CDD Skills: controlled autonomous coding

Plan once. Approve the run. Let agents work while you sleep.

Skills for Codex and Claude Code that keep long agent runs on a leash: you plan into TODO steps, approve the kickoff, and get back an isolated branch plus a mission report.

Master Chef is the multi-step runner. It walks those TODO steps with recovery, commits, and checks so you can review in the morning instead of babysitting every edit.

Powered by cdd-boilerplate.

---

30-second path

# 1) Install (Codex + Claude defaults)
bash <(curl -fsSL https://raw.githubusercontent.com/ruphware/cdd-skills/main/install-remote.sh) --all

# 2) Runtime permissions so the run does not stall on tool prompts:
#    Codex  → Permissions: approve for me
#    Claude → permission mode: Auto

# 3) In a CDD repo: plan → readiness check → autonomous run
$cdd-plan docs/specs/blueprint.md into TODO.md
$cdd-audit TODO.md and confirm cdd-master-chef readiness
$cdd-master-chef TODO.md all steps, session gpt-5.6 max builder gpt-5.6 high, worktree and branch todo-123

Use $cdd-* in Codex; /cdd-* (or your runtime's skill entry) in Claude Code.

Swap the model names for whatever your session actually runs. Session model must match the live session.

---

Master Chef

Built for unattended runs you can still merge after review.

What it does	Walks runnable TODO steps end to end in a branch-backed worktree
Why trust it	One kickoff approval, paced execution, checks, commits, mission report
Budget	Predictable token use from careful pacing, not a free-form agent ramble
Required input	A CDD TODO from $cdd-plan (or equivalent chunked steps). No plan, no run.

Prerequisites, then how the run stays reviewable

1. $cdd-plan first so TODO.md has bounded, checkable steps. Master Chef executes a plan; it does not invent one.
2. $cdd-audit (recommended) to confirm step size and Master Chef readiness before a long run.
3. One kickoff approval sets how far the run can go.
4. Branch-backed worktrees keep changes isolated until you review.
5. Persistent Builder context carries knowledge across steps instead of starting cold each time.
6. Oversized work is reviewed first; split only when the split cost is justified.
7. Wave-parallel (opt-in): TODO steps annotated with deps: / touches: can run declared-disjoint work as bounded waves with a serial merge queue.
8. Mission report records completed steps, checks, pushes, decisions, and next actions.

$cdd-plan docs/specs/blueprint.md into TODO.md
$cdd-audit TODO.md and confirm cdd-master-chef readiness
$cdd-master-chef TODO.md all steps, session gpt-5.6 max builder gpt-5.6 high, worktree and branch todo-123

Note

Session model must match the running session. Example above uses gpt-5.6 max + builder gpt-5.6 high; change both to your stack.

Runtime tool permissions are separate from Master Chef's one-time kickoff approval. For long runs, set Codex Permissions → approve for me, or Claude permission mode → Auto, so shell/tool calls do not block overnight.

Runtime-native alternatives

Codex: /goal $cdd-implement TODO.md all undone steps
Claude: /workflows $cdd-implement TODO.md all undone steps

These skip Master Chef's orchestration (worktree lifecycle, recovery, mission report). Prefer Master Chef for multi-step unattended delivery; watch spend on long native loops.

Start Master Chef via $cdd-master-chef (Codex) or /cdd-master-chef (Claude Code).

---

When to use what

You want	Use
Overnight / multi-step autonomous delivery	Plan (+ audit), then Master Chef
One bounded change with you in the loop	Boot → Plan → Implement
Review shipped work or a proposal	Audit → Plan approved follow-up
Doc drift, cleanup, index, refactor sweep	Maintain
New or non-CDD repo	Init Project

Human-guided: Boot → Plan → Implement
Review: Audit → Plan
Autonomous: Plan → Audit (readiness) → Master Chef

---

Skill map

Core loop: boot context, plan work, implement one step, audit the result. Master Chef runs the loop unattended once the plan is ready.

	Skill	When
[CDD-0]	cdd-boot	Load repo context; route to the right next skill or vanilla work
[CDD-1]	cdd-init-project	Adopt CDD in the current repo (gh helps on GitHub)
[CDD-2]	cdd-plan	Turn a change request or audit findings into TODO steps
[CDD-3]	cdd-implement	Ship exactly one bounded task (TODO-backed preferred)
[CDD-4]	cdd-audit	Goal-fit review with proportional proof; route follow-up to plan
[CDD-5]	cdd-maintain	Doc drift, cleanup, indexes, refactor candidates
[CDD-6]	cdd-master-chef	Controlled multi-step autonomous execution

---

Install

Fastest (no clone)

bash <(curl -fsSL https://raw.githubusercontent.com/ruphware/cdd-skills/main/install-remote.sh) --all

Update:

bash <(curl -fsSL https://raw.githubusercontent.com/ruphware/cdd-skills/main/install-remote.sh) --all --update --yes

Uninstall:

bash <(curl -fsSL https://raw.githubusercontent.com/ruphware/cdd-skills/main/uninstall-remote.sh) --all

Local clone (specific runtime or target)

git clone git@github.com:ruphware/cdd-skills.git
cd cdd-skills
./scripts/install.sh --all          # or --runtime claude
./scripts/install.sh --help

Optional NPX (first install only)

Codex, Claude Code, and Gemini CLI. No managed update/uninstall or prune semantics; prefer the scripts above for day-to-day.

npx skills add https://github.com/ruphware/cdd-skills/ \
  --skill cdd-boot --skill cdd-init-project --skill cdd-plan \
  --skill cdd-implement --skill cdd-audit --skill cdd-maintain \
  --skill cdd-master-chef -a codex -a claude-code -a gemini-cli -g

Gemini gets the core skills path; concrete Master Chef adapters today are Codex and Claude Code.

---

Adapters & runtimes

Runtime	Docs
Codex	CODEX-ADAPTER.md · CODEX-RUNBOOK.md
Claude Code	CLAUDE-ADAPTER.md · CLAUDE-RUNBOOK.md

Shared:

• RUNBOOK.md
• RUNTIME-CAPABILITIES.md

---

License

Free-to-use-adjust-just-don't-blame-me-for-anything-license. ✌️

---

Structured CDD Project + CDD Skills workflow via local AGENTS.md.