retry the previous incomplete Kiro response.";
-  repaired.systemPrompt = repaired.systemPrompt
-    ? `${repaired.systemPrompt}\n\n${instruction}`
-    : instruction;
+  const msg = repaired?.conversationState?.currentMessage?.userInputMessage;
+  if (msg) {
+    const content = typeof msg.content === "string" ? msg.content : "";
+    msg.content = content ? `${content}\n\n${instruction}` : instruction;
+  }
  return repaired;
 }

@@ -259,6 +265,19 @@ export class KiroExecutor extends BaseExecutor {
      }
    }

+    // CLIRO parity for the Amazon surfaces: the Kiro runtime accepts the
+    // SSO bearer header + agent-mode marker. Without these the deprecated
+    // path gateway answers REQUEST_BODY_INVALID for modern payloads.
+    if (credentials?.accessToken) {
+      headers["x-amz-sso-bearer"] = credentials.accessToken;
+    }
+    headers["x-amzn-kiro-agent-mode"] = "spec";
+    headers["x-amzn-codewhisperer-machine-id"] = "kiro-desktop";
+    const profileArn = credentials?.providerSpecificData?.profileArn;
+    if (profileArn) {
+      headers["x-amzn-codewhisperer-profile-arn"] = profileArn;
+    }
+
    return headers;
  }

@@ -285,9 +304,13 @@ export class KiroExecutor extends BaseExecutor {
    // 403 "bearer token invalid", so they must hit the CodeWhisperer
    // *.amazonaws.com surface, and in the region the token was minted in
    // (the baseUrls are hardcoded us-east-1).
-    const isCodeWhispererSurface =
-      authMethod === "api_key" || authMethod === "external_idp" || authMethod === "idc";
-    if (!isCodeWhispererSurface) return baseUrls;
+    // Kiro deprecated the legacy path-style GenerateAssistantResponse on
+    // runtime.*.kiro.dev (IDE 1.0.228+ moved to POST / + x-amz-target). The
+    // path gateway now answers valid modern payloads with 400
+    // REQUEST_BODY_INVALID, and 400 is terminal in BaseExecutor, so kiro.dev
+    // must never be the first surface for any auth method. Amazon surfaces
+    // reject foreign tokens with 401/403, which DO fall through, so trying
+    // q/codewhisperer first is safe for every auth method (CLIRO parity).

    const region = (credentials?.providerSpecificData?.region || "us-east-1").trim();
    const regionalize = (u) =>
@@ -297,20 +320,17 @@ export class KiroExecutor extends BaseExecutor {

    const amazon = baseUrls.filter((u) => u.includes("amazonaws.com")).map(regionalize);
    const others = baseUrls.filter((u) => !u.includes("amazonaws.com"));
-    if (authMethod === "api_key") {
-      const q = amazon.filter((u) => u.includes("://q."));
-      const remaining = amazon.filter((u) => !u.includes("://q."));
-      return q.length > 0
-        ? [...q, ...remaining, ...others]
-        : [...amazon, ...others];
-    }
-
-    return amazon.length > 0 ? [...amazon, ...others] : baseUrls;
+    const q = amazon.filter((u) => u.includes("://q."));
+    const remaining = amazon.filter((u) => !u.includes("://q."));
+    return q.length > 0
+      ? [...q, ...remaining, ...others]
+      : [...amazon, ...others];
  }

  buildUrl(model, stream, urlIndex = 0, credentials = null) {
    const baseUrls = this.getOrderedBaseUrls(credentials);
-    return baseUrls[urlIndex] || baseUrls[0] || this.config.baseUrl;
+    const url = baseUrls[urlIndex] || baseUrls[0] || this.config.baseUrl;
+    return url;
  }

  // Retry only endpoint/auth-surface failures. Payload-invalid HTTP 400 must be
--- a/open-sse/executors/qoder.js
+++ b/open-sse/executors/qoder.js
@@ -32,14 +32,16 @@ import { SSE_DONE } from "../utils/sseConstants.js";
 import { FETCH_CONNECT_TIMEOUT_MS } from "../config/runtimeConfig.js";
 import { resolveProviderTimeoutMs } from "../services/providerTimeout.js";
 import {
-  QODER_CHAT_URL_ENCODED,
-  QODER_CHAT_BASE_ALT,
  QODER_CHAT_SIG_PATH,
-  QODER_MODEL_MAP,
+  QODER_CONTEXT_TIER_ENV,
+  qoderInferenceBase,
 } from "../shared/qoder/constants.js";
 import { getQoderModelConfig, resolveQoderModels, isQoderPat, resolveQoderCredentials } from "../services/qoderModels.js";
 import { OPENAI_BLOCK, CLAUDE_BLOCK } from "../translator/schema/blocks.js";
 import { encodeDataUri } from "../translator/concerns/image.js";
+import { createQoderSseCoalescer } from "../shared/qoder/sse.js";
+import { rewriteQoderMessageAttachments } from "../shared/qoder/attachments.js";
+import { resolveQoderContextTier, applyQoderContextTier } from "../shared/qoder/contextTier.js";

 /**
 * Hoist role:"system" messages out of the messages array (Qoder rejects
@@ -71,15 +73,16 @@ function normalizeMessages(messages) {
 *
 * Text-only content is flattened to a plain string (Qoder's historical
 * shape). When images are present the content stays an array and image
- * blocks are kept as OpenAI-style `image_url` parts — verified against the
- * upstream: it accepts both http(s) URLs and inline base64 data: URIs
- * directly, no pre-upload to the /image/upload OSS flow required (that is
- * a qodercli client-side choice, not a protocol requirement). The legacy
+ * blocks are kept as OpenAI-style `image_url` parts. Native qodercli
+ * uploads inlined bytes to `/api/v2/image/upload` first and then sends
+ * the OSS URL — `buildQoderRequestBody` does that rewrite before this
+ * runs. Tiny leftover data URIs are still accepted. The legacy
 * top-level `image_urls` / `chat_context.imageUrls` slots stay null —
 * qodercli leaves them null too.
 *
 * Claude-style `{type:"image", source:{...}}` blocks are converted to
- * `image_url` so claude-format clients also round-trip.
+ * `image_url`. File/document blocks that survived rewrite become short
+ * stubs so 30MB PDFs never land in agent_chat_generation.
 */
 function normalizeContent(content) {
  if (typeof content === "string") return content;
@@ -89,10 +92,24 @@ function normalizeContent(content) {
  const blocks = [];
  const textParts = [];
  let hasImage = false;
+
+  const pushText = (text) => {
+    if (!text) return;
+    if (hasImage || blocks.length) blocks.push({ type: OPENAI_BLOCK.TEXT, text });
+    else textParts.push(text);
+  };
+
+  const imageUrlOf = (item) => {
+    if (typeof item.image_url === "string" && item.image_url) return item.image_url;
+    if (typeof item.image_url?.url === "string" && item.image_url.url) return item.image_url.url;
+    return null;
+  };
+
  for (const item of content) {
    if (!item || typeof item !== "object") continue;
-    if (item.type === OPENAI_BLOCK.IMAGE_URL && typeof item.image_url?.url === "string" && item.image_url.url) {
-      blocks.push({ type: OPENAI_BLOCK.IMAGE_URL, image_url: { url: item.image_url.url } });
+    const imageUrl = item.type === OPENAI_BLOCK.IMAGE_URL ? imageUrlOf(item) : null;
+    if (imageUrl) {
+      blocks.push({ type: OPENAI_BLOCK.IMAGE_URL, image_url: { url: imageUrl } });
      hasImage = true;
    } else if (item.type === CLAUDE_BLOCK.IMAGE && item.source) {
      // Claude base64/url image → OpenAI image_url equivalent.
@@ -104,13 +121,14 @@ function normalizeContent(content) {
        blocks.push({ type: OPENAI_BLOCK.IMAGE_URL, image_url: { url } });
        hasImage = true;
      }
+    } else if (item.type === OPENAI_BLOCK.FILE) {
+      const name = item.file?.filename || item.file?.name || "file";
+      pushText(`[file omitted: ${name} — Qoder reads documents via its file API, not inlined bytes]`);
+    } else if (item.type === CLAUDE_BLOCK.DOCUMENT) {
+      const name = item.title || "document";
+      pushText(`[file omitted: ${name} — Qoder reads documents via its file API, not inlined bytes]`);
    } else if (typeof item.text === "string" && item.text) {
-      if (hasImage || blocks.length) {
-        // Keep ordering faithful once images are in play.
-        blocks.push({ type: OPENAI_BLOCK.TEXT, text: item.text });
-      } else {
-        textParts.push(item.text);
-      }
+      pushText(item.text);
    }
  }

@@ -190,7 +208,7 @@ function truncate(s, n) {
 /**
 * Map the OpenAI-style request body into the exact shape Qoder expects.
 */
-async function buildQoderRequestBody({ model, body, credentials, log, proxyOptions, signal }) {
+async function buildQoderRequestBody({ model, body, credentials, log, proxyOptions, signal, uploadFn = null }) {
  const qoderKey = String(model || "").replace(/^qoder\//, "");
  
  // Fetch model config from dynamic API instead of relying on static QODER_MODEL_MAP.
@@ -209,7 +227,30 @@ async function buildQoderRequestBody({ model, body, credentials, log, proxyOptio
    modelConfig = { ...retried, key: qoderKey };
  }

-  const { messages, systemText } = normalizeMessages(body.messages || []);
+  const incoming = Array.isArray(body.messages)
+    ? body.messages.map((m) => {
+      if (!m || typeof m !== "object") return m;
+      return {
+        ...m,
+        content: Array.isArray(m.content)
+          ? m.content.map((b) => (b && typeof b === "object" ? { ...b } : b))
+          : m.content,
+      };
+    })
+    : [];
+  try {
+    await rewriteQoderMessageAttachments(incoming, {
+      credentials,
+      log,
+      proxyOptions,
+      signal,
+      uploadFn,
+    });
+  } catch (err) {
+    log?.warn?.("QODER", `attachment rewrite failed: ${err.message}`);
+  }
+
+  const { messages, systemText } = normalizeMessages(incoming);
  const tools = body.tools;
  const isReasoning = !!modelConfig.is_reasoning;
  const maxOutputTokens = Number(modelConfig.max_output_tokens) || 0;
@@ -228,7 +269,21 @@ async function buildQoderRequestBody({ model, body, credentials, log, proxyOptio
  const sessionId = stableHash("qoder-session", psd.userId, qoderKey);
  const recordId = stableChatRecordId(qoderKey, messages, tools, maxTokens);

-  return {
+  // Context-window tier (200K/400K/1M): the IDE picks one from model_config.context_config;
+  // qodercli-style requests default to the smallest. Escalate when the prompt no longer fits.
+  const tierChoice = resolveQoderContextTier(
+    modelConfig,
+    { system: systemText, messages, tools },
+    { preference: process.env[QODER_CONTEXT_TIER_ENV] },
+  );
+  if (tierChoice) {
+    log?.info?.(
+      "QODER",
+      `context tier ${tierChoice.tier.name} (${tierChoice.tier.tokenCount} tokens, ${tierChoice.reason}) for ~${tierChoice.estimatedTokens} prompt tokens`,
+    );
+  }
+
+  const built = {
    qoderKey,
    payload: {
      request_id: uuidv4(),
@@ -276,6 +331,8 @@ async function buildQoderRequestBody({ model, body, credentials, log, proxyOptio
    },
    modelConfig,
  };
+  if (tierChoice) applyQoderContextTier(built.payload, tierChoice.tier);
+  return built;
 }

 /**
@@ -339,6 +396,11 @@ async function peekFirstQoderFrame(reader, decoder) {
 * response.text() which hangs until the socket closes — so on terminal
 * events we cancel the upstream reader and close our stream immediately.
 *
+ * Usage: Qoder puts finish_reason on `delta` and sends token counts on a
+ * later `choices: []` frame. Downstream OpenAI/Claude clients only read
+ * usage from the finish chunk, so we coalesce those two frames (see
+ * createQoderSseCoalescer) before forwarding.
+ *
 * NEW: Peek first frame to detect billing blocks (code 112/10605/pricingUrl).
 * If detected, return 403 response so chatCore marks connection unavailable
 * and triggers combo fallback instead of leaking error text into chat.
@@ -365,6 +427,11 @@ async function wrapQoderSSE(response, model) {
  const upstreamDrained = peek.upstreamDone === true;
  const encoder = new TextEncoder();
  let doneEmitted = false;
+  const coalescer = createQoderSseCoalescer({ model, encoder, sseDone: SSE_DONE });
+
+  const syncDone = () => {
+    if (coalescer.doneEmitted) doneEmitted = true;
+  };

  // Process one already-extracted SSE line (no trailing newline).
  const processLine = (line, controller) => {
@@ -375,15 +442,17 @@ async function wrapQoderSSE(response, model) {

    const data = trimmed.slice(5).trimStart();
    if (data === "[DONE]") {
-      controller.enqueue(encoder.encode(SSE_DONE));
-      doneEmitted = true;
+      coalescer.flush(controller);
+      syncDone();
      return;
    }

    let envelope;
    try { envelope = JSON.parse(data); } catch { return; }
    const statusVal = typeof envelope.statusCodeValue === "number" ? envelope.statusCodeValue : 200;
-    const inner = typeof envelope.body === "string" ? envelope.body : "";
+    const inner = typeof envelope.body === "string"
+      ? envelope.body
+      : envelope.body != null ? JSON.stringify(envelope.body) : "";
    if (statusVal !== 200) {
      const msg = inner || `upstream status ${statusVal}`;
      const errChunk = JSON.stringify({
@@ -399,14 +468,8 @@ async function wrapQoderSSE(response, model) {
      return;
    }
    if (!inner) return;
-    if (inner === "[DONE]") {
-      controller.enqueue(encoder.encode(SSE_DONE));
-      doneEmitted = true;
-      return;
-    }
-    // Strip embedded newlines so the SSE frame stays a single event.
-    const sanitized = inner.replace(/\r?\n/g, "");
-    controller.enqueue(encoder.encode(`data: ${sanitized}\n\n`));
+    coalescer.handleInner(inner, controller);
+    syncDone();
  };

  const stream = new ReadableStream({
@@ -465,7 +528,7 @@ async function wrapQoderSSE(response, model) {
      } finally {
        if (!doneEmitted) {
          try {
-            controller.enqueue(encoder.encode(SSE_DONE));
+            coalescer.flush(controller);
            doneEmitted = true;
          } catch { /* already closed */ }
        }
@@ -494,13 +557,7 @@ export class QoderExecutor extends BaseExecutor {
  }

  buildUrl(credentials) {
-    // Job-token (jt-...) traffic must hit api2.qoder.sh — api3 rejects jt-
-    // with "Login expired" (403). Device tokens (dt-...) stay on api3.
-    const raw = credentials?.apiKey || credentials?.accessToken;
-    if (typeof raw === "string" && !raw.startsWith("pt-") && (raw.startsWith("jt-") || (credentials?.accessToken || "").startsWith("jt-"))) {
-      return `${QODER_CHAT_BASE_ALT}/algo${QODER_CHAT_SIG_PATH}?FetchKeys=llm_model_result&AgentId=agent_common&Encode=1`;
-    }
-    return QODER_CHAT_URL_ENCODED;
+    return `${qoderInferenceBase(credentials)}/algo${QODER_CHAT_SIG_PATH}?FetchKeys=llm_model_result&AgentId=agent_common&Encode=1`;
  }

  // Override execute entirely — Qoder needs:
--- a/open-sse/executors/xiaomi-mimo.js
+++ b/open-sse/executors/xiaomi-mimo.js
@@ -0,0 +1,99 @@
+import { DefaultExecutor } from "./default.js";
+import { getMimoAccountCookie, invalidateMimoAccountCookieCache, MIMO_API_BASE, MIMO_API_UA } from "../shared/mimoAccount.js";
+
+// Desktop-exclusive Preview models. These are served by the account service's
+// /api/route proxy, authorized by the Xiaomi account session (NOT the sk- key).
+// See shared/mimoAccount.js for the session handshake.
+const PREVIEW_MODELS = new Set(["mimo-x-pro-preview", "mimo-x-flash-preview"]);
+
+// Session cookie resolved in execute() (async) and read back by buildHeaders()
+// (sync — BaseExecutor.execute does not await it). Carried on the per-request
+// credentials object, same as runtimeTransport.
+const COOKIE_KEY = "__mimoAccountCookie";
+
+// Upstream calls may hand us either the bare id or a `provider/model` ref.
+function bareModel(model) {
+  const s = String(model || "");
+  const i = s.indexOf("/");
+  return i >= 0 ? s.slice(i + 1) : s;
+}
+
+export class XiaomiMimoExecutor extends DefaultExecutor {
+  constructor() {
+    super("xiaomi-mimo");
+  }
+
+  static isPreviewModel(model) {
+    return PREVIEW_MODELS.has(bareModel(model));
+  }
+
+  buildUrl(model, stream, urlIndex = 0, credentials = null) {
+    // Preview models live on the account-service route, which is not one of the
+    // declared transports — resolve it before the default runtimeTransport path.
+    if (XiaomiMimoExecutor.isPreviewModel(model)) {
+      return `${MIMO_API_BASE}/api/route/chat/completions`;
+    }
+    // Cloud API models keep default handling, so a Claude-format client reaches
+    // the /anthropic/v1/messages transport.
+    return super.buildUrl(model, stream, urlIndex, credentials);
+  }
+
+  buildHeaders(credentials, stream = true, url, model) {
+    if (XiaomiMimoExecutor.isPreviewModel(model) && credentials?.[COOKIE_KEY]) {
+      // Preview models authenticate with the account-session cookie, not the key.
+      return {
+        "Content-Type": "application/json",
+        Accept: stream ? "text/event-stream" : "application/json",
+        "User-Agent": MIMO_API_UA,
+        Cookie: credentials[COOKIE_KEY],
+      };
+    }
+    return super.buildHeaders(credentials, stream, url, model);
+  }
+
+  transformRequest(model, body, stream, credentials) {
+    // super runs stripUnsupportedParams, which flattens Preview content-part
+    // arrays (see the xiaomi-mimo rule in translator/concerns/paramSupport.js).
+    const out = super.transformRequest(model, body, stream, credentials);
+
+    // Preview models: thinking/params get defaults only — never override what the
+    // caller set explicitly. (body.model is already `xiaomi/<id>` via upstreamModelId.)
+    if (XiaomiMimoExecutor.isPreviewModel(model)) {
+      if (out.thinking == null) out.thinking = { type: "enabled" };
+      if (out.temperature == null) out.temperature = 1.0;
+      if (out.top_p == null) out.top_p = 0.95;
+      if (!out.max_tokens) out.max_tokens = 4096;
+    }
+
+    return out;
+  }
+
+  async execute(args) {
+    const { model, credentials, proxyOptions = null } = args;
+    if (!XiaomiMimoExecutor.isPreviewModel(model)) return super.execute(args);
+
+    const cookie = await getMimoAccountCookie(credentials?.providerSpecificData, proxyOptions);
+    if (!cookie) {
+      throw new Error(
+        "Xiaomi MiMo account session unavailable. Sign in to MiMo Desktop once so its passToken is present, then retry.",
+      );
+    }
+    credentials[COOKIE_KEY] = cookie;
+    const result = await super.execute(args);
+
+    // A cached session can expire early — drop it and retry once with a fresh one.
+    if (result.response.status === 401) {
+      invalidateMimoAccountCookieCache();
+      const fresh = await getMimoAccountCookie(credentials?.providerSpecificData, proxyOptions).catch(() => null);
+      if (fresh) {
+        credentials[COOKIE_KEY] = fresh;
+        return super.execute(args);
+      }
+    }
+    return result;
+  }
+}
+
+export const __test__ = { PREVIEW_MODELS, bareModel, COOKIE_KEY };
+
+export default XiaomiMimoExecutor;
--- a/open-sse/handlers/chatCore.js
+++ b/open-sse/handlers/chatCore.js
@@ -28,7 +28,7 @@ import { compressWithPxpipe } from "../rtk/pxpipe.js";
 import { getCapabilitiesForModel } from "../providers/capabilities.js";
 import { stripUnsupportedModalities } from "../translator/concerns/modality.js";
 import { prefetchRemoteImages } from "../translator/concerns/prefetch.js";
-import { defaultClaudeToolType } from "../translator/concerns/toolCall.js";
+import { defaultClaudeToolType, shouldDefaultClaudeToolType } from "../translator/concerns/toolCall.js";
 import { resolveSessionId } from "../utils/sessionManager.js";
 import { maybeRejectEarlyStreamError } from "../utils/streamErrorPeek.js";

@@ -246,7 +246,11 @@ export async function handleChatCore({ body, modelInfo, credentials, log, onCred

  // Claude tool schema requires `type` to be explicitly set; strict gateways (e.g., MiniMax)
  // reject legacy payloads that omit it with HTTP 400. Default to "custom" when missing.
-  if (finalFormat === FORMATS.CLAUDE && Array.isArray(translatedBody.tools)) {
+  // Provider-scoped via quirks (shouldDefaultClaudeToolType): only gateways that declare
+  // requireClaudeToolType get the explicit type. Applying it unconditionally breaks
+  // Claude-format endpoints that only accept the legacy typeless tool shape — DeepSeek's
+  // Anthropic-compatible endpoint 400s with "unknown variant `custom`" (#3905).
+  if (shouldDefaultClaudeToolType(provider, finalFormat, translatedBody.tools, PROVIDERS)) {
    translatedBody.tools = defaultClaudeToolType(translatedBody.tools);
  }

--- a/open-sse/handlers/chatCore/nonStreamingHandler.js
+++ b/open-sse/handlers/chatCore/nonStreamingHandler.js
@@ -6,6 +6,7 @@ import { addBufferToUsage, filterUsageForFormat } from "../../utils/usageTrackin
 import { createErrorResult } from "../../utils/error.js";
 import { HTTP_STATUS } from "../../config/runtimeConfig.js";
 import { parseSSEToOpenAIResponse } from "./sseToJsonHandler.js";
+import { unwrapClineEnvelope } from "../../shared/clineEnvelope.js";
 import { buildRequestDetail, extractRequestConfig, extractUsageFromResponse, saveUsageStats, formatDoneLine, tokensForDetail, shouldPersistRequestDetail } from "./requestDetail.js";
 import { saveRequestDetail } from "@/lib/usageDb.js";
 import { matchStreamErrorPatterns } from "../../utils/streamErrorPatterns.js";
@@ -305,6 +306,11 @@ export async function handleNonStreamingResponse({ providerResponse, provider, m
    }
  }

+  // Unwrap before any consumer reads choices/usage so non-stream clients get a
+  // bare OpenAI body and usage tracking sees data.usage. No-op unless the
+  // provider opts in via transport.quirks.clineEnvelope.
+  responseBody = unwrapClineEnvelope(responseBody, provider);
+
  reqLogger.logProviderResponse(providerResponse.status, providerResponse.statusText, providerResponse.headers, responseBody);
  if (onRequestSuccess) {
    Promise.resolve()
--- a/open-sse/handlers/imageProviders/codex.js
+++ b/open-sse/handlers/imageProviders/codex.js
@@ -2,13 +2,21 @@
 import { randomUUID } from "node:crypto";
 import { nowSec } from "./_base.js";
 import { PROVIDERS } from "../../config/providers.js";
+import { CODEX_CLI_VERSION } from "../../config/appConstants.js";

 const CODEX_RESPONSES_URL = PROVIDERS["codex"].baseUrl;
-const CODEX_USER_AGENT = "codex_cli_rs/0.136.0";
-const CODEX_VERSION = "0.136.0";
+const CODEX_USER_AGENT = `codex_cli_rs/${CODEX_CLI_VERSION}`;
 const CODEX_ORIGINATOR = "codex_cli_rs";
 const CODEX_MODEL_SUFFIX = "-image";
 const CODEX_REF_DETAIL = "high";
+const CODEX_IMAGES_MAIN_MODEL = "gpt-5.5";
+const CODEX_TOOL_IMAGE_MODELS = new Set([
+  "gpt-image-1.5",
+  "gpt-image-2",
+  "gpt-image-2.5",
+  "gpt-image-2.5-flare",
+  "gpt-image-2.5-sunburst",
+]);

 function decodeAccountId(idToken) {
  try {
@@ -27,6 +35,13 @@ function stripImageSuffix(model) {
  return model.endsWith(CODEX_MODEL_SUFFIX) ? model.slice(0, -CODEX_MODEL_SUFFIX.length) : model;
 }

+function resolveCodexImageModels(model) {
+  if (CODEX_TOOL_IMAGE_MODELS.has(model)) {
+    return { responsesModel: CODEX_IMAGES_MAIN_MODEL, toolModel: model };
+  }
+  return { responsesModel: stripImageSuffix(model), toolModel: null };
+}
+
 function toDataUrl(input) {
  if (!input || typeof input !== "string") return null;
  if (/^data:image\//i.test(input) || /^https?:\/\//i.test(input)) return input;
@@ -157,7 +172,7 @@ export default {
      "originator": CODEX_ORIGINATOR,
      "session_id": randomUUID(),
      "user-agent": CODEX_USER_AGENT,
-      "version": CODEX_VERSION,
+      "version": CODEX_CLI_VERSION,
      "x-client-request-id": randomUUID(),
    };
  },
@@ -167,21 +182,26 @@ export default {
    const single = toDataUrl(body.image);
    if (single) refs.push(single);
    const detail = body.image_detail || CODEX_REF_DETAIL;
+    const { responsesModel, toolModel } = resolveCodexImageModels(model);
    const imgTool = { type: "image_generation", output_format: (body.output_format || "png").toLowerCase() };
+    if (toolModel) {
+      imgTool.action = refs.length > 0 ? "edit" : "generate";
+      imgTool.model = toolModel;
+    }
    if (body.size && body.size !== "") imgTool.size = body.size;
    if (body.quality && body.quality !== "") imgTool.quality = body.quality;
    if (body.background && body.background !== "") imgTool.background = body.background;
    return {
-      model: stripImageSuffix(model),
+      model: responsesModel,
      instructions: "",
      input: [{ type: "message", role: "user", content: buildContent(body.prompt, refs, detail) }],
      tools: [imgTool],
-      tool_choice: "auto",
+      tool_choice: toolModel ? { type: "image_generation" } : "auto",
      parallel_tool_calls: false,
      prompt_cache_key: randomUUID(),
      stream: true,
      store: false,
-      reasoning: null,
+      reasoning: toolModel ? { effort: "medium", summary: "auto" } : null,
    };
  },
  // Custom: codex parses SSE → either pipe to client or collect b64
--- a/open-sse/handlers/videoCore.js
+++ b/open-sse/handlers/videoCore.js
@@ -2,6 +2,7 @@ import { createErrorResult } from "../utils/error.js";
 import { HTTP_STATUS } from "../config/runtimeConfig.js";
 import { refreshTokenByProvider } from "../services/tokenRefresh.js";
 import { PROVIDER_MEDIA } from "../providers/index.js";
+import { getVideoAdapter } from "./videoProviders/index.js";

 // Upstream fetch deadline for video job submission/polling (the job itself is
 // async upstream — this only bounds the HTTP round-trip, not video rendering).
@@ -94,21 +95,49 @@ export async function handleVideoProxyCore({
    return createErrorResult(HTTP_STATUS.BAD_REQUEST, `Unknown video action: ${action}`);
  }

-  const method = requestId ? "GET" : "POST";
-  const url = buildUpstreamUrl(config, action, requestId);
+  const adapter = getVideoAdapter(provider);
  const fetchSignal = combineSignals(signal, timeoutMs);

-  const doFetch = (token) =>
-    fetch(url, {
+  // Default (xAI shape) request plan; adapters override URL/method/headers/body.
+  const defaultPlan = () => {
+    const method = requestId ? "GET" : "POST";
+    return {
      method,
-      headers: buildHeaders({ token, contentType: method === "POST" ? contentType : null, idempotencyKey: method === "POST" ? idempotencyKey : null }),
+      url: buildUpstreamUrl(config, action, requestId),
+      headers: buildHeaders({
+        token: credentials?.accessToken || credentials?.apiKey,
+        contentType: method === "POST" ? contentType : null,
+        idempotencyKey: method === "POST" ? idempotencyKey : null,
+      }),
      body: method === "POST" ? rawBody : undefined,
-      signal: fetchSignal,
-    });
+    };
+  };

+  // Rebuilt per attempt so the auth retry below picks up the refreshed token.
+  const doFetch = async () => {
+    const plan = adapter
+      ? await adapter.buildRequest({
+          config, action, requestId, rawBody, contentType, idempotencyKey, credentials, log,
+          token: credentials?.accessToken || credentials?.apiKey,
+        })
+      : defaultPlan();
+    if (plan.error) return { planError: plan.error };
+    return {
+      response: await fetch(plan.url, {
+        method: plan.method,
+        headers: plan.headers,
+        body: plan.body,
+        signal: fetchSignal,
+      }),
+    };
+  };
+
+  const method = requestId ? "GET" : "POST";
  let upstream;
  try {
-    upstream = await doFetch(credentials?.accessToken || credentials?.apiKey);
+    const first = await doFetch();
+    if (first.planError) return createErrorResult(HTTP_STATUS.BAD_REQUEST, `[${provider}] ${first.planError}`);
+    upstream = first.response;
  } catch (error) {
    if (error?.name === "AbortError" || error?.name === "TimeoutError") {
      return createErrorResult(HTTP_STATUS.REQUEST_TIMEOUT, `[${provider}] video ${method} aborted: ${error.message}`);
@@ -136,7 +165,9 @@ export async function handleVideoProxyCore({
        await upstream.body?.cancel?.();
      } catch { /* noop */ }
      try {
-        upstream = await doFetch(credentials.accessToken || credentials.apiKey);
+        const retry = await doFetch();
+        if (retry.planError) return createErrorResult(HTTP_STATUS.BAD_REQUEST, `[${provider}] ${retry.planError}`);
+        upstream = retry.response;
      } catch (error) {
        return createErrorResult(HTTP_STATUS.BAD_GATEWAY, sanitizeSecrets(`[${provider}] video retry after refresh failed: ${error.message}`, credentials));
      }
@@ -152,13 +183,25 @@ export async function handleVideoProxyCore({
    return createErrorResult(upstream.status, `[${provider}] ${message.slice(0, 2000)}`);
  }

-  // Success: pass the upstream JSON through untouched (request_id / status / video.url).
+  // Success: pass the upstream JSON through untouched (request_id / status / video.url),
+  // unless the adapter maps a provider-native shape onto it (Vertex operations).
+  let outBody = bodyText;
+  let outType = upstream.headers.get("content-type") || "application/json";
+  if (adapter?.transformResponse) {
+    try {
+      outBody = JSON.stringify(adapter.transformResponse(JSON.parse(bodyText)));
+      outType = "application/json";
+    } catch {
+      // Non-JSON or unexpected shape — fall back to the raw upstream body.
+    }
+  }
+
  return {
    success: true,
-    response: new Response(bodyText, {
+    response: new Response(outBody, {
      status: upstream.status,
      headers: {
-        "Content-Type": upstream.headers.get("content-type") || "application/json",
+        "Content-Type": outType,
        "Access-Control-Allow-Origin": "*",
      },
    }),
--- a/open-sse/handlers/videoProviders/index.js
+++ b/open-sse/handlers/videoProviders/index.js
@@ -0,0 +1,13 @@
+// Video provider adapters.
+//
+// Default (no adapter) = xAI shape: raw body forwarded to {baseUrl}/{action},
+// polled at {baseUrl}/{id}, upstream JSON passed through verbatim.
+// A provider only needs an adapter when its wire format differs from that.
+import openrouter from "./openrouter.js";
+import vertex from "./vertex.js";
+
+const ADAPTERS = { openrouter, vertex };
+
+export function getVideoAdapter(provider) {
+  return ADAPTERS[provider] || null;
+}
--- a/open-sse/handlers/videoProviders/openrouter.js
+++ b/open-sse/handlers/videoProviders/openrouter.js
@@ -0,0 +1,39 @@
+// OpenRouter video jobs — https://openrouter.ai/docs/api/api-reference/videos
+//
+// Same async shape as xAI (POST → { id, status }, GET → status/unsigned_urls),
+// two differences only: creation POSTs to the collection root (no `/generations`
+// suffix) and the account headers come from the registry entry.
+// Response bodies are passed through verbatim.
+
+// ponytail: generations only — OpenRouter has no edits/extensions endpoint today.
+const SUPPORTED_ACTIONS = new Set(["generations"]);
+
+function headers(config, token) {
+  return {
+    Accept: "application/json",
+    ...(config.headers || {}),
+    ...(token ? { Authorization: `Bearer ${token}` } : {}),
+  };
+}
+
+export default {
+  buildRequest({ config, action, requestId, rawBody, contentType, token }) {
+    const base = config.baseUrl.replace(/\/$/, "");
+
+    if (requestId) {
+      return { method: "GET", url: `${base}/${encodeURIComponent(requestId)}`, headers: headers(config, token) };
+    }
+    if (!SUPPORTED_ACTIONS.has(action)) {
+      return { error: `OpenRouter video supports 'generations' only (got '${action}')` };
+    }
+    if (contentType && !contentType.includes("application/json")) {
+      return { error: "OpenRouter video requires an application/json body" };
+    }
+    return {
+      method: "POST",
+      url: base,
+      headers: { ...headers(config, token), "Content-Type": "application/json" },
+      body: rawBody,
+    };
+  },
+};
--- a/open-sse/handlers/videoProviders/vertex.js
+++ b/open-sse/handlers/videoProviders/vertex.js
@@ -0,0 +1,159 @@
+// Vertex AI (Veo) video jobs.
+//
+// Vertex does NOT speak the OpenAI-ish /v1/videos shape, so unlike OpenRouter
+// this adapter translates both directions:
+//   create → POST {model}:predictLongRunning  { instances[], parameters{} } → { name }
+//   poll   → POST {model}:fetchPredictOperation { operationName } → { done, response }
+// Docs: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/veo-video-generation
+//
+// The operation name is a resource path (contains "/"), so it is base64url-encoded
+// into the job id returned to the client — GET /v1/videos/{id} stays a flat path.
+import { parseVertexSaJson, refreshVertexToken } from "../../services/tokenRefresh.js";
+
+const DEFAULT_LOCATION = "us-central1";
+
+const encodeJobId = (name) => Buffer.from(name, "utf8").toString("base64url");
+
+// Operation name shape: projects/{p}/locations/{l}/publishers/{pub}/models/{m}/operations/{op}.
+// Anchored and single-segment-per-field so a decoded path can never carry `..` or a
+// host-changing prefix into the request URL.
+const OPERATION_NAME_RE = /^projects\/[^/]+\/locations\/[^/]+\/publishers\/[^/]+\/models\/[^/]+\/operations\/[^/]+$/;
+
+function modelPathOf(operationName) {
+  return operationName.slice(0, operationName.indexOf("/operations/"));
+}
+
+function decodeJobId(id) {
+  const raw = String(id ?? "");
+  // Buffer.from(x, "base64url") silently drops invalid characters instead of
+  // throwing, so only ids that re-encode byte-for-byte are accepted.
+  if (!raw || raw.length > 1024 || !/^[A-Za-z0-9_-]+$/.test(raw)) return null;
+  const decoded = Buffer.from(raw, "base64url").toString("utf8");
+  if (Buffer.from(decoded, "utf8").toString("base64url") !== raw) return null;
+  return OPERATION_NAME_RE.test(decoded) ? decoded : null;
+}
+
+async function resolveAuth(credentials, log) {
+  const saJson = parseVertexSaJson(credentials?.apiKey);
+  const projectId =
+    saJson?.project_id ||
+    credentials?.projectId ||
+    credentials?.providerSpecificData?.projectId;
+  const location = credentials?.providerSpecificData?.location || DEFAULT_LOCATION;
+
+  if (!projectId) {
+    return { error: "Vertex video requires a project_id — use Service Account JSON or set providerSpecificData.projectId" };
+  }
+
+  let token = credentials?.accessToken;
+  if (saJson) {
+    const minted = await refreshVertexToken(saJson, log);
+    if (!minted?.accessToken) return { error: "Vertex video: failed to mint access token from service account JSON" };
+    token = minted.accessToken;
+  }
+  if (!token) return { error: "Vertex video requires Service Account JSON or an OAuth access token (raw API keys are not supported)" };
+
+  return { token, projectId, location };
+}
+
+/** OpenAI-ish video body → Vertex predictLongRunning body. */
+function toVertexBody(body) {
+  const instance = { prompt: body.prompt };
+  // Image-to-video: accept the Vertex-native shape or a bare data URL / base64 string.
+  const image = body.image ?? body.image_url;
+  if (image && typeof image === "object") {
+    instance.image = image;
+  } else if (typeof image === "string") {
+    const match = image.match(/^data:([^;]+);base64,(.*)$/s);
+    instance.image = match
+      ? { bytesBase64Encoded: match[2], mimeType: match[1] }
+      : { gcsUri: image };
+  }
+  if (body.video && typeof body.video === "object") instance.video = body.video;
+
+  const parameters = {};
+  if (body.n != null) parameters.sampleCount = Number(body.n);
+  if (body.duration != null) parameters.durationSeconds = Number(body.duration);
+  if (body.aspect_ratio) parameters.aspectRatio = body.aspect_ratio;
+  if (body.resolution) parameters.resolution = body.resolution;
+  if (body.seed != null) parameters.seed = body.seed;
+  if (body.negative_prompt) parameters.negativePrompt = body.negative_prompt;
+  // Without storageUri Vertex returns inline base64 bytes; a GCS bucket keeps
+  // the poll response small and is what production callers want.
+  if (body.storage_uri) parameters.storageUri = body.storage_uri;
+  if (body.generate_audio != null) parameters.generateAudio = !!body.generate_audio;
+
+  return { instances: [instance], ...(Object.keys(parameters).length ? { parameters } : {}) };
+}
+
+/** Vertex operation → the async-job shape 9Router clients already poll for. */
+function fromVertexOperation(json) {
+  if (!json?.name) return json;
+  const id = encodeJobId(json.name);
+  if (json.error) {
+    return { id, request_id: id, status: "failed", error: json.error };
+  }
+  if (!json.done) {
+    return { id, request_id: id, status: "pending" };
+  }
+  const samples =
+    json.response?.videos ||
+    json.response?.generateVideoResponse?.generatedSamples ||
+    [];
+  const videos = samples.map((s) => ({
+    url: s.gcsUri || s.video?.uri || s.uri || null,
+    b64_json: s.bytesBase64Encoded || s.video?.bytesBase64Encoded || null,
+    mime_type: s.mimeType || s.video?.mimeType || "video/mp4",
+  }));
+  return { id, request_id: id, status: "completed", video: videos[0] || null, videos };
+}
+
+export default {
+  async buildRequest({ config, action, requestId, rawBody, contentType, credentials, log }) {
+    if (contentType && !contentType.includes("application/json")) {
+      return { error: "Vertex video requires an application/json body" };
+    }
+
+    const auth = await resolveAuth(credentials, log);
+    if (auth.error) return { error: auth.error };
+    const { token, projectId, location } = auth;
+    const base = (config.baseUrl || "https://aiplatform.googleapis.com").replace(/\/$/, "");
+    const headers = { Accept: "application/json", "Content-Type": "application/json", Authorization: `Bearer ${token}` };
+
+    if (requestId) {
+      const operationName = decodeJobId(requestId);
+      if (!operationName) return { error: "Invalid Vertex video job id" };
+      return {
+        method: "POST",
+        url: `${base}/v1/${modelPathOf(operationName)}:fetchPredictOperation`,
+        headers,
+        body: JSON.stringify({ operationName }),
+      };
+    }
+
+    if (action !== "generations") {
+      // ponytail: Veo extend/edit go through generations with `video`/`image` in the body.
+      return { error: `Vertex video supports 'generations' only (got '${action}')` };
+    }
+
+    let body;
+    try {
+      body = JSON.parse(typeof rawBody === "string" ? rawBody : rawBody.toString("utf8"));
+    } catch {
+      return { error: "Invalid JSON body" };
+    }
+    if (!body.model) return { error: "Vertex video requires a model (e.g. vertex/veo-3.1-generate-preview)" };
+    // Plain model id only — a path segment carrying "/" or ".." would rewrite the URL.
+    if (!/^[A-Za-z0-9._-]+$/.test(body.model)) return { error: "Invalid Vertex video model id" };
+    if (!body.prompt && !body.image && !body.image_url) return { error: "Vertex video requires a prompt or an image" };
+
+    return {
+      method: "POST",
+      url: `${base}/v1/projects/${projectId}/locations/${location}/publishers/google/models/${body.model}:predictLongRunning`,
+      headers,
+      body: JSON.stringify(toVertexBody(body)),
+    };
+  },
+
+  transformResponse: fromVertexOperation,
+};
--- a/open-sse/providers/capabilities.js
+++ b/open-sse/providers/capabilities.js
@@ -209,7 +209,10 @@ export const PROVIDER_CAPABILITIES = {
    "glm-5.3-flash":      { vision: true, reasoning: true, thinkingFormat: "openai", thinkingCanDisable: true, contextWindow: 1000000, maxOutput: 32000 },
    "kimi-k3-1":          { vision: true, reasoning: true, thinkingFormat: "openai", thinkingCanDisable: false, contextWindow: 1000000, maxOutput: 32000 },
    "deepseek-v4-pro":    { vision: true, reasoning: true, thinkingFormat: "openai", thinkingCanDisable: true, contextWindow: 1000000, maxOutput: 50000 },
-    "deepseek-v4-flash":  { vision: true, reasoning: true, thinkingFormat: "openai", thinkingCanDisable: true, contextWindow: 1000000, maxOutput: 50000 },
+    // deepseek-v4.1-flash replaces v4-flash (dropped from the server list;
+    // the old endpoint still answers 200 but the published list is the
+    // contract). maxOutput 128000 per the server's product-config payload.
+    "deepseek-v4.1-flash": { vision: true, reasoning: true, thinkingFormat: "openai", thinkingCanDisable: true, contextWindow: 1000000, maxOutput: 128000 },
  },
  // Qoder — upstream exposes opaque internal ids (dfmodel, kmodel, …); the
  // registry `name` is display-only and capability lookup matches on the raw
@@ -219,9 +222,9 @@ export const PROVIDER_CAPABILITIES = {
  // windows (GLM-5.3 / Kimi-K3 / Qwen3.8-Max claim 180K but accept more).
  // max_output_tokens arrives as 0 for every model, so outputs are
  // best-guess from the real model family. Vision tags below follow the
-  // upstream is_vl flag per explicit request, even though the executor
-  // currently sends image_urls:null (image pass-through over the agent_chat
-  // SSE protocol is unverified). reasoning:true on all of them — every model can
+  // upstream is_vl flag. The executor uploads inlined images to
+  // /api/v2/image/upload and leaves image_urls/chat_context.imageUrls null
+  // (same as qodercli). reasoning:true on all of them — every model can
  // reason; the upstream is_reasoning flag only drives model_config selection.
  // thinkingFormat keeps the true-model family for documentation/UI, but
  // thinkingCanDisable:false everywhere: the executor only forwards
--- a/open-sse/providers/registry/antigravity.js
+++ b/open-sse/providers/registry/antigravity.js
@@ -37,6 +37,7 @@ export default {
    },
    usage: {
      quotaApiUrl: `${ANTIGRAVITY_IDE_BASE_URL}/v1internal:fetchAvailableModels`,
+      quotaSummaryApiUrl: `${ANTIGRAVITY_IDE_BASE_URL}/v1internal:retrieveUserQuotaSummary`,
      loadProjectApiUrl: "https://cloudcode-pa.googleapis.com/v1internal:loadCodeAssist",
      tokenUrl: "https://oauth2.googleapis.com/token",
    },
--- a/open-sse/providers/registry/api-airforce.js
+++ b/open-sse/providers/registry/api-airforce.js
@@ -20,6 +20,8 @@ export default {
  authModes: [
    "apikey",
  ],
+  passthroughModels: true,
+  modelsFetcher: { url: "https://api.airforce/v1/models", type: "airforce-free" },
  transport: {
    baseUrl: "https://api.airforce/v1/chat/completions",
    validateUrl: "https://api.airforce/v1/models",
@@ -27,10 +29,11 @@ export default {
      "HTTP-Referer": "https://endpoint-proxy.local",
      "X-Title": "Endpoint Proxy",
    },
+    forceStream: true,
  },
  models: [
-    { id: "anthropic/claude-3.7-sonnet", name: "Claude 3.7 Sonnet (Free)", contextLength: 200000 },
-    { id: "moonshot/kimi-k2.6", name: "Kimi K2.6 (Free)", contextLength: 262144 },
-    { id: "google/gemini-2.5-flash", name: "Gemini 2.5 Flash (Free)", contextLength: 1048576 },
+    { id: "gpt-oss-120b", name: "GPT-OSS 120B (Free)", contextLength: 131072 },
+    { id: "gpt-oss-20b", name: "GPT-OSS 20B (Free)", contextLength: 131072 },
+    { id: "kimi-k2.7-code", name: "Kimi K2.7 Code (Free)", contextLength: 262144 },
  ],
 };
--- a/open-sse/providers/registry/cline.js
+++ b/open-sse/providers/registry/cline.js
@@ -14,12 +14,16 @@ export default {
    },
  },
  category: "oauth",
+  authModes: ["oauth"],
+  hasOAuth: true,
  transport: {
    baseUrl: "https://api.cline.bot/api/v1/chat/completions",
    headers: {
      "HTTP-Referer": "https://cline.bot",
      "X-Title": "Cline",
    },
+    // Non-stream chat completions come back wrapped in {"success":true,"data":{...}}
+    quirks: { clineEnvelope: true },
    tokenUrl: "https://api.cline.bot/api/v1/auth/token",
    refreshUrl: "https://api.cline.bot/api/v1/auth/refresh",
    auth: {
--- a/open-sse/providers/registry/clinepass.js
+++ b/open-sse/providers/registry/clinepass.js
@@ -14,7 +14,10 @@ export default {
    },
  },
  category: "oauth",
-  authModes: ["oauth", "apikey"],
+  // ClinePass authenticates with a plain API key from app.cline.bot/settings/api-keys
+  // (category "apikey"). The OAuth extension flow used by Cline does not issue
+  // tokens that the ClinePass API consumer endpoint accepts (HTTP 401) — see #2333.
+  authModes: ["apikey", "oauth"],
  hasOAuth: true,
  transport: {
    baseUrl: "https://api.cline.bot/api/v1/chat/completions",
@@ -22,6 +25,8 @@ export default {
      "HTTP-Referer": "https://cline.bot",
      "X-Title": "Cline",
    },
+    // Non-stream chat completions come back wrapped in {"success":true,"data":{...}}
+    quirks: { clineEnvelope: true },
    auth: {
      combined: true,
      header: "Authorization",
--- a/open-sse/providers/registry/codebuddy-cn.js
+++ b/open-sse/providers/registry/codebuddy-cn.js
@@ -58,7 +58,9 @@ export default {
    // (endpoint returns 11102 "model service info not found"), plus
    // glm-5.0-turbo / minimax-m2.7 / kimi-k2.5 / hy3-preview /
    // deepseek-v3-2-volc (absent from the server list, though still answering
-    // 200) and hy3-x (paid tier, not used here).
+    // 200) and hy3-x (paid tier, not used here). deepseek-v4-flash removed
+    // 2026-09: replaced server-side by deepseek-v4.1-flash (same low/high/
+    // xhigh efforts; endpoint still answers 200 but the list is the contract).
    // "-x" suffix = paid tier of the same model (free id rides the promo quota).
    { id: "hy3", name: "Hy3" },
    { id: "hy4-preview", name: "Hy4-Preview" },
@@ -66,7 +68,7 @@ export default {
    { id: "glm-5.3-flash", name: "GLM-5.3-Flash" },
    { id: "kimi-k3-1", name: "Kimi-K3" },
    { id: "deepseek-v4-pro", name: "DeepSeek-V4-Pro" },
-    { id: "deepseek-v4-flash", name: "DeepSeek-V4-Flash" },
+    { id: "deepseek-v4.1-flash", name: "DeepSeek-V4.1-Flash" },
  ],
  oauth: {
    baseUrl: "https://copilot.tencent.com",
--- a/open-sse/providers/registry/codex.js
+++ b/open-sse/providers/registry/codex.js
@@ -1,5 +1,9 @@
 import { withCodexReviewModels } from "../models/helpers.js";

+// Codex CLI version seen by OpenAI's backend — single source for the Version /
+// User-Agent identity headers. Bump when the installed codex CLI is upgraded.
+const CODEX_CLI_VERSION = "0.154.0";
+
 export default {
  id: "codex",
  priority: 30,
@@ -34,9 +38,10 @@ export default {
    baseUrl: "https://chatgpt.com/backend-api/codex/responses",
    format: "openai-responses",
    forceStream: true,
+    cliVersion: CODEX_CLI_VERSION,
    headers: {
      originator: "codex_cli_rs",
-      "User-Agent": "codex_cli_rs/0.136.0",
+      "User-Agent": `codex_cli_rs/${CODEX_CLI_VERSION}`,
    },
    usage: {
      url: "https://chatgpt.com/backend-api/wham/usage",
@@ -60,6 +65,11 @@ export default {
    { id: "gpt-5.4-mini-review", name: "GPT 5.4 Mini Review", upstreamModelId: "gpt-5.4-mini", quotaFamily: "review" },
    { id: "gpt-5.3-codex-spark", name: "GPT 5.3 Codex Spark" },
    { id: "gpt-5.3-codex-spark-review", name: "GPT 5.3 Codex Spark Review", upstreamModelId: "gpt-5.3-codex-spark", quotaFamily: "review" },
+    { id: "gpt-image-2.5", name: "GPT Image 2.5", capabilities: ["text2img","edit","multiImage"], params: ["size","quality","background","image_detail","output_format"], kind: "image" },
+    { id: "gpt-image-2.5-flare", name: "GPT Image 2.5 Flare", capabilities: ["text2img","edit","multiImage"], params: ["size","quality","background","image_detail","output_format"], kind: "image" },
+    { id: "gpt-image-2.5-sunburst", name: "GPT Image 2.5 Sunburst", capabilities: ["text2img","edit","multiImage"], params: ["size","quality","background","image_detail","output_format"], kind: "image" },
+    { id: "gpt-image-2", name: "GPT Image 2", capabilities: ["text2img","edit","multiImage"], params: ["size","quality","background","image_detail","output_format"], kind: "image" },
+    { id: "gpt-image-1.5", name: "GPT Image 1.5", capabilities: ["text2img","edit","multiImage"], params: ["size","quality","background","image_detail","output_format"], kind: "image" },
    { id: "gpt-5.6-sol-image", name: "GPT 5.6 Sol Image", capabilities: ["text2img","edit"], params: ["size","quality","background","image_detail","output_format"], kind: "image" },
    { id: "gpt-5.6-terra-image", name: "GPT 5.6 Terra Image", capabilities: ["text2img","edit"], params: ["size","quality","background","image_detail","output_format"], kind: "image" },
    { id: "gpt-5.6-luna-image", name: "GPT 5.6 Luna Image", capabilities: ["text2img","edit"], params: ["size","quality","background","image_detail","output_format"], kind: "image" },
--- a/open-sse/providers/registry/deepseek.js
+++ b/open-sse/providers/registry/deepseek.js
@@ -25,6 +25,21 @@ export default {
    reasoningInject: {
      scope: "all",
    },
+    quirks: {
+      // DeepSeek's Anthropic-compatible endpoint
+      // (https://api.deepseek.com/anthropic/v1/messages) accepts ONLY the
+      // built-in web_search_* tools and rejects client-defined `custom` tools
+      // (MCP / Read / Bash / etc.) with HTTP 400
+      //   "tools[0]: unknown variant `custom`, expected
+      //    `web_search_20250305` or `web_search_20260209`".
+      //
+      // Declaring this whitelist makes prepareClaudeRequest() forward only
+      // web_search_* tools and strip everything else before sending, so MCP /
+      // function tools are dropped instead of failing the whole request.
+      // DeepSeek's OpenAI-compatible transport is unaffected (targetFormat
+      // there is "openai", not "claude", so prepareClaudeRequest is not run).
+      claudeSupportedToolTypes: ["web_search_20250305", "web_search_20260209"],
+    },
  },
  // Multi-endpoint: pick the transport matching client sourceFormat to skip translation.
  transports: [
--- a/open-sse/providers/registry/index.js
+++ b/open-sse/providers/registry/index.js
@@ -123,7 +123,6 @@ import p119 from "./selfhosted-embedding.js";
 import p120 from "./fish-audio.js";
 import p121 from "./alitp-intl.js";
 import p122 from "./xquik.js";
-
 export default [
  p0,
  p1,
--- a/open-sse/providers/registry/minimax-cn.js
+++ b/open-sse/providers/registry/minimax-cn.js
@@ -22,6 +22,7 @@ export default {
    headers: { ...CLAUDE_API_HEADERS },
    quirks: {
      dropOutputConfig: true,
+      requireClaudeToolType: true,
    },
    reasoningInject: {
      scope: "all",
--- a/open-sse/providers/registry/minimax.js
+++ b/open-sse/providers/registry/minimax.js
@@ -22,6 +22,7 @@ export default {
    headers: { ...CLAUDE_API_HEADERS },
    quirks: {
      dropOutputConfig: true,
+      requireClaudeToolType: true,
    },
    reasoningInject: {
      scope: "all",
--- a/open-sse/providers/registry/openai.js
+++ b/open-sse/providers/registry/openai.js
@@ -57,6 +57,9 @@ export default {
    { id: "whisper-1", name: "Whisper 1", params: ["language","response_format","temperature","prompt"], kind: "stt" },
    { id: "gpt-4o-transcribe", name: "GPT-4o Transcribe", params: ["language","response_format","temperature","prompt"], kind: "stt" },
    { id: "gpt-4o-mini-transcribe", name: "GPT-4o Mini Transcribe", params: ["language","response_format","temperature","prompt"], kind: "stt" },
+    { id: "gpt-image-2.5", name: "GPT Image 2.5", params: ["n","size","quality","response_format"], kind: "image" },
+    { id: "gpt-image-2.5-flare", name: "GPT Image 2.5 Flare", params: ["n","size","quality","response_format"], kind: "image" },
+    { id: "gpt-image-2.5-sunburst", name: "GPT Image 2.5 Sunburst", params: ["n","size","quality","response_format"], kind: "image" },
    { id: "gpt-image-1", name: "GPT Image 1", params: ["n","size","quality","response_format"], kind: "image" },
    { id: "dall-e-3", name: "DALL-E 3", params: ["size","quality","style","response_format"], kind: "image" },
    { id: "dall-e-2", name: "DALL-E 2", params: ["n","size","response_format"], kind: "image" },
--- a/open-sse/providers/registry/opencode-go.js
+++ b/open-sse/providers/registry/opencode-go.js
@@ -33,25 +33,36 @@ export default {
    { format: "claude", baseUrl: "https://opencode.ai/zen/go/v1/messages", auth: { combined: true, header: "x-api-key", scheme: "raw", anthropicVersion: true } },
    { format: "openai-responses", baseUrl: "https://opencode.ai/zen/go/v1/responses", auth: { combined: true, header: "Authorization", scheme: "bearer" } },
  ],
+  // supportedFormats follow the endpoint table in https://opencode.ai/docs/go/
  models: [
+    { id: "deepseek-flash", name: "DeepSeek V4.1 Flash", supportedFormats: ["openai"] },
    { id: "glm-5.3-flash", name: "GLM 5.3 Flash (Vision)", supportedFormats: ["openai"] },
+    { id: "glm-5.3", name: "GLM 5.3", supportedFormats: ["openai"] },
    { id: "glm-5.2", name: "GLM 5.2", supportedFormats: ["openai"] },
    { id: "glm-5.1", name: "GLM 5.1", supportedFormats: ["openai"] },
    { id: "kimi-k2.7-code", name: "Kimi K2.7 Code", supportedFormats: ["openai"] },
    { id: "kimi-k2.6", name: "Kimi K2.6", supportedFormats: ["openai"] },
+    { id: "kimi-k3", name: "Kimi K3", supportedFormats: ["openai"] },
    { id: "deepseek-v4-pro", name: "DeepSeek V4 Pro", supportedFormats: ["openai", "claude", "openai-responses"] },
    { id: "deepseek-v4-flash", name: "DeepSeek V4 Flash", supportedFormats: ["openai", "claude", "openai-responses"] },
    { id: "deepseek-v4-flash-vision-exp", name: "DeepSeek V4 Flash Vision (Exp)", supportedFormats: ["openai", "claude", "openai-responses"] },
+    { id: "longcat-2.0", name: "LongCat 2.0", supportedFormats: ["openai"] },
    { id: "mimo-v2.5", name: "MiMo V2.5", supportedFormats: ["openai"] },
    { id: "mimo-v2.5-pro", name: "MiMo V2.5 Pro", supportedFormats: ["openai"] },
    { id: "minimax-m3", name: "MiniMax M3", supportedFormats: ["openai", "claude"] },
    { id: "minimax-m2.7", name: "MiniMax M2.7", supportedFormats: ["openai", "claude"] },
    { id: "minimax-m2.5", name: "MiniMax M2.5", supportedFormats: ["openai", "claude"] },
+    { id: "qwen3.8-max", name: "Qwen 3.8 Max", supportedFormats: ["openai", "claude"] },
+    { id: "qwen3.8-flash", name: "Qwen 3.8 Flash", supportedFormats: ["openai", "claude"] },
    { id: "qwen3.7-max", name: "Qwen 3.7 Max", supportedFormats: ["openai", "claude"] },
    { id: "qwen3.7-plus", name: "Qwen 3.7 Plus", supportedFormats: ["openai", "claude"] },
    { id: "qwen3.6-plus", name: "Qwen 3.6 Plus", supportedFormats: ["openai", "claude"] },
-    // Muse Spark is served by /zen/go/v1/responses only — responses-only entry forces
-    // chatCore past the sourceFormat-matched transports into translation (see chatCore guard).
+    { id: "hy4-preview", name: "Hy4 Preview", supportedFormats: ["openai"] },
+    { id: "hy3", name: "Hy3", supportedFormats: ["openai"] },
+    // Served by /zen/go/v1/responses only — the responses-only entry forces chatCore
+    // past the sourceFormat-matched transports into translation (see chatCore guard).
+    { id: "grok-4.6", name: "Grok 4.6", targetFormat: "openai-responses", supportedFormats: ["openai-responses"] },
+    { id: "gpt-5.6-luna", name: "GPT 5.6 Luna", targetFormat: "openai-responses", supportedFormats: ["openai-responses"] },
    { id: "muse-spark-1.2-contributor", name: "Muse Spark 1.2 Contributor", targetFormat: "openai-responses", supportedFormats: ["openai-responses"] },
    { id: "muse-spark-1.3-contributor", name: "Muse Spark 1.3 Contributor", targetFormat: "openai-responses", supportedFormats: ["openai-responses"] },
  ],
--- a/open-sse/providers/registry/openrouter.js
+++ b/open-sse/providers/registry/openrouter.js
@@ -40,8 +40,11 @@ export default {
    { id: "openai/gpt-image-1", name: "GPT Image 1 (via OpenRouter)", params: ["n","size","quality","response_format"], kind: "image" },
    { id: "google/imagen-3.0-generate-002", name: "Imagen 3 (via OpenRouter)", params: ["n","size"], kind: "image" },
    { id: "black-forest-labs/FLUX.1-schnell", name: "FLUX.1 Schnell (via OpenRouter)", params: ["n","size"], kind: "image" },
+    { id: "google/veo-3.1", name: "Veo 3.1 (via OpenRouter)", params: ["duration","aspect_ratio","resolution"], kind: "video" },
+    { id: "openai/sora-2-pro", name: "Sora 2 Pro (via OpenRouter)", params: ["duration","aspect_ratio","resolution"], kind: "video" },
+    { id: "bytedance/seedance-2.0", name: "Seedance 2.0 (via OpenRouter)", params: ["duration","aspect_ratio","resolution"], kind: "video" },
  ],
-  serviceKinds: ["llm","embedding","tts","imageToText"],
+  serviceKinds: ["llm","embedding","tts","imageToText","video"],
  ttsConfig: {
    baseUrl: "https://openrouter.ai/api/v1/chat/completions",
    defaultModel: "openai/gpt-4o-mini-tts",
@@ -57,6 +60,12 @@ export default {
    baseUrl: "https://openrouter.ai/api/v1/images/generations",
    headers: {"HTTP-Referer":"https://endpoint-proxy.local","X-Title":"Endpoint Proxy"},
  },
+  // Async video jobs (POST /videos → { id, status }, GET /videos/{id} polls).
+  // Docs: https://openrouter.ai/docs/api/api-reference/videos
+  videoConfig: {
+    baseUrl: "https://openrouter.ai/api/v1/videos",
+    headers: {"HTTP-Referer":"https://endpoint-proxy.local","X-Title":"Endpoint Proxy"},
+  },
  modelsFetcher: { url: "https://openrouter.ai/api/v1/models", type: "openrouter-free" },
  passthroughModels: true,
 };
--- a/open-sse/providers/registry/vertex.js
+++ b/open-sse/providers/registry/vertex.js
@@ -27,6 +27,13 @@ export default {
    { id: "gemini-3.1-flash-lite-preview", name: "Gemini 3.1 Flash Lite Preview" },
    { id: "gemini-3-flash-preview", name: "Gemini 3 Flash Preview" },
    { id: "gemini-2.5-flash", name: "Gemini 2.5 Flash" },
+    { id: "veo-3.1-generate-preview", name: "Veo 3.1 (Preview)", params: ["duration","aspect_ratio","resolution","negative_prompt","seed","storage_uri","generate_audio"], kind: "video" },
+    { id: "veo-3.1-fast-generate-preview", name: "Veo 3.1 Fast (Preview)", params: ["duration","aspect_ratio","resolution","negative_prompt","seed","storage_uri","generate_audio"], kind: "video" },
+    { id: "veo-3.0-generate-001", name: "Veo 3", params: ["duration","aspect_ratio","resolution","negative_prompt","seed","storage_uri","generate_audio"], kind: "video" },
+    { id: "veo-2.0-generate-001", name: "Veo 2", params: ["duration","aspect_ratio","negative_prompt","seed","storage_uri"], kind: "video" },
  ],
-  serviceKinds: ["llm","imageToText"],
+  serviceKinds: ["llm","imageToText","video"],
+  // Veo via predictLongRunning + fetchPredictOperation (adapter: handlers/videoProviders/vertex.js).
+  // Docs: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/veo-video-generation
+  videoConfig: { baseUrl: "https://aiplatform.googleapis.com" },
 };
--- a/open-sse/providers/registry/xiaomi-mimo.js
+++ b/open-sse/providers/registry/xiaomi-mimo.js
@@ -1,11 +1,19 @@
 import { CLAUDE_API_HEADERS } from "../shared.js";

+// Dual auth (same pattern as kimi):
+//   - API key (sk-...)      → cloud API on api.xiaomimimo.com
+//   - Desktop account/OAuth → same cloud host, plus the Desktop-exclusive Preview
+//     models served by the account-service route on mimo-server-cn.xiaomimimo.com
+//     (authorized by a Xiaomi account session cookie, not the key).
+// Endpoint is picked per model in the executor, same as opencode-go's /responses split.
 export default {
  id: "xiaomi-mimo",
  priority: 290,
  alias: "xiaomi-mimo",
  aliases: [
    "mimo",
+    "mimo-desktop",
+    "xmd",
  ],
  uiAlias: "mimo",
  display: {
@@ -16,9 +24,12 @@ export default {
    website: "https://xiaomimimo.com",
    notice: {
      apiKeyUrl: "https://platform.xiaomimimo.com/console/api-keys",
+      signupUrl: "https://mimo.xiaomimimo.com/desktop/invite/",
    },
  },
-  category: "apikey",
+  category: "oauth",
+  authModes: ["oauth", "apikey"],
+  hasOAuth: true,
  serviceKinds: ["llm", "tts"],
  transport: {
    baseUrl: "https://api.xiaomimimo.com/v1/chat/completions",
@@ -39,6 +50,11 @@ export default {
    },
  ],
  models: [
+    // Desktop-exclusive — served by the account-service route, which only accepts
+    // OpenAI format, so supportedFormats pins them to the openai transport.
+    { id: "mimo-x-pro-preview", name: "MiMo-X-Pro-Preview", upstreamModelId: "xiaomi/mimo-x-pro-preview", supportedFormats: ["openai"] },
+    { id: "mimo-x-flash-preview", name: "MiMo-X-Flash-Preview", upstreamModelId: "xiaomi/mimo-x-flash-preview", supportedFormats: ["openai"] },
+    // Cloud API models (api.xiaomimimo.com/v1)
    { id: "mimo-v2.5-pro", name: "MiMo V2.5 Pro" },
    { id: "mimo-v2.5", name: "MiMo V2.5" },
    { id: "mimo-v2-omni", name: "MiMo V2 Omni" },
@@ -51,4 +67,18 @@ export default {
    authHeader: "bearer",
    format: "xiaomi-mimo-tts",
  },
+  features: {
+    usage: true,
+    usageApikey: true,
+  },
+  // Custom OAuth — non-standard ECDH encrypted-callback flow.
+  // Handled by the Xiaomi MiMo OAuth service, not the generic PKCE pipeline.
+  oauth: {
+    custom: true,
+    authorizeUrl: "https://platform.xiaomimimo.com/authorize",
+    // The callback carries ?u=<ECDH-encrypted payload> instead of ?code=.
+    // Decryption yields { uid, sk, url }.
+    callbackParam: "u",
+    kn: "mimocode",
+  },
 };
--- a/open-sse/rtk/systemInject.js
+++ b/open-sse/rtk/systemInject.js
@@ -13,7 +13,7 @@ export function injectSystemPrompt(body, format, prompt) {
    if (!body || !prompt) return;
    if (typeof body !== "object") return;

-    // Kiro wire shape is unique (conversationState/systemPrompt) — handle directly.
+    // Kiro wire shape is unique (conversationState) — handle directly.
    if (isKiroBody(body) || format === FORMATS.KIRO) {
      injectKiroSystem(body, prompt);
      return;
@@ -61,10 +61,13 @@ export function injectSystemPrompt(body, format, prompt) {

 function isKiroBody(body) {
  if (!body || typeof body !== "object") return false;
-  if (typeof body.systemPrompt !== "string") return false;
  const cs = body.conversationState;
  if (!cs || typeof cs !== "object") return false;
-  return Array.isArray(cs.history) || !!(cs.currentMessage && typeof cs.currentMessage === "object");
+  // A top-level `systemPrompt` used to be the marker, but the Kiro translator no
+  // longer emits it (kiro.dev rejects the field), so gate on the turn shape.
+  const historyTurn = Array.isArray(cs.history)
+    && cs.history.some(it => it && (it.userInputMessage || it.assistantResponseMessage));
+  return historyTurn || !!(cs.currentMessage && cs.currentMessage.userInputMessage);
 }

 // Exact idempotency: prompt present as its own SEP-delimited segment (or the
@@ -258,80 +261,33 @@ function injectGeminiSystem(body, prompt) {
 }

 // ---- Kiro ----
-// Updates top-level systemPrompt and only the mirrored leading prefix of the
-// first user history turn, else current user. next = old + SEP + prompt.
-// Replace old leading prefix only; preserve time context and user tail.
+// The prompt is appended to the first user turn's content — the same place the
+// Kiro translator already mirrors the system text via its contentPrefix.
+//
+// A top-level `systemPrompt` is deliberately NOT written: the kiro.dev gateway
+// answers any body carrying that field with
+//   400 {"message":"Improperly formed request.","reason":"REQUEST_BODY_INVALID"}
+// The translator stopped emitting it in v0.5.59, but this injector kept adding
+// it back, so every kr/ model failed whenever an RTK prompt (caveman, ponytail)
+// was active.
 function injectKiroSystem(body, prompt) {
  try {
-    let oldPrompt = typeof body.systemPrompt === "string" ? body.systemPrompt : "";
-    // Repair path: a previous partial write left systemPrompt updated but user
-    // content still mirroring the pre-write prefix. Re-derive the effective old
-    // prefix from content so this pass converges instead of early-returning.
-    const cs0 = body.conversationState;
-    let firstUser0 = cs0 && Array.isArray(cs0.history)
-      ? (cs0.history.find(it => it && it.userInputMessage)?.userInputMessage ?? null)
-      : null;
-    if (!firstUser0 && cs0?.currentMessage?.userInputMessage) firstUser0 = cs0.currentMessage.userInputMessage;
-
-    if (firstUser0 && typeof firstUser0.content === "string" && oldPrompt && !hasPrompt(oldPrompt, prompt)) {
-      const c0 = firstUser0.content;
-      if (c0 === oldPrompt || (c0.startsWith(oldPrompt) && !c0.startsWith(`${oldPrompt}${SEP}`))) {
-        // systemPrompt advanced past mirrored prefix → stale; treat as un-mirrored
-        oldPrompt = "";
-      }
-    }
-    if (oldPrompt && hasPrompt(oldPrompt, prompt)) return;
-    const next = oldPrompt ? `${oldPrompt}${SEP}${prompt}` : prompt;
-
-    // Atomicity: write user content first, then systemPrompt only if content
-    // write succeeded (or was a no-op). If systemPrompt write then fails, the
-    // repair heuristic above re-derives from content on retry — no permanent
-    // half-applied state.
    const cs = body.conversationState;
    let targetMsg = null;
-    try {
-      const hist = Array.isArray(cs?.history) ? cs.history : null;
-      if (hist) {
-        for (const item of hist) {
-          if (item && item.userInputMessage) { targetMsg = item.userInputMessage; break; }
-        }
-      }
-      if (!targetMsg && cs?.currentMessage?.userInputMessage) {
-        targetMsg = cs.currentMessage.userInputMessage;
-      }
-    } catch (_) { targetMsg = null; }
-
-    let sysWritten = false;
-    try { body.systemPrompt = next; sysWritten = true; } catch (_) {}
-
-    const applyContent = () => {
-      const content = typeof targetMsg.content === "string" ? targetMsg.content : "";
-      if (oldPrompt === "") {
-        // Empty old prompt: prepend unless already at head (exact, not substring)
-        if (content.startsWith(prompt) || content.startsWith(next)) return;
-        const newContent = content ? `${next}${SEP}${content}` : next;
-        try { targetMsg.content = newContent; } catch (_) {}
-        return;
-      }
-      if (!content.startsWith(oldPrompt)) return; // not mirrored at head — leave alone
-      if (content.startsWith(next)) return; // already applied → idempotent
-      const tail = content.slice(oldPrompt.length);
-      try { targetMsg.content = `${next}${tail}`; } catch (_) {}
-    };
-
-    try {
-      if (targetMsg) applyContent();
-    } catch (_) {}
-    if (sysWritten && targetMsg) {
-      // verify convergence: content should now start with next (or be un-mirrored)
-      let ok = false;
-      try {
-        const c = targetMsg.content;
-        ok = typeof c !== "string" || c.startsWith(next) || !c.startsWith(oldPrompt);
-      } catch (_) {}
-      if (!ok) {
-        try { body.systemPrompt = oldPrompt; } catch (_) {} // rollback
+    const hist = Array.isArray(cs?.history) ? cs.history : null;
+    if (hist) {
+      for (const item of hist) {
+        if (item && item.userInputMessage) { targetMsg = item.userInputMessage; break; }
      }
    }
+    if (!targetMsg && cs?.currentMessage?.userInputMessage) {
+      targetMsg = cs.currentMessage.userInputMessage;
+    }
+    if (!targetMsg) return;
+
+    const content = typeof targetMsg.content === "string" ? targetMsg.content : "";
+    const next = dedupStringAppend(content, prompt);
+    if (next === content) return; // already injected — idempotent across retries
+    try { targetMsg.content = next; } catch (_) { /* frozen/proxy fail-open */ }
  } catch (_) {}
 }
--- a/open-sse/services/clinepassModels.js
+++ b/open-sse/services/clinepassModels.js
@@ -19,12 +19,10 @@ function buildModelListHeaders(token, isApiKey) {
 }

 /**
- * Fetch ClinePass live model catalog from Cline's /models endpoint.
- *
- * @param {object} credentials - Connection credentials ({ accessToken, apiKey })
- * @returns {Promise<{ models: { id: string, name: string }[] } | null>}
+ * Internal: fetch the raw model list from Cline's /models endpoint.
+ * Returns the parsed array or null on any failure.
 */
-export async function resolveClinepassModels(credentials) {
+async function fetchClineRawModels(credentials) {
  const isApiKey = Boolean(credentials?.apiKey);
  const token = isApiKey ? credentials.apiKey : credentials?.accessToken;
  if (!token) return null;
@@ -45,19 +43,53 @@ export async function resolveClinepassModels(credentials) {

    const json = await response.json();
    const rawList = Array.isArray(json) ? json : json?.data;
-    if (!Array.isArray(rawList)) return null;
-
-    const models = rawList
-      .filter((m) => typeof m?.id === "string" && m.id.startsWith("cline-pass/"))
-      .map((m) => ({
-        id: m.id,
-        name: m.name || m.id,
-      }));
-
-    return models.length ? { models } : null;
+    return Array.isArray(rawList) ? rawList : null;
  } catch {
    return null;
  } finally {
    clearTimeout(timer);
  }
 }
+
+/**
+ * Fetch ClinePass live model catalog from Cline's /models endpoint.
+ * Returns only models with the cline-pass/ prefix.
+ *
+ * @param {object} credentials - Connection credentials ({ accessToken, apiKey })
+ * @returns {Promise<{ models: { id: string, name: string }[] } | null>}
+ */
+export async function resolveClinepassModels(credentials) {
+  const rawList = await fetchClineRawModels(credentials);
+  if (!rawList) return null;
+
+  const models = rawList
+    .filter((m) => typeof m?.id === "string" && m.id.startsWith("cline-pass/"))
+    .map((m) => ({
+      id: m.id,
+      name: m.name || m.id,
+    }));
+
+  return models.length ? { models } : null;
+}
+
+/**
+ * Fetch Cline live model catalog from Cline's /models endpoint.
+ * Unlike resolveClinepassModels, this returns ALL models (including
+ * free-tier models like z-ai/glm-5.3-flash) without the cline-pass/ prefix filter.
+ *
+ * @param {object} credentials - Connection credentials ({ accessToken, apiKey })
+ * @returns {Promise<{ models: { id: string, name: string }[] } | null>}
+ */
+export async function resolveClineModels(credentials) {
+  const rawList = await fetchClineRawModels(credentials);
+  if (!rawList) return null;
+
+  const models = rawList
+    .filter((m) => typeof m?.id === "string" && m.id.trim() !== "")
+    .map((m) => ({
+      id: m.id,
+      name: m.name || m.id,
+    }));
+
+  return models.length ? { models } : null;
+}
--- a/open-sse/services/qoderModels.js
+++ b/open-sse/services/qoderModels.js
@@ -343,6 +343,30 @@ export async function resolveQoderModels(credentials, options = {}) {
  }
 }

+/**
+ * Every model key the chat endpoint accepts for this credential: the IDE-visible
+ * models first, then catalog entries flagged `enable:false` (hidden in the IDE
+ * picker, e.g. by an account policy, but still served by agent_chat_generation —
+ * see fetchQoderCatalogRaw). /v1/models uses this so the advertised list matches
+ * what the router will actually route instead of collapsing to one or two keys.
+ */
+export function routableQoderModels(catalog) {
+  if (!catalog) return [];
+  const out = [];
+  const seen = new Set();
+  for (const m of catalog.models || []) {
+    if (!m?.id || seen.has(m.id)) continue;
+    seen.add(m.id);
+    out.push({ id: m.id, name: m.name || m.id, hidden: false });
+  }
+  for (const [key, cfg] of catalog.rawConfigs || []) {
+    if (!key || seen.has(key)) continue;
+    seen.add(key);
+    out.push({ id: key, name: cfg?.display_name || key, hidden: true });
+  }
+  return out;
+}
+
 export function invalidateQoderCatalog(credentials) {
  if (!credentials) return;
  catalogCache.delete(cacheKey(credentials));
--- a/open-sse/services/tokenRefresh.js
+++ b/open-sse/services/tokenRefresh.js
@@ -148,6 +148,8 @@ const REFRESH_HANDLERS = {
  "codebuddy-intl": (c, log) => refreshCodebuddyIntlToken(c.refreshToken, log),
  trae: (c, log) => refreshTraeToken(c.refreshToken, c, log),
  cline: (c, log) => refreshClineToken(c.refreshToken, log),
+  // ClinePass shares Cline's WorkOS auth endpoints, so the same refresh works.
+  clinepass: (c, log) => refreshClineToken(c.refreshToken, log),
  zed: () => refreshZedToken(),
  windsurf: (c, log) => refreshWindsurfToken(c, log),
  // Kimi Code OAuth (merged into id `kimi`); legacy id still routes here
--- a/open-sse/services/usage.js
+++ b/open-sse/services/usage.js
@@ -19,6 +19,7 @@ import { getCommandCodeUsage } from "./usage/commandcode.js";
 import { getOpenCodeGoUsage } from "./usage/opencode-go.js";
 import { getGroqUsage } from "./usage/groq.js";
 import { getZedUsage } from "./usage/zed.js";
+import { getXiaomiMimoUsage } from "./usage/xiaomi-mimo.js";
 import { resolveQoderCredentials } from "./qoderModels.js";
 import { getGlmUsage } from "./usage/glm.js";
 import {
@@ -64,6 +65,7 @@ const USAGE_HANDLERS = {
  commandcode: (c) => getCommandCodeUsage(c.apiKey, c.proxyOptions),
  groq: (c) => getGroqUsage(c.apiKey, c.proxyOptions),
  zed: (c) => getZedUsage(c.accessToken, c.providerSpecificData, c.proxyOptions),
+  "xiaomi-mimo": (c) => getXiaomiMimoUsage(c.accessToken, c.providerSpecificData, c.proxyOptions),
 };

 export async function getUsageForProvider(connection, proxyOptions = null, options = {}) {
--- a/open-sse/services/usage/antigravity-weekly.js
+++ b/open-sse/services/usage/antigravity-weekly.js
@@ -0,0 +1,150 @@
+/**
+ * Antigravity weekly quota — best-effort retrieval from retrieveUserQuotaSummary.
+ * Failure never breaks existing per-model quota display.
+ */
+
+import { U, parseResetTime, fetchWithTimeout } from "./shared.js";
+import { ANTIGRAVITY_IDE_USER_AGENT, ANTIGRAVITY_IDE_VERSION } from "../../providers/shared.js";
+
+// — Weekly quota summary config ——————————————————————————————
+const WEEKLY_CONFIG = {
+  ...U("antigravity"),
+  userAgent: ANTIGRAVITY_IDE_USER_AGENT,
+};
+
+// — Cache: TTL + in-flight dedup per project ———————————————
+const WEEKLY_CACHE_TTL_MS = 180_000; // 3 minutes
+const weeklyCache = new Map(); // cacheKey -> { result, expiresAt } | { promise }
+
+function cacheKey(accessToken, projectId) {
+  return `${accessToken}::${projectId || ""}`;
+}
+
+// Exported for tests only
+export function _clearWeeklyCache() {
+  weeklyCache.clear();
+}
+
+// — Group-name to stable key mapping ——————————————————————
+const GROUP_MATCHERS = [
+  { pattern: /gemini/i, key: "gemini_weekly", displayName: "Gemini (Weekly)" },
+  { pattern: /claude|gpt/i, key: "claude_gpt_weekly", displayName: "Claude & GPT (Weekly)" },
+];
+
+/**
+ * Parse a retrieveUserQuotaSummary response into normalized weekly quotas.
+ * Pure function — safe to unit-test without network.
+ *
+ * @param {Object|null} data  Raw JSON response
+ * @returns {Object}  e.g. { gemini_weekly: { used, total, ... }, claude_gpt_weekly: { ... } }
+ */
+export function parseWeeklyQuotaSummary(data) {
+  if (!data || typeof data !== "object") return {};
+
+  // Groups may live at data.groups or data.quotaSummary.groups
+  const groups = Array.isArray(data.groups)
+    ? data.groups
+    : Array.isArray(data.quotaSummary?.groups)
+      ? data.quotaSummary.groups
+      : null;
+
+  if (!groups) return {};
+
+  const result = {};
+
+  for (const group of groups) {
+    if (!group || typeof group !== "object") continue;
+    const displayName = group.displayName || "";
+
+    const buckets = Array.isArray(group.buckets) ? group.buckets : [];
+    for (const bucket of buckets) {
+      if (!bucket || typeof bucket !== "object") continue;
+
+      // Identify weekly buckets by checking bucketId + displayName for "weekly"
+      const bucketText = `${bucket.bucketId || ""} ${bucket.displayName || ""}`.toLowerCase();
+      if (!bucketText.includes("weekly")) continue;
+
+      // Skip disabled buckets
+      if (bucket.disabled === true) continue;
+
+      const remainingFraction = Number(bucket.remainingFraction);
+      if (!Number.isFinite(remainingFraction)) continue;
+
+      // Match group to a known family
+      for (const matcher of GROUP_MATCHERS) {
+        if (matcher.pattern.test(displayName)) {
+          const total = 1000;
+          const remaining = Math.round(total * remainingFraction);
+          const used = Math.max(0, total - remaining);
+
+          result[matcher.key] = {
+            used,
+            total,
+            resetAt: parseResetTime(bucket.resetTime),
+            remainingPercentage: remainingFraction * 100,
+            unlimited: false,
+            displayName: matcher.displayName,
+          };
+          break; // first matching bucket per family wins
+        }
+      }
+    }
+  }
+
+  return result;
+}
+
+/**
+ * Fetch weekly quota summary — cached, deduped, never throws.
+ */
+export async function fetchAntigravityWeeklyQuota(accessToken, projectId, proxyOptions = null) {
+  const key = cacheKey(accessToken, projectId);
+
+  // Serve in-flight or cached
+  const hit = weeklyCache.get(key);
+  if (hit?.promise) return hit.promise;
+  if (hit && hit.expiresAt > Date.now()) return hit.result;
+
+  const promise = (async () => {
+    try {
+      const url = WEEKLY_CONFIG.quotaSummaryApiUrl;
+      if (!url) return {};
+
+      const response = await fetchWithTimeout(url, {
+        method: "POST",
+        headers: {
+          "Authorization": `Bearer ${accessToken}`,
+          "User-Agent": WEEKLY_CONFIG.userAgent,
+          "Content-Type": "application/json",
+          "X-Client-Name": "antigravity",
+          "X-Client-Version": ANTIGRAVITY_IDE_VERSION,
+        },
+        body: JSON.stringify({
+          ...(projectId ? { project: projectId } : {}),
+        }),
+      }, 10000, proxyOptions);
+
+      if (!response.ok) return {};
+
+      const data = await response.json();
+      return parseWeeklyQuotaSummary(data);
+    } catch {
+      return {};
+    }
+  })();
+
+  weeklyCache.set(key, { promise });
+
+  try {
+    const result = await promise;
+    if (result && Object.keys(result).length > 0) {
+      weeklyCache.set(key, { result, expiresAt: Date.now() + WEEKLY_CACHE_TTL_MS });
+    } else {
+      weeklyCache.delete(key);
+    }
+    return result;
+  } catch {
+    weeklyCache.delete(key);
+    return {};
+  }
+}
--- a/open-sse/services/usage/claude.js
+++ b/open-sse/services/usage/claude.js
@@ -102,32 +102,28 @@ async function fetchClaudeUsageRaw(accessToken, proxyOptions = null) {
        quotas["weekly (7d)"] = createQuotaObject(data.seven_day);
      }

-      // Parse model-specific weekly windows (e.g. seven_day_sonnet, seven_day_opus, seven_day_fable)
-      const MODEL_DISPLAY_NAMES = {
-        fable_5_1: "fable",
-        fable_5: "fable",
-      };
-
+      // Parse model-specific weekly windows (e.g. seven_day_sonnet, seven_day_opus)
      for (const [key, value] of Object.entries(data)) {
        if (key.startsWith("seven_day_") && key !== "seven_day" && hasUtilization(value)) {
-          const rawName = key.replace("seven_day_", "");
-          const modelName = MODEL_DISPLAY_NAMES[rawName] || rawName;
+          const modelName = key.replace("seven_day_", "");
          quotas[`weekly ${modelName} (7d)`] = createQuotaObject(value);
-        } else if ((key === "fable" || key === "fable_5" || key === "fable_5_1") && hasUtilization(value)) {
-          quotas["weekly fable (7d)"] = createQuotaObject(value);
        }
      }

-      // Fallback: surface Fable quota row if weekly window exists but Fable was not returned yet
-      if (!quotas["weekly fable (7d)"] && hasUtilization(data.seven_day)) {
-        quotas["weekly fable (7d)"] = {
-          used: 0,
-          total: 100,
-          remaining: 100,
-          remainingPercentage: 100,
-          resetAt: parseResetTime(data.seven_day.resets_at),
-          unlimited: false,
-        };
+      // Model-scoped weekly limits (e.g. Fable) arrive in limits[], not as
+      // seven_day_* keys: { kind: "weekly_scoped", percent, resets_at,
+      // scope: { model: { display_name: "Fable" } } }. No limits entry means
+      // the account has no such window — omit the row, never fabricate one.
+      if (Array.isArray(data.limits)) {
+        for (const limit of data.limits) {
+          if (limit?.kind !== "weekly_scoped") continue;
+          const modelName = String(limit?.scope?.model?.display_name || "").trim().toLowerCase();
+          if (!modelName || typeof limit.percent !== "number") continue;
+          quotas[`weekly ${modelName} (7d)`] = createQuotaObject({
+            utilization: Math.max(0, Math.min(100, limit.percent)),
+            resets_at: limit.resets_at,
+          });
+        }
      }

      return {
--- a/open-sse/services/usage/google.js
+++ b/open-sse/services/usage/google.js
@@ -5,6 +5,7 @@
 import { CLIENT_METADATA } from "../../config/appConstants.js";
 import { ANTIGRAVITY_IDE_USER_AGENT, ANTIGRAVITY_IDE_VERSION, ANTIGRAVITY_OAUTH_CLIENT } from "../../providers/shared.js";
 import { U, parseResetTime, normalizeCloudCodeProjectId, fetchWithTimeout } from "./shared.js";
+import { fetchAntigravityWeeklyQuota } from "./antigravity-weekly.js";

 // Antigravity API config (from Quotio) — urls from registry, oauth client + dynamic UA kept here
 const ANTIGRAVITY_CONFIG = {
@@ -157,8 +158,15 @@ export async function getAntigravityUsage(accessToken, providerSpecificData, pro
    const data = await response.json();
    const quotas = {};

-    // Parse model quotas (inspired by vscode-antigravity-cockpit)
-    if (data.models) {
+    // Detect tier: free-tier accounts only have weekly quotas (no separate 5h window).
+    // On free-tier, fetchAvailableModels returns misleading per-model quota info
+    // (missing remainingFraction defaults to 0, or reflects the weekly limit not a 5h window).
+    const paidTierId = subscriptionInfo?.paidTier?.id;
+    const isFreeTier = !paidTierId || paidTierId === "free-tier";
+
+    // Parse model quotas only for paid-tier accounts.
+    // Free-tier accounts skip this — their only meaningful quota is the weekly limit.
+    if (!isFreeTier && data.models) {
      // Filter only recommended/important models (must match PROVIDER_MODELS ag ids)
      const importantModels = [
        'gemini-3.8-flash-high',
@@ -212,6 +220,56 @@ export async function getAntigravityUsage(accessToken, providerSpecificData, pro
      }
    }

+    // Best-effort weekly quota overlay — never blocks or breaks per-model results
+    try {
+      const weeklyQuotas = await fetchAntigravityWeeklyQuota(
+        accessToken,
+        projectId,
+        proxyOptions
+      );
+
+      // Reconcile weekly quota against model family status:
+      // If every model in a family is locked/exhausted (remainingPercentage === 0)
+      // until a future reset time, the weekly limit cannot be 100% available.
+      // On Google's Free Starter tier, retrieveUserQuotaSummary buggily reports
+      // remainingFraction: 1 even after the starter quota is depleted and all models 429.
+      const entries = Object.entries(quotas);
+      const geminiModels = entries.filter(([k]) => k.startsWith("gemini-") && !k.includes("image"));
+      const claudeModels = entries.filter(([k]) => k.startsWith("claude-"));
+
+      if (weeklyQuotas.gemini_weekly && geminiModels.length > 0) {
+        const allGeminiExhausted = geminiModels.every(([, q]) => (q.remainingPercentage ?? 0) === 0);
+        if (allGeminiExhausted && weeklyQuotas.gemini_weekly.remainingPercentage > 0) {
+          const maxResetAt = geminiModels.reduce((max, [, q]) =>
+            !max || (q.resetAt && new Date(q.resetAt) > new Date(max)) ? q.resetAt : max, null
+          );
+          weeklyQuotas.gemini_weekly.used = weeklyQuotas.gemini_weekly.total;
+          weeklyQuotas.gemini_weekly.remainingPercentage = 0;
+          if (maxResetAt) {
+            weeklyQuotas.gemini_weekly.resetAt = maxResetAt;
+          }
+        }
+      }
+
+      if (weeklyQuotas.claude_gpt_weekly && claudeModels.length > 0) {
+        const allClaudeExhausted = claudeModels.every(([, q]) => (q.remainingPercentage ?? 0) === 0);
+        if (allClaudeExhausted && weeklyQuotas.claude_gpt_weekly.remainingPercentage > 0) {
+          const maxResetAt = claudeModels.reduce((max, [, q]) =>
+            !max || (q.resetAt && new Date(q.resetAt) > new Date(max)) ? q.resetAt : max, null
+          );
+          weeklyQuotas.claude_gpt_weekly.used = weeklyQuotas.claude_gpt_weekly.total;
+          weeklyQuotas.claude_gpt_weekly.remainingPercentage = 0;
+          if (maxResetAt) {
+            weeklyQuotas.claude_gpt_weekly.resetAt = maxResetAt;
+          }
+        }
+      }
+
+      Object.assign(quotas, weeklyQuotas);
+    } catch {
+      // Silently ignore — weekly is best-effort
+    }
+
    return {
      plan: subscriptionInfo?.currentTier?.name || "Unknown",
      quotas,
--- a/open-sse/services/usage/xiaomi-mimo.js
+++ b/open-sse/services/usage/xiaomi-mimo.js
@@ -0,0 +1,125 @@
+/**
+ * Xiaomi MiMo usage — weekly quota from the Xiaomi account session.
+ *
+ * Primary path: GET {mimo-server}/api/user/usage authorized by the account-session
+ * cookie (see shared/mimoAccount.js). Response: { code: 0, data: { percent (remaining
+ * %), resetDate, resetAt } }.
+ *
+ * Fallback: the sk- API key cannot read the quota, so when no account session is
+ * available we surface a graceful message instead of failing.
+ */
+
+import { proxyAwareFetch } from "../../utils/proxyFetch.js";
+import { getMimoAccountUsage } from "../../shared/mimoAccount.js";
+
+const USAGE_URL = "https://aistudio.xiaomimimo.com/open-apis/v1/user/usage";
+
+/**
+ * @param {string|null|undefined} accessToken - sk- API key
+ * @param {object|null} providerSpecificData - may contain mimoPassToken, uid, etc.
+ * @param {object|null} proxyOptions
+ */
+export async function getXiaomiMimoUsage(accessToken = null, providerSpecificData = null, proxyOptions = null) {
+  // Preferred path: the weekly quota comes from the account service session
+  // (mimo-server /api/user/usage), which the sk- key cannot reach. The session is
+  // derived from MiMo Desktop's persisted passToken via the SSO/sts handshake.
+  const account = await getMimoAccountUsage(providerSpecificData, proxyOptions);
+  if (typeof account.percent === "number" && Number.isFinite(account.percent)) {
+    const remaining = Math.max(0, Math.min(100, Math.round(account.percent)));
+    const used = 100 - remaining;
+    let resetAt = null;
+    if (typeof account.resetAt === "number" && account.resetAt > 0) {
+      resetAt = new Date(account.resetAt * 1000).toISOString();
+    } else if (typeof account.resetDate === "string") {
+      const parsed = new Date(`${account.resetDate}T00:00:00Z`);
+      if (!Number.isNaN(parsed.getTime())) resetAt = parsed.toISOString();
+    }
+    return {
+      plan: "Xiaomi MiMo Desktop",
+      quotas: {
+        Weekly: { used, total: 100, remainingPercentage: remaining, resetAt, unlimited: false },
+      },
+    };
+  }
+
+  // Fallback: no account session available (Desktop never logged in, or its cookie
+  // store is locked). The sk- key cannot read the quota, so surface a clear message.
+  const key = accessToken || providerSpecificData?.apiKey;
+  if (!key || typeof key !== "string" || !key.trim()) {
+    return { message: "Xiaomi MiMo Desktop not connected. Add credentials to view usage." };
+  }
+
+  try {
+    const response = await proxyAwareFetch(
+      USAGE_URL,
+      {
+        method: "GET",
+        headers: {
+          Authorization: `Bearer ${key.trim()}`,
+          "X-Mimo-Source": "mimocode-cli",
+          Accept: "application/json",
+        },
+        signal: AbortSignal.timeout(10000),
+      },
+      proxyOptions,
+    );
+
+    if (response.status === 401) {
+      return {
+        plan: "Xiaomi MiMo Desktop",
+        message: "Weekly quota requires Xiaomi account session. API key alone is insufficient.",
+      };
+    }
+
+    if (!response.ok) {
+      return {
+        plan: "Xiaomi MiMo Desktop",
+        message: `Usage API error (${response.status})`,
+      };
+    }
+
+    const data = await response.json().catch(() => null);
+    if (!data || data.code !== 0 || !data.data) {
+      return {
+        plan: "Xiaomi MiMo Desktop",
+        message: "Usage endpoint returned unexpected response.",
+      };
+    }
+
+    const { percent, resetDate } = data.data;
+    if (typeof percent !== "number" || !Number.isFinite(percent)) {
+      return {
+        plan: "Xiaomi MiMo Desktop",
+        message: "Usage data missing percent field.",
+      };
+    }
+
+    // percent = remaining percentage (e.g. 94 means 94% remaining)
+    const remaining = Math.max(0, Math.min(100, Math.round(percent)));
+    const used = 100 - remaining;
+
+    // Parse resetDate — expected format "2026-09-16"
+    let resetAt = null;
+    if (resetDate && typeof resetDate === "string") {
+      const parsed = new Date(`${resetDate}T00:00:00Z`);
+      if (!Number.isNaN(parsed.getTime())) {
+        resetAt = parsed.toISOString();
+      }
+    }
+
+    return {
+      plan: "Xiaomi MiMo Desktop",
+      quotas: {
+        Weekly: {
+          used,
+          total: 100,
+          remainingPercentage: remaining,
+          resetAt,
+          unlimited: false,
+        },
+      },
+    };
+  } catch (error) {
+    return { message: `Xiaomi MiMo Desktop usage error: ${error.message}` };
+  }
+}
--- a/open-sse/shared/clineAuth.js
+++ b/open-sse/shared/clineAuth.js
@@ -6,7 +6,14 @@ export function getClineAccessToken(token) {
  if (typeof token !== "string") return "";
  const trimmed = token.trim();
  if (!trimmed) return "";
-  return trimmed.startsWith("workos:") ? trimmed : `workos:${trimmed}`;
+  if (trimmed.toLowerCase().startsWith("workos:")) return trimmed;
+  // Cline OAuth access tokens are WorkOS JWTs (base64url `eyJ…` header).
+  // ClinePass API keys (category "apikey", e.g. `clp_…`) are NOT JWTs and must
+  // be sent verbatim — prefixing them with `workos:` makes the Cline API reject
+  // the request with HTTP 401 ("Please make sure you're using the latest
+  // version of Cline and re-authenticate your Cline account.").
+  const isWorkOsJwt = /^eyJ[A-Za-z0-9_-]+\.[A-Za-z0-9_-]+/.test(trimmed);
+  return isWorkOsJwt ? `workos:${trimmed}` : trimmed;
 }

 export function getClineAuthorizationHeader(token) {
--- a/open-sse/shared/clineEnvelope.js
+++ b/open-sse/shared/clineEnvelope.js
@@ -0,0 +1,19 @@
+import { PROVIDERS } from "../providers/index.js";
+
+/**
+ * Unwrap Cline's non-stream envelope: {"success":true,"data":{...choices...}}.
+ *
+ * Scoped to providers opting in via `transport.quirks.clineEnvelope` so no other
+ * provider's body is ever rewritten. The error envelope ({"success":false,...})
+ * never matches and passes through untouched.
+ *
+ * @param {object} body - Parsed upstream response body
+ * @param {string} provider - Provider id or alias
+ * @returns {object} The inner `data` object, or `body` unchanged
+ */
+export function unwrapClineEnvelope(body, provider) {
+  if (!provider || !PROVIDERS[provider]?.quirks?.clineEnvelope) return body;
+  const { success, data } = body || {};
+  if (success !== true || !data || typeof data !== "object" || Array.isArray(data)) return body;
+  return data;
+}
--- a/open-sse/shared/mimoAccount.js
+++ b/open-sse/shared/mimoAccount.js
@@ -0,0 +1,264 @@
+import fs from "node:fs";
+import os from "node:os";
+import path from "node:path";
+import crypto from "node:crypto";
+import { proxyAwareFetch } from "../utils/proxyFetch.js";
+
+/**
+ * Xiaomi MiMo account-session helpers (used for weekly quota).
+ *
+ * The weekly quota endpoint lives on the account service domain and is authorized
+ * by an account session cookie, NOT the sk- API key. Acquiring that cookie mirrors
+ * MiMo Desktop: a passToken (persisted in Desktop's cookie store) is exchanged via
+ * the passportapi SSO, then authorized for the `mimopc` service, and finally stamped
+ * by the mimo-server /api/sts callback into a `serviceToken` cookie.
+ *
+ * Flow (verified against MiMo Desktop traffic):
+ *   1. GET  {api}/api/user/xiaomi/me           -> 302 to account SSO (sid=mimopc)
+ *   2. GET  account /pass/serviceLogin?sid=passportapi&_json=true   -> nonce/ssecurity
+ *   3. GET  {location}&clientSign=...          -> account-level serviceToken
+ *   4. GET  account /pass/serviceLogin?sid=mimopc&callback=<sts>&_json=true
+ *   5. GET  {api}/api/sts?...&ticket...        -> Set-Cookie: serviceToken (mimopc scope)
+ */
+
+const API_BASE = "https://mimo-server-cn.xiaomimimo.com";
+const ACCOUNT_HOST = "account.xiaomi.com";
+const API_UA =
+  "miNative PC/Normal Windows_NT/10.0.19045 SDKV/1.0.0 DEVT/PC DEVS/Windows APP/miaccount_desktop APPV/0.1.0";
+const SSO_UA = "MiClaw/1.0";
+const COOKIE_TTL_MS = 30 * 60 * 1000;
+
+// Per-account session caches (keyed by passToken hash) so multiple Xiaomi
+// accounts / connections can rotate without clobbering each other.
+const _cache = new Map(); // key -> { cookie, at }
+const _inflight = new Map(); // key -> Promise<cookie|null>
+
+function desktopCookiePath() {
+  const home = os.homedir();
+  if (process.platform === "win32") {
+    return path.join(home, "AppData", "Roaming", "Xiaomi MiMo", "Partitions", "xiaomi-account", "Network", "Cookies");
+  }
+  if (process.platform === "darwin") {
+    return path.join(home, "Library", "Application Support", "Xiaomi MiMo", "Partitions", "xiaomi-account", "Network", "Cookies");
+  }
+  return path.join(home, ".config", "Xiaomi MiMo", "Partitions", "xiaomi-account", "Network", "Cookies");
+}
+
+/**
+ * Read the persisted Xiaomi account cookies from MiMo Desktop's Electron profile.
+ * The Chromium cookie DB is held with an exclusive lock while Desktop runs, so we
+ * copy it first and bail (return null) if that fails.
+ * @returns {Promise<Record<string,string>|null>}
+ */
+async function readDesktopAccountCookies() {
+  const src = desktopCookiePath();
+  if (!fs.existsSync(src)) return null;
+  const tmp = path.join(os.tmpdir(), `9r-mimo-cookies-${process.pid}-${crypto.randomBytes(4).toString("hex")}.db`);
+  try {
+    fs.copyFileSync(src, tmp);
+  } catch {
+    return null; // locked by a running Desktop
+  }
+  try {
+    const { DatabaseSync } = await import("node:sqlite");
+    const db = new DatabaseSync(tmp, { readOnly: true });
+    const rows = db.prepare("SELECT name, value FROM cookies WHERE host_key = ?").all("." + ACCOUNT_HOST);
+    db.close();
+    const jar = Object.fromEntries(rows.map((r) => [r.name, r.value]));
+    return jar.passToken ? jar : null;
+  } catch {
+    return null;
+  } finally {
+    try {
+      fs.unlinkSync(tmp);
+    } catch {
+      /* ignore */
+    }
+  }
+}
+
+/**
+ * Read just the passToken + identity cookies from Desktop's profile.
+ * Exported so the connect flow can persist a per-account passToken into the
+ * connection's providerSpecificData — this is what enables multi-account rotation.
+ * @returns {Promise<{passToken:string, userId:string|null, cUserId:string|null}|null>}
+ */
+export async function readDesktopPassToken() {
+  try {
+    const jar = await readDesktopAccountCookies();
+    if (!jar?.passToken) return null;
+    return { passToken: jar.passToken, userId: jar.userId || null, cUserId: jar.cUserId || null };
+  } catch {
+    return null;
+  }
+}
+
+function signatureClientSign(nonce, ssecurity) {
+  const input = `nonce=${nonce}` + (ssecurity && ssecurity.trim() ? `&${ssecurity}` : "");
+  return encodeURIComponent(crypto.createHash("sha1").update(input).digest("base64"));
+}
+
+function absorbSetCookie(jar, res) {
+  for (const c of res.headers.getSetCookie?.() || []) {
+    const m = /^([^=]+)=([^;]*)/.exec(c.trim());
+    if (m && m[2]) jar[m[1]] = m[2];
+  }
+}
+
+function cookieHeader(jar) {
+  return Object.entries(jar)
+    .filter(([, v]) => v)
+    .map(([k, v]) => `${k}=${v}`)
+    .join("; ");
+}
+
+/**
+ * Exchange a passToken for a mimo-server service session cookie.
+ * @returns {Promise<string|null>} Cookie header value, or null on failure.
+ */
+async function acquireServiceCookie(passJar, proxyOptions) {
+  const jar = { ...passJar };
+  const ck = () => cookieHeader(jar);
+
+  // 1. Unauthenticated API call -> 302 carrying the sts callback (sid=mimopc)
+  const r1 = await proxyAwareFetch(
+    `${API_BASE}/api/user/xiaomi/me`,
+    { redirect: "manual", headers: { "User-Agent": API_UA, Cookie: ck() } },
+    proxyOptions,
+  );
+  const redirect = r1.headers.get("location");
+  if (!redirect) return null;
+  const stsCallback = new URL(redirect).searchParams.get("callback");
+  if (!stsCallback) return null;
+
+  // 2. passportapi SSO phase 1 -> nonce + ssecurity
+  const sso1 = await proxyAwareFetch(
+    `https://${ACCOUNT_HOST}/pass/serviceLogin?sid=passportapi&_json=true`,
+    { headers: { Cookie: ck(), "User-Agent": SSO_UA, Accept: "application/json" } },
+    proxyOptions,
+  );
+  const j1 = JSON.parse((await sso1.text()).replace(/^&&&START&&&/, ""));
+  const nonce = j1.nonce || (j1.location ? new URL(j1.location).searchParams.get("nonce") : null);
+  if (!nonce || !j1.location) return null;
+
+  // 3. passportapi SSO phase 2 -> account-level serviceToken
+  const sso2 = await proxyAwareFetch(
+    `${j1.location}&clientSign=${signatureClientSign(nonce, j1.ssecurity)}`,
+    { redirect: "manual", headers: { Cookie: ck(), "User-Agent": SSO_UA } },
+    proxyOptions,
+  );
+  absorbSetCookie(jar, sso2);
+
+  // 4. mimopc SSO -> sts callback carrying a ticket
+  const sso3 = await proxyAwareFetch(
+    `https://${ACCOUNT_HOST}/pass/serviceLogin?sid=mimopc&callback=${encodeURIComponent(stsCallback)}&_json=true`,
+    { headers: { Cookie: ck(), "User-Agent": SSO_UA, Accept: "application/json" } },
+    proxyOptions,
+  );
+  const j3 = JSON.parse((await sso3.text()).replace(/^&&&START&&&/, ""));
+  absorbSetCookie(jar, sso3);
+  if (!j3?.location || !/\/api\/sts/.test(j3.location)) return null;
+
+  // 5. sts callback -> Set-Cookie: serviceToken (mimopc scope)
+  const sts = await proxyAwareFetch(
+    j3.location,
+    { redirect: "manual", headers: { "User-Agent": API_UA, Cookie: ck() } },
+    proxyOptions,
+  );
+  absorbSetCookie(jar, sts);
+
+  const needed = ["serviceToken", "mimopc_ph", "mimopc_slh", "userId"];
+  if (!jar.serviceToken) return null;
+  const out = {};
+  for (const k of needed) if (jar[k]) out[k] = jar[k];
+  return cookieHeader(out);
+}
+
+/**
+ * Get (and cache) the mimo-server account cookie.
+ * @param {object|null} providerSpecificData - may carry `mimoPassToken` override
+ */
+async function getServiceCookie(providerSpecificData, proxyOptions) {
+  const passJar = providerSpecificData?.mimoPassToken
+    ? { passToken: providerSpecificData.mimoPassToken, userId: providerSpecificData.mimoUserId, cUserId: providerSpecificData.mimoCUserId }
+    : await readDesktopAccountCookies();
+  if (!passJar) return { cookie: null, reason: "no-pass-token" };
+
+  // One cached session per passToken — accounts/connections rotate independently.
+  const key = crypto.createHash("sha256").update(passJar.passToken).digest("hex");
+
+  const cached = _cache.get(key);
+  if (cached && Date.now() - cached.at < COOKIE_TTL_MS) {
+    return { cookie: cached.cookie };
+  }
+
+  // De-dupe concurrent handshakes for the same account: a burst of requests must
+  // not each run the full 5-step SSO chain.
+  const inflight = _inflight.get(key);
+  if (inflight) {
+    const cookie = await inflight;
+    return cookie ? { cookie } : { cookie: null, reason: "sso-failed" };
+  }
+
+  const promise = (async () => {
+    try {
+      return await acquireServiceCookie(passJar, proxyOptions);
+    } catch {
+      return null; // network/parse failure — callers degrade, never throw
+    } finally {
+      _inflight.delete(key);
+    }
+  })();
+  _inflight.set(key, promise);
+
+  const cookie = await promise;
+  if (!cookie) return { cookie: null, reason: "sso-failed" };
+  _cache.set(key, { cookie, at: Date.now() });
+  return { cookie };
+}
+
+/** Drop cached sessions so the next call re-runs the handshake (e.g. after a 401). */
+export function invalidateMimoAccountCookieCache() {
+  _cache.clear();
+}
+
+/** mimo-server account API base + the User-Agent its backend expects. */
+export const MIMO_API_BASE = API_BASE;
+export const MIMO_API_UA = API_UA;
+
+/**
+ * Resolve the mimo-server account-session cookie, for upstream /api/route/* calls.
+ * @returns {Promise<string|null>} Cookie header value, or null when unavailable.
+ */
+export async function getMimoAccountCookie(providerSpecificData = null, proxyOptions = null) {
+  try {
+    const { cookie } = await getServiceCookie(providerSpecificData, proxyOptions);
+    return cookie;
+  } catch {
+    return null;
+  }
+}
+
+/**
+ * Fetch the weekly quota from the account service.
+ * @returns {Promise<{percent?:number, resetDate?:string, resetAt?:number, error?:string}>}
+ */
+export async function getMimoAccountUsage(providerSpecificData = null, proxyOptions = null) {
+  const { cookie, reason } = await getServiceCookie(providerSpecificData, proxyOptions);
+  if (!cookie) {
+    return { error: reason === "no-pass-token" ? "no-session" : "session-failed" };
+  }
+  try {
+    const res = await proxyAwareFetch(
+      `${API_BASE}/api/user/usage`,
+      { headers: { "User-Agent": API_UA, Cookie: cookie, Accept: "application/json" }, signal: AbortSignal.timeout(10000) },
+      proxyOptions,
+    );
+    if (!res.ok) return { error: `http-${res.status}` };
+    const data = await res.json().catch(() => null);
+    if (!data || data.code !== 0 || !data.data) return { error: "bad-response" };
+    return { percent: data.data.percent, resetDate: data.data.resetDate, resetAt: data.data.resetAt };
+  } catch (e) {
+    return { error: e.message };
+  }
+}
--- a/open-sse/shared/qoder/attachments.js
+++ b/open-sse/shared/qoder/attachments.js
@@ -0,0 +1,341 @@
+/**
+ * Native qodercli does NOT stuff image/PDF bytes into agent_chat_generation.
+ * It PUTs them to /algo/api/v2/image/upload (COSY-signed multipart) and then
+ * sends the returned OSS URL. Agents like Claude Code send OpenAI/Claude
+ * data-URIs instead, which 9router previously forwarded verbatim — 10MB
+ * images become 30MB+ JSON and upstream 413s even though the model window
+ * is ~200k tokens.
+ *
+ * This module:
+ *   1. Uploads inlined images to Qoder's file API (cached by sha256).
+ *   2. Replaces huge non-image file blocks with a short stub.
+ *   3. Caps leftover data-URIs so the chat JSON stays small.
+ */
+
+import { createHash } from "crypto";
+import { v4 as uuidv4 } from "uuid";
+
+import { proxyAwareFetch } from "../../utils/proxyFetch.js";
+import { parseDataUri } from "../../translator/concerns/image.js";
+import { OPENAI_BLOCK, CLAUDE_BLOCK } from "../../translator/schema/blocks.js";
+import { MAX_IMAGE_BYTES } from "../../config/mediaConfig.js";
+import { buildCosyHeaders } from "./cosy.js";
+import {
+  QODER_IMAGE_UPLOAD_SIG_PATH,
+  QODER_INLINE_FALLBACK_MAX_BYTES,
+  QODER_MAX_PAYLOAD_BYTES,
+  qoderInferenceBase,
+} from "./constants.js";
+
+const IMAGE_MIME_RE = /^image\//i;
+const DATA_URI_RE = /data:[^;]+;base64,[A-Za-z0-9+/=\s]+/g;
+
+function mimeExt(mime) {
+  const m = String(mime || "").toLowerCase();
+  if (m.includes("png")) return "png";
+  if (m.includes("jpeg") || m.includes("jpg")) return "jpg";
+  if (m.includes("gif")) return "gif";
+  if (m.includes("webp")) return "webp";
+  if (m.includes("bmp")) return "bmp";
+  if (m.includes("pdf")) return "pdf";
+  return "bin";
+}
+
+function decodedBytes(b64) {
+  if (typeof b64 !== "string" || !b64) return 0;
+  const compact = b64.replace(/\s/g, "");
+  return Math.floor(compact.length * 3 / 4);
+}
+
+function stubText({ name, mime, bytes, reason }) {
+  const label = name || mime || "attachment";
+  const size = bytes ? `, ${bytes} bytes` : "";
+  return `[file omitted: ${label}${size} — ${reason}]`;
+}
+
+export function buildMultipartFile(buffer, { fieldName = "file", fileName, mediaType } = {}) {
+  const boundary = `----9routerQoder${Date.now().toString(16)}${Math.random().toString(16).slice(2)}`;
+  const filename = fileName || `upload.${mimeExt(mediaType)}`;
+  const head = Buffer.from(
+    `--${boundary}\r\nContent-Disposition: form-data; name="${fieldName}"; filename="${filename}"\r\nContent-Type: ${mediaType || "application/octet-stream"}\r\n\r\n`,
+  );
+  const tail = Buffer.from(`\r\n--${boundary}--\r\n`);
+  const body = Buffer.concat([head, buffer, tail]);
+  return { boundary, body };
+}
+
+function extractUrlFromUploadResponse(json) {
+  if (!json || typeof json !== "object") return null;
+  const result = json.result && typeof json.result === "object" ? json.result : json;
+  const arrays = [result.imageUrls, result.image_urls, json.imageUrls, json.image_urls];
+  for (const arr of arrays) {
+    if (Array.isArray(arr) && typeof arr[0] === "string" && arr[0]) return arr[0];
+  }
+  const keys = ["imageUrl", "image_url", "url", "ossUrl", "oss_url", "originalUrl", "originUrl", "link", "image"];
+  for (const key of keys) {
+    const v = result[key] ?? json[key];
+    if (typeof v === "string" && v) return v;
+  }
+  if (typeof json.body === "string") {
+    try { return extractUrlFromUploadResponse(JSON.parse(json.body)); } catch { /* ignore */ }
+  }
+  return null;
+}
+
+async function defaultUploadImage({ buffer, mediaType, credentials, proxyOptions, signal }) {
+  const requestId = uuidv4();
+  const url = `${qoderInferenceBase(credentials)}${`/algo${QODER_IMAGE_UPLOAD_SIG_PATH}`}?request_id=${requestId}`;
+  const { boundary, body } = buildMultipartFile(buffer, {
+    fileName: `image.${mimeExt(mediaType)}`,
+    mediaType: mediaType || "application/octet-stream",
+  });
+  const psd = credentials?.providerSpecificData || {};
+  const cosyHeaders = buildCosyHeaders(body, url, {
+    userId: psd.userId,
+    authToken: credentials.accessToken,
+    name: credentials.displayName || "",
+    email: credentials.email || "",
+    machineId: psd.machineId || "",
+  });
+  const headers = {
+    ...cosyHeaders,
+    Accept: "application/json",
+    "Content-Type": `multipart/form-data; boundary=${boundary}`,
+    "Content-Length": String(body.length),
+    "AI-CLIENT-TIMESTAMP": String(Math.floor(Date.now() / 1000)),
+    "Accept-Encoding": "identity",
+  };
+  const res = await proxyAwareFetch(
+    url,
+    { method: "PUT", headers, body, signal },
+    proxyOptions,
+  );
+  if (!res.ok) {
+    const text = await res.text().catch(() => "");
+    throw new Error(`HTTP ${res.status}${text ? `: ${text.slice(0, 180)}` : ""}`);
+  }
+  const json = await res.json().catch(() => null);
+  const uploaded = extractUrlFromUploadResponse(json);
+  if (!uploaded) throw new Error("upload response missing url");
+  return uploaded;
+}
+
+async function uploadImageData({ base64, mediaType, credentials, proxyOptions, signal, log, uploadFn, cache }) {
+  const compact = String(base64 || "").replace(/\s/g, "");
+  if (!compact) return null;
+  const bytes = decodedBytes(compact);
+  if (bytes > MAX_IMAGE_BYTES) {
+    log?.warn?.("QODER", `image ${bytes} bytes exceeds upload cap, stubbing`);
+    return { stub: true, bytes, mime: mediaType };
+  }
+  let buffer;
+  try {
+    buffer = Buffer.from(compact, "base64");
+  } catch {
+    return { stub: true, bytes, mime: mediaType };
+  }
+  const digest = createHash("sha256").update(buffer).digest("hex");
+  if (cache?.has(digest)) return { url: cache.get(digest), bytes, mime: mediaType };
+
+  const doUpload = uploadFn || defaultUploadImage;
+  try {
+    const url = await doUpload({ buffer, mediaType, credentials, proxyOptions, signal });
+    if (typeof url === "string" && url) {
+      cache?.set(digest, url);
+      return { url, bytes, mime: mediaType };
+    }
+  } catch (err) {
+    log?.warn?.("QODER", `image upload failed (${err.message}); ${bytes <= QODER_INLINE_FALLBACK_MAX_BYTES ? "keeping inline" : "stubbing"}`);
+  }
+  if (bytes <= QODER_INLINE_FALLBACK_MAX_BYTES) return { keep: true, bytes, mime: mediaType };
+  return { stub: true, bytes, mime: mediaType };
+}
+
+function imageUrlBlock(url) {
+  return { type: OPENAI_BLOCK.IMAGE_URL, image_url: { url } };
+}
+
+async function rewriteBlock(block, ctx) {
+  if (!block || typeof block !== "object") return block;
+
+  if (block.type === OPENAI_BLOCK.IMAGE_URL) {
+    const raw = typeof block.image_url === "string" ? block.image_url : block.image_url?.url;
+    if (typeof raw !== "string" || !raw) return null;
+    if (raw.startsWith("http://") || raw.startsWith("https://")) return imageUrlBlock(raw);
+    const parsed = parseDataUri(raw);
+    if (!parsed) return { type: OPENAI_BLOCK.TEXT, text: stubText({ name: "attachment", reason: "unreadable data URI" }) };
+    if (!IMAGE_MIME_RE.test(parsed.mimeType)) {
+      return { type: OPENAI_BLOCK.TEXT, text: stubText({ name: "file", mime: parsed.mimeType, bytes: decodedBytes(parsed.base64), reason: "non-image bytes are not inlined into Qoder context" }) };
+    }
+    const up = await uploadImageData({ ...ctx, base64: parsed.base64, mediaType: parsed.mimeType });
+    if (up?.url) return imageUrlBlock(up.url);
+    if (up?.keep) return imageUrlBlock(raw);
+    return { type: OPENAI_BLOCK.TEXT, text: stubText({ name: "image", mime: parsed.mimeType, bytes: up?.bytes, reason: "upload failed; not inlined" }) };
+  }
+
+  if (block.type === OPENAI_BLOCK.IMAGE || block.type === CLAUDE_BLOCK.IMAGE) {
+    const src = block.source || {};
+    if (src.type === "url" && typeof src.url === "string") return imageUrlBlock(src.url);
+    if (src.type === "base64" && src.data) {
+      const mime = src.media_type || "image/png";
+      const up = await uploadImageData({ ...ctx, base64: src.data, mediaType: mime });
+      if (up?.url) return imageUrlBlock(up.url);
+      if (up?.keep) return imageUrlBlock(`data:${mime};base64,${src.data}`);
+      return { type: OPENAI_BLOCK.TEXT, text: stubText({ name: "image", mime, bytes: up?.bytes, reason: "upload failed; not inlined" }) };
+    }
+  }
+
+  if (block.type === OPENAI_BLOCK.FILE && block.file) {
+    const file = block.file;
+    const name = file.filename || file.name || "file";
+    const dataUri = typeof file.file_data === "string" ? file.file_data : null;
+    const parsed = dataUri ? parseDataUri(dataUri) : null;
+    const b64 = parsed?.base64 || (typeof file.file_data === "string" && !file.file_data.startsWith("data:") ? file.file_data : null);
+    const mime = parsed?.mimeType || file.format || "application/octet-stream";
+    if (b64 && IMAGE_MIME_RE.test(mime)) {
+      const up = await uploadImageData({ ...ctx, base64: b64, mediaType: mime });
+      if (up?.url) return imageUrlBlock(up.url);
+    }
+    return { type: OPENAI_BLOCK.TEXT, text: stubText({ name, mime, bytes: decodedBytes(b64 || ""), reason: "Qoder reads documents via its file API, not inlined bytes" }) };
+  }
+
+  if (block.type === CLAUDE_BLOCK.DOCUMENT && block.source) {
+    const src = block.source;
+    const name = block.title || "document";
+    if (src.type === "base64" && src.data) {
+      const mime = src.media_type || "application/pdf";
+      if (IMAGE_MIME_RE.test(mime)) {
+        const up = await uploadImageData({ ...ctx, base64: src.data, mediaType: mime });
+        if (up?.url) return imageUrlBlock(up.url);
+      }
+      return { type: OPENAI_BLOCK.TEXT, text: stubText({ name, mime, bytes: decodedBytes(src.data), reason: "Qoder reads documents via its file API, not inlined bytes" }) };
+    }
+  }
+
+  if (typeof block.text === "string" && block.text.includes("data:") && block.text.length > 8192) {
+    const next = block.text.replace(DATA_URI_RE, (m) => {
+      const parsed = parseDataUri(m.trim());
+      const bytes = parsed ? decodedBytes(parsed.base64) : m.length;
+      if (bytes <= QODER_INLINE_FALLBACK_MAX_BYTES) return m;
+      return stubText({ mime: parsed?.mimeType, bytes, reason: "inlined data URI stripped from Qoder context" });
+    });
+    return { ...block, text: next };
+  }
+
+  return block;
+}
+
+async function rewriteContent(content, ctx) {
+  if (typeof content === "string") {
+    if (content.includes("data:") && content.length > 8192) {
+      return content.replace(DATA_URI_RE, (m) => {
+        const parsed = parseDataUri(m.trim());
+        const bytes = parsed ? decodedBytes(parsed.base64) : m.length;
+        if (bytes <= QODER_INLINE_FALLBACK_MAX_BYTES) return m;
+        return stubText({ mime: parsed?.mimeType, bytes, reason: "inlined data URI stripped from Qoder context" });
+      });
+    }
+    return content;
+  }
+  if (!Array.isArray(content)) return content;
+  const out = [];
+  for (const block of content) {
+    const next = await rewriteBlock(block, ctx);
+    if (next == null) continue;
+    out.push(next);
+  }
+  return out.length ? out : "";
+}
+
+function payloadBytes(messages) {
+  try {
+    return Buffer.byteLength(JSON.stringify(messages), "utf8");
+  } catch {
+    return 0;
+  }
+}
+
+function stripRemainingDataUris(messages) {
+  for (const msg of messages || []) {
+    if (typeof msg?.content === "string" && msg.content.includes("data:")) {
+      msg.content = msg.content.replace(DATA_URI_RE, (m) =>
+        stubText({ bytes: m.length, reason: "payload over Qoder size budget" }),
+      );
+    } else if (Array.isArray(msg?.content)) {
+      msg.content = msg.content.map((block) => {
+        if (block?.type === OPENAI_BLOCK.IMAGE_URL) {
+          const raw = typeof block.image_url === "string" ? block.image_url : block.image_url?.url;
+          if (typeof raw === "string" && raw.startsWith("data:")) {
+            return { type: OPENAI_BLOCK.TEXT, text: stubText({ name: "image", reason: "payload over Qoder size budget" }) };
+          }
+        }
+        if (typeof block?.text === "string" && block.text.includes("data:")) {
+          return { ...block, text: block.text.replace(DATA_URI_RE, (m) =>
+            stubText({ bytes: m.length, reason: "payload over Qoder size budget" }),
+          ) };
+        }
+        return block;
+      });
+    }
+  }
+}
+
+/**
+ * Rewrite OpenAI-shaped messages in place: upload images, stub huge files.
+ * @returns {Promise<{imageUrls: string[], uploaded: number, stubbed: number}>}
+ */
+export async function rewriteQoderMessageAttachments(messages, {
+  credentials,
+  log,
+  proxyOptions = null,
+  signal = null,
+  uploadFn = null,
+} = {}) {
+  const stats = { imageUrls: [], uploaded: 0, stubbed: 0 };
+  if (!Array.isArray(messages) || messages.length === 0) return stats;
+
+  const ctx = { credentials, log, proxyOptions, signal, uploadFn, cache: new Map() };
+
+  for (const msg of messages) {
+    if (!msg || typeof msg !== "object") continue;
+    if (Array.isArray(msg.images)) {
+      // Ollama-style sidecar; fold into content so normalizeMessages can see them.
+      const extras = msg.images.map((url) => imageUrlBlock(String(url)));
+      msg.content = Array.isArray(msg.content)
+        ? [...msg.content, ...extras]
+        : [{ type: OPENAI_BLOCK.TEXT, text: typeof msg.content === "string" ? msg.content : "" }, ...extras];
+      delete msg.images;
+    }
+    msg.content = await rewriteContent(msg.content, ctx);
+  }
+
+  // Collect surviving http(s) image URLs for callers that want image_urls.
+  for (const msg of messages) {
+    if (!Array.isArray(msg?.content)) continue;
+    for (const block of msg.content) {
+      const url = block?.type === OPENAI_BLOCK.IMAGE_URL
+        ? (typeof block.image_url === "string" ? block.image_url : block.image_url?.url)
+        : null;
+      if (typeof url === "string" && /^https?:\/\//i.test(url)) stats.imageUrls.push(url);
+      if (block?.type === OPENAI_BLOCK.TEXT && typeof block.text === "string" && block.text.startsWith("[file omitted:")) stats.stubbed += 1;
+    }
+  }
+  stats.uploaded = stats.imageUrls.length;
+
+  if (payloadBytes(messages) > QODER_MAX_PAYLOAD_BYTES) {
+    log?.warn?.("QODER", `request still ${payloadBytes(messages)} bytes after rewrite; stripping leftover data URIs`);
+    stripRemainingDataUris(messages);
+  }
+
+  return stats;
+}
+
+/** Test helper kept for callers; upload memo is now per-request. */
+export function clearQoderUploadCache() {}
+
+export const __test__ = {
+  extractUrlFromUploadResponse,
+  decodedBytes,
+  stubText,
+  payloadBytes,
+};
--- a/open-sse/shared/qoder/constants.js
+++ b/open-sse/shared/qoder/constants.js
@@ -33,6 +33,39 @@ export const QODER_CHAT_SIG_PATH = "/api/v2/service/pro/sse/agent_chat_generatio
 export const QODER_CHAT_URL = `${QODER_CHAT_BASE}/algo${QODER_CHAT_SIG_PATH}?FetchKeys=llm_model_result&AgentId=agent_common`;
 export const QODER_CHAT_URL_ENCODED = `${QODER_CHAT_URL}&Encode=1`;
 export const QODER_MODEL_LIST_URL = `${QODER_CHAT_BASE}/algo/api/v2/model/list`;
+// Official qodercli uploads images here (COSY-signed PUT multipart, field "file")
+// instead of inlining base64 into agent_chat_generation.
+export const QODER_IMAGE_UPLOAD_SIG_PATH = "/api/v2/image/upload";
+
+// Drop remaining inlined binaries if the Qoder JSON body would still exceed this.
+// 30MB+ payloads are what blow past Claude-Code's ~200k context on the wire.
+export const QODER_MAX_PAYLOAD_BYTES = 6 * 1024 * 1024;
+// If OSS upload fails, keep tiny data-URIs; anything larger is stubbed.
+export const QODER_INLINE_FALLBACK_MAX_BYTES = 512 * 1024;
+
+// Context-window tier selection (see shared/qoder/contextTier.js). The IDE exposes the
+// model's context_config tiers (200K/400K/1M); we auto-escalate when the estimated prompt
+// (+ headroom, tokenizer variance) no longer fits the current max_input_tokens.
+export const QODER_CONTEXT_TIER_HEADROOM = 0.15;
+export const QODER_CONTEXT_TIER_ENV = "QODER_CONTEXT_TIER";
+export const QODER_CONTEXT_TIER_MODES = Object.freeze({ AUTO: "auto", MAX: "max", DEFAULT: "default" });
+
+/**
+ * Job-token (jt-...) traffic must hit api2.qoder.sh — api3 rejects jt- with
+ * "Login expired" (403). Device tokens (dt-...) stay on api3. PATs (pt-...)
+ * are exchanged for jt- before this is consulted.
+ */
+export function qoderInferenceBase(credentials) {
+  const raw = credentials?.apiKey || credentials?.accessToken;
+  if (
+    typeof raw === "string" &&
+    !raw.startsWith("pt-") &&
+    (raw.startsWith("jt-") || (credentials?.accessToken || "").startsWith("jt-"))
+  ) {
+    return QODER_CHAT_BASE_ALT;
+  }
+  return QODER_CHAT_BASE;
+}

 // COSY header constants. These are not arbitrary — the upstream signature
 // validation matches them against the values used at signing time.
--- a/open-sse/shared/qoder/contextTier.js
+++ b/open-sse/shared/qoder/contextTier.js
@@ -0,0 +1,160 @@
+/**
+ * Qoder context-window tiers.
+ *
+ * Each Qoder model_config ships a `context_config` list (e.g. 200K / 400K / 1M for
+ * qmodel_38max) while `max_input_tokens` only carries the tier the IDE currently has
+ * selected (~180K by default). The Qoder IDE lets the user switch tiers from the model
+ * picker; a qodercli-style client (which is what 9router impersonates) has no picker,
+ * so a long Claude-Code / Codex session that grew past the default tier is rejected
+ * upstream even though the model itself supports 1M.
+ *
+ * This module emulates the IDE: estimate the prompt size, pick the smallest advertised
+ * tier that fits (never below the model's current default), and mirror the choice into
+ * the same three places the IDE writes:
+ *   parameters.context_length
+ *   chat_context.extra.ideModelConfigOverride.max_input_tokens
+ *   model_config.max_input_tokens
+ *
+ * Override with QODER_CONTEXT_TIER = auto (default) | max | default | <tier name, e.g. 1M>.
+ * Pure functions, no I/O — the executor wires them into buildQoderRequestBody.
+ */
+
+import { QODER_CONTEXT_TIER_HEADROOM, QODER_CONTEXT_TIER_MODES } from "./constants.js";
+
+const UNIT = { K: 1_000, M: 1_000_000 };
+
+/** "200K" | "1M" | "204800" | 204800 → integer token count (0 when unparseable). */
+export function parseTierTokenCount(value) {
+  if (typeof value === "number") return Number.isFinite(value) && value > 0 ? Math.floor(value) : 0;
+  if (typeof value !== "string") return 0;
+  const m = value.trim().toUpperCase().match(/^(\d+(?:\.\d+)?)\s*([KM])?$/);
+  if (!m) return 0;
+  const n = Number(m[1]) * (UNIT[m[2]] || 1);
+  return Number.isFinite(n) && n > 0 ? Math.floor(n) : 0;
+}
+
+function tierName(entry, tokenCount) {
+  const raw = entry.name ?? entry.label ?? entry.display_name ?? entry.displayName ?? entry.key ?? entry.id;
+  if (typeof raw === "string" && raw.trim()) return raw.trim();
+  if (tokenCount >= UNIT.M && tokenCount % UNIT.M === 0) return `${tokenCount / UNIT.M}M`;
+  if (tokenCount >= UNIT.K && tokenCount % UNIT.K === 0) return `${tokenCount / UNIT.K}K`;
+  return String(tokenCount);
+}
+
+/**
+ * Normalize a model_config into sorted tiers: [{ name, tokenCount, isDefault }] ascending.
+ * Accepts snake_case and camelCase shapes; returns [] when the model has no tiers.
+ */
+export function getQoderContextTiers(modelConfig) {
+  const list = modelConfig?.context_config ?? modelConfig?.contextConfig;
+  if (!Array.isArray(list)) return [];
+  const byCount = new Map();
+  for (const entry of list) {
+    if (!entry || typeof entry !== "object") continue;
+    const tokenCount = parseTierTokenCount(
+      entry.tokenCount ?? entry.token_count ?? entry.max_input_tokens ?? entry.maxInputTokens ?? entry.contextLength ?? entry.context_length,
+    );
+    if (!tokenCount) continue;
+    const isDefault = entry.isDefault === true || entry.is_default === true || entry.default === true;
+    const prev = byCount.get(tokenCount);
+    byCount.set(tokenCount, {
+      name: tierName(entry, tokenCount),
+      tokenCount,
+      isDefault: (prev?.isDefault || false) || isDefault,
+    });
+  }
+  return [...byCount.values()].sort((a, b) => a.tokenCount - b.tokenCount);
+}
+
+const CJK_RE = /[\u1100-\u11ff\u2e80-\u9fff\uac00-\ud7af\uf900-\ufaff\uff00-\uffef]/g;
+
+/**
+ * Rough prompt-size estimate in tokens. CJK characters count ~1 token each, everything
+ * else ~4 chars/token — the plain chars/4 rule underestimates Chinese/Japanese by up to
+ * 4x, which is exactly when a tier decision matters.
+ */
+export function estimateQoderPromptTokens({ system, messages, tools } = {}) {
+  let text = "";
+  try {
+    text = JSON.stringify({ system: system || "", messages: messages || [], tools: tools || [] }) || "";
+  } catch {
+    return 0;
+  }
+  const cjk = (text.match(CJK_RE) || []).length;
+  return Math.ceil(cjk + (text.length - cjk) / 4);
+}
+
+function normalizeMode(preference) {
+  const p = String(preference ?? "").trim();
+  return p ? p : QODER_CONTEXT_TIER_MODES.AUTO;
+}
+
+function findNamedTier(tiers, name) {
+  const wanted = name.replace(/\s+/g, "").toUpperCase();
+  const asCount = parseTierTokenCount(wanted);
+  return tiers.find((t) => t.name.replace(/\s+/g, "").toUpperCase() === wanted || (asCount && t.tokenCount === asCount)) || null;
+}
+
+/**
+ * Decide which tier a request should run under.
+ *
+ * @param {object} modelConfig   raw Qoder model_config (has context_config + max_input_tokens)
+ * @param {{system?: string, messages?: any[], tools?: any[]}} prompt  what will be sent
+ * @param {{preference?: string, headroom?: number}} [options]
+ * @returns {{ tier: {name, tokenCount, isDefault}, estimatedTokens: number, reason: string } | null}
+ *   null → leave the payload exactly as before (no tiers, or the default already fits).
+ */
+export function resolveQoderContextTier(modelConfig, prompt, options = {}) {
+  const tiers = getQoderContextTiers(modelConfig);
+  if (!tiers.length) return null;
+
+  const mode = normalizeMode(options.preference);
+  const largest = tiers[tiers.length - 1];
+  const defaultTier = tiers.find((t) => t.isDefault) || tiers[0];
+  const estimatedTokens = estimateQoderPromptTokens(prompt);
+  const headroom = typeof options.headroom === "number" ? options.headroom : QODER_CONTEXT_TIER_HEADROOM;
+  const need = Math.ceil(estimatedTokens * (1 + headroom));
+
+  if (mode.toLowerCase() === QODER_CONTEXT_TIER_MODES.MAX) {
+    return { tier: largest, estimatedTokens, reason: "forced:max" };
+  }
+  if (mode.toLowerCase() === QODER_CONTEXT_TIER_MODES.DEFAULT) {
+    return { tier: defaultTier, estimatedTokens, reason: "forced:default" };
+  }
+  if (mode.toLowerCase() !== QODER_CONTEXT_TIER_MODES.AUTO) {
+    const named = findNamedTier(tiers, mode);
+    if (named) return { tier: named, estimatedTokens, reason: `forced:${named.name}` };
+    // Unknown tier name → fall through to auto rather than silently breaking requests.
+  }
+
+  // auto: keep the upstream default (current behaviour) while the prompt fits in it.
+  const currentMax = parseTierTokenCount(modelConfig?.max_input_tokens ?? modelConfig?.maxInputTokens);
+  const currentLimit = currentMax || defaultTier.tokenCount;
+  if (need <= currentLimit) return null;
+
+  const fits = tiers.find((t) => t.tokenCount >= need && t.tokenCount > currentLimit);
+  const tier = fits || largest;
+  if (tier.tokenCount <= currentLimit) return null; // nothing bigger to escalate to
+  return { tier, estimatedTokens, reason: fits ? "auto:fits" : "auto:largest" };
+}
+
+/**
+ * Write the chosen tier into a Qoder chat payload (mutates + returns it).
+ * Mirrors the IDE: parameters.context_length, ideModelConfigOverride, model_config.
+ */
+export function applyQoderContextTier(payload, tier) {
+  if (!payload || !tier?.tokenCount) return payload;
+  payload.parameters = { ...(payload.parameters || {}), context_length: tier.tokenCount };
+  payload.chat_context = payload.chat_context || {};
+  payload.chat_context.extra = {
+    ...(payload.chat_context.extra || {}),
+    ideModelConfigOverride: {
+      ...(payload.chat_context.extra?.ideModelConfigOverride || {}),
+      max_input_tokens: tier.tokenCount,
+    },
+  };
+  if (payload.model_config && typeof payload.model_config === "object") {
+    payload.model_config = { ...payload.model_config, max_input_tokens: tier.tokenCount };
+  }
+  return payload;
+}
--- a/open-sse/shared/qoder/sse.js
+++ b/open-sse/shared/qoder/sse.js
@@ -0,0 +1,208 @@
+/**
+ * Qoder SSE is OpenAI-shaped inside `{statusCodeValue, body}` envelopes, but
+ * usage arrives on a later `choices: []` frame — after finish_reason, which
+ * itself often lives on `delta.finish_reason` rather than the choice.
+ *
+ * Downstream (Claude translator, OpenAI clients, Claude Code) look for usage
+ * on the finish chunk or drop `choices: []` entirely. 9router's own dashboard
+ * still sees tokens because extractUsage runs on every forwarded frame.
+ *
+ * Coalesce: hold empty finish + usage-only frames, then emit one OpenAI
+ * include_usage-style chunk: `{choices:[{delta:{}, finish_reason}], usage}`.
+ */
+
+function num(v) {
+  const n = Number(v);
+  return Number.isFinite(n) ? n : null;
+}
+
+/**
+ * Normalize Qoder/OpenAI usage into the shape stream.js + Claude translation
+ * already understand (prompt_tokens + prompt_tokens_details.cached_tokens).
+ */
+export function canonicalizeQoderUsage(usage) {
+  if (!usage || typeof usage !== "object" || Array.isArray(usage)) return null;
+
+  const prompt = num(usage.prompt_tokens ?? usage.input_tokens);
+  const completion = num(usage.completion_tokens ?? usage.output_tokens);
+  if (prompt == null && completion == null) return null;
+
+  const details = (usage.prompt_tokens_details && typeof usage.prompt_tokens_details === "object")
+    ? { ...usage.prompt_tokens_details }
+    : {};
+  const cached = num(
+    details.cached_tokens ??
+    usage.cached_tokens ??
+    usage.prompt_cache_hit_tokens ??
+    usage.cache_read_input_tokens,
+  );
+  const cacheCreation = num(
+    details.cache_creation_tokens ??
+    usage.cache_creation_input_tokens,
+  );
+
+  const promptTokens = prompt || 0;
+  const completionTokens = completion || 0;
+  const out = {
+    prompt_tokens: promptTokens,
+    completion_tokens: completionTokens,
+    total_tokens: num(usage.total_tokens) ?? (promptTokens + completionTokens),
+  };
+
+  if (cached != null) {
+    out.cached_tokens = cached;
+    details.cached_tokens = cached;
+  }
+  if (cacheCreation != null) {
+    details.cache_creation_tokens = cacheCreation;
+  }
+  if (Object.keys(details).length) out.prompt_tokens_details = details;
+
+  if (usage.completion_tokens_details && typeof usage.completion_tokens_details === "object") {
+    out.completion_tokens_details = usage.completion_tokens_details;
+  }
+  const reasoning = num(usage.reasoning_tokens ?? usage.completion_tokens_details?.reasoning_tokens);
+  if (reasoning != null) out.reasoning_tokens = reasoning;
+
+  return out;
+}
+
+function finishReasonOf(parsed) {
+  const choice = parsed?.choices?.[0];
+  return choice?.finish_reason || choice?.delta?.finish_reason || parsed?.finish_reason || null;
+}
+
+function hasValuableDelta(parsed) {
+  const delta = parsed?.choices?.[0]?.delta;
+  if (!delta || typeof delta !== "object") return false;
+  if (typeof delta.content === "string" && delta.content.length > 0) return true;
+  if (typeof delta.reasoning_content === "string" && delta.reasoning_content.length > 0) return true;
+  if (Array.isArray(delta.tool_calls) && delta.tool_calls.length > 0) return true;
+  if (delta.role) return true;
+  return false;
+}
+
+function parseInner(inner) {
+  if (inner == null || inner === "") return { raw: false, parsed: null };
+  if (inner === "[DONE]") return { done: true };
+  if (typeof inner !== "string") {
+    if (typeof inner === "object") return { parsed: inner };
+    return { raw: true, text: String(inner) };
+  }
+  try {
+    return { parsed: JSON.parse(inner) };
+  } catch {
+    return { raw: true, text: inner };
+  }
+}
+
+/**
+ * @param {object} opts
+ * @param {string} opts.model
+ * @param {TextEncoder} opts.encoder
+ * @param {string} opts.sseDone  "data: [DONE]\\n\\n"
+ */
+export function createQoderSseCoalescer({ model, encoder, sseDone }) {
+  let pendingFinish = null;
+  let pendingUsage = null;
+  let lastMeta = { id: null, created: null, model };
+  let doneEmitted = false;
+  let finishAlreadyForwarded = false;
+
+  const emitJson = (controller, obj) => {
+    const sanitized = JSON.stringify(obj).replace(/\r?\n/g, "");
+    controller.enqueue(encoder.encode(`data: ${sanitized}\n\n`));
+  };
+
+  const emitRaw = (controller, text) => {
+    controller.enqueue(encoder.encode(`data: ${String(text).replace(/\r?\n/g, "")}\n\n`));
+  };
+
+  const emitDone = (controller) => {
+    if (doneEmitted) return;
+    controller.enqueue(encoder.encode(sseDone));
+    doneEmitted = true;
+  };
+
+  const emitTerminal = (controller) => {
+    if (!pendingFinish && !pendingUsage) return;
+    emitJson(controller, {
+      id: lastMeta.id || `qoder-${Date.now()}`,
+      object: "chat.completion.chunk",
+      created: lastMeta.created || Math.floor(Date.now() / 1000),
+      model: lastMeta.model || model,
+      choices: [{ index: 0, delta: {}, finish_reason: pendingFinish || "stop" }],
+      ...(pendingUsage ? { usage: pendingUsage } : {}),
+    });
+    pendingFinish = null;
+    pendingUsage = null;
+  };
+
+  const flush = (controller) => {
+    if (doneEmitted) return;
+    if (pendingUsage || (pendingFinish && !finishAlreadyForwarded)) {
+      emitTerminal(controller);
+    }
+    emitDone(controller);
+  };
+
+  const handleInner = (inner, controller) => {
+    if (doneEmitted) return { terminal: true };
+
+    const parsedInner = parseInner(inner);
+    if (parsedInner.done) {
+      flush(controller);
+      return { terminal: true };
+    }
+    if (parsedInner.raw) {
+      emitRaw(controller, parsedInner.text);
+      return {};
+    }
+    const parsed = parsedInner.parsed;
+    if (!parsed || typeof parsed !== "object") return {};
+
+    if (typeof parsed.id === "string" && parsed.id) lastMeta.id = parsed.id;
+    if (typeof parsed.created === "number") lastMeta.created = parsed.created;
+    if (typeof parsed.model === "string" && parsed.model) lastMeta.model = parsed.model;
+
+    const usage = canonicalizeQoderUsage(parsed.usage);
+    if (usage) pendingUsage = usage;
+
+    const finish = finishReasonOf(parsed);
+    if (hasValuableDelta(parsed)) {
+      // Stream content as-is (preserves upstream JSON for tests/clients).
+      emitRaw(controller, typeof inner === "string" ? inner : JSON.stringify(parsed));
+      if (finish) {
+        finishAlreadyForwarded = true;
+        // Keep finish around only if we still need a usage trailer.
+        pendingFinish = pendingUsage ? finish : null;
+      }
+      if (pendingFinish && pendingUsage) {
+        emitTerminal(controller);
+        emitDone(controller);
+        return { terminal: true };
+      }
+      return {};
+    }
+
+    if (finish) pendingFinish = finish;
+
+    // Empty finish and/or usage-only: emit as soon as we have both (Qoder
+    // order is finish then usage). Don't wait for the later [DONE]/keepalive.
+    if ((pendingFinish || finishAlreadyForwarded) && pendingUsage) {
+      if (!pendingFinish) pendingFinish = "stop";
+      emitTerminal(controller);
+      emitDone(controller);
+      return { terminal: true };
+    }
+    return {};
+  };
+
+  return {
+    handleInner,
+    flush,
+    get doneEmitted() {
+      return doneEmitted;
+    },
+  };
+}
--- a/open-sse/translator/concerns/paramSupport.js
+++ b/open-sse/translator/concerns/paramSupport.js
@@ -14,6 +14,9 @@ const STRIP_RULES = [
  { 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 },
+  // MiMo Desktop Preview models (account-service route): content must be plain string,
+  // rejects OpenAI content-part array. Cloud models keep their parts (mimo-v2-omni is multi-modal).
+  { provider: "xiaomi-mimo", match: /preview/i, flattenContent: true },
  { provider: "volcengine-ark", match: /glm-5/i, clampToModelMaxOutput: true },
  // VolcEngine Ark caps the Kimi family at max_tokens <= 32768, but the model's
  // advertised ceiling is far higher (Kimi-K2.7-Code resolves to maxOutput 262144),
--- a/open-sse/translator/concerns/toolCall.js
+++ b/open-sse/translator/concerns/toolCall.js
@@ -1,5 +1,7 @@
 // Tool call helper functions for translator

+import { FORMATS } from "../formats.js";
+
 // Anthropic tool_use.id must match: ^[a-zA-Z0-9_-]+$
 const TOOL_ID_PATTERN = /^[a-zA-Z0-9_-]+$/;

@@ -165,3 +167,16 @@ export function defaultClaudeToolType(tools) {
  return tools.map(tool => tool?.type ? tool : { ...tool, type: "custom" });
 }

+// Whether Claude-format tools need explicit `type` defaulting before dispatch.
+// Only gateways that declare the `requireClaudeToolType` quirk (MiniMax) reject typeless
+// tools. Applying the default globally breaks Claude-format endpoints that only accept the
+// legacy typeless tool shape — DeepSeek's Anthropic-compatible endpoint answers HTTP 400
+// "unknown variant `custom`" and every Claude Code request routed there fails (#3905).
+export function shouldDefaultClaudeToolType(provider, finalFormat, tools, PROVIDERS) {
+  return (
+    finalFormat === FORMATS.CLAUDE
+    && Array.isArray(tools)
+    && PROVIDERS?.[provider]?.quirks?.requireClaudeToolType === true
+  );
+}
+
--- a/open-sse/translator/formats/claude.js
+++ b/open-sse/translator/formats/claude.js
@@ -27,6 +27,14 @@ export function lastCacheableToolIndex(tools) {
 // Check if message has valid non-empty content
 export function hasValidContent(msg) {
  if (typeof msg.content === "string" && msg.content.trim()) return true;
+  if (msg.content && typeof msg.content === "object" && !Array.isArray(msg.content)) {
+    const block = msg.content;
+    return !!((block.type === CLAUDE_BLOCK.TEXT && block.text?.trim()) ||
+      block.type === CLAUDE_BLOCK.TOOL_USE ||
+      block.type === CLAUDE_BLOCK.TOOL_RESULT ||
+      block.type === CLAUDE_BLOCK.IMAGE ||
+      block.type === CLAUDE_BLOCK.DOCUMENT);
+  }
  if (Array.isArray(msg.content)) {
    return msg.content.some(block =>
      (block.type === CLAUDE_BLOCK.TEXT && block.text?.trim()) ||
@@ -38,6 +46,60 @@ export function hasValidContent(msg) {
  }
  return false;
 }
+// Content may arrive as a single content block object (spec allows string | array;
+// some clients send the bare object). Wrap it as a one-block array and strip any
+// client-placed cache_control: a bare-object marker must never survive
+// normalization, on any path, guard or no guard.
+function normalizeMessageContent(msg) {
+  const c = msg?.content;
+  if (c && typeof c === "object" && !Array.isArray(c)) {
+    delete c.cache_control;
+    msg.content = [c];
+  }
+  return msg;
+}
+
+// Total blocks carrying cache_control across system, tools, and messages — the
+// upstream Messages API allows at most 4 markers per request.
+function countCacheControlBlocks(body) {
+  let n = 0;
+  if (Array.isArray(body?.system)) for (const b of body.system) if (b?.cache_control) n++;
+  if (Array.isArray(body?.tools)) for (const t of body.tools) if (t?.cache_control) n++;
+  if (Array.isArray(body?.messages)) {
+    for (const m of body.messages) {
+      if (Array.isArray(m?.content)) {
+        for (const b of m.content) if (b?.cache_control) n++;
+      } else if (m?.content && typeof m.content === "object" && m.content.cache_control) n++;
+    }
+  }
+  return n;
+}
+// Trim every marker past the 4-marker budget. The head anchors (last system
+// block, last cacheable tool) are held; the remaining slots go to the tail-most
+// of the other markers in document order. A plain "keep the last 4 in document
+// order" rule would drop the head anchors first — they lead document order, yet
+// they are exactly what re-anchoring exists to pin.
+function capCacheControlBlocks(body) {
+  const isHead = (b) => {
+    const sys = Array.isArray(body?.system) ? body.system : [];
+    if (sys.length && sys[sys.length - 1] === b) return true;
+    const tools = Array.isArray(body?.tools) ? body.tools : [];
+    const lastTool = lastCacheableToolIndex(tools);
+    return lastTool >= 0 && tools[lastTool] === b;
+  };
+  const marked = [];
+  if (Array.isArray(body?.system)) for (const b of body.system) if (b?.cache_control) marked.push(b);
+  if (Array.isArray(body?.tools)) for (const t of body.tools) if (t?.cache_control) marked.push(t);
+  if (Array.isArray(body?.messages)) {
+    for (const m of body.messages) {
+      if (Array.isArray(m?.content)) for (const b of m.content) if (b?.cache_control) marked.push(b);
+    }
+  }
+  const head = marked.filter(isHead);
+  const rest = marked.filter(b => !isHead(b));
+  const keep = Math.max(0, 4 - head.length);
+  for (const b of rest.slice(0, Math.max(0, rest.length - keep))) delete b.cache_control;
+}

 // Fix tool_use/tool_result ordering for Claude API
 // 1. Assistant message with tool_use: remove text AFTER tool_use (Claude doesn't allow)
@@ -136,8 +198,9 @@ function hasForeignServerToolUseId(block) {
 // Newer Cowork/Claude Code clients emit beta-only shapes that OAuth endpoints reject:
 // 1. thinking.type "adaptive" → unsupported on Haiku
 // 2. output_config.effort → unsupported on Haiku
-// 3. role "system" messages (mid-conversation-system beta) → only top-level system is allowed
-// 4. server_tool_use blocks carrying a foreign (non-srvtoolu_) id → rejected outright
+// 3. bare content-block objects (content: {block} instead of [{block}]) → wrapped first
+// 4. role "system" messages (mid-conversation-system beta) → only top-level system is allowed
+// 5. server_tool_use blocks carrying a foreign (non-srvtoolu_) id → rejected outright
 export function normalizeClaudePassthrough(body, model = "") {
  if (!body || typeof body !== "object") return body;

@@ -152,7 +215,15 @@ export function normalizeClaudePassthrough(body, model = "") {
    if (Object.keys(body.output_config).length === 0) delete body.output_config;
  }

-  // 2. Fold mid-conversation system messages into the neighbouring turn.
+  // 3. Wrap bare content-block objects as one-element arrays before folding.
+  // Some clients send content: {block} instead of content: [{block}]; the
+  // mid-conversation-system fold below assumes the array shape, so it must
+  // run first — a bare-object neighbor would otherwise be zeroed to [].
+  if (Array.isArray(body.messages)) {
+    for (const msg of body.messages) normalizeMessageContent(msg);
+  }
+
+  // 4. Fold mid-conversation system messages into the neighbouring turn.
  // Hoisting them into body.system would insert volatile content (token counters,
  // reminders) ahead of the whole conversation and invalidate the prefix cache on
  // every request. Folding in place keeps the cached prefix stable.
@@ -186,7 +257,7 @@ export function normalizeClaudePassthrough(body, model = "") {
    body.messages = messages;
  }

-  // 3. Drop thinking blocks whose signature is not Claude's (combo mixes models,
+  // 5. Drop thinking blocks whose signature is not Claude's (combo mixes models,
  // so foreign signatures leak into history and Anthropic rejects them).
  const thinkingEnabled = body.thinking?.type === "enabled";
  const droppedServerToolUseIds = new Set();
@@ -233,7 +304,7 @@ export function normalizeClaudePassthrough(body, model = "") {
    }
  }

-  // 5. Drop empty text blocks and any message left with no content at all.
+  // 6. Drop empty text blocks and any message left with no content at all.
  // Anthropic rejects `messages.N.content` blocks with empty text (400
  // "text content blocks must be non-empty"); a message whose blocks were all
  // stripped above must be dropped, not padded with an empty placeholder.
@@ -271,7 +342,22 @@ function markLastCacheableBlock(msg) {
 // (normalize, tool dedupe, token savers) — otherwise the anchor drifts off the tail.
 export function anchorClaudeCache(body) {
  if (!body || typeof body !== "object") return body;
+  if (Array.isArray(body.messages)) {
+    for (const msg of body.messages) normalizeMessageContent(msg);
+  }
+  // Invalid markers first, whatever the budget: Anthropic rejects a tool that
+  // carries BOTH defer_loading and cache_control (#3567). The re-anchor path
+  // below strips them anyway; the over-budget early return used to forward
+  // them untouched.
+  if (Array.isArray(body.tools)) {
+    for (const t of body.tools) {
+      if (t?.defer_loading === true) delete t.cache_control;
+    }
+  }

+  // Head anchors first, before any budget guard: the 1h TTL on system/tools is
+  // the point of re-anchoring, and skipping it because the client spent its
+  // budget would silently downgrade a cache hit to the 5m default.
  if (Array.isArray(body.system)) {
    const last = body.system.length - 1;
    body.system.forEach((block, i) => {
@@ -289,6 +375,15 @@ export function anchorClaudeCache(body) {
    });
  }

+  // Budget guard AFTER the head anchors: with the last system block and last
+  // tool pinned, at most 2 slots remain. At >= 4 markers the client has spent
+  // the rest of the budget and every remaining marker is itself a valid
+  // breakpoint — re-anchoring the tail could only exceed 4, so trim instead.
+  if (countCacheControlBlocks(body) >= 4) {
+    capCacheControlBlocks(body);
+    return body;
+  }
+
  if (Array.isArray(body.messages)) {
    let anchored = null;
    for (let i = body.messages.length - 1; i >= 0; i--) {
@@ -368,6 +463,7 @@ export function prepareClaudeRequest(body, provider = null, apiKey = null, conne
    // Pass 1: remove cache_control + filter empty messages
    for (let i = 0; i < len; i++) {
      const msg = body.messages[i];
+      normalizeMessageContent(msg);

      // Remove cache_control from content blocks
      if (Array.isArray(msg.content)) {
@@ -461,8 +557,21 @@ export function prepareClaudeRequest(body, provider = null, apiKey = null, conne
    // Strip built-in tools (e.g. web_search_20250305) and normalize to Anthropic-native shape
    // (drop `type` field, fold `function.{name,description,parameters}`) for non-Anthropic providers
    if (provider !== "claude") {
+      // Provider-specific whitelist of Anthropic tool `type` values that the
+      // upstream actually accepts. When the provider declares it
+      // (e.g. DeepSeek — only web_search_*), keep only listed types; otherwise
+      // keep the prior behaviour of dropping every non-function tool, which is
+      // correct for OpenAI-compatible targets reached through this Claude-format
+      // pass (their tools get normalized below to function-style).
+      const supportedTypes = PROVIDERS[provider]?.quirks?.claudeSupportedToolTypes;
+      const hasWhitelist = Array.isArray(supportedTypes);
      body.tools = body.tools
-        .filter(tool => !tool.type || tool.type === "function")
+        .filter(tool => {
+          const t = tool?.type;
+          if (!t || t === "function") return true;
+          if (hasWhitelist) return supportedTypes.includes(t);
+          return false;
+        })
        .map(tool => {
          if (tool.function) {
            return {
@@ -471,6 +580,13 @@ export function prepareClaudeRequest(body, provider = null, apiKey = null, conne
              input_schema: tool.function.parameters,
            };
          }
+          // When the provider declared a supportedToolTypes whitelist, keep
+          // the surviving tools' `type` field intact — the upstream
+          // Anthropic-compatible endpoint (e.g. DeepSeek) requires it to
+          // route built-ins like web_search_* correctly. Without a
+          // whitelist, preserve prior behaviour and strip `type` so the
+          // tool is normalized to plain Anthropic shape.
+          if (hasWhitelist) return tool;
          const { type, ...rest } = tool;
          return rest;
        });
--- a/open-sse/translator/formats/gemini.js
+++ b/open-sse/translator/formats/gemini.js
@@ -432,3 +432,21 @@ export function cleanJSONSchemaForAntigravity(schema) {
  return cleaned;
 }

+// Merge adjacent same-role messages, strip empty parts, ensure initial user turn
+export function normalizeGeminiContents(contents) {
+  const out = [];
+  for (const c of contents || []) {
+    if (!c?.role || !Array.isArray(c.parts)) continue;
+    const parts = c.parts.filter(p => p && Object.keys(p).length > 0);
+    if (parts.length === 0) continue;
+    const last = out.at(-1);
+    if (last?.role === c.role) last.parts.push(...parts);
+    else out.push({ ...c, parts: [...parts] });
+  }
+  if (out.length > 0 && out[0].role !== "user") {
+    out.unshift({ role: "user", parts: [{ text: "..." }] });
+  }
+  return out;
+}
+
+
--- a/open-sse/translator/request/claude-to-kiro.js
+++ b/open-sse/translator/request/claude-to-kiro.js
@@ -242,9 +242,9 @@ export function claudeToKiroRequest(model, body, stream, credentials) {
    ? (credentials?.providerSpecificData?.profileArn || "")
    : (credentials?.providerSpecificData?.profileArn || resolveDefaultProfileArn(authMethod));

-  // Kiro CLI/KAS sends system prompt as top-level `systemPrompt`. Keep a
-  // content fallback too because the CodeWhisperer surface does not always
-  // enforce top-level systemPrompt for direct calls.
+  // The system prompt travels inside the first user turn's content (contentPrefix):
+  // the CodeWhisperer surface rejects a top-level `systemPrompt` with
+  // 400 REQUEST_BODY_INVALID, so the value below is only a replay cache key.
  const timestamp = new Date().toISOString();
  const systemPromptParts = [];
  if (thinkingBudget !== null && !usesNativeGptEffort) {
@@ -316,14 +316,11 @@ export function claudeToKiroRequest(model, body, stream, credentials) {
    conversationState: {
      chatTriggerType: "MANUAL",
      conversationId,
-      agentContinuationId: continuationId,
-      agentTaskType: "vibe",
      currentMessage: {
        userInputMessage,
      },
      history: canonical.history,
    },
-    agentMode: "vibe",
  };

  if (profileArn) payload.profileArn = profileArn;
--- a/open-sse/translator/request/claude-to-openai.js
+++ b/open-sse/translator/request/claude-to-openai.js
@@ -142,6 +142,13 @@ function systemReminderText(content) {

 // Convert single Claude message - returns single message or array of messages
 function convertClaudeMessage(msg) {
+  // Some clients send content as a single block object; normalize to the
+  // one-element array every branch below (the system-reminder fold included)
+  // expects. Must run BEFORE the role branch: systemReminderText only reads
+  // arrays and strings, so a bare-object system turn was dropped outright.
+  if (msg.content && typeof msg.content === "object" && !Array.isArray(msg.content)) {
+    msg.content = [msg.content];
+  }
  // Mid-conversation system message -> user (per Anthropic placement rules)
  if (msg.role === ROLE.SYSTEM) {
    const text = systemReminderText(msg.content);
--- a/open-sse/translator/request/openai-to-gemini.js
+++ b/open-sse/translator/request/openai-to-gemini.js
@@ -15,7 +15,8 @@ import {
  generateRequestId,
  generateSessionId,
  generateProjectId,
-  cleanJSONSchemaForAntigravity
+  cleanJSONSchemaForAntigravity,
+  normalizeGeminiContents
 } from "../formats/gemini.js";
 import { deriveSessionId, toNumericSessionId } from "../../utils/sessionManager.js";
 import { ROLE, GEMINI_ROLE, OPENAI_BLOCK, CLAUDE_BLOCK } from "../schema/index.js";
@@ -35,17 +36,6 @@ function sanitizeGeminiFunctionName(name) {
  return sanitized.substring(0, 64);
 }

-function normalizeGeminiContents(contents) {
-  const out = [];
-  for (const c of contents || []) {
-    if (!c?.role || !Array.isArray(c.parts) || c.parts.length === 0) continue;
-    const last = out.at(-1);
-    if (last?.role === c.role) last.parts.push(...c.parts);
-    else out.push({ ...c, parts: [...c.parts] });
-  }
-  return out;
-}
-
 // Core: Convert OpenAI request to Gemini format (base for all variants)
 function openaiToGeminiBase(model, body, stream, signature = DEFAULT_THINKING_AG_SIGNATURE, sessionId = null) {
  const result = {
@@ -163,12 +153,14 @@ function openaiToGeminiBase(model, body, stream, signature = DEFAULT_THINKING_AG
          }

          // Check if there are actual tool responses in the next messages
-          const hasActualResponses = toolCallIds.some(fid => toolResponses[fid]);
+          const isIntermediate = i < body.messages.length - 1;
+          const hasActualResponses = toolCallIds.some(fid => toolResponses[fid] !== undefined);

-          if (hasActualResponses) {
+          if (hasActualResponses || isIntermediate) {
            const toolParts = [];
            for (const fid of toolCallIds) {
-              if (!toolResponses[fid]) continue;
+              let resp = toolResponses[fid];
+              if (resp === undefined) resp = "";

              let name = tcID2Name[fid];
              if (!name) {
@@ -180,7 +172,6 @@ function openaiToGeminiBase(model, body, stream, signature = DEFAULT_THINKING_AG
                }
              }

-              let resp = toolResponses[fid];
              let parsedResp = tryParseJSON(resp);
              if (parsedResp === null) {
                parsedResp = { result: resp };
--- a/open-sse/translator/request/openai-to-kiro.js
+++ b/open-sse/translator/request/openai-to-kiro.js
@@ -340,9 +340,9 @@ export function openaiToKiroRequest(model, body, stream, credentials) {

  const timestamp = new Date().toISOString();

-  // Kiro CLI/KAS sends these as top-level systemPrompt. Keep a content fallback
-  // too because the CodeWhisperer surface does not always enforce top-level
-  // systemPrompt for direct calls.
+  // The system prompt travels inside the first user turn's content (contentPrefix):
+  // the CodeWhisperer surface rejects a top-level `systemPrompt` with
+  // 400 REQUEST_BODY_INVALID, so the value below is only a replay cache key.
  const systemPromptParts = [];
  if (thinkingBudget !== null && !usesNativeGptEffort) {
    systemPromptParts.push(buildThinkingSystemPrefix(thinkingBudget));
@@ -397,8 +397,6 @@ export function openaiToKiroRequest(model, body, stream, credentials) {
    conversationState: {
      chatTriggerType: "MANUAL",
      conversationId,
-      agentContinuationId: continuationId,
-      agentTaskType: "vibe",
      currentMessage: {
        userInputMessage: {
          content: replayCurrent.content || "",
@@ -414,7 +412,6 @@ export function openaiToKiroRequest(model, body, stream, credentials) {
      },
      history: canonical.history
    },
-    agentMode: "vibe",
  };

  if (profileArn) {
--- a/open-sse/utils/codexToolSchema.js
+++ b/open-sse/utils/codexToolSchema.js
@@ -0,0 +1,80 @@
+// Codex-specific tool JSON Schema compatibility.
+//
+// `https://chatgpt.com/backend-api/codex/responses` validates every function
+// tool's `parameters` with a regex engine that does not implement Unicode
+// property escapes. A `pattern` such as
+//
+//   "^(?!__.*__$)[^\\p{Cc}\\p{Cf}\\p{Zl}\\p{Zp}\"\\\\./\\[\\]]{1,200}$"
+//
+// is a perfectly valid ECMAScript `u`-mode regex, but Codex answers
+//
+//   400 Invalid schema for function 'Artifact': '^\p{Cc}...' is not a 'regex'
+//   param: tools[0].parameters
+//
+// The request is deterministically malformed for this provider, so every
+// account fails identically and the combo pays a full failover before landing
+// somewhere that accepts it (#3922).
+//
+// Scope guardrail (#3667): this is NOT a global schema sanitizer. Providers
+// that do support `\p{...}` keep the constraint untouched — the strip runs only
+// on the Codex dispatch path, and only on `pattern` strings that actually
+// contain a property escape. Everything else in the schema (including valid
+// patterns) passes through byte-identical.
+
+// `\p{...}` / `\P{...}` with an odd number of preceding backslashes — an even
+// count means the backslash itself is escaped, so `\\p{Cc}` is a literal "p".
+const UNICODE_PROPERTY_ESCAPE = /(^|[^\\])(\\\\)*\\[pP]\{/;
+
+export function hasUnicodePropertyEscape(pattern) {
+  return typeof pattern === "string" && UNICODE_PROPERTY_ESCAPE.test(pattern);
+}
+
+// Copy-on-write walk: returns the original reference when nothing changed, so
+// untouched schemas keep object identity and callers can cheaply detect a no-op.
+// `properties` is special-cased because its keys are arbitrary property *names*
+// (which may themselves be "pattern" or "properties") and must never be read as
+// schema keywords; every other key recurses as an ordinary schema node.
+function stripNode(node, stats) {
+  if (Array.isArray(node)) {
+    let changed = false;
+    const next = node.map((item) => {
+      const cleaned = stripNode(item, stats);
+      if (cleaned !== item) changed = true;
+      return cleaned;
+    });
+    return changed ? next : node;
+  }
+  if (!node || typeof node !== "object") return node;
+
+  let changed = false;
+  const next = {};
+  for (const [key, value] of Object.entries(node)) {
+    if (key === "pattern" && hasUnicodePropertyEscape(value)) {
+      stats.removed++;
+      changed = true;
+      continue;
+    }
+    if (key === "properties" && value && typeof value === "object" && !Array.isArray(value)) {
+      let propsChanged = false;
+      const props = {};
+      for (const [propName, propSchema] of Object.entries(value)) {
+        const cleaned = stripNode(propSchema, stats);
+        if (cleaned !== propSchema) propsChanged = true;
+        props[propName] = cleaned;
+      }
+      if (propsChanged) changed = true;
+      next[key] = propsChanged ? props : value;
+      continue;
+    }
+    const cleaned = stripNode(value, stats);
+    if (cleaned !== value) changed = true;
+    next[key] = cleaned;
+  }
+  return changed ? next : node;
+}
+
+// Remove only the `pattern` constraints Codex's validator rejects.
+// Returns the same reference when the schema is already compatible.
+export function stripCodexUnsupportedPatterns(schema, stats = { removed: 0 }) {
+  return stripNode(schema, stats);
+}