Introduction
What is PRECC?
PRECC (Claude Code 預測性錯誤糾正) is a Rust tool that intercepts Claude Code bash commands via the official PreToolUse hook mechanism. It fixes errors before they happen, saving tokens and eliminating retry loops.
對社區用戶免費。
The Problem
Claude Code wastes significant tokens on preventable mistakes:
- Wrong-directory errors – Running
cargo buildin a parent directory that has noCargo.toml, then retrying after reading the error. - Retry loops – A failed command produces verbose output, Claude reads it, reasons about it, and retries. Each cycle burns hundreds of tokens.
- Verbose output – Commands like
findorls -Rdump thousands of lines that Claude must process.
The Four Pillars
上下文修復 (cd-prepend)
檢測到 cargo build 或 npm test 等命令在錯誤的目錄中運行時,在執行前添加 cd /正確/路徑 &&。
GDB Debugging
Detects opportunities to attach GDB for deeper debugging of segfaults and crashes, providing structured debug information instead of raw core dumps.
Session Mining
Mines Claude Code session logs for failure-fix pairs. When the same mistake recurs, PRECC already knows the fix and applies it automatically.
Automation Skills
A library of built-in and mined skills that match command patterns and rewrite them. Skills are defined as TOML files or SQLite rows, making them easy to inspect, edit, and share.
How It Works (30-Second Version)
- Claude Code is about to run a bash command.
- The PreToolUse hook sends the command to
precc-hookas JSON on stdin. precc-hookruns the command through the pipeline (skills, directory correction, compression) in under 3 milliseconds.- The corrected command is returned as JSON on stdout.
- Claude Code executes the corrected command instead.
Claude never sees the error. No tokens wasted.
Adaptive Compression
如果命令在壓縮後失敗,PRECC會自動在重試時跳過壓縮,以便Claude獲得完整的未壓縮輸出來除錯。
Live Usage Statistics
| Metric | Value |
|---|---|
| Hook invocations | – |
| Tokens saved | – |
| Saving ratio | –% |
| RTK rewrites | – |
| CD corrections | – |
| Hook latency | – ms (p50) |
數字為估算值。每次預防的失敗避免了完整的重試循環:錯誤輸出、模型推理和重試命令。 These numbers update automatically from anonymized telemetry.
Links
- GitHub: https://github.com/peria-ai/precc-cc
- Website: https://peria.ai
- Documentation: https://precc.cc