Merge remote-tracking branch 'origin/master' into gitea/new_feature

# Conflicts: # open-sse/handlers/chatCore.js # open-sse/services/combo.js # src/app/(dashboard)/dashboard/profile/page.js # src/app/api/v1/models/route.js # src/lib/db/repos/settingsRepo.js
2026-08-17 00:21:51 +07:00
parent de9e00c66d 699edac327
commit 1256f29d92
103 changed files with 11250 additions and 3514 deletions
--- a/open-sse/config/providerModels.js
+++ b/open-sse/config/providerModels.js
@@ -2,7 +2,7 @@ import { PROVIDERS } from "./providers.js";
 import REGISTRY from "../providers/registry/index.js";
 // PROVIDER_MODELS now built from providers/registry (transport + models co-located)
 import { PROVIDER_MODELS } from "../providers/index.js";
-import { modelQuotaFamily, modelStrip, modelTargetFormat, normalizeModelId } from "../providers/models/schema.js";
+import { modelQuotaFamily, modelStrip, modelTargetFormat, modelSupportedFormats, normalizeModelId } from "../providers/models/schema.js";
 import { CODEX_REVIEW_SUFFIX } from "../providers/models/helpers.js";
 export { PROVIDER_MODELS };

@@ -54,6 +54,14 @@ export function getModelTargetFormat(aliasOrId, modelId) {
  return modelTargetFormat(findModel(models, modelId, aliasOrId));
 }

+// Declared upstream formats for a model (registry `supportedFormats`). Drives the
+// per-model guard on the sourceFormat-matched transport; null when undeclared.
+export function getModelSupportedFormats(aliasOrId, modelId) {
+  const models = PROVIDER_MODELS[aliasOrId];
+  if (!models) return null;
+  return modelSupportedFormats(findModel(models, modelId, aliasOrId));
+}
+
 export function getModelType(aliasOrId, modelId) {
  const models = PROVIDER_MODELS[aliasOrId];
  if (!models) return null;
--- a/open-sse/executors/antigravity.js
+++ b/open-sse/executors/antigravity.js
@@ -245,6 +245,18 @@ export class AntigravityExecutor extends BaseExecutor {
    // Strip tools/toolConfig (handled separately) and blacklisted fields that Google rejects
    const { tools: _originalTools, toolConfig: _originalToolConfig, ...requestWithoutTools } = body.request || {};
    stripBlacklisted(requestWithoutTools);
+    
+    // Rewrite competitive system prompts (e.g. Zed IDE's Claude prompt) to prevent Antigravity from 
+    // flagging the request and immediately blocking it with a 429 Quota Exhausted response.
+    if (requestWithoutTools.systemInstruction?.parts) {
+      const oldText = "You are a Claude agent, built on Anthropic's Claude Agent SDK.";
+      for (const part of requestWithoutTools.systemInstruction.parts) {
+        if (typeof part.text === "string" && part.text.includes(oldText)) {
+          part.text = part.text.split(oldText).join("");
+        }
+      }
+    }
+
    const generationConfig = { ...(requestWithoutTools.generationConfig || {}) };
    if (generationConfig.maxOutputTokens > MAX_ANTIGRAVITY_OUTPUT_TOKENS) {
      generationConfig.maxOutputTokens = MAX_ANTIGRAVITY_OUTPUT_TOKENS;
--- a/open-sse/executors/index.js
+++ b/open-sse/executors/index.js
@@ -10,7 +10,6 @@ import { CodexExecutor } from "./codex.js";
 import { CursorExecutor } from "./cursor.js";
 import { VertexExecutor } from "./vertex.js";
 import { OpenCodeExecutor } from "./opencode.js";
-import { OpenCodeGoExecutor } from "./opencode-go.js";
 import { GrokWebExecutor } from "./grok-web.js";
 import { GrokCliExecutor } from "./grok-cli.js";
 import { PerplexityWebExecutor } from "./perplexity-web.js";
@@ -41,7 +40,6 @@ const executors = {
  vertex: new VertexExecutor("vertex"),
  "vertex-partner": new VertexExecutor("vertex-partner"),
  opencode: new OpenCodeExecutor(),
-  "opencode-go": new OpenCodeGoExecutor(),
  "grok-web": new GrokWebExecutor(),
  "grok-cli": new GrokCliExecutor(),
  gcli: new GrokCliExecutor(), // Alias
@@ -86,7 +84,6 @@ export { CursorExecutor } from "./cursor.js";
 export { VertexExecutor } from "./vertex.js";
 export { DefaultExecutor } from "./default.js";
 export { OpenCodeExecutor } from "./opencode.js";
-export { OpenCodeGoExecutor } from "./opencode-go.js";
 export { GrokWebExecutor } from "./grok-web.js";
 export { GrokCliExecutor } from "./grok-cli.js";
 export { PerplexityWebExecutor } from "./perplexity-web.js";
--- a/open-sse/executors/kiro.js
+++ b/open-sse/executors/kiro.js
@@ -144,6 +144,12 @@ function normalizeStopReason(value) {
  return reason || null;
 }

+// Of the reasons stopDisposition() folds into "terminal_incomplete", only these
+// mean "usable as far as it got, then the budget ran out" -- the case
+// finish_reason "length" exists for. cancelled / pause_turn are abandoned turns
+// whose partial content must stay private, so they are deliberately absent.
+const KIRO_TRUNCATION_STOP_REASONS = new Set(["model_context_window_exceeded", "max_tokens"]);
+
 function stopDisposition(stopReason, hasToolCalls) {
  if (["malformed_model_output", "invalid_model_output"].includes(stopReason)) return "retryable_protocol_failure";
  if (["cancelled", "pause_turn", "model_context_window_exceeded"].includes(stopReason)) return "terminal_incomplete";
@@ -711,14 +717,25 @@ export class KiroExecutor extends BaseExecutor {
    };
    const emitTools = (controller) => {
      for (const tool of state.tools.values()) {
-        const input = parsedToolInput(tool);
-        if (tool.name === "tool_call") {
-          if (typeof input.name !== "string" || !input.name.trim()) {
-            throw new Error("Invalid Kiro tool_call payload: missing nested MCP tool name");
-          }
-          if (!Object.prototype.hasOwnProperty.call(input, "arguments")) {
-            throw new Error("Invalid Kiro tool_call payload: missing nested MCP tool arguments");
+        // Validate per tool, not per turn: one unusable fragment used to throw out
+        // of emitTools and take every other complete tool call in the same turn
+        // with it, which the client saw as a turn that answered nothing.
+        let input;
+        try {
+          input = parsedToolInput(tool);
+          if (tool.name === "tool_call") {
+            if (typeof input.name !== "string" || !input.name.trim()) {
+              throw new Error("Invalid Kiro tool_call payload: missing nested MCP tool name");
+            }
+            if (!Object.prototype.hasOwnProperty.call(input, "arguments")) {
+              throw new Error("Invalid Kiro tool_call payload: missing nested MCP tool arguments");
+            }
          }
+        } catch (error) {
+          state.droppedTools = (state.droppedTools || 0) + 1;
+          state.toolValidationError ||= error.message;
+          console.error(`[Kiro] dropping unusable tool call ${tool.id} (${tool.name}): ${error.message}`);
+          continue;
        }
        const index = state.toolCounter++;
        emitDelta(controller, {
@@ -729,14 +746,26 @@ export class KiroExecutor extends BaseExecutor {
            function: { name: tool.name, arguments: "" }
          }]
        });
+        const serializedInput = JSON.stringify(input);
        emitDelta(controller, {
-          tool_calls: [{ index, function: { arguments: JSON.stringify(input) } }]
+          tool_calls: [{ index, function: { arguments: serializedInput } }]
        });
+        // Tool arguments are billed output like any other completion bytes. They
+        // were never added to totalContentLength, so the /4 estimator in finish()
+        // reported OUT 0 -- or the Math.max floor of 1 -- for every turn whose
+        // entire answer was a tool call.
+        state.totalContentLength += tool.name.length + serializedInput.length;
        state.hasToolCalls = true;
      }
      state.tools.clear();
      state.bufferedToolBytes = 0;
-      if (state.stopReason === "tool_use" && !state.hasToolCalls) {
+      // A declared tool turn that emitted no usable call is only fatal when the
+      // turn produced nothing else. Throwing unconditionally here escaped
+      // emitTools() with provenance "invalid_tool_call", which the integrity gate
+      // re-derived into a repair retry -- discarding text the client had already
+      // been promised.
+      if (state.stopReason === "tool_use" && !state.hasToolCalls &&
+          !state.hasText && !state.hasReasoning && !state.hasCode) {
        throw new Error("Kiro tool_use stop reason did not include a complete tool call");
      }
    };
@@ -796,7 +825,6 @@ export class KiroExecutor extends BaseExecutor {
        emitDelta(controller, { content: event.payload.content });
      } else if (eventType === "toolUseEvent") {
        state.sawToolUse = true;
-        if (state.toolValidationError) return true;
        const values = Array.isArray(event.payload) ? event.payload : [event.payload];
        if (!values[0]) throw new Error("Kiro toolUseEvent is empty");
        for (const value of values) {
@@ -924,9 +952,10 @@ export class KiroExecutor extends BaseExecutor {
        } catch (error) {
          const bufferExceeded = error.code === "KIRO_BUFFER_EXCEEDED";
          if (!bufferExceeded) {
+            // Keep whatever is already buffered: the rejected fragment belongs to
+            // one tool, and clearing the map dropped the complete calls too.
            state.toolValidationError ||= error.message;
-            state.tools.clear();
-            state.bufferedToolBytes = 0;
+            console.error(`[Kiro] tool fragment rejected, keeping ${state.tools.size} buffered tool(s): ${error.message}`);
            continue;
          }
          fail(
@@ -958,7 +987,16 @@ export class KiroExecutor extends BaseExecutor {
      }
      state.transportState = "clean_eof";
      const declaredDisposition = stopDisposition(state.stopReason, state.sawToolUse);
-      if (["retryable_protocol_failure", "terminal_incomplete", "terminal_refusal", "unknown_failure"].includes(declaredDisposition)) {
+      // model_context_window_exceeded / max_tokens map to terminal_incomplete. When
+      // they arrive after the model already streamed content, fail() threw away a
+      // complete-enough answer; a truncated turn is what finish_reason "length" is
+      // for. chunkIndex > 0 means at least one delta already reached the client.
+      const declaredTruncatedAfterOutput = declaredDisposition === "terminal_incomplete" &&
+        KIRO_TRUNCATION_STOP_REASONS.has(state.stopReason) && state.chunkIndex > 0;
+      if (declaredTruncatedAfterOutput) {
+        console.error(`[Kiro] truncated after ${state.chunkIndex} chunk(s) (stop_reason=${state.stopReason}); keeping output`);
+      }
+      if (!declaredTruncatedAfterOutput && ["retryable_protocol_failure", "terminal_incomplete", "terminal_refusal", "unknown_failure"].includes(declaredDisposition)) {
        const code = declaredDisposition === "retryable_protocol_failure"
          ? "kiro_retryable_protocol_failure"
          : declaredDisposition === "terminal_refusal"
@@ -975,16 +1013,6 @@ export class KiroExecutor extends BaseExecutor {
        );
        return;
      }
-      if (state.toolValidationError) {
-        fail(
-          controller,
-          "invalid_tool_call",
-          "invalid_kiro_tool_call",
-          state.toolValidationError,
-          { transport_state: state.transportState, stop_disposition: "retryable_protocol_failure" }
-        );
-        return;
-      }
      try {
        emitTools(controller);
      } catch (error) {
@@ -997,6 +1025,22 @@ export class KiroExecutor extends BaseExecutor {
        );
        return;
      }
+      // Fail only when the turn has nothing usable left. emitTools() validates
+      // per tool and drops just the unusable ones, so this has to run AFTER it:
+      // before, the rejected tool was still buffered and tools.size was never 0.
+      // A turn that also produced text keeps that text -- the dropped call is
+      // logged, not fatal.
+      if (state.toolValidationError && !state.hasToolCalls &&
+          !state.hasText && !state.hasReasoning && !state.hasCode) {
+        fail(
+          controller,
+          "invalid_tool_call",
+          "invalid_kiro_tool_call",
+          state.toolValidationError,
+          { transport_state: state.transportState, stop_disposition: "retryable_protocol_failure" }
+        );
+        return;
+      }

      const hasOutput = state.hasText || state.hasReasoning || state.hasCode || state.hasToolCalls;
      if (!hasOutput && !state.explicitStop) {
@@ -1011,7 +1055,13 @@ export class KiroExecutor extends BaseExecutor {
      }

      const disposition = stopDisposition(state.stopReason, state.hasToolCalls);
-      if (["retryable_protocol_failure", "terminal_incomplete", "terminal_refusal", "unknown_failure"].includes(disposition)) {
+      // Same reasoning as declaredTruncatedAfterOutput above.
+      const truncatedAfterOutput = disposition === "terminal_incomplete" &&
+        KIRO_TRUNCATION_STOP_REASONS.has(state.stopReason) && state.chunkIndex > 0;
+      if (truncatedAfterOutput) {
+        console.error(`[Kiro] truncated after ${state.chunkIndex} chunk(s) (stop_reason=${state.stopReason}); closing as length`);
+      }
+      if (!truncatedAfterOutput && ["retryable_protocol_failure", "terminal_incomplete", "terminal_refusal", "unknown_failure"].includes(disposition)) {
        const code = disposition === "retryable_protocol_failure"
          ? "kiro_retryable_protocol_failure"
          : disposition === "terminal_refusal"
@@ -1041,18 +1091,24 @@ export class KiroExecutor extends BaseExecutor {
          total_tokens: prompt + completion
        };
      }
-      const finishReason = state.hasToolCalls
-        ? "tool_calls"
-        : disposition === "length"
-          ? "length"
-          : "stop";
+      const finishReason = truncatedAfterOutput
+        ? "length"
+        : state.hasToolCalls
+          ? "tool_calls"
+          : disposition === "length"
+            ? "length"
+            : "stop";
      controller.enqueue(sseChunk({}, finishReason, state.usage));
      controller.enqueue(encoder.encode(SSE_DONE));
      state.finished = true;
      options.onTerminalState?.(diagnostics({
        terminal_provenance: state.terminalProvenance || "clean_eventstream_eof",
        transport_state: state.transportState,
-        stop_disposition: disposition
+        // Report what this exit actually did, not the raw disposition. The
+        // integrity gate re-derives its verdict from stop_disposition, so
+        // reporting "terminal_incomplete" for a turn we deliberately kept made
+        // it discard the very bytes we just released to the client.
+        stop_disposition: truncatedAfterOutput ? "length" : disposition
      }));
    };

--- a/open-sse/executors/opencode-go.js
+++ b/open-sse/executors/opencode-go.js
@@ -1,49 +0,0 @@
-import { BaseExecutor } from "./base.js";
-import { PROVIDERS } from "../config/providers.js";
-import { injectReasoningContent } from "../utils/reasoningContentInjector.js";
-import { ANTHROPIC_API_VERSION } from "../providers/shared.js";
-
-// Models that use /zen/go/v1/messages (Anthropic/Claude format + x-api-key auth)
-const MESSAGES_FORMAT_MODELS = new Set([
-  "minimax-m3",
-  "minimax-m2.7",
-  "minimax-m2.5",
-  "qwen3.7-max",
-  "qwen3.7-plus",
-  "qwen3.6-plus",
-]);
-
-const BASE = "https://opencode.ai/zen/go/v1";
-
-export class OpenCodeGoExecutor extends BaseExecutor {
-  constructor() {
-    super("opencode-go", PROVIDERS["opencode-go"]);
-  }
-
-  // buildUrl runs before buildHeaders in BaseExecutor.execute, cache model here
-  buildUrl(model) {
-    this._lastModel = model;
-    return MESSAGES_FORMAT_MODELS.has(model)
-      ? `${BASE}/messages`
-      : `${BASE}/chat/completions`;
-  }
-
-  buildHeaders(credentials, stream = true) {
-    const key = credentials?.apiKey || credentials?.accessToken;
-    const headers = { "Content-Type": "application/json" };
-
-    if (MESSAGES_FORMAT_MODELS.has(this._lastModel)) {
-      headers["x-api-key"] = key;
-      headers["anthropic-version"] = ANTHROPIC_API_VERSION;
-    } else {
-      headers["Authorization"] = `Bearer ${key}`;
-    }
-
-    if (stream) headers["Accept"] = "text/event-stream";
-    return headers;
-  }
-
-  transformRequest(model, body) {
-    return injectReasoningContent({ provider: this.provider, model, body });
-  }
-}
--- a/open-sse/executors/opencode.js
+++ b/open-sse/executors/opencode.js
@@ -1,16 +1,43 @@
+import crypto from "crypto";
 import { BaseExecutor } from "./base.js";
 import { PROVIDERS } from "../config/providers.js";
 import { injectReasoningContent } from "../utils/reasoningContentInjector.js";
+import { resolveSessionId } from "../utils/sessionManager.js";

-// Models that use /zen/v1/messages (claude format)
+const OPENCODE_UA = "opencode";
 const MESSAGES_MODELS = new Set();

+function generateRequestId() {
+  return `msg_${crypto.randomUUID().replace(/-/g, "")}`;
+}
+
+function generateSessionId() {
+  return `ses_${crypto.randomUUID().replace(/-/g, "")}`;
+}
+
+// Normalize any resolved id into opencode's ses_ format (stable per-conversation)
+function toOpencodeSession(id) {
+  const stripped = String(id || "").replace(/^ses_/, "").replace(/-/g, "");
+  return stripped ? `ses_${stripped}` : null;
+}
+
+function resolveOpencodeSession(body, credentials) {
+  return toOpencodeSession(resolveSessionId({
+    headers: credentials?.rawHeaders,
+    body,
+    connectionId: credentials?.connectionId,
+    scope: "opencode",
+  }));
+}
+
 export class OpenCodeExecutor extends BaseExecutor {
  constructor() {
    super("opencode", PROVIDERS.opencode);
+    this._currentSessionId = null;
  }

-  transformRequest(model, body) {
+  transformRequest(model, body, stream, credentials) {
+    this._currentSessionId = resolveOpencodeSession(body, credentials);
    return injectReasoningContent({ provider: this.provider, model, body });
  }

@@ -21,12 +48,23 @@ export class OpenCodeExecutor extends BaseExecutor {
      : `${base}/zen/v1/chat/completions`;
  }

-  buildHeaders() {
+  buildHeaders(credentials, stream = true) {
+    const raw = credentials?.rawHeaders || {};
+    const lower = {};
+    for (const [k, v] of Object.entries(raw)) lower[k.toLowerCase()] = v;
+
+    const downstreamUa = lower["user-agent"] || "";
+    const isOpencodeDownstream = downstreamUa.toLowerCase().includes("opencode");
+
    return {
      "Content-Type": "application/json",
      "Authorization": "Bearer public",
-      "x-opencode-client": "desktop",
-      "Accept": "text/event-stream"
+      "User-Agent": isOpencodeDownstream ? downstreamUa : OPENCODE_UA,
+      "x-opencode-client": lower["x-opencode-client"] || "desktop",
+      "x-opencode-session": lower["x-opencode-session"] || this._currentSessionId || generateSessionId(),
+      "x-opencode-request": lower["x-opencode-request"] || generateRequestId(),
+      "x-opencode-project": lower["x-opencode-project"] || "global",
+      "Accept": stream ? "text/event-stream" : "*/*",
    };
  }
 }
--- a/open-sse/executors/qoder.js
+++ b/open-sse/executors/qoder.js
@@ -216,6 +216,52 @@ async function buildQoderRequestBody({ model, body, credentials, log, proxyOptio
  };
 }

+/**
+ * Check if a qoder error message indicates a billing/quota block.
+ * Signatures: code 112 (quota exhausted), code 10605 (queue throttle), pricingUrl field.
+ */
+function isBillingBlock(inner) {
+  if (!inner || typeof inner !== "string") return false;
+  const lowerMsg = inner.toLowerCase();
+  // Match: {"code":"112",...}, {"code":"10605",...}, or pricingUrl field
+  return /\"code\"\s*:\s*\"(112|10605)\"/.test(inner) || lowerMsg.includes("pricingurl");
+}
+
+/**
+ * Peek the first SSE frame to detect billing errors before piping.
+ * Returns { isBilling, statusVal, message, consumed } — `consumed` is every
+ * byte read so far (including the peeked line) so the caller can re-process
+ * it and nothing is dropped from the stream.
+ */
+async function peekFirstQoderFrame(reader, decoder) {
+  let consumed = "";
+  while (true) {
+    const { done, value } = await reader.read();
+    if (done) return { isBilling: false, consumed, upstreamDone: true };
+
+    consumed += decoder.decode(value, { stream: true });
+    const nl = consumed.indexOf("\n");
+    if (nl === -1) continue; // need a full line first
+
+    const line = consumed.slice(0, nl).replace(/\r$/, "").trim();
+    if (!line.startsWith("data:")) continue;
+
+    const data = line.slice(5).trimStart();
+    if (data === "[DONE]") return { isBilling: false, consumed };
+
+    let envelope;
+    try { envelope = JSON.parse(data); } catch { return { isBilling: false, consumed }; }
+
+    const statusVal = typeof envelope.statusCodeValue === "number" ? envelope.statusCodeValue : 200;
+    const inner = typeof envelope.body === "string" ? envelope.body : "";
+
+    if (statusVal !== 200 && isBillingBlock(inner)) {
+      return { isBilling: true, statusVal, message: inner || `qoder billing block (${statusVal})` };
+    }
+    return { isBilling: false, consumed };
+  }
+}
+
 /**
 * Wrap the upstream's `{statusCodeValue, body}` SSE envelope into plain
 * OpenAI SSE chunks the rest of the chatCore pipeline understands.
@@ -230,16 +276,34 @@ async function buildQoderRequestBody({ model, body, credentials, log, proxyOptio
 * [DONE]/error frame (agent keepalive). Non-streaming clients drain via
 * response.text() which hangs until the socket closes — so on terminal
 * events we cancel the upstream reader and close our stream immediately.
+ *
+ * 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.
 */
-function wrapQoderSSE(response, model) {
+async function wrapQoderSSE(response, model) {
  if (!response.ok || !response.body) return response;

  const decoder = new TextDecoder();
-  const encoder = new TextEncoder();
-  let buffer = "";
-  let doneEmitted = false;
  const reader = response.body.getReader();

+  // Peek first frame to detect billing block
+  const peek = await peekFirstQoderFrame(reader, decoder);
+  if (peek?.isBilling) {
+    // Billing block detected — return 403 so chatCore fails this connection
+    await reader.cancel().catch(() => {});
+    return new Response(
+      JSON.stringify({ error: { message: peek.message, code: peek.statusVal } }),
+      { status: 403, headers: { "Content-Type": "application/json" } }
+    );
+  }
+
+  // Normal flow: re-process every byte the peek consumed, then continue.
+  let buffer = peek.consumed || "";
+  const upstreamDrained = peek.upstreamDone === true;
+  const encoder = new TextEncoder();
+  let doneEmitted = false;
+
  // Process one already-extracted SSE line (no trailing newline).
  const processLine = (line, controller) => {
    const trimmed = line.replace(/\r$/, "").trim();
@@ -288,7 +352,28 @@ function wrapQoderSSE(response, model) {
    // enqueueing would never be re-invoked, hanging consumers like .text().
    async start(controller) {
      try {
-        while (!doneEmitted) {
+        // Drain whatever the peek already pulled off the socket first.
+        let nlSeed;
+        while ((nlSeed = buffer.indexOf("\n")) !== -1) {
+          const line = buffer.slice(0, nlSeed);
+          buffer = buffer.slice(nlSeed + 1);
+          processLine(line, controller);
+          if (doneEmitted) {
+            await reader.cancel().catch(() => {});
+            controller.close();
+            return;
+          }
+        }
+        if (upstreamDrained) {
+          // Peek hit end-of-stream: flush any trailing partial line.
+          buffer += decoder.decode();
+          if (buffer.length > 0) {
+            processLine(buffer, controller);
+            buffer = "";
+          }
+        }
+
+        while (!doneEmitted && !upstreamDrained) {
          const { done, value } = await reader.read();
          if (done) {
            buffer += decoder.decode();
@@ -473,7 +558,7 @@ export class QoderExecutor extends BaseExecutor {
      return { response, url, headers, transformedBody: payload };
    }

-    const wrapped = wrapQoderSSE(response, `qoder/${qoderKey}`);
+    const wrapped = await wrapQoderSSE(response, `qoder/${qoderKey}`);
    return { response: wrapped, url, headers, transformedBody: payload };
  }

@@ -497,4 +582,5 @@ export const __test__ = {
  normalizeMessages,
  wrapQoderSSE,
  buildQoderRequestBody,
+  isBillingBlock,
 };
--- a/open-sse/handlers/chatCore.js
+++ b/open-sse/handlers/chatCore.js
--- a/open-sse/handlers/chatCore/requestDetail.js
+++ b/open-sse/handlers/chatCore/requestDetail.js
@@ -44,13 +44,14 @@ export function extractUsageFromResponse(responseBody) {
    };
  }

-  // Gemini format
-  if (responseBody.usageMetadata) {
+  // Gemini format. Antigravity / gemini-cli wrap the payload in { response: {...} }.
+  const usageMetadata = responseBody.usageMetadata || responseBody.response?.usageMetadata;
+  if (usageMetadata) {
    return {
-      prompt_tokens: responseBody.usageMetadata.promptTokenCount || 0,
-      completion_tokens: responseBody.usageMetadata.candidatesTokenCount || 0,
-      cached_tokens: responseBody.usageMetadata.cachedContentTokenCount || 0,
-      reasoning_tokens: responseBody.usageMetadata.thoughtsTokenCount || 0
+      prompt_tokens: usageMetadata.promptTokenCount || 0,
+      completion_tokens: usageMetadata.candidatesTokenCount || 0,
+      cached_tokens: usageMetadata.cachedContentTokenCount || 0,
+      reasoning_tokens: usageMetadata.thoughtsTokenCount || 0
    };
  }

--- a/open-sse/handlers/search/callers.js
+++ b/open-sse/handlers/search/callers.js
@@ -29,6 +29,8 @@
 * @property {Record<string,unknown>} [providerSpecificData]
 */

+import { assertPublicUrl } from "../../../src/shared/utils/ssrfGuard.js";
+
 // ── Helpers ─────────────────────────────────────────────────────────────

 /**
@@ -63,12 +65,31 @@ export function getProviderSetting(params, key) {

 /**
 * Resolve base URL with optional override from providerOptions.baseUrl.
+ *
+ * The override is client-controlled and therefore SSRF-hardened: only public
+ * http(s) URLs are accepted (internal/private/loopback/metadata addresses are
+ * rejected via assertPublicUrl). The provider's own configured baseUrl is
+ * trusted as-is (admin-controlled).
+ *
 * @param {SearchProviderConfig} config
 * @param {SearchRequestParams} params
 * @returns {string}
 */
 export function resolveBaseUrl(config, params) {
  const override = getProviderSetting(params, "baseUrl");
+  if (override) {
+    // SSRF guard: client-supplied base URLs must be public http(s) only.
+    let parsed;
+    try {
+      parsed = new URL(override);
+    } catch {
+      throw new Error(`Invalid baseUrl: ${override}`);
+    }
+    if (parsed.protocol !== "http:" && parsed.protocol !== "https:") {
+      throw new Error(`Invalid baseUrl protocol: ${parsed.protocol}`);
+    }
+    assertPublicUrl(override);
+  }
  return (override || config.baseUrl).replace(/\/+$/, "");
 }

--- a/open-sse/handlers/ttsProviders/genericFormats.js
+++ b/open-sse/handlers/ttsProviders/genericFormats.js
@@ -51,6 +51,25 @@ async function huggingface({ baseUrl, apiKey, text, modelId }) {
  return responseToBase64(res, "wav");
 }

+// Fish Audio: model travels in an HTTP header, the voice is a reference_id, returns binary
+async function fishAudio({ baseUrl, apiKey, text, modelId, voiceId }) {
+  const res = await fetch(baseUrl, {
+    method: "POST",
+    headers: {
+      "Content-Type": "application/json",
+      "Authorization": `Bearer ${apiKey}`,
+      "model": modelId || "s2.1-pro-free",
+    },
+    body: JSON.stringify({
+      text,
+      format: "mp3",
+      ...(voiceId ? { reference_id: voiceId } : {}),
+    }),
+  });
+  if (!res.ok) await throwUpstreamError(res);
+  return responseToBase64(res, "mp3");
+}
+
 // Inworld: Basic auth, JSON { audioContent }
 async function inworld({ baseUrl, apiKey, text, modelId, voiceId }) {
  const res = await fetch(baseUrl, {
@@ -166,4 +185,5 @@ export const FORMAT_HANDLERS = {
  tortoise,
  openai: openaiCompat,
  "minimax-tts": minimaxTts,
+  "fish-audio": fishAudio,
 };
--- a/open-sse/providers/capabilities.js
+++ b/open-sse/providers/capabilities.js
@@ -205,6 +205,7 @@ export const PATTERN_CAPABILITIES = [

  // ── Gemini (all 2.0+ multimodal + google_search grounding, 1M ctx) ─
  { pattern: "*gemini*image*",  caps: { vision: true, imageOutput: true, contextWindow: 1048576 } },
+  { pattern: "*gemini-3.7*",    caps: { vision: true, audioInput: true, videoInput: true, reasoning: true, search: true, thinkingFormat: "gemini-level", thinkingCanDisable: false, contextWindow: 1048576, maxOutput: 65536 } },
  { pattern: "*gemini-3*pro*",  caps: { vision: true, audioInput: true, videoInput: true, reasoning: true, search: true, thinkingFormat: "gemini-level", thinkingCanDisable: false, contextWindow: 1048576, maxOutput: 65535 } },
  { pattern: "*gemini-3*",      caps: { vision: true, audioInput: true, videoInput: true, reasoning: true, search: true, thinkingFormat: "gemini-level", thinkingCanDisable: false, contextWindow: 1048576, maxOutput: 65536 } },
  { pattern: "*gemini-2.5*",    caps: { vision: true, audioInput: true, videoInput: true, reasoning: true, search: true, thinkingFormat: "gemini-budget", thinkingRange: { min: 0, max: 24576 }, contextWindow: 1048576, maxOutput: 65536 } },
--- a/open-sse/providers/models/schema.js
+++ b/open-sse/providers/models/schema.js
@@ -38,3 +38,11 @@ export function modelStrip(model) {
 export function modelTargetFormat(model) {
  return model?.targetFormat || MODEL_DEFAULTS.targetFormat;
 }
+
+// Per-model declared upstream formats (e.g. ["openai", "claude"]). Guards the
+// sourceFormat-matched transport for multi-endpoint providers whose models differ
+// in endpoint support (opencode-go: kimi/glm only do /chat/completions, minimax/qwen
+// also do /messages, deepseek also does /responses).
+export function modelSupportedFormats(model) {
+  return model?.supportedFormats || null;
+}
--- a/open-sse/providers/pricing.js
+++ b/open-sse/providers/pricing.js
@@ -57,6 +57,10 @@ export const MODEL_PRICING = {
  "o1-mini":                      { input: 3.00,  output: 12.00, cached: 1.50,  reasoning: 18.00,  cache_creation: 3.00  },

  // === Gemini ===
+  "gemini-3.7-flash":              { input: 1.50,  output: 7.50,  cached: 0.15,  reasoning: 11.25,  cache_creation: 1.875 },
+  "gemini-3.7-flash-high":         { input: 1.50,  output: 7.50,  cached: 0.15,  reasoning: 11.25,  cache_creation: 1.875 },
+  "gemini-3.7-flash-medium":       { input: 1.50,  output: 7.50,  cached: 0.15,  reasoning: 11.25,  cache_creation: 1.875 },
+  "gemini-3.7-flash-low":          { input: 1.50,  output: 7.50,  cached: 0.15,  reasoning: 11.25,  cache_creation: 1.875 },
  "gemini-3.6-flash":              { input: 1.50,  output: 7.50,  cached: 0.15,  reasoning: 11.25,  cache_creation: 1.875 },
  "gemini-3.6-flash-high":         { input: 1.50,  output: 7.50,  cached: 0.15,  reasoning: 11.25,  cache_creation: 1.875 },
  "gemini-3.6-flash-medium":       { input: 1.50,  output: 7.50,  cached: 0.15,  reasoning: 11.25,  cache_creation: 1.875 },
--- a/open-sse/providers/registry/alitp-intl.js
+++ b/open-sse/providers/registry/alitp-intl.js
@@ -0,0 +1,35 @@
+// Token Plan — credit subscription keys on token-plan.<region>.maas.aliyuncs.com.
+// Fourth Alibaba key type: Coding Plan (alicode/alicode-intl) and Model Studio
+// (alims-intl) both reject these keys, and they reject Model Studio keys back.
+// Singapore is the only region that serves the plan; eu-central-1 answers
+// IllegalEndpoint. The Anthropic surface (/apps/anthropic/v1/messages) is not
+// authorized for this plan, so OpenAI-compatible mode is the only transport.
+export default {
+  id: "alitp-intl",
+  priority: 11,
+  alias: "alitp-intl",
+  display: {
+    name: "Alibaba Token Plan",
+    icon: "cloud",
+    color: "#FF6A00",
+    textIcon: "ATP",
+    website: "https://www.alibabacloud.com/campaign/ai-landing-page-token",
+    notice: {
+      apiKeyUrl: "https://modelstudio.console.alibabacloud.com/?apiKey=1",
+    },
+  },
+  category: "apikey",
+  transport: {
+    baseUrl: "https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1/chat/completions",
+    headers: {},
+    quirks: { preserveCacheControl: true },
+  },
+  models: [
+    { id: "qwen3.8-max-preview", name: "Qwen3.8 Max Preview" },
+    { id: "qwen3.7-max", name: "Qwen3.7 Max" },
+    { id: "qwen3.7-plus", name: "Qwen3.7 Plus" },
+    { id: "qwen3.6-flash", name: "Qwen3.6 Flash" },
+    { id: "glm-5.2", name: "GLM 5.2" },
+    { id: "deepseek-v4-pro", name: "DeepSeek V4 Pro" },
+  ],
+};
--- a/open-sse/providers/registry/antigravity.js
+++ b/open-sse/providers/registry/antigravity.js
@@ -45,6 +45,9 @@ export default {
    clientSecret: "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",
  },
  models: [
+    { id: "gemini-3.7-flash-high", name: "Gemini 3.7 Flash (High)", upstreamModelId: "gemini-3.7-flash-tiered(high)" },
+    { id: "gemini-3.7-flash-medium", name: "Gemini 3.7 Flash (Medium)", upstreamModelId: "gemini-3.7-flash-tiered(medium)" },
+    { id: "gemini-3.7-flash-low", name: "Gemini 3.7 Flash (Low)", upstreamModelId: "gemini-3.7-flash-tiered(low)" },
    { id: "gemini-3.6-flash-high", name: "Gemini 3.6 Flash (High)", upstreamModelId: "gemini-3.6-flash-tiered(high)" },
    { id: "gemini-3.6-flash-medium", name: "Gemini 3.6 Flash (Medium)", upstreamModelId: "gemini-3.6-flash-tiered(medium)" },
    { id: "gemini-3.6-flash-low", name: "Gemini 3.6 Flash (Low)", upstreamModelId: "gemini-3.6-flash-tiered(low)" },
--- a/open-sse/providers/registry/fish-audio.js
+++ b/open-sse/providers/registry/fish-audio.js
@@ -0,0 +1,31 @@
+// Fish Audio TTS — the model id travels in an HTTP `model` header rather than the
+// JSON body, and the voice is a reference_id (a cloned or preset voice model).
+export default {
+  id: "fish-audio",
+  alias: "fish",
+  display: {
+    name: "Fish Audio",
+    icon: "record_voice_over",
+    color: "#1E9BF0",
+    textIcon: "FA",
+    website: "https://fish.audio",
+    notice: {
+      apiKeyUrl: "https://fish.audio/app/api-keys/",
+    },
+  },
+  category: "apikey",
+  authType: "apikey",
+  serviceKinds: ["tts"],
+  ttsConfig: {
+    baseUrl: "https://api.fish.audio/v1/tts",
+    authType: "apikey",
+    authHeader: "bearer",
+    format: "fish-audio",
+    models: [
+      { id: "s2.1-pro-free", name: "S2.1 Pro Free" },
+      { id: "s2.1-pro", name: "S2.1 Pro" },
+      { id: "s2-pro", name: "S2 Pro" },
+      { id: "s1", name: "S1" },
+    ],
+  },
+};
--- a/open-sse/providers/registry/gemini.js
+++ b/open-sse/providers/registry/gemini.js
@@ -36,6 +36,7 @@ export default {
    },
  },
  models: [
+    { id: "gemini-3.7-flash", name: "Gemini 3.7 Flash" },
    { id: "gemini-3.6-flash", name: "Gemini 3.6 Flash" },
    { id: "gemini-3.5-flash-lite", name: "Gemini 3.5 Flash Lite" },
    { id: "gemini-3.1-pro-preview", name: "Gemini 3.1 Pro Preview" },
--- a/open-sse/providers/registry/glm-cn.js
+++ b/open-sse/providers/registry/glm-cn.js
@@ -21,6 +21,7 @@ export default {
    },
  },
  models: [
+    { id: "glm-5.3", name: "GLM 5.3" },
    { id: "glm-5.2", name: "GLM 5.2" },
    { id: "glm-5.1", name: "GLM 5.1" },
    { id: "glm-5", name: "GLM 5" },
--- a/open-sse/providers/registry/glm.js
+++ b/open-sse/providers/registry/glm.js
@@ -45,6 +45,7 @@ export default {
    },
  ],
  models: [
+    { id: "glm-5.3", name: "GLM 5.3" },
    { id: "glm-5.2", name: "GLM 5.2" },
    { id: "glm-5.1", name: "GLM 5.1" },
    { id: "glm-5", name: "GLM 5" },
--- a/open-sse/providers/registry/index.js
+++ b/open-sse/providers/registry/index.js
@@ -119,6 +119,8 @@ import p116 from "./tokenrouter.js";
 import p117 from "./selfhosted-stt.js";
 import p118 from "./selfhosted-tts.js";
 import p119 from "./selfhosted-embedding.js";
+import p120 from "./fish-audio.js";
+import p121 from "./alitp-intl.js";

 export default [
  p0,
@@ -239,4 +241,6 @@ export default [
  p117,
  p118,
  p119,
+  p120,
+  p121,
 ];
--- a/open-sse/providers/registry/kimchi.js
+++ b/open-sse/providers/registry/kimchi.js
@@ -14,7 +14,7 @@ export default {
    },
  },
  category: "freeTier",
-  authModes: ["oauth"],
+  authModes: ["oauth", "apikey"],
  hasOAuth: true,
  transport: {
    baseUrl: "https://llm.kimchi.dev/openai/v1/chat/completions",
--- a/open-sse/providers/registry/opencode-go.js
+++ b/open-sse/providers/registry/opencode-go.js
@@ -22,20 +22,28 @@ export default {
    baseUrl: "https://opencode.ai/zen/go/v1/chat/completions",
    headers: {},
  },
+  // Multi-endpoint: pick the transport matching the client sourceFormat to skip
+  // translation. Guarded per-model by `supportedFormats` (see chatCore) because
+  // opencode-go models differ in endpoint support.
+  transports: [
+    { format: "openai", baseUrl: "https://opencode.ai/zen/go/v1/chat/completions", auth: { combined: true, header: "Authorization", scheme: "bearer" } },
+    { 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" } },
+  ],
  models: [
-    { id: "glm-5.2", name: "GLM 5.2" },
-    { id: "glm-5.1", name: "GLM 5.1" },
-    { id: "kimi-k2.7-code", name: "Kimi K2.7 Code" },
-    { id: "kimi-k2.6", name: "Kimi K2.6" },
-    { id: "deepseek-v4-pro", name: "DeepSeek V4 Pro" },
-    { id: "deepseek-v4-flash", name: "DeepSeek V4 Flash" },
-    { id: "mimo-v2.5", name: "MiMo V2.5" },
-    { id: "mimo-v2.5-pro", name: "MiMo V2.5 Pro" },
-    { id: "minimax-m3", name: "MiniMax M3", targetFormat: "claude" },
-    { id: "minimax-m2.7", name: "MiniMax M2.7", targetFormat: "claude" },
-    { id: "minimax-m2.5", name: "MiniMax M2.5", targetFormat: "claude" },
-    { id: "qwen3.7-max", name: "Qwen 3.7 Max", targetFormat: "claude" },
-    { id: "qwen3.7-plus", name: "Qwen 3.7 Plus", targetFormat: "claude" },
-    { id: "qwen3.6-plus", name: "Qwen 3.6 Plus", targetFormat: "claude" },
+    { 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: "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: "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: "qwen3.7-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"] },
  ],
 };
--- a/open-sse/services/combo.js
+++ b/open-sse/services/combo.js
--- a/open-sse/services/usage.js
+++ b/open-sse/services/usage.js
@@ -35,7 +35,7 @@ const USAGE_HANDLERS = {
  github: (c) => getGitHubUsage(c.accessToken, c.providerSpecificData, c.proxyOptions),
  "gemini-cli": (c) => getGeminiUsage(c.accessToken, c.providerDataWithProjectId, c.proxyOptions),
  antigravity: (c) => getAntigravityUsage(c.accessToken, c.providerSpecificData, c.proxyOptions),
-  claude: (c) => getClaudeUsage(c.accessToken, c.proxyOptions),
+  claude: (c) => getClaudeUsage(c.accessToken, c.proxyOptions, { force: c.force }),
  codex: (c) => getCodexUsage(c.accessToken, c.proxyOptions),
  kiro: (c) => getKiroUsage(c.accessToken, c.providerSpecificData, c.proxyOptions),
  qoder: async (c) => {
@@ -60,7 +60,7 @@ const USAGE_HANDLERS = {
  commandcode: (c) => getCommandCodeUsage(c.apiKey, c.proxyOptions),
 };

-export async function getUsageForProvider(connection, proxyOptions = null) {
+export async function getUsageForProvider(connection, proxyOptions = null, options = {}) {
  const { provider, accessToken, apiKey, providerSpecificData, projectId } = connection;
  const providerDataWithProjectId = {
    ...(providerSpecificData || {}),
@@ -69,5 +69,13 @@ export async function getUsageForProvider(connection, proxyOptions = null) {

  const handler = USAGE_HANDLERS[provider];
  if (!handler) return { message: `Usage API not implemented for ${provider}` };
-  return await handler({ provider, accessToken, apiKey, providerSpecificData, providerDataWithProjectId, proxyOptions });
+  return await handler({
+    provider,
+    accessToken,
+    apiKey,
+    providerSpecificData,
+    providerDataWithProjectId,
+    proxyOptions,
+    force: options.force === true,
+  });
 }
--- a/open-sse/services/usage/claude.js
+++ b/open-sse/services/usage/claude.js
@@ -19,7 +19,43 @@ const CLAUDE_CONFIG = {
 const OAUTH_429_COOLDOWN_MS = 180000;
 const oauthCooldown = new Map();

-export async function getClaudeUsage(accessToken, proxyOptions = null) {
+// Dedup + short TTL cache per access token. Many tabs / many accounts / auto-refresh
+// all funnel through here; without this each call hits Anthropic and triggers 429.
+const USAGE_CACHE_TTL_MS = 300000;
+const usageCache = new Map(); // token -> { promise } | { result, expiresAt }
+
+export async function getClaudeUsage(accessToken, proxyOptions = null, options = {}) {
+  const force = options?.force === true;
+
+  // Serve in-flight or fresh cached result (skip on manual force)
+  if (!force && accessToken) {
+    const hit = usageCache.get(accessToken);
+    if (hit?.promise) return hit.promise;
+    if (hit && hit.expiresAt > Date.now()) return hit.result;
+  }
+
+  const stale = (!force && accessToken && usageCache.get(accessToken)?.result) || null;
+
+  const promise = (async () => {
+    const result = await fetchClaudeUsageRaw(accessToken, proxyOptions);
+    // Only cache real quota data, not soft-failure {message: ...} payloads
+    if (accessToken && result?.quotas) {
+      usageCache.set(accessToken, {
+        result,
+        expiresAt: Date.now() + USAGE_CACHE_TTL_MS,
+      });
+      return result;
+    }
+    // Soft failure (429/error): prefer the last good read over a transient error
+    if (stale) return stale;
+    return result;
+  })();
+
+  if (accessToken) usageCache.set(accessToken, { promise });
+  return promise;
+}
+
+async function fetchClaudeUsageRaw(accessToken, proxyOptions = null) {
  try {
    // Skip OAuth usage call while this token is cooling down from a recent 429
    const cooldownUntil = oauthCooldown.get(accessToken);
--- a/open-sse/services/usage/google.js
+++ b/open-sse/services/usage/google.js
@@ -161,6 +161,9 @@ export async function getAntigravityUsage(accessToken, providerSpecificData, pro
    if (data.models) {
      // Filter only recommended/important models (must match PROVIDER_MODELS ag ids)
      const importantModels = [
+        'gemini-3.7-flash-high',
+        'gemini-3.7-flash-medium',
+        'gemini-3.7-flash-low',
        'gemini-3.6-flash-high',
        'gemini-3.6-flash-medium',
        'gemini-3.6-flash-low',
--- a/open-sse/translator/concerns/modality.js
+++ b/open-sse/translator/concerns/modality.js
@@ -62,6 +62,19 @@ function stripOpenAI(body, caps) {
  if (!Array.isArray(body.messages)) return;
  const last = body.messages.length - 1;
  body.messages.forEach((msg, i) => {
+    if (caps.vision === false) {
+      if (Array.isArray(msg.images)) delete msg.images;
+      if (Array.isArray(msg.experimental_attachments)) {
+        msg.experimental_attachments = msg.experimental_attachments.filter(
+          (a) => !(a?.contentType?.startsWith("image/") || (typeof a?.url === "string" && a.url.startsWith("data:image/")))
+        );
+      }
+      if (Array.isArray(msg.attachments)) {
+        msg.attachments = msg.attachments.filter(
+          (a) => !(a?.contentType?.startsWith("image/") || (typeof a?.url === "string" && a.url.startsWith("data:image/")))
+        );
+      }
+    }
    if (!Array.isArray(msg.content)) return;
    const removed = new Set();
    msg.content = filterBlocks(msg.content, capForOpenAIBlock, caps, removed, i === last);
--- a/open-sse/translator/formats/claude.js
+++ b/open-sse/translator/formats/claude.js
@@ -9,6 +9,9 @@ import { PROVIDERS } from "../../providers/index.js";
 import { getCapabilitiesForModel } from "../../providers/capabilities.js";
 import { DEFAULT_MAX_TOKENS } from "../../config/runtimeConfig.js";

+const CACHE_CONTROL_5M = { type: "ephemeral" };
+const CACHE_CONTROL_1H = { type: "ephemeral", ttl: "1h" };
+
 // Check if message has valid non-empty content
 export function hasValidContent(msg) {
  if (typeof msg.content === "string" && msg.content.trim()) return true;
@@ -124,32 +127,38 @@ export function normalizeClaudePassthrough(body, model = "") {
    if (Object.keys(body.output_config).length === 0) delete body.output_config;
  }

-  // 2. Hoist mid-conversation system messages into the top-level system field
+  // 2. 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.
  if (Array.isArray(body.messages)) {
-    const systemBlocks = [];
    const messages = [];
    for (const msg of body.messages) {
-      if (msg.role === ROLE.SYSTEM) {
-        const text = typeof msg.content === "string"
-          ? msg.content
-          : Array.isArray(msg.content)
-            ? msg.content.map(b => (typeof b === "string" ? b : b?.text || "")).join("\n")
-            : "";
-        if (text.trim()) systemBlocks.push({ type: CLAUDE_BLOCK.TEXT, text });
+      if (msg.role !== ROLE.SYSTEM) {
+        messages.push(msg);
        continue;
      }
-      messages.push(msg);
-    }
+      const text = typeof msg.content === "string"
+        ? msg.content
+        : Array.isArray(msg.content)
+          ? msg.content.map(b => (typeof b === "string" ? b : b?.text || "")).join("\n")
+          : "";
+      if (!text.trim()) continue;

-    if (systemBlocks.length > 0) {
-      const existing = Array.isArray(body.system)
-        ? body.system
-        : typeof body.system === "string" && body.system.trim()
-          ? [{ type: "text", text: body.system }]
-          : [];
-      body.system = [...existing, ...systemBlocks];
-      body.messages = messages;
+      // Copy-on-write: the caller's body is reused across account-fallback
+      // attempts, so folding must never mutate the original message.
+      const block = { type: CLAUDE_BLOCK.TEXT, text };
+      const prev = messages[messages.length - 1];
+      if (prev?.role === ROLE.USER) {
+        const content = typeof prev.content === "string"
+          ? [{ type: CLAUDE_BLOCK.TEXT, text: prev.content }]
+          : Array.isArray(prev.content) ? [...prev.content] : [];
+        messages[messages.length - 1] = { ...prev, content: [...content, block] };
+        continue;
+      }
+      messages.push({ role: ROLE.USER, content: [block] });
    }
+    body.messages = messages;
  }

  // 3. Drop thinking blocks whose signature is not Claude's (combo mixes models,
@@ -182,6 +191,70 @@ export function normalizeClaudePassthrough(body, model = "") {
  return body;
 }

+// Put a 5m breakpoint on the last cache-eligible block of a message.
+// thinking/redacted_thinking blocks do not accept cache_control.
+function markLastCacheableBlock(msg) {
+  if (!Array.isArray(msg?.content)) return false;
+  for (let i = msg.content.length - 1; i >= 0; i--) {
+    const block = msg.content[i];
+    if (typeof block !== "object" || block === null) continue;
+    if (block.type === CLAUDE_BLOCK.THINKING || block.type === CLAUDE_BLOCK.REDACTED_THINKING) continue;
+    block.cache_control = { ...CACHE_CONTROL_5M };
+    return true;
+  }
+  return false;
+}
+
+// Re-anchor cache breakpoints on a Claude passthrough body (same policy as
+// prepareClaudeRequest): last tool + last system block at 1h, last assistant at 5m.
+// The client's own markers point at pre-normalization offsets, so they are dropped.
+// Must run LAST, after every step that can reshape system/tools/messages
+// (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.system)) {
+    const last = body.system.length - 1;
+    body.system.forEach((block, i) => {
+      if (typeof block !== "object" || block === null) return;
+      if (i === last) block.cache_control = { ...CACHE_CONTROL_1H };
+      else delete block.cache_control;
+    });
+  }
+
+  if (Array.isArray(body.tools)) {
+    const last = body.tools.length - 1;
+    body.tools.forEach((tool, i) => {
+      if (i === last) tool.cache_control = { ...CACHE_CONTROL_1H };
+      else delete tool.cache_control;
+    });
+  }
+
+  if (Array.isArray(body.messages)) {
+    let anchored = null;
+    for (let i = body.messages.length - 1; i >= 0; i--) {
+      const msg = body.messages[i];
+      if (!Array.isArray(msg.content)) continue;
+      for (const block of msg.content) delete block.cache_control;
+
+      // Prefer the last assistant turn: it ends a completed exchange, so the
+      // prefix up to it stays byte-stable across the following requests.
+      if (anchored || msg.role !== ROLE.ASSISTANT) continue;
+      anchored = markLastCacheableBlock(msg);
+    }
+
+    // First turn of a conversation has no assistant yet — anchor the final
+    // message instead, so the opening prompt is cached rather than paid twice.
+    if (!anchored) {
+      for (let i = body.messages.length - 1; i >= 0 && !anchored; i--) {
+        anchored = markLastCacheableBlock(body.messages[i]);
+      }
+    }
+  }
+
+  return body;
+}
+
 // Prepare request for Claude format endpoints
 // - Cleanup cache_control
 // - Filter empty messages
--- a/open-sse/translator/request/claude-to-kiro.js
+++ b/open-sse/translator/request/claude-to-kiro.js
@@ -287,6 +287,18 @@ export function claudeToKiroRequest(model, body, stream, credentials) {
    toolSpecs,
    nameMap,
  });
+  // canonicalizeKiroConversation() already ran its second-chance repair (flatten
+  // every structured tool turn to text, then re-validate). A body that is STILL
+  // invalid here cannot be made shippable, and Kiro answers it with
+  // 400 {"message":"Improperly formed request.","reason":"REQUEST_BODY_INVALID"}.
+  // Fail locally instead: chatCore turns a falsy return into a 400 without
+  // spending an upstream call or a per-account cooldown. The taxonomy
+  // (role:N | pair:N | id:N | spec:N | orphan:0 | current) names the offending
+  // turn so the shape can be diagnosed from the log alone.
+  if (!canonical.valid) {
+    console.error(`[Kiro] refusing invalid conversation (claude → kiro): ${(canonical.errors || []).join(", ") || "unknown"} | turns=${(canonical.history || []).length + 1}`);
+    return null;
+  }
  const replayCurrent = canonical.currentMessage.userInputMessage;
  const userInputMessage = {
    content: replayCurrent.content || "",
--- a/open-sse/translator/request/openai-responses.js
+++ b/open-sse/translator/request/openai-responses.js
@@ -421,6 +421,7 @@ export function openaiToOpenAIResponsesRequest(model, body, stream, credentials)
  if (body.reasoning !== undefined) result.reasoning = body.reasoning;
  if (body.reasoning_effort !== undefined) result.reasoning = { effort: body.reasoning_effort, summary: "auto" };
  if (body.service_tier !== undefined) result.service_tier = body.service_tier;
+  if (body.prompt_cache_key !== undefined) result.prompt_cache_key = body.prompt_cache_key;

  return result;
 }
--- a/open-sse/translator/request/openai-to-kiro.js
+++ b/open-sse/translator/request/openai-to-kiro.js
@@ -379,6 +379,18 @@ export function openaiToKiroRequest(model, body, stream, credentials) {
    toolSpecs,
    nameMap,
  });
+  // canonicalizeKiroConversation() already ran its second-chance repair (flatten
+  // every structured tool turn to text, then re-validate). A body that is STILL
+  // invalid here cannot be made shippable, and Kiro answers it with
+  // 400 {"message":"Improperly formed request.","reason":"REQUEST_BODY_INVALID"}.
+  // Fail locally instead: chatCore turns a falsy return into a 400 without
+  // spending an upstream call or a per-account cooldown. The taxonomy
+  // (role:N | pair:N | id:N | spec:N | orphan:0 | current) names the offending
+  // turn so the shape can be diagnosed from the log alone.
+  if (!canonical.valid) {
+    console.error(`[Kiro] refusing invalid conversation (openai → kiro): ${(canonical.errors || []).join(", ") || "unknown"} | turns=${(canonical.history || []).length + 1}`);
+    return null;
+  }
  const replayCurrent = canonical.currentMessage.userInputMessage;

  const payload = {
--- a/open-sse/translator/response/kiro-to-claude.js
+++ b/open-sse/translator/response/kiro-to-claude.js
@@ -75,6 +75,15 @@ export function kiroToClaudeResponse(chunk, state) {
        ? data.usage.completion_tokens
        : 0;
    state.usage = { input_tokens: promptTokens, output_tokens: outputTokens };
+    // Claude clients read cache_read/cache_creation to price a turn and to size
+    // their prompt cache. Both spellings are accepted because the Kiro executor
+    // emits the Chat shape and passthrough responses use the nested details form.
+    const cacheRead = data.usage.cache_read_input_tokens
+      ?? data.usage.prompt_tokens_details?.cached_tokens;
+    const cacheCreation = data.usage.cache_creation_input_tokens
+      ?? data.usage.prompt_tokens_details?.cache_creation_tokens;
+    if (typeof cacheRead === "number") state.usage.cache_read_input_tokens = cacheRead;
+    if (typeof cacheCreation === "number") state.usage.cache_creation_input_tokens = cacheCreation;
  }

  // First chunk → emit message_start.
@@ -254,6 +263,13 @@ export function kiroToClaudeNonStreaming(data) {
    usage: {
      input_tokens: usage.prompt_tokens || 0,
      output_tokens: usage.completion_tokens || 0,
+      // Same cache preservation as the streaming path above.
+      ...(typeof (usage.cache_read_input_tokens ?? usage.prompt_tokens_details?.cached_tokens) === "number"
+        ? { cache_read_input_tokens: usage.cache_read_input_tokens ?? usage.prompt_tokens_details.cached_tokens }
+        : {}),
+      ...(typeof (usage.cache_creation_input_tokens ?? usage.prompt_tokens_details?.cache_creation_tokens) === "number"
+        ? { cache_creation_input_tokens: usage.cache_creation_input_tokens ?? usage.prompt_tokens_details.cache_creation_tokens }
+        : {}),
    },
  };
 }
--- a/open-sse/translator/response/openai-responses.js
+++ b/open-sse/translator/response/openai-responses.js
@@ -99,8 +99,8 @@ export function openaiToOpenAIResponsesResponse(chunk, state) {
    }
  }

-  // Handle tool_calls
-  if (delta.tool_calls) {
+  // Handle tool_calls (empty array is truthy; require a real call)
+  if (delta.tool_calls && delta.tool_calls.length) {
    closeMessage(state, emit, idx);
    for (const tc of delta.tool_calls) {
      emitToolCall(state, emit, tc);