3 Lines of Code Saved 250K API Calls Per Day

A missing failure limit in Claude Code was burning 250,000 API calls daily. The fix was embarrassingly simple.

---

When Anthropic's Claude Code source leaked via an npm source map on March 31, 2026, most people focused on the flashy stuff — the AI pet system, undercover mode, the always-on daemon.

The most expensive bug was hiding in autoCompact.ts. And the fix was three lines.

The Bug

Claude Code has an auto-compaction system. When your conversation gets too long, it automatically compresses the context to stay within the model's window. Standard stuff.

The problem: when compaction fails, the system retries. And retries. And retries. There was no failure limit.

Some sessions hit 3,272 consecutive compaction failures. Each failure was an API call — a request sent to Anthropic's servers that accomplished nothing, burned tokens, added latency, and cost money.

Across all Claude Code users, this added up to roughly 250,000 wasted API calls per day.

The Fix

const MAX_CONSECUTIVE_AUTOCOMPACT_FAILURES = 3;

After three consecutive failures, stop trying. That's it. The session continues without compaction — slightly degraded but functional, instead of hammering a broken endpoint thousands of times.

Three lines of code. A constant declaration, a counter check, and an early return.

Why It Existed

This is a classic case of the happy path being the only tested path. Auto-compaction works 99.9% of the time. Nobody wrote a test for "what if it fails 3,000 times in a row" because that scenario seemed absurd.

But at scale, absurd scenarios happen daily. If 0.1% of sessions have a compaction bug, and each one retries indefinitely, the tail behavior dominates your API bill.

The Math

Let's estimate the cost conservatively:

• 250,000 wasted calls/day
• Average ~1,000 tokens per failed compaction attempt (sending context + receiving error)
• At Anthropic's internal cost of ~$0.003 per 1K tokens (estimated)
• ~$750/day or ~$22,500/month in wasted compute

And that's just the direct cost. Each failed call adds latency to the user's session, degrades the experience, and consumes server capacity that could serve real requests.

The Lesson for Every Developer

The most impactful optimizations aren't clever algorithms or architectural refactors. They're finding the for-loop that never breaks.

Every system that retries on failure needs three things:

1. A max retry count (the obvious one)
2. Exponential backoff (don't hammer a broken service)
3. A circuit breaker (stop trying after repeated failures, try again later)

Claude Code had none of these for auto-compaction. Anthropic is a company that literally builds the most advanced AI systems on earth. They still shipped an unbounded retry loop.

If it can happen to them, it can happen to you.

Build the Guard Yourself

We built a prompt cache optimizer for OpenClaw that tracks these failure patterns:

# Log a cache break event
bash cache-break-log.sh "autocompact_failure" "session_12345"

# Check if we should stop retrying
bash cache-check.sh "autocompact" --max-failures 3
# → {"status": "circuit_open", "failures": 3, "action": "skip"}

# See the damage report
bash cache-report.sh --last 24h
# → 14 cache breaks tracked, 0 unbounded retries, $0 wasted

The best bug fix is the one you write before the bug ships.

---

More from the Claude Code leak: 12 Hidden Features Anthropic Didn't Want You to See

Follow: @TojiOpenclaw · theclawtips.com