@cyberuni/agent-harness

A toolkit for working with AI agent harnesses. Detect which one is running — Claude Code, Cursor, Codex, GitHub Copilot CLI, OpenCode, Kilo Code, Gemini CLI, Qwen Code, GitHub Copilot in VS Code, Cline, Crush, OpenHands, or the Auggie CLI — query what it holds (its managed-policy locations, plugin storage, enabled plugins, how it names plugin skills, and how to run it headless), and resolve the reference documents an agent reads on demand.

It depends on no other agent tool, so any of them can build on it.

Status: pre-release. Every fact this library encodes rests on .research/harness-detection.

Detect the harness

import { detectHarness } from '@cyberuni/agent-harness'

const result = detectHarness()
// {
//   harness: 'claude-code',
//   candidates: ['claude-code'],
//   evidence: [
//     { harness: 'claude-code', signal: 'CLAUDECODE', research: 'E-CC-D1', description: '…' },
//     { harness: 'claude-code', signal: 'CLAUDE_CODE_CHILD_SESSION', research: 'E-CC-D3', description: '…' },
//   ],
// }

Detection reads environment variables that each harness sets on the processes it spawns:

It reports unknown instead of guessing:

• when no harness matches, including CLAUDECODE=1 on its own, which IDE extensions also set;
• when more than one harness matches, as when one harness runs inside another's shell. The matches are listed in candidates;
• for AGENT and AI_AGENT, which several vendors set (only the exact value AI_AGENT=openhands counts), and for OPENCODE=1 alone, which Kilo Code, an OpenCode fork, also sets;
• in contexts where a harness sets nothing of its own: Codex, Gemini CLI, and Qwen Code hooks, Qwen Code, Crush, OpenHands, and Auggie MCP servers, Auggie shell commands, OpenHands CLI 1.16.0 shell commands (its SDK predates AI_AGENT=openhands), and the Cline CLI;
• for harnesses with no verified signal, such as Goose, Antigravity, Rovo Dev, Kiro, Amp, Factory Droid, Devin Desktop, and Warp. The research records why for each one.

Pass { env } to inspect an environment other than process.env.

Managed policy

import { managedPolicyLocations } from '@cyberuni/agent-harness'

managedPolicyLocations('claude-code', { platform: 'linux' })
// [
//   { kind: 'server', location: 'claude.ai admin console or a self-hosted Claude apps gateway', … },
//   { kind: 'file', location: '/etc/claude-code/managed-settings.json', … },
//   { kind: 'directory', location: '/etc/claude-code/managed-settings.d', … },
//   { kind: 'file', location: '/etc/claude-code/managed-mcp.json', … },
// ]

kind is file or directory for what a caller can read from disk, macos-managed-preferences or windows-registry for MDM delivery, and server for policy held by the vendor or the organization. A server entry means the local files are not the whole policy.

Plugin storage

import { pluginStorage } from '@cyberuni/agent-harness'

pluginStorage('codex')
// {
//   harness: 'codex',
//   configDir: '/home/me/.codex',
//   locations: [
//     { kind: 'plugin-cache', path: '/home/me/.codex/plugins/cache', … },
//     { kind: 'enabled-record', path: '/home/me/.codex/config.toml', … },
//   ],
// }

Paths are computed, not checked, so a location may not exist until the first plugin is installed. Cursor has no enabled-record: no file records which of its plugins are enabled.

Enabled plugins

import { enabledPlugins } from '@cyberuni/agent-harness'

const result = await enabledPlugins('claude-code')
// {
//   harness: 'claude-code',
//   supported: true,
//   plugins: [{ id: 'my-plugin@my-marketplace', enabled: true, scope: 'project', source: '/repo/.claude/settings.json' }],
//   sources: [{ scope: 'managed', path: '/etc/claude-code/managed-settings.json', found: false }, …],
//   unread: ['server-managed settings (claude.ai admin console or a self-hosted Claude apps gateway)', …],
//   research: ['E-CC-P3', …],
// }

Only plugins with an explicit entry are reported. A plugin with no entry falls back to a default this package does not know. When unread is not empty, MDM or server-managed policy can still override the answer.

Installed plugins

import { enabledPlugins, installedPlugins } from '@cyberuni/agent-harness'

const installed = await installedPlugins('claude-code')
// {
//   harness: 'claude-code',
//   supported: true,
//   plugins: [{ id: 'my-plugin@my-marketplace', path: '/home/me/.claude/plugins/cache/my-marketplace/my-plugin/1.0.0', version: '1.0.0', scope: 'user', research: ['E-CC-P2'] }],
//   unresolved: [],
//   sources: [{ path: '/home/me/.claude/plugins/installed_plugins.json', found: true }],
//   unread: [],
//   research: ['E-CC-P2'],
// }

// The folder of each enabled plugin:
const enabled = new Set((await enabledPlugins('claude-code')).plugins.filter((p) => p.enabled).map((p) => p.id))
const folders = installed.plugins.filter((p) => enabled.has(p.id)).map((p) => p.path)

unresolved lists Codex plugins whose cache mixes semver and non-semver version folders, where the folder Codex picks depends on its sort. Copilot CLI loads plugins from a marketplace added by local path in place, without an install record, so they appear in unread instead.

Skill naming and plugin dependencies

import { skillInvocation, supportsPluginDependencies } from '@cyberuni/agent-harness'

skillInvocation('claude-code', { plugin: 'my-plugin', skill: 'review' })?.text // '/my-plugin:review'
skillInvocation('copilot-cli', { plugin: 'my-plugin', skill: 'review' })?.text // '/review'
skillInvocation('codex', { plugin: 'my-plugin', skill: 'review' })?.text // '$review'
skillInvocation('gemini-cli', { plugin: 'my-plugin', skill: 'review' }) // undefined: the model loads skills itself

supportsPluginDependencies('claude-code') // true; false for the others

Skill directories

import { skillsDirectories } from '@cyberuni/agent-harness'

skillsDirectories('claude-code') // { project: ['.claude/skills'], user: ['.claude/skills'], research: ['E-CC-L1'] }
skillsDirectories('codex')?.project.includes('.agents/skills') // true
skillsDirectories('qwen-code') // undefined: not confirmed

Project directories are relative to the project root, user directories to the home directory. Only directories read by default are listed; admin, bundled, and plugin skills are left out.

Headless runs

How to run a harness on one prompt with no user present: the command, the model and permission flags, the output formats and where token usage and cost appear in them, the transcript location, and the documented exit codes. The library only describes the run; starting the process is up to you.

import { headlessCommand, headlessInvocation } from '@cyberuni/agent-harness'

headlessCommand('claude-code', {
  prompt: 'Fix the failing test',
  model: 'sonnet',
  permission: 'unattended',
  outputFormat: 'stream-json',
})
// { executable: 'claude', args: ['-p', '--model', 'sonnet', '--permission-mode', 'bypassPermissions',
//   '--output-format', 'stream-json', '--verbose', 'Fix the failing test'] }

const codex = headlessInvocation('codex')
// codex.outputFormats[1]: { name: 'jsonl', args: ['--json'],
//   usage: { event: 'turn.completed', tokens: 'usage', accumulate: 'sum' } } (no dollar cost)

headlessInvocation('vscode-copilot').supported // false: no headless mode
headlessInvocation('opencode').supported // 'unknown': a one-shot mode exists, but its flags are not verified

Claude Code, Codex, Copilot CLI, Cursor, Gemini CLI, and Qwen Code are covered. permission: 'unattended' lets the run edit files and run commands without asking, so use it only in a throwaway checkout. Exit codes are a weak signal: Claude Code documents only zero and non-zero, and a run stopped by a turn or budget cap is told apart by the result message's subtype.

Reference documents

A reference is a named Markdown document an agent reads on demand, such as a checklist or a set of weights a skill applies. Anyone can override one without editing the plugin that ships it: a project in .agents/references/<name>.md, a person in ~/.agents/references/<name>.md.

import { loadReference } from '@cyberuni/agent-harness'

const resolved = await loadReference('agent-readiness-weights', {
    root: process.cwd(),
    // The plugin calling the resolver: its own `references/` is the first plugin layer.
    plugin: { name: 'my-plugin', root: pluginRoot },
})
if (resolved.status === 'found') console.log(resolved.content)

status is found, missing, or ambiguous (two plugins hold the name; ask for one of resolved.plugins, such as my-plugin/agent-readiness-weights). trace says what each layer held.

Layers are read highest precedence first:

In each folder, <name>.md answers first, then <name>/README.md, <name>/index.md, and <name>/SKILL.md. Legacy governances/ folders are still read, below the references/ beside them.

A document's frontmatter merge says how it combines with the layers below:

• first-wins (the default): it replaces everything below.
• combine: it is placed above the layers below, each kept whole.
• merge-sections: its sections replace the same-titled sections below, matched by heading path, and the sections it does not redefine are kept. Under a heading, <!-- merge: combine --> appends to that section instead, and <!-- merge: remove --> deletes it.

Frontmatter is read as YAML. merge, description, and tags are the keys the resolver reads; resolved.metadata holds every key, the higher layer winning. Frontmatter that is not a YAML mapping is ignored with a warning.

The reference command

The package ships a reference command with show, list, search, where, create, and delete subcommands:

npx -y @cyberuni/agent-harness reference show <name>... --root <repository root>
npx -y @cyberuni/agent-harness reference list
npx -y @cyberuni/agent-harness reference list 'project-tracker.capabilities.*'
npx -y @cyberuni/agent-harness reference create <name> --scope project
npx -y @cyberuni/agent-harness reference delete <name> --scope project --dry-run

delete removes only the copy in .agents/references/ (--scope project) or ~/.agents/references/ (--scope user), and reports which copy answers the name afterwards: the plugin's copy, once a project override of it is gone. It refuses a plugin-shipped or managed copy.

Output is TOON by default, for an agent to parse; pass --format json or --format text (show, create, and delete default to text). The formats follow clibuilder 12, and any other value is a usage error (exit code 2). A CLI built on clibuilder 12 can host the same command under its own plugin name, imported from the ./commands subpath. --format is the host's global option, so the host's cli({ format }) sets the default for every subcommand; the text default of show, create, and delete belongs to the agent-harness CLI:

import { createReferenceCommand } from '@cyberuni/agent-harness/commands'

app.command(createReferenceCommand({ plugin: { name: 'my-plugin', root: pluginRoot } }))

The reference skill

The package is also an agent plugin, cyber-agent-harness, that ships a reference skill. Type /reference (/cyber-agent-harness:reference in Claude Code) to load, create, update, delete, find, or inspect a reference, or to wire a skill to load one. The skill writes or deletes a file only after you approve it. It runs the command from its bundled scripts/reference.mjs, which needs no node_modules.

/plugin marketplace add cyberuni/cyberplace
/plugin install cyber-agent-harness@cyberplace

A skill that needs a reference names it in one line, as the skill's README.md shows:

Load `skill-design` with the `reference` skill in the `cyber-agent-harness` plugin.

Why this exists

universal-plugin, buddy-agent-harness, repobuddy, and cyberlegion/cyber-mux each need to know which agent harness they are running under, and each was re-implementing that detection separately. This package exists so they share one implementation instead.

License

MIT