revert prompt-too-long fallback and keep inline directives

src/auto-reply/command-reply.test.ts

@@ -251,35 +251,6 @@ describe("runCommandReply (pi)", () => {
     expect(meta.killed).toBe(true);
   });
 
-  it("surfaces prompt-too-long and flags meta for session reset", async () => {
-    mockPiRpc({
-      stdout:
-        '{"type":"agent_end","message":{"role":"assistant","content":[],"errorMessage":"400 {\\"type\\":\\"error\\",\\"error\\":{\\"type\\":\\"invalid_request_error\\",\\"message\\":\\"prompt is too long: 200333 tokens > 200000 maximum\\"}}"}}',
-      stderr: "",
-      code: 0,
-    });
-
-    const { payloads, meta } = await runCommandReply({
-      reply: {
-        mode: "command",
-        command: ["pi", "{{Body}}"],
-        agent: { kind: "pi" },
-      },
-      templatingCtx: noopTemplateCtx,
-      sendSystemOnce: false,
-      isNewSession: false,
-      isFirstTurnInSession: false,
-      systemSent: false,
-      timeoutMs: 1000,
-      timeoutSeconds: 1,
-      commandRunner: vi.fn(),
-      enqueue: enqueueImmediate,
-    });
-
-    expect(payloads?.[0]?.text).toMatch(/history is too long/i);
-    expect(meta.agentMeta?.extra?.promptTooLong).toBe(true);
-  });
-
   it("collapses rpc deltas instead of emitting raw JSON spam", async () => {
     mockPiRpc({
       stdout: [

src/auto-reply/command-reply.ts

@@ -684,7 +684,6 @@ export async function runCommandReply(
   if (stderr?.trim()) {
     logVerbose(`Command auto-reply stderr: ${stderr.trim()}`);
   }
-  const promptTooLong = rawStdout.includes("prompt is too long");
 
   const logFailure = () => {
     const truncate = (s?: string) =>
@@ -703,21 +702,6 @@ export async function runCommandReply(
       );
     };
 
-    if (promptTooLong) {
-      const text =
-        "⚠️ Session history is too long. Starting a fresh session — please resend your last message.";
-      const meta: CommandReplyMeta = {
-        durationMs: Date.now() - started,
-        queuedMs,
-        queuedAhead,
-        exitCode: code,
-        signal,
-        killed,
-        agentMeta: { extra: { promptTooLong: true } },
-      };
-      return { payloads: [{ text }], meta };
-    }
-
     const parsed = trimmed ? agent.parseOutput(trimmed) : undefined;
 
     // Collect assistant texts and tool results from parseOutput (tau RPC can emit many).

src/auto-reply/reply.ts

@@ -701,26 +701,6 @@ export async function getReplyFromConfig(
       });
       const payloadArray = runResult.payloads ?? [];
       const meta = runResult.meta;
-      const promptTooLong =
-        meta.agentMeta &&
-        typeof meta.agentMeta.extra === "object" &&
-        (meta.agentMeta.extra as { promptTooLong?: boolean }).promptTooLong;
-
-      if (promptTooLong && sessionCfg && sessionStore && sessionKey) {
-        // Rotate to a new session to avoid hitting context limits again.
-        const newId = crypto.randomUUID();
-        sessionEntry = {
-          sessionId: newId,
-          updatedAt: Date.now(),
-          systemSent,
-          abortedLastRun: false,
-        };
-        sessionStore[sessionKey] = sessionEntry;
-        await saveSessionStore(storePath, sessionStore);
-        sessionId = newId;
-        isNewSession = true;
-      }
-
       let finalPayloads = payloadArray;
       if (!finalPayloads || finalPayloads.length === 0) {
         return undefined;