OpenCode 智能体包 agents-opencode：10 个智能体 + 24 项按需技能

Ten specialist agents. 24 on-demand skills. One install.

Turn a single prompt into a full delivery loop — plan → implement → review → document — without leaving the terminal you already use. Built for OpenCode. Already on Claude Code? The same 24 skills ship as a plugin.

Why this pack is different

• One prompt, four disciplines. @orchestrator scopes the work, then hands off to @codebase for implementation, @review for quality, and @docs for documentation — in one session.
• Skills that don't tax context. All 24 skills load only when invoked, so an idle skill costs nothing.
• Least privilege by default. Every agent starts at "*": "deny" and grants narrow allows for the tools and skills it actually needs.
• Verified, not vibes. Agent contracts, command matrices, docs links, and plugin guards are checked in CI on every change.

Quick Start

Requires: Node.js (current LTS).

npx agents-opencode --global      # install into your global OpenCode config
opencode                          # start the TUI
/init                             # initialize a new session
@orchestrator Build a REST API with JWT auth

That last line triggers the orchestrator: it plans the work, delegates implementation to @codebase, quality checks to @review, and documentation to @docs — all in one session.

Install options (filter, update, uninstall, status)

# Lighter install: keep only selected language skills
npx agents-opencode --global --languages python,typescript

# Update an existing install (auto-detects scope; --all, --global, or --project [dir] to be explicit)
npx agents-opencode --update

# Uninstall (current project by default; --global or --all for explicit scope)
npx agents-opencode --uninstall

# Show detected installation scopes
npx agents-opencode --status

--languages prunes non-requested language skill directories only — shared workflow skills always stay.

Install: npm package and installer command: agents-opencode. OpenCode runtime command: opencode. Uninstall: targets the current project by default; timestamped backups are created before removal. Config: the installer only manages $schema, plugin, and missing permission defaults in opencode.json. It never rewrites your provider, model, instructions, share, or compaction settings.

A session, end to end

@orchestrator coordinates instead of just answering:

1. Plan — breaks the request into phases and surfaces risks.
2. Implement — delegates build work to @codebase, which follows your repo's conventions.
3. Review — @review checks for bugs, security holes, and performance regressions before you ship.
4. Document — @docs writes the README, API reference, and ADRs to match.

You approve the checkpoints; the specialists handle the handoffs.

Agents

Ten agents ship in the pack: seven delegation targets that the orchestrator calls, and three primary agents you run directly.

@blogger and @em-advisor are primary agents — switch to them with Tab or mention them directly. They are not Task-invocable subagents, so the orchestrator will not delegate to them.

Canonical source for exact allowlists and skill triggers: Skills Matrix.

Commands

Type /command-name in the TUI to run:

Skills

24 skills live under .opencode/skills/, each defined in a SKILL.md:

• 11 language skills — .NET, Flutter, Go, Java/Spring, Node/Express, Python, React/Next.js, Ruby/Rails, Rust, SQL migrations, TypeScript.
• 13 workflow skills — API docs, ADRs, docs validation, project bootstrap, agent diagnostics, refactoring, security audit, code-change impact, UX/responsive, decision grounding, blogging, brutal critique, legal compliance.

Skills load on demand via the skill tool. Coding standards and language guidance live in these skills — not in a sprawling instruction directory. A small set of shared reference instructions stays in .opencode/instructions/.

Scope is core-only: additions pass demand, clear-gap, ownership, and licensing/provenance checks.

Permission configuration (least-privilege patterns)

Skill permissions (prevent unrelated skill loads):

permission:
  skill:
    "*": "deny"
    "python": "allow"
    "sql-migrations": "allow"

Task permissions (control which subagents each agent can invoke):

permission:
  task:
    "*": "deny"
    "explore": "allow"
    "review": "allow"

Start with "*": "deny", then add explicit allows. Rules match in order — last match wins.

Real-world efficiency

Metrics from production usage (May–July 2026) on the deepseek-v4-pro model. Persistent context reuse keeps the bulk of input tokens cached across agent sessions.

Across 2,803 requests, 97.3% of input tokens were served from cache — less re-processing and faster session starts. That is the practical payoff of an on-demand skill model.

Installation

npx (recommended)

npx agents-opencode --global

Omit --global to install into the current project instead.

Claude Code plugin

# Add marketplace (one-time)
/plugin marketplace add shahboura/agents-opencode-claude

# Install
/plugin install agents-opencode@shahboura

# Update
/plugin update agents-opencode@shahboura

This gives Claude Code access to the same 24 on-demand skills, which load only when invoked — no context cost until you use them. See adapters/claude-code/ for the plugin manifest and generator script.

Validation

Run npm run doctor for the complete local validation suite (agent contracts, markdown linting, docs links, session state, eval trends, and more). For the full check mapping (local commands ↔ CI gates), see Compatibility.

Agent evals: npm run eval:agents runs deterministic contract checks for agent and command metadata. npm run eval:agents:json writes machine-readable output.

Docs

• Getting Started
• Approval Gates
• Compatibility
• Deprecation & Migration Policy
• State Management
• Skills Matrix
• Decision Grounding
• Full Documentation

License

MIT © 2025–2026 Shehab Elhadidy.

If this pack saves you time, star the repo so others can find it.