Gentleman-Programming/gentle-ai

Gentle-AI is a Go binary that configures existing AI coding agents — Claude Code, Cursor, Codex, OpenCode, Pi and others — with a shared workflow layer instead of replacing the agent or model. It targets developers frustrated by agents that forget decisions between sessions and produce work they cannot verify. Its two most important additions are Engram, a project memory the agent consults before asking you again, and Organic Driven Development (ODD), which sizes work by understanding, risk and resumability rather than file count README.

How it changes a session

Engram stores what the agent learns so a later session or a post-compaction context can recall past decisions instead of re-reading the repository or re-interviewing you Engram docs. ODD keeps small, understood changes inline with no task document, while substantial authorized work gets one recoverable feature document that holds your spec verbatim; handoffs point at that document rather than paraphrasing it ODD docs.

Delegation is also gated. A subagent is used only when the orchestrator can name a reason: exploration beyond roughly 3 calls and 10k tokens, parallel independent writers, a context backstop, or an independent verifier for high-risk changes. With no named reason, the parent works inline routing rules.

Review and determinism

Receipt-Driven Development (RDD) is on by default and opt-out via gentle-ai review mode disable. The candidate change is frozen before review, so evidence belongs to that exact version; review depth comes from a risk assessment, ranging from a structural readback to the full 4R lenses, with at most one bounded correction. Reviews are informational — commit, push and release remain your decision review docs. The gentle-ai binary, reading change state from disk, owns the next valid transition rather than a model voting on it README.

The README reports benchmark results from the separate Gentle Shell project: a 3-line bug went from 335 s / 56 turns / $0.339 to 77 s / 10 turns / $0.105, and a 73-line spec with 26 hidden checks passed after previously failing. These are single runs (n=1) on a local benchmark with API-equivalent cost in a different package, not end-to-end guarantees for every agent integration README.

Requirements and limits

Install on macOS, Linux or Windows; Homebrew and curl scripts exist, while the Windows PowerShell path is a source install needing Go 1.25.10+ and does not update PATH automatically README. Gentle-AI deliberately never installs an agent — it configures one you already have, and features differ per agent because each integration uses native capabilities README. Configs are backed up before every write, a security deny-list blocks paths such as ~/.ssh and .env, and gentle-ai doctor gives a read-only health report. Gentle Shell is a separate Pi package whose runtime behavior is not established by installing this binary README. The code is MIT-licensed; the Gentle AI and Engram names and logos are trademarks, and the licence does not permit implying endorsement README.