Clawdio: Router de Agentes SDLC
jasonmadigan/clawdio
Router agent com especialistas para automação nativa do SDLC no Claude
Ver no GitHubPlugin Claude Code que orquestra automação de ciclo de vida de software (SDLC) através de um agente router que despacha para subagentes especialistas. Implementa skills de workflow e hooks de guardrails, focando em confiabilidade de agentes nativos do Claude Code em vez de infraestrutura de orquestração complexa.
Instalação
npx convez-ai jasonmadigan/clawdio1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Orquestração inteligente de tarefas de desenvolvimento entre múltiplos agentes especialistas
- Automação de fluxo de código (review, testes, merge, documentação)
- Integração com GitHub, Jira e workflows de equipe de desenvolvimento
Veredito

Solução focada em confiabilidade de agentes para SDLC com arquitetura sensata (router → especialistas → skills). Requer setup complexo de dependências externas (plugins, MCP servers, CLI tools) e é mais adequada para equipes que já usam Claude Code e ferramentas como GitHub/Jira.
README
A tradução para o português deste README ainda está na fila.
clawdio
Claude Code plugin for SDLC automation. A router agent dispatches to specialist subagents based on the task. Skills provide cross-cutting workflow knowledge. Hooks enforce guardrails.
The premise: the bottleneck is never orchestration infrastructure -- it's agent reliability. This plugin invests in good agents, skills, and hooks that work natively in Claude Code, replacing a custom Go orchestrator that was over-built for what it did.
Install
# add the marketplace
claude plugin marketplace add jasonmadigan/clawdio
# install the plugin (user scope, available in all repos)
claude plugin install clawdio
For local development:
claude --plugin-dir /path/to/clawdio
# or with a specific agent and permissions
claude --plugin-dir ~/Work/clawdio --dangerously-skip-permissions --agent clawdio:router
Reload after changes without restarting Claude:
/reload-plugins
Dependencies
Install these separately -- clawdio agents and skills reference them.
Plugins
| Plugin | Install | What it provides |
|---|---|---|
| agent-skills | claude plugin marketplace add addyosmani/agent-skills && claude plugin install agent-skills | Security hardening, code review, TDD, debugging, git workflow, spec-driven development |
| dev-team-plugin | claude plugin marketplace add kuadrant/dev-team-plugin && claude plugin install kdt | Design docs, feature lifecycle, Go PR review, doc verification, external contribs |
| playwright | claude plugin install playwright | Browser automation for UI test verification |
Clawdio handles SDLC orchestration (router, specialists, shipping). agent-skills handles cross-cutting development practices. kdt provides design doc workflows and feature lifecycle management. The router dispatches to all three.
CLI tools
| Tool | Purpose | Used by |
|---|---|---|
gh | GitHub issue/PR operations | implement, review, triage, refine, address-feedback, next, ship, worktree-worker, issues |
Must be authenticated (gh auth login).
MCP servers
| Server | Purpose | Used by |
|---|---|---|
| GitHub MCP | Issue/PR comments, review threads | address-feedback |
| Atlassian MCP | Jira issue search, creation, updates | next, triage, router |
Install Atlassian MCP via claude mcp add atlassian -s user -e JIRA_URL=https://your-site.atlassian.net -e JIRA_USERNAME=you@company.com -e JIRA_API_TOKEN=your-token -- uvx mcp-atlassian --jira-url https://your-site.atlassian.net. Requires uv installed.
How it works
Talk to the router agent. It classifies your request and dispatches the right specialist.
graph LR
You -->|request| Router
Router -->|classify & dispatch| Agents[agents: implement, triage, refine, address-feedback, release-notes, test-writer, docs, worktree-worker]
Router -->|invoke| Skills[skills: next, ship, pluck, issues, pr-description, doc-sync, review-coordination, merge-gate, worktree-recovery, parallel-ship]
Router -->|review fanout| Review[code-reviewer + test-verifier + domain specialists]
Agents -->|result| Router
Review -->|findings| Router
Router -->|present & confirm| You
Hooks -.->|guardrails| Agents
Review flow
The router owns the review fanout. Subagents can't spawn sub-subagents, so the router dispatches all specialists in parallel.
graph TD
A[User: review PR] --> B[Router: classify files]
B --> C{File types?}
C -->|always| D[code-reviewer]
C -->|always| E[test-verifier]
C -->|*.go| F[go-k8s-reviewer]
C -->|*auth*| G[auth-reviewer]
C -->|*crypto*| H[security-auditor]
D & E & F & G & H -->|findings| I[Router: merge across axes]
I --> J{Verdict}
J -->|APPROVE| K[offer merge]
J -->|CHANGES REQUESTED| L[draft PR comment]
J -->|BLOCKED| L
L --> M{User: post?}
M -->|post| N[gh pr comment]
M -->|edit| L
M -->|don't post| O{Next?}
N --> O
K --> O
O -->|address feedback| P[address-feedback agent]
O -->|merge| Q[merge gate]
O -->|done| R[next]
P --> S{Next?}
S -->|re-review| B
S -->|merge| Q
S -->|next| R
Ship flow
Full lifecycle from issue to merged PR. Supports --resume (pick up mid-flow), --skip-review, and --draft.
graph TD
A[User: ship #N] --> AA{resume?}
AA -->|existing state| AB[resume from saved phase]
AA -->|fresh| B[assign issue + in-progress label]
AB --> B
B --> C[implement agent]
C --> D{diff gate}
D -->|no changes| E[blocked: comment on issue, remove label]
D -->|changes exist| F[pre-ship checks]
F --> G[self-review locally]
G --> H{findings?}
H -->|critical/important| I[fix + recommit]
I --> G
H -->|clean| J[push + gh pr create --draft]
J --> K{CI checks}
K -->|pass| L{repo type?}
K -->|fail| N[report failures, offer to fix]
K -->|running| O[report status, check back later]
N --> I
L -->|personal + user says merge| P[gh pr merge --squash]
L -->|team| Q[report draft PR link]
For multiple issues ("ship #1, #2, #3"), the router dispatches worktree-worker agents in parallel, each in an isolated git worktree.
What's on flow
Scoped to the current repo by default. Checks OWNERS files for component ownership.
graph TD
A[User: what's on] --> B[detect repo]
B --> C[query issues + PRs assigned/requesting review]
B --> D[check OWNERS file]
D -->|user is owner| E[query unassigned issues + open PRs]
D -->|not owner| F[skip]
B --> G[query Jira if available]
C & E & F & G --> H{group by priority}
H --> I[address feedback]
H --> J[review requested]
H --> K[merge ready]
H --> L[my PRs awaiting review]
H --> M[implement]
H --> N[component owner]
H --> O[Jira]
I & J & K & L & M & N & O --> P[present prioritised table]
P --> Q[suggest top action]
Typical commands
"Pluck" -- invokes the pluck skill. Shows unassigned issues in the current repo. Pick ones to claim without starting implementation.
"What's on?" -- invokes the next skill. Queries GitHub for issues, PRs, and feedback in the current repo. Returns a prioritised table.
"Ship #42" -- invokes the ship skill. Implements, pushes, creates PR, self-reviews, fixes findings, reports back.
"Review this PR" -- classifies files, dispatches specialist reviewers in parallel, collects findings, drafts PR comment.
"Triage this issue" -- dispatches the triage agent. Assesses readiness, recommends workflow (implement, refine, split, or human review).
"This issue is vague" -- dispatches the refine agent. Produces a structured spec with testable acceptance criteria.
"Ship #1, #2, #3" -- dispatches worktree-worker agents in parallel, each in its own worktree. Collects results and presents a summary table.
"Create an issue" -- invokes the issues skill. Creates issues with acceptance criteria, manages state, links PRs.
Agents
| Agent | Purpose |
|---|---|
| router | Task intake, classification, delegation. Coordinates review fanout. Never writes code. |
| implement | Takes a well-defined issue, writes code, runs tests, commits |
| code-reviewer | General code quality: correctness, readability, architecture, naming |
| security-auditor | Security review: injection, auth bypasses, secrets, crypto, OWASP |
| go-k8s-reviewer | Go idioms, concurrency, controller patterns, CRD conventions, RBAC |
| auth-reviewer | OAuth2/OIDC flows, token handling, policy evaluation, standards compliance |
| triage | Assesses issue readiness, labels, prioritises, recommends workflow |
| refine | Turns vague issues into implementable specs with acceptance criteria |
| address-feedback | Reads PR review comments, categorises, fixes, reports what needs human input |
| release-notes | Generates grouped release notes between git tags |
| test-writer | Finds coverage gaps, writes targeted tests matching project patterns |
| test-verifier | Verifies PR test plans: runs tests, checks criteria, drives browser for UI checks |
| docs | Writes and updates documentation. Verifies every example and path. |
| worktree-worker | Self-contained implement-to-PR in an isolated worktree. For parallel multi-issue dispatch. |
Skills
Clawdio provides its own skills for SDLC orchestration. For cross-cutting development practices (TDD, debugging, security hardening, code review), it leans on the companion agent-skills plugin. Design doc workflows and feature lifecycle come from dev-team-plugin (kdt).
Clawdio skills
| Skill | Trigger | Args | Purpose |
|---|---|---|---|
| next | "what's on?", "what next?" | none | Scans GitHub and Jira for issues, PRs, and feedback across repos |
| ship | "ship #42" | <issue>, --resume, --skip-review, --ready | Full lifecycle: implement > push > draft PR > self-review > fix |
| pluck | "pluck", "claim issue", "grab issue" | none | Claim unassigned issues from the repo backlog without implementing |
| pr-description | Creating a PR | none | PR body template: summary, linked issue, test evidence |
| issues | "create issue", "update issue" | create, update, close, link, --repo | Create, update, close issues. Manages PR-issue links and lifecycle state. |
| doc-sync | "check docs", "are docs up to date" | none | Verify and fix documentation accuracy against actual repo contents |
| review-coordination | PR review dispatch | none | Coordinates multi-specialist PR review fanout |
| merge-gate | pre-merge checks | none | Pre-merge safety checks before any merge |
| worktree-recovery | worktree recovery | none | Recovers in-progress worktree workers before dispatching new ones |
| parallel-ship | "ship #1, #2, #3" | <issues> | Dispatches multiple worktree-workers in parallel for multi-issue ship |
agent-skills (companion plugin)
Clawdio agents invoke agent-skills skills at key workflow points. These provide the development discipline that clawdio orchestrates.
| agent-skills skill | Used by | When |
|---|---|---|
| code-review-and-quality | code-reviewer, go-k8s-reviewer | Five-axis review (correctness, readability, architecture, security, performance) |
| code-simplification | code-reviewer | Identify simplification opportunities in reviewed code |
| security-and-hardening | security-auditor, auth-reviewer | OWASP, secrets, auth/authz checks |
| test-driven-development | implement, worktree-worker, test-verifier, test-writer | RED-GREEN-REFACTOR loop |
| incremental-implementation | implement, worktree-worker, address-feedback | One logical change per commit, vertical slices |
| debugging-and-error-recovery | implement, worktree-worker, address-feedback | When tests fail and cause is unclear |
| spec-driven-development | implement, refine | Spec-first for non-trivial changes |
| planning-and-task-breakdown | refine, triage | Break scope into ordered tasks |
| api-and-interface-design | go-k8s-reviewer, auth-reviewer | API surface stability, Hyrum's Law checks |
| performance-optimization | go-k8s-reviewer | N+1 queries, unbounded ops, async patterns |
| git-workflow-and-versioning | ship, worktree-worker, address-feedback | Commit conventions, branch hygiene |
| shipping-and-launch | ship | Pre-ship checklist in phase 2 |
| documentation-and-adrs | docs | ADR structure and documentation patterns |
| browser-testing-with-devtools | test-verifier | Drive browser for UI verification |
kdt (companion plugin)
The router dispatches dev-team-plugin skills for design and feature lifecycle work.
| kdt skill | Trigger | Purpose |
|---|---|---|
| feature-design | "design doc", "feature design" | Create and review design docs, generate issues from TODOs |
| feature-implement | "pick up", "implement from design" | Pick up issues from design docs, manage feature lifecycle |
| pr-closes-issue | "does the PR close the issue" | Verify PR changes match issue requirements |
| external-contribs | "external contribs", "community PRs" | Find external contributions needing attention |
Hooks
| Hook | Trigger | Purpose |
|---|---|---|
| block-env-writes | Before Write/Edit | Blocks writes to .env, credentials, .pem, .key files |
| doc-sync-reminder | After Write/Edit | Reminds to update docs when agent/skill/hook files change |
| format-on-save | After Write/Edit | Runs project formatter if configured (prettier, gofmt, clang-format) |
| lint-on-edit | After Write/Edit | Runs project linter if configured (eslint, golangci-lint) |
Structure
agents/ subagent definitions (one .md per agent)
skills/ on-demand skills (SKILL.md per directory)
hooks/ lifecycle hooks (hooks.json)
references/ supporting docs agents can read
docs/ architecture decisions and project context
.claude-plugin/ plugin manifest and marketplace config
Personalisation
The go-k8s-reviewer and auth-reviewer ship with generic definitions suitable for any Go/K8s or auth project. For domain-specific review depth, override them with personal versions in ~/.claude/agents/ -- personal agents take precedence over plugin agents.
Personal agent overrides stay out of this repo. They're your competitive advantage, not a shared concern.
Development
See docs/contributing.md for how to write agents, skills, and hooks.
Edit, test locally with claude --plugin-dir ., push, then claude plugin update clawdio@jasonmadigan-clawdio.
Design
See docs/architecture.md for the full design rationale, including why this is a plugin and not an orchestrator, the three-tier primitive location model, and future Agent SDK migration path.
See docs/grill-findings.md for the structured interview that informed these decisions.