@xtruder/opencode-claude-max-plugin 0.3.1 → 0.4.0

package/README.md

@@ -27,7 +27,7 @@ Add the plugin to your `opencode.json` (project-level or `~/.config/opencode/ope
 
 That's it. The plugin self-registers the `anthropic-sdk` provider and its models (Haiku 4.5, Sonnet 4.6, Opus 4.6, Opus 4.7, Opus 4.8) at startup via the OpenCode config hook. No separate `provider` block is needed.
 
-Then open OpenCode and run `/connect` → Other → `anthropic-sdk`. If Claude Code is installed and you're logged in, credentials are read automatically from `~/.claude/.credentials.json` — no API key needed.
+Then open OpenCode and models will automatically be available under `anthropic-sdk` provider.
 
 ### TUI Plugin (sidebar + /usage command)
 

package/build/index.d.ts

@@ -1,14 +1,22 @@
 import type { LanguageModelV3 } from "@ai-sdk/provider";
 import type { Plugin } from "@opencode-ai/plugin";
 export declare const CLAUDE_CODE_SYSTEM_PROMPT: string;
-export declare const CLAUDE_CODE_OPUS48_SYSTEM_PROMPT: string;
+export declare const CLAUDE_CODE_NEW_SYSTEM_PROMPT: string;
 /**
  * Select the Claude Code system prompt that matches the given model.
  *
- * Claude Code 2.1.154 ships per-model prompts: Opus 4.8 gets a condensed
- * ~6KB "Harness" prompt that trusts the model's defaults; everything else
- * (Opus 4.7, Sonnet 4.6, Haiku 4.5, Opus 4.6) gets the long-form ~11KB
- * prompt with explicit behavior rules.
+ * Claude Code ships per-model prompts: the newer flagships (Opus 4.8,
+ * Fable 5) get a condensed "Harness" prompt that trusts the model's
+ * defaults; everything else (Opus 4.7, Sonnet 4.6, Haiku 4.5, Opus 4.6)
+ * gets the long-form prompt with explicit behavior rules.
+ *
+ * Opus 4.8 and Fable 5 share one condensed prompt here. Claude Code 2.1.173
+ * sends Fable 5 a longer variant (security steering, a verbose
+ * "Communicating" section, a Fable 5 identity blurb, and CLI-only session
+ * items), but the safety classifiers are server-side rather than
+ * prompt-driven and the CLI-specific guidance does not apply to OpenCode —
+ * so we distill it down to the shared base. The verbatim capture lives in
+ * src/fixtures/claudecode-system-fable5-original.txt.
  */
 export declare function selectClaudePromptForModel(modelId: string): string;
 export interface AnthropicSDKProviderOptions {

package/build/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,kBAAkB,CAAA;AACvD,OAAO,KAAK,EAAE,MAAM,EAAE,MAAM,qBAAqB,CAAA;AASjD,eAAO,MAAM,yBAAyB,QAA4B,CAAA;AAClE,eAAO,MAAM,gCAAgC,QAA8B,CAAA;AAE3E;;;;;;;GAOG;AACH,wBAAgB,0BAA0B,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM,CAGlE;AAqND,MAAM,WAAW,2BAA2B;IAC1C;;OAEG;IACH,MAAM,CAAC,EAAE,MAAM,CAAA;IAEf;;OAEG;IACH,OAAO,CAAC,EAAE,MAAM,CAAA;IAEhB;;OAEG;IACH,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IAEhC;;OAEG;IACH,KAAK,CAAC,EAAE,OAAO,UAAU,CAAC,KAAK,CAAA;IAE/B;;OAEG;IACH,IAAI,CAAC,EAAE,MAAM,CAAA;IAEb;;;OAGG;IACH,eAAe,CAAC,EAAE,MAAM,CAAA;IAExB;;OAEG;IACH,CAAC,GAAG,EAAE,MAAM,GAAG,OAAO,CAAA;CACvB;AAED,MAAM,WAAW,oBAAoB;IACnC;;OAEG;IACH,aAAa,CAAC,OAAO,EAAE,MAAM,GAAG,eAAe,CAAA;CAChD;AA4BD;;;;;;;;;GASG;AACH,wBAAgB,kBAAkB,CAChC,OAAO,GAAE,2BAAgC,GACxC,oBAAoB,CAqJtB;AAED;;;;;;;;;;;;;;;;GAgBG;AACH,eAAO,MAAM,kBAAkB,EAAE,MAoDhC,CAAA;AAID,eAAe,kBAAkB,CAAA"}
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,kBAAkB,CAAA;AACvD,OAAO,KAAK,EAAE,MAAM,EAAE,MAAM,qBAAqB,CAAA;AASjD,eAAO,MAAM,yBAAyB,QAA4B,CAAA;AAClE,eAAO,MAAM,6BAA6B,QAA2B,CAAA;AAErE;;;;;;;;;;;;;;;GAeG;AACH,wBAAgB,0BAA0B,CAAC,OAAO,EAAE,MAAM,GAAG,MAAM,CAKlE;AAwPD,MAAM,WAAW,2BAA2B;IAC1C;;OAEG;IACH,MAAM,CAAC,EAAE,MAAM,CAAA;IAEf;;OAEG;IACH,OAAO,CAAC,EAAE,MAAM,CAAA;IAEhB;;OAEG;IACH,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;IAEhC;;OAEG;IACH,KAAK,CAAC,EAAE,OAAO,UAAU,CAAC,KAAK,CAAA;IAE/B;;OAEG;IACH,IAAI,CAAC,EAAE,MAAM,CAAA;IAEb;;;OAGG;IACH,eAAe,CAAC,EAAE,MAAM,CAAA;IAExB;;OAEG;IACH,CAAC,GAAG,EAAE,MAAM,GAAG,OAAO,CAAA;CACvB;AAED,MAAM,WAAW,oBAAoB;IACnC;;OAEG;IACH,aAAa,CAAC,OAAO,EAAE,MAAM,GAAG,eAAe,CAAA;CAChD;AA4BD;;;;;;;;;GASG;AACH,wBAAgB,kBAAkB,CAChC,OAAO,GAAE,2BAAgC,GACxC,oBAAoB,CA6KtB;AAED;;;;;;;;;;;;;;;;GAgBG;AACH,eAAO,MAAM,kBAAkB,EAAE,MAwEhC,CAAA;AAID,eAAe,kBAAkB,CAAA"}

package/build/index.js

@@ -48,8 +48,8 @@ function hasCchPlaceholder(body) {
   return block?.text?.includes("cch=00000") ?? false;
 }
 
-// src/claudecode-system-opus48.txt
-var claudecode_system_opus48_default = `
+// src/claudecode-system-new.txt
+var claudecode_system_new_default = `
 You are an interactive agent that helps users with software engineering tasks.
 
 # Harness
@@ -382,6 +382,14 @@ function convertUserMessage(message) {
 function convertAssistantMessage(message) {
   const content = [];
   for (const part of message.content) {
+    const fallback = part.providerMetadata?.anthropic?.fallback ?? part.providerOptions?.anthropic?.fallback;
+    if (fallback?.from?.model && fallback?.to?.model) {
+      content.push({
+        type: "fallback",
+        from: { model: fallback.from.model },
+        to: { model: fallback.to.model }
+      });
+    }
     switch (part.type) {
       case "text":
         if (part.text.length > 0) {
@@ -460,6 +468,14 @@ var idCounter = 0;
 function generateId() {
   return `block-${Date.now()}-${idCounter++}`;
 }
+function buildRefusalError(stopDetails, fallbackModel) {
+  const category = stopDetails?.category ?? null;
+  const explanation = stopDetails?.explanation ?? null;
+  const categoryNote = category ? ` (category: ${category})` : "";
+  const detail = explanation ? explanation : "Claude Fable 5's safety classifiers declined this request.";
+  const hint = fallbackModel ? `The configured fallback model (${fallbackModel}) also refused.` : `Retry with a fallback model such as anthropic-sdk/claude-opus-4-8.`;
+  return new Error(`Claude Fable 5 refused this request${categoryNote}. ${detail} ${hint}`);
+}
 function mapFinishReason(stopReason) {
   const unified = (() => {
     switch (stopReason) {
@@ -476,8 +492,13 @@ function mapFinishReason(stopReason) {
   })();
   return { unified, raw: stopReason ?? undefined };
 }
-function convertStream(anthropicStream, modelId) {
+function convertStream(anthropicStream, modelId, context) {
   const blockStates = new Map;
+  const ctx = {
+    apiModelId: context?.apiModelId ?? modelId,
+    fallbacksEnabled: context?.fallbacksEnabled ?? false,
+    fallbackModel: context?.fallbackModel
+  };
   let inputTokens;
   let outputTokens;
   let cachedInputTokens;
@@ -486,7 +507,7 @@ function convertStream(anthropicStream, modelId) {
     async start(controller) {
       try {
         for await (const event of anthropicStream) {
-          const parts = processEvent(event, blockStates, modelId);
+          const parts = processEvent(event, blockStates, modelId, ctx);
           if (event.type === "message_start" && event.message.usage) {
             const u = event.message.usage;
             inputTokens = u.input_tokens;
@@ -496,6 +517,12 @@ function convertStream(anthropicStream, modelId) {
           if (event.type === "message_delta") {
             const delta = event;
             outputTokens = delta.usage?.output_tokens;
+            if (delta.delta?.stop_reason === "refusal") {
+              controller.enqueue({
+                type: "error",
+                error: buildRefusalError(delta.delta?.stop_details, ctx.fallbacksEnabled ? ctx.fallbackModel : undefined)
+              });
+            }
             const finishReason = mapFinishReason(delta.delta?.stop_reason);
             controller.enqueue({
               type: "finish",
@@ -532,10 +559,16 @@ function convertStream(anthropicStream, modelId) {
     }
   });
 }
-function processEvent(event, blockStates, modelId) {
+function processEvent(event, blockStates, modelId, ctx) {
   const parts = [];
   switch (event.type) {
     case "message_start": {
+      const servedModel = event.message.model;
+      if (ctx.fallbacksEnabled && servedModel && servedModel !== ctx.apiModelId) {
+        ctx.pendingMeta = {
+          servedBy: { from: ctx.apiModelId, to: servedModel, kind: "sticky" }
+        };
+      }
       parts.push({
         type: "response-metadata",
         id: event.message.id,
@@ -547,14 +580,38 @@ function processEvent(event, blockStates, modelId) {
     case "content_block_start": {
       const index = event.index;
       const block = event.content_block;
-      if (block.type === "text") {
+      const takeMeta = () => {
+        const meta = ctx.pendingMeta;
+        ctx.pendingMeta = undefined;
+        return meta;
+      };
+      if (block.type === "fallback") {
+        ctx.pendingMeta = {
+          fallback: { from: block.from, to: block.to },
+          servedBy: {
+            from: block.from?.model ?? ctx.apiModelId,
+            to: block.to?.model ?? "unknown",
+            kind: "fallback"
+          }
+        };
+      } else if (block.type === "text") {
         const id = generateId();
-        blockStates.set(index, { type: "text", id });
-        parts.push({ type: "text-start", id });
+        const meta = takeMeta();
+        blockStates.set(index, { type: "text", id, fallbackMeta: meta });
+        parts.push({
+          type: "text-start",
+          id,
+          ...meta ? { providerMetadata: { anthropic: meta } } : {}
+        });
       } else if (block.type === "thinking") {
         const id = generateId();
-        blockStates.set(index, { type: "thinking", id });
-        parts.push({ type: "reasoning-start", id });
+        const meta = takeMeta();
+        blockStates.set(index, { type: "thinking", id, fallbackMeta: meta });
+        parts.push({
+          type: "reasoning-start",
+          id,
+          ...meta ? { providerMetadata: { anthropic: meta } } : {}
+        });
       } else if (block.type === "tool_use") {
         const toolId = block.id;
         const toolName = toOpencodeToolName(block.name);
@@ -563,7 +620,8 @@ function processEvent(event, blockStates, modelId) {
           id: toolId,
           toolUseId: toolId,
           toolName,
-          argsText: ""
+          argsText: "",
+          fallbackMeta: takeMeta()
         });
         parts.push({
           type: "tool-input-start",
@@ -611,12 +669,14 @@ function processEvent(event, blockStates, modelId) {
       if (state.type === "text") {
         parts.push({ type: "text-end", id: state.id });
       } else if (state.type === "thinking") {
+        const meta = {
+          ...state.signature ? { signature: state.signature } : {},
+          ...state.fallbackMeta
+        };
         parts.push({
           type: "reasoning-end",
           id: state.id,
-          ...state.signature ? {
-            providerMetadata: { anthropic: { signature: state.signature } }
-          } : {}
+          ...Object.keys(meta).length > 0 ? { providerMetadata: { anthropic: meta } } : {}
         });
       } else if (state.type === "tool_use") {
         parts.push({ type: "tool-input-end", id: state.id });
@@ -624,7 +684,8 @@ function processEvent(event, blockStates, modelId) {
           type: "tool-call",
           toolCallId: state.toolUseId,
           toolName: state.toolName,
-          input: state.argsText || "{}"
+          input: state.argsText || "{}",
+          ...state.fallbackMeta ? { providerMetadata: { anthropic: state.fallbackMeta } } : {}
         });
       }
       blockStates.delete(index);
@@ -695,6 +756,17 @@ function handleApiError(error) {
   }
   throw error;
 }
+function isRefusal(stopReason) {
+  return stopReason === "refusal";
+}
+function refusalError(stopDetails, fallbackModel) {
+  const category = stopDetails?.category ?? null;
+  const explanation = stopDetails?.explanation ?? null;
+  const categoryNote = category ? ` (category: ${category})` : "";
+  const detail = explanation ? explanation : "Claude Fable 5's safety classifiers declined this request.";
+  const hint = fallbackModel ? `The configured fallback model (${fallbackModel}) also refused.` : `Retry with a fallback model such as anthropic-sdk/claude-opus-4-8.`;
+  return new Error(`Claude Fable 5 refused this request${categoryNote}. ${detail} ${hint}`);
+}
 function mapFinishReason2(stopReason) {
   const unified = (() => {
     switch (stopReason) {
@@ -711,9 +783,10 @@ function mapFinishReason2(stopReason) {
   })();
   return { unified, raw: stopReason ?? undefined };
 }
+var FALLBACK_BETAS_HEADER = "x-anthropic-sdk-fallback-betas";
 var BILLING_SYSTEM_BLOCK = {
   type: "text",
-  text: "x-anthropic-billing-header: cc_version=2.1.154.cea; cc_entrypoint=sdk-cli; cch=00000;"
+  text: "x-anthropic-billing-header: cc_version=2.1.173.d11; cc_entrypoint=sdk-cli; cch=00000;"
 };
 var IDENTITY_SYSTEM_BLOCK = {
   type: "text",
@@ -847,6 +920,9 @@ class AnthropicSDKModel {
       }
       if (anthropicOptions.metadata)
         params.metadata = anthropicOptions.metadata;
+      if (typeof anthropicOptions.refusalFallback === "string" && anthropicOptions.refusalFallback.length > 0) {
+        params.fallbacks = [{ model: anthropicOptions.refusalFallback }];
+      }
     }
     if (this.isOAuth && supportsContextManagement(this.apiModelId)) {
       const edits = [];
@@ -861,16 +937,19 @@ class AnthropicSDKModel {
         params.context_management = { edits };
       }
     }
-    return { params, warnings };
+    const needsFallbackBetas = params.fallbacks != null || messages.some((msg) => Array.isArray(msg.content) && msg.content.some((block) => block?.type === "fallback"));
+    return { params, warnings, needsFallbackBetas };
   }
-  buildRequestOptions(signal) {
+  buildRequestOptions(signal, needsFallbackBetas) {
     const opts = {};
     if (signal)
       opts.signal = signal;
+    if (needsFallbackBetas)
+      opts.headers = { [FALLBACK_BETAS_HEADER]: "1" };
     return opts;
   }
   async doGenerate(options) {
-    const { params, warnings } = this.buildParams(options);
+    const { params, warnings, needsFallbackBetas } = this.buildParams(options);
     if (!options.maxOutputTokens && params.max_tokens > 4096) {
       params.max_tokens = 4096;
     }
@@ -879,31 +958,74 @@ class AnthropicSDKModel {
       response = await this.client.messages.create({
         ...params,
         stream: false
-      }, this.buildRequestOptions(options.abortSignal));
+      }, this.buildRequestOptions(options.abortSignal, needsFallbackBetas));
     } catch (error) {
       handleApiError(error);
     }
+    const fallbacksSent = Array.isArray(params.fallbacks) && params.fallbacks.length > 0;
+    if (isRefusal(response.stop_reason)) {
+      throw refusalError(response.stop_details, fallbacksSent ? params.fallbacks[0].model : undefined);
+    }
+    const fallbackBlock = response.content.find((b) => b.type === "fallback");
+    const sticky = !fallbackBlock && fallbacksSent && response.model && response.model !== this.apiModelId;
+    let pendingMeta;
+    if (fallbackBlock) {
+      pendingMeta = {
+        fallback: { from: fallbackBlock.from, to: fallbackBlock.to },
+        servedBy: {
+          from: fallbackBlock.from?.model ?? this.apiModelId,
+          to: fallbackBlock.to?.model ?? "unknown",
+          kind: "fallback"
+        }
+      };
+    } else if (sticky) {
+      pendingMeta = {
+        servedBy: { from: this.apiModelId, to: response.model, kind: "sticky" }
+      };
+    }
     const content = [];
+    const takeMeta = () => {
+      const meta = pendingMeta;
+      pendingMeta = undefined;
+      return meta;
+    };
+    let seenFallbackBlock = false;
     for (const block of response.content) {
       switch (block.type) {
-        case "text":
-          content.push({ type: "text", text: block.text });
+        case "text": {
+          const meta = seenFallbackBlock || sticky ? takeMeta() : undefined;
+          content.push({
+            type: "text",
+            text: block.text,
+            ...meta ? { providerMetadata: { anthropic: meta } } : {}
+          });
           break;
-        case "tool_use":
+        }
+        case "tool_use": {
+          const meta = seenFallbackBlock || sticky ? takeMeta() : undefined;
           content.push({
             type: "tool-call",
             toolCallId: block.id,
             toolName: toOpencodeToolName(block.name),
-            input: JSON.stringify(block.input)
+            input: JSON.stringify(block.input),
+            ...meta ? { providerMetadata: { anthropic: meta } } : {}
           });
           break;
+        }
         default: {
           const anyBlock = block;
-          if (anyBlock.type === "thinking" && anyBlock.thinking) {
+          if (anyBlock.type === "fallback") {
+            seenFallbackBlock = true;
+          } else if (anyBlock.type === "thinking" && anyBlock.thinking) {
+            const extra = seenFallbackBlock || sticky ? takeMeta() : undefined;
+            const meta = {
+              ...anyBlock.signature ? { signature: anyBlock.signature } : {},
+              ...extra
+            };
             content.push({
               type: "reasoning",
               text: anyBlock.thinking,
-              providerMetadata: anyBlock.signature ? { anthropic: { signature: anyBlock.signature } } : undefined
+              providerMetadata: Object.keys(meta).length > 0 ? { anthropic: meta } : undefined
             });
           }
           break;
@@ -942,13 +1064,13 @@ class AnthropicSDKModel {
     };
   }
   async doStream(options) {
-    const { params, warnings } = this.buildParams(options);
+    const { params, warnings, needsFallbackBetas } = this.buildParams(options);
     let anthropicStream;
     try {
       anthropicStream = await this.client.messages.create({
         ...params,
         stream: true
-      }, this.buildRequestOptions(options.abortSignal));
+      }, this.buildRequestOptions(options.abortSignal, needsFallbackBetas));
     } catch (error) {
       if (error instanceof APIError && error.status === 400 && error.message?.includes("must end with a user message")) {
         return {
@@ -976,7 +1098,12 @@ class AnthropicSDKModel {
       }
       handleApiError(error);
     }
-    const stream = convertStream(anthropicStream, this.modelId);
+    const fallbacksSent = Array.isArray(params.fallbacks) && params.fallbacks.length > 0;
+    const stream = convertStream(anthropicStream, this.modelId, {
+      apiModelId: this.apiModelId,
+      fallbacksEnabled: fallbacksSent,
+      fallbackModel: fallbacksSent ? params.fallbacks[0].model : undefined
+    });
     const wrappedStream = new ReadableStream({
       async start(controller) {
         controller.enqueue({
@@ -1099,13 +1226,14 @@ var API_TTL_MS = 5 * 60000;
 
 // src/index.ts
 var CLAUDE_CODE_SYSTEM_PROMPT = claudecode_system_default;
-var CLAUDE_CODE_OPUS48_SYSTEM_PROMPT = claudecode_system_opus48_default;
+var CLAUDE_CODE_NEW_SYSTEM_PROMPT = claudecode_system_new_default;
 function selectClaudePromptForModel(modelId) {
-  if (modelId.includes("opus-4-8"))
-    return claudecode_system_opus48_default;
+  if (modelId.includes("opus-4-8") || modelId.includes("fable-5")) {
+    return claudecode_system_new_default;
+  }
   return claudecode_system_default;
 }
-var CLAUDE_CODE_VERSION = "2.1.154";
+var CLAUDE_CODE_VERSION = "2.1.173";
 var OAUTH_BETAS = [
   "claude-code-20250219",
   "oauth-2025-04-20",
@@ -1122,7 +1250,8 @@ var PROVIDER_ID = "anthropic-sdk";
 var API_KEY_COSTS = {
   haiku: { input: 1, output: 5, cache_read: 0.1, cache_write: 1.25 },
   sonnet: { input: 3, output: 15, cache_read: 0.3, cache_write: 3.75 },
-  opus: { input: 5, output: 25, cache_read: 0.5, cache_write: 6.25 }
+  opus: { input: 5, output: 25, cache_read: 0.5, cache_write: 6.25 },
+  fable: { input: 10, output: 50, cache_read: 1, cache_write: 12.5 }
 };
 var ZERO_COST = { input: 0, output: 0, cache_read: 0, cache_write: 0 };
 function buildPluginModels(isOAuth) {
@@ -1236,6 +1365,21 @@ function buildPluginModels(isOAuth) {
       },
       options: { effort: "medium" },
       ...opus47Variants
+    },
+    "claude-fable-5": {
+      name: "Claude Fable 5",
+      reasoning: true,
+      tool_call: true,
+      attachment: true,
+      temperature: true,
+      limit: { context: 1e6, output: 128000 },
+      cost: cost("fable"),
+      modalities: {
+        input: ["text", "image", "pdf"],
+        output: ["text"]
+      },
+      options: { effort: "medium", refusalFallback: "claude-opus-4-8" },
+      ...opus47Variants
     }
   };
 }
@@ -1300,7 +1444,7 @@ function createAnthropicSDK(options = {}) {
         if (!betas.includes("effort-2025-11-24"))
           betas.push("effort-2025-11-24");
       }
-      if (model.includes("opus-4-8")) {
+      if (model.includes("opus-4-8") || model.includes("fable-5")) {
         if (!betas.includes("mid-conversation-system-2026-04-07")) {
           betas.push("mid-conversation-system-2026-04-07");
         }
@@ -1308,6 +1452,19 @@ function createAnthropicSDK(options = {}) {
       reqHeaders.set("anthropic-beta", betas.join(","));
       init = { ...init, headers: reqHeaders };
     }
+    {
+      const reqHeaders = new Headers(init?.headers);
+      if (reqHeaders.has(FALLBACK_BETAS_HEADER)) {
+        reqHeaders.delete(FALLBACK_BETAS_HEADER);
+        const betas = (reqHeaders.get("anthropic-beta") ?? "").split(",").map((beta) => beta.trim()).filter(Boolean);
+        for (const beta of ["server-side-fallback-2026-06-01", "fallback-credit-2026-06-01"]) {
+          if (!betas.includes(beta))
+            betas.push(beta);
+        }
+        reqHeaders.set("anthropic-beta", betas.join(","));
+        init = { ...init, headers: reqHeaders };
+      }
+    }
     const resp = await baseFetch(url, init);
     const h5 = resp.headers.get("anthropic-ratelimit-unified-5h-utilization");
     if (h5 != null) {
@@ -1356,10 +1513,22 @@ var anthropicSDKPlugin = async () => {
       if (!cfg.provider) {
         cfg.provider = {};
       }
+      const userProvider = cfg.provider[PROVIDER_ID] ?? {};
+      const defaultModels = buildPluginModels(isOAuth);
+      const models = { ...defaultModels };
+      for (const [modelID, userModel] of Object.entries(userProvider.models ?? {})) {
+        const base = defaultModels[modelID];
+        models[modelID] = base ? {
+          ...base,
+          ...userModel,
+          options: { ...base.options, ...userModel?.options },
+          variants: { ...base.variants, ...userModel?.variants }
+        } : userModel;
+      }
       cfg.provider[PROVIDER_ID] = {
         npm: PACKAGE_NAME,
-        models: buildPluginModels(isOAuth),
-        ...cfg.provider[PROVIDER_ID]
+        ...userProvider,
+        models
       };
     },
     "experimental.chat.system.transform": async (input, output) => {
@@ -1391,5 +1560,5 @@ export {
   createAnthropicSDK,
   anthropicSDKPlugin,
   CLAUDE_CODE_SYSTEM_PROMPT,
-  CLAUDE_CODE_OPUS48_SYSTEM_PROMPT
+  CLAUDE_CODE_NEW_SYSTEM_PROMPT
 };

package/build/model.d.ts

@@ -2,6 +2,31 @@ import type { LanguageModelV3, LanguageModelV3CallOptions, LanguageModelV3Genera
 import type Anthropic from "@anthropic-ai/sdk";
 type DoGenerateResult = LanguageModelV3GenerateResult;
 type DoStreamResult = LanguageModelV3StreamResult;
+/**
+ * Build a clear error for a Fable 5 / Mythos-class safety refusal.
+ *
+ * Fable 5's safety classifiers can decline a request — the Messages API
+ * returns this as a *successful* HTTP 200 with `stop_reason: "refusal"`,
+ * empty `content`, and a `stop_details` object naming the policy area.
+ * Categories: cyber, bio, frontier_llm, reasoning_extraction (may be null).
+ *
+ * For now we surface this as an error so the user gets a clear message
+ * instead of an empty/confusing response. A future enhancement is to
+ * automatically retry on a fallback model (Claude Opus 4.8), per Anthropic's
+ * recommended pattern. See RESEARCH.md.
+ */
+export declare function isRefusal(stopReason: string | null | undefined): boolean;
+export declare function refusalError(stopDetails: any, fallbackModel?: string): Error;
+/**
+ * Internal marker header: set on requests that need the server-side-fallback
+ * betas (request carries a `fallbacks` chain, or the conversation history
+ * echoes a `fallback` block from a previous fallback response). The fetch
+ * wrapper in index.ts translates it into the real `anthropic-beta` flags and
+ * strips it before the request leaves the process. This keeps the decision
+ * here — where the request is built — instead of sniffing the serialized
+ * body in the fetch wrapper (cf. RESEARCH.md Discovery #24/#25).
+ */
+export declare const FALLBACK_BETAS_HEADER = "x-anthropic-sdk-fallback-betas";
 export declare class AnthropicSDKModel implements LanguageModelV3 {
     private client;
     private providerName;

package/build/model.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"model.d.ts","sourceRoot":"","sources":["../src/model.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EACV,eAAe,EACf,0BAA0B,EAG1B,6BAA6B,EAE7B,2BAA2B,EAE5B,MAAM,kBAAkB,CAAA;AACzB,OAAO,KAAK,SAAS,MAAM,mBAAmB,CAAA;AAQ9C,KAAK,gBAAgB,GAAG,6BAA6B,CAAA;AACrD,KAAK,cAAc,GAAG,2BAA2B,CAAA;AA8HjD,qBAAa,iBAAkB,YAAW,eAAe;IAarD,OAAO,CAAC,MAAM;IACd,OAAO,CAAC,YAAY;IACpB,OAAO,CAAC,OAAO;IAdjB,QAAQ,CAAC,oBAAoB,EAAG,IAAI,CAAS;IAC7C,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAA;IACzB,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAA;IACxB,QAAQ,CAAC,aAAa,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,EAAE,CAAC,CAAK;IAErD,mEAAmE;IACnE,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAQ;IACnC,+FAA+F;IAC/F,OAAO,CAAC,QAAQ,CAAC,WAAW,CAAS;gBAGnC,OAAO,EAAE,MAAM,EACP,MAAM,EAAE,SAAS,EACjB,YAAY,EAAE,MAAM,EACpB,OAAO,GAAE,OAAe;IASlC,OAAO,CAAC,WAAW;IA6KnB;;;;OAIG;IACH,OAAO,CAAC,mBAAmB;IAMrB,UAAU,CAAC,OAAO,EAAE,0BAA0B,GAAG,OAAO,CAAC,gBAAgB,CAAC;IAsF1E,QAAQ,CAAC,OAAO,EAAE,0BAA0B,GAAG,OAAO,CAAC,cAAc,CAAC;CA4E7E"}
+{"version":3,"file":"model.d.ts","sourceRoot":"","sources":["../src/model.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EACV,eAAe,EACf,0BAA0B,EAG1B,6BAA6B,EAE7B,2BAA2B,EAE5B,MAAM,kBAAkB,CAAA;AACzB,OAAO,KAAK,SAAS,MAAM,mBAAmB,CAAA;AAQ9C,KAAK,gBAAgB,GAAG,6BAA6B,CAAA;AACrD,KAAK,cAAc,GAAG,2BAA2B,CAAA;AAkDjD;;;;;;;;;;;;GAYG;AACH,wBAAgB,SAAS,CAAC,UAAU,EAAE,MAAM,GAAG,IAAI,GAAG,SAAS,GAAG,OAAO,CAExE;AAED,wBAAgB,YAAY,CAAC,WAAW,EAAE,GAAG,EAAE,aAAa,CAAC,EAAE,MAAM,GAAG,KAAK,CAW5E;AAmBD;;;;;;;;GAQG;AACH,eAAO,MAAM,qBAAqB,mCAAmC,CAAA;AA6DrE,qBAAa,iBAAkB,YAAW,eAAe;IAarD,OAAO,CAAC,MAAM;IACd,OAAO,CAAC,YAAY;IACpB,OAAO,CAAC,OAAO;IAdjB,QAAQ,CAAC,oBAAoB,EAAG,IAAI,CAAS;IAC7C,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAA;IACzB,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAA;IACxB,QAAQ,CAAC,aAAa,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,EAAE,CAAC,CAAK;IAErD,mEAAmE;IACnE,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAQ;IACnC,+FAA+F;IAC/F,OAAO,CAAC,QAAQ,CAAC,WAAW,CAAS;gBAGnC,OAAO,EAAE,MAAM,EACP,MAAM,EAAE,SAAS,EACjB,YAAY,EAAE,MAAM,EACpB,OAAO,GAAE,OAAe;IASlC,OAAO,CAAC,WAAW;IAuMnB;;;;OAIG;IACH,OAAO,CAAC,mBAAmB;IAUrB,UAAU,CAAC,OAAO,EAAE,0BAA0B,GAAG,OAAO,CAAC,gBAAgB,CAAC;IAyJ1E,QAAQ,CAAC,OAAO,EAAE,0BAA0B,GAAG,OAAO,CAAC,cAAc,CAAC;CAiF7E"}

package/build/server.js

@@ -48,8 +48,8 @@ function hasCchPlaceholder(body) {
   return block?.text?.includes("cch=00000") ?? false;
 }
 
-// src/claudecode-system-opus48.txt
-var claudecode_system_opus48_default = `
+// src/claudecode-system-new.txt
+var claudecode_system_new_default = `
 You are an interactive agent that helps users with software engineering tasks.
 
 # Harness
@@ -382,6 +382,14 @@ function convertUserMessage(message) {
 function convertAssistantMessage(message) {
   const content = [];
   for (const part of message.content) {
+    const fallback = part.providerMetadata?.anthropic?.fallback ?? part.providerOptions?.anthropic?.fallback;
+    if (fallback?.from?.model && fallback?.to?.model) {
+      content.push({
+        type: "fallback",
+        from: { model: fallback.from.model },
+        to: { model: fallback.to.model }
+      });
+    }
     switch (part.type) {
       case "text":
         if (part.text.length > 0) {
@@ -460,6 +468,14 @@ var idCounter = 0;
 function generateId() {
   return `block-${Date.now()}-${idCounter++}`;
 }
+function buildRefusalError(stopDetails, fallbackModel) {
+  const category = stopDetails?.category ?? null;
+  const explanation = stopDetails?.explanation ?? null;
+  const categoryNote = category ? ` (category: ${category})` : "";
+  const detail = explanation ? explanation : "Claude Fable 5's safety classifiers declined this request.";
+  const hint = fallbackModel ? `The configured fallback model (${fallbackModel}) also refused.` : `Retry with a fallback model such as anthropic-sdk/claude-opus-4-8.`;
+  return new Error(`Claude Fable 5 refused this request${categoryNote}. ${detail} ${hint}`);
+}
 function mapFinishReason(stopReason) {
   const unified = (() => {
     switch (stopReason) {
@@ -476,8 +492,13 @@ function mapFinishReason(stopReason) {
   })();
   return { unified, raw: stopReason ?? undefined };
 }
-function convertStream(anthropicStream, modelId) {
+function convertStream(anthropicStream, modelId, context) {
   const blockStates = new Map;
+  const ctx = {
+    apiModelId: context?.apiModelId ?? modelId,
+    fallbacksEnabled: context?.fallbacksEnabled ?? false,
+    fallbackModel: context?.fallbackModel
+  };
   let inputTokens;
   let outputTokens;
   let cachedInputTokens;
@@ -486,7 +507,7 @@ function convertStream(anthropicStream, modelId) {
     async start(controller) {
       try {
         for await (const event of anthropicStream) {
-          const parts = processEvent(event, blockStates, modelId);
+          const parts = processEvent(event, blockStates, modelId, ctx);
           if (event.type === "message_start" && event.message.usage) {
             const u = event.message.usage;
             inputTokens = u.input_tokens;
@@ -496,6 +517,12 @@ function convertStream(anthropicStream, modelId) {
           if (event.type === "message_delta") {
             const delta = event;
             outputTokens = delta.usage?.output_tokens;
+            if (delta.delta?.stop_reason === "refusal") {
+              controller.enqueue({
+                type: "error",
+                error: buildRefusalError(delta.delta?.stop_details, ctx.fallbacksEnabled ? ctx.fallbackModel : undefined)
+              });
+            }
             const finishReason = mapFinishReason(delta.delta?.stop_reason);
             controller.enqueue({
               type: "finish",
@@ -532,10 +559,16 @@ function convertStream(anthropicStream, modelId) {
     }
   });
 }
-function processEvent(event, blockStates, modelId) {
+function processEvent(event, blockStates, modelId, ctx) {
   const parts = [];
   switch (event.type) {
     case "message_start": {
+      const servedModel = event.message.model;
+      if (ctx.fallbacksEnabled && servedModel && servedModel !== ctx.apiModelId) {
+        ctx.pendingMeta = {
+          servedBy: { from: ctx.apiModelId, to: servedModel, kind: "sticky" }
+        };
+      }
       parts.push({
         type: "response-metadata",
         id: event.message.id,
@@ -547,14 +580,38 @@ function processEvent(event, blockStates, modelId) {
     case "content_block_start": {
       const index = event.index;
       const block = event.content_block;
-      if (block.type === "text") {
+      const takeMeta = () => {
+        const meta = ctx.pendingMeta;
+        ctx.pendingMeta = undefined;
+        return meta;
+      };
+      if (block.type === "fallback") {
+        ctx.pendingMeta = {
+          fallback: { from: block.from, to: block.to },
+          servedBy: {
+            from: block.from?.model ?? ctx.apiModelId,
+            to: block.to?.model ?? "unknown",
+            kind: "fallback"
+          }
+        };
+      } else if (block.type === "text") {
         const id = generateId();
-        blockStates.set(index, { type: "text", id });
-        parts.push({ type: "text-start", id });
+        const meta = takeMeta();
+        blockStates.set(index, { type: "text", id, fallbackMeta: meta });
+        parts.push({
+          type: "text-start",
+          id,
+          ...meta ? { providerMetadata: { anthropic: meta } } : {}
+        });
       } else if (block.type === "thinking") {
         const id = generateId();
-        blockStates.set(index, { type: "thinking", id });
-        parts.push({ type: "reasoning-start", id });
+        const meta = takeMeta();
+        blockStates.set(index, { type: "thinking", id, fallbackMeta: meta });
+        parts.push({
+          type: "reasoning-start",
+          id,
+          ...meta ? { providerMetadata: { anthropic: meta } } : {}
+        });
       } else if (block.type === "tool_use") {
         const toolId = block.id;
         const toolName = toOpencodeToolName(block.name);
@@ -563,7 +620,8 @@ function processEvent(event, blockStates, modelId) {
           id: toolId,
           toolUseId: toolId,
           toolName,
-          argsText: ""
+          argsText: "",
+          fallbackMeta: takeMeta()
         });
         parts.push({
           type: "tool-input-start",
@@ -611,12 +669,14 @@ function processEvent(event, blockStates, modelId) {
       if (state.type === "text") {
         parts.push({ type: "text-end", id: state.id });
       } else if (state.type === "thinking") {
+        const meta = {
+          ...state.signature ? { signature: state.signature } : {},
+          ...state.fallbackMeta
+        };
         parts.push({
           type: "reasoning-end",
           id: state.id,
-          ...state.signature ? {
-            providerMetadata: { anthropic: { signature: state.signature } }
-          } : {}
+          ...Object.keys(meta).length > 0 ? { providerMetadata: { anthropic: meta } } : {}
         });
       } else if (state.type === "tool_use") {
         parts.push({ type: "tool-input-end", id: state.id });
@@ -624,7 +684,8 @@ function processEvent(event, blockStates, modelId) {
           type: "tool-call",
           toolCallId: state.toolUseId,
           toolName: state.toolName,
-          input: state.argsText || "{}"
+          input: state.argsText || "{}",
+          ...state.fallbackMeta ? { providerMetadata: { anthropic: state.fallbackMeta } } : {}
         });
       }
       blockStates.delete(index);
@@ -695,6 +756,17 @@ function handleApiError(error) {
   }
   throw error;
 }
+function isRefusal(stopReason) {
+  return stopReason === "refusal";
+}
+function refusalError(stopDetails, fallbackModel) {
+  const category = stopDetails?.category ?? null;
+  const explanation = stopDetails?.explanation ?? null;
+  const categoryNote = category ? ` (category: ${category})` : "";
+  const detail = explanation ? explanation : "Claude Fable 5's safety classifiers declined this request.";
+  const hint = fallbackModel ? `The configured fallback model (${fallbackModel}) also refused.` : `Retry with a fallback model such as anthropic-sdk/claude-opus-4-8.`;
+  return new Error(`Claude Fable 5 refused this request${categoryNote}. ${detail} ${hint}`);
+}
 function mapFinishReason2(stopReason) {
   const unified = (() => {
     switch (stopReason) {
@@ -711,9 +783,10 @@ function mapFinishReason2(stopReason) {
   })();
   return { unified, raw: stopReason ?? undefined };
 }
+var FALLBACK_BETAS_HEADER = "x-anthropic-sdk-fallback-betas";
 var BILLING_SYSTEM_BLOCK = {
   type: "text",
-  text: "x-anthropic-billing-header: cc_version=2.1.154.cea; cc_entrypoint=sdk-cli; cch=00000;"
+  text: "x-anthropic-billing-header: cc_version=2.1.173.d11; cc_entrypoint=sdk-cli; cch=00000;"
 };
 var IDENTITY_SYSTEM_BLOCK = {
   type: "text",
@@ -847,6 +920,9 @@ class AnthropicSDKModel {
       }
       if (anthropicOptions.metadata)
         params.metadata = anthropicOptions.metadata;
+      if (typeof anthropicOptions.refusalFallback === "string" && anthropicOptions.refusalFallback.length > 0) {
+        params.fallbacks = [{ model: anthropicOptions.refusalFallback }];
+      }
     }
     if (this.isOAuth && supportsContextManagement(this.apiModelId)) {
       const edits = [];
@@ -861,16 +937,19 @@ class AnthropicSDKModel {
         params.context_management = { edits };
       }
     }
-    return { params, warnings };
+    const needsFallbackBetas = params.fallbacks != null || messages.some((msg) => Array.isArray(msg.content) && msg.content.some((block) => block?.type === "fallback"));
+    return { params, warnings, needsFallbackBetas };
   }
-  buildRequestOptions(signal) {
+  buildRequestOptions(signal, needsFallbackBetas) {
     const opts = {};
     if (signal)
       opts.signal = signal;
+    if (needsFallbackBetas)
+      opts.headers = { [FALLBACK_BETAS_HEADER]: "1" };
     return opts;
   }
   async doGenerate(options) {
-    const { params, warnings } = this.buildParams(options);
+    const { params, warnings, needsFallbackBetas } = this.buildParams(options);
     if (!options.maxOutputTokens && params.max_tokens > 4096) {
       params.max_tokens = 4096;
     }
@@ -879,31 +958,74 @@ class AnthropicSDKModel {
       response = await this.client.messages.create({
         ...params,
         stream: false
-      }, this.buildRequestOptions(options.abortSignal));
+      }, this.buildRequestOptions(options.abortSignal, needsFallbackBetas));
     } catch (error) {
       handleApiError(error);
     }
+    const fallbacksSent = Array.isArray(params.fallbacks) && params.fallbacks.length > 0;
+    if (isRefusal(response.stop_reason)) {
+      throw refusalError(response.stop_details, fallbacksSent ? params.fallbacks[0].model : undefined);
+    }
+    const fallbackBlock = response.content.find((b) => b.type === "fallback");
+    const sticky = !fallbackBlock && fallbacksSent && response.model && response.model !== this.apiModelId;
+    let pendingMeta;
+    if (fallbackBlock) {
+      pendingMeta = {
+        fallback: { from: fallbackBlock.from, to: fallbackBlock.to },
+        servedBy: {
+          from: fallbackBlock.from?.model ?? this.apiModelId,
+          to: fallbackBlock.to?.model ?? "unknown",
+          kind: "fallback"
+        }
+      };
+    } else if (sticky) {
+      pendingMeta = {
+        servedBy: { from: this.apiModelId, to: response.model, kind: "sticky" }
+      };
+    }
     const content = [];
+    const takeMeta = () => {
+      const meta = pendingMeta;
+      pendingMeta = undefined;
+      return meta;
+    };
+    let seenFallbackBlock = false;
     for (const block of response.content) {
       switch (block.type) {
-        case "text":
-          content.push({ type: "text", text: block.text });
+        case "text": {
+          const meta = seenFallbackBlock || sticky ? takeMeta() : undefined;
+          content.push({
+            type: "text",
+            text: block.text,
+            ...meta ? { providerMetadata: { anthropic: meta } } : {}
+          });
           break;
-        case "tool_use":
+        }
+        case "tool_use": {
+          const meta = seenFallbackBlock || sticky ? takeMeta() : undefined;
           content.push({
             type: "tool-call",
             toolCallId: block.id,
             toolName: toOpencodeToolName(block.name),
-            input: JSON.stringify(block.input)
+            input: JSON.stringify(block.input),
+            ...meta ? { providerMetadata: { anthropic: meta } } : {}
           });
           break;
+        }
         default: {
           const anyBlock = block;
-          if (anyBlock.type === "thinking" && anyBlock.thinking) {
+          if (anyBlock.type === "fallback") {
+            seenFallbackBlock = true;
+          } else if (anyBlock.type === "thinking" && anyBlock.thinking) {
+            const extra = seenFallbackBlock || sticky ? takeMeta() : undefined;
+            const meta = {
+              ...anyBlock.signature ? { signature: anyBlock.signature } : {},
+              ...extra
+            };
             content.push({
               type: "reasoning",
               text: anyBlock.thinking,
-              providerMetadata: anyBlock.signature ? { anthropic: { signature: anyBlock.signature } } : undefined
+              providerMetadata: Object.keys(meta).length > 0 ? { anthropic: meta } : undefined
             });
           }
           break;
@@ -942,13 +1064,13 @@ class AnthropicSDKModel {
     };
   }
   async doStream(options) {
-    const { params, warnings } = this.buildParams(options);
+    const { params, warnings, needsFallbackBetas } = this.buildParams(options);
     let anthropicStream;
     try {
       anthropicStream = await this.client.messages.create({
         ...params,
         stream: true
-      }, this.buildRequestOptions(options.abortSignal));
+      }, this.buildRequestOptions(options.abortSignal, needsFallbackBetas));
     } catch (error) {
       if (error instanceof APIError && error.status === 400 && error.message?.includes("must end with a user message")) {
         return {
@@ -976,7 +1098,12 @@ class AnthropicSDKModel {
       }
       handleApiError(error);
     }
-    const stream = convertStream(anthropicStream, this.modelId);
+    const fallbacksSent = Array.isArray(params.fallbacks) && params.fallbacks.length > 0;
+    const stream = convertStream(anthropicStream, this.modelId, {
+      apiModelId: this.apiModelId,
+      fallbacksEnabled: fallbacksSent,
+      fallbackModel: fallbacksSent ? params.fallbacks[0].model : undefined
+    });
     const wrappedStream = new ReadableStream({
       async start(controller) {
         controller.enqueue({
@@ -1099,13 +1226,14 @@ var API_TTL_MS = 5 * 60000;
 
 // src/index.ts
 var CLAUDE_CODE_SYSTEM_PROMPT = claudecode_system_default;
-var CLAUDE_CODE_OPUS48_SYSTEM_PROMPT = claudecode_system_opus48_default;
+var CLAUDE_CODE_NEW_SYSTEM_PROMPT = claudecode_system_new_default;
 function selectClaudePromptForModel(modelId) {
-  if (modelId.includes("opus-4-8"))
-    return claudecode_system_opus48_default;
+  if (modelId.includes("opus-4-8") || modelId.includes("fable-5")) {
+    return claudecode_system_new_default;
+  }
   return claudecode_system_default;
 }
-var CLAUDE_CODE_VERSION = "2.1.154";
+var CLAUDE_CODE_VERSION = "2.1.173";
 var OAUTH_BETAS = [
   "claude-code-20250219",
   "oauth-2025-04-20",
@@ -1122,7 +1250,8 @@ var PROVIDER_ID = "anthropic-sdk";
 var API_KEY_COSTS = {
   haiku: { input: 1, output: 5, cache_read: 0.1, cache_write: 1.25 },
   sonnet: { input: 3, output: 15, cache_read: 0.3, cache_write: 3.75 },
-  opus: { input: 5, output: 25, cache_read: 0.5, cache_write: 6.25 }
+  opus: { input: 5, output: 25, cache_read: 0.5, cache_write: 6.25 },
+  fable: { input: 10, output: 50, cache_read: 1, cache_write: 12.5 }
 };
 var ZERO_COST = { input: 0, output: 0, cache_read: 0, cache_write: 0 };
 function buildPluginModels(isOAuth) {
@@ -1236,6 +1365,21 @@ function buildPluginModels(isOAuth) {
       },
       options: { effort: "medium" },
       ...opus47Variants
+    },
+    "claude-fable-5": {
+      name: "Claude Fable 5",
+      reasoning: true,
+      tool_call: true,
+      attachment: true,
+      temperature: true,
+      limit: { context: 1e6, output: 128000 },
+      cost: cost("fable"),
+      modalities: {
+        input: ["text", "image", "pdf"],
+        output: ["text"]
+      },
+      options: { effort: "medium", refusalFallback: "claude-opus-4-8" },
+      ...opus47Variants
     }
   };
 }
@@ -1300,7 +1444,7 @@ function createAnthropicSDK(options = {}) {
         if (!betas.includes("effort-2025-11-24"))
           betas.push("effort-2025-11-24");
       }
-      if (model.includes("opus-4-8")) {
+      if (model.includes("opus-4-8") || model.includes("fable-5")) {
         if (!betas.includes("mid-conversation-system-2026-04-07")) {
           betas.push("mid-conversation-system-2026-04-07");
         }
@@ -1308,6 +1452,19 @@ function createAnthropicSDK(options = {}) {
       reqHeaders.set("anthropic-beta", betas.join(","));
       init = { ...init, headers: reqHeaders };
     }
+    {
+      const reqHeaders = new Headers(init?.headers);
+      if (reqHeaders.has(FALLBACK_BETAS_HEADER)) {
+        reqHeaders.delete(FALLBACK_BETAS_HEADER);
+        const betas = (reqHeaders.get("anthropic-beta") ?? "").split(",").map((beta) => beta.trim()).filter(Boolean);
+        for (const beta of ["server-side-fallback-2026-06-01", "fallback-credit-2026-06-01"]) {
+          if (!betas.includes(beta))
+            betas.push(beta);
+        }
+        reqHeaders.set("anthropic-beta", betas.join(","));
+        init = { ...init, headers: reqHeaders };
+      }
+    }
     const resp = await baseFetch(url, init);
     const h5 = resp.headers.get("anthropic-ratelimit-unified-5h-utilization");
     if (h5 != null) {
@@ -1356,10 +1513,22 @@ var anthropicSDKPlugin = async () => {
       if (!cfg.provider) {
         cfg.provider = {};
       }
+      const userProvider = cfg.provider[PROVIDER_ID] ?? {};
+      const defaultModels = buildPluginModels(isOAuth);
+      const models = { ...defaultModels };
+      for (const [modelID, userModel] of Object.entries(userProvider.models ?? {})) {
+        const base = defaultModels[modelID];
+        models[modelID] = base ? {
+          ...base,
+          ...userModel,
+          options: { ...base.options, ...userModel?.options },
+          variants: { ...base.variants, ...userModel?.variants }
+        } : userModel;
+      }
       cfg.provider[PROVIDER_ID] = {
         npm: PACKAGE_NAME,
-        models: buildPluginModels(isOAuth),
-        ...cfg.provider[PROVIDER_ID]
+        ...userProvider,
+        models
       };
     },
     "experimental.chat.system.transform": async (input, output) => {

package/build/stream.d.ts

@@ -1,6 +1,36 @@
 import type { LanguageModelV3StreamPart } from "@ai-sdk/provider";
 import type Anthropic from "@anthropic-ai/sdk";
 type MessageStreamEvent = Anthropic.MessageStreamEvent;
-export declare function convertStream(anthropicStream: AsyncIterable<MessageStreamEvent>, modelId: string): ReadableStream<LanguageModelV3StreamPart>;
+/**
+ * Options and mutable state for one streamed response.
+ *
+ * Server-side fallback (Fable 5 safety refusals → Opus 4.8) surfaces in two
+ * ways that we translate into part metadata:
+ *
+ *   - A `fallback` content block (content_block_start/stop, no deltas) marks
+ *     the hop where the declining model's output gives way to the fallback
+ *     model. We don't emit a part for it; instead we carry two keys on the
+ *     NEXT block's part metadata (OpenCode persists part metadata and syncs
+ *     it to the TUI — the same round trip thinking signatures use):
+ *       anthropic.fallback — {from, to}; prompt.ts re-inserts the block into
+ *         history verbatim. Position is load-bearing for thinking
+ *         verification, so it must precede exactly this block.
+ *       anthropic.servedBy — {from, to, kind} display-only notice for the
+ *         TUI plugin (toast + sidebar). Never echoed back.
+ *   - Sticky-routed follow-up turns carry no block; message_start announcing
+ *     a different model than requested is the signal. servedBy (kind
+ *     "sticky") attaches to the first content block.
+ */
+interface StreamContext {
+    /** Model ID requested by the caller (without -1m suffix). */
+    apiModelId: string;
+    /** Whether the request carried a `fallbacks` chain. */
+    fallbacksEnabled: boolean;
+    /** First fallback model — used for error wording when the chain refuses. */
+    fallbackModel?: string;
+    /** Metadata waiting to be attached to the next content block. */
+    pendingMeta?: Record<string, any>;
+}
+export declare function convertStream(anthropicStream: AsyncIterable<MessageStreamEvent>, modelId: string, context?: Partial<StreamContext>): ReadableStream<LanguageModelV3StreamPart>;
 export {};
 //# sourceMappingURL=stream.d.ts.map

package/build/stream.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"stream.d.ts","sourceRoot":"","sources":["../src/stream.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAA+B,yBAAyB,EAAE,MAAM,kBAAkB,CAAA;AAC9F,OAAO,KAAK,SAAS,MAAM,mBAAmB,CAAA;AAG9C,KAAK,kBAAkB,GAAG,SAAS,CAAC,kBAAkB,CAAA;AAmCtD,wBAAgB,aAAa,CAC3B,eAAe,EAAE,aAAa,CAAC,kBAAkB,CAAC,EAClD,OAAO,EAAE,MAAM,GACd,cAAc,CAAC,yBAAyB,CAAC,CAwD3C"}
+{"version":3,"file":"stream.d.ts","sourceRoot":"","sources":["../src/stream.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAA+B,yBAAyB,EAAE,MAAM,kBAAkB,CAAA;AAC9F,OAAO,KAAK,SAAS,MAAM,mBAAmB,CAAA;AAG9C,KAAK,kBAAkB,GAAG,SAAS,CAAC,kBAAkB,CAAA;AAetD;;;;;;;;;;;;;;;;;;;GAmBG;AACH,UAAU,aAAa;IACrB,6DAA6D;IAC7D,UAAU,EAAE,MAAM,CAAA;IAClB,uDAAuD;IACvD,gBAAgB,EAAE,OAAO,CAAA;IACzB,4EAA4E;IAC5E,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB,iEAAiE;IACjE,WAAW,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,GAAG,CAAC,CAAA;CAClC;AA0CD,wBAAgB,aAAa,CAC3B,eAAe,EAAE,aAAa,CAAC,kBAAkB,CAAC,EAClD,OAAO,EAAE,MAAM,EACf,OAAO,CAAC,EAAE,OAAO,CAAC,aAAa,CAAC,GAC/B,cAAc,CAAC,yBAAyB,CAAC,CA0E3C"}

package/build/tui.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"tui.d.ts","sourceRoot":"","sources":["../src/tui.tsx"],"names":[],"mappings":"AAAA,sCAAsC;AACtC,OAAO,KAAK,EAGV,eAAe,EAEhB,MAAM,yBAAyB,CAAA;AAoYhC,QAAA,MAAM,MAAM,EAAE,eAAe,GAAG;IAAE,EAAE,EAAE,MAAM,CAAA;CAG3C,CAAA;AAED,eAAe,MAAM,CAAA"}
+{"version":3,"file":"tui.d.ts","sourceRoot":"","sources":["../src/tui.tsx"],"names":[],"mappings":"AAAA,sCAAsC;AACtC,OAAO,KAAK,EAGV,eAAe,EAEhB,MAAM,yBAAyB,CAAA;AAujBhC,QAAA,MAAM,MAAM,EAAE,eAAe,GAAG;IAAE,EAAE,EAAE,MAAM,CAAA;CAG3C,CAAA;AAED,eAAe,MAAM,CAAA"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@xtruder/opencode-claude-max-plugin",
-  "version": "0.3.1",
+  "version": "0.4.0",
   "description": "OpenCode plugin for Claude Max/Pro subscription via @anthropic-ai/sdk - uses Claude Code OAuth credentials",
   "keywords": [
     "ai-sdk",
@@ -49,7 +49,7 @@
     "prepare": "simple-git-hooks"
   },
   "dependencies": {
-    "@anthropic-ai/sdk": "0.94.0",
+    "@anthropic-ai/sdk": "0.104.1",
     "xxhash-wasm": "^1.1.0"
   },
   "devDependencies": {

package/src/tui.tsx

@@ -132,6 +132,144 @@ function createUsageStore(pollInterval: number) {
   }
 }
 
+// ─── Fallback store ──────────────────────────────────────────────────
+
+/**
+ * Watches for server-side model fallback (Fable 5 safety refusal → Opus 4.8).
+ *
+ * The provider attaches a display-only `anthropic.servedBy` object to part
+ * metadata when a response was served by a fallback model (see stream.ts).
+ * OpenCode persists part metadata and forwards it here via
+ * `message.part.updated` — no side-channel store is needed.
+ *
+ * Sticky routing is keyed on the conversation prefix server-side, so the
+ * fallback state is PER SESSION — a refusal in one session says nothing
+ * about another.
+ *
+ * The sidebar state is DERIVED from the session's persisted messages (via
+ * api.state, OpenCode's synced replica) rather than accumulated from live
+ * events — so it is correct after a TUI restart or when opening an old
+ * session. The event subscription only drives the live toast, deduped with
+ * an in-memory Set (part events never replay across restarts).
+ */
+type FallbackInfo = {
+  from: string
+  to: string
+  kind: "fallback" | "sticky"
+  at: number
+}
+
+const shortModel = (model: string) => model.replace(/^claude-/, "")
+
+// Absolute time — stable without periodic re-rendering ("ago" phrasing
+// would freeze between reactive updates).
+const atTime = (at: number) =>
+  new Date(at).toLocaleTimeString(undefined, { hour: "numeric", minute: "2-digit" })
+
+function createFallbackStore(api: Api) {
+  // Dedupe the repeated part-update events that fire as a part streams.
+  const seenParts = new Set<string>()
+
+  /** Extract the servedBy marker from one assistant message's parts. */
+  const markerOf = (msg: any): FallbackInfo | null => {
+    let latest: FallbackInfo | null = null
+    for (const part of api.state.part(msg.id)) {
+      const servedBy = (part as any)?.metadata?.anthropic?.servedBy
+      if (!servedBy?.to) continue
+      const time = (part as any)?.time
+      const at = time?.end ?? time?.start ?? msg.time?.created ?? 0
+      latest = {
+        from: String(servedBy.from ?? "unknown"),
+        to: String(servedBy.to),
+        kind: servedBy.kind === "sticky" ? "sticky" : "fallback",
+        at,
+      }
+    }
+    return latest
+  }
+
+  /**
+   * Walk a session's messages newest-first, yielding real assistant turns
+   * (skipping title-generation / compaction summaries, which never carry
+   * routing markers and would mask the real last turn).
+   */
+  function* assistantTurns(sessionID: string, beforeMessageID?: string) {
+    const messages = api.state.session.messages(sessionID)
+    let skipping = beforeMessageID !== undefined
+    for (let i = messages.length - 1; i >= 0; i--) {
+      const msg = messages[i] as any
+      if (skipping) {
+        if (msg?.id === beforeMessageID) skipping = false
+        continue
+      }
+      if (msg?.role !== "assistant") continue
+      if (msg.summary) continue
+      yield msg
+    }
+  }
+
+  // Live toast — only on the TRANSITION into fallback. A conversation that
+  // trips the classifier every turn would otherwise toast on every message;
+  // the sidebar line covers the continuing state. Sticky follow-ups never
+  // toast for the same reason.
+  const off = api.event.on("message.part.updated", (event) => {
+    const part = (event as any).properties?.part
+    const sessionID = (event as any).properties?.sessionID ?? part?.sessionID
+    const servedBy = part?.metadata?.anthropic?.servedBy
+    if (!servedBy?.to || !part?.id || !part?.messageID || !sessionID) return
+    if (seenParts.has(part.id)) return
+    seenParts.add(part.id)
+    if (servedBy.kind === "sticky") return
+
+    // Was the previous completed assistant turn already fallback-served?
+    for (const msg of assistantTurns(sessionID, part.messageID)) {
+      if (markerOf(msg)) return // continuation, not a transition
+      if (msg.time?.completed) break // last known state was the primary model
+    }
+
+    api.ui.toast({
+      variant: "warning",
+      title: "Model fallback",
+      message: `${shortModel(String(servedBy.from ?? "unknown"))} refused — answered by ${shortModel(String(servedBy.to))}`,
+      duration: 6000,
+    })
+  })
+
+  /**
+   * Fallback state for one session, derived from its message parts — a pure
+   * reactive derivation that re-evaluates only when session state changes.
+   *
+   * Only the MOST RECENT *settled* assistant turn is consulted: sticky
+   * routing re-emits a servedBy marker on every turn while active, and a
+   * turn that routed back to the primary model carries no marker. Looking
+   * further back would resurrect a stale fallback after the session has
+   * already returned to the requested model.
+   *
+   * A still-streaming turn without a marker is "unknown", not "no
+   * fallback" — the previous completed turn's state keeps showing until
+   * the new turn settles (otherwise the line flickers off on every send
+   * and back on when the refusal hits mid-stream).
+   *
+   * No time window: "the latest turn was served by X" is a fact that stays
+   * true until the next turn says otherwise. Whether the server's sticky
+   * routing is still active is unobservable — we don't guess.
+   */
+  const active = (sessionID: string | undefined): FallbackInfo | null => {
+    if (!sessionID) return null
+    for (const msg of assistantTurns(sessionID)) {
+      const marker = markerOf(msg)
+      if (marker) return marker
+      // No marker on a finished turn → currently served by the requested
+      // model. No marker on a streaming turn → not known yet, keep looking
+      // at the previous turn.
+      if (msg.time?.completed) return null
+    }
+    return null
+  }
+
+  return { active, dispose: off }
+}
+
 // ─── Helpers ─────────────────────────────────────────────────────────
 
 const clamp01 = (v: number) => Math.max(0, Math.min(1, v))
@@ -242,6 +380,34 @@ function UsageSidebar(props: {
   )
 }
 
+// ─── Fallback Sidebar Component ──────────────────────────────────────
+
+function FallbackSidebar(props: {
+  theme: TuiThemeCurrent
+  sessionID: string | undefined
+  active: (sessionID: string | undefined) => FallbackInfo | null
+}) {
+  return (
+    <Show when={props.active(props.sessionID)}>
+      {(info: Accessor<FallbackInfo>) => (
+        <box width="100%" flexDirection="column">
+          <text fg={props.theme.warning}>
+            <b>Model Fallback</b>
+          </text>
+          <text fg={props.theme.text}>
+            {shortModel(info().from)} → {shortModel(info().to)}
+          </text>
+          <text fg={props.theme.textMuted}>
+            {info().kind === "sticky"
+              ? `sticky-routed at ${atTime(info().at)}`
+              : `refused at ${atTime(info().at)}`}
+          </text>
+        </box>
+      )}
+    </Show>
+  )
+}
+
 // ─── Dialog Component ────────────────────────────────────────────────
 
 function UsageDialog(props: {
@@ -353,6 +519,10 @@ const tui: TuiPlugin = async (api, options) => {
   const store = createUsageStore(config.poll_interval)
   api.lifecycle.onDispose(store.dispose)
 
+  // Model fallback notifications (toast) + sidebar status line
+  const fallback = createFallbackStore(api)
+  api.lifecycle.onDispose(fallback.dispose)
+
   // Refresh when session goes idle (inference completed — server just wrote new headers)
   const offEvent = api.event.on("session.idle", () => {
     setTimeout(store.refresh, 1000)
@@ -360,31 +530,40 @@ const tui: TuiPlugin = async (api, options) => {
   api.lifecycle.onDispose(offEvent)
 
   // Register /usage command
-  api.command?.register(() => [
-    {
-      title: "Claude subscription usage",
-      value: "anthropic-sdk.usage",
-      category: "Claude",
-      slash: {
-        name: "usage",
-      },
-      onSelect() {
-        store.refresh()
-        api.ui.dialog.replace(() => (
-          <UsageDialog api={api} usage={store.usage} status={store.status} />
-        ))
+  const offCommand = api.keymap.registerLayer({
+    commands: [
+      {
+        namespace: "palette",
+        name: "anthropic-sdk.usage",
+        title: "Claude subscription usage",
+        category: "Claude",
+        slashName: "usage",
+        run() {
+          store.refresh()
+          api.ui.dialog.replace(() => (
+            <UsageDialog api={api} usage={store.usage} status={store.status} />
+          ))
+        },
       },
-    },
-  ])
+    ],
+  })
+  api.lifecycle.onDispose(offCommand)
 
   // Register sidebar slot
   if (config.sidebar) {
     api.slots.register({
       order: 110,
       slots: {
-        sidebar_content(ctx) {
+        sidebar_content(ctx, props) {
           return (
-            <UsageSidebar theme={ctx.theme.current} usage={store.usage} status={store.status} />
+            <>
+              <UsageSidebar theme={ctx.theme.current} usage={store.usage} status={store.status} />
+              <FallbackSidebar
+                theme={ctx.theme.current}
+                sessionID={(props as { session_id?: string } | undefined)?.session_id}
+                active={fallback.active}
+              />
+            </>
           )
         },
       },