diff --git a/packages/opencode/src/provider/transform.ts b/packages/opencode/src/provider/transform.ts
index cd29e40822da..5a2abc4ef881 100644
--- a/packages/opencode/src/provider/transform.ts
+++ b/packages/opencode/src/provider/transform.ts
@@ -156,8 +156,11 @@ function normalizeMessages(
         }
         if (!Array.isArray(msg.content)) return msg
         const filtered = msg.content.filter((part) => {
-          if (part.type === "text" || part.type === "reasoning") {
-            return part.text !== ""
+          if (part.type === "text") return part.text !== ""
+          if (part.type === "reasoning") {
+            // Keep reasoning parts that carry providerOptions (e.g. redacted_thinking
+            // blocks which have empty text but contain redactedData that must be preserved)
+            return part.text !== "" || part.providerOptions != null
           }
           return true
         })
@@ -213,6 +216,7 @@ function normalizeMessages(
       const parts = msg.content
       const first = parts.findIndex((part) => part.type === "tool-call")
       if (first === -1) return [msg]
+      if (parts.some((part) => part.type === "reasoning")) return [msg]
       if (!parts.slice(first).some((part) => part.type !== "tool-call")) return [msg]
       return [
         { ...msg, content: parts.filter((part) => part.type !== "tool-call") },
@@ -358,13 +362,17 @@ function applyCaching(msgs: ModelMessage[], model: Provider.Model): ModelMessage
     const shouldUseContentOptions = !useMessageLevelOptions && Array.isArray(msg.content) && msg.content.length > 0
 
     if (shouldUseContentOptions) {
-      const lastContent = msg.content[msg.content.length - 1]
-      if (
-        lastContent &&
-        typeof lastContent === "object" &&
-        lastContent.type !== "tool-approval-request" &&
-        lastContent.type !== "tool-approval-response"
-      ) {
+      const lastContent = [...msg.content]
+        .reverse()
+        .find(
+          (part) =>
+            part &&
+            typeof part === "object" &&
+            part.type !== "tool-approval-request" &&
+            part.type !== "tool-approval-response" &&
+            part.type !== "reasoning",
+        ) as any
+      if (lastContent) {
         lastContent.providerOptions = mergeDeep(lastContent.providerOptions ?? {}, providerOptions)
         continue
       }
diff --git a/packages/opencode/src/session/compaction.ts b/packages/opencode/src/session/compaction.ts
index 067d43da2e25..58490ec8740f 100644
--- a/packages/opencode/src/session/compaction.ts
+++ b/packages/opencode/src/session/compaction.ts
@@ -407,6 +407,7 @@ export const layer: Layer.Layer<
       yield* plugin.trigger("experimental.chat.messages.transform", {}, { messages: msgs })
       const modelMessages = yield* MessageV2.toModelMessagesEffect(msgs, model, {
         stripMedia: true,
+        stripReasoning: true,
         toolOutputMaxChars: TOOL_OUTPUT_MAX_CHARS,
       })
       const ctx = yield* InstanceState.context
@@ -449,7 +450,7 @@ export const layer: Layer.Layer<
         tools: {},
         system: [],
         messages: [
-          ...modelMessages,
+          ...(modelMessages ?? []),
           {
             role: "user",
             content: [{ type: "text", text: nextPrompt }],
diff --git a/packages/opencode/src/session/message-v2.ts b/packages/opencode/src/session/message-v2.ts
index ed09262d0efe..8a36a414ea65 100644
--- a/packages/opencode/src/session/message-v2.ts
+++ b/packages/opencode/src/session/message-v2.ts
@@ -729,7 +729,7 @@ function providerMeta(metadata: Record<string, any> | undefined) {
 export const toModelMessagesEffect = Effect.fnUntraced(function* (
   input: WithParts[],
   model: Provider.Model,
-  options?: { stripMedia?: boolean; toolOutputMaxChars?: number },
+  options?: { stripMedia?: boolean; stripReasoning?: boolean; toolOutputMaxChars?: number },
 ) {
   const result: UIMessage[] = []
   const toolNames = new Set<string>()
@@ -956,6 +956,7 @@ export const toModelMessagesEffect = Effect.fnUntraced(function* (
             })
         }
         if (part.type === "reasoning") {
+          if (options?.stripReasoning) continue
           if (differentModel) {
             if (part.text.trim().length > 0)
               assistantMessage.parts.push({
@@ -1012,9 +1013,9 @@ export const toModelMessagesEffect = Effect.fnUntraced(function* (
 export function toModelMessages(
   input: WithParts[],
   model: Provider.Model,
-  options?: { stripMedia?: boolean; toolOutputMaxChars?: number },
+  options?: { stripMedia?: boolean; stripReasoning?: boolean; toolOutputMaxChars?: number },
 ): Promise<ModelMessage[]> {
-  return Effect.runPromise(toModelMessagesEffect(input, model, options).pipe(Effect.provide(EffectLogger.layer)))
+  return Effect.runPromise(toModelMessagesEffect(input, model, options).pipe(Effect.provide(EffectLogger.layer), Effect.map((msgs) => msgs ?? [])))
 }
 
 export function page(input: { sessionID: SessionID; limit: number; before?: string }) {
diff --git a/packages/opencode/src/session/prompt.ts b/packages/opencode/src/session/prompt.ts
index fef8c438366c..78d3b374162f 100644
--- a/packages/opencode/src/session/prompt.ts
+++ b/packages/opencode/src/session/prompt.ts
@@ -208,7 +208,7 @@ export const layer = Layer.effect(
           model: mdl,
           sessionID: input.session.id,
           retries: 2,
-          messages: [{ role: "user", content: "Generate a title for this conversation:\n" }, ...msgs],
+          messages: [{ role: "user", content: "Generate a title for this conversation:\n" }, ...(msgs ?? [])],
         })
         .pipe(
           Stream.filter((e): e is Extract<LLM.Event, { type: "text-delta" }> => e.type === "text-delta"),
@@ -1581,7 +1581,7 @@ NOTE: At any point in time through this workflow you should feel free to ask the
               sessionID,
               parentSessionID: session.parentID,
               system,
-              messages: [...modelMsgs, ...(isLastStep ? [{ role: "assistant" as const, content: MAX_STEPS }] : [])],
+               messages: [...(modelMsgs ?? []), ...(isLastStep ? [{ role: "assistant" as const, content: MAX_STEPS }] : [])],
               tools,
               model,
               toolChoice: format.type === "json_schema" ? "required" : undefined,
diff --git a/packages/opencode/test/session/message-v2.test.ts b/packages/opencode/test/session/message-v2.test.ts
index 999b61b48e56..3fd0fc4e3b36 100644
--- a/packages/opencode/test/session/message-v2.test.ts
+++ b/packages/opencode/test/session/message-v2.test.ts
@@ -526,6 +526,49 @@ describe("session.message-v2.toModelMessage", () => {
     ])
   })
 
+  test("converts reasoning to text when assistant model differs", async () => {
+    const userID = "m-user"
+    const assistantID = "m-assistant"
+
+    const input: MessageV2.WithParts[] = [
+      {
+        info: userInfo(userID),
+        parts: [
+          {
+            ...basePart(userID, "u1"),
+            type: "text",
+            text: "think about this",
+          },
+        ] as MessageV2.Part[],
+      },
+      {
+        info: assistantInfo(assistantID, userID, undefined, { providerID: "other", modelID: "other" }),
+        parts: [
+          {
+            ...basePart(assistantID, "a1"),
+            type: "reasoning",
+            text: "reasoning trace",
+            time: { start: 0 },
+            metadata: { anthropic: { signature: "sig-abc" } },
+          },
+        ] as MessageV2.Part[],
+      } as MessageV2.WithParts,
+    ]
+
+    expect(await MessageV2.toModelMessages(input, model)).toStrictEqual([
+      {
+        role: "user",
+        content: [{ type: "text", text: "think about this" }],
+      },
+      {
+        role: "assistant",
+        content: [
+          { type: "text", text: "reasoning trace" },
+        ],
+      },
+    ])
+  })
+
   test("replaces compacted tool output with placeholder", async () => {
     const userID = "m-user"
     const assistantID = "m-assistant"