What Is PRAXIS?
PRAXIS is a local Truth Kernel for agentic coding tools. It verifies coding-agent outputs using human-approved acceptance criteria, local evidence, deterministic gates, and repair packets.
PRAXIS is not a coding agent. It does not write code. It does not compete with Claude Code, OpenCode, or any coding agent. It sits above them and answers one question: did the agent actually do what it claimed?
The Problem
AI coding agents are powerful but unreliable completion reporters.
| Problem | Symptom |
|---|---|
| False Done | Agent says "done" but the diff is empty |
| Echo Chamber | Agent writes the acceptance criteria AND passes them |
| Missing Evidence | Agent claims tests passed but never ran them |
| Self-Reported Truth | Agent's own status messages treated as completion |
| Scattered Verification | Evidence spread across messages, files, and terminal output |
PRAXIS solves these by being an independent verification authority. It does not trust agent claims. It checks evidence.
The Three Laws
LAW 1 — COMPLETION AUTHORITY
Agent says done ≠ done.
Truth Kernel FinalGate PASS = done.
Nothing else counts.
LAW 2 — WRITE AUTHORITY
No worker writes to shared integration files.
The Deterministic Assembler is the only shared writer.
(Future scope — single-session only in v0.1.)
LAW 3 — VERIFICATION AUTHORITY
FinalGate criteria come from human-authored TaskSpec only.
An agent cannot define or verify its own completion criteria.
The Three Gates
| Gate | Question | Detects |
|---|---|---|
| EvidenceGate | Does evidence exist? | Empty diff, missing command logs, missing test output |
| ExecGate | Did commands/tests actually run? | Zero tests ran, commands not executed, test failures |
| FinalGate | Do results meet human criteria? | Criteria not met, task not human-approved, agent claims vs evidence |
Current Status
- Identity: Local Truth Kernel for agentic coding tools
- Design progress: ~45% (D0-D1 complete)
- Implementation progress: 0% (not authorized yet)
- Primary interface: Claude Code plugin + praxis CLI
- License: MIT
What PRAXIS Is Not
- ❌ A coding agent (does not write code, does not run its own agent loop)
- ❌ A Claude Code clone or competitor
- ❌ An OpenCode/MiMo clone
- ❌ "Only a Claude Code plugin" — the kernel is independent; the plugin is a bridge
- ❌ A desktop-first multi-agent orchestrator (not in v0.1)
- ❌ A server/runtime platform (not in v0.1)
Roadmap
| Stage | Name | Status |
|---|---|---|
| D0 | Pivot Decision Lock | ✅ Complete |
| D1 | Plugin-First Design Pack | ✅ Complete |
| D2 | Truth Kernel Proof Design | 🔜 Next |
| D3 | Claude Code Plugin Spike Spec | Future |
| D4 | Final Design Lock Audit | Future |
| I0–I4 | Implementation | ⛔ Not authorized |
Project Layout
praxis/
├── README.md
├── docs/
│ ├── decisions.md
│ ├── adr/
│ ├── product-scope.md
│ ├── phase-map.md
│ └── ...
├── artifacts/
└── pi/ # Old Pi monorepo (reference only)