No No Debug

no-no-debug — Self-Evolution System for AI Coding Assistants

Overview

Six mechanisms work together to eliminate repeated mistakes permanently:

• Mechanism 1: Real-time logging — errors, corrections, and failures auto-appended to ~/.claude/memory/error_log.md
• Mechanism 2: Silent 3-gate checkpoint before every code change (no output when all clear)
• Mechanism 3: Periodic review that reads error_log.md, classifies errors, updates a persistent tracker
• Mechanism 4: Rule accumulation — new error types auto-added, repeat offenses strengthen gates, 4 clean periods = cured
• Mechanism 5: Confirmation gate — must ask user before new features, DB/env/deploy changes, external publishing, or new directions
• Mechanism 6: Auto hooks — passive capture configured at install time

Configuration

To change review frequency, add a comment to the top of error_tracker.md:

<!-- review_frequency: 1 -->

Mechanism 1 — Real-time Logging

Always active. Runs passively. No user action required.

Append a line to ~/.claude/memory/error_log.md whenever any of the following occur:

Trigger conditions

Log format

[YYYY-MM-DD HH:MM] TYPE | description

Examples:

[2024-04-08 14:32] BUILD_FAIL | npm run build failed — cannot find module './utils/auth'
[2024-04-08 15:01] USER_CORRECTION | User said "wrong" — had used cached data instead of live fetch
[2024-04-08 16:45] REPEATED_FIX | Applied the same null-check fix to userProfile.ts for the second time
[2024-04-09 09:12] AUTH_FAIL | Login redirect broken after updating next.config.js

File initialization

If ~/.claude/memory/error_log.md does not exist, create the directory and file with this header:

# Error Log
<!-- Auto-maintained by no-no-debug -->
<!-- Format: [YYYY-MM-DD HH:MM] TYPE | description -->

Mechanism 2 — Three-Gate Checkpoint (Silent)

Always active. No output when all gates pass.

Before ANY code change, file edit, configuration update, or deployment action, silently verify all three gates:

Gate 1 — Before Making the Change

• What exactly does this change affect?
• Could authentication or login break?
• Does this touch .env, database schema, connection strings, or credentials?
• Does this modify shared infrastructure used by other features?
• Is this a new feature or a bug fix? (If new feature → trigger Mechanism 5 first)

Gate 2 — After Making the Change

• Did I actually verify the result, or just check that a command ran without error?
• Did I verify using the real production command string (the exact hook/CLI/route handler users will hit), not a simplified test harness or sandbox invocation?
• Did I walk the flow as a real user would, end to end?
• Did I check the failure path, not just the success path?
• Does the output shown to the user match what is actually in the code?
• For shell/hook changes with multiple escape layers (JSON → shell → interpreter), did I run the exact string from settings.json rather than a locally-equivalent variant?

Gate 3 — Before Deploying or Publishing

• Did I test with a non-admin account?
• Is the permission table complete for all new routes and actions?
• Are all required database migrations included?
• Are there cache invalidation requirements?
• For changes reviewed by a second agent (e.g. another model, sandboxed reviewer): did I independently re-run the verification in my own environment rather than trusting the reviewer's "pass" report? A sandbox pass is not a real-environment pass.
• Does this publish to an external platform? (If yes → trigger Mechanism 5 first)

Output rule: Stay completely silent when all gates pass. When any gate raises a concern, surface it to the user before proceeding.

Mechanism 3 — Periodic Review

Auto-trigger condition

A UserPromptSubmit hook (hooks/review_reminder.py) checks ~/.claude/memory/error_tracker.md for Last Review Date on each user message. If elapsed days since that date are greater than or equal to the configured review frequency (default: 3), the hook outputs a reminder. When you see this reminder, run the review automatically. A 4-hour cooldown prevents repeated reminders within the same session.

Manual trigger

User says any of: error review, 错误追踪, 进化报告, evolution report

Step 1 — Initialize files if missing

If ~/.claude/memory/error_log.md does not exist, create it (see Mechanism 1 initialization).

If ~/.claude/memory/error_tracker.md does not exist:

1. Create the directory ~/.claude/memory/ if needed
2. Write the following template to ~/.claude/memory/error_tracker.md:

# Error Tracker
<!-- review_frequency: 3 -->

Last Review Date: {TODAY}
Review Count: 0
Total Lifetime Errors: 0

## Active Dimensions

| Dimension | Total | Last Seen | Clean Periods | Status |
|-----------|-------|-----------|---------------|--------|
| 数据准确性 / Data Accuracy | 0 | — | 0 | Active |
| 环境安全 / Environment Safety | 0 | — | 0 | Active |
| 预见性 / Foresight | 0 | — | 0 | Active |
| 用户视角 / User Perspective | 0 | — | 0 | Active |
| 验证完整性 / Verification | 0 | — | 0 | Active |
| 记忆一致性 / Memory Consistency | 0 | — | 0 | Active |
| 工具判断 / Tool Judgment | 0 | — | 0 | Active |
| 审查覆盖 / Review Completeness | 0 | — | 0 | Active |
| 操作精准 / Operational Precision | 0 | — | 0 | Active |
| 先查后做 / Check Before Doing | 0 | — | 0 | Active |
| 简洁性 / Conciseness | 0 | — | 0 | Active |
| 回归意识 / Regression Awareness | 0 | — | 0 | Active |
| 风格一致性 / Style Consistency | 0 | — | 0 | Active |
| 独立判断 / Independent Judgment | 0 | — | 0 | Active |
| 真实环境验证 / Real-env Verification | 0 | — | 0 | Active |
| 跨 agent 采信 / Cross-agent Trust | 0 | — | 0 | Active |
| 人类将要犯的蠢 / Dumb things humans will do | 0 | — | 0 | Active |
| AI 将要犯的蠢 / Dumb things AI will do | 0 | — | 0 | Active |

## Cured Dimensions (4+ consecutive clean periods)

_None yet_

## Prevention Rules

_Auto-populated as errors are discovered_

## History

| Period | Date Range | New Errors | Notes |
|--------|------------|------------|-------|

3. Output the first-run message (see First-Run Output section below)

Step 2 — Read error_log.md

Read ~/.claude/memory/error_log.md and collect all entries from Last Review Date to today.

If the file is empty or has no entries in the review window, note that and proceed with zero counts.

If a session search tool (e.g. claude-mem) is also available, supplement with a search for:

correction mistake wrong error broken fix failed incorrect fabricated

If both the log file has no entries and no search tool is available, ask the user:

"Can you briefly describe any mistakes or corrections from this review period? (Or say 'none' to skip.)"

Step 3 — Classify each entry

Map each log entry or incident to one of the active dimensions:

If an entry does not match any existing dimension, treat it as a new dimension (see Mechanism 4).

Step 4 — Update the tracker

For each dimension:

• Errors found this period: increment Total, reset Clean Periods to 0, update Last Seen to today
• No errors this period: increment Clean Periods by 1
• Clean Periods reaches 4: move dimension row to "Cured Dimensions" section

Update:

• Last Review Date to today
• Review Count + 1
• Total Lifetime Errors + sum of new errors this period
• Add a row to the History table

Write all changes to disk before reporting.

Step 5 — Output the report

Detect language from conversation context. Use the matching format:

Chinese format:

进化报告 R{N}（{开始日期} – {结束日期}）
━━━━━━━━━━━━━━━━━━━━━━━━━━━━

维度                  本期    累计    状态
──────────────────────────────────────────
数据准确性              {n}     {total}   {status}
环境安全                {n}     {total}   {status}
预见性                  {n}     {total}   {status}
用户视角                {n}     {total}   {status}
验证完整性              {n}     {total}   {status}
记忆一致性              {n}     {total}   {status}
工具判断                {n}     {total}   {status}
审查覆盖                {n}     {total}   {status}
操作精准                {n}     {total}   {status}
先查后做                {n}     {total}   {status}
简洁性                  {n}     {total}   {status}
回归意识                {n}     {total}   {status}
风格一致性              {n}     {total}   {status}
独立判断                {n}     {total}   {status}
真实环境验证            {n}     {total}   {status}
跨 agent 采信           {n}     {total}   {status}
人类将要犯的蠢          {n}     {total}   {status}
AI 将要犯的蠢           {n}     {total}   {status}

合计: {sum} | 已根治: {cured}/{total_dimensions}
━━━━━━━━━━━━━━━━━━━━━━━━━━━━

重灾区: {top dimension}({n}) > {second}({n})
下次审查: {date}

English format:

Evolution Report R{N} ({start date} – {end date})
━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Dimension                    This period  Total   Status
─────────────────────────────────────────────────────────
Data Accuracy                    {n}      {total}  {status}
Environment Safety               {n}      {total}  {status}
Foresight                        {n}      {total}  {status}
User Perspective                 {n}      {total}  {status}
Verification                     {n}      {total}  {status}
Memory Consistency               {n}      {total}  {status}
Tool Judgment                    {n}      {total}  {status}
Review Completeness              {n}      {total}  {status}
Operational Precision            {n}      {total}  {status}
Check Before Doing               {n}      {total}  {status}
Conciseness                      {n}      {total}  {status}
Regression Awareness             {n}      {total}  {status}
Style Consistency                {n}      {total}  {status}
Independent Judgment             {n}      {total}  {status}
Real-env Verification            {n}      {total}  {status}
Cross-agent Trust                {n}      {total}  {status}
Dumb things humans will do       {n}      {total}  {status}
Dumb things AI will do           {n}      {total}  {status}

Total: {sum} | Cured: {cured}/{total_dimensions}
━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Hot spots: {top dimension}({n}) > {second}({n})
Next review: {date}

Show cured dimensions separately at the end of the report with Moved to Cured: {dimension name} when a new one crosses the threshold this period.

Mechanism 4 — Rule Accumulation and Self-Evolution

New error type discovered

When an entry cannot be classified into any existing dimension:

1. Add a new row to the "Active Dimensions" table with:

   • A short descriptive name (bilingual if possible)
   • Total: 1, Last Seen: today, Clean Periods: 0, Status: Active

2. Add a new entry to the "Prevention Rules" section of the tracker:

   ### {Dimension Name}
   - **What it is**: {one-sentence description}
   - **How to detect**: {observable signal}
   - **Prevention**: {specific check to add to Gate 1, 2, or 3}
   - **Added**: {today}
   - **Occurrences**: 1

3. Include in the report: New dimension added: {name}

Repeat offense — rule strengthening

When any dimension records 3 or more total errors:

• Review the Prevention Rule for that dimension
• Add one specific, concrete check to the appropriate Gate in Mechanism 2
• Note the strengthening in the tracker's "Prevention Rules" section