デバッグ

INIT

調査・計測・捕獲・修正・検証・清掃

第一版二〇二五

工具 MCP Server for AI Coding Agents

Arm your AGENTS.
Hunt every bug.虫狩り

One MCP server. Full runtime visibility. Your agent sees the code execute, not just reads it.

$ npx debug-toolkit init

$ npx debug-toolkit demo

━━━ Step 1: debug_investigate ━━━

✓ Error type: TypeError — type

✓ Source: src/api.ts:8

>> 8 | const names = users.map(u => u.name); // BUG

✓ Suggestion: Check for null/undefined values

━━━ Step 3: debug_capture ━━━

✓ [DBG_001] users = undefined

✓ Hypothesis confirmed → CONFIRMED

━━━ Step 5: debug_verify ━━━

✓ Exit code: 0 ✓ Errors: 0 ✓ Verdict: PASSED

$ ▊

問題点

問題 The Problem

AI agents
debug blind

When an AI agent hits a bug, it reads code and guesses a fix. You run it, paste the error back, the agent guesses again.

✗ Without debug-toolkit

8–12 conversation turns per bug

Paste error → guess fix → repeat. No runtime data, no learning, no context.

✓ With debug-toolkit

1–2 turns with full context

Investigate → instrument → capture → fix → verify → cleanup. Diagnosis saved for next time.

道具 Tools

Eight tools.
Zero config.

Everything an agent needs to go from error to verified fix.

debug_investigate

Start here. Give it an error — get classification, source code at crash site, git context, environment, and past solutions.

Entry point

debug_instrument

Add tagged logging to source files. Supports JS/TS, Python, Go, and Rust. Each marker links to a hypothesis.

Multi-language

debug_capture

Run a command and capture output. Tagged output is linked to hypotheses. Drains terminal, browser, and Tauri logs.

Runtime

debug_verify

After applying a fix, run the test command and get a clear pass/fail with exit code and error output.

Validation

debug_cleanup

Remove all instrumentation, verify files are restored, save diagnosis + causal chain to memory.

Persistence

debug_recall

Search past sessions for similar errors. Returns diagnoses ranked by relevance with staleness tracking.

Memory

debug_patterns

Detect recurring errors, hot files, regressions, and error clusters across all past sessions.

Analytics

debug_session

Lightweight view of current state: hypotheses, active instruments, recent captures.

Status

言語 Language Support

Polyglot from
day one

First-class support where it matters. Growing with every release.

Feature	JS / TS	Python	Go	Rust / Tauri
Stack trace parsing	✓	✓	—	✓
Error classification	✓	—	—	✓
Code instrumentation	✓	✓	✓	✓
Source extraction	✓	✓	✓	✓
Log file tailing	—	—	—	✓
Environment detection	✓	✓	—	✓

記憶 Memory System

Learns from
every session

debug-toolkit remembers what worked. Next time the same bug appears, the agent already knows the fix.

Recall

Auto-searches past diagnoses when investigating new errors. Similar errors get previous solutions and causal chains.

Staleness

Every diagnosis tagged with git SHA. Recalled diagnoses checked against current code. Stale ones are flagged.

Causal Chains

Records where the error appeared vs. where the actual bug was. Next time, go straight to the cause file.

Patterns

Recurring errors, hot files, regressions, error clusters — systemic issues detected across all sessions.

構造

構造 Architecture

Small surface.
Big capability.

Files

3,451

Lines of TypeScript

npm dependencies

mcp.ts473

context.ts420

demo.ts398

memory.ts369

capture.ts276

proxy.ts200

index.ts204

security.ts185

session.ts162

injected.js151

instrument.ts140

cleanup.ts126

cli.ts125

hook.ts122

methodology.ts100

安全 Security

Locked down
by default

—

Path traversal protection

All file ops validated against project root

—

Expression validation

Blocks eval, require, exec in instrumentation

—

Secret redaction

Tokens, API keys, passwords auto-redacted

—

Localhost-only proxy

Binds to 127.0.0.1

—

Pre-commit hook

Blocks commits containing debug markers

—

Atomic writes

Temp file + rename, no corruption on crash

Arm your AGENTS.Hunt every bug.虫狩り