DEV Community

anicca
anicca

Posted on

How I Discovered My AI App Builder Has No 50-Iteration Limit

#ai #debugging #automation #claude

TL;DR

My Mobile App Factory kept stopping around 50 iterations. After digging through logs, I found there's no actual limit. Claude Code completes 2+ tasks too efficiently → ralph.sh enforces "1 iteration = 1 task" rule → auto-resets extra completions → prd.json flags become unreliable → progress.txt (immutable log) is the true source of truth.

The Mystery: Why 50 Iterations?

For weeks I wondered: Why does my Factory always stop around iteration 50?

Looking at Desk Stretch Timer's prd.json, I saw US-008b/c/d/e all marked passes: false with notes saying [reset: 1 US per iteration rule].

{
  "id": "US-008e",
  "title": "Preflight + TestFlight",
  "passes": false,
  "notes": "[reset: 1 US per iteration rule]"
}
Enter fullscreen mode Exit fullscreen mode

But checking progress.txt (an append-log), US-008e was completed on 2026-03-06. Contradiction.

Investigation: Three Sources

1. progress.txt (append-only log) 

2026-03-06 US-008e GREENLIT (0 blocking)
2026-03-06 US-008e asc validate Errors=0
2026-03-06 US-008e TestFlight beta review WAITING_FOR_REVIEW
2026-03-06 US-008e Public link: https://testflight.apple.com/join/78sH7CE9
Enter fullscreen mode Exit fullscreen mode

→ US-008e is complete

2. prd.json 

{
  "id": "US-008e",
  "passes": false
}
Enter fullscreen mode Exit fullscreen mode

→ Completion flag is false (contradiction)

3. ralph.sh (control script)

Here's where I found the truth:

if NEW_COUNT > 1:
    echo "🔴 1 US enforcement violation: $NEW_COUNT US completed in 1 iteration"
    # Keep the first completion, reset the rest
    KEEP_FIRST=$(echo "$NEW_PASSES" | xargs | awk '{print $1}')
    # Reset remaining to passes: false
Enter fullscreen mode Exit fullscreen mode

Root Cause: 1 Iteration = 1 US Enforcement

Claude Code is too efficient. It completes 2+ User Stories in a single iteration.

ralph.sh enforces "1 iteration = 1 US" for quality control. Any extra completions get auto-reset to passes: false.

Result:

  • prd.json: Auto-reset target → unreliable
  • progress.txt: Immutable log → true source of truth

Fix: Trust progress.txt 

# Check progress this way
cat progress.txt | grep "US-008e"

# prd.json passes flag is just a reference
cat prd.json | jq '.userStories[] | select(.id=="US-008e")'
Enter fullscreen mode Exit fullscreen mode

Actual Desk Stretch Timer status:

  • US-001 ~ US-008e: Complete (confirmed via progress.txt)
  • US-009 (App Privacy + submit): Not started

Key Takeaways

Lesson Detail
Immutable logs are truth Overwritable flags (prd.json) get auto-reset. Immutable logs (progress.txt) are the reliable completion record
Efficiency has side effects Claude Code's speed → 1 iteration = 1 US enforcement → auto-resets → appears "stuck"
No limit exists "50-iteration limit" was a misunderstanding. Real issue: quality control design (1 US enforcement)

Summary

There was no "max 50" limit. The real cause: Claude Code's efficiency × quality control's 1-iteration-per-US rule creating auto-reset behavior.

Debugging takeaways:

  1. Read 3+ sources (prd.json, progress.txt, scripts)
  2. Trust immutable logs over overwritable flags
  3. Look for design intent behind "limits"

A quiet Sunday spent deep-diving past logs revealed system behavior I hadn't seen before.

Top comments (1)

Some comments may only be visible to logged-in visitors. Sign in to view all comments.