opencode-anthropic-fix 0.0.49 → 0.1.1

package/README.md

@@ -44,7 +44,7 @@ The [original plugin](https://github.com/anomalyco/opencode-anthropic-auth) prov
 - **Configurable strategies** — sticky, round-robin, or hybrid account selection
 - **Claude Code signature emulation** — full HTTP header, system prompt, beta flag, and metadata mimicry derived from Claude Code's open source code. System prompt cache scoping follows the real CC's three-path `splitSysPromptPrefix()` architecture (boundary/global/org) with exact marker detection and scope-aware `cache_control` generation
 - **OAuth endpoint fingerprint parity** — matches the real CLI's bundled axios 1.13.6 HTTP client signature (`Accept`, `User-Agent`, `Content-Type`) on all OAuth token endpoint calls, required since 2026-03-21 server-side enforcement
-- **Billing header fingerprint parity** — `cc_version` suffix uses the real CLI's 3-char fingerprint hash (SHA-256 of salt + first user message chars + version), `cch` matches the Bun native client attestation placeholder, and `X-Claude-Code-Session-Id` header is sent on all requests
+- **Billing header fingerprint parity** — `cc_version` suffix uses the real CLI's 3-char fingerprint hash (SHA-256 of salt + first user message chars + version), `cch=00000` is a static placeholder (xxHash64 attestation was removed in Claude Code v2.1.97), system prompt is sanitized to match Claude Code's pattern validation, and `X-Claude-Code-Session-Id` header is sent on all requests
 - **Adaptive thinking for Opus/Sonnet 4.6** — automatically normalizes thinking to `{type: "adaptive"}` for supported models, with `effort-2025-11-24` beta
 - **Upstream-aligned auto betas** — 13+ always-on betas matching Claude Code 2.1.92 (`redact-thinking-2026-02-12` available as opt-in to preserve thinking block visibility)
 - **1M context limit override** — patches `model.limit.context` so OpenCode compacts at the right threshold while `models.dev` catches up
@@ -700,6 +700,19 @@ Make sure `~/.local/bin` is on your PATH:
 export PATH="$HOME/.local/bin:$PATH"
 ```
 
+## Credits & Acknowledgments
+
+### Reverse Engineering & Research
+
+- **[CCH Attestation Reverse Engineering](https://a10k.co/b/reverse-engineering-claude-code-cch.html)** — Deep dive into the xxHash64 algorithm and Bun binary attestation mechanism
+- **[OpenClaw Billing Proxy](https://github.com/zacdcook/openclaw-billing-proxy)** — Early exploration of billing headers and trigger phrase discovery
+- **[Free Code CCH Implementation](https://github.com/paoloanzn/free-code/pull/9)** — Reference implementation of xxHash64 in JavaScript
+- **[rmk40](https://github.com/rmk40/opencode-anthropic-auth)** — Co-author on fingerprint extraction work and multi-account architecture
+
+### Original Work
+
+- [AnomalyCo](https://github.com/anomalyco/opencode-anthropic-auth) — Original OpenCode Anthropic auth plugin
+
 ## License
 
 This project is licensed under the [GNU General Public License v3.0](LICENSE) (GPLv3).

package/index.mjs

@@ -3,6 +3,7 @@ import { stdin, stdout } from "node:process";
 import { randomBytes, randomUUID, createHash as createHashCrypto } from "node:crypto";
 import { existsSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
 import { join, resolve, basename } from "node:path";
+// xxhash-wasm import removed: CCH attestation was removed in CC v2.1.97
 import { AccountManager } from "./lib/accounts.mjs";
 import { authorize as oauthAuthorize, exchange as oauthExchange, refreshToken } from "./lib/oauth.mjs";
 import { loadConfig, loadConfigFresh, saveConfig, CLIENT_ID, getConfigDir } from "./lib/config.mjs";
@@ -448,6 +449,8 @@ export async function AnthropicAuthPlugin({ client, project, directory, worktree
         acc.access = credentials.access;
         acc.expires = credentials.expires;
         if (credentials.email) acc.email = credentials.email;
+        if (credentials.accountUuid) acc.accountUuid = credentials.accountUuid;
+        if (credentials.organizationUuid) acc.organizationUuid = credentials.organizationUuid;
         acc.enabled = true;
         acc.consecutiveFailures = 0;
         acc.lastFailureTime = null;
@@ -469,6 +472,8 @@ export async function AnthropicAuthPlugin({ client, project, directory, worktree
       stored.accounts.push({
         id: `${now}:${credentials.refresh.slice(0, 12)}`,
         email: credentials.email,
+        accountUuid: credentials.accountUuid,
+        organizationUuid: credentials.organizationUuid,
         refreshToken: credentials.refresh,
         access: credentials.access,
         expires: credentials.expires,
@@ -481,8 +486,25 @@ export async function AnthropicAuthPlugin({ client, project, directory, worktree
         lastFailureTime: null,
         stats: createDefaultStats(now),
       });
-      await saveAccounts(stored);
+      // If accountUuid wasn't in the token exchange response, fetch from profile API
       const newAccount = stored.accounts[stored.accounts.length - 1];
+      if (!newAccount.accountUuid && newAccount.access) {
+        try {
+          const profileResp = await globalThis.fetch("https://api.anthropic.com/api/oauth/profile", {
+            method: "GET",
+            headers: { Authorization: `Bearer ${newAccount.access}`, "Content-Type": "application/json" },
+            signal: AbortSignal.timeout(10_000),
+          });
+          if (profileResp.ok) {
+            const profile = await profileResp.json();
+            if (profile.account?.uuid) newAccount.accountUuid = profile.account.uuid;
+            if (profile.organization?.uuid) newAccount.organizationUuid = profile.organization.uuid;
+          }
+        } catch {
+          /* Best-effort — don't fail account creation */
+        }
+      }
+      await saveAccounts(stored);
       await persistOpenCodeAuth(newAccount.refreshToken, newAccount.access, newAccount.expires);
       await reloadAccountManagerFromDisk();
       pendingSlashOAuth.delete(sessionID);
@@ -2793,12 +2815,16 @@ export async function AnthropicAuthPlugin({ client, project, directory, worktree
                   _adaptiveOverride,
                   _tokenEconomy,
                 );
+                // v2.1.97: cch=00000 is now static (xxHash64 attestation removed).
+                // Send body as-is without cch replacement.
+                const finalBody = body;
+
                 // Execute the request
                 let response;
                 try {
                   response = await fetch(requestInput, {
                     ...requestInit,
-                    body,
+                    body: finalBody,
                     headers: requestHeaders,
                     // Disable keepalive when a previous ECONNRESET/EPIPE was detected
                     // to force a fresh TCP connection and avoid stale socket reuse.
@@ -3225,11 +3251,14 @@ export async function AnthropicAuthPlugin({ client, project, directory, worktree
                     }
                   }
 
-                  // Auto-disable extended cache TTL if the API rejects it with a 400
+                  // Auto-disable extended cache TTL ONLY if the API explicitly says TTL is
+                  // not supported. Do NOT disable on TTL ordering errors (which are fixable).
                   if (
                     response.status === 400 &&
                     errorBody &&
-                    (errorBody.includes("ttl") || errorBody.includes("cache_control"))
+                    errorBody.includes("cache_control") &&
+                    !errorBody.includes("must not come after") &&
+                    !errorBody.includes("maximum of")
                   ) {
                     if (config.cache_policy && config.cache_policy.ttl_supported !== false) {
                       config.cache_policy.ttl_supported = false;
@@ -4929,17 +4958,22 @@ process.once("beforeExit", _beforeExitHandler);
 // Request building helpers (extracted from original fetch interceptor)
 // ---------------------------------------------------------------------------
 
-const FALLBACK_CLAUDE_CLI_VERSION = "2.1.92";
+const FALLBACK_CLAUDE_CLI_VERSION = "2.1.97";
 const CLAUDE_CODE_NPM_LATEST_URL = "https://registry.npmjs.org/@anthropic-ai/claude-code/latest";
-const CLAUDE_CODE_BUILD_TIME = "2026-04-03T23:25:15Z";
+const CLAUDE_CODE_BUILD_TIME = "2026-04-08T20:46:46Z";
 
-// The @anthropic-ai/sdk version bundled with Claude Code v2.1.92.
+// The @anthropic-ai/sdk version bundled with Claude Code v2.1.97.
 // This is distinct from the CLI version and goes in X-Stainless-Package-Version.
-// Verified by extracting VERSION="0.208.0" from the bundled cli.js of all versions .80-.92.
+// Verified by extracting VERSION="0.208.0" from the bundled cli.js of all versions .80-.97.
 const ANTHROPIC_SDK_VERSION = "0.208.0";
 
 // Map of CLI version → bundled SDK version (update when CLI version changes)
 const CLI_TO_SDK_VERSION = new Map([
+  ["2.1.97", "0.208.0"],
+  ["2.1.96", "0.208.0"],
+  ["2.1.95", "0.208.0"],
+  ["2.1.94", "0.208.0"],
+  ["2.1.93", "0.208.0"],
   ["2.1.92", "0.208.0"],
   ["2.1.91", "0.208.0"],
   ["2.1.90", "0.208.0"],
@@ -4966,6 +5000,12 @@ function getSdkVersion(cliVersion) {
 const BILLING_HASH_SALT = "59cf53e54c78";
 const BILLING_HASH_INDICES = [4, 7, 20];
 
+// cch attestation: REMOVED in v2.1.97.
+// Previously (v2.1.96), CC computed xxHash64 of the serialized body and replaced
+// "cch=00000" with the 20-bit masked hash. In v2.1.97, the xxHash64 computation
+// was completely removed — "cch=00000" is now sent as a static placeholder.
+// Sending a computed cch value would now be detected as non-genuine.
+
 /**
  * Compute the billing cache hash (cch) matching Claude Code's NP1() function.
  * SHA256(salt + chars_at_indices[4,7,20]_from_first_user_msg + version).slice(0,3)
@@ -5192,24 +5232,23 @@ const BEDROCK_UNSUPPORTED_BETAS = new Set([
   "interleaved-thinking-2025-05-14",
   "context-1m-2025-08-07",
   "tool-search-tool-2025-10-19",
-  "code-execution-2025-08-25",
-  "files-api-2025-04-14",
-  "fine-grained-tool-streaming-2025-05-14",
 ]);
 const EXPERIMENTAL_BETA_FLAGS = new Set([
   "adaptive-thinking-2026-01-28",
   "advanced-tool-use-2025-11-20",
+  "advisor-tool-2026-03-01",
   "afk-mode-2026-01-31",
   "code-execution-2025-08-25",
+  "compact-2026-01-12",
   "context-1m-2025-08-07",
   "context-management-2025-06-27",
   "fast-mode-2026-02-01",
   "files-api-2025-04-14",
-  "fine-grained-tool-streaming-2025-05-14",
   "interleaved-thinking-2025-05-14",
   "prompt-caching-scope-2026-01-05",
   "redact-thinking-2026-02-12",
   "structured-outputs-2025-12-15",
+  "task-budgets-2026-03-13",
   "tool-search-tool-2025-10-19",
   "web-search-2025-03-05",
 ]);
@@ -5321,7 +5360,8 @@ function isNonInteractiveMode() {
 
 /**
  * Build the extended User-Agent for API calls.
- * Claude Code sends "claude-cli/{version} (external, {entrypoint})".
+ * Real CC v96 sends "claude-cli/{version} (external, {entrypoint})" — confirmed via
+ * proxy capture of real CC on Windows/Node.js.
  * @param {string} version
  * @returns {string}
  */
@@ -5439,12 +5479,17 @@ function isAdaptiveThinkingModel(model) {
 
 /**
  * Check if a model is eligible for 1M context (can receive context-1m beta).
- * This includes models with explicit "1m" in the name AND Opus 4.6.
+ * Real CC v2.1.97 U01(): claude-sonnet-4* || opus-4-6 are eligible.
+ * Also matches explicit "1m" in the name (e.g. "claude-opus-4-6[1m]").
  * @param {string} model
  * @returns {boolean}
  */
 function isEligibleFor1MContext(model) {
-  return /(^|[-_ ])1m($|[-_ ])|context[-_]?1m/i.test(model) || isOpus46Model(model);
+  if (!model) return false;
+  // Explicit 1m suffix/tag in model name
+  if (/(^|[-_ ])1m($|[-_ ])|context[-_]?1m|\[1m\]/i.test(model)) return true;
+  // CC v2.1.97 U01: claude-sonnet-4* (any Sonnet 4.x) or opus-4-6
+  return /claude-sonnet-4|sonnet[._-]4/i.test(model) || isOpus46Model(model);
 }
 
 /**
@@ -5477,11 +5522,20 @@ function supportsWebSearch(model) {
 
 /**
  * @param {URL | null} requestUrl
- * @returns {"anthropic" | "bedrock" | "vertex" | "foundry"}
+ * @returns {"anthropic" | "bedrock" | "vertex" | "foundry" | "anthropicAws" | "mantle"}
  */
 function detectProvider(requestUrl) {
+  // Match Claude Code provider precedence first (env-driven), then URL fallback.
+  if (isTruthyEnv(process.env.CLAUDE_CODE_USE_BEDROCK)) return "bedrock";
+  if (isTruthyEnv(process.env.CLAUDE_CODE_USE_FOUNDRY)) return "foundry";
+  if (isTruthyEnv(process.env.CLAUDE_CODE_USE_ANTHROPIC_AWS)) return "anthropicAws";
+  if (isTruthyEnv(process.env.CLAUDE_CODE_USE_MANTLE)) return "mantle";
+  if (isTruthyEnv(process.env.CLAUDE_CODE_USE_VERTEX)) return "vertex";
+
   if (!requestUrl) return "anthropic";
   const host = requestUrl.hostname.toLowerCase();
+  if (host.includes("mantle")) return "mantle";
+  if (host.includes("anthropicaws")) return "anthropicAws";
   if (host.includes("bedrock") || host.includes("amazonaws.com")) return "bedrock";
   if (host.includes("aiplatform") || host.includes("vertex")) return "vertex";
   if (host.includes("foundry") || host.includes("azure")) return "foundry";
@@ -5596,20 +5650,21 @@ function buildRequestMetadata(input) {
 
 /**
  * Build the billing header block for Claude Code system prompt injection.
- * Claude Code v2.1.92: cc_version includes 3-char fingerprint hash (not model ID).
- * cch is a static "00000" placeholder for Bun native client attestation.
+ * Claude Code v2.1.97: cc_version includes 3-char fingerprint hash (not model ID).
+ * cch is a static "00000" placeholder (xxHash64 attestation removed in v2.1.97).
  *
  * Real CC (system.ts:78): version = `${MACRO.VERSION}.${fingerprint}`
- * Real CC (system.ts:82): cch = feature('NATIVE_CLIENT_ATTESTATION') ? ' cch=00000;' : ''
+ * Real CC (system.ts:82): cch = ' cch=00000;' (static, no longer computed)
  *
- * @param {string} version - CLI version (e.g., "2.1.92")
+ * @param {string} version - CLI version (e.g., "2.1.97")
  * @param {string} [firstUserMessage] - First user message text for fingerprint computation
- * @param {string} [provider] - API provider ("anthropic" | "bedrock" | "vertex" | "foundry")
+ * @param {string} [provider] - API provider ("anthropic" | "bedrock" | "vertex" | "foundry" | "anthropicAws" | "mantle")
  * @returns {string}
  */
 function buildAnthropicBillingHeader(version, firstUserMessage, provider) {
   if (isFalsyEnv(process.env.CLAUDE_CODE_ATTRIBUTION_HEADER)) return "";
-  const entrypoint = process.env.CLAUDE_CODE_ENTRYPOINT || "unknown";
+  // Real CC sends cc_entrypoint=cli (confirmed via proxy capture).
+  const entrypoint = process.env.CLAUDE_CODE_ENTRYPOINT || "cli";
   // Fix #1: cc_version suffix is the 3-char fingerprint hash, NOT the model ID.
   // computeBillingCacheHash() computes SHA256(salt + msg[4]+msg[7]+msg[20] + version)[:3]
   // which matches computeFingerprint() in the real CC source (utils/fingerprint.ts).
@@ -5617,29 +5672,44 @@ function buildAnthropicBillingHeader(version, firstUserMessage, provider) {
   // the hash from "000" chars (indices 4,7,20 all missing → fallback "0").
   const fingerprint = computeBillingCacheHash(firstUserMessage || "", version);
   const ccVersion = `${version}.${fingerprint}`;
-  // Fix #4: cch is a static "00000" placeholder for Bun's native client attestation.
-  // Real CC v92: cch is included for all providers EXCEPT bedrock/anthropicAws.
-  // The real Bun binary overwrites these zeros in the serialized body bytes.
-  // For non-Bun runtimes, the server sees "00000" and skips attestation verification.
-  const isBedrock = provider === "bedrock" || provider === "anthropicAws";
-  const cchPart = isBedrock ? "" : " cch=00000;";
-  let header = `x-anthropic-billing-header: cc_version=${ccVersion}; cc_entrypoint=${entrypoint};${cchPart}`;
+  // cch: v2.1.97 sends static "cch=00000" — xxHash64 attestation was removed.
+  // The server uses the PRESENCE of cch=00000 as a CC identification signal.
+  const cchDisabled = provider === "bedrock" || provider === "anthropicAws" || provider === "mantle";
+  const cchPart = cchDisabled ? "" : " cch=00000;";
+  // Build workload part (upstream concatenates directly, no regex replace)
+  let workloadPart = "";
   const workload = process.env.CLAUDE_CODE_WORKLOAD;
   if (workload) {
     // QA fix M5: sanitize workload value to prevent header injection
     const safeWorkload = workload.replace(/[;\s\r\n]/g, "_");
-    header = header.replace(/;$/, ` cc_workload=${safeWorkload};`);
+    workloadPart = ` cc_workload=${safeWorkload};`;
   }
-  return header;
+  return `x-anthropic-billing-header: cc_version=${ccVersion}; cc_entrypoint=${entrypoint};${cchPart}${workloadPart}`;
 }
 
 /**
  * @param {string} text
  * @returns {string}
  */
+// Max system prompt length that passes CC billing validation.
+// The server pattern-matches the system prompt against the real CC prompt.
+// Opencode's customizations after ~5800 chars diverge and trigger extra usage billing.
+const MAX_SAFE_SYSTEM_TEXT_LENGTH = 5000;
+
 function sanitizeSystemText(text) {
   // QA fix M4: use word boundaries to avoid mangling URLs and code identifiers
-  return text.replace(/\bOpenCode\b/g, "Claude Code").replace(/\bopencode\b/gi, "Claude");
+  let sanitized = text.replace(/\bOpenCode\b/g, "Claude Code").replace(/\bopencode\b/gi, "Claude");
+  // Strip non-CC custom prefixes before the standard CC prompt.
+  const ccStandardStart = sanitized.indexOf("You are an interactive");
+  if (ccStandardStart > 0) {
+    sanitized = sanitized.slice(ccStandardStart);
+  }
+  // Truncate to safe length — opencode customizations beyond this point
+  // diverge from real CC and trigger extra usage billing detection.
+  if (sanitized.length > MAX_SAFE_SYSTEM_TEXT_LENGTH) {
+    sanitized = sanitized.slice(0, MAX_SAFE_SYSTEM_TEXT_LENGTH);
+  }
+  return sanitized;
 }
 
 /**
@@ -5904,9 +5974,9 @@ function splitSysPromptPrefix(blocks, attributionHeader, identityString, useBoun
       if (attributionHeader) result.push({ text: attributionHeader, cacheScope: null });
       // Identity: cacheScope null in boundary mode (real CC behavior)
       result.push({ text: identityString, cacheScope: null });
-      const staticJoined = staticBlocks.join("\n\n");
+      const staticJoined = staticBlocks.join("\n");
       if (staticJoined) result.push({ text: staticJoined, cacheScope: "global" });
-      const dynamicJoined = dynamicBlocks.join("\n\n");
+      const dynamicJoined = dynamicBlocks.join("\n");
       if (dynamicJoined) result.push({ text: dynamicJoined, cacheScope: null });
       return result;
     }
@@ -5921,7 +5991,7 @@ function splitSysPromptPrefix(blocks, attributionHeader, identityString, useBoun
   const result = [];
   if (attributionHeader) result.push({ text: attributionHeader, cacheScope: null });
   result.push({ text: identityString, cacheScope: "org" });
-  const restJoined = rest.join("\n\n");
+  const restJoined = rest.join("\n");
   if (restJoined) result.push({ text: restJoined, cacheScope: "org" });
   return result;
 }
@@ -5987,7 +6057,7 @@ function buildSystemPromptBlocks(system, signature) {
  * @param {string} incomingBeta
  * @param {boolean} signatureEnabled
  * @param {string} model
- * @param {"anthropic" | "bedrock" | "vertex" | "foundry"} provider
+ * @param {"anthropic" | "bedrock" | "vertex" | "foundry" | "anthropicAws" | "mantle"} provider
  * @param {string[]} [customBetas]
  * @param {import('./lib/config.mjs').AccountSelectionStrategy} [strategy]
  * @param {string} [requestPath]
@@ -6044,7 +6114,7 @@ function buildAnthropicBetaHeader(
   // Tool search: use provider-aware header.
   // 1P/Foundry u2192 advanced-tool-use-2025-11-20 (enables broader tool capabilities)
   // Vertex/Bedrock u2192 tool-search-tool-2025-10-19 (3P-compatible subset)
-  if (provider === "vertex" || provider === "bedrock") {
+  if (provider === "vertex" || provider === "bedrock" || provider === "mantle") {
     betas.push("tool-search-tool-2025-10-19");
   } else {
     betas.push(ADVANCED_TOOL_USE_BETA_FLAG); // "advanced-tool-use-2025-11-20"
@@ -6339,7 +6409,9 @@ function buildRequestHeaders(
     requestHeaders.set("x-stainless-arch", getStainlessArch(process.arch));
     requestHeaders.set("x-stainless-lang", "js");
     requestHeaders.set("x-stainless-os", getStainlessOs(process.platform));
-    requestHeaders.set("x-stainless-package-version", getSdkVersion(signature.claudeCliVersion));
+    // Real CC sends 0.81.0 (confirmed via proxy capture), not the internal 0.208.0.
+    requestHeaders.set("x-stainless-package-version", "0.81.0");
+    // Real CC on Windows/Node reports "node" — confirmed via proxy capture.
     requestHeaders.set("x-stainless-runtime", "node");
     requestHeaders.set("x-stainless-runtime-version", process.version);
     const incomingRetryCount = requestHeaders.get("x-stainless-retry-count");
@@ -6347,21 +6419,10 @@ function buildRequestHeaders(
       "x-stainless-retry-count",
       incomingRetryCount && !isFalsyEnv(incomingRetryCount) ? incomingRetryCount : "0",
     );
-    // x-stainless-timeout: sent only for non-streaming requests.
-    // Claude Code sends 600 (10 minutes) as the default timeout in seconds.
-    // For streaming requests, the SDK omits this header entirely.
-    if (requestBody) {
-      try {
-        const parsed = JSON.parse(requestBody);
-        // stream defaults to true in Claude Code; only explicitly false means non-streaming
-        if (parsed.stream === false) {
-          requestHeaders.set("x-stainless-timeout", "600");
-        }
-        // Streaming requests: omit x-stainless-timeout (real SDK behavior)
-      } catch {
-        // Non-JSON body or parse error — omit header (safe default)
-      }
-    }
+    // x-stainless-timeout: real CC sends 600 on ALL requests (confirmed via proxy capture).
+    requestHeaders.set("x-stainless-timeout", "600");
+    // anthropic-dangerous-direct-browser-access: real CC sends this on all requests.
+    requestHeaders.set("anthropic-dangerous-direct-browser-access", "true");
     const stainlessHelpers = buildStainlessHelperHeader(tools, messages);
     if (stainlessHelpers) {
       requestHeaders.set("x-stainless-helper", stainlessHelpers);
@@ -6383,10 +6444,11 @@ function buildRequestHeaders(
       requestHeaders.set("x-anthropic-additional-protection", "true");
     }
 
-    // Claude Code v2.1.84: x-client-request-id — unique UUID per request for debugging timeouts.
-    requestHeaders.set("x-client-request-id", randomUUID());
+    // x-client-request-id: NOT sent by real CC (confirmed via proxy capture). Removed.
   }
   requestHeaders.delete("x-api-key");
+  // x-session-affinity: set by opencode SDK but NOT in real CC. Strip it.
+  requestHeaders.delete("x-session-affinity");
 
   return requestHeaders;
 }
@@ -6545,17 +6607,33 @@ function transformRequestBody(body, signature, runtime, betaHeader, config) {
       signature.cachePolicy?.ttl_supported !== false &&
       !isTitleGen
     ) {
-      const strategy = signature.strategy || "sticky";
-      if (strategy !== "round-robin" && Array.isArray(parsed.messages)) {
-        const cacheControl = { type: "ephemeral", ttl: signature.cachePolicy?.ttl || "1h" };
+      // Strip ALL incoming cache_control from tools and messages to prevent
+      // TTL ordering violations (host SDK may set ttl=5m which conflicts with
+      // our system prompt ttl=1h). Then add our own 1h to the last user message
+      // (matching real CC behavior seen in proxy capture).
+      const ccTtl = signature.cachePolicy?.ttl || "1h";
+      if (Array.isArray(parsed.tools)) {
+        for (const tool of parsed.tools) {
+          if (tool.cache_control) delete tool.cache_control;
+        }
+      }
+      if (Array.isArray(parsed.messages)) {
         for (const msg of parsed.messages) {
-          if (!msg.content || !Array.isArray(msg.content) || msg.content.length === 0) continue;
-          // Only cache user and assistant messages (not tool results which change frequently)
-          if (msg.role !== "user" && msg.role !== "assistant") continue;
+          if (Array.isArray(msg.content)) {
+            for (const block of msg.content) {
+              if (block.cache_control) delete block.cache_control;
+            }
+          }
+        }
+        // Add cache_control to last user message (real CC does this)
+        for (let i = parsed.messages.length - 1; i >= 0; i--) {
+          const msg = parsed.messages[i];
+          if (msg.role !== "user" || !Array.isArray(msg.content) || msg.content.length === 0) continue;
           const lastBlock = msg.content[msg.content.length - 1];
           if (lastBlock && typeof lastBlock === "object") {
-            lastBlock.cache_control = cacheControl;
+            lastBlock.cache_control = { type: "ephemeral", ttl: ccTtl };
           }
+          break;
         }
       }
     }
@@ -6595,9 +6673,10 @@ function transformRequestBody(body, signature, runtime, betaHeader, config) {
       }
     }
 
-    // Fast mode: inject speed parameter for supported models
+    // Fast mode: inject speed parameter for Opus 4.6 only (v2.1.97 restriction).
+    // Real CC v2.1.97 xJ() checks: model.includes("opus-4-6") — Sonnet is NOT eligible.
     const fastModeEnabled = signature.fastMode && !isFalsyEnv(process.env.OPENCODE_ANTHROPIC_DISABLE_FAST_MODE);
-    if (fastModeEnabled && parsed.model && (isOpus46Model(parsed.model) || isSonnet46Model(parsed.model))) {
+    if (fastModeEnabled && parsed.model && isOpus46Model(parsed.model)) {
       parsed.speed = "fast";
     }
 
@@ -6662,13 +6741,19 @@ function transformRequestUrl(input) {
     requestUrl = null;
   }
 
-  if (
-    requestUrl &&
-    (requestUrl.pathname === "/v1/messages" || requestUrl.pathname === "/v1/messages/count_tokens") &&
-    !requestUrl.searchParams.has("beta")
-  ) {
-    requestUrl.searchParams.set("beta", "true");
-    requestInput = input instanceof Request ? new Request(requestUrl.toString(), input) : requestUrl;
+  if (requestUrl && !requestUrl.searchParams.has("beta")) {
+    const p = requestUrl.pathname;
+    // SDK may send to /messages (base URL includes /v1) or /v1/messages (base URL is root)
+    const isMessages =
+      p === "/v1/messages" || p === "/messages" || p === "/v1/messages/count_tokens" || p === "/messages/count_tokens";
+    if (isMessages) {
+      // Normalize path to /v1/messages (required by API and proxies)
+      if (!p.startsWith("/v1/")) {
+        requestUrl.pathname = "/v1" + p;
+      }
+      requestUrl.searchParams.set("beta", "true");
+      requestInput = input instanceof Request ? new Request(requestUrl.toString(), input) : requestUrl;
+    }
   }
 
   return { requestInput, requestUrl };
@@ -7149,6 +7234,13 @@ async function refreshAccountToken(account, client, source = "foreground", { onT
     if (json.refresh_token) {
       account.refreshToken = json.refresh_token;
     }
+    // Extract account UUID from token refresh response if present
+    if (json.account?.uuid) {
+      account.accountUuid = json.account.uuid;
+    }
+    if (json.organization?.uuid) {
+      account.organizationUuid = json.organization.uuid;
+    }
     markTokenStateUpdated(account);
 
     // Persist new tokens to disk BEFORE releasing the cross-process lock.

package/lib/accounts.mjs

@@ -107,6 +107,8 @@ export class AccountManager {
         id: acc.id || `${acc.addedAt}:${hashTokenFragment(acc.refreshToken)}`,
         index,
         email: acc.email,
+        accountUuid: acc.accountUuid,
+        organizationUuid: acc.organizationUuid,
         refreshToken: acc.refreshToken,
         access: acc.access,
         expires: acc.expires,
@@ -736,6 +738,8 @@ export class AccountManager {
         return {
           id: acc.id,
           email: acc.email,
+          accountUuid: acc.accountUuid,
+          organizationUuid: acc.organizationUuid,
           refreshToken: freshestAuth.refreshToken,
           access: freshestAuth.access,
           expires: freshestAuth.expires,

package/lib/oauth.mjs

@@ -292,6 +292,10 @@ export async function exchange(code, verifier) {
     access: json.access_token,
     expires: Date.now() + json.expires_in * 1000,
     email: json.account?.email_address || undefined,
+    // Real CC extracts account UUID from token response (oauth/client.ts:253)
+    // and uses it in metadata.user_id.account_uuid for billing correlation.
+    accountUuid: json.account?.uuid || undefined,
+    organizationUuid: json.organization?.uuid || undefined,
   };
 }
 

package/lib/storage.mjs

@@ -163,6 +163,8 @@ function validateAccount(raw, now) {
   return {
     id,
     email: typeof acc.email === "string" ? acc.email : undefined,
+    accountUuid: typeof acc.accountUuid === "string" ? acc.accountUuid : undefined,
+    organizationUuid: typeof acc.organizationUuid === "string" ? acc.organizationUuid : undefined,
     refreshToken: acc.refreshToken,
     access: typeof acc.access === "string" ? acc.access : undefined,
     expires: typeof acc.expires === "number" && Number.isFinite(acc.expires) ? acc.expires : undefined,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "opencode-anthropic-fix",
-  "version": "0.0.49",
+  "version": "0.1.1",
   "license": "GPL-3.0-or-later",
   "main": "./index.mjs",
   "files": [