feat: add two-layer thinking block validation (proactive + reactive) (#248)

- Add thinking-block-validator hook for proactive prevention before API calls
- Enhance session-recovery to include previous thinking content
- Fix hook registration to actually invoke the validator

Addresses extended thinking errors with Claude Opus/Sonnet 4.5 using tool calls.

Related: https://github.com/vercel/ai/issues/7729
Related: https://github.com/sst/opencode/issues/2599

assets/oh-my-opencode.schema.json

@@ -59,7 +59,8 @@
           "agent-usage-reminder",
           "non-interactive-env",
           "interactive-bash-session",
-          "empty-message-sanitizer"
+          "empty-message-sanitizer",
+          "thinking-block-validator"
         ]
       }
     },

src/config/schema.ts

@@ -64,6 +64,7 @@ export const HookNameSchema = z.enum([
   "non-interactive-env",
   "interactive-bash-session",
   "empty-message-sanitizer",
+  "thinking-block-validator",
 ])
 
 export const AgentOverrideConfigSchema = z.object({

src/hooks/index.ts

@@ -21,3 +21,4 @@ export { createKeywordDetectorHook } from "./keyword-detector";
 export { createNonInteractiveEnvHook } from "./non-interactive-env";
 export { createInteractiveBashSessionHook } from "./interactive-bash-session";
 export { createEmptyMessageSanitizerHook } from "./empty-message-sanitizer";
+export { createThinkingBlockValidatorHook } from "./thinking-block-validator";

src/hooks/session-recovery/storage.ts

@@ -223,6 +223,41 @@ export function findMessagesWithOrphanThinking(sessionID: string): string[] {
   return result
 }
 
+/**
+ * Find the most recent thinking content from previous assistant messages
+ * Following Anthropic's recommendation to include thinking blocks from previous turns
+ */
+function findLastThinkingContent(sessionID: string, beforeMessageID: string): string {
+  const messages = readMessages(sessionID)
+
+  // Find the index of the current message
+  const currentIndex = messages.findIndex(m => m.id === beforeMessageID)
+  if (currentIndex === -1) return ""
+
+  // Search backwards through previous assistant messages
+  for (let i = currentIndex - 1; i >= 0; i--) {
+    const msg = messages[i]
+    if (msg.role !== "assistant") continue
+
+    // Look for thinking parts in this message
+    const parts = readParts(msg.id)
+    for (const part of parts) {
+      if (THINKING_TYPES.has(part.type)) {
+        // Found thinking content - return it
+        // Note: 'thinking' type uses 'thinking' property, 'reasoning' type uses 'text' property
+        const thinking = (part as { thinking?: string; text?: string }).thinking
+        const reasoning = (part as { thinking?: string; text?: string }).text
+        const content = thinking || reasoning
+        if (content && content.trim().length > 0) {
+          return content
+        }
+      }
+    }
+  }
+
+  return ""
+}
+
 export function prependThinkingPart(sessionID: string, messageID: string): boolean {
   const partDir = join(PART_STORAGE, messageID)
 
@@ -230,13 +265,16 @@ export function prependThinkingPart(sessionID: string, messageID: string): boole
     mkdirSync(partDir, { recursive: true })
   }
 
+  // Try to get thinking content from previous turns (Anthropic's recommendation)
+  const previousThinking = findLastThinkingContent(sessionID, messageID)
+
   const partId = `prt_0000000000_thinking`
   const part = {
     id: partId,
     sessionID,
     messageID,
     type: "thinking",
-    thinking: "",
+    thinking: previousThinking || "[Continuing from previous reasoning]",
     synthetic: true,
   }
 

src/hooks/thinking-block-validator/index.ts

@@ -0,0 +1,170 @@
+/**
+ * Proactive Thinking Block Validator Hook
+ *
+ * Prevents "Expected thinking/redacted_thinking but found tool_use" errors
+ * by validating and fixing message structure BEFORE sending to Anthropic API.
+ *
+ * This hook runs on the "experimental.chat.messages.transform" hook point,
+ * which is called before messages are converted to ModelMessage format and
+ * sent to the API.
+ *
+ * Key differences from session-recovery hook:
+ * - PROACTIVE (prevents error) vs REACTIVE (fixes after error)
+ * - Runs BEFORE API call vs AFTER API error
+ * - User never sees the error vs User sees error then recovery
+ */
+
+import type { Message, Part } from "@opencode-ai/sdk"
+
+interface MessageWithParts {
+  info: Message
+  parts: Part[]
+}
+
+type MessagesTransformHook = {
+  "experimental.chat.messages.transform"?: (
+    input: Record<string, never>,
+    output: { messages: MessageWithParts[] }
+  ) => Promise<void>
+}
+
+/**
+ * Check if a model has extended thinking enabled
+ * Uses patterns from think-mode/switcher.ts for consistency
+ */
+function isExtendedThinkingModel(modelID: string): boolean {
+  if (!modelID) return false
+  const lower = modelID.toLowerCase()
+
+  // Check for explicit thinking/high variants (always enabled)
+  if (lower.includes("thinking") || lower.endsWith("-high")) {
+    return true
+  }
+
+  // Check for thinking-capable models (claude-4 family, claude-3)
+  // Aligns with THINKING_CAPABLE_MODELS in think-mode/switcher.ts
+  return (
+    lower.includes("claude-sonnet-4") ||
+    lower.includes("claude-opus-4") ||
+    lower.includes("claude-3")
+  )
+}
+
+/**
+ * Check if a message has tool parts (tool_use)
+ */
+function hasToolParts(parts: Part[]): boolean {
+  if (!parts || parts.length === 0) return false
+
+  return parts.some((part: Part) => {
+    const type = part.type as string
+    return type === "tool" || type === "tool_use"
+  })
+}
+
+/**
+ * Check if a message starts with a thinking/reasoning block
+ */
+function startsWithThinkingBlock(parts: Part[]): boolean {
+  if (!parts || parts.length === 0) return false
+
+  const firstPart = parts[0]
+  const type = firstPart.type as string
+  return type === "thinking" || type === "reasoning"
+}
+
+/**
+ * Find the most recent thinking content from previous assistant messages
+ */
+function findPreviousThinkingContent(
+  messages: MessageWithParts[],
+  currentIndex: number
+): string {
+  // Search backwards from current message
+  for (let i = currentIndex - 1; i >= 0; i--) {
+    const msg = messages[i]
+    if (msg.info.role !== "assistant") continue
+
+    // Look for thinking parts
+    if (!msg.parts) continue
+    for (const part of msg.parts) {
+      const type = part.type as string
+      if (type === "thinking" || type === "reasoning") {
+        const thinking = (part as any).thinking || (part as any).text
+        if (thinking && typeof thinking === "string" && thinking.trim().length > 0) {
+          return thinking
+        }
+      }
+    }
+  }
+
+  return ""
+}
+
+/**
+ * Prepend a thinking block to a message's parts array
+ */
+function prependThinkingBlock(
+  message: MessageWithParts,
+  thinkingContent: string
+): void {
+  if (!message.parts) {
+    message.parts = []
+  }
+
+  // Create synthetic thinking part
+  const thinkingPart = {
+    type: "thinking" as const,
+    id: `prt_0000000000_synthetic_thinking`,
+    sessionID: (message.info as any).sessionID || "",
+    messageID: message.info.id,
+    thinking: thinkingContent,
+    synthetic: true,
+  }
+
+  // Prepend to parts array
+  message.parts.unshift(thinkingPart as unknown as Part)
+}
+
+/**
+ * Validate and fix assistant messages that have tool_use but no thinking block
+ */
+export function createThinkingBlockValidatorHook(): MessagesTransformHook {
+  return {
+    "experimental.chat.messages.transform": async (_input, output) => {
+      const { messages } = output
+
+      if (!messages || messages.length === 0) {
+        return
+      }
+
+      // Get the model info from the last user message
+      const lastUserMessage = messages.findLast(m => m.info.role === "user")
+      const modelID = (lastUserMessage?.info as any)?.modelID || ""
+
+      // Only process if extended thinking might be enabled
+      if (!isExtendedThinkingModel(modelID)) {
+        return
+      }
+
+      // Process all assistant messages
+      for (let i = 0; i < messages.length; i++) {
+        const msg = messages[i]
+
+        // Only check assistant messages
+        if (msg.info.role !== "assistant") continue
+
+        // Check if message has tool parts but doesn't start with thinking
+        if (hasToolParts(msg.parts) && !startsWithThinkingBlock(msg.parts)) {
+          // Find thinking content from previous turns
+          const previousThinking = findPreviousThinkingContent(messages, i)
+
+          // Prepend thinking block with content from previous turn or placeholder
+          const thinkingContent = previousThinking || "[Continuing from previous reasoning]"
+
+          prependThinkingBlock(msg, thinkingContent)
+        }
+      }
+    },
+  }
+}

src/index.ts

@@ -23,6 +23,7 @@ import {
   createNonInteractiveEnvHook,
   createInteractiveBashSessionHook,
   createEmptyMessageSanitizerHook,
+  createThinkingBlockValidatorHook,
 } from "./hooks";
 import { createGoogleAntigravityAuthPlugin } from "./auth/antigravity";
 import {
@@ -319,6 +320,9 @@ const OhMyOpenCodePlugin: Plugin = async (ctx) => {
   const emptyMessageSanitizer = isHookEnabled("empty-message-sanitizer")
     ? createEmptyMessageSanitizerHook()
     : null;
+  const thinkingBlockValidator = isHookEnabled("thinking-block-validator")
+    ? createThinkingBlockValidatorHook()
+    : null;
 
   const backgroundManager = new BackgroundManager(ctx);
 
@@ -365,6 +369,8 @@ const OhMyOpenCodePlugin: Plugin = async (ctx) => {
       input: Record<string, never>,
       output: { messages: Array<{ info: unknown; parts: unknown[] }> }
     ) => {
+      // eslint-disable-next-line @typescript-eslint/no-explicit-any
+      await thinkingBlockValidator?.["experimental.chat.messages.transform"]?.(input, output as any);
       // eslint-disable-next-line @typescript-eslint/no-explicit-any
       await emptyMessageSanitizer?.["experimental.chat.messages.transform"]?.(input, output as any);
     },