claude-ledger 0.2.0

Portfolio-level project tracking for Claude Code

claude-ledger

Portfolio-level project tracking for Claude Code.

Why This Exists

Claude Code is excellent at per-project memory. Each project gets its own CLAUDE.md, auto-memory files, and todos. Within a single project, context persists well.

But if you work across many projects — 10, 20, 50+ — there's a fundamental gap: Claude starts every session ignorant of what happened in your other projects, even on the same machine. There's no portfolio layer.

This is the problem I hit running 30+ projects with Claude Code as my primary development tool. The symptoms:

• Constant re-explaining. "Where were we?" became the opening line of every session. Not just for the current project — for the portfolio. Which projects are active? What did I ship yesterday? What's been neglected?
• Invisible progress. I built a session tracking system but never wired the hooks. Six days of work on a client project went completely untracked. I only discovered this during an audit.
• Lost cross-project awareness. Pausing one project without realising it blocks three others. Changing shared infrastructure without knowing what depends on it. No cascade warnings.
• Session amnesia after compaction. Mid-session context compression would wipe portfolio awareness entirely. The project I was currently working on survived, but everything else vanished.

Per-project memory is solved. Portfolio-level intelligence is not. That's what claude-ledger addresses.

How It Fits Into Claude Code's Memory System

Claude Code already has a layered memory system. claude-ledger adds the missing layer:

Layer 1: CLAUDE.md          Per-project instructions, vocabulary, protocols
Layer 2: Auto-memory        Per-project facts, preferences, feedback
Layer 3: Todos/Tasks        Per-project work tracking
─────────────────────────────────────────────────────────
Layer 4: claude-ledger      Cross-project portfolio awareness  ← NEW

Layers 1-3 answer: "What do I need to know about this project?"

Layer 4 answers: "What's happening across all my projects? What's active, what's stale, what depends on what, and what did I do yesterday?"

claude-ledger sits on top of the existing system — it doesn't replace anything. Your CLAUDE.md files, memory, and todos keep working exactly as they do now. The ledger adds the portfolio view that was missing.

How It Works

claude-ledger is a set of Claude Code hooks + a CLI that automatically:

1. Tracks activity — every file edit and git commit is recorded to a per-project ledger file, without you doing anything
2. Generates briefings — on session start, Claude gets a portfolio summary with priority tiers and staleness detection
3. Maps workstreams — groups related projects together and warns when changes might cascade

The key design principle: capture must be fully automatic. Manual triggers have already proven unreliable — if it requires a human to remember to run something, it won't happen. Hooks solve this.

What Claude Sees at Session Start

When you start a Claude Code session, the SessionStart hook generates a briefing like:

Portfolio briefing ready — 32 projects tracked, 7 P1 active, 3 stale.
Read ~/.claude/ledger/_portfolio.md and ~/.claude/ledger/_workstreams.md for full context.

Claude can then read the full briefing — grouped by priority, with staleness warnings and workstream cascade alerts — before you even type your first message.

What Happens During a Session

As you work, hooks fire silently in the background:

• Edit a file → the ledger records which project was touched
• Make a git commit → the commit message and SHA are captured to the project's activity log
• Claude responds → the session summary is stored
• Session ends → everything is finalised and the ledger repo is committed (giving you time-travel via git log)

You never interact with the ledger directly. It just accumulates.

Quick Start

pip install claude-ledger
claude-ledger init --scan-dirs ~/Code --github-user your-username
claude-ledger scan
claude-ledger bootstrap

That's it. Every Claude Code session now starts with a portfolio briefing, and activity is tracked automatically.

What Gets Created

~/.claude/ledger/
├── ledger.yaml              # Your configuration
├── _portfolio.md            # Auto-generated briefing (Claude reads this at session start)
├── _workstreams.md          # Cross-project dependency map with cascade warnings
├── .git/                    # Time-travel — git log shows portfolio state over time
├── my-project.md            # One ledger file per project
└── another-project.md       #   (YAML frontmatter + activity log)

The whole thing is plain markdown files in ~/.claude/. No database, no daemon, no external service. Git-backed for time-travel. Human-readable. Works offline.

Commands

Command	What it does
claude-ledger init	Set up ledger directory and install Claude Code hooks
claude-ledger scan	Discover projects in your configured directories
claude-ledger bootstrap	Create ledger files from scan results
claude-ledger briefing	Generate portfolio + workstream briefings
claude-ledger status	Quick portfolio summary
claude-ledger uninstall	Remove hooks (optionally delete ledger data)

Configuration

After running init, edit ~/.claude/ledger/ledger.yaml:

version: 1

# Where to scan for projects
scan_dirs:
  - ~/Code
  - ~/Projects

# GitHub username for remote repo discovery (optional)
github_user: your-username

# Directories to scan for stray project files
stray_scan_dirs:
  - ~/Downloads

# Days of inactivity before a project is "stale"
stale_days: 7

# Projects to skip during scan/bootstrap
skip_slugs: []
no_track: []

# Group related projects into workstreams (optional)
workstreams:
  backend:
    display_name: "Backend Services"
    members:
      - api-server
      - auth-service
      - worker-queue
  frontend:
    display_name: "Frontend Apps"
    members:
      - web-app
      - mobile-app

Ledger File Format

Each project gets a markdown file with YAML frontmatter:

---
name: My Project
slug: my-project
directory: /Users/you/Code/my-project
repo_url: https://github.com/you/my-project.git
status: active        # active, paused, dormant, archived, completed
priority: P1          # P1, P2, P3 (inferred from activity)
vision: A short description of what this project does
current_phase: building
last_session: '2026-03-20T15:03:39Z'
last_activity: Add user authentication
systems: []
tags: [next.js, typescript, tailwind]
workstreams: [frontend]
---

## Activity Log

### 20 March 2026
- Add user authentication (a1b2c3d)
- Fix login redirect loop (e4f5g6h)

### 18 March 2026
- Initial project scaffold (1234567)

These are just markdown files. Edit them freely — change priorities, update the vision, add notes. The hooks only ever append to the activity log and update timestamps.

How Hooks Work

claude-ledger installs five hooks into your Claude Code settings:

Hook	Event	What it does
PostToolUse (Edit/Write)	File edited	Records which project was touched
PostToolUse (Bash)	Command run	Captures git commit metadata
Stop	Response complete	Stores session summary
SessionEnd	Session closes	Finalises activity, commits ledger repo
SessionStart	Session opens	Generates portfolio briefing

All hooks have tight timeouts (2-5 seconds) and fail silently — they never block your work.

Concurrent Sessions

If you run multiple Claude Code sessions at once (different terminals, different projects), the ledger handles it safely. Each project's ledger file has its own lock file, and session state is isolated per session ID. Two sessions touching the same project won't corrupt each other's data.

Heuristic Inference

When you run bootstrap, claude-ledger infers project metadata from what it finds:

Signal	Inference
>=10 commits/30d + CLAUDE.md	P1 (high priority)
>=3 commits/30d or has .mcp.json	P2 (medium)
Everything else	P3 (low)
>=1 commit in 30 days	Active
No commits but <90 days old	Paused
<365 days old	Dormant
Older	Archived

You can override any of these in the ledger files — they're just markdown.

Design Decisions

Why plain files, not a database? Markdown files are human-readable, git-compatible, and always available. No MCP dependency, no daemon, no external service. Each file is independently lockable for thread-safe concurrent writes.

Why YAML frontmatter? Structured enough for machines to parse (priority sorting, staleness detection), readable enough for humans to edit. The format is forward-compatible — a future daemon or dashboard could ingest these files directly.

Why hooks, not manual tracking? Because manual triggers don't work. I built a session tracking script and forgot to wire the hooks for 6 days. If it requires a human to remember, it won't happen.

Why the workstreams layer? Because pausing one project can silently block three others. Cascade warnings surface cross-project risks before they cause problems.

Requirements

• Python >= 3.9
• Claude Code (for the hooks integration)
• gh CLI (optional, for GitHub repo discovery)

License

MIT