Config

An OpenCode plugin that bundles orchestration agents, skills, commands, instructions, and MCP configurations.

How It Works

When installed, this plugin:

1. Discovers bundled assets in the npm package
2. Installs bundled skills, plugins, and MCPs via symlinks to your OpenCode config
3. Injects bundled agents, commands, and instructions into OpenCode's runtime config
4. Respects your local overrides and explicit disables

Priority Order

1. Your local configs — always win
2. Bundled configs — fill in the gaps

Installation

1. Add to your OpenCode config

In ~/.config/opencode/opencode.json:

{
  "plugins": ["@unfrgivn/opencode-config"]
}

2. Restart OpenCode

The plugin will automatically discover and install bundled assets on startup.

Optional: Enable the TUI sidebar

The main plugin entry in opencode.json installs and injects Fellowship agents, commands, skills, MCPs, orchestration, and companion support. The TUI sidebar is loaded separately by OpenCode's TUI config.

To show the Fellowship sidebar/status panel, add the package to ~/.config/opencode/cli.json:

{
  "plugins": ["@unfrgivn/opencode-config/tui"]
}

If you already have TUI settings such as a theme, keep them and add the plugin entry:

{
  "theme": {
    "name": "catppuccin-macchiato"
  },
  "plugins": ["@unfrgivn/opencode-config/tui"]
}

Use the /tui subpath here. The package root exports the server plugin for opencode.json, while @unfrgivn/opencode-config/tui exports the TUI sidebar plugin. Restart OpenCode after changing cli.json. If you do not want the sidebar, you can omit the TUI plugin entry; the main Fellowship plugin still works from opencode.json.

Configuration

Create ~/.config/opencode/fellowship.json (or for project-level .opencode/fellowship.json):

{
  "disable": ["skill:example", "agent:verbose-debugger"],
  "installMethod": "link"
}

Options

Option	Type	Default	Description
disable	string[]	[]	Assets to disable (see below)
installMethod	"link" | "copy"	"link"	How to install skills/plugins/MCPs
agentNames	"lotr" | "descriptive"	"descriptive"	Agent naming theme (see below)
models	Record<string, string>	{}	Override the model used by built-in bundled agents
companion	object	{ enabled: false }	Desktop companion overlay configuration
orchestration	object	defaults	Runtime orchestration admission, limits, and provider backpressure policy

Agent Name Themes

Agents can be displayed with either plain role-based names or Tolkien character names. Set agentNames in your config:

{
  "agentNames": "lotr"
}

Role	"descriptive" (default)	"lotr"
Orchestrator	orchestrator	gandalf
Explorer	explorer	legolas
Researcher	researcher	radagast
Planner	planner	treebeard
Implementer	implementer	celebrimbor
Architect	architect	elrond
Analyst	analyst	galadriel
TDD	tdd	aragorn

The theme affects agent names everywhere: in the agent list, slash commands, delegation routing, and cross-references between agents. Agent markdown files use template variables (e.g. {{explorer}}) so cross-references resolve correctly regardless of theme.

When using the disable list, use the themed name (the name that appears after applying your chosen theme). For example, with "descriptive" theme, disable with "agent:orchestrator" instead of "agent:gandalf".

Override Agent Models

Use models in fellowship.json to override the model for built-in bundled agents without replacing the rest of the agent config.

// ~/.config/opencode/fellowship.json
{
  "models": {
    "orchestrator": "anthropic/claude-opus-4",
    "explorer": "github-copilot/claude-sonnet-4.6"
  }
}

Built-in bundled agents accept either naming style as the key, regardless of your active agentNames theme. These two examples are equivalent:

{
  "models": {
    "orchestrator": "anthropic/claude-opus-4"
  }
}

{
  "models": {
    "gandalf": "anthropic/claude-opus-4"
  }
}

Notes:

• models only affects built-in bundled agents from this plugin
• local opencode.json agent entries still take priority over bundled agent injection
• omitted agents keep their plugin default model
• empty string values fall back to the plugin default model
• if both aliases are provided for the same built-in agent, the exact key match wins

Orchestration Runtime Configuration

All fields below show their defaults. You only need to include fields you want to override.

{
  "orchestration": {
    "providerMode": "copilot",
    "features": {
      "delegation": true,
      "diagnostics": true,
      "vision": true,
      "continuousLoop": true,
      "selfHealing": true
    },
    "limits": {
      "maxConcurrent": 3,
      "maxConcurrentGlobal": 5,
      "providerMaxConcurrent": {},
      "maxRetries": 2,
      "maxIterations": 25,
      "retryBackoffMs": 1500,
      "taskTimeoutMs": 120000
    },
    "runtimeFallback": {
      "enabled": true,
      "retryOnErrors": [400, 401, 403, 408, 429, 500, 502, 503, 504, 529],
      "maxFallbackAttempts": 2,
      "cooldownSeconds": 30,
      "timeoutSeconds": 30,
      "notifyOnFallback": true,
      "providerOrder": ["copilot", "native"]
    },
    "hooks": {
      "disabled": [],
      "telemetry": true
    },
    "notifications": {
      "enabled": true,
      "maxItems": 100
    },
    "categories": {
      "routing": {},
      "maxConcurrent": {}
    },
    "recovery": {
      "autoResumeOnStart": true
    },
    "tmux": {
      "enabled": false,
      "sessionPrefix": "fellowship"
    }
  }
}

Tmux support is optional and only active when orchestration.tmux.enabled is true and OpenCode is running inside a tmux session. When enabled, a dedicated tmux session ({sessionPrefix}-{opencode-session-id}) is created for each OpenCode session, and each subagent task gets its own window inside that session. Windows are automatically cleaned up when tasks complete, fail, stop, or parent sessions end. Use tmux switch-client -t fellowship-* to peek at agent work.

Feature flags (orchestration.features) toggle individual orchestration capabilities. Set any to false to disable that feature while keeping the rest active.

limits.maxConcurrent is the per-parent/session active cap. maxConcurrentGlobal caps active work across local OpenCode processes, and providerMaxConcurrent can lower that cap for copilot or native. Admission is coordinated through the local orchestration cache. A rejected task does not launch and does not create a queued record.

Category policies are optional. Use orchestration.categories.routing to override provider mode per task category (for example, "quick": "native") and orchestration.categories.maxConcurrent to cap active tasks per category (for example, "quick": 1). Congestion responses (408, 429, 503, and 504 when listed in runtimeFallback.retryOnErrors) establish a shared provider cooldown and reduce effective provider capacity. Successful completions restore capacity one slot at a time.

Recovery policies are optional. orchestration.recovery.autoResumeOnStart checks persisted work on startup. Tasks owned by a live process remain active. Abandoned running records and legacy queued/retrying records become terminal interrupted records because OpenCode does not expose a safe API for replaying the original task call.

The runtime does not auto-dispatch queues or retry failed task calls. maxRetries, retryBackoffMs, runtimeFallback.maxFallbackAttempts, timeoutSeconds, notifyOnFallback, and providerOrder remain accepted for configuration compatibility but do not schedule work.

Note: Configuration changes take effect on restart. After modifying your config, restart OpenCode to apply the changes.

Desktop Companion

The package includes a bundled companion helper that can show small desktop tiles while agents are active or waiting for input. No separate install is required for the default setup, the helper and themed artwork ship inside @unfrgivn/opencode-config under companion/.

To enable it, add companion.enabled to ~/.config/opencode/fellowship.json or a project-level .opencode/fellowship.json:

{
  "$schema": "./schemas/fellowship.schema.json",
  "companion": {
    "enabled": true,
    "position": "bottom-right",
    "size": "medium",
    "gifPack": "default",
    "loopStyle": "classic",
    "speed": 1,
    "debug": false
  }
}

Restart OpenCode after saving the config. On startup, Fellowship writes session state to OpenCode storage, launches the companion process, and the bundled macOS overlay opens chromeless tiles for busy or waiting agents. Idle state stays hidden.

Binary resolution order:

1. companion.binaryPath, when you configure a custom companion binary
2. $XDG_DATA_HOME/opencode/storage/fellowship/bin/fellowship-companion, or ~/.local/share/opencode/storage/fellowship/bin/fellowship-companion
3. The bundled companion/bin/fellowship-companion from this npm package

Only set binaryPath when replacing the bundled helper:

{
  "companion": {
    "enabled": true,
    "binaryPath": "/usr/local/bin/fellowship-companion"
  }
}

Companion labels and artwork follow agentNames. The default "descriptive" theme uses role names such as orchestrator and researcher; "lotr" uses gandalf, radagast, and the rest of the Fellowship agent theme. Available positions are bottom-right, bottom-left, top-right, and top-left; sizes are small, medium, and large; speed accepts 0.25 through 4.

Disable Syntax

{
  "disable": [
    "skill:verbose-logger", // Disable specific skill
    "agent:heavy-model", // Disable specific agent
    "command:dangerous-cmd", // Disable specific command
    "plugin:auto-formatter", // Disable specific plugin
    "instruction:old-guide", // Disable specific instruction
    "mcp:expensive-api", // Disable specific MCP
    "deprecated-thing" // Disable across ALL types (bare name)
  ]
}

Overriding Bundled Configs

The plugin respects local configurations. Your local settings always take precedence over bundled defaults.

Override Priority

1. Your local OpenCode config (opencode.json)     ← Highest priority
2. Explicit disable list (fellowship.json)
3. Local skill directories (~/.config/opencode/skill/<name>/)
4. Bundled configuration from this repo               ← Lowest priority

By Asset Type

Asset Type	How to Override	How to Disable
Skill	Create ~/.config/opencode/skill/<name>/SKILL.md	"skill:<name>" in disable list
Agent	Recommended: fellowship.json agentOverrides for partial changes. Use opencode.json agent only for full replacement.	"agent:<name>" in disable list
Command	Add to opencode.json: { "command": { "<name>": {...} } }	"command:<name>" in disable list
Plugin	Disable bundled plugin, add your own to plugins directory	"plugin:<name>" in disable list
Instruction	Cannot override (only append). Disable and add your own.	"instruction:<name>" in disable list
MCP	Add to opencode.json MCP config section	"mcp:<name>" in disable list

Examples

Recommended: partially override a bundled agent without replacing everything:

// ~/.config/opencode/fellowship.json
{
  "agentOverrides": {
    "planner": {
      "temperature": 0.2
    }
  }
}

agentOverrides applies patch-style merge to bundled agent configs before injection. This is the default/safest way to tweak one value while preserving the rest of the bundled config. Built-in bundled agents accept either naming style as the key, so planner and treebeard both work.

Advanced: fully replace a bundled agent definition:

// ~/.config/opencode/opencode.json
{
  "agent": {
    "bundled-agent": {
      "model": "claude-sonnet-4-20250514",
      "temperature": 0.5
    }
  }
}

⚠️ opencode.json agent entries are full replacements, not merges. If you set only one field (for example model), you may unintentionally drop bundled defaults like permission and prompt for that agent.

Disable multiple bundled assets:

// ~/.config/opencode/fellowship.json
{
  "disable": [
    "skill:verbose-logger",
    "agent:heavy-model",
    "instruction:old-conventions"
  ]
}

Override a skill with your own version:

# Create your local skill directory
mkdir -p ~/.config/opencode/skill/my-skill

# Add your SKILL.md
cat > ~/.config/opencode/skill/my-skill/SKILL.md << 'EOF'
---
name: my-skill
description: My customized version
---
Your custom skill content here...
EOF

Completely disable a skill you don't use:

// ~/.config/opencode/fellowship.json
{
  "disable": ["skill:unused-skill"]
}

What's Included

Agents (agent/)

Pre-configured agent definitions with model, temperature, and system prompts. There are two primary-capable agents (you can talk to them directly) and six dedicated subagents (they run in the background, delegated by the primary agent). The planner is available as both. Names shown below use the default "descriptive" theme, with "lotr" names in parentheses.

Primary Agents

Agent	Model	What it does
orchestrator (gandalf)	Claude Opus 4	The default orchestrator. Classifies your request, breaks it into tasks, delegates to the right subagents in parallel, and assembles the final result. Handles everything from quick fixes to multi-step projects.
planner (treebeard)	GPT-5.4	Read-only planning and analysis agent. Can explore codebases, search the web, and run safe shell commands, but cannot edit files or run mutating commands. Use directly when you want to think through a problem, review a plan, or explore before committing to changes. Also available as a subagent for delegation.

Subagents

Agent	Model	When called	What it does
explorer (legolas)	Claude Sonnet 4.6	"Where is X in the codebase?"	Searches the repo fast. Uses LSP, grep, glob, and git to find files, symbols, call paths, and code patterns. Returns structured results with absolute paths and a direct answer, not just a file list. Read-only by default.
researcher (radagast)	GPT-5.2	"How does library X work?"	Researches external docs, OSS repos, and the web. Classifies the question (conceptual, implementation, history, or comprehensive), fetches official sources, and returns evidence-backed recommendations with links.
implementer (celebrimbor)	GPT-5.2 Codex	"Build this end-to-end."	The workhorse implementation agent. Operates like a senior staff engineer: explores context, plans edits, executes changes, runs tests, and iterates until done. Does not stop at partial progress or ask permission for normal engineering work.
architect (elrond)	GPT-5.2	"What's the right architecture?"	Read-only strategic advisor. Analyzes codebases, evaluates tradeoffs, and returns a bottom-line recommendation with an action plan and effort estimate. Biases toward simplicity and existing patterns over speculative complexity.
analyst (galadriel)	Gemini 3.8 Flash	"What's in this image?"	Multimodal analysis specialist. Handles PNG, JPEG, and WebP images, diagrams, and charts that other agents cannot process as plain text. Extracts targeted information and returns concise findings with confidence notes. Also serves as the backend for the vision tool.
tdd (aragorn)	GPT-5.2 Codex	"Write tests for this."	Strict TDD specialist following Red-Green-Refactor. Writes a failing test first, makes it pass with minimal code, then refactors under green. Produces a test plan, TDD log, execution results, and a quality scorecard. Never writes production code without a failing test.

Commands (command/)

Slash commands available to all users.

Command	Description
/continue	Enable continuous orchestration loop for current session
/stop	Stop continuous loop and active tracked orchestration work
/tasks	List active and recent subagent orchestration tasks
/task	Show details for one subagent orchestration task
/diagnostics	Run orchestration diagnostics with live delegation probes
/doctor	Run deep orchestration doctor checks with actionable remediation
/test-orchestration	Run a parallel subagent delegation torture test audit
/fellowship-status	Show status for Fellowship status
/fellowship-list	List all bundled assets from the Fellowship configuration
/fellowship-cleanup	Remove stale cache files, broken symlinks, and unused plugins
/example	Example command demonstrating the expected format

Skills (skill/)

Reusable skill files that agents can load. Each skill is a directory containing SKILL.md with YAML frontmatter.

Skill	Description
orchestration-core	Continuous loop orchestration, delegation visibility, and diagnostics workflow
reflect	Review repeated workflows and recommend the smallest useful reusable asset
codemap	Create or update concise repository maps for unfamiliar codebases
readme-editor	Craft professional README.md files for GitHub open source projects
example	Example skill demonstrating the expected format

Runtime Plugins (src/)

Native V2 plugins compiled into the package root and activated during setup.

Plugin	Description
fellowship-orchestration	Orchestration runtime: admission limits, task tracking, tmux sessions, provider backpressure, recovery
vision-tool	PNG/JPEG/WebP analysis in a child session on the analyst agent's model, keeping image tokens out of the caller

Instructions (instruction/)

Markdown files automatically appended to agent context (registered in manifest.json).

Instruction	Description
getting-started	Quick start guide for Fellowship OpenCode
orchestration-runtime	Orchestration controls, delegation discipline, tmux runtime

MCPs (mcp/)

Model Context Protocol server configurations.

MCP	Type	Endpoint
chrome-devtools	local	npx chrome-devtools-mcp@latest
context7	remote	https://mcp.context7.com/mcp
exa	remote	https://mcp.exa.ai/mcp

Development

# Install dependencies
bun install

# Build
bun run build

# Run tests
bun test

# Type check
bun run typecheck

# Load the packed plugin and its TUI sidebar in the matching OpenCode release (needs tmux)
bun run verify:release

# Move the OpenCode SDK pins to the latest release
bun run update:opencode

# Link for local testing
npm link

Adding New Configs

See CONTRIBUTING.md for how to add new skills, agents, commands, and more.

Version Pinning

The bundled asset version is tied to the installed plugin version. To pin to a specific version:

npm install -g @unfrgivn/opencode-config@1.2.0

Check the releases for available versions.

Installed Asset Locations

• Skills: ~/.config/opencode/skill/_plugins/opencode-config/
• MCPs: ~/.config/opencode/mcp/_plugins/opencode-config/

Each generated install directory contains its own .gitignore so generated content stays out of version control.

Troubleshooting

Quick fixes for common issues:

Issue	Solution
Asset not appearing	Check disable list in fellowship.json
Override not working	Verify exact name match (case-sensitive)
Stale installs	Delete ~/.config/opencode/skill/_plugins and restart
Plugin not loading	Verify @unfrgivn/opencode-config is in opencode.json plugins

For detailed troubleshooting, see TROUBLESHOOTING.md.

License

MIT, see LICENSE