diff --git a/CHANGELOG.md b/CHANGELOG.md
index ce7eb5574..dfd13a1ae 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -29,6 +29,7 @@
 - Models/Providers: treat credential validation failures as auth errors to trigger fallback; normalize `${ENV_VAR}` apiKey values and auto-fill missing provider keys; preserve explicit GitHub Copilot provider config + agent-dir auth profiles.
 - Auth: drop invalid auth profiles from ordering so environment keys can still be used for providers like MiniMax.
 - Gemini: normalize Gemini 3 ids to preview variants; strip Gemini CLI tool call/response ids; downgrade missing `thought_signature`; strip Claude `msg_*` thought_signature fields to avoid base64 decode errors.
+- Agents: auto-recover from compaction context overflow by resetting the session and retrying; propagate overflow details from embedded runs so callers can recover.
 - MiniMax: strip malformed tool invocation XML; include `MiniMax-VL-01` in implicit provider for image pairing.
 - Onboarding/Auth: honor `CLAWDBOT_AGENT_DIR` / `PI_CODING_AGENT_DIR` when writing auth profiles (MiniMax). (#829) — thanks @roshanasingh4.
 - Anthropic: merge consecutive user turns (preserve newest metadata) before validation to avoid incorrect role errors.
diff --git a/src/agents/pi-embedded-helpers.ts b/src/agents/pi-embedded-helpers.ts
index 02a05d4a5..4a1c4fcc5 100644
--- a/src/agents/pi-embedded-helpers.ts
+++ b/src/agents/pi-embedded-helpers.ts
@@ -370,8 +370,8 @@ export function formatAssistantErrorText(
   // Check for context overflow (413) errors
   if (isContextOverflowError(raw)) {
     return (
-      "Context overflow: the conversation history is too large. " +
-      "Use /new or /reset to start a fresh session."
+      "Context overflow: prompt too large for the model. " +
+      "Try again with less input or a larger-context model."
     );
   }
 
diff --git a/src/agents/pi-embedded-runner.ts b/src/agents/pi-embedded-runner.ts
index 40eed6ea3..c01579392 100644
--- a/src/agents/pi-embedded-runner.ts
+++ b/src/agents/pi-embedded-runner.ts
@@ -362,6 +362,10 @@ export type EmbeddedPiRunMeta = {
   durationMs: number;
   agentMeta?: EmbeddedPiAgentMeta;
   aborted?: boolean;
+  error?: {
+    kind: "context_overflow" | "compaction_failure";
+    message: string;
+  };
 };
 
 function buildModelAliasLines(cfg?: ClawdbotConfig) {
@@ -1976,12 +1980,15 @@ export async function runEmbeddedPiAgent(params: {
           if (promptError && !aborted) {
             const errorText = describeUnknownError(promptError);
             if (isContextOverflowError(errorText)) {
+              const kind = isCompactionFailureError(errorText)
+                ? "compaction_failure"
+                : "context_overflow";
               return {
                 payloads: [
                   {
                     text:
-                      "Context overflow: the conversation history is too large for the model. " +
-                      "Use /new or /reset to start a fresh session, or try a model with a larger context window.",
+                      "Context overflow: prompt too large for the model. " +
+                      "Try again with less input or a larger-context model.",
                     isError: true,
                   },
                 ],
@@ -1992,6 +1999,7 @@ export async function runEmbeddedPiAgent(params: {
                     provider,
                     model: model.id,
                   },
+                  error: { kind, message: errorText },
                 },
               };
             }
diff --git a/src/auto-reply/reply/agent-runner.heartbeat-typing.test.ts b/src/auto-reply/reply/agent-runner.heartbeat-typing.test.ts
index 89544d618..5b3d9b562 100644
--- a/src/auto-reply/reply/agent-runner.heartbeat-typing.test.ts
+++ b/src/auto-reply/reply/agent-runner.heartbeat-typing.test.ts
@@ -525,6 +525,65 @@ describe("runReplyAgent typing (heartbeat)", () => {
     }
   });
 
+  it("retries after context overflow payload by resetting the session", async () => {
+    const prevStateDir = process.env.CLAWDBOT_STATE_DIR;
+    const stateDir = await fs.mkdtemp(
+      path.join(tmpdir(), "clawdbot-session-overflow-reset-"),
+    );
+    process.env.CLAWDBOT_STATE_DIR = stateDir;
+    try {
+      const sessionId = "session";
+      const storePath = path.join(stateDir, "sessions", "sessions.json");
+      const sessionEntry = { sessionId, updatedAt: Date.now() };
+      const sessionStore = { main: sessionEntry };
+
+      await fs.mkdir(path.dirname(storePath), { recursive: true });
+      await fs.writeFile(storePath, JSON.stringify(sessionStore), "utf-8");
+
+      runEmbeddedPiAgentMock
+        .mockImplementationOnce(async () => ({
+          payloads: [
+            { text: "Context overflow: prompt too large", isError: true },
+          ],
+          meta: {
+            durationMs: 1,
+            error: {
+              kind: "context_overflow",
+              message:
+                'Context overflow: Summarization failed: 400 {"message":"prompt is too long"}',
+            },
+          },
+        }))
+        .mockImplementationOnce(async () => ({
+          payloads: [{ text: "ok" }],
+          meta: { durationMs: 1 },
+        }));
+
+      const callsBefore = runEmbeddedPiAgentMock.mock.calls.length;
+      const { run } = createMinimalRun({
+        sessionEntry,
+        sessionStore,
+        sessionKey: "main",
+        storePath,
+      });
+      const res = await run();
+
+      expect(runEmbeddedPiAgentMock.mock.calls.length - callsBefore).toBe(2);
+      const payload = Array.isArray(res) ? res[0] : res;
+      expect(payload).toMatchObject({ text: "ok" });
+      expect(sessionStore.main.sessionId).not.toBe(sessionId);
+
+      const persisted = JSON.parse(await fs.readFile(storePath, "utf-8"));
+      expect(persisted.main.sessionId).toBe(sessionStore.main.sessionId);
+    } finally {
+      if (prevStateDir) {
+        process.env.CLAWDBOT_STATE_DIR = prevStateDir;
+      } else {
+        delete process.env.CLAWDBOT_STATE_DIR;
+      }
+    }
+  });
+
   it("still replies even if session reset fails to persist", async () => {
     const prevStateDir = process.env.CLAWDBOT_STATE_DIR;
     const stateDir = await fs.mkdtemp(
diff --git a/src/auto-reply/reply/agent-runner.ts b/src/auto-reply/reply/agent-runner.ts
index a25c55d92..3478aeb62 100644
--- a/src/auto-reply/reply/agent-runner.ts
+++ b/src/auto-reply/reply/agent-runner.ts
@@ -834,6 +834,20 @@ export async function runReplyAgent(params: {
         runResult = fallbackResult.result;
         fallbackProvider = fallbackResult.provider;
         fallbackModel = fallbackResult.model;
+
+        // Some embedded runs surface context overflow as an error payload instead of throwing.
+        // Treat those as a session-level failure and auto-recover by starting a fresh session.
+        const embeddedError = runResult.meta?.error;
+        if (
+          embeddedError &&
+          isContextOverflowError(embeddedError.message) &&
+          !didResetAfterCompactionFailure &&
+          (await resetSessionAfterCompactionFailure(embeddedError.message))
+        ) {
+          didResetAfterCompactionFailure = true;
+          continue;
+        }
+
         break;
       } catch (err) {
         const message = err instanceof Error ? err.message : String(err);
@@ -894,7 +908,7 @@ export async function runReplyAgent(params: {
         defaultRuntime.error(`Embedded agent failed before reply: ${message}`);
         return finalizeWithFollowup({
           text: isContextOverflow
-            ? "⚠️ Context overflow - conversation too long. Starting fresh might help!"
+            ? "⚠️ Context overflow — prompt too large for this model. Try a shorter message or a larger-context model."
             : `⚠️ Agent failed before reply: ${message}. Check gateway logs for details.`,
         });
       }