feat(FN-705): add rate-limit retry with exponential backoff for AI agents

- Add rate-limit retry utility with exponential backoff, jitter, and configurable limits
- Integrate retry logic into executor, triage, and merger agent sessions
- Export retryOnRateLimit from engine package index
- Remove unused DirectoryPicker component, stale CSS, and dead dashboard routes
- Update SetupWizardModal and usage tests to reflect cleanup

packages/engine/src/executor.test.ts

@@ -51,6 +51,9 @@ vi.mock("node:child_process", () => ({
 vi.mock("node:fs", () => ({
   existsSync: vi.fn().mockReturnValue(true),
 }));
+vi.mock("./rate-limit-retry.js", () => ({
+  withRateLimitRetry: (fn: () => Promise<any>) => fn(),
+}));
 
 import { TaskExecutor, buildExecutionPrompt } from "./executor.js";
 import { createKbAgent } from "./pi.js";

packages/engine/src/executor.ts

@@ -14,6 +14,7 @@ import { AgentLogger } from "./agent-logger.js";
 import { executorLog, reviewerLog } from "./logger.js";
 import { isUsageLimitError, checkSessionError, type UsageLimitPauser } from "./usage-limit-detector.js";
 import { isTransientError } from "./transient-error-detector.js";
+import { withRateLimitRetry } from "./rate-limit-retry.js";
 import type { StuckTaskDetector } from "./stuck-task-detector.js";
 
 // Re-export for backward compatibility (tests import from executor.ts)
@@ -653,10 +654,18 @@ export class TaskExecutor {
         }
       };
 
+      const retryableWork = () => withRateLimitRetry(agentWork, {
+        onRetry: (attempt, delayMs, error) => {
+          const delaySec = Math.round(delayMs / 1000);
+          executorLog.warn(`⏳ ${task.id} rate limited — retry ${attempt} in ${delaySec}s: ${error.message}`);
+          this.store.logEntry(task.id, `Rate limited — retry ${attempt} in ${delaySec}s`).catch(() => {});
+        },
+      });
+
       if (this.options.semaphore) {
-        await this.options.semaphore.run(agentWork, PRIORITY_EXECUTE);
+        await this.options.semaphore.run(retryableWork, PRIORITY_EXECUTE);
       } else {
-        await agentWork();
+        await retryableWork();
       }
     } catch (err: any) {
       if (this.depAborted.has(task.id)) {

packages/engine/src/index.ts

@@ -9,6 +9,7 @@ export { createKbAgent, type AgentOptions, type AgentResult } from "./pi.js";
 export { WorktreePool, scanIdleWorktrees, cleanupOrphanedWorktrees } from "./worktree-pool.js";
 export { createLogger, type Logger } from "./logger.js";
 export { isUsageLimitError, UsageLimitPauser } from "./usage-limit-detector.js";
+export { withRateLimitRetry } from "./rate-limit-retry.js";
 export { PrMonitor, type PrComment, type TrackedPr, type OnNewCommentsCallback } from "./pr-monitor.js";
 export { PrCommentHandler } from "./pr-comment-handler.js";
 export { NtfyNotifier, type NtfyNotifierOptions } from "./notifier.js";

packages/engine/src/merger.test.ts

@@ -14,6 +14,10 @@ vi.mock("node:fs", () => ({
   readFileSync: vi.fn(),
 }));
 
+vi.mock("./rate-limit-retry.js", () => ({
+  withRateLimitRetry: (fn: () => Promise<any>) => fn(),
+}));
+
 import {
   aiMergeTask,
   findWorktreeUser,

packages/engine/src/merger.ts

@@ -6,6 +6,7 @@ import type { WorktreePool } from "./worktree-pool.js";
 import { AgentLogger } from "./agent-logger.js";
 import { mergerLog } from "./logger.js";
 import { isUsageLimitError, checkSessionError, type UsageLimitPauser } from "./usage-limit-detector.js";
+import { withRateLimitRetry } from "./rate-limit-retry.js";
 import type { ToolDefinition } from "@mariozechner/pi-coding-agent";
 import { Type } from "@sinclair/typebox";
 
@@ -1153,9 +1154,15 @@ async function runAiAgentForCommit(params: AiAgentParams): Promise<{ success: bo
       simplifiedContext,
       buildCommand,
     });
-    await session.prompt(prompt);
-
-    checkSessionError(session);
+    await withRateLimitRetry(async () => {
+      await session.prompt(prompt);
+      checkSessionError(session);
+    }, {
+      onRetry: (attempt, delayMs, error) => {
+        const delaySec = Math.round(delayMs / 1000);
+        mergerLog.warn(`⏳ ${taskId} rate limited — retry ${attempt} in ${delaySec}s: ${error.message}`);
+      },
+    });
 
     // Check if build failed
     if (buildFailed) {

packages/engine/src/rate-limit-retry.test.ts

@@ -0,0 +1,217 @@
+import { describe, it, expect, vi, beforeEach, afterEach } from "vitest";
+import { withRateLimitRetry } from "./rate-limit-retry.js";
+
+describe("withRateLimitRetry", () => {
+  beforeEach(() => {
+    vi.useFakeTimers();
+  });
+  afterEach(() => {
+    vi.useRealTimers();
+  });
+
+  it("returns the result when fn succeeds on first call", async () => {
+    const fn = vi.fn().mockResolvedValue("ok");
+    const promise = withRateLimitRetry(fn);
+    const result = await promise;
+    expect(result).toBe("ok");
+    expect(fn).toHaveBeenCalledTimes(1);
+  });
+
+  it("retries on rate limit error and succeeds", async () => {
+    const fn = vi
+      .fn()
+      .mockRejectedValueOnce(new Error("429 too many requests"))
+      .mockResolvedValueOnce("recovered");
+
+    const onRetry = vi.fn();
+    const promise = withRateLimitRetry(fn, {
+      baseDelayMs: 1000,
+      maxDelayMs: 10000,
+      onRetry,
+    });
+
+    // Advance past the first backoff delay (1000ms base + jitter)
+    await vi.advanceTimersByTimeAsync(1500);
+
+    const result = await promise;
+    expect(result).toBe("recovered");
+    expect(fn).toHaveBeenCalledTimes(2);
+    expect(onRetry).toHaveBeenCalledTimes(1);
+    expect(onRetry).toHaveBeenCalledWith(1, expect.any(Number), expect.any(Error));
+  });
+
+  it("throws after all retries are exhausted", async () => {
+    const rateLimitErr = new Error("rate_limit exceeded");
+    const fn = vi.fn().mockRejectedValue(rateLimitErr);
+    const onRetry = vi.fn();
+
+    const promise = withRateLimitRetry(fn, {
+      maxRetries: 2,
+      baseDelayMs: 100,
+      maxDelayMs: 1000,
+      onRetry,
+    });
+
+    // Advance enough to cover all backoff delays
+    for (let i = 0; i < 10; i++) {
+      await vi.advanceTimersByTimeAsync(500);
+    }
+
+    await expect(promise).rejects.toThrow("rate_limit exceeded");
+    expect(fn).toHaveBeenCalledTimes(3); // initial + 2 retries
+    expect(onRetry).toHaveBeenCalledTimes(2);
+  });
+
+  it("re-throws non-rate-limit errors immediately without retry", async () => {
+    const fn = vi.fn().mockRejectedValue(new Error("ENOENT: file not found"));
+    const onRetry = vi.fn();
+
+    await expect(
+      withRateLimitRetry(fn, { baseDelayMs: 1000, onRetry }),
+    ).rejects.toThrow("ENOENT: file not found");
+
+    expect(fn).toHaveBeenCalledTimes(1);
+    expect(onRetry).not.toHaveBeenCalled();
+  });
+
+  it("applies exponential backoff with increasing delays", async () => {
+    const fn = vi
+      .fn()
+      .mockRejectedValueOnce(new Error("429"))
+      .mockRejectedValueOnce(new Error("429"))
+      .mockResolvedValueOnce("ok");
+
+    const delays: number[] = [];
+    const onRetry = (_attempt: number, delayMs: number) => {
+      delays.push(delayMs);
+    };
+
+    // Use deterministic random for jitter
+    vi.spyOn(Math, "random").mockReturnValue(0.5); // jitter = 0
+
+    const promise = withRateLimitRetry(fn, {
+      baseDelayMs: 1000,
+      maxDelayMs: 10000,
+      onRetry,
+    });
+
+    await vi.advanceTimersByTimeAsync(1100); // 1st delay: 1000ms
+    await vi.advanceTimersByTimeAsync(2100); // 2nd delay: 2000ms
+
+    await promise;
+
+    expect(delays).toEqual([1000, 2000]);
+    expect(fn).toHaveBeenCalledTimes(3);
+
+    vi.spyOn(Math, "random").mockRestore();
+  });
+
+  it("caps delay at maxDelayMs", async () => {
+    const fn = vi
+      .fn()
+      .mockRejectedValueOnce(new Error("overloaded"))
+      .mockResolvedValueOnce("ok");
+
+    const delays: number[] = [];
+    vi.spyOn(Math, "random").mockReturnValue(0.5);
+
+    const promise = withRateLimitRetry(fn, {
+      baseDelayMs: 100000, // would exceed maxDelayMs
+      maxDelayMs: 5000,
+      onRetry: (_a, d) => delays.push(d),
+    });
+
+    await vi.advanceTimersByTimeAsync(6000);
+    await promise;
+
+    // baseDelayMs * 2^0 = 100000, capped to 5000
+    expect(delays[0]).toBe(5000);
+
+    vi.spyOn(Math, "random").mockRestore();
+  });
+
+  it("cancels backoff sleep when abort signal fires", async () => {
+    const fn = vi.fn().mockRejectedValue(new Error("429 rate limited"));
+    const ac = new AbortController();
+
+    const promise = withRateLimitRetry(fn, {
+      baseDelayMs: 60000,
+      maxDelayMs: 120000,
+      signal: ac.signal,
+    });
+
+    // Let first call fail and start sleeping
+    await vi.advanceTimersByTimeAsync(10);
+
+    // Abort during backoff
+    ac.abort(new Error("Task paused"));
+
+    await expect(promise).rejects.toThrow("Task paused");
+    expect(fn).toHaveBeenCalledTimes(1); // only initial call, no retry
+  });
+
+  it("does not retry if abort signal is already aborted", async () => {
+    const fn = vi.fn().mockRejectedValue(new Error("too many requests"));
+    const ac = new AbortController();
+    ac.abort(new Error("Already cancelled"));
+
+    await expect(
+      withRateLimitRetry(fn, { signal: ac.signal }),
+    ).rejects.toThrow("too many requests");
+
+    // fn called once, then abort check triggers throw before sleep
+    expect(fn).toHaveBeenCalledTimes(1);
+  });
+
+  it("classifies various rate limit error patterns correctly", async () => {
+    const patterns = [
+      "overloaded",
+      "rate limit exceeded",
+      "429 Too Many Requests",
+      "quota exceeded",
+      "billing limit reached",
+      "insufficient credit",
+    ];
+
+    for (const msg of patterns) {
+      const fn = vi
+        .fn()
+        .mockRejectedValueOnce(new Error(msg))
+        .mockResolvedValueOnce("ok");
+
+      const promise = withRateLimitRetry(fn, {
+        baseDelayMs: 100,
+        maxDelayMs: 100,
+      });
+      await vi.advanceTimersByTimeAsync(200);
+      const result = await promise;
+      expect(result).toBe("ok");
+      expect(fn).toHaveBeenCalledTimes(2);
+    }
+  });
+
+  it("handles non-Error thrown values", async () => {
+    const fn = vi.fn().mockRejectedValue("string error");
+
+    await expect(
+      withRateLimitRetry(fn, { baseDelayMs: 100 }),
+    ).rejects.toThrow("string error");
+
+    expect(fn).toHaveBeenCalledTimes(1); // not a rate limit error string
+  });
+
+  it("uses default options when none provided", async () => {
+    const fn = vi
+      .fn()
+      .mockRejectedValueOnce(new Error("429"))
+      .mockResolvedValueOnce("ok");
+
+    const promise = withRateLimitRetry(fn);
+
+    // Default baseDelayMs is 30000
+    await vi.advanceTimersByTimeAsync(35000);
+
+    const result = await promise;
+    expect(result).toBe("ok");
+  });
+});

packages/engine/src/rate-limit-retry.ts

@@ -0,0 +1,142 @@
+/**
+ * Rate Limit Retry — wraps async agent work with exponential backoff
+ * specifically for rate-limit / usage-limit errors.
+ *
+ * When an AI model returns a rate limit error (429, overloaded, quota, etc.),
+ * this utility retries the operation with exponential backoff before letting
+ * the error propagate to the caller's catch block, which triggers a global
+ * pause via `UsageLimitPauser`.
+ *
+ * **Backoff strategy:** `delay = min(baseDelayMs × 2^attempt, maxDelayMs)` with
+ * ±10 % jitter to avoid thundering-herd effects across concurrent agents.
+ *
+ * **Abort support:** An optional `AbortSignal` allows the engine to cancel
+ * pending retries when a task is paused, cancelled, or the engine is shutting
+ * down — so agents don't sit in a 2-minute sleep unnecessarily.
+ *
+ * **Scope:** Only rate-limit errors (as classified by `isUsageLimitError`) are
+ * retried. All other error types are re-thrown immediately so existing error
+ * handling (transient-error retry, failure marking, etc.) is unaffected.
+ */
+
+import { isUsageLimitError } from "./usage-limit-detector.js";
+
+export interface RateLimitRetryOptions {
+  /** Maximum number of retry attempts before re-throwing (default: 3). */
+  maxRetries?: number;
+  /** Initial backoff delay in milliseconds (default: 30 000 — 30 s). */
+  baseDelayMs?: number;
+  /** Upper bound on backoff delay in milliseconds (default: 120 000 — 2 min). */
+  maxDelayMs?: number;
+  /**
+   * Called before each retry with the attempt number (1-based) and the
+   * computed delay. Use this to log retry activity to the task and agent logs.
+   */
+  onRetry?: (attempt: number, delayMs: number, error: Error) => void;
+  /**
+   * Abort signal that, when triggered, cancels any pending backoff sleep and
+   * re-throws the last error immediately. Essential for paused / cancelled tasks.
+   */
+  signal?: AbortSignal;
+}
+
+/**
+ * Wrap an async function with rate-limit-aware exponential backoff.
+ *
+ * The wrapper calls `fn()`. If it throws a rate-limit error (detected via
+ * `isUsageLimitError`), it sleeps with exponential backoff and retries up to
+ * `maxRetries` times. Non-rate-limit errors are re-thrown immediately.
+ *
+ * After all retries are exhausted, the **original** error is thrown so the
+ * caller's existing catch block can trigger the global pause via
+ * `UsageLimitPauser`.
+ *
+ * @example
+ * ```ts
+ * await withRateLimitRetry(() => agentWork(), {
+ *   onRetry: (attempt, delayMs) =>
+ *     store.logEntry(taskId, `Rate limited — retry ${attempt} in ${delayMs}ms`),
+ *   signal: abortController.signal,
+ * });
+ * ```
+ */
+export async function withRateLimitRetry<T>(
+  fn: () => Promise<T>,
+  options: RateLimitRetryOptions = {},
+): Promise<T> {
+  const {
+    maxRetries = 3,
+    baseDelayMs = 30_000,
+    maxDelayMs = 120_000,
+    onRetry,
+    signal,
+  } = options;
+
+  let lastError: Error | undefined;
+
+  for (let attempt = 0; attempt <= maxRetries; attempt++) {
+    try {
+      return await fn();
+    } catch (err: any) {
+      const error = err instanceof Error ? err : new Error(String(err));
+
+      // Non-rate-limit errors: re-throw immediately — no retry
+      if (!isUsageLimitError(error.message)) {
+        throw error;
+      }
+
+      lastError = error;
+
+      // All retries exhausted — throw so caller can trigger global pause
+      if (attempt >= maxRetries) {
+        throw lastError;
+      }
+
+      // Check abort before sleeping
+      if (signal?.aborted) {
+        throw lastError;
+      }
+
+      // Exponential backoff with ±10 % jitter
+      const rawDelay = Math.min(baseDelayMs * 2 ** attempt, maxDelayMs);
+      const jitter = rawDelay * 0.1 * (2 * Math.random() - 1); // ±10 %
+      const delay = Math.max(0, Math.round(rawDelay + jitter));
+
+      onRetry?.(attempt + 1, delay, error);
+
+      await sleep(delay, signal);
+    }
+  }
+
+  // Unreachable, but satisfies TypeScript
+  throw lastError ?? new Error("withRateLimitRetry: unexpected state");
+}
+
+/**
+ * Sleep for `ms` milliseconds, cancellable via an `AbortSignal`.
+ * @internal exported for testing only
+ */
+export function sleep(ms: number, signal?: AbortSignal): Promise<void> {
+  return new Promise<void>((resolve, reject) => {
+    if (signal?.aborted) {
+      reject(signal.reason ?? new Error("Aborted"));
+      return;
+    }
+
+    const timer = setTimeout(resolve, ms);
+
+    if (signal) {
+      const onAbort = () => {
+        clearTimeout(timer);
+        reject(signal.reason ?? new Error("Aborted"));
+      };
+      signal.addEventListener("abort", onAbort, { once: true });
+      // Clean up listener when timer fires normally
+      const origResolve = resolve;
+      resolve = () => {
+        signal.removeEventListener("abort", onAbort);
+        origResolve();
+      };
+    }
+  });
+}

packages/engine/src/triage.ts

@@ -22,6 +22,7 @@ import {
   type UsageLimitPauser,
 } from "./usage-limit-detector.js";
 import { isTransientError } from "./transient-error-detector.js";
+import { withRateLimitRetry } from "./rate-limit-retry.js";
 
 export const TRIAGE_SYSTEM_PROMPT = `You are a task specification agent for "kb", an AI-orchestrated task board.
 
@@ -640,10 +641,18 @@ export class TriageProcessor {
         }
       };
 
+      const retryableWork = () => withRateLimitRetry(agentWork, {
+        onRetry: (attempt, delayMs, error) => {
+          const delaySec = Math.round(delayMs / 1000);
+          triageLog.warn(`⏳ ${task.id} rate limited — retry ${attempt} in ${delaySec}s: ${error.message}`);
+          this.store.logEntry(task.id, `Rate limited — retry ${attempt} in ${delaySec}s`).catch(() => {});
+        },
+      });
+
       if (this.options.semaphore) {
-        await this.options.semaphore.run(agentWork, PRIORITY_SPECIFY);
+        await this.options.semaphore.run(retryableWork, PRIORITY_SPECIFY);
       } else {
-        await agentWork();
+        await retryableWork();
       }
     } catch (err: any) {
       // Race condition: task was deleted (e.g. as a duplicate) between listTasks()