diff --git a/packages/opencode/src/session/message-v2.ts b/packages/opencode/src/session/message-v2.ts
index 1bea9f52c3ec..c36f1a6e99aa 100644
--- a/packages/opencode/src/session/message-v2.ts
+++ b/packages/opencode/src/session/message-v2.ts
@@ -259,27 +259,27 @@ export const toModelMessagesEffect = Effect.fnUntraced(function* (
         role: "assistant",
         parts: [],
       }
-      // Anthropic adaptive thinking can persist assistant turns like:
-      // step-start, reasoning(signature), text(""), step-start,
-      // reasoning(signature). The empty text part is a structural separator,
-      // but it does not carry the signature metadata itself. Dropping it shifts
-      // signed thinking positions after step-start splitting/provider regrouping;
-      // keeping it as "" is filtered by the AI SDK and rejected by Anthropic.
-      // It is unclear whether this shape originates in our stream processing,
-      // a proxy, or a lower-level library, but preserving a non-empty separator
-      // here is the only safe replay point we have.
-      // Use a single space so the separator survives replay without changing
-      // the neighboring signed reasoning blocks.
+      // Anthropic extended thinking signs reasoning blocks and rejects any
+      // modification to the assistant message that contains them — including
+      // mutations to adjacent text parts that shift content-block indices.
+      // An empty text part ("") can appear as a structural separator between
+      // reasoning groups (e.g. step-start, reasoning(sig), text(""),
+      // step-start, reasoning(sig)).  Previously this was rewritten to a
+      // single space (" ") to survive AI-SDK filtering, but that mutation
+      // itself causes Anthropic to reject the signed reasoning blocks as
+      // "modified".  Dropping the empty separator is safe because step-start
+      // parts already delimit reasoning groups, and the absence of the text
+      // part does not change the signed blocks themselves.
       const hasSignedReasoning = msg.parts.some((part) => {
         if (part.type !== "reasoning") return false
         return part.metadata?.anthropic?.signature != null
       })
       for (const part of msg.parts) {
         if (part.type === "text") {
-          const text = part.text === "" && hasSignedReasoning ? " " : part.text
+          if (part.text === "" && hasSignedReasoning) continue
           assistantMessage.parts.push({
             type: "text",
-            text,
+            text: part.text,
             ...(differentModel ? {} : { providerMetadata: part.metadata }),
           })
         }
@@ -361,6 +361,12 @@ export const toModelMessagesEffect = Effect.fnUntraced(function* (
         }
         if (part.type === "reasoning") {
           if (differentModel) {
+            // When replaying to a different model, signed reasoning blocks
+            // cannot be preserved verbatim (the signature is Anthropic-specific).
+            // Downgrade readable thinking to plain text and silently drop
+            // redacted thinking (which has no readable content).  This is a
+            // one-way transformation for cross-provider replay only; same-model
+            // replay preserves the signed block below.
             if (part.text.trim().length > 0)
               assistantMessage.parts.push({
                 type: "text",
diff --git a/packages/opencode/test/session/message-v2.test.ts b/packages/opencode/test/session/message-v2.test.ts
index 9bb688aeddb0..23af0971931d 100644
--- a/packages/opencode/test/session/message-v2.test.ts
+++ b/packages/opencode/test/session/message-v2.test.ts
@@ -1259,8 +1259,14 @@ describe("session.message-v2.toModelMessage", () => {
     ])
   })
 
-  test("substitutes space for empty text between signed reasoning blocks", async () => {
-    // Reproduces the bug pattern: [reasoning(sig), text(""), reasoning(sig), text(full)]
+  test("drops empty text between signed reasoning blocks to preserve signature validity", async () => {
+    // Anthropic signs reasoning blocks and rejects any modification to the
+    // assistant message that contains them.  An empty text part ("") can
+    // appear as a structural separator between reasoning groups.  Previously
+    // it was rewritten to " " to survive AI-SDK filtering, but that mutation
+    // itself causes Anthropic to reject the signed blocks as "modified".
+    // Dropping the empty separator is safe because step-start parts already
+    // delimit reasoning groups.
     const assistantID = "m-assistant"
     const input: SessionV1.WithParts[] = [
       {
@@ -1288,12 +1294,64 @@ describe("session.message-v2.toModelMessage", () => {
 
     const result = await MessageV2.toModelMessages(input, model)
 
-    // step-start splits into two assistant messages; SDK's groupIntoBlocks merges them later
     expect(result).toHaveLength(2)
-    expect((result[0].content as any[]).find((p) => p.type === "text").text).toBe(" ")
+    // First reasoning group: no text part at all (empty separator dropped)
+    expect((result[0].content as any[]).find((p) => p.type === "text")).toBeUndefined()
+    const reasoning0 = (result[0].content as any[]).find((p) => p.type === "reasoning")
+    expect(reasoning0).toBeDefined()
+    expect(reasoning0.providerOptions?.anthropic?.signature).toBe("sig1")
+    // Second reasoning group: non-empty text preserved
+    const reasoning1 = (result[1].content as any[]).find((p) => p.type === "reasoning")
+    expect(reasoning1).toBeDefined()
+    expect(reasoning1.providerOptions?.anthropic?.signature).toBe("sig2")
     expect((result[1].content as any[]).find((p) => p.type === "text").text).toBe("the answer")
   })
 
+  test("preserves redacted thinking blocks with empty text during replay", async () => {
+    // Redacted thinking (text: "" with anthropic.redactedData) must survive
+    // toModelMessages without mutation. The empty text separator between
+    // redacted reasoning and a normal text part must be dropped (not mutated
+    // to " ") so the signed redactedData block is not invalidated.
+    const assistantID = "m-assistant-redacted"
+    const input: SessionV1.WithParts[] = [
+      {
+        info: assistantInfo(assistantID, "m-parent"),
+        parts: [
+          { ...basePart(assistantID, "p1"), type: "step-start" },
+          {
+            ...basePart(assistantID, "p2"),
+            type: "reasoning",
+            text: "",
+            metadata: { anthropic: { redactedData: "encrypted-blob" } },
+          },
+          { ...basePart(assistantID, "p3"), type: "text", text: "" },
+          { ...basePart(assistantID, "p4"), type: "step-start" },
+          {
+            ...basePart(assistantID, "p5"),
+            type: "reasoning",
+            text: "thinking-two",
+            metadata: { anthropic: { signature: "sig2" } },
+          },
+          { ...basePart(assistantID, "p6"), type: "text", text: "response" },
+        ] as SessionV1.Part[],
+      },
+    ]
+
+    const result = await MessageV2.toModelMessages(input, model)
+
+    expect(result).toHaveLength(2)
+    // First group: redacted reasoning present, no text separator
+    const reasoning0 = (result[0].content as any[]).find((p) => p.type === "reasoning")
+    expect(reasoning0).toBeDefined()
+    expect(reasoning0.providerOptions?.anthropic?.redactedData).toBe("encrypted-blob")
+    expect((result[0].content as any[]).find((p) => p.type === "text")).toBeUndefined()
+    // Second group: signed reasoning + normal text
+    const reasoning1 = (result[1].content as any[]).find((p) => p.type === "reasoning")
+    expect(reasoning1).toBeDefined()
+    expect(reasoning1.providerOptions?.anthropic?.signature).toBe("sig2")
+    expect((result[1].content as any[]).find((p) => p.type === "text").text).toBe("response")
+  })
+
   test("leaves empty text alone when reasoning signature is under 'bedrock' namespace", async () => {
     // Bedrock signed reasoning is preserved as reasoning metadata, but unlike the
     // direct Anthropic path we do not preserve empty text separators for Bedrock.