LeanCode Engineer is an open-source Agent Skill for disciplined software engineering. Its tiny entrypoint routes an agent to focused references and domain modules only when they change the next decision.
The project optimizes in this order:
Correctness → Security → Verification → Minimal Change → Context Efficiency → Brevity
Token savings are a constraint, never a reason to skip evidence or weaken safeguards.
Large “do everything” prompts consume context on every task. LeanCode Engineer keeps the shared operating contract in SKILL.md, then uses progressive disclosure for debugging, security, architecture, verification, and optional domain modules.
- Small, durable core that remains useful across stacks.
- Specialist references loaded only for relevant decisions.
- Risk-based verification with honest reporting of gaps.
- Single-agent default; delegation and Agent Teams are selective and optional.
- No service, API key, telemetry, or runtime package required.
From this repository:
claude --plugin-dir .Then invoke /lean-code-engineer:lean-code-engineer, or describe a coding task and allow automatic activation.
./scripts/install.shThe installer defaults to ~/.claude/skills/lean-code-engineer, refuses to overwrite an existing install, and preserves the old directory as a timestamped backup when --force is explicitly used.
./scripts/install.sh --dry-run
./scripts/install.sh --target codex
./scripts/install.sh --dest /absolute/custom/skills/lean-code-engineerSee installation and compatibility for supported and experimental paths.
flowchart LR
U[Software task] --> C[SKILL.md core]
C --> D{Decision needs detail?}
D -->|No| E[Focused change]
D -->|Failure| R1[debugging.md]
D -->|Risk boundary| R2[security.md]
D -->|Confidence| R3[verification.md]
D -->|Cross-system| R4[architecture.md]
D -->|Domain-specific| M[one module]
R1 --> E
R2 --> E
R3 --> E
R4 --> E
M --> E
E --> V[Evidence-backed completion]
The complete design is documented in architecture.
SKILL.md Lean entrypoint and router
references/ Conditional engineering playbooks
modules/ Optional domain guidance
scripts/ Installer, validator, router, benchmark
tests/ Standard-library behavioral checks
examples/ Realistic routing examples
docs/ Architecture, compatibility, security, release docs
agents/openai.yaml Codex-compatible UI metadata
.claude-plugin/ Claude Code plugin manifest
.github/ CI and community health files
assets/ Logo, icon, and GitHub artwork
Python 3.9+ is needed only for contributor tooling:
./scripts/validate.sh
./scripts/test.sh
./scripts/benchmark.sh --checkThe benchmark reports estimated context cost and route-specific savings. It is a structural proxy, not a claim about model billing or answer quality; see benchmark methodology.
v0.1.0 is the first locally validated foundation. Claude Code is the primary host. Codex metadata and installation are included as an experimental adapter; Cursor and OpenCode remain roadmap targets until their adapters and compatibility tests are implemented.
Start with CONTRIBUTING.md. Report suspected vulnerabilities through the private process in SECURITY.md, not a public issue.
Before publication, apply the reviewed defaults in GitHub repository setup.
MIT © 2026 ALJANEF.