agentic-dev-team

Agentic Android Dev Team

A reusable PM → Architect → Coder → Tester pipeline for Android projects, working identically in Claude Code and Antigravity. Distributed as per-file symlinks into each consuming project — no whole-dir symlinks, no forced migration of your existing .claude/ or .agents/ content.

How To Use It

Once installed in a project, open it in Claude Code or Antigravity and run a slash command from the chat prompt. The build-* commands run the full pipeline end to end (the suffix tells you whether a human is in the loop, and whether automated reviewers gate each phase); the two plan-* commands run just one early phase and stop, so you can research or design without committing to a build.

/build-guided <vague idea>

Human-in-the-loop variant. PM → Architect → Coder → Tester with approval gates between phases. Use when the idea is rough and you want the PM to refine it before any code is written.

/build-guided add a recently-played carousel to the home screen

The PM agent turns that into a concrete feature spec at pipeline_artifacts/<slug>/feature.md and pauses for your sign-off. After you approve, the Architect produces an implementation plan, the Coder implements it, and the Tester verifies — each phase stops for review.

/build-auto <specified feature>

Fully automatic variant. Architect → Coder → Tester, no gates. Use when the feature is already specified.

/build-auto add a "Save draft on background" hook to ComposeViewModel that persists the current input to Room every 2s and restores it on launch

The Architect writes the plan, the Coder implements it, the Tester runs verification on a device via the auto-mobile MCP server. Artifacts for every run land under pipeline_artifacts/<slug>/.

/build-auto-reviewed <specified feature>

Same shape as /build-auto (no PM phase, no human gates), but each producing phase is followed by an automated reviewer that can send the work back. Use when you want a higher-quality unattended run and are willing to spend more tokens for it.

/build-auto-reviewed add a "Save draft on background" hook to ComposeViewModel that persists the current input to Room every 2s and restores it on launch

The Architect writes the plan, then @adt-android-architect-reviewer reviews it; the Coder implements, then @adt-android-code-reviewer reviews the diff. On each gate, if the reviewer requests changes, the producing agent is re-run with the feedback — at most twice per gate. If a reviewer still isn't satisfied after the second re-run, the pipeline stops and reports rather than shipping work the reviewer rejected. A clean run then hands off to the Tester as usual.

Planning-only commands

When you want a plan but not a build, run just one phase — each runs a single phase, stops after writing its artifact, and works identically in Claude Code and Antigravity. Their output chains into a build command (or each other) later.

/plan-research <vague idea>

Runs only the PM. Turns a rough idea into a concrete, unambiguous feature spec at pipeline_artifacts/<slug>/feature.md, asking clarifying questions along the way — then stops.

/plan-research add a recently-played carousel to the home screen

Feed the resulting feature.md to /plan-design, /build-auto, or /build-guided when you're ready.

/plan-design <feature.md path | specified feature>

Runs only the Architect. Produces an implementation plan at pipeline_artifacts/<slug>/implementation-plan.md from either a feature.md (e.g. from /plan-research) or a clear feature description — then stops.

/plan-design pipeline_artifacts/recently-played-carousel/feature.md

Feed the resulting plan to /build-auto or /build-guided to implement and verify it.

Installation

Two paths:

• Claude Code plugin — installable from the Claude Code CLI via the plugin marketplace. Installs the adt-* agents and the /build-auto, /build-auto-reviewed, /build-guided, /plan-research, and /plan-design slash commands without per-project setup. Does not wire up Antigravity (no .agents/workflows/ files, no agents.md persona stubs, no .gitignore block).
• install.sh per-project (local repo install) — works for both Claude Code and Antigravity. Materializes per-file symlinks inside each consuming project, manages a .gitignore block, and inlines persona stubs into .agents/agents.md for Antigravity. This is the only supported path for Antigravity — Antigravity has no plugin marketplace.

The two are not mutually exclusive — you can install the plugin in Claude Code and still run install.sh in Antigravity projects.

Install via Claude Code marketplace (plugin)

From inside the Claude Code CLI:

/plugin marketplace add jaxvy/agentic-dev-team
/plugin install agentic-dev-team@adt-pipeline