Developer Skill Pack
dennisonbertram/developer-skill-pack
Toolkit completo: UX testing, multi-agente, TDD e scaffolding repo
Ver no GitHubKit completo de desenvolvimento para programação assistida por IA, reunindo testes de UX, coordenação multi-agente, workflow TDD, templates GitHub, research spikes e scaffolding de repositórios. Integra-se com Claude Code e OpenAI Codex CLI para automatizar ciclos inteiros de desenvolvimento.
Instalação
npx convez-ai dennisonbertram/developer-skill-pack1 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
- Testes de jornada do usuário com swarms de sub-agentes
- Coordenação autônoma de backlog (grooming e implementação de issues)
- Automação de workflow TDD com hooks preventivos
- Research spikes estruturados e documentados
- Scaffolding inicial de repositórios com templates e CI
Veredito

Excelente para times que querem automação pesada em todo o ciclo de desenvolvimento. Requer experiência com shell scripting, GitHub Actions e orquestração de agentes — não é para iniciantes, mas oferece ROI alto em produtividade para equipes maduras.
README
A tradução para o português deste README ainda está na fila.
Developer Skill Pack
A complete development toolkit for AI-assisted coding — UX testing, multi-agent coordination, TDD workflow, GitHub templates, research spikes, and repo scaffolding.
Works with Claude Code and OpenAI Codex CLI.
What's Inside
UX Testing (from ux-toolkit)
| Skill | Description |
|---|---|
| ux-paths | Generate exhaustive user journey stories by analyzing the codebase. Produces short, medium, and long stories with a swarm of parallel sub-agents. |
| ux-walker | Walk the UX story catalog through a real browser, testing each journey for correctness, visual quality, and UX excellence. Auto-fixes small issues, files GitHub issues for larger ones. |
| walk-the-issues | Groom all open GitHub issues, then loop through them — branching, researching, implementing with swarms, testing, and creating PRs until all issues are complete. |
| deploy-check | Instrument a frontend with version fingerprinting (meta tag + /__version endpoint) so LLM-based testing can verify the correct build is deployed before walking. |
Coordination (from claude-coordinator)
| Skill | Description |
|---|---|
| coordinator | Pure-delegation control plane with a 10-phase state machine. Plans, delegates, reviews, and learns. 14 specialized agents with JSON Schema output contracts. |
| issue-groomer | Autonomous backlog groomer. Claims open status-less issues, researches each deeply (codebase + product docs + external APIs), fills the target repo's issue template exhaustively, and applies status:ready (clear path) or status:blocked (exhaustively groomed, path-decision needed). Loops; WATCH mode re-checks every poll_interval. Sequential v1 — one instance per repo. |
| issue-implementer | Autonomous backlog agent. Claims status:ready issues, drives each to a tested PR (red → green → regression commits), then loops until the backlog is clear or a stop condition trips. Sequential v1 — one instance per repo. |
| mine-transcripts | Sweep sub-agent JSONL transcripts for novel learnings and promote the strongest to durable docs. Parallel mining with dedupe baseline. |
Development Process (from partyline + synthetix)
| Skill | Description |
|---|---|
| setup-repo | Scaffold a repo with GitHub issue templates, PR template, TDD hooks, process docs, and CI configuration. |
| debug | Deterministic 7-phase debugging workflow — symptom capture through regression test commit. |
| regression-guard | Post-fix test-coverage analysis — generates unit and regression tests, audits how the bug could have been caught earlier. |
| research-spike | Time-boxed research with Context7 integration — produces structured findings in docs/research/. |
Hooks
| Hook | Event | Description |
|---|---|---|
| tdd-reminder | PreToolUse (Edit/Write) | Warns when modifying src/ without a failing test first |
| session-start | SessionStart | Announces TDD-first requirement |
Launcher shortcuts (claude-groom / claude-implement)
claude-groom and claude-implement are shell shortcuts that invoke cco
(the Claude Code sandbox wrapper) with the correct Gamut proxy settings and
agent names — no flags to remember, no token in any script.
How it works
Both names are symlinks to bin/claude-coord in this repo. The script reads
your Gamut auth token at runtime from one of two secure locations (never
hardcoded), maps its invocation name to the right agent, and execs cco.
Install the symlinks
mkdir -p ~/.local/bin ~/.config/cco-gamut
chmod 700 ~/.config/cco-gamut
ln -sf /path/to/developer-skill-pack/bin/claude-coord ~/.local/bin/claude-groom
ln -sf /path/to/developer-skill-pack/bin/claude-coord ~/.local/bin/claude-implement
Replace /path/to/developer-skill-pack with the absolute path where you cloned
this repo (e.g. /Users/you/develop/developer-skill-pack).
Make sure ~/.local/bin is on your PATH. Add to ~/.zshrc if needed:
export PATH="$HOME/.local/bin:$PATH"
Provide your token securely (after rotating any old token)
Option 1 — environment variable (per-session or in your shell profile):
export GAMUT_AUTH_TOKEN='<your-rotated-token>'
Option 2 — token file (persists across sessions):
echo '<your-rotated-token>' > ~/.config/cco-gamut/token
chmod 600 ~/.config/cco-gamut/token
The token is read at runtime. It is never written into any script or committed
to this repo. The test suite (test/launcher.test.sh) will fail if a token
literal ever appears in bin/claude-coord.
Usage
Run from inside the target repo (agents operate on the current working
directory unless you pass target_repo):
# Groom the issue backlog (all open status-less issues)
claude-groom
# Cautious first run (one issue only)
claude-groom -p "max_issues_per_run=1"
# Work the backlog to tested PRs
claude-implement
# One issue only
claude-implement -p "max_issues_per_run=1"
Any extra arguments are forwarded directly to cco.
User-scope agent requirement
claude-groom uses --agent issue-groomer and claude-implement uses
--agent issue-implementer. These agents must be installed at user scope
(~/.claude/agents/) — install them with:
npx skills add dennisonbertram/developer-skill-pack
Or use the plugin-namespaced form by editing bin/claude-coord to pass
--agent developer-skill-pack:issue-groomer (after claude plugin install).
Installation
Install as a Claude Code plugin (agents + skills)
The plugin installs all 20 coordinator agents as developer-skill-pack: namespaced subagents plus all skills. Use this path if you want --agent developer-skill-pack:issue-groomer style dispatch and persistent plugin management.
Step 1 — Add this repo as a marketplace source (one-time per machine):
claude plugin marketplace add dennisonbertram/developer-skill-pack
Step 2 — Install the plugin:
claude plugin install developer-skill-pack@developer-skill-pack
Use the namespaced agents:
# Groom the issue backlog
claude --agent developer-skill-pack:issue-groomer -p "groom all open issues"
# Work the backlog to tested PRs
claude --agent developer-skill-pack:issue-implementer
# Run the coordinator control plane
claude --agent developer-skill-pack:coordinator -p "plan the sprint"
# Other agents: worker, reviewer, planner, briefer, researcher, scribe,
# distiller, knowledge-packager, learning-extractor, system-tester,
# ui-tester, ux-tester, worker-investigation, worker-refactor, worker-test,
# evidence-auditor, intent-validator
Note: user-scope agents in ~/.claude/agents/ take precedence over plugin agents. If you have any of these agents installed at user scope (e.g. via npx skills add), use the namespaced form developer-skill-pack:agent-name to invoke the plugin version explicitly.
Local session-only load (no install required):
claude --plugin-dir /path/to/developer-skill-pack --agent developer-skill-pack:issue-groomer -p "say READY"
Install skills only via npx skills (recommended for most users)
npx skills add dennisonbertram/developer-skill-pack
Or install specific skills:
npx skills add dennisonbertram/developer-skill-pack --skill coordinator debug setup-repo
This installs skills (invoked as /skill-name) and copies the coordinator agent .md files to your user-scope ~/.claude/agents/ as unnamespaced agents. Use this path for the simplest setup. To use agents after skills install:
# Unnamespaced (from user scope after npx skills install)
claude --agent issue-groomer
# Namespaced (only available after plugin install)
claude --agent developer-skill-pack:issue-groomer
OpenAI Codex CLI
# Clone and copy skills to your project
git clone https://github.com/dennisonbertram/developer-skill-pack
cp -r developer-skill-pack/skills/ .agents/skills/
cp developer-skill-pack/codex/AGENTS.md AGENTS.md
cp developer-skill-pack/codex/config.toml .codex/config.toml
cp developer-skill-pack/hooks/ .codex/hooks/
Manual (any agent)
git clone https://github.com/dennisonbertram/developer-skill-pack
# Copy skills/ to your agent's skill directory
# Copy hooks/ for TDD enforcement
# Copy docs/process/ for development workflow documentation
Quick Start
1. Scaffold a new repo
/setup-repo
Creates GitHub issue templates, PR template, hooks, and process docs.
2. Plan and coordinate work
/coordinator
Enters the 10-phase state machine: intake → plan → delegate → integrate → review → test → validate → close.
3. Groom the backlog
--agent issue-groomer
Claims open status-less issues one at a time → researches each deeply → fills the issue template exhaustively → applies status:ready or status:blocked → loops until the backlog is groomed (or max_issues_per_run is reached). Unattended after invocation. See the issue-groomer runbook for preconditions, label setup, stop conditions, and first-run guidance.
4. Work the backlog autonomously
--agent issue-implementer
Claims status:ready issues one at a time → TDD implementation → tested PR → loops until the backlog is clear (or max_issues_per_run is reached). Unattended after invocation. See the issue-implementer runbook for preconditions, label setup, stop conditions, and first-run guidance.
5. Generate UX stories
/ux-paths
Analyzes the codebase and produces a catalog of user journey stories.
6. Add deploy verification to a frontend
/deploy-check --install
Instruments the app with a <meta name="build-version"> tag and a /__version JSON endpoint so ux-walker can confirm the correct build is deployed before testing.
7. Walk the stories through a browser
/ux-walker http://localhost:3000
Tests each story for correctness, visual quality, and UX excellence. Automatically checks the deploy version first if fingerprinting is installed.
8. Research before adopting
/research-spike "Should we use Prisma or Drizzle for the ORM?"
9. Debug a bug
/debug
7-phase deterministic workflow: symptom → archaeology → reproduce → logs → hypothesis → fix → regression test.
10. Mine session transcripts for learnings
/mine-transcripts
Discovers sub-agent JSONLs, mines them in parallel for dead ends, silently-recovered errors, and design sub-decisions, then promotes high-confidence findings to durable docs.
11. Guard against regressions
/regression-guard
Generates unit and regression tests for the last fix.
Autonomous Backlog Implementation (issue-groomer + issue-implementer)
Two autonomous agents form a continuous pipeline from raw idea to merged, tested PR — with no human involvement between invocation and completion.
Pipeline: Groomer → Implementer
open (no status)
→ [issue-groomer] status:grooming
→ [issue-groomer] status:ready ← HANDOFF LABEL
OR status:blocked (exhaustively groomed, path-decision needed)
OR status:grooming removed + unassigned (failed — operational, auto-retried next run)
status:ready
→ [issue-implementer] status:in-progress
→ PR (Closes #N) → merged
status:ready is the handoff label between the two agents. The groomer writes it; the implementer reads it.
| Agent | Trigger | Output |
|---|---|---|
issue-groomer | Open issues with no status:* label | status:ready (claimable), status:blocked (groomed, decision pending), or groom_status:failed (operational error, auto-retried) |
issue-implementer | Issues labeled status:ready | Tested PR with red → green → regression audit trail; issue closed on merge |
Blocked path: status:blocked means the issue is fully groomed — not half-done. A genuine product/path decision needs a human (or review agent) before implementation can proceed. Once resolved, the issue transitions back to status:ready and the implementer picks it up.
Failed path: groom_status:failed is an OPERATIONAL outcome — a tool error, network failure, or retry-budget exhaustion. It is NOT a product decision and requires NO human action. The groomer releases the claim (removes status:grooming, unassigns) and the issue returns to status-less for automatic retry on the next run. Distinguish from blocked: blocked = human decision needed; failed = operational, auto-retried.
issue-groomer
Claims status-less issues one at a time → researches each deeply (codebase + product docs + external APIs) → fills the issue template exhaustively with behavior contract, verified file paths, assumptions, and alternatives → applies status:ready or status:blocked → loops until the backlog is groomed. Unattended after invocation.
--agent issue-groomer
For a cautious first run (one issue only):
--agent issue-groomer max_issues_per_run=1
V1 constraint: sequential and single-instance only. Never run two issue-groomer instances against the same repo simultaneously.
For the full runbook — preconditions, exact gh label create commands, stop conditions, WATCH mode, kill switch, and first-run guidance — see skills/coordinator/docs/issue-groomer-runbook.md.
issue-implementer
Claims the oldest status:ready issue → grounds itself against the codebase → delegates a TDD worker (red → green → regression commits) → runs review and test passes → opens a PR whose body contains the full DoD checklist, audit-trail commit links, and a collapsible machine-readable report → closes the issue → loops to the next ready issue. Stops when the backlog is clear, max_issues_per_run is reached, or a stop condition trips.
--agent issue-implementer
For a cautious first run (one issue only):
--agent issue-implementer max_issues_per_run=1
V1 constraint: sequential and single-instance only. Never run two issue-implementer instances against the same repo simultaneously.
For the full runbook — preconditions, label setup with exact gh label create commands, stop conditions, kill switch, and first-run guidance — see skills/coordinator/docs/issue-implementer-runbook.md.
Workflow Scripts
The workflows/ directory contains deterministic multi-agent orchestration scripts that leverage Claude Code's Workflow runtime. These provide typed outputs, parallel execution, and adversarial verification — capabilities beyond what prompt-based skills can achieve.
| Workflow | Agents | Pattern | Description |
|---|---|---|---|
| mine-transcripts.js | 5-8 | Discover → parallel pipeline → consolidate | Sweeps sub-agent JSONLs with typed schemas, dedupes across slices, persists to learning docs |
| code-review.js | 11-16 | 5 parallel reviewers → adversarial verify per finding | Reviews across correctness, security, performance, concurrency, and API contracts. Skeptics refute false positives before reporting. |
| ux-walker.js | 3-20+ | Preflight → parallel walks → triage → report | Deploy verification, parallel story walking, structured findings with severity classification |
Using Workflows
Workflows run when you ask Claude Code to "use a workflow" or include "ultracode" in your prompt:
use a workflow for mine-transcripts
Or copy to your project's .claude/workflows/ for direct /command invocation:
cp workflows/mine-transcripts.js .claude/workflows/
# Now available as /mine-transcripts in Claude Code
Key Advantages Over Prompt-Based Skills
- Typed outputs: JSON Schema contracts on every agent call — no parsing, automatic retry on mismatch
- Parallel execution:
pipeline()streams items through stages without barriers;parallel()for fan-out - Adversarial verification: Independent skeptics refute findings before they're reported
- Deterministic control flow: Loops, conditionals, and error handling in JavaScript, not natural language
- Resumable: Completed agents return cached results if the script is re-run
Codex CLI Coordinator
The coordinator pattern works with Codex CLI via its MCP server mode. Codex supports:
- Hooks: Same lifecycle events as Claude Code (SessionStart, PreToolUse, PostToolUse, Stop)
- Skills: First-class
SKILL.mdsupport with progressive disclosure - AGENTS.md: Hierarchical instructions (equivalent to CLAUDE.md)
- Subagents: Native spawning with
max_threadsandmax_depthconfig - MCP Server:
codex mcp-serverexposescodexandcodex-replytools for external orchestration
See codex/ directory for Codex-specific configuration files.
Known Codex Limitations
type: "agent"hook handlers are parsed but not yet functionalasync: trueon hooks is parsed but skipped- Automatic delegation (
delegation_mode = "skill_auto") is not yet implemented (tracked in openai/codex#18193) - Project-local hook discovery may require global config in some versions
GitHub Issue Templates
The setup-repo skill installs three structured issue templates:
Feature Slice
Behavior contracts, design specs, TDD audit trail, agent todo checklist, regression risks, deploy impact, documentation handoff.
Bug Regression
Reproduction steps, regression test plan (smallest test + behavior proof), blast radius, TDD audit trail.
Research Spike
Question, options to compare, Context7 documentation sources, agent todo checklist, expected output.
Development Process
The skill pack enforces a behavior-first TDD workflow:
- RED — Write a failing test that defines the desired behavior
- GREEN — Write the minimum code to make the test pass
- REFACTOR — Clean up while keeping tests green
Every feature/bugfix produces three auditable commits:
test(red): TASK-042 failing tests for rate-limit
feat: TASK-042 implement per-IP rate limiting
test(regression): TASK-042 regression coverage
See docs/process/ for the full workflow documentation.
Directory Structure
developer-skill-pack/
├── .claude-plugin/
│ ├── plugin.json # Plugin manifest
│ └── marketplace.json # Marketplace entry (enables /plugin marketplace add)
├── agents/ # Symlinks -> skills/coordinator/agents/ (plugin agent exposure)
├── skills/
│ ├── ux-paths/ # User journey story generator
│ │ ├── SKILL.md
│ │ └── references/
│ ├── ux-walker/ # Browser-based UX testing
│ │ ├── SKILL.md
│ │ ├── references/
│ │ └── templates/
│ ├── walk-the-issues/ # Issue grooming + implementation loop
│ │ ├── SKILL.md
│ │ └── references/
│ ├── coordinator/ # Multi-agent orchestration
│ │ ├── SKILL.md
│ │ ├── agents/ # 14 agent definitions
│ │ ├── schemas/ # JSON Schema output contracts
│ │ └── templates/
│ ├── deploy-check/ # Version fingerprinting for deploy verification
│ │ └── SKILL.md
│ ├── mine-transcripts/ # Session transcript mining
│ │ └── SKILL.md
│ ├── setup-repo/ # Repo scaffolding
│ │ ├── SKILL.md
│ │ └── references/ # GitHub templates, hooks, PR template
│ ├── debug/ # 7-phase debugging workflow
│ │ └── SKILL.md
│ ├── regression-guard/ # Post-fix test coverage
│ │ └── SKILL.md
│ └── research-spike/ # Structured research
│ └── SKILL.md
├── workflows/ # Deterministic multi-agent workflow scripts
│ ├── mine-transcripts.js # Parallel transcript mining with dedupe
│ ├── code-review.js # 5-dimension review + adversarial verify
│ └── ux-walker.js # Parallel story walking with deploy check
├── hooks/ # TDD enforcement hooks
│ ├── tdd-reminder.sh
│ └── session-start.sh
├── codex/ # OpenAI Codex CLI configuration
│ ├── AGENTS.md
│ └── config.toml
├── docs/
│ └── process/ # Development workflow documentation
│ ├── development-workflow.md
│ ├── github-build-process.md
│ ├── behavior-tdd.md
│ └── regression-discipline.md
├── skills.sh.json # skills.sh display configuration
├── README.md
└── LICENSE
Sources
This skill pack combines and extends:
- ux-toolkit — UX story generation, browser testing, issue implementation
- claude-coordinator — Multi-agent orchestration with JSON Schema contracts
- partyline — GitHub templates, TDD hooks, development process, learning system
- synthetix — Behavior-first TDD, issue templates with agent checklists
- agent-university — Hook architecture research, fleet-bus substrate, Codex compatibility
License
MIT — see LICENSE.