chensi 2.0.0

Chensi — Advisory Safety Layer for AI agents: auto-generates repo context, command policies, and Claude Code deny-lists

Chensi

Advisory Safety Layer for AI agents. Run one command. Your AI agent gets an auto-generated repo map, typed command policies, and a .claude/settings.json deny-list — without writing a single line of agent config.

Chensi is advisory, not a full enforcement layer yet. It generates structured context that Claude Code, Codex CLI, and other agents read at startup. Hard enforcement via preToolUse hooks is planned for Phase 5. See roadmap.

Meet Agenty 🤖 — Chensi's mascot, your repo's safety advisor.

---

Status

Branch studio — активная разработка Chensi v2.0.

Версия	Статус
v0.1.0	Released — markdown-only framework
v2.0	In development — Python CLI, Phase 2 in progress (settings.json enforcement + PyPI release)

Публичный релиз v2.0 будет на ветке main.

---

What Chensi does

chensi activate my-repo/

Result — your AI agent opens the repo and sees:

# Project Context (generated by Chensi v2.0)
Stack: python + langchain + langgraph + pandas + redis

## Safe commands
- pytest
- ruff check
- mypy src/

## Off-limits
- .env*  (secrets)
- *.key  (secrets)
- migrations/  (requires approval)

Plus a .chensi-os/adapters/claude/settings.json deny-list that Claude Code reads to physically block destructive commands:

{
  "permissions": {
    "deny": [
      "Bash(rm -rf *)",
      "Bash(git push --force*)",
      "Bash(git reset --hard*)",
      "Bash(chmod 777*)"
    ]
  }
}

No manual config. No hand-written CLAUDE.md.

---

Quick start (v2.0 — dev build)

# Install
pip install -e .

# Activate repo
chensi activate path/to/your-repo

# Check results
cat path/to/your-repo/.chensi-os/adapters/CLAUDE.md

# Or with explicit output dir
chensi activate path/to/your-repo --output-dir /tmp/test --force

# Validate
chensi validate path/to/your-repo

---

What gets generated

.chensi-os/
├── adapters/
│   ├── CLAUDE.md                    ← Claude Code reads this
│   ├── AGENTS.md                    ← Codex CLI reads this
│   └── claude/
│       └── settings.json            ← Claude Code deny-list (copy to .claude/)
├── project/
│   ├── architecture-map.md
│   └── boundaries.md
├── verification/
│   ├── commands.lock.yaml           ← typed safe/ask/deny rules
│   └── commands.md
├── repo-profile.yaml                ← detected languages, frameworks, test runners
└── boundaries.policy.yaml           ← off-limits paths

---

Advisory vs Enforcement

Layer	Phase	What it does
Advisory (CLAUDE.md / AGENTS.md)	Phase 1 ✅	Agent reads context at startup — soft guidance
Deny-list (settings.json)	Phase 2 ✅	Claude Code physically blocks listed commands
Hook interception (preToolUse)	Phase 5 🔜	Runtime enforcement for all tool calls

Phase 5 notice: Full hook-based interception is planned for Phase 5. Until then, settings.json provides hard blocking for Claude Code; advisory files handle all other agents.

---

Repository structure

.agent-os/          ← studio operational layer (Director, PM, PO, Engineering, QA)
template/           ← distribution template for users
  .chensi-os/       ← what gets copied into the user's repo
  .adapters/        ← Claude, Codex adapters
src/chensi/         ← Python CLI (Scanner → Analyzer → PolicyCompiler → ChensiLayer)
sandbox/            ← test repos for QA
  targets/trading-agents/
.github/workflows/  ← CI (pytest + ruff, Python 3.11/3.12)
docs/               ← design document + business description

---

Platform support

Platform	Status
Claude Code	Supported (CLAUDE.md + settings.json)
OpenAI Codex CLI	Supported (AGENTS.md)
Amazon Q Developer	Planned
Cursor	Planned

---

Roadmap

Phase	Description	Status
Phase 1	chensi activate core pipeline	✅ Done
Phase 2	settings.json deny-list + PyPI v2.0.0 release	🔄 In progress
Phase 3	--policy-profile org flag	Planned
Phase 4	Windows support + test suite hardening	Planned
Phase 5	Full hook-based interception (preToolUse)	Planned
Phase 6	Enterprise: OCI, Helm, SBOM, Artifactory	Planned

---

License

MIT — see LICENSE