Ancienttwo/repo-harness

File-backed workflow harness for reliable Claude Code and Codex sessions.

What it solves

repo-harness provides a structured, file-backed workflow for AI agents (specifically Claude and Codex) to manage software development. It solves the problem of "chat memory loss," where agents lose context between sessions, by storing plans, task contracts, and handoffs directly in the repository files. It also addresses the risk of unattended AI autonomy by introducing a system of authorizations, budgets, and leases for long-running programs.

How it works

The system operates in two layers:

  1. Session Contract: A human-driven layer where tasks are managed through a lifecycle of Plan $\rightarrow$ Contract $\rightarrow$ Review. Hooks route agent events into typed handlers that enforce guards at the edit boundary.
  2. Authorized Programs: An automation layer for unattended work (like refactor campaigns or sprint controllers) that requires operator-minted authorizations and operates under a strict budget ledger.

It integrates with a CodeGraph index for structural queries and uses progressive context loading to keep token usage low by loading only the necessary capability blocks for the files being edited.

Who it’s for

  • Developers using AI agents like Claude or Codex for complex, multi-session coding tasks.
  • Teams wanting to implement agentic development flows with durable, reviewable evidence of work.
  • Operators who need to run unattended AI automation with hard caps on budget and duration.

Highlights

  • File-backed state: Resumes sessions from repository artifacts rather than chat history.
  • Bounded autonomy: Uses authorizations, budgets, and renewable leases to keep unattended programs accountable.
  • Token efficiency: Employs a pre-built CodeGraph and progressive context loading to minimize token overhead.
  • Structured evidence: Every task generates a contract and review card, making human verification concrete rather than based on agent claims.
  • MCP sidecar: Allows ChatGPT to plan PRDs and Sprints while Codex executes them without direct source-code write access.

Related

  • Project
    jsmastery-pro/skills

    A set of modular AI agent skills that implement a professional engineering workflow, moving project state from chat sessions into repository files for persistence and team collaboration.

  • Project
    Storybloq/storybloq

    A project memory and workflow system for AI coding agents that persists stories, plans, and handovers in Git-tracked files to maintain context across sessions.

  • Project
    aiblueprinthq/ai-blueprint

    A file-backed, spec-driven workflow framework that gives AI coding agents a structured process for planning, building, and verifying software features one at a time.

  • Project
    akitaonrails/ai-memory

    ai‑memory is a Rust‑based server that records every observation a coding‑assistant makes (prompts, tool calls, session boundaries) as plain markdown files in a git‑backed wiki. The server can be run locally, in Docker/Podman, or on a homelab box and is designed to be shared across many agents (Claude Code, Codex, Cursor, Gemini CLI, etc.), machines and team members. Hooks posted by agents are consolidated into readable wiki pages at session end; the next session receives a concise handoff so it can continue exactly where the previous one stopped. Storage is human‑readable markdown plus a SQLite index (FTS5, entity and optional vector search). By default it needs no LLM or embedding keys; optional Anthropic/OpenAI integrations add LLM‑generated summaries and semantic search. Multi‑user authentication, audit logs and per‑person attribution are built in. Install via Arch AUR, Docker wrapper, or native binary, start the server, register a client and its hooks, and the system works transparently.