context-map

name	context-map
description	Generates a compressed project context map to avoid expensive Read/Grep calls. Use at session start or before implementing features in an unfamiliar codebase.
category	optimization
tags	["context","scanning","token-conservation","exploration","import-graph","routes"]
tools	[]
complexity	low
model_hint	fast
estimated_tokens	300

Context Map

Generate a compressed context map for the current project. The map pre-compiles structural knowledge that AI assistants would otherwise discover through expensive Read/Grep calls, saving thousands of tokens per session.

When To Use

• At the start of a session to understand project layout
• Before implementing features to identify entry points
• When exploring an unfamiliar codebase
• To reduce token waste from Read calls
• To identify hot files (high blast radius) before changes

When NOT To Use

• Context is already over budget mid-session (use conserve:clear-context)
• Auditing the codebase for bloat (use conserve:bloat-detector)

What It Detects

Category	Description
Structure	Directory layout with file counts and languages
Dependencies	Multi-ecosystem: Python, Node, Rust, Go, Java
Frameworks	Framework detection from dependency analysis
Entry Points	main.py, index.ts, CLI scripts, etc.
Import Graph	File-to-file import relationships
Hot Files	Files imported by 3+ others (high blast radius)
Routes	FastAPI, Flask, Express, Hono API endpoints
Env Vars	Environment variable references with defaults
Middleware	Auth, CORS, rate-limit, logging patterns
Models/Schemas	SQLAlchemy, Django, Pydantic, Prisma definitions
Token Savings	Estimated tokens saved vs manual exploration

Procedure

1. Run the scanner on the project root:

PYTHONPATH="$(find . -path '*/conserve/scripts' -type d \
  -print -quit 2>/dev/null || \
  echo 'plugins/conserve/scripts')" \
  python3 -m context_scanner .

2. Present the output to the user as the project overview.

3. Use the context map to guide subsequent file reads. Prioritize hot files and entry points first.

Options

Output

• --format json for structured output
• --max-tokens N to adjust output size (default: 5000)
• --output FILE to save to a file

Modes

• --blast FILE to show blast radius for a specific file
• --section NAME to output a single section (routes, deps, env, hot-files, models, structure, middleware, frameworks)
• --wiki-only to generate wiki articles without stdout

Opt-out

• --no-cache to force a fresh scan
• --no-wiki to skip wiki article generation

Wiki Articles

The scanner generates per-topic knowledge articles in .codesight/ for selective context loading:

python3 scanner.py .
# Creates .codesight/INDEX.md, auth.md, database.md, etc.

Load only what you need per session instead of the full map:

python3 scanner.py --section routes .
# ~200 tokens vs ~5,000 for the full map

Example Output

# Context Map: myproject
Files: 127

## Structure
  src                  42 files (Python)
  tests                18 files (Python)
  docs                  5 files (Markdown)

## Dependencies (Python)
Package manager: uv
  - fastapi 0.104.0
  - pydantic 2.5.0
  - sqlalchemy 2.0.0
  ...12 more

## Frameworks Detected
  - FastAPI
  - SQLAlchemy
  - Pytest

## Routes
  GET    /users          (src/routes/users.py)
  POST   /users          (src/routes/users.py)
  GET    /users/{id}     (src/routes/users.py)

## Hot Files (high blast radius)
  - src/models/base.py (12 importers)
  - src/utils/auth.py (8 importers)

## Environment Variables
  - DATABASE_URL (required)
  - SECRET_KEY (has default)

## Token Savings: ~12,600 tokens saved
  Routes: ~1,200
  Hot files: ~300
  Env vars: ~200
  File scanning: ~10,200

Exit Criteria

• Scanner produces output covering at minimum: file count, directory structure, detected frameworks, and hot files (imported by 3+ others); output appears in the session before any feature implementation reads begin
• "Token Savings" line is present in the output with a numeric estimate (e.g., ~12,600 tokens saved)
• If --blast FILE is used, blast-radius output names the specific file and lists its importers by count
• Context map guides subsequent reads: hot files and entry points are consulted before any other file read in the session