Agentic Delivery Playbook is a small, evidence-first delivery kernel for Codex, ChatGPT Work, and Pi: classify the work by consequence, choose the minimum safe GPT-6 route, and close from recorded evidence.
Direct and Controlled are process modes, not model choices. Use the decision conditions from playbook.md exactly:
| Mode | Decision condition | Delivery shape |
|---|---|---|
| Direct | Use only when all are true: intent and acceptance are clear; effects are low-consequence or readily reversible; no material security, privacy, auth, payment, financial, destructive, public-contract, or external-authority concern exists; one owner can make the change without unsafe coordination; and completion is objectively verifiable. | State the change, make the smallest correct edit, run deterministic validation, and report evidence. |
| Controlled | Use when any Direct condition is false, or when broad delegated work needs ownership and synthesis controls. | State a compact contract, preserve acceptance evidence, and review the contract, actual diff, and evidence in fresh context. |
GPT-6 Luna, Sol, and Astra are selected for task shape, independently of the delivery mode.
| Route | Use for | Topology |
|---|---|---|
| Luna | Focused, repeatable work; start with High reasoning. | Single agent; increase effort only when needed and supported. |
| Sol | Complex coding, architecture-sensitive judgment, risk, escalation, and fresh-context review; start with Medium, use High selectively. | Single agent by default. |
| Astra | Broad code, research, or computer-use work; start with Light/low effort. | Single agent or bounded independent lanes. |
| Ultra | Broad work with genuinely independent lanes. | Maximum reasoning with one synthesis barrier; it is not another process mode. |
- Read adapters/codex/AGENTS.md and merge it into the target repository's existing AGENTS.md.
- Optionally review profiles/codex/, then copy the selected project settings into the target repository's .codex directory without overwriting local choices.
- Start Codex in that repository and give it the task.
- Paste adapters/chatgpt/instructions.md into ChatGPT Work project instructions.
- Optionally attach profiles/gpt-6.md, templates/contract.md, and templates/run.json for controlled work and handoffs.
- Provide only the sources and authority needed for the current task.
Install the package as a skill resource from a local checkout with pi install /path/to/agentic-delivery-playbook. After merge, install the default Git branch with pi install git:github.com/arcayne/agentic-delivery-playbook; after publication, use pi install npm:agentic-delivery-playbook. Installation adds the canonical agentic-delivery-playbook skill; it does not set project/global model defaults or apply the optional settings template. See Pi install, update, and use notes.
npx agentic-delivery-playbook show codex
npx agentic-delivery-playbook show chatgpt
npx agentic-delivery-playbook show profile
The CLI only prints maintained artifacts; it does not install or overwrite local or global configuration.
- Delivery modes
- GPT-6 routing profile
- Compact contract template
- Durable run-record template
- Evaluation protocol
- End-to-end walkthrough
- Maintained adapters
- Active controlled example
Only the GPT-6 profile is maintained. Prior material is unsupported under legacy/ and excluded from the npm package.
For 0.2 users, the mutating install pi and install claude commands were removed; the current Pi support is a skill-only package resource, not a restored mutating CLI command. Existing installations are not updated automatically. Review the maintained adapter instructions and install the Pi package separately if wanted.
See CONTRIBUTING.md, SECURITY.md, and CHANGELOG.md.