@xtruder/opencode-claude-max-plugin 0.3.1 → 0.4.1

package/README.md

@@ -9,8 +9,9 @@ An [OpenCode](https://opencode.ai/) plugin that enables Claude Pro/Max subscript
 - **Use your Claude subscription** — Automatically reads OAuth credentials from Claude Code, no separate API key needed
 - **Matches Claude Code 2.1.154** — Same request format and behavior as the official CLI
 - **Prompt caching** — Multi-turn conversations cache properly, keeping costs and latency low
-- **All Claude models** — Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 4.6, Haiku 4.5
+- **All Claude models** — Fable 5, Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 4.6, Haiku 4.5
 - **Extended / adaptive thinking** — Full reasoning support across models, including Opus 4.7+'s adaptive thinking
+- **Safety-refusal fallback** — Fable 5 refusals are transparently answered by Opus 4.8 in the same request, with TUI notification
 - **Usage tracking** — Sidebar widget with live progress bars + `/usage` command
 - **Self-registering** — Models are registered automatically, no manual provider config needed
 
@@ -25,9 +26,9 @@ 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.
+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, Fable 5) 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)
 
@@ -54,6 +55,7 @@ The TUI plugin provides:
 - **Sidebar widget** — Compact progress bars for 5-hour session and 7-day weekly usage
 - **`/usage` command** — Opens a dialog with full usage breakdown (per-model, extra usage)
 - **Auto-refresh** — Polls the usage API every 60s and after each inference call
+- **Fallback indicator** — Toast when a Fable 5 refusal falls back to Opus 4.8, plus a sidebar line showing which model served the latest turn
 
 #### TUI Configuration
 
@@ -108,6 +110,36 @@ If you want to override model settings (e.g. thinking budgets, variants), you ca
 
 Config-level settings are merged with plugin defaults — you only need to specify what you want to override.
 
+### Claude Fable 5 and safety-refusal fallback
+
+Claude Fable 5 (`claude-fable-5`) ships with stricter safety classifiers that can refuse a request at the API level (`stop_reason: "refusal"`) — even for benign follow-ups if the conversation contains a flagged topic. To keep sessions usable, the plugin enables Anthropic's server-side fallback by default: when Fable 5 refuses, **Opus 4.8 answers the same request in the same round trip**. Tool loops keep running, thinking chains stay verified, and prompt caching is unaffected.
+
+When a fallback happens:
+
+- The TUI shows a toast (`fable-5 refused — answered by opus-4-8`) on the first fallback turn
+- The sidebar shows a `Model Fallback` line while the latest turn was served by the fallback model
+- The served model is recorded in part metadata (`anthropic.servedBy`) — the model's own self-report will still say `claude-fable-5`, since identity comes from the prompt, not the serving model
+
+Configure via the `refusalFallback` model option:
+
+```json
+{
+  "provider": {
+    "anthropic-sdk": {
+      "models": {
+        "claude-fable-5": {
+          "options": {
+            "refusalFallback": false
+          }
+        }
+      }
+    }
+  }
+}
+```
+
+Set it to another model ID to change the fallback target, or `false` to disable (refusals then surface as errors with the refusal category).
+
 ## Authentication
 
 Credentials are resolved in order:
@@ -124,6 +156,7 @@ For Claude Code credentials, log in via `claude` CLI first (`claude auth login`)
 - MCP tool name remapping (`server_tool` → `mcp__server__tool`)
 - Extended thinking (Sonnet/Opus 4.6) and adaptive thinking (Opus 4.7+) with effort levels and multi-turn signature passthrough
 - Prompt caching that holds across long, tool-heavy conversations
+- Server-side safety-refusal fallback for Fable 5 (configurable, on by default)
 - Subscription rate limit detection — fails fast with a clear message instead of hanging
 - Long-context auto-detection for large prompts
 - TUI sidebar with live usage bars + `/usage` slash 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
@@ -222,6 +222,84 @@ function getCachedCredentials(path) {
 import { APIError, RateLimitError } from "@anthropic-ai/sdk";
 import { createHash, randomBytes } from "node:crypto";
 
+// src/pause-turn.ts
+var MAX_PAUSE_TURN_CONTINUATIONS = 5;
+function isPauseTurn(stopReason) {
+  return stopReason === "pause_turn";
+}
+
+class BlockAccumulator {
+  blocks = [];
+  open = new Map;
+  push(event) {
+    switch (event.type) {
+      case "content_block_start": {
+        this.open.set(event.index, {
+          block: JSON.parse(JSON.stringify(event.content_block))
+        });
+        break;
+      }
+      case "content_block_delta": {
+        const entry = this.open.get(event.index);
+        if (!entry)
+          break;
+        const delta = event.delta;
+        if (delta.type === "text_delta") {
+          entry.block.text = (entry.block.text ?? "") + delta.text;
+        } else if (delta.type === "thinking_delta") {
+          entry.block.thinking = (entry.block.thinking ?? "") + delta.thinking;
+        } else if (delta.type === "signature_delta") {
+          entry.block.signature = (entry.block.signature ?? "") + delta.signature;
+        } else if (delta.type === "input_json_delta") {
+          entry.argsText = (entry.argsText ?? "") + delta.partial_json;
+        }
+        break;
+      }
+      case "content_block_stop": {
+        const entry = this.open.get(event.index);
+        if (!entry)
+          break;
+        this.open.delete(event.index);
+        if (entry.argsText !== undefined) {
+          try {
+            entry.block.input = JSON.parse(entry.argsText || "{}");
+          } catch {
+            entry.block.input = {};
+          }
+        }
+        this.blocks.push(entry.block);
+        break;
+      }
+    }
+  }
+}
+function continuationParams(params, blocks) {
+  if (blocks.length === 0)
+    return params;
+  return {
+    ...params,
+    messages: [...params.messages ?? [], { role: "assistant", content: blocks }]
+  };
+}
+async function* withPauseTurnContinuation(initial, params, request, maxContinuations = MAX_PAUSE_TURN_CONTINUATIONS) {
+  const acc = new BlockAccumulator;
+  let stream = initial;
+  let continuations = 0;
+  outer:
+    while (true) {
+      for await (const event of stream) {
+        if (event.type === "message_delta" && isPauseTurn(event.delta?.stop_reason) && continuations < maxContinuations) {
+          continuations++;
+          stream = await request(continuationParams(params, acc.blocks));
+          continue outer;
+        }
+        acc.push(event);
+        yield event;
+      }
+      return;
+    }
+}
+
 // src/tool-names.ts
 import { existsSync as existsSync2, readFileSync as readFileSync2 } from "node:fs";
 import { homedir as homedir2 } from "node:os";
@@ -382,6 +460,14 @@ 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 +546,14 @@ 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 +570,13 @@ 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 +585,7 @@ 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 +595,18 @@ 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)
+              });
+            }
+            if (delta.delta?.stop_reason === "pause_turn") {
+              controller.enqueue({
+                type: "error",
+                error: new Error(`Anthropic kept pausing this turn (stop_reason "pause_turn") ` + `after the provider's auto-continuation attempts. Retry the request.`)
+              });
+            }
             const finishReason = mapFinishReason(delta.delta?.stop_reason);
             controller.enqueue({
               type: "finish",
@@ -532,10 +643,16 @@ 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 +664,38 @@ 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 +704,8 @@ function processEvent(event, blockStates, modelId) {
           id: toolId,
           toolUseId: toolId,
           toolName,
-          argsText: ""
+          argsText: "",
+          fallbackMeta: takeMeta()
         });
         parts.push({
           type: "tool-input-start",
@@ -611,12 +753,14 @@ 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 +768,8 @@ 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 +840,17 @@ 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 +867,10 @@ 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 +1004,9 @@ 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 +1021,19 @@ 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 +1042,92 @@ 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 pausedBlocks = [];
+    let continuations = 0;
+    while (isPauseTurn(response.stop_reason) && continuations < MAX_PAUSE_TURN_CONTINUATIONS) {
+      continuations++;
+      pausedBlocks.push(...response.content);
+      try {
+        response = await this.client.messages.create({
+          ...continuationParams(params, pausedBlocks),
+          stream: false
+        }, this.buildRequestOptions(options.abortSignal, needsFallbackBetas));
+      } catch (error) {
+        handleApiError(error);
+      }
+    }
+    if (isPauseTurn(response.stop_reason)) {
+      throw new Error(`Anthropic paused this turn (stop_reason "pause_turn") more than ` + `${MAX_PAUSE_TURN_CONTINUATIONS} times. Retry the request.`);
+    }
+    const responseContent = [...pausedBlocks, ...response.content];
+    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 = responseContent.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 = [];
-    for (const block of response.content) {
+    const takeMeta = () => {
+      const meta = pendingMeta;
+      pendingMeta = undefined;
+      return meta;
+    };
+    let seenFallbackBlock = false;
+    for (const block of responseContent) {
       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 +1166,13 @@ 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 +1200,14 @@ class AnthropicSDKModel {
       }
       handleApiError(error);
     }
-    const stream = convertStream(anthropicStream, this.modelId);
+    const makeRequest = (p) => this.client.messages.create({ ...p, stream: true }, this.buildRequestOptions(options.abortSignal, needsFallbackBetas));
+    const continued = withPauseTurnContinuation(anthropicStream, params, makeRequest);
+    const fallbacksSent = Array.isArray(params.fallbacks) && params.fallbacks.length > 0;
+    const stream = convertStream(continued, 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 +1330,14 @@ 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 +1354,8 @@ 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 +1469,21 @@ 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 +1548,7 @@ 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 +1556,19 @@ 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 +1617,22 @@ 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 +1664,5 @@ export {
   createAnthropicSDK,
   anthropicSDKPlugin,
   CLAUDE_CODE_SYSTEM_PROMPT,
-  CLAUDE_CODE_OPUS48_SYSTEM_PROMPT
+  CLAUDE_CODE_NEW_SYSTEM_PROMPT
 };