gentle-ai

Gentle-AI is a deterministic engineering environment layered on top of AI coding agents you already have. Rather than shipping its own agent, it configures existing ones — the README lists 16 integrations including Pi, OpenCode, Claude Code, Codex, Cursor, VS Code Copilot, Gemini CLI, Kilo Code, Kimi Code, Kiro IDE, Qwen Code, Hermes, Antigravity, Windsurf, OpenClaw and Trae. Each integration uses that agent's native capabilities, so features such as delegation and review can differ between agents. Core ideas described in the README: - Engram: persistent project memory. The agent records decisions as it works and consults that memory before asking the user again, so context accumulates across sessions and context compaction instead of resetting. - ODD (Organic Driven Development): keeps small, understood changes lightweight while giving substantial authorized work a single recoverable feature document, so work can resume without rebuilding a plan. - Strict TDD: when enabled, a failing behavior test is captured before implementation, then made to pass, then refactored with tests green. When disabled, applicable functional checks still run. - RDD (Receipt-Driven Development): enabled by default and opt-out via `gentle-ai review mode disable`. The candidate change is frozen before review, review depth is derived from that frozen candidate (passive structural readback, one focused lens, or a four-lens Risk/Resilience/Readability/Reliability review), at most one bounded correction is allowed, and results are informational — commit, push and release remain the user's decision. - Deterministic transitions: a `gentle-ai` binary reads change state from files on disk and returns the next valid transition, with four public states (Working, Checking, Ready, Needs your decision) rather than model judgment. - Gentle Shell: a separate Pi integration package with its own runtime, agents and interface; installing the binary does not by itself establish Pi behavior parity. Additional bundled components listed: a skills library loaded when tasks match, optional Context7 MCP for live framework/library documentation, a read-only CodeGraph symbol graph, a security deny-list blocking access to `~/.ssh`, `.env` and credential files, config snapshots before every write, a read-only `gentle-ai doctor` health report, optional personas (Gentleman), themes, and model assignment for supported agents and review roles. Installation options shown are Homebrew on macOS, a curl script for macOS/Linux, and a Go source install on Windows requiring Go 1.25.10+. After install, running `gentle-ai` prompts for agents, components and persona; the README states the tool never installs an AI agent itself and snapshots configs before writes. Documentation covers intended usage, quickstart, per-agent feature matrix, ODD routing, review architecture, Engram, components, contributing, telemetry (with an opt-out) and a codebase guide. The project is MIT licensed, with trademark notices for the Gentle AI and Engram names.