jjcfatras-tools — Marketplace de Skills Claude
jjcfatras/Claude
16 skills prontos para Git, testes, review e docs no Claude Code
Ver no GitHubColeção de 16 skills (slash commands) em Go para potencializar fluxos de Git, testes, code review, documentação e raciocínio no Claude Code. Inclui automação de cherry-pick, merge, commits inteligentes, test-driven fixes, revisão multi-especialista de PRs e auditoria de documentação.
Instalação
npx convez-ai jjcfatras/Claude1 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
- Automatizar workflows Git complexos (cherry-pick, merge com resolução de conflitos)
- Executar code reviews multi-especialista (segurança, TypeScript, React, infra, perf)
- Corrigir bugs via loop autônomo de patch → teste → revert
- Auditar documentação e arquitetura em busca de informações desatualizadas
- Facilitar debate adversarial e argumentação estruturada de claims técnicos
Veredito

Portfolio maduro de skills para engenheiros Go/Claude que buscam automatizar operações Git e code review em escala. Bem documentado com exemplos de CLI claros, mas exige familiaridade com Claude Code e fluxos de desenvolvimento estruturados.
README
A tradução para o português deste README ainda está na fila.
jjcfatras-tools — Claude Code marketplace
A Claude Code plugin marketplace shipping sixteen slash commands the author uses for everyday Git, testing, code-review, documentation, and reasoning workflows.
Install
/plugin marketplace add jjcfatras/Claude
/plugin install <plugin-name>@jjcfatras-tools
Plugins
| Plugin | Slash command | What it does |
|---|---|---|
cherry-pick | /cherry-pick <source-branch> [commit-sha or sha1..sha2] | Cherry-picks one or more commits from a source branch into the current branch and resolves conflicts intelligently. |
merge | /merge <source-branch> | Merges a source branch into the current branch with conflict resolution. |
git | /git:commit · /git:commit-push · /git:commit-push-pr · /git:clean_gone | Git workflow: auto-message commit, commit + push (refreshes an open PR), commit + push + open PR, and prune local [gone] branches. |
test-driven-fix | /test-driven-fix <spec-or-bug> | Autonomous patch → test → revert-on-regression loop, hard-capped at 10 iterations. |
respond-to-review | /respond-to-review <pr-number> [comment-id] | Triages every flagged issue on a PR — inline comments and review-body findings — dismissing false positives and fixing valid ones. |
code-review-AT | /code-review-AT [pr-number] | Multi-specialist PR review (security, typescript, react, infra, errors, perf, quality, claude-md) coordinated via a sub-agent team. Posts inline comments. |
code-review | /code-review [pr-number] | Same multi-specialist PR review using parallel native Claude Code subagents — no Agent SDK, no agent team, no cross-agent verification. Posts inline comments. |
doc-audit | /audit-docs | Scans CLAUDE.md / READMEs / .claude/commands / .claude/skills / architecture docs for stale claims about the codebase and reports findings with suggested fixes. |
debate | /debate <claim> | Adversarial pro/con debate — opening arguments, then up to 5 rounds of attack/defend, then an inline markdown report of surviving, negated, and disputed arguments. |
simplify | /simplify [path|--staged|--since=<ref>] | Proposes targeted, behavior-preserving simplifications to recently modified code; shows diffs per file and applies only on approval. |
transcript | /transcript | Prints the filepath of the current Claude Code session's .jsonl transcript, with size and line count. |
jira-ticket | /jira-ticket | Generates a structured JIRA ticket (summary, acceptance criteria, QA testing steps) from a diff, PR, or description and files it in JIRA after confirmation. |
jira-implement | /jira-implement <JIRA-key> | Picks up a JIRA ticket by key, verifies every claim it makes about the codebase, designs and reconciles an implementation plan, presents it for approval, then implements it. |
Install only the plugins you want — each is independent.
Three additional hook-only plugins ship no slash command: tool-discipline (durable PreToolUse guardrails — no-cd-chaining, prefer-builtin-tools), tool-discipline-lsp (the prefer-LSP PreToolUse guardrail plus a PostToolUse advisory that nudges a retry when workspaceSymbol returns empty), and grep-glob-fallback (temporary ToolSearch shim for the Grep/Glob outage, deletable once those tools are restored).
Using the plugins
Each subsection below covers how to invoke the plugin, what to have ready first, and what to expect step by step. The full flow lives in each plugin's command file under plugins/<name>/commands/.
/cherry-pick
Invoke: /cherry-pick <source-branch> [commit-sha or sha1..sha2]
Prereqs: Clean working tree (commit or stash first); source branch exists locally or as origin/<branch>.
What happens:
- Preflight: validates the working tree is clean and the source branch exists.
- Determines commits to apply — uses the SHA / range you passed, or lists the 15 most recent commits on the source branch and asks you to pick.
- Shows a summary (target, source, commit list) and asks you to confirm.
- Applies commits one at a time in chronological order.
- On conflict: reads each conflicted file, resolves it by combining intent from both sides, strips conflict markers,
git adds, then runsgit cherry-pick --continue. - Reports a final
git logsummary and lists any conflicts that were resolved.
Escape hatch: git cherry-pick --abort restores the original state if you want out mid-run.
/merge
Invoke: /merge <source-branch>
Prereqs: Clean working tree; source branch exists locally or as origin/<branch>.
What happens:
- Preflight: validates the working tree is clean and the source ref resolves.
- Classifies the merge as already up to date, fast-forward, or divergent (merge commit) and shows the incoming commits.
- Asks you to confirm.
- Runs
git merge <source-ref>with no flags — git picks fast-forward vs. merge commit based on history. - On conflict: same auto-resolution flow as
/cherry-pick, finishing withgit merge --continue. - Reports a final
git log --graphsummary and agit diff ORIG_HEAD..HEAD --stat.
Escape hatch: git merge --abort.
/test-driven-fix
Invoke: /test-driven-fix <spec-path-or-bug-description> — a path that resolves to an existing file is treated as a spec; anything else is treated as a free-text bug description.
Prereqs: A detectable test stack — package.json, pyproject.toml / pytest.ini, Cargo.toml, go.mod, or a Makefile exposing test / lint / typecheck. A dirty working tree is auto-stashed under tdf-baseline before the loop starts.
What happens:
- Detects test/lint/typecheck commands from the project metadata.
- Runs the baseline and parses failures into a tracked task list.
- Iterates up to 10 times: locate the symbol → propose a minimal patch → narrow re-run → full re-run → revert any patch that regresses a previously-green test → repeat. Never prompts mid-loop.
- On full green: stages the touched files and creates a
fix(<scope>): …commit with a body listing the failures that moved red → green. - On exhaustion (10 iterations, still red): leaves best-effort patches in the working tree and does not commit. The baseline stash is preserved so you can
git stash show -p stash@{…}to diff.
/respond-to-review
Invoke: /respond-to-review <pr-number> [comment-id] — passing a comment ID scopes the run to one inline thread and skips review-body parsing.
Prereqs: gh CLI authenticated for the repo; the PR exists.
What happens:
- Fetches the PR diff, all inline comments, and all review bodies.
- Filters to actionable items only — drops replies, your own comments, anything you've already replied to, and trivial acks like "LGTM".
- Parses review bodies into discrete findings (one per bullet / heading / paragraph).
- Triages each item as false positive, preexisting code (not introduced by this PR), or valid issue.
- Implements fixes for the valid items and replies confirming the change; replies to the others with an explanation dismissing the finding.
/code-review-AT
Invoke: /code-review-AT [pr-number] — omit the argument to review the PR for the current branch.
Prereqs: export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 must be set before launching Claude Code (see Requirements below); /tmp must be writable for the agent's Write tool, or grant Write to $HOME/.claude/tmp/pr-review-* as a fallback.
What happens:
- Preps the PR diff, prior reviews, project
CLAUDE.md, and a one-paragraph PR summary into a temp workspace under/tmp/pr-review-…. - Picks an applicable subset of specialists — security, typescript, react, infra, errors, perf, quality, claude-md — and spawns one subagent per category in parallel.
- Specialists scan their domain and cross-verify with peer DMs; findings are written to a shared task list.
- The Go
code-review-helperfinalizes, dedupes, and gates findings, then assembles the review payload. - You're shown the inline + summary findings and asked to approve before anything is posted.
- On approval: posts the review with inline comments, then cleans up the temp workspace.
/code-review
Invoke: /code-review <pr-number> — the PR number is required (the command treats an absent or non-integer argument as a hard error).
Prereqs: gh CLI authenticated for the repo; /tmp (or $TMPDIR) writable for the scratch workspace under pr-review-<number>-<epoch>/. Unlike /code-review-AT, this plugin does not require CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 — it uses native Claude Code subagents (Agent with subagent_type) only.
What happens:
- Fetches PR metadata, the full diff, and the most recent prior Claude-Code review (for dedup) via
gh. - Parses the diff with the bundled Go
code-review-helperand builds the specialist roster — always-on:security,quality,errors,perf; conditional by changed-file extension/path:typescript(.ts/.tsx/.cts/.mts),react(.tsx/.jsxor component/pages paths),infra(.sql,.tf,.hcl,Dockerfile,docker-compose,k8s/,terraform/, …). - The helper synthesizes the
## Summarysection deterministically from changed-file metadata — no pre-pass subagent runs. - Spawns all roster specialists in parallel against a shared
spawn-context.mdbundle and a separaterubric.md. - The helper finalizes — dedup, severity/confidence gating, line-snapping, payload rendering — and you're shown the findings summary with a
Post review? [Y]es/[n]o/[i]ds <csv>prompt. You can post all, skip, or filter to specific finding IDs. - On approval: posts via a three-tier fallback (batched review → pending-then-submit → plain PR comment), then removes the scratch dir.
Hook: the plugin bundles a PreToolUse hook that auto-approves Grep and Glob so specialist subagents never stall on permission prompts during parallel scans.
/audit-docs
Invoke: /audit-docs (no arguments).
Prereqs: Run inside a git repo. Outside of one, the command falls back to the current working directory and warns that path-relative claim verification is less reliable.
What happens:
- Locates every doc file matching
CLAUDE.md,README.md,.claude/commands/*.md,.claude/skills/*/SKILL.md, and*[Aa]rchitecture*.md(skippingnode_modules,dist, etc.). - If more than 50 files match, lists them and asks whether to proceed, narrow scope, or skip directories.
- Extracts only concrete claims from each file — file paths, versions, scripts, symbol names, cross-doc links — not subjective prose.
- Verifies each claim against the current codebase.
- Reports findings grouped by source file with suggested fixes.
- Offers to apply fixes one at a time. Read-only until you approve a specific fix.
/simplify
Invoke: /simplify [path|--staged|--since=<ref>] — no argument scopes to files modified in this session plus the working tree / staged set.
Prereqs: Inside a git repo for --staged / --since= modes. The plugin's PostToolUse formatter hooks (Prettier, gofmt) handle reformatting after edits, so don't run formatters manually.
What happens:
- Resolves scope per the argument (path/glob,
--staged,--since=<ref>, or default session + working tree). - Drops generated files, binaries, and whitespace-only changes from the candidate set; lists the survivors and asks you to confirm.
- Loads the root
CLAUDE.md, nearest non-rootCLAUDE.mdancestors per file, and the plugin's four-pillarstandards.mdrubric. - Analyzes each file and proposes hunks — unified diff + one-sentence rationale citing the pillar. Hard guardrails drop hunks that change behavior, alter public API, or duplicate formatter work.
- For each file with surviving hunks: shows diffs and prompts
apply all/apply some/skip file/edit and apply. - Applies approved hunks via
Edit. Never commits — that's your call.
Notes: single-threaded by design (no subagents); no --auto mode — the propose-then-apply checkpoint is the whole point.
/transcript
Invoke: /transcript (no arguments).
Prereqs: CLAUDE_CODE_SESSION_ID must be set (Claude Code exports it automatically).
What happens:
- Resolves the transcript path as
$HOME/.claude/projects/<encoded-cwd>/$CLAUDE_CODE_SESSION_ID.jsonl, where<encoded-cwd>replaces every/and.in$PWDwith-. - If the expected file is missing but the encoded directory exists, falls back to the most-recently-modified
*.jsonlin that directory and labels the outputpath (fallback):. - Prints three lines: the path, size in MB, and line count. Nothing else.
Use this to hand the file to /plugin-session-auditor or any other transcript-consuming workflow.
/jira-implement
Invoke: /jira-implement <JIRA-key> (e.g. /jira-implement PROJ-1234).
Prereqs: Authenticated Atlassian MCP access (the command resolves the site via getAccessibleAtlassianResources and fetches the ticket via getJiraIssue); run inside the repo the ticket targets.
What happens:
- Fetch the ticket — description, comments, linked issues, and acceptance criteria (later comments weighted over the original description on conflict).
- Understand intent — distills the underlying goal and sorts the ticket into three buckets: goal/problem, claims about the codebase, and any proposed solution.
- Verify every codebase claim against the actual current code (Grep/Glob/Read/LSP), classifying each as confirmed / false / partial / unverifiable with
file:lineevidence. If a false claim knocks out the ticket's premise, it stops and surfaces the finding instead of building on a phantom. - Design a plan independently from the goal and verified facts first, then reconciles it against the ticket's proposed solution (if any), recording where each approach won.
- Present for approval — intent, claim-verification table (false claims headlined), the synthesized plan, and reconciliation rationale. Gates with
AskUserQuestion; edits no file before you approve. - Implement the approved plan — test-first for testable changes, directly for config/docs — then runs the existing suite and reports results honestly.
code-review-AT — extras
- Bundles a Go helper (
code-review-helper) used to deterministically parse diffs and assemble review payloads. The plugin ships prebuilt binaries fordarwin-amd64,darwin-arm64,linux-amd64, andlinux-arm64; abin/code-review-helpershell wrapper dispatches to the right one. - Installs eight
code-review-*review specialists into the team that runs the review.
Requirements
/code-review-AT orchestrates multiple specialist subagents via Claude Code's agent-team APIs (TeamCreate, SendMessage, concurrent Agent spawns). Those tools are gated behind an experimental flag — without it, the command's preflight aborts with an allowlist error. Enable agent teams in your shell before running /code-review-AT:
export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
The other plugins (cherry-pick, merge, git, test-driven-fix, respond-to-review, doc-audit, debate, simplify, transcript, jira-ticket, jira-implement) do not need this flag.
Building the helper from source
cd "${CLAUDE_PLUGIN_ROOT}/tools/code-review-helper"
make release # cross-compile all 4 platforms into ../../bin/
make test
make release is what the author runs before tagging a new plugin version. End users do not need a Go toolchain.
Maintainer scripts
All cross-plugin scripts live in the repo-root package.json and are invoked with pnpm <script>. End users installing the plugins via the marketplace never need these — they exist for contributors editing this repo.
| Script | Command | Purpose |
|---|---|---|
pnpm install | Installs dependencies; also fires prepare. | Standard install. |
pnpm format | prettier --write . | Format every file the prettier config matches. Uses prettier-plugin-sh for shell. |
pnpm format:go | for d in plugins/code-review/tools/code-review-helper plugins/code-review-AT/tools/code-review-helper .claude/skills/plugin-session-auditor/tools/session-parser; do gofmt -w "$d" && go -C "$d" mod edit -fmt || exit 1; done | gofmt -w and go mod edit -fmt across all three Go modules (both code-review helpers + plugin-session-auditor session-parser). |
pnpm build:go | for d in plugins/code-review/tools/code-review-helper plugins/code-review-AT/tools/code-review-helper; do make -C "$d" release || exit 1; done | make release for both code-review helpers — cross-compiles darwin/linux × amd64/arm64 prebuilts into each plugin's bin/. Skips session-parser (no prebuilt shipped). |
pnpm check-types | tsc --noEmit | TypeScript type-check using the root tsconfig.json (also extended by plugins/code-review-AT/tsconfig.json). |
pnpm test | echo "Error: no test specified" && exit 1 | Stub — no JS/TS test suite. Go tests live in each helper's make test. |
pnpm prepare | husky | Installs the Husky git hooks (runs automatically after pnpm install). The repo's pre-commit runs pnpm exec lint-staged; the lint-staged config lives at lint-staged.config.mjs. |
To build a single Go helper, run make release (or make test) directly from inside plugins/code-review-AT/tools/code-review-helper/ or plugins/code-review/tools/code-review-helper/.
Repo-internal skills
The repo also ships skills under .claude/skills/. These are not part of the marketplace and are not installed by /plugin install — they activate only when Claude Code is run inside a clone of this repo. They exist for maintainers and contributors working on the plugins themselves.
plugin-session-auditor
Triggers: auto-invokes when you hand Claude Code a .jsonl session log and ask to audit, review, analyze, or find issues in a plugin run. Phrases like "look at this transcript", "what went wrong in this session", or "the plugins were misbehaving in this run" also trigger it.
What it does:
- Parses the jsonl transcript(s) into structured event JSON via a bundled Go tool (
tools/session-parser/). - Detects which plugins under
plugins/were exercised in the session. - Spawns four specialist subagents in parallel —
permissions,errors,tool-failures,orchestration— each writing findings to a shared$RUN_DIR/findings/<category>.md. - Consolidates findings into
proposals.mdwith evidence (timestamps + tool_use_ids), 2+ fix options per issue, and a recommendation. - Asks which proposals to implement before touching any
plugins/<name>/source. On approval, applies fixes and bumps the affectedplugin.jsonversionper the repo's SemVer rules.
Where it lives: .claude/skills/plugin-session-auditor/ (SKILL.md, agents/, references/, tools/, evals/).
Repository layout
.claude-plugin/marketplace.json # marketplace manifest
plugins/<name>/ # one directory per plugin
.claude-plugin/plugin.json
commands/<file>.md # usually <plugin>.md; doc-audit ships audit-docs.md
agents/, references/, bin/, tools/, hooks/ # only where the plugin needs them
src/, dist/, package.json, tsconfig.json # code-review-AT only (SDK build)
.claude/skills/<name>/ # repo-internal skills (not marketplace plugins)