Agent Roles

Role-based coding agents for architecture, product, implementation, verification, security, and delivery judgment.

This plugin is intentionally agent-only. It gives OMP and compatible harnesses crisp senior engineering roles that can be invoked independently or composed by a lead agent. It does not add an orchestration workflow, command router, or role hierarchy.

Core invariant:

Each role owns one kind of judgment and must hand off when the question belongs to another role.

Roles

Agent	Owns	Use when
product-architect	Product intent and acceptance behavior	A request is fuzzy, acceptance is unclear, or implementation could redefine intent.
systems-architect	System shape, boundaries, interfaces, and trade-offs	A change affects modules, APIs, data flow, or long-term architecture.
principal-engineer	Correct, boring, maintainable implementation	Code must be changed or reviewed with clean cutover discipline.
verification-lead	Proof, tests, smoke checks, and acceptance evidence	Work needs behavioral evidence before it can be called complete.
security-architect	Threat model, trust boundaries, auth, secrets, data exposure	A change touches users, data, auth, external input, deployment, or secrets.
delivery-lead	Merge, release, rollout, rollback, and reviewer readiness	Work is ready to leave the branch or needs a readiness packet.

Handoff chains

Complex feature:

product-architect -> systems-architect -> principal-engineer -> verification-lead -> delivery-lead

Sensitive feature:

product-architect -> systems-architect -> security-architect -> principal-engineer -> verification-lead -> delivery-lead

Bug fix:

principal-engineer -> verification-lead -> delivery-lead

Architecture decision:

product-architect when behavior is unclear -> systems-architect -> verification-lead for proof strategy

Install

OMP

Install from GitHub as an OMP plugin marketplace:

omp plugin marketplace add dmltdev/agent-roles
omp plugin install agent-roles@agent-roles

For local development:

omp plugin marketplace add /path/to/agent-roles
omp plugin install agent-roles@agent-roles --force

Claude Code compatible plugin marketplace

/plugin marketplace add dmltdev/agent-roles
/plugin install agent-roles@agent-roles

Codex compatible plugin

codex plugin marketplace add dmltdev/agent-roles --ref main
codex plugin add agent-roles@agent-roles

Dispatch examples

Use product-architect to turn this feature idea into behavior and acceptance criteria. Do not design the implementation yet.

Use systems-architect to review this proposed API boundary. Return APPROVE, REVISE, or REJECT with trade-offs and the smallest safe implementation slice.

Use principal-engineer to implement the approved plan. Keep a clean cutover, migrate callers, and remove obsolete paths.

Use verification-lead to define the narrowest proof that this change works against the acceptance criteria.

Use security-architect to threat-model this endpoint and identify realistic abuse paths before implementation.

Use delivery-lead to decide whether this branch is ready to push or merge. Include blockers, checks, and reviewer packet.

Instruction sources

The role contracts are source-grounded syntheses from related skills.sh candidates. Each shortlisted candidate was followed to raw GitHub SKILL.md instructions and license evidence where available. The agents do not copy upstream text, command names, repository paths, or tool assumptions.

Auditable source matrix: docs/research/skills-sh-role-candidates.md. OMP install verification: docs/research/omp-install-verification.md.

Role	Raw-source-backed synthesis
product-architect	Anthropic write-spec, Warp write-product-spec, and acceptance-criteria-designer: consumer behavior, measurable outcomes, non-goals, and observable criteria.
systems-architect	architecture-review: explicit decisions, boundaries, trade-offs, guardrails, and first safe slices.
principal-engineer	code-review-and-quality, code-simplification, and refactor-safely: correctness, boring maintainability, clean cutover, behavior preservation, and targeted proof.
verification-lead	test-driven-development and verification-before-completion: behavioral tests where durable contracts changed and fresh evidence before claims.
security-architect	OpenAI security-threat-model and OWASP api-security-review: trust boundaries, assets, abuse paths, API risks, assumptions, and testable controls.
delivery-lead	ship, skrrt release, and pre-release: explicit git/release boundaries, pre-launch gates, real-history release notes, immutable release artifacts, blockers, rollback, and reviewer evidence.

Boundary

• This plugin supplies roles, not a mandated lifecycle.
• Roles must not silently steal each other's decisions.
• Read-only roles must not edit files even if their tool list would technically allow it.
• Output must be evidence-backed when it mentions files, commands, tests, URLs, or runtime behavior.
• Missing product intent, architecture constraints, proof evidence, or release policy should be surfaced as a handoff or blocker, not guessed.