lao 0.1.5

Local preflight for cloud AI coding agents

LAO — Local AI Orchestrator

Stop paying cloud agents to grep your repo.
Local compute first. Cloud reasoning last.

---

The Problem

Every time you run Claude Code, Cursor, or Codex on a task, the agent either:

• Gets the entire repo dumped into context (expensive, noisy)
• Has to guess which files are relevant (misses things, wastes tokens)
• Sees your .env files and secrets (dangerous)

LAO solves this locally — before anything reaches the cloud.

---

What LAO Does

lao prep "fix failing auth test" --repo ./my-repo

In under 2 seconds, locally:

1. Scans the repo with ripgrep
2. Scores files by relevance (keyword match, import graph, path match, stacktrace parsing)
3. Blocks secrets and .env files from leaving your machine
4. Builds a token-optimized evidence pack for the agent

Result: Claude gets 1.7k tokens of the right context instead of 72k tokens of everything.

---

Quick Start

# Install
pip install lao
# or: pipx install lao

# Initialize in your repo
cd your-repo
lao init

# Prepare context for a task
lao prep "fix the failing auth session test"

# Output goes to .lao/context-pack.md
# Copy to Claude: cat .lao/context-pack.md | pbcopy

# See how much you save
lao bench

lao init auto-installs ripgrep and gitleaks if missing — just confirm when prompted.

---

Example Output

lao prep "fix failing auth session test" --output compact

TASK: fix failing auth session test
TYPE:bug_diagnosis DIFFICULTY:cloud_recommended CONFIDENCE:70%
REPO:nextjs/typescript FILES:61 SELECTED:6
TOKENS:1.7k (raw:72k saved:98%)
BLOCKED:.env.local

=== EVIDENCE ===
--- src/auth/session.ts [keyword:15 path:12 error:10] ---
# L38-82
```ts
export function getSession(sessionId: string): Session | null {
  const session = sessionStore.get(sessionId)
  if (!session) return null
  
  const now = Math.floor(Date.now() / 1000)
  // BUG: should be <= not <
  if (session.expiresAt < now) {
    sessionStore.delete(sessionId)
    return null
  }
  return session
}

--- tests/auth/session.test.ts [keyword:15 path:12 test:4] ---

L28-55

it('returns null for expired session', () => {
  vi.setSystemTime(new Date(Date.now() + 3601 * 1000))
  const result = getSession(session.sessionId)
  expect(result).toBeNull()  // FAILS
})

SEND_TO:claude-sonnet

---

## Commands

| Command | Description |
|---------|-------------|
| `lao init` | Initialize lao in current repo |
| `lao prep "task"` | Prepare evidence pack for a task |
| `lao map` | Generate `.ai/` repo infrastructure |
| `lao mcp start` | Start MCP server (for Claude Code/Cursor) |
| `lao mcp config` | Show MCP configuration snippets |
| `lao check` | Check tool availability |
| `lao bench` | Benchmark token savings in dollar terms |
| `lao index` | Build persistent file index (faster prep) |
| `lao stats` | Show last run statistics |
| `lao show` | Show last context pack |

### `lao prep` options

```bash
lao prep "task" [options]

  --repo, -r PATH       Repo path (default: current directory)
  --context, -c TEXT    Extra context (paste error output, stacktrace)
  --output, -o FORMAT   markdown (default) | compact | json
  --no-cloud            Disable cloud routing recommendation
  --verbose, -v         Debug logging

Output formats

# Markdown file (default) — for manual copy/paste
lao prep "fix auth bug"

# Compact — token-efficient, for piping to agents
lao prep "fix auth bug" --output compact

# JSON — structured, for tooling
lao prep "fix auth bug" --output json

---

MCP Integration

LAO exposes three MCP tools so Claude Code, Cursor, and any MCP-compatible agent can call it directly.

Claude Code Setup

Add to .claude/settings.json in your repo:

{
  "mcpServers": {
    "lao": {
      "command": "uv",
      "args": ["run", "lao", "mcp", "start"],
      "cwd": "/path/to/lao"
    }
  }
}

Or run lao mcp config to get the exact snippet for your environment.

Available MCP Tools

Tool	Description
lao_prep(task, repo_path?, extra_context?, output_format?)	Prepare evidence pack
lao_map(repo_path?, force?)	Generate .ai/ infrastructure
lao_status(repo_path?)	Show last run stats

Cursor Setup

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "lao": {
      "command": "uv",
      "args": ["run", "lao", "mcp", "start"]
    }
  }
}

Codex / Other Agents

# Pipe compact output directly
lao prep "fix auth bug" --output compact | your-agent-cli

---

.ai/ Repo Infrastructure

lao map generates a .ai/ directory that makes your repo permanently agent-ready.

lao map

Creates:

.ai/
  manifest.yaml      # Stack, framework, entry points
  repo-map.yaml      # Module/feature → files mapping
  test-map.yaml      # Source file → test file mapping
  risk-zones.yaml    # Sensitive paths and why
  agent-rules.md     # Rules for AI agents (editable)

Commit .ai/ to git. Every agent that works in the repo benefits immediately.

Example manifest.yaml

repo:
  name: my-app
  framework: nextjs
  primary_language: typescript
  test_runner: vitest
  package_manager: pnpm
entry_points:
  - path: src/app/layout.tsx
    type: application
  - path: src/cli/index.ts
    type: cli
ai_hints:
  test_command: npx vitest run
  install_command: pnpm install

Example agent-rules.md

# Agent Rules for my-app

## Always
- Run tests before declaring a fix complete
- Keep changes minimal — do not refactor unless asked
- Check .ai/risk-zones.yaml before reading sensitive files

## Never
- Send .env* files or private keys to cloud context
- Add new dependencies without asking

---

Security

LAO is local-first and privacy-preserving by design:

• .env* files are always blocked from leaving your machine
• Private keys (*.pem, *.key, id_rsa) are always blocked
• Secret scanning runs locally before any file is included
• risk-zones.yaml documents exactly which paths are sensitive and why
• Nothing is sent anywhere — LAO only writes to .lao/ and .ai/ locally

Default blocked patterns

.env, .env.*, .env.local, .env.production
**/*.pem, **/*.key, **/*.p12
**/id_rsa, **/id_ed25519
**/credentials*, **/*secret*, **/wallet*

Customize in .lao/config.yaml:

security:
  cloud_deny:
    - ".env*"
    - "**/*.pem"
    - "src/internal/pricing/**"   # add your own

---

Configuration

.lao/config.yaml is created by lao init. All fields are optional.

# Token budget for context pack
limits:
  max_files_in_context: 8        # Max files selected per run
  max_file_tokens: 6000          # Max tokens per file
  max_total_context_tokens: 30000

# Cloud routing
cloud:
  enabled: true
  preferred_model: claude-sonnet

# Local models (optional)
local:
  embeddings_enabled: false      # Enable with: lao index --with-embeddings
  local_llm_enabled: false

---

How It Works

lao prep "fix failing auth test"
    │
    ├── 1. Classify task          → bug_diagnosis, cloud_recommended
    ├── 2. ripgrep search         → 251 matches across 61 files
    ├── 3. Import graph           → session.ts imported by test file (+3.0)
    ├── 4. Stacktrace parsing     → error paths extracted from --context
    ├── 5. Score files            → path_match + keyword_match + entry_point
    ├── 6. Security policy        → .env.local BLOCKED
    ├── 7. Secret scan            → 0 secrets in selected files
    ├── 8. Build context pack     → 6 files, 1.7k tokens
    └── 9. Write outputs          → .lao/context-pack.md + report.json

Scoring signals

Signal	Weight	Source
keyword_match	3.0/match, cap 15	ripgrep hits in file
path_match	6.0/keyword	keyword stem in file path
error_path	10.0	file mentioned in stacktrace
entry_point	8.0 / 4.0	__init__.py, main.py etc.
import_reference	3.0/ref	imported by high-scoring file
repo_map_boost	5.0	.ai/repo-map.yaml match
doc_penalty	-8.0	README, CHANGELOG etc.
generated_penalty	-8.0	auto-generated files

---

Eval Results

Measured on 20 tasks across two repos (Next.js auth fixture + KMF Python library):

Metric	Baseline	LAO v0.1	Target
Context Recall	60.4%	78.3%	≥70%
Safety (no secrets)	100%	100%	100%
Tasks ≥75% recall	40%	50%	—

---

Roadmap

Phase	Focus	Status
v0.1	lao prep, ripgrep, security, MCP, index	✅ Done
v0.2	VS Code Extension, lao index --watch	🔜 Next
v0.3	GitHub Action, CI/CD integration	📋 Planned
v0.4	Team dashboard, org policies	📋 Planned
v1.0	.ai/ community standard, public launch	📋 Planned

---

Requirements

• Python 3.11+
• uv (recommended) or pip

Optional (improves quality):

• ripgrep — faster search
• gitleaks — better secret detection
• fastmcp — MCP server (pip install fastmcp)

---

License

MIT