Wisegit

wisegit

*"Don't take a fence down until you know the reason it was put up."*

— G.K. Chesterton

wisegit is a local MCP server that extracts decision intent from git history and protects intentional code from AI modification.

When Claude Code (or any MCP-compatible agent) is about to edit a file, wisegit injects a decision manifest showing which functions are frozen, stable, or open — so the AI respects what was intentional, not just what compiles.

Zero config. Zero external services. Everything local.

Install

# Set up any repo (one command)

npx @sandip124/wisegit setup

Or add as MCP server globally

claude mcp add wisegit -- npx @sandip124/wisegit serve

Published on

• npm: @sandip124/wisegit
• MCP Registry: io.github.Sandip124/wisegit
• GitHub: Sandip124/wisegit

The Problem

LLMs have no concept of intentional code. A manually-tested fix and a broken stub look identical — both are just text. Real scenario:

1. You fix a Stripe race condition with sleep(350) — manually tested, committed.
2. Next session: "find bugs." Claude removes sleep(350) — looks like dead code.
3. Production incident.

Root cause: git history contains proof of intention. Nobody extracts it.

How It Works

Git History → Tree-sitter AST → Intent Extraction → SQLite Event Store → MCP Tools

1. Indexes your git history — walks every commit, parses diffs at the AST level (function boundaries, not line counts)
2. Classifies commits — STRUCTURED (fix:, feat:), DESCRIPTIVE (plain sentences), or NOISE (wip, x)
3. Extracts intent — rule-based for structured/descriptive commits, LLM for noise (Phase 2)
4. Computes freeze scores — 0–1 per function, derived from protection signals (git, issue, code structure, test, structural, Naur, Aranda) minus adaptive obsolescence penalty (8 signals, entropy-calibrated). Age uses Weibull survival [18]; expertise uses DOE model [19]
5. Serves decision manifests via MCP — Claude Code calls get_file_decisions before editing any file

What the AI Sees

[DECISION MANIFEST: payment.service.cs]

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ FROZEN: ProcessPayment() [score: 0.89] [Recovery: L1]

• sleep(350) → Stripe race condition. Won't Fix. HIGH — commit a3f19b2

STABLE: ValidateOrder() [score: 0.55] [Recovery: L2]

• Fixed null reference on Safari iOS WebKit. MEDIUM — commit 7c14694

OPEN: FormatReceipt() [score: 0.12] [Recovery: L3] ← safe to modify ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

• FROZEN (score >= 0.80): Do not modify without explicit user approval
• STABLE (score 0.50-0.79): Proceed with caution, review intent first
• OPEN (score < 0.50): Safe to modify freely

Quick Start

Prerequisites

• Node.js >= 20

That's it. No Docker, no PostgreSQL, no external services.

1. Set Up a Repository (one command)

cd /path/to/your/repo

npx @sandip124/wisegit setup

This single command:

• Creates a local SQLite database at ~/.wisegit/wisegit.db
• Indexes your entire git history (462 commits in ~13 seconds)
• Creates .mcp.json for Claude Code auto-discovery
• Creates CLAUDE.md rules that instruct AI to check before editing
• Adds .mcp.json to .gitignore

2. Enrich with Issue Context (optional)

# Fetch issue/PR details from GitHub/GitLab

GITHUB_TOKEN=ghp_... npx @sandip124/wisegit enrich

This fetches referenced issues (e.g., #134 in commit messages), detects Won't Fix / By Design decisions, and boosts freeze scores for functions linked to those issues.

3. Done

Open the repo in Claude Code. It will automatically:

1. Start the wisegit MCP server (via .mcp.json)
2. Read the protection rules (via CLAUDE.md)
3. Call get_file_decisions before editing any file

MCP Tools

MCP Resource: wisegit://manifest/{filePath} — decision manifest as auto-discoverable resource

MCP Prompt: check_before_edit — mandatory workflow prompt that returns the decision manifest before editing any file

LLM Intent Extraction Strategy

wisegit uses a smart fallback chain for extracting intent from NOISE commits:

Inside Claude Code, call extract_intent to retroactively recover intent for NOISE commits — uses Claude itself, no Ollama installation needed.

CLI Commands

wisegit setup [--path <dir>] [--global]         # One-command repo setup

wisegit init [--full-history] [--path ] # Index git history wisegit enrich [--path ] # Fetch issue/PR context from GitHub/GitLab wisegit audit # Show decision manifest wisegit history [--file ] # Show decision timeline wisegit recompute [--path ] # Recompute scores with PageRank + theory gaps wisegit override --file --reason "..." # Override a frozen function wisegit overrides # List active overrides wisegit sync # Rebuild local cache from git + .wisegit/ wisegit config list # View team configuration wisegit config set # Modify team policy wisegit team-status # Team overview: enrichments, overrides, contributors wisegit team-health # Theory health: healthy/fragile/critical functions wisegit branch-capture # Capture branch context from last merge wisegit branch-list # List all captured branch snapshots wisegit branch-recover # Recover context from old merge commit wisegit calibrate # Show adaptive obsolescence weights vs defaults wisegit report [--output ] # Generate HTML report with scores + insights wisegit serve # Start MCP server (stdio) wisegit hook install|uninstall # Manage git hooks (post-commit + post-merge)

Configure for Claude Code

Option A: Per-repo (recommended)

Run npx @sandip124/wisegit setup in any repo. It creates .mcp.json automatically.

Option B: Global registration

claude mcp add wisegit -- npx @sandip124/wisegit serve

Option C: Manual .mcp.json

Create .mcp.json in your repo root:

{

"wisegit": { "command": "npx", "args": ["@sandip124/wisegit", "serve"] } }

Supported Languages

More languages can be added via Tree-sitter grammar configs in src/ast/languages/.

Issue Enrichment

A commit saying fix: handle null token #134 points to an issue containing reproduction steps, root cause, and explicit decision rationale — everything the commit message never says.

# Fetch issue context from GitHub/GitLab

wisegit enrich --path /path/to/repo

With auth (5000 req/hr instead of 60)

GITHUB_TOKEN=ghp_... wisegit enrich

Supported platforms: GitHub, GitLab (Azure DevOps, Jira, Bitbucket planned)

Auth tokens: GITHUB_TOKEN / GH_TOKEN for GitHub, GITLAB_TOKEN for GitLab. Never stored by wisegit.

Issue-derived freeze signals

Freeze Score Signals

The freeze score is never stored directly — it's derived by replaying the event stream for each function. Signal categories:

Freeze score formula:

freeze_score = base_score x (1 - obsolescence_penalty)

Protection signals produce the base score (weighted average of present signal categories, not additive sum); obsolescence signals produce the penalty. Obsolescence weights are adaptive — calibrated per-repository using Shannon entropy and Bayesian feedback. Falls back to hardcoded defaults when < 20 functions have signals. Age signal uses Weibull survival model [18] (k=0.7, lambda=2.4y); contributor expertise uses DOE model [19] (4 variables: contribution share, recency, duration, frequency).

Academic grounding: 24 published papers. See REFERENCE.md for full citations.

Cross-Repo Validation

Tested on 3 real-world open-source codebases:

Flask's core APIs (__init__, run, wsgi_app, url_for) c