harness-dev-plugins
vitoliu93/harness-dev-plugins
Global Claude Code / Cursor / Codex skills, agents, and harness plugins
Instalação
npx convez-ai vitoliu93/harness-dev-plugins1 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
README
A tradução para o português deste README ainda está na fila.
vito-agent-plugins
Vito's global skill collection, packaged as a Claude Code plugin. The repo root
is the plugin root — all 14 skills under skills/, 9 subagents under
agents/, and the hooks under hooks/ are auto-discovered.
Install
This repo doubles as its own marketplace. From any machine with the repo checked
out (e.g. at ~/codebase/projects/agent-plugins):
/plugin marketplace add ~/codebase/projects/agent-plugins
/plugin install vito-agent-plugins@vito-agents
Then restart the session. Skills become available as vito-agent-plugins:<skill>
(e.g. /vito-agent-plugins:advanced-plan).
Skills
| Skill | What it does |
|---|---|
| advanced-plan | Track a non-trivial dev task as a mini-project that survives context resets and handoff. |
| ask-ai | Clean-context second opinion (当局者迷,旁观者清) via opus subagent / Codex gpt-5.5 / ultra (gpt-5.5 @ xhigh or opus-4.8 @ max). |
| audit-context | Audit, prune, and lean-refactor session context (CLAUDE.md, memory, imports). |
| bilibili-cli | Search B站 and read videos by subtitle + AI summary + comment 舆情 (via bili). |
| cc-reflection | Research-grounded reflection report on collaboration over a date range. |
| exa-code | Search the web for code examples, docs, and programming solutions via Exa. |
| gemini-media | Base skill: understand any audio/video file via Gemini Flash-Lite (no-subtitle fallback for the video skills). |
| handoff | Save / pick up task state for cross-session, cross-agent transfer. |
| harness-loop | Autonomous orient → plan → execute → observe → decide → persist loop methodology. |
| html-doc | Produce a single self-contained, infographic-style HTML explainer. |
| self-learn | Extract knowledge from the session into Chinese learning notes. |
| tech-selection | Track an open research/feasibility question as a resumable study. |
| xiaohongshu-cli | Search 小红书 and read 图文 notes + comment 舆情 for real Chinese content (via xhs). |
| youtube-cli | Search YouTube and read videos by transcript + comment sentiment (via yt-dlp). |
Subagents
Eight skills also ship a companion subagent (spawnable via the Agent/Task tool as
vito-agent-plugins:<agent>). Each wraps its skill in an isolated context so the
noisy intermediate work — transcripts, search dumps, upload logs — never enters
the main session; only a distilled answer comes back. The other six skills are
deliberately not wrapped: they either need the live session conversation
(handoff, self-learn), are interactive end-to-end (audit-context), or are
methodologies the main agent itself must drive (advanced-plan, harness-loop,
tech-selection).
| Agent | Wraps skill | Returns |
|---|---|---|
| ai-second-opinion | ask-ai | Distilled verdict from the chosen engine (opus/codex/ultra) + verification note. |
| bilibili-researcher | bilibili-cli | Quality-assessed digest of B站 videos, cited by BVID. |
| cc-reflector | cc-reflection | Report path + Lark URL + top findings + critic tally. |
| exa-searcher | exa-code | Synthesized answer + sources from Exa web/code search. |
| html-visualizer | html-doc | Path of the written HTML artifact + pattern rationale. |
| media-reader | gemini-media | Structured digest (or verbatim transcript on request). |
| xiaohongshu-researcher | xiaohongshu-cli | 小红书 consensus + 避坑/软广 flags, cited by note. |
| youtube-researcher | youtube-cli | Per-video verdicts with timestamps + synthesis. |
A ninth subagent, lark-operator (pinned to sonnet), wraps the global
lark-* skill family (installed via npx skills, not part of this repo). It
executes 飞书 operations — notify, archive, schedule, read-back — and returns
only a confirmation (链接/ID) or a digest, keeping skill bodies and lark-cli
JSON out of the main session.
Note: Claude Code subagents cannot spawn nested subagents, so the wrappers read content inline with strict distillation discipline instead of fanning out the per-video/per-note readers their skills describe for main-session use.
Hooks
hooks/hooks.json registers one PreToolUse command hook (lark-guard.sh)
matching Skill|Read. In the main context it denies loading any lark-*
skill (and reading */skills/lark-* files), with a reason that redirects the
agent to delegate to vito-agent-plugins:lark-operator — most Lark work is
task-irrelevant post-hoc 通知/留档. Inside any subagent (the hook input
carries agent_id there) lark-* runs freely, so the delegation itself is never
blocked. Exemptions: lark-skill-maker (skill development is a main-context
task); typing a /lark-* slash command directly also bypasses the guard since
that path doesn't go through the Skill tool. Hook changes load at session
start — restart the session after editing.
Layout
.claude-plugin/
plugin.json # plugin manifest (name: vito-agent-plugins)
marketplace.json # marketplace (name: vito-agents), plugin source "./"
skills/ # 14 skills, auto-discovered
agents/ # 9 companion subagents, auto-discovered
hooks/
hooks.json # PreToolUse: lark-guard
scripts/lark-guard.sh
Note on portability
Several skills (cc-reflection, advanced-plan, tech-selection, exa-code)
reference their own scripts/assets via hardcoded ~/.claude/skills/<name>/... (or
repo-relative skills/<name>/...) paths. When loaded as a plugin, skills live in
the plugin cache, not ~/.claude/skills, so those hardcoded paths do not
resolve — those skills need ${CLAUDE_PLUGIN_ROOT}-relative paths to be fully
portable. The packaging here does not modify skill internals; fix those paths if
you hit a missing-file error when a skill tries to read its own assets.