feat: allow skipping tasks that exceed time/cost threshold in auto mode

[rubenmarcus]

Feature: Interactive controls during auto mode loops

Problem

In auto mode, once the loop starts, the user has no way to interact with it. If a task is stuck, taking too long, or the user realizes it should be skipped, they have to kill the entire process and restart.

Proposed: Live interaction during loops

While auto mode is running, the user should be able to interact with the loop in real-time via the terminal:

Interactive keybinds (during loop execution)

Key	Action
s	Skip — stop current task, move to next in queue
p	Pause — pause after current iteration, wait for resume
r	Resume — continue after pause
d	Debug — toggle verbose output (agent commands, file reads/writes, tool calls, errors in real-time)
i	Info — show current task, iteration count, elapsed time, cost so far
h	Hint — pause and let user type a one-line instruction for the agent to consider next iteration (e.g., "try using esm imports instead")
q	Quit — graceful shutdown after current iteration completes

Debug mode detail

When d is pressed, toggle between:

• Normal mode (default): shows iteration number, pass/fail status, cost per iteration
• Debug mode: streams the agent's activity in real-time:
  • Files being read/written
  • Commands being executed + their output
  • Tool calls and their results
  • Validation output (build/test results)
  • Error messages and stack traces

Debug mode is especially useful when:

• A task keeps failing and you can't tell why from the summary
• You want to verify the agent isn't going in circles
• You want to learn what approach the agent is taking before deciding to skip

Toggle on/off anytime with d. Debug state persists across iterations until toggled off.

UX during the loop

🔄 Task 3/7: feat: scaffold Next.js landing page
   Iteration 4/10 | 2m 13s | $0.84
   [s] skip  [p] pause  [d] debug  [i] info  [h] hint  [q] quit

With debug on:

🔄 Task 3/7: feat: scaffold Next.js landing page  [DEBUG ON]
   Iteration 4/10 | 2m 13s | $0.84
   │ READ package.json
   │ EDIT src/app/page.tsx (lines 12-34)
   │ BASH npm run build
   │ └─ error TS2307: Cannot find module './components/Hero'
   │ EDIT src/app/page.tsx (line 2) — fixing import
   │ BASH npm run build
   │ └─ ✓ compiled successfully
   [s] skip  [p] pause  [d] debug off  [i] info  [h] hint  [q] quit

Additionally: pre-configured thresholds (optional)

For unattended runs (overnight, CI), set automatic skip thresholds via CLI flags:

• --max-time-per-task 30m — skip task if it exceeds N minutes
• --max-cost-per-task 5 — skip task if it exceeds $N
• --debug — start with debug mode on from the beginning

These work alongside interactive controls — if the user is watching, they can interact manually. If running unattended, the thresholds kick in automatically.

When a task is skipped (manual or automatic)

• Log the skip reason in activity.md (user-skipped vs timeout vs cost-limit)
• Comment on the issue explaining it was skipped and include last error if available
• Optionally label the issue (e.g., needs-human-review)
• Move on to the next task in the queue

Use cases

1. Watching the loop — user sees a task is struggling, presses s to skip and move on
2. Debugging a stuck task — press d to see what the agent is doing, realize it's looping on the same error, press h to give it a hint
3. Quick check — press i to see elapsed time and cost
4. Overnight batch — set --max-time-per-task 20m and let it run unattended
5. Redirecting the agent — press h and type "the test needs jsdom, add it to vitest config" to nudge it in the right direction