Coffee Debt is a hook-based system that turns Claude Code mistakes into gamified observability data. Every AI error (tool failure, user correction, blocked destructive command) adds a bean. 5 beans = 1 coffee debt. The debt is cumulative and never resets. The goal isn't punishment, it's data collection.

Which errors add beans to Coffee Debt?

Three sources: (1) Tool failures: when an Edit command can't find a match, when a Bash command returns a non-zero exit code, when a Write operation fails. (2) User corrections: when Turkish or English correction phrases are detected (yanlis, hatali, that's wrong, you hallucinated, etc.). (3) Blocked commands: when destructive commands like rm -rf, git reset --hard, git push --force are stopped by the hook.

What does deep analysis do?

It cross-references 4 data sources. coffee-log.jsonl (error records), hermes.db (web search history, was there a search within 5 minutes before the error), state.db (file complexity and static analysis results, are error-prone files also complex), knowledge.db (file interaction history, which files are touched most often). This cross-referencing provides deeper clues about why errors occur.

What are context proxies for?

The prompt_count and session_kb fields added to each error record show the conditions under which the error occurred. prompt_count is the number of prompts in the session, session_kb is the size of the session JSONL file (a proxy for context window usage). With this data, hypotheses like 'does the error rate increase as context grows' or 'do errors cluster as the session gets longer' become testable.

AI Pair Programming: Coffee Debt, Gamifying Error Tracking

AI coding assistants make mistakes. Edit commands fail to match, bash scripts crash, users have to step in and correct. The point isn't getting frustrated at errors, it's logging them and extracting patterns. The Coffee Debt system does exactly that.

TL;DR

Every AI mistake is a bean, five beans make a coffee debt. Four hook scripts capture errors automatically: tool failures, user corrections, blocked destructive commands. Everything is logged as JSONL alongside context proxies. Fourteen days of data show the errors are not spread randomly; they cluster in specific tools and specific hours. Data instead of frustration.

View Subscription Sign In

Subscribers Only

The full article plus 21 more subscriber-only posts come with a subscription.

ai developer-tools

Subscribers Only

RELATED

Argus: Make Your AI Coding Assistant's Web Searches Visible

Which Files Do You Actually Read? Developer Interaction Tracking

Living Architecture Documentation for AI Coding Agents: Research, Approaches, and Tools