Agent Workflow

This repo uses a small repo-native agent layer. It is intentionally narrower than larger orchestration frameworks.

Precedence

Use the following order when shaping or reviewing work:

AGENTS.md defines repo-wide architecture, invariants, and trust-boundary rules.
Path-scoped .cursor/rules/*.mdc files override the generic builder workflow inside their scope.
The relevant repo skill under script/agent-runtime/skills/ supplies workflow guidance and expected validation commands for that class of change.
node script/agent-runtime/verify-change.js <paths...> or pnpm agent:verify-change -- <paths...> selects the required verification commands and manual review flags from the touched paths.

Bundles a small set of high-signal repo skills for common high-risk work.
Keeps skill metadata local to the repo instead of depending on an external marketplace or control plane.
Maps changed paths to explicit validation commands so contributors do not need to guess which checks matter.

No multi-agent orchestration or task delegation framework.
No persistent memory, instincts, or automatic learning loop.
No cross-tool compatibility abstraction beyond the repo’s existing Cursor/Codex posture.
No automatic write-time hooks that mutate files or block commits.

Read AGENTS.md and any path-scoped rule files that cover the files you are changing.
Check the matching skill in script/agent-runtime/skills/ for risk-specific guidance.
Run pnpm agent:verify-change -- <paths...> with repo-relative changed paths.
Execute the listed commands and address any manual review flags before opening a PR.

Each SKILL.md file under script/agent-runtime/skills/ must include frontmatter with these required fields:

Malformed or duplicate skill metadata fails discovery tests on purpose. Silent degradation is not allowed.