pact

PACT — Programmatic Agent Constraint Toolkit

A framework for governing AI coding agents through infrastructure instead of instructions.

The common assumption is that Claude's problem is memory. It forgets facts, decisions, and other nuances between sessions, so the go-to "solution" is to make it remember more by applying bandaids like session logs or plans in endless markdown files, classic databases, spawning dozens of subagents to argue it out with little context, and attempting marathon sessions to prove you've avoided compaction; that last one is actually nightmare fuel, not an accomplishment. I'll take "context rot" for $500, Alex.

The uncomfortable truth is that Claude makes bad decisions even when it remembers everything. It will edit a file it hasn't read. Fix one layer and break three downstream. Guess at a package API instead of checking docs. Declare "done" while half the work is missing. Full context doesn't fix any of these. They're reasoning failures, not recall failures.

Memory plugins ask: "How do we help Claude remember?" PACT asks: "How do we help Claude think?"

Who Is PACT For?

Solo developers and small teams. PACT was built by a solo developer for solo developers, then designed to scale up. The smaller your team, the more you need it. There's no code reviewer to notice the stale cache, no teammate to check the dependency chain, no QA to catch the UI regression. Large teams have humans for that. You have PACT.

Small project? Good. PACT scales down with you — the Seed tier gives you reasoning quality without structural overhead, and delegation lets small projects inherit knowledge from a larger one or a shared stack-level instance. PACT's subsystems activate on the patterns they detect, not on project size.

What Is PACT?

PACT is a modular governance framework for AI coding agents (Claude Code, Cursor, Copilot Workspace, etc.). Use all of it or just the parts that fill gaps in your existing setup — if you already have a memory layer, a task manager, or a workflow orchestrator, PACT detects that and only scaffolds what's missing.

Every recommendation in Anthropic's Claude Prompting Best Practices guide is addressed by PACT — clear instructions, context with rationale, structured XML, role-setting, examples, anti-hallucination guards, investigation-before-answering mandates, state management, subagent orchestration, autonomy/safety balancing, and anti-overengineering. PACT was built from production failures before these were published as best practices.

PACT has twelve features. Take what you need: