coding-agent-setup

A portable, self-installing kit that sets up any complex codebase for token-efficient, high-velocity development by coding agents.

Point a coding agent (Claude Code, GitHub Copilot, Cursor, or any MCP-capable host) at SETUP.md and it installs and wires up the systems that make large, unfamiliar codebases cheap and fast to work on:

• Graphify — a local tree-sitter code knowledge graph (with an optional MCP server). One graphify query replaces the grep→read→re-read loop that burns the most tokens on big repos.
• RTK (Rust Token Killer) — filters noisy shell output (tests, lint, git, docker) by 60–90% before it reaches the model's context.
• Review + verify discipline — adversarial review and end-to-end verification that catch the plausible-but-wrong changes shallow checks miss.
• Project instructions + DOX contract — encode the repo's conventions once, then a hierarchy of per-directory AGENTS.md contracts that tells the agent what each directory owns and where a new feature belongs — so work lands in the right architectural layer instead of wherever a file was already open.
• Planning + multi-agent orchestration — a supervisor-worker loop that takes a goal to merged PRs: decompose → wireframe → spec → audit, then dispatch → build → verify → review. The plan lives in the repo as a dispatch board, the owner gates live in a hook that enforces them deterministically, and the loop stops at "PR opened" so merge and deploy stay human.

Quickstart

For an agent: read SETUP.md and execute it.

For a human:

# from your target repo's root
bash /path/to/coding-agent-setup/install.sh       # macOS / Linux / Git-Bash
pwsh /path/to/coding-agent-setup/install.ps1       # Windows PowerShell

The installers are idempotent and non-destructive — they merge into existing config, never overwrite, and no-op anything already installed.

What's here

File	Purpose
SETUP.md	Agent-facing step-by-step auto-install playbook
PLAYBOOK.md	The methodology — why these tools, the principles behind them, hard-won lessons
ORCHESTRATION.md	Multi-agent planning + dispatch — the two loops, the seven roles, where wireframes and evals fit, what stays human, and what it costs
HARNESS-PRACTICES.md	Harder harness practices (tool risk annotations, fail-closed permissions, trajectory evals, prompt-cache discipline, CI gates) for keeping a codebase agent-developable over time
install.sh / install.ps1	Runnable installers (Unix / Windows)
templates/	Ready-to-copy config for Claude Code + VS Code Copilot
CONTRIBUTING.md	How we push new learnings back into this kit

Host support

Works across Claude Code, VS Code GitHub Copilot (Chat + CLI), Cursor, and other MCP hosts. Each capability degrades gracefully where a host lacks it — see the support matrix in SETUP.md §8. Two platform gaps worth knowing: VS Code Copilot Chat cannot transparently rewrite shell commands from a hook, so RTK there works via an instruction nudge instead of a silent rewrite; and hosts without subagents run the orchestration loop as sequential roles in fresh sessions rather than as parallel agents.

Origin

Distilled from real work on a large multi-agent platform running two agent runtimes side by side — the kind of codebase where naive exploration is ruinously expensive. Every principle in PLAYBOOK.md was learned by hitting the problem it solves. This kit is a living document: as we learn things that make development faster, better, or cheaper, they land here.

License

MIT — see LICENSE.