Edictum

This is an outline of a project generated by AI based on my concept. I plan to refine it further, so please use it at your own risk and with caution at this stage.

Global install warning

The default installer writes to your global ~/.claude directory and can affect Claude Code behavior for every project on this machine. Use --dry-run first, or use --project-only <path> to install into one project's .claude/ directory instead of changing global state.

Token-efficient delegation for Claude Code. Edictum turns your Claude Code main session into a commander that plans and judges, and pushes the actual coding down to OpenAI Codex (which runs on your ChatGPT subscription and costs zero Claude tokens) and to cheap sonnet subagents. Frontier models (Fable 5, Opus 4.8) stay expensive — Edictum keeps their tokens on judgment, not on diffs and fix loops.

The problem it solves

Running a frontier model as the everyday driver burns tokens fast: it reads source, writes code, re-reads diffs, and loops on fixes — all at frontier rates. Edictum restructures that work:

• Commander (main session, any tier) decides what to build and writes a cold-executable task spec. It does not read source in bulk or write the code.
• Codex executes the spec end-to-end (branch → commit → push → draft PR → CI), on your ChatGPT plan's quota — not Claude's.
• Sonnet subagents draft the specs, verify acceptance criteria, and run the execute→verify→correct→review loop, so build logs / diffs / review prose never reach the commander's context.
• When a frontier commander (Fable 5) has set direction, it hands control down to Opus/Sonnet — within a session or across sessions — so routine progression doesn't run on frontier tokens.

How it works

Three Claude Code primitives plus the Codex bridge:

Piece	Type	Role
edictum	Skill	The on-demand playbook: delegation loop, spec format, Codex-invocation rules, model/effort selection, calibration.
Edictum block in ~/.claude/CLAUDE.md	always-on policy	~30 lines: role split, commander tiers, token-budget rules. The lean always-on core; detail lives in the skill.
spec-builder, acceptance-checker, pipeline-runner	Subagents (sonnet)	Draft specs, verify acceptance (CI-first), and own the full spec-batch loop.
/delegate	Command	Runs the whole decompose → spec → execute → verify → review → merge pipeline.
codex-plugin-cc	MCP plugin (external)	The Codex bridge. Edictum drives it; it is not bundled here.

Prerequisites

• Node.js 18+
• Claude Code CLI — npm i -g @anthropic-ai/claude-code
• Codex CLI, logged in — npm i -g @openai/codex then codex login
• codex-plugin-cc — claude plugin marketplace add openai/codex-plugin-cc then claude plugin install codex@openai-codex

The installer checks these and prints what's missing; it won't overwrite your work. Node.js, Claude Code CLI, Codex CLI, and codex-plugin-cc are required for install. Codex login status is warning-only. If a required item is missing, the installer stops before changing files unless you pass --force.

Install

Clone, then run the installer for your shell:

# Windows (PowerShell)
git clone https://github.com/4i7/Edictum.git
cd Edictum
./install.ps1

# macOS / Linux / WSL
git clone https://github.com/4i7/Edictum.git
cd Edictum
bash install.sh

The installer:

1. Verifies prerequisites and stops before changes when required items are missing unless --force is set.
2. Copies home-claude/{agents,commands,skills} into ~/.claude/.
3. Merges the Edictum policy block into ~/.claude/CLAUDE.md between <!-- EDICTUM:START --> / <!-- EDICTUM:END --> markers, backing up your existing CLAUDE.md to CLAUDE.md.bak.YYYYMMDD-HHMMSS first and replacing only the marked block on re-runs (idempotent).

Useful installer flags:

./install.ps1 --dry-run
./install.ps1 --project-only C:\path\to\project
./install.ps1 --force
./install.ps1 --uninstall

bash install.sh --dry-run
bash install.sh --project-only /path/to/project
bash install.sh --force
bash install.sh --uninstall

• --dry-run prints prerequisite results, copy targets, CLAUDE.md actions, backup names, and uninstall delete targets without creating, writing, copying, or deleting anything.
• --project-only <path> targets <path>/.claude instead of global ~/.claude. Use this when you want Edictum available only inside one project.
• --force continues installation even if a required prerequisite is missing.
• --uninstall removes only Edictum-installed files and the Edictum block from the selected CLAUDE.md. Combine with --dry-run to preview removal.

Backups use local timestamps, for example CLAUDE.md.bak.20260613-153045. Existing CLAUDE.md files are backed up before merge/uninstall edits; a first-time install also writes an initial backup after creating the Edictum block. Restore with:

cp ~/.claude/CLAUDE.md.bak.<timestamp> ~/.claude/CLAUDE.md

Copy-Item -LiteralPath "$env:USERPROFILE\.claude\CLAUDE.md.bak.<timestamp>" -Destination "$env:USERPROFILE\.claude\CLAUDE.md" -Force

Restart your Claude Code session afterward so the new skill/agents/command load.

Validating / reviewing Edictum

Run the repository self-check before reviewing installer, skill, agent, or command changes:

node scripts/validate-edictum.mjs

Exit 0 means the repo is healthy. The validator uses only Node built-ins and checks front matter, task-path hygiene, installer markers, companion-script sentinels, and delegation artifact ignores.

Do not judge file health by rendering raw.githubusercontent.com content as HTML/Markdown. Rendering can strip <!-- --> comments and collapse whitespace, which creates false positives that make healthy files look empty or malformed. Instead, read the raw bytes, for example:

gh api repos/<owner>/Edictum/contents/<path> --jq .content | base64 -d

Or clone the repository and run node scripts/validate-edictum.mjs.

Repo hygiene

Edictum writes cold-executable specs and verdicts under local, ephemeral work directories that may contain verbatim source or secrets. Add these paths to the target project's .gitignore:

.claude/tasks/
.claude/tasks/results/

Usage

• Just ask for multi-step implementation work normally — the commander policy + skill steer it into delegation automatically.
• Or invoke explicitly: /delegate <what you want built>.
• Borderline-small task you still want delegated? /delegate it. Frontier models implement directly only when checkable acceptance criteria can't be written at all.

Calibrate to your project

Edictum ships with the author's calibration as an example (TS+Rust, n=7). Make it yours: put your project's env/test workarounds in that project's CLAUDE.md so specs inherit them, and after a few real runs update ~/.claude/skills/edictum/reference/model-benchmarks.md with what Codex actually handled at what effort. Label anything unverified as 仮説 / 推定 / 要検証.

Uninstall

Run the installer in uninstall mode:

./install.ps1 --uninstall

bash install.sh --uninstall

This removes only ~/.claude/skills/edictum/, the three Edictum agents, ~/.claude/commands/delegate.md, and the <!-- EDICTUM:START -->…<!-- EDICTUM:END --> block from ~/.claude/CLAUDE.md. It creates a timestamped CLAUDE.md backup first when the block is present.

Layout

home-claude/                      # mirrors ~/.claude/ ; the installer copies this in
  CLAUDE-policy-snippet.md        # the always-on block merged into ~/.claude/CLAUDE.md
  agents/{spec-builder,acceptance-checker,pipeline-runner}.md
  commands/delegate.md
  skills/edictum/
    SKILL.md                      # the on-demand playbook
    reference/{task-spec-template,handoff-template,model-benchmarks}.md
install.ps1 / install.sh

License

MIT © 4i7. See LICENSE.