Dev Plugin
Accelerate development with seven specialist agents, nineteen skills, multi-agent team orchestration, and behavioral skill nudges that watch your work across every installed plugin. The dev plugin is the most-used Agronod plugin — most other plugins compose against its agents and skills.
At a Glance
- 7 specialist agents — architect, engineer, researcher, critic, code-writer, experience-analyst, storyteller
- 19 skills — 6 you invoke directly, 13 that agents load for process and discipline
- Team orchestration — assemble multi-agent teams with leadership protocols and a live dashboard
- Skill awareness — watches your tool usage across all installed plugins and suggests skills when patterns emerge
- 5 hooks — auto-approve reads, detect LSP tools, inject linting config, behavioral skill nudges, cross-plugin skill index
Install
/plugin marketplace add agronod/claude-plugins
/plugin install dev@agronod-plugins
/dev:setup/dev:setup validates your environment and installs the claude-server binary required by team skills.
Daily Surface Area
The skills you invoke directly:
/dev:research # Research a library, API, or tech decision
/dev:review # Review changed code for quality and patterns
/dev:spawn-team # Ad-hoc team for analysis or deliberation
/dev:team-feature # Plan-then-implement a feature with a team
/dev:team-dashboard # Watch agent teams work in real time
/dev:setup # Validate plugin environmentAgents and Skills
Agents define who does the work. Skills define how it gets done. Any agent can load any skill — they compose freely.
Agents
| Agent | Sees | Contributes |
|---|---|---|
| architect | System boundaries, data models, contracts | Integration strategy, design trade-offs |
| engineer | Tech stack, conventions, core logic, testing | Implementation patterns, validation |
| code-writer | Focused implementation scope | Detailed, clean code from a blueprint |
| researcher | Web, docs, source code | Answers, options, evidence |
| experience-analyst | User flows, error states, edge cases | UX and accessibility concerns |
| critic | Failure modes, assumptions, pre-mortems | Risk analysis, adversarial review |
| storyteller | Decision context, narrative flow | ADRs, changelogs, design docs, post-mortems |
Invocable Skills
The six skills you invoke directly cover the most common workflows:
/dev:research — Assembles current docs and code examples for a library, API, or architecture question. Start here when the documentation is stale or you need to verify behavior across versions.
/dev:review — Reviews changed code for quality, patterns, and maintainability. Load this before committing multi-file changes or opening a PR.
/dev:spawn-team — Assembles agents with different expertise (architect, researcher, critic) to explore a question or decision from multiple angles. Use when the right approach isn't obvious and benefits from debate.
/dev:team-feature — Orchestrates a full team for features spanning 3+ files: a leader that plans, parallel implementers, and a critic that spots risks. Both planning and implementation happen together with live status.
/dev:team-dashboard — Displays real-time status of running teams in a browser window. Pair this with /dev:team-feature or /dev:spawn-team to watch progress and debug blockages as they happen.
/dev:setup — Validates your environment and installs the claude-server binary required by team skills. Run this once after install or if team skills fail to start.
Agent-Loaded Skills
These don't surface as slash commands — agents load them when their work calls for the protocol.
| Skill | Enforces |
|---|---|
code | Production quality: no TODOs, clarity, completeness |
explore | Pre-implementation investigation: trace data flow, map contracts |
test-writing | Integration-first tests, mock externals only |
readme | README quality standards: progressive disclosure, format selection |
technical-writing | Audience identification, progressive disclosure, editorial standards |
prompt-writing | Calibrated system prompts that don't over-prescribe |
team-leader | Delegation, synthesis, decision-making |
team-member | Reporting, disagreement, collaboration |
propose-skill | Workflow-gap detection and skill proposals |
skill-description-writer | Trigger optimization for autonomous activation |
skill-health | Triage for skill-ecosystem problems (gap, drift, overlap, trigger-blindness) |
update-skill | Body updates when documented workflow drifts from actual behavior |
reconcile-skills | Overlap resolution — merge, differentiate, or absorb |
Team Orchestration
/dev:team-feature assembles a team that plans collectively, implements in parallel, and reports through a live dashboard:
/dev:spawn-team is the lighter sibling — ad-hoc teams for analysis, debate, or research without the plan-then-implement structure. Both require the CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 environment variable that /dev:setup sets up.
Deciding Between Single Agent and Team
For focused work — a bug fix, small refactor, or feature in one or two files — a single agent or no agent is faster and cheaper. Use /dev:spawn-team when a decision or analysis would benefit from multiple expert perspectives: architect reviewing boundaries, researcher exploring unfamiliar tech, critic stress-testing assumptions. Features spanning 3+ files where both planning and implementation matter should go through /dev:team-feature, which ensures the plan is sound before agents write code in parallel. Each agent in a team can ask others for clarification or critique; spawning multiple isolated agents and reconciling results yourself loses that collaboration.
Hooks
Five hooks run automatically:
| Hook | Event | What it does |
|---|---|---|
| Auto-approve | PreToolUse | Skips confirmation for read-only operations (git status, GitHub queries, docs lookups). Writes still confirm. |
| Linting injection | PreToolUse (matcher: Task) | Loads project lint config into spawned subagents so they respect your rules. |
| LSP detection | SessionStart | Detects LSP tool providers and injects guidance for using them over Grep/Glob. |
| Skill index | SessionStart | Indexes skills across all installed plugins, not just dev. |
| Skill awareness | PostToolUse + UserPromptSubmit | Watches tool patterns and nudges when a skill would help. |
The skill awareness system fires across plugins. It runs the same logic for dev:code, meta:challenge, or any third-party plugin's skills — once installed, they're all candidates for nudging.
The system watches patterns in your tool use and suggests a skill when a pattern emerges. Here are the patterns it recognizes:
| Pattern | Threshold | Suggests |
|---|---|---|
| Source file edits without a coding skill | 3+ edits | Coding discipline |
| Grep/Glob searches without exploration | 5+ searches | Systematic investigation |
| WebFetch/WebSearch without research | 2+ calls | Research protocol |
| Test file edits without testing skill | 2+ edits | Integration testing |
| Edits across many files | 4+ files | Feature orchestration |
| Reading many source files | 5+ files | Code review |
You receive a maximum of 3 nudges per session — one per category — and the system stays silent if a relevant skill is already loaded.
Auto-Approve Matrix
Operations that skip confirmation
| Category | Auto-approved | Scope |
|---|---|---|
| Git | status, log, diff, branch, show, rev-parse, remote, describe, tag, merge-base, ls-files, ls-tree | Current repo only |
| GitHub CLI | pr list/view/status/checks/diff, issue list/view/status, repo view/list | Current repo or Agronod org |
| GitHub MCP | get_*, list_*, search_*, pull_request_read, issue_read | Agronod org only |
| Documentation MCP | All lookups | None |
| IDE MCP | getDiagnostics, getOpenFiles, getActiveFile | None |
| HTTP | curl GET only | No upload flags |
| Built-in | WebSearch, ListMcpResourcesTool, ReadMcpResourceTool | None |
| Skills & Subagents | All invocations | None |
Never auto-approved: git push/commit/reset, file edits, GitHub ops on non-Agronod repos, POST/PUT/DELETE requests.
Cross-Plugin Composition
Dev's agents are reusable across all plugins. The specs plugin chains dev:researcher during planning and dev:code-writer during build; any skill can spawn dev:critic for review or dev:storyteller for documentation. Reference agents by their qualified name: subagent_type: "dev:engineer".
Related
- Component Model — How commands, skills, and agents fit together
- Skill Activation — Why skill descriptions matter for autonomous use
- Hooks & Lifecycle — The hook events the dev plugin listens on
- Specs Workflow — The spec-driven workflow that composes dev agents
- Using Subagents — Briefing single agents and their validation