A small Codex workflow pack for completing observable software outcomes. One main agent owns exploration, necessary design, implementation, integration, and verification. Use subagents only when independent context, parallel work, isolated ownership, or review adds value.
This replaces the historical AgentTeam experiment. The former skills, role graph, slice-plan formats, validators, and duplicate templates have been removed. There are no compatibility aliases or migration adapters. Historical versions remain in Git. MIT License.
| Skill | Use it for |
|---|---|
| software-delivery | Implementing or fixing a capability through its end-to-end acceptance. Default for ordinary development. |
| requirements-design | Design-only work and material product, architecture, or contract decisions. Reuse adequate existing documents. |
| release-readiness | Assessing the exact release candidate and separating repository checks, prerequisites, deployment, and business acceptance. |
A clear small fix can be completed directly without a new planning document or agent. Large work may use a short plan of independently useful outcomes; it does not require a dedicated implementer or a commit for each internal step.
- Read the target repository's own instructions. Copy the needed folders from
.agents/skills/into its.agents/skills/, with their resources intact. Each skill is independently usable. - Optionally copy delivery-reviewer.toml into
.codex/agents/for independent review. The built-in agents remain usable; this pack does not set models, thread limits, sandbox policies, or Goals. - Merge relevant workflow guidance from AGENTS.md, preserving the target's architecture, authority, permissions, and release rules. Discover commands and paths from actual evidence as needed; no full-profile bootstrap is required.
- Ask for the outcome and acceptance, for example:
Use $software-delivery to fix the failing save workflow and verify the affected success and failure paths.
Skill templates live only under their owning skill's assets/. Adapt them when useful; do not create documents just to fill a template. Role instructions request behavior and do not create an independent read-only sandbox.
- Integrate the smallest runnable end-to-end path early, then finish required behavior and meaningful failure checks.
- Delegate bounded independent tasks rather than copying an organization chart into agents. Keep one writer per worktree and account for shared runtime state.
- Use independent review for consequential risk or when required. Include new files and distinguish the user's pre-existing changes. Acceptance applies to the actual reviewed content.
- Reuse evidence while relevant code/diff, configuration, environment, and data assumptions hold. The main agent performs the final critical check; avoid automatically running every suite in every role.
- Keep a short checkpoint only for useful recovery/handoff. Update on material events; store raw evidence in artifacts and batch routine status updates with the work.
- Missing external evidence blocks dependent work or claims only. Deployment and live mutations require explicit authority; local tests do not prove production acceptance.
Requires Python 3.11+ and the development dependencies. From the repository root:
python -m pip install -r requirements-dev.txt
python scripts/check_pack.py
python -m unittest discover -s tests -vTo check a separately copied skill: python scripts/check_pack.py --skill /path/to/software-delivery.
The checker validates packaging, metadata, and local resource links. It does not judge design semantics, authorize operations, or establish business acceptance. Behavioral evaluation cases cover task routing, evidence reuse, external gaps, and coordination behavior. Requirements and design explain this replacement.
Measure future workflow changes by cost/time per accepted outcome, first end-to-end pass, and rework/regressions under matched conditions. This revision does not claim measured token savings or delivery speedups.