Skip to content

dev-team

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

dev-team

README.md

dev-team

🌐 English · Français

An agentic development team for Oh-My-Pi, ported from Bryan Finster's bdfinst/agentic-dev-team. An orchestrator routes work through a spec → plan → build → PR workflow with a forced plan gate (pre-analysis → plan → build → review), tests required, and human gates, backed by 32 specialist and critic agents and blocking guard extensions.

Philosophy (Finster): AI is a high-pass filter for engineering discipline. The win comes from continuous delivery and defining what you build before you build it — not from picking a model. This plugin encodes that discipline as a forced plan gate plus tests as the safety net. (Test-first/TDD ordering is not enforced — it adds little for AI agents; the leverage is the plan gate.)

Install

omp plugin marketplace add outofrange-consulting/omp-dev-team
omp plugin install dev-team@omp-dev-team

Then paste config.snippet.yml into ~/.omp/agent/config.yml (model routing + skills.enableSkillCommands + the task graph). The plugin's prerequisite checker:

bash plugins/dev-team/install.sh        # checks OMP/git/optional tools, can apply config

The workflow

/specs → /plan → /build → /pr:

  1. /specs — capture a feature as Intent + BDD scenarios + Architecture + Acceptance Criteria. Human approves.
  2. /plan — turn the approved spec into a step-plan (each slice names its tests), with per-slice Depends-on metadata grouping slices into build waves. Five plan-review critics (acceptance-test, design, UX, strategic, parallelization) challenge it in parallel before you see it. Approve with /plan-approve to unlock the build (plan-gate); for a genuinely trivial task, /scope --trivial instead.
  3. /build — execute the approved plan, building independent slices concurrently wave by wave. Tests are required per unit (test-first optional), verified green by /impl-verify.
  4. /pr — run the quality gates and open the pull request.

For complex work the orchestrator runs Research → Plan → Implement with a human gate between phases. Plus /code-review (/review), /review-agent, /continue, /triage, /design-doc, /issues-from-plan, and the /routing diagnostic. Every skill is also available as /skill:<name>.

Model tiers (all cloud)

Agents declare a tier in model: frontmatter, resolved natively by your modelRoles:

Tier Frontmatter For
small pi/smol (cheap cloud, default Haiku) lexical/structural checks, checklist reviews — high volume
balanced claude-sonnet-4-6 most team & review agents, orchestrator
deep claude-opus-4-8 cross-file reasoning, design synthesis, threat modeling, recon

The high-volume small tier is where token spend concentrates — keep it cheap. Point modelRoles.smol at claude-haiku-4-5, or (with the copilot-preset plugin) at github-copilot/gpt-5-mini to run it on your Copilot license. Pair with token-diet to cut tokens further. Source of truth: skills/dev-team-knowledge/model-routing.json; diagnose with /routing.

Guardrails (extensions)

plan-gate (blocks edits to production source until the task is scoped and a plan is approved — enforces pre-analysis → plan → build → review; /scope, /trivial, /plan-approve, /plan-reset), path-guard (secrets), destructive-guard + /careful, freeze-guard (/freeze /unfreeze), spec-guard (blocks edits to existing .feature specs — fix code, not the spec; /allow-feature-edits to override), review-gate (blocks git commit until /code-review + /review-approve), impl-verify (/impl-verify strict build + tests), telemetry + /cost-report, model-routing (dispatch tier log

  • /routing). They intercept tool_call and block with a reason — OMP's native blocking mechanism.

Layout

.claude-plugin/plugin.json · package.json (omp.extensions)
agents/  skills/  commands/  rules/  extensions/  .mcp.json
skills/dev-team-knowledge/   # registries, rubrics, model-routing.json
config.snippet.yml  install.sh  install.ps1