Agent Toolchain

SWE Agent

**A local-first operating layer for reliable Claude Code and Codex work. It keeps repository instructions, roles, and verification evidence from drifting, without becoming a runtime, framework, or hosted service.**

What it is	Markdown conventions, installable skills, session and Git hooks, and Python scripts with no dependencies outside the standard library.
What you get	A validated task route, eleven active personas plus three compatibility definitions, declared writer scopes, structural no-write restrictions for judges, a bounded review loop, and local proof you can re-run.
How to start	Install once, then explicitly invoke methodology-management in your project. See Quickstart.
What it is not	A runtime, a package, an API, or a hosted service. See What this is not.

[Quickstart](#quickstart) · [Architecture](#architecture) · [Components](#components) · [Current state](#current-state) · [The problem](#the-problem) · [Documentation](#documentation) · [Working here](#working-in-this-repository) · [Improvements](#recent-improvements)

Architecture

flowchart LR
    S1["1. Repository route"] --> S2["2. Shared capabilities"]
    S2 --> S3["3. Harness layer"]
    S3 --> S4["4. Controlled work loop"]
    S4 --> S5["5. Local proof"]
    S5 --> S6["6. PR and audit trail"]
    S5 -- "a failed gate returns the task" --> S4
    S6 -. "lessons and session signals" .-> S1

Repository knowledge is the durable source of truth. Shared skills accelerate both harnesses; session signals are Claude-specific. Neither replaces repository knowledge.

Stage	What happens	What it buys
1. Repository route	A repository declares its contract and task routes in AGENTS.md and docs/agents/; the repository standard supplies the common taxonomy and a migrator. Entry points: AGENTS.md, docs/agents/, [validate_disclosure.py](install/skills/p