diff --git a/packages/opencode/src/session/compaction.ts b/packages/opencode/src/session/compaction.ts
index fa439e4efffa..4aa1ccdec65d 100644
--- a/packages/opencode/src/session/compaction.ts
+++ b/packages/opencode/src/session/compaction.ts
@@ -412,10 +412,12 @@ const layer = Layer.effect(
         return "stop"
       }
 
-      if (compactionPart && selected.tail_start_id && compactionPart.tail_start_id !== selected.tail_start_id) {
+      // A replay re-asks the overflowing message, so nothing before this marker may survive next to it.
+      const tailStartID = replay ? input.parentID : selected.tail_start_id
+      if (compactionPart && tailStartID && compactionPart.tail_start_id !== tailStartID) {
         yield* session.updatePart({
           ...compactionPart,
-          tail_start_id: selected.tail_start_id,
+          tail_start_id: tailStartID,
         })
       }
 
diff --git a/packages/opencode/src/session/prompt.ts b/packages/opencode/src/session/prompt.ts
index 472c1aecf0ab..0fc8e54dbbf2 100644
--- a/packages/opencode/src/session/prompt.ts
+++ b/packages/opencode/src/session/prompt.ts
@@ -61,6 +61,8 @@ import { LLMEvent } from "@opencode-ai/llm"
 // @ts-ignore
 globalThis.AI_SDK_LOG_WARNINGS = false
 
+const MAX_OVERFLOW_COMPACTIONS_PER_TURN = 1
+
 const decodeMessageInfo = Schema.decodeUnknownExit(SessionV1.Info)
 const decodeMessagePart = Schema.decodeUnknownExit(SessionV1.Part)
 const MAX_MCP_RESOURCE_BLOB_BYTES = 10 * 1024 * 1024
@@ -1087,6 +1089,7 @@ const layer = Layer.effect(
         const ctx = yield* InstanceState.context
         let structured: unknown
         let step = 0
+        let overflowCompactions = 0
         const turnStart = Date.now()
         let firstTokenAt: number | undefined
         let firstTokenReqStart: number | undefined
@@ -1376,15 +1379,28 @@ const layer = Layer.effect(
             }
 
             if (result === "stop") return "break" as const
-            if (result === "compact") {
-              yield* compaction.create({
-                sessionID,
-                agent: lastUser.agent,
-                model: lastUser.model,
-                auto: true,
-                overflow: !handle.message.finish,
-              })
+            if (result === "continue") {
+              overflowCompactions = 0
+              return "continue" as const
+            }
+            const overflow = !handle.message.finish
+            // Compaction already replayed this message once; overflowing again means it can never fit.
+            if (overflow && overflowCompactions >= MAX_OVERFLOW_COMPACTIONS_PER_TURN) {
+              handle.message.error = new SessionV1.ContextOverflowError({
+                message: "Message exceeds the model context limit even after compaction",
+              }).toObject()
+              handle.message.finish = "error"
+              yield* sessions.updateMessage(handle.message)
+              return "break" as const
             }
+            if (overflow) overflowCompactions++
+            yield* compaction.create({
+              sessionID,
+              agent: lastUser.agent,
+              model: lastUser.model,
+              auto: true,
+              overflow,
+            })
             return "continue" as const
           }).pipe(
             Effect.ensuring(instruction.clear(handle.message.id)),
diff --git a/packages/opencode/test/session/compaction.test.ts b/packages/opencode/test/session/compaction.test.ts
index 008f230a2272..b1ce8e55d9ca 100644
--- a/packages/opencode/test/session/compaction.test.ts
+++ b/packages/opencode/test/session/compaction.test.ts
@@ -1163,6 +1163,45 @@ describe("session.compaction.process", () => {
     }),
   )
 
+  it.instance(
+    "overflow replay drops the overflowing message from the retained window",
+    Effect.gen(function* () {
+      const ssn = yield* SessionNs.Service
+      const session = yield* ssn.create({})
+      const big = "x".repeat(600_000)
+      const first = yield* createUserMessage(session.id, big)
+      yield* createAssistantMessage(session.id, first.id, "/")
+      const resent = yield* createUserMessage(session.id, big)
+      yield* SessionCompaction.use.create({
+        sessionID: session.id,
+        agent: "build",
+        model: ref,
+        auto: true,
+        overflow: true,
+      })
+      const msgs = yield* MessageV2.filterCompactedEffect(session.id)
+      const marker = msgs.at(-1)!
+
+      const result = yield* SessionCompaction.use.process({
+        parentID: marker.info.id,
+        messages: msgs,
+        sessionID: session.id,
+        auto: true,
+        overflow: true,
+      })
+      expect(result).toBe("continue")
+      // The fake processor never stamps finish; filterCompacted only honours a finished summary.
+      const all = yield* ssn.messages({ sessionID: session.id })
+      const summary = all.find((m) => m.info.role === "assistant" && m.info.parentID === marker.info.id)!
+      yield* ssn.updateMessage({ ...(summary.info as SessionV1.Assistant), finish: "stop" })
+
+      const context = yield* MessageV2.filterCompactedEffect(session.id)
+      expect(context.map((m) => m.info.role)).toEqual(["user", "assistant", "user"])
+      expect(context.some((m) => m.info.id === resent.id)).toBe(false)
+      expect(context.at(-1)?.parts.some((part) => part.type === "text" && part.text === big)).toBe(true)
+    }),
+  )
+
   it.instance(
     "falls back to overflow guidance when no replayable turn exists",
     Effect.gen(function* () {
diff --git a/packages/opencode/test/session/prompt.test.ts b/packages/opencode/test/session/prompt.test.ts
index 9272720dcb07..0650f9a19ca2 100644
--- a/packages/opencode/test/session/prompt.test.ts
+++ b/packages/opencode/test/session/prompt.test.ts
@@ -668,6 +668,36 @@ it.instance("loop stops provider overflow instead of auto-compacting when disabl
   }),
 )
 
+it.instance("loop stops after one overflow compaction when the replay still overflows", () =>
+  Effect.gen(function* () {
+    const { llm } = yield* useServerConfig(providerCfg)
+    const prompt = yield* SessionPrompt.Service
+    const sessions = yield* Session.Service
+    const chat = yield* sessions.create({ title: "Pinned" })
+
+    yield* llm.error(413, { error: { message: "prompt is too long" } })
+    yield* llm.text("## Summary")
+    yield* llm.error(413, { error: { message: "prompt is too long" } })
+    yield* prompt.prompt({
+      sessionID: chat.id,
+      agent: "build",
+      noReply: true,
+      parts: [{ type: "text", text: "hello" }],
+    })
+
+    const result = yield* prompt.loop({ sessionID: chat.id })
+    const messages = yield* sessions.messages({ sessionID: chat.id })
+
+    expect(yield* llm.hits).toHaveLength(3)
+    expect(result.info.role).toBe("assistant")
+    if (result.info.role === "assistant") {
+      expect(result.info.error?.name).toBe("ContextOverflowError")
+      expect(result.info.finish).toBe("error")
+    }
+    expect(messages.filter((message) => message.parts.some((part) => part.type === "compaction"))).toHaveLength(1)
+  }),
+)
+
 noLLMServer.instance.skip(
   "prompt emits v2 prompted and synthetic events (v2 projector disabled)",
   () =>
@@ -836,10 +866,12 @@ it.instance("turn.done logs prep_ms from the first request even when ttft captur
     const capture = Logger.make((options) => {
       captured.push(options.message)
     })
-    yield* prompt.loop({ sessionID: session.id }).pipe(
-      Effect.provide(Logger.layer([capture], { mergeWithExisting: true })),
-      Effect.provide(Layer.succeed(References.MinimumLogLevel, "Debug")),
-    )
+    yield* prompt
+      .loop({ sessionID: session.id })
+      .pipe(
+        Effect.provide(Logger.layer([capture], { mergeWithExisting: true })),
+        Effect.provide(Layer.succeed(References.MinimumLogLevel, "Debug")),
+      )
 
     const message = captured.find((item) => Array.isArray(item) && item[0] === "turn.done")
     expect(message).toBeDefined()