/init

description: Initialize .autocontext/ in the current project for knowledge persistence

Initialize the autocontext knowledge directory in the current project. Use AskUserQuestion for setup decisions.

Pre-initialization check

First, the script checks if `.autocontext/` already exists in the project root. If it does, you'll be informed and asked:

**Question:** This project already has autocontext initialized. Reinitialize?

**Options:**

• Yes — set up fresh configuration (existing lessons are NOT deleted)
• No — exit and use current setup

If you choose to reinitialize, existing lessons in `.autocontext/lessons.json` will be preserved but all config files will be reset.

Setup questions

Step 1: Seed from CLAUDE.md (if applicable)

If a `CLAUDE.md` file exists in the project root:

**Question:** This project has a CLAUDE.md. Seed initial lessons from it?

**Options:**

• Yes, extract and let me review each one — parses CLAUDE.md and shows each extracted lesson for approval
• Yes, extract automatically — extracts all lessons without review
• No, start fresh — begins with empty lessons

This helps bootstrap project knowledge from existing documentation.

Step 2: Cross-developer sharing

**Question:** Will other developers use Claude Code on this repo?

**Options:**

• Yes, set up cross-developer sharing — enables the union merge driver for collaborative lesson management
• No, just me — single-developer setup, simpler merge strategy

Cross-developer sharing allows multiple Claude instances to work on the same lessons.json without conflicts.

Step 3: Performance baselines

**Question:** Track test/build performance baselines?

**Options:**

• Yes — initialize baseline tracking for test and build performance
• No — skip baseline tracking

Performance baselines help detect regressions in test suite and build times over time.

Directory structure creation

The script will create the following structure in `.autocontext/`:

.autocontext/
├── config.json              # Project configuration (shared)
├── config.local.json        # Local developer identity
├── lessons.json             # Accumulated lessons (git-tracked)
├── playbook.md              # Curated lesson playbook (auto-generated)
├── .gitignore               # Ignore cache and local files
├── .gitattributes           # (if sharing enabled) Configure merge driver
├── cache/                   # Temporary files (not git-tracked)
│   ├── pending-lessons.json # (if ask_before_persist) Awaiting approval
│   └── curated-pending.json # From /autocontext:review sessions
└── archive/                 # Tombstoned lessons (git-tracked)
    └── superseded.json      # Deleted lessons for reference

Configuration templates

The script copies template files from `${CLAUDE_PLUGIN_ROOT}/templates/`:

• `config.json` — project-level settings
• `config.local.json` — developer identity and preferences
• `lessons.json` — empty initial lessons
• `.gitignore` — excludes cache and local files
• `.gitattributes` — (if sharing) configures union merge driver

The `project_name` in config.json will be set to the git repository name or directory basename if not in a git repo.

Global config inheritance

After copying the config template, merge values from `~/.claude/autocontext.json` (if it exists) into the project's `config.json`. Global settings provide defaults; per-project config can override them later.

Mapping from global to project config:

• `lesson_loading.max_lessons` → `max_session_lessons`
• `lesson_loading.min_confidence` → `confidence_threshold`
• `staleness.days` → `staleness_days`
• `performance_baselines.enabled` → `performance_baselines`
• `performance_baselines.baseline_commands` → `baseline_commands` (only if user chose custom commands)
• `pretooluse_injection.mode` → `pretooluse_hook`
• `lesson_persistence.strategy` → `persistence_mode` (`auto_persist_with_review` → `auto_curated`, `ask_before_persist` → `ask_before_persist`, `auto_persist_all` → `auto_all`)
• `correction_sensitivity` → `correction_sensitivity`
• `playbook.generation` → `playbook_generation`
• `multi_machine` → `multi_machine`
• `test_quality_rules` → `builtin_rules`
• `identity` → written to `config.local.json`

Cross-developer merge driver

If you selected cross-developer sharing, the script will:

1. Copy `.gitattributes` template 2. Install the union merge driver via git config:

   git config merge.autocontext-union.name "Autocontext lessons.json union merge"
   git config merge.autocontext-union.driver "python3 ${CLAUDE_PLUGIN_ROOT}/scripts/merge-driver.py %O %A %B"

This allows `lessons.json` to be merged non-destructively when multiple developers commit lessons in parallel.

CLAUDE.md seeding (if enabled)

If you selected CLAUDE.md seeding, the script runs:

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/seed-from-claude-md.py CLAUDE.md .autocontext/lessons.json

This parses the CLAUDE.md file and extracts structured lessons. If you chose "extract and review," each lesson will be presented for approval before being added.

Playbook generation

After all setup is complete, the script generates an initial playbook:

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/generate-playbook.py .autocontext/lessons.json .autocontext/playbook.md

The playbook is a curated, readable summary of your project's lessons. It's automatically regenerated after each `/autocontext:review` session.

Completion summary

Once initialization is complete, you'll see a summary of what was created:

• Number of lessons (if seeded from CLAUDE.md)
• Configuration files created
• Merge driver status (if applicable)
• Path to playbook.md for reading
• Next steps: run `/autocontext:review` to curate lessons, or just start coding

You can now use the autocontext plugin in this project. Lessons will automatically accumulate during sessions.