Codex
--target codexAGENTS.mdMain instructions.agents/skills/Workflow skills, loaded on demand.codex/agents/Judge, comprehension, and implementer agents, when enabled
Judge, comprehension, and implementer run as native Codex subagents.
A harness for coding agents
Wepwawet
Install instructions for planning tasks, writing acceptance criteria, and reviewing changes. Keep the project context and task records in your repository.
MIT licensed · CLI: wep
How it works
From setup to a reviewed task
Wepwawet installs workflow instructions and a shared context structure into your project. Your coding agent reads those files to plan tasks, choose a working method, and check results against explicit acceptance criteria.
Install the harness
wep initcreates shared context, task templates, and agent instructions. Existing root instruction files receive a suggested merge file.Define the task
The agent records the goal, resolves material ambiguity, selects a workflow, and writes acceptance criteria and validation commands.
Implement and validate
The agent follows the selected workflow, updates the task checklist, and runs the agreed checks. Independent review is available when enabled.
A first task, in your agent
Run wep task workflow for invoice API examples. Use SDD+EDD.
The agent creates or updates a task capsule under .wep/tasks/TASK-XXX/. Root instructions stay small; detailed workflows load when relevant.
What it gives your agent
What gets installed
Project notes and task records live alongside the code. Later sessions can read the same files, including when you switch to another supported coding agent.
Task capsules
One directory per task under
.wep/tasks/, with the goal, relevant context, workflow, acceptance criteria, and validation commands. It stays for the next session or another supported agent.Workflow selection
fast-patch,research-first,SDD,TDD,EDD, or a combination such asSDD+EDD. The capsule records which one and why.Project wiki
Architecture, conventions, testing commands, and decisions in
.wep/wiki/. Plain Markdown you can read and version.Agent adapters
Shared context lives in
.wep/. Each adapter installs instructions and skills where the selected coding agent expects them.Doctor
wep doctorchecks required files, configuration, policy, merge conflicts, and wiki integrity. The installed doctor skill lets the agent review whether instructions are useful, duplicated, or stale.Template packs
Package team conventions as config, policy, wiki pages, and target files. Apply them at init or later with
wep templates apply.
Review, comprehension, and tracing
Choose what to enable
A standard init enables the judge. Comprehension checkpoints, the implementer, and tracing are off by default. Continuation records are written separately for each task.
- Enabled by a standard init
Independent judge
A separate subagent reviews the change against the task capsule and reports APPROVED, NEEDS CHANGES, or BLOCKED. It reads without modifying product code or Git state, and does not implement its own feedback.
wep judge status - Optional · off by default
Comprehension checkpoints
After an approved change, a separate agent shows a change map and a test matrix, then asks you about the code one question at a time. It does not block merges or replace correctness review.
wep comprehension enable - Optional · off by default
Implementation agent
An implementer agent works from a task capsule and records each attempt in a local log. The CLI enforces the configured log-entry budget; your coding agent invokes the implementer.
wep impl enable - Opt-in
Session tracing
Trace hooks observe tool calls without blocking them. Records omit raw commands, arguments, file paths, and results, but retain lifecycle metadata and command digests.
wep trace reportsummarizes recorded sessions.wep trace enable - Optional file
Portable continuation records
A task can carry a
continuation.jsonthat travels with the repository. The reader is read-only: it does not resume work, run commands, or authorize anything.wep task continuation TASK-001
Works with
Codex, Claude Code, GitHub Copilot, and Pi.
All targets share the .wep/ files. Instructions and available agents vary by target. Use --target all to install every adapter, or--dry-run to preview an install.
Claude Code
--target claudeCLAUDE.mdMain instructions.claude/commands/Slash-command shortcuts.claude/skills/Workflow skills, loaded on demand.claude/agents/Judge, comprehension, and implementer agents, when enabled
Judge, comprehension, and implementer run as native Claude Code subagents.
GitHub Copilot
--target copilot.github/copilot-instructions.mdMain instructions.github/instructions/Rules scoped to file patterns.github/prompts/Reusable prompt files.github/skills/Workflow skills.github/agents/Judge, comprehension, and implementer agents, when enabled
Judge, comprehension, and implementer run as Copilot custom agents.
Pi
--target pi.pi/prompts/Prompt files.pi/skills/Workflow skills
Pi receives workflow prompts and skills. Judge, comprehension, and implementer agent files are not installed for Pi.
Safety and limits
Before you rely on it
The CLI checks files and review records. Your coding agent follows the instructions and performs the work. See the security model andjudge documentation for the exact checks and limits.
Existing root instructions
If
AGENTS.md,CLAUDE.md,.github/copilot-instructions.md, or.pi/README.mdalready exists, init writes a.wep.suggested.mdfile for human review instead of replacing it.Policy is guidance, not a sandbox
Policy files guide a cooperative agent. They do not resist a malicious or compromised agent, prompt injection, or a template pack built to weaken them. Use your coding agent's own permissions to restrict access and execution.
Review verification
Verification checks a structured review record against a task and snapshot. Pin the validation receipt digest from your own run, or use
--run-teststo run validation again. Verification cannot establish which agent or model wrote the review.No model inside the CLI
The CLI generates files that your coding agent reads. It does not call an LLM or execute an agent itself.
Get started
Install with npm, then initialize your project.
Requires Node.js 20+, npm, Git, and a supported coding agent. See installation for updates and development builds.
Install the CLI
$ npm install -g wepwawet $ wep --version $ wep --help
Initialize your project
# or: claude, copilot, pi, all $ wep init --target codex # check setup and report gaps $ wep doctor # installed targets and task status $ wep status
Open your agent and describe a task
Ask it to use the wep task workflow. The agent owns the workflow after initialization; CLI task commands are also available for scripts and headless use.
MIT licensed · CLI: wep