A local product-development harness that carries explicit evidence and human decisions from discovery through definition, design, specification and a reviewed implementation branch. It operates against a separate, authorized target project. It retains failed attempts and keeps verification, human acceptance, integration, release and product validation distinct.
Start with the synthetic onboarding walkthrough. It creates a disposable sample, reaches a read-only dry-run plan, and demonstrates restart using the same event identity without a model call or GitHub publication.
Prerequisites: Node 20+, Git 2+, Python 3 for artifact validation, and native macOS
Seatbelt for execution. Claude Code is required for model execution; gh is
needed only for GitHub mode. Browser checks additionally use an operator-installed
Playwright module and Chrome, configured outside target source. This repository's
runner has no npm dependencies. Provider authentication and spend require the
operator's explicit setup; doctor never installs or authenticates anything.
Run node runner/bin/ateam-runner.mjs doctor --path /absolute/target --json to
inspect local setup. Whole-process-tree containment is an accepted limitation documented in #37;
native availability checks do not establish that guarantee.
The skill bundle lives in .claude/skills/; SKILLS.md lists each
entry. Point the chosen agent's skill discovery at the reviewed checkout using
that agent's supported configuration, then confirm it can read the feature
entry and its relative resources. Keep the complete pinned bundle together;
copying one SKILL.md without its references is insufficient. Skill installation
is an explicit operator step, never a diagnostic side effect. Record the full
harness commit and location in the target's A-Team Config.
See RUNNER.md for commands, execution boundaries for configuration and native isolation, CONTRACT.md for product and authority rules, and combined verification for the evidence required before delivery. No command or skill grants client project authority, merges a PR by implication, or substitutes automated tests for a target-user study.
For an opt-in comparison with role agents, see the agent workflow trial.
/feature keeps the existing skill-based harness by default; select
--orchestration-mode agent on a new feature to trial the complete agent workflow.