convez.ai
SkillDevOps & InfraestruturaIntermediário
astra-sh

Quiver (qvr)

astra-sh/qvr

Gerenciador git-nativo de skills para agentes IA com lockfile reprodutível

Ver no GitHub
InicialPluginScore · de 0 a 100

Gerenciador de skills para agentes IA escrito em Go, com modelo git-nativo e lockfile-first. Oferece versionamento, reprodutibilidade e segurança para skills sem necessidade de servidor, similar ao uv para Python.

Instalação

$npx convez-ai astra-sh/qvr

1 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

Estrelas
26
Forks
0
Acompanhando
0
Saúde
50%
Último commit
há 4 meses
Versão
v0.30.0
Issues abertas
5
Contribuidores
2

Casos de uso

  • Versionamento e reprodutibilidade de skills em projetos multi-agente
  • Governança e auditoria de dependências de skills com SHA-keyed immutability
  • Instalação ultra-rápida de skills em cold-start (~1.3s) e warm-cache (~0.02s)

Veredito

Ferramenta madura e pragmática que trata skills como cidadãos de primeira classe no ecossistema de agentes. A abordagem git-nativa sem daemon é inteligente para CI/CD e colaboração em equipe.

README

A tradução para o português deste README ainda está na fila.

quiver(qvr)

An extremely fast skills manager for your agents, written in Go.


Agent skills are the new packages of AI — it's time to manage them the right way!

A skill is a folder of instructions and scripts your agent loads and executes on your behalf. That makes it a dependency, with all the same questions npm and uv taught us to ask: Where did it come from? What version is pinned? Has it been scanned? Can I reproduce this exact set on another machine?

Today most skills are copy-pasted into .agent/skills/ by hand — unversioned, unscanned, unattributable. Quiver gives agent skills the lifecycle software packages already have: a source, a lock, a gate, and an audit trail — with no server to run and no runtime in the read path.

qvr is to agent skills what uv is to Python packages: a Git-native, zero-service CLI to install, version, lint, scan, and govern [agent skills] across every coding agent — Claude Code, Cursor, Copilot, Codex, Windsurf, anything that reads skills from a directory.


Why Quiver?

1. Built for agentic usage

Developed in Go on native git — no daemon, no service, no language runtime in the read path. On a named-subset install from anthropics/skills, qvr lands a cold (first-time) install in ~1.3s and a warm (cached) install in ~0.02s:

Each skill is installed as a git worktree — a sparse checkout off a single bare clone of the source registry. The worktrees are immutable and SHA-keyed, so content is shared by construction: switching versions is a symlink repoint, not a re-clone, and two projects pinned to the same SHA share one copy on disk.

2. qvr.toml + qvr.lock — intent and proof, kept in sync

A project's skill set lives in two committed files that move together: qvr.toml declares the intent — the skills you want and the agents they install into — and qvr.lock records the resolved proof — the exact bytes, pins, and gate verdicts behind that intent. One you write; one Quiver writes.

qvr.toml is the front door. Human-authored and hand-editable, it carries the skills a project depends on ([skills], a coordinate → ref map) and its default agent targets ([project].default-targets, set by qvr target add). It's the declarative source you edit and review in a PR, and it's where routing policy travels with the repo — clone, qvr sync, and skills land in the same agent dirs with no machine-local drift.

# qvr.toml — declarative intent (hand-editable, committed)
[project]
name            = 'my-project'
version         = '0.1.0'
default-targets = ['claude', 'codex']   # agents a bare `qvr add` installs into

[skills]
'anthropics/skills/frontend-design' = 'main'   # coordinate -> ref

qvr.lock is the proof. Machine-generated, it's the heart of the trust layer and the unit of reproducibility: every teammate and CI runner resolves it to the byte-identical, already-vetted skill set. It doesn't just pin a SHA — each entry records the resolved commit, a subtree hash of the exact bytes, the scan report hash + decision, and the commit author, so the lock is the audit trail.

# qvr.lock — one resolved, vetted entry per skill package (machine-generated)
[[skill]]
name        = 'frontend-design'
registry    = 'anthropics/skills'
source      = 'https://github.com/anthropics/skills.git'
path        = 'skills/frontend-design'
ref         = 'main'
commit      = 'da20c92503b2e8ff1cf28ca81a0df4673debdbf7'    # resolved SHA
subtreeHash = 'sha256:21dce9699042…'                        # exact bytes installed
targets     = ['claude']
scan        = {reportSHA = 'sha256:8e861d1c…', decision = 'allowed', counts = {…}}
provenance  = {commitAuthor = 'Keith Lazuka <klazuka@anthropic.com>', signatureStatus = 'none'}

They stay in sync — both directions. Every mutating command (qvr add, switch, remove, target, …) writes through to both files at once, so they never drift in normal use. When they do diverge — a hand-edit, a merge — two explicit verbs reconcile them: qvr sync resolves toward the lock (the lock wins; it's the reproducible truth) and qvr lock --from-toml pushes your qvr.toml edits into the lock (intent wins). And the lock is self-sufficient: qvr sync rebuilds the whole set from qvr.lock alone, so CI never needs qvr.toml and a lost qvr.toml is regenerated from the lock — routing policy and all.

[!NOTE] Portability: the pair travels with the repo. Clone on any machine, run qvr sync, and you get the byte-identical, already-vetted skill set — not "whatever the registry serves today."

Provenance & governance: the lock is a verifiable record. qvr lock verify --strict and qvr sync --frozen are CI gates that fail on any drift between what's on disk and what the lock attests. Only skills in the lock are visible to the agent — no ambient surprises.

Provenance surfaces through qvr provenance; invalid signatures always block, and qvr trust pin enforces per-registry commit-author policy. And because the lock pins git refs rather than opaque archives, version control is a first-class feature:

qvr add code-review@v1.2.0
qvr add code-review@v1.3.0-rc1 --as code-review-rc   # both coexist for A/B

3. Traceability — the foundation for optimizing and evaluating skills

You can't optimize what you can't measure. Traceability is the foundational block for evaluating and improving skills: once every run is attributable to the exact skill bytes that produced it, you can tell which version actually moved the needle and close the authoring loop on evidence rather than guesswork.

Skills are software, so qvr gives them the inspection surface software gets. qvr audit captures each agent's native transcript verbatim (the lossless source of truth), then projects it into OpenTelemetry spans — Turn / Tool / Skill — using the GenAI semantic conventions. Spans serialize to standard OTLP, so qvr audit spans --otlp feeds any consumer (Jaeger, Tempo, Honeycomb, an OTel Collector) unchanged. A skill.* attribute family tags which skill each span belongs to and whether that identity was proven from the artifact the agent actually loaded — making skill attribution a first-class, queryable dimension of every trace. Pair that proven attribution with the lock's per-ref pinning and an A/B test stops being anecdotal: each variant's spans trace back to a specific SHA.

The embedded dashboard (qvr ui, baked into the binary) drills from a registry down to a single skill: its files, agent targets, scan results, version history, provenance, and recorded sessions.


Installation

1. Prebuilt binary (recommended)

A single self-contained binary with the dashboard baked in — no Go or Node required.

# Linux / macOS
curl -fsSL https://raw.githubusercontent.com/astra-sh/qvr/main/install.sh | sh
# Windows (PowerShell)
irm https://raw.githubusercontent.com/astra-sh/qvr/main/install.ps1 | iex

The installer detects your OS/arch, downloads the matching release, verifies its checksum, and drops qvr on your PATH. Tune it with environment variables:

VariableEffect
QVR_VERSIONInstall a specific release (e.g. v0.13.0).
QVR_INSTALL_DIRInstall location (default: a dir on your PATH).

Verify the install:

qvr --version
qvr doctor          # sanity-check the environment and any existing installs

2. From source

For contributors, or to build the latest main. Requires Go 1.25+ and Node 20+.

git clone https://github.com/astra-sh/qvr.git
cd qvr
make build          # builds the React UI, then embeds it into the binary
make install        # -> /usr/local/bin/qvr  (use sudo if needed)

[!NOTE] Plain go install is intentionally not a supported path — it can't run the npm build, so it ships without the dashboard. Use the prebuilt binary or make build.

Updating

Once qvr is on your PATH, update it in place — no need to re-run the installer:

qvr upgrade                     # download + verify + atomic swap to latest
qvr upgrade --check             # report whether a newer release exists
qvr upgrade --version v0.13.0   # pin a specific release

The release binary carries the dashboard embedded, so qvr upgrade brings the UI current too. The download, checksum verify, and swap all happen in-process — no curl/tar dependency on any platform.


Quick start

# register a source — any git clone URL, one skill or fifty
qvr registry add git@github.com:acme/skills.git
qvr registry list

# find skills
qvr search deploy
qvr version list code-review

# add a skill into the current project (scans, then writes qvr.toml + qvr.lock)
qvr add code-review                 # latest semver tag, or default branch
qvr add code-review@v1.2.0          # pin a tag, branch, or SHA
qvr sync                            # reconcile the project against qvr.toml + qvr.lock

Anything under .claude/skills/ that isn't in qvr.lock is hidden from the agent on qvr sync. The lockfile is the only source of truth for what your agent loads.


Commands

Skills are software, so Quiver runs them through a software lifecycle — and it closes. Authoring a new version isn't a fresh start; it re-enters the same gate every consumer's install went through.

source ─► registry add ─► scan ─► lint ─► add ─► edit ─► publish ─────┐
                          ▲                                           │
                          └──────────────── re-gate ◄─────────────────┘
              (authoring a new version re-enters the gate at scan)

Register a source — qvr registry add

Point qvr at where skills live. Any git clone URL works — the indexer walks the repo and finds the skills, whether it holds one skill at its root or fifty under skills/.

qvr registry add https://github.com/acme-labs/agent-skills   # -> acme-labs/agent-skills
qvr registry list acme-labs/agent-skills                     # skills inside one source
qvr registry update                                          # fetch + rebuild the index

Install a skill — qvr add

Add a skill from a registered source into the current project. The skill is scanned, the lockfile pins the resolved SHA + verdict, and it's symlinked into every agent dir you target.

qvr add code-review                 # latest semver tag, or default branch
qvr add code-review@v1.2.0          # pin a tag, branch, or SHA
qvr add --global diagnose           # ambient: available in every session
qvr add code-review --target cursor # install into a specific agent dir

Author & edit — qvr create / qvr edit

Scaffold a new skill, or eject an installed one from its immutable worktree into a real, editable directory with its own git history. Other agent targets are repointed at the edit dir so they stay in sync.

qvr init                                   # bootstrap a project (writes qvr.toml; uv-init style)
qvr create my-skill                        # scaffold a spec-valid skeleton
qvr lint my-skill                          # check it against agentskills.io
qvr edit code-review                       # symlink -> real, editable dir
qvr diff code-review                       # local changes vs. HEAD

qvr init is the project front door: it writes a qvr.toml into the current directory and infers default-targets from any agent dirs already present (.claude/skills, .github/skills, a shared .agents/skills → the universal project target). It never creates a skill or a lockfile — use qvr create to author a skill and qvr add to install one.

Publish upstream — qvr publish

Push your edits back to the skill's origin. qvr publish re-runs the full lint + scan gate locally and never touches the remote until it passes — closing the loop.

qvr publish code-review -m "tighten checklist"      # push HEAD upstream
qvr publish code-review --tag v1.3.0 -m "v1.3.0"    # cut a release, auto un-eject to the tag
qvr publish code-review --fork <git-url> --migrate  # push to your fork and track it
qvr publish code-review --dry-run                   # lint + scan, report target, no push

qvr switch <skill> --latest follows the latest semver tag; qvr switch <skill> <ref> flips to any ref without branching; qvr pull <skill> fast-forwards the current ref to its upstream tip. (qvr upgrade is separate — it updates the qvr CLI itself.)

Audit & trace — qvr audit

qvr audit enable                    # opt in (creates ~/.quiver/skillops.db)
qvr audit install-hooks             # wire your agents' native hooks
qvr audit sessions                  # recorded agent sessions
qvr audit logs                      # turn / tool / skill spans
qvr audit export > traces.jsonl     # raw traces as JSONL (OTLP-ready)

The capture is verbatim and projects to OpenTelemetry spans — see the audit & tracing guide.

Inspect & verify

qvr list                    # skills in the project lock (--all unions global)
qvr info <skill>            # frontmatter, refs, targets
qvr status [skill...]       # per-skill state: clean / dirty / drift
qvr provenance <skill>      # where it came from, signature + author trust
qvr outdated                # ls-remote vs. pinned SHAs
qvr doctor                  # broken installs, orphan artifacts (--strict to fail)
qvr lock verify --strict    # CI gate: every entry is verifiably the recorded state

Garbage-collect the worktree store

The SHA-keyed worktree store under ~/.quiver/worktrees/ is derived state — qvr sync rebuilds any missing worktree from the lock — so it GCs freely (verbs mirror uv cache):

qvr cache list              # reachable + orphan worktrees with sizes
qvr cache prune             # drop worktrees no project lock references
qvr cache clean             # wipe the store and the registry index cache

Dashboard — qvr ui

qvr ui                      # serve the local dashboard (embedded in the binary)

How it's wired

Storage: bare git clones → per-skill worktrees (sparse checkout) → symlinks into agent dirs.

 ~/.quiver/                              shared worktree store + index cache
 ├── config.yaml                         registered sources (URLs + names)
 ├── qvr.lock                            global ambient lock (--global lane)
 ├── registries/<org>/<repo>.git/        bare clones (source of truth)
 ├── worktrees/<org>/<repo>/<skill>/<sha7>/   immutable, SHA-keyed, shared
 └── cache/index/<name>.json             TTL'd registry index cache

 <project>/
 ├── qvr.toml                            declarative intent (skills + default targets)
 ├── qvr.lock                            resolved lock — source of truth
 ├── .claude/skills/<skill>   -> symlink into ~/.quiver/worktrees/
 ├── .agents/skills/<skill>   -> symlink into ~/.quiver/worktrees/
 └── .github/skills/<skill>   -> symlink into ~/.quiver/worktrees/

Agent targets

Targets are a data-driven registry (~60 agents) compiled into the binary. Run qvr target list for the full set with directories and aliases. Paths are sourced from each tool's official docs; many newer CLIs share the AGENTS.md .agents/skills project location. Common core targets:

TargetLocal dirGlobal dirAliases
claude.claude/skills~/.claude/skillsclaude-code
codex.agents/skills~/.codex/skills
cursor.agents/skills~/.cursor/skills
copilot.github/skills~/.copilot/skillsgithub-copilot
gemini.agents/skills~/.gemini/skillsantigravity
windsurf.windsurf/skills~/.codeium/windsurf/skills
project.agents/skills~/.agents/skillsagents

Pick which agents a project installs into with qvr target add <name>...; the choice is recorded in qvr.toml ([project].default-targets) so it travels with the repo. Selection order is --target flag > qvr.toml default-targets > config default_target.


Standards

Quiver builds on open standards rather than inventing its own formats.

  • Skills follow agentskills.io. A skill is a directory with a SKILL.md whose YAML frontmatter carries name + description (plus optional license, compatibility, allowed-tools, metadata). Quiver's parser (pkg/skillspec) and linter enforce the spec — nothing Quiver-proprietary is required to author a skill.
  • Traces follow OpenTelemetry. Captures are stored verbatim and projected into OTLP spans, stamped with a deriver_version so an improved deriver can re-run over old captures (qvr audit rederive) without re-capturing.

Documentation

In-depth docs live under documentation/:


Built and maintained by SRP · MIT licensed · Issues and PRs welcome

Ver o original no GitHub

Bom para

Relacionados