open-multi-agent

open-multi-agent

open-multi-agent is an OpenCode plugin for terminal-first, lead-driven development. It combines typed agent roles, command orchestration, tool catalogs, and workspace-backed plan state for durable execution across small to large workspaces.

What it is

A multi-agent orchestration plugin built on OpenCode's plugin interface. The system routes every user intent through a three-tier executive hierarchy — cto (strategic), lead (user-facing), sub-lead (operations) — and fans out to a catalog of embedded specialists (coders, explorers, reviewers, testers, security auditors, and more). All plan state lives in .agent/plans/ and .agent/roadmap/, giving the workspace a persistent memory across sessions.

• Turkish README: README.tr.md
• Plugin schema: schemas/open-multi-agent.schema.json
• Config template: templates/open-multi-agent.jsonc

v2 public command surface

Command	Agent	Description
/plan	lead	Analyze scope, gaps, and constraints; produce a durable plan in .agent/plans/active.md
/implement	lead	Execute the active plan by dispatching bounded work to sub-lead and coder agents
/validate	lead	Run smallest sufficient project validation (build, test, lint, typecheck, format)
/roadmap	cto	Define a multi-phase roadmap with milestones and dependency mapping
/roadmap-status	cto	Summarize current roadmap phase, progress, and blockers
/roadmap-continue	cto	Execute remaining items in the current roadmap phase sequentially

Legacy commands (subject to redesign)

/review, /status, /analyze — these commands are still available but will be reclassified or replaced in a future release.

Internal-only commands

/audit, /quality-gate, /instinct-status, /projects, /git-clean — these route to internal specialists and are not considered part of the stable public surface.

Roadmap / plan / implement / validate flow

/roadmap → cto
    ↓ defines phases & milestones → .agent/roadmap/roadmap.md + roadmap.json

/plan → lead
    ↓ repo scan → gap analysis → advisor → validator
    → .agent/plans/active.md (named plans: .agent/plans/<name>/active.md)

/implement → lead
    ↓ reads active plan → validates readiness
    → EXECUTE mode → sub-lead dispatch → coder agents (sequential or bounded parallel)
    → update_plan_task → completion / archival

/validate → lead
    ↓ detect project type → smallest sufficient validation scope
    → build / test / lint / typecheck / format-check
    → failure classification → remediation

Named plans follow the same lifecycle as active.md but live under .agent/plans/<plan-name>/.

High-level actor topology

cto (executive)
├── Roadmap-scale coordination
├── Delegates: sub-lead, internal-plan-advisor, internal-plan-validator,
│             internal-roadmap-writer, internal-docs, internal-gitmaster
└── Owns background tasks (with lead)

lead (user-facing)
├── All public commands (/plan, /implement, /validate, /roadmap*)
├── Quality gates and normalization
└── Owns background tasks (with cto)

sub-lead (operations)
├── Bounded work fan-out to coder agents
├── Routes to: internal-coder-*, internal-reviewer, internal-tester,
│             internal-debugger, internal-build-resolver
└── Does not own background tasks; must route such requests through lead or cto

Coder family

Agent	Scope
internal-coder-small	2 files / 300 lines max
internal-coder-medium	6 files / 1 500 lines max
internal-coder-strong	15 files / 5 000 lines max
internal-coder-ultrabrain	30 files / 10 000 lines max
internal-coder-ui-low	UI fixes (2 files / 500 lines)
internal-coder-ui-high	Complex UI (8 files / 3 000 lines)

Workflow specialists

internal-plan-advisor · internal-plan-validator · internal-reviewer · internal-tester · internal-debugger · internal-build-resolver · internal-security-reviewer · internal-architect · internal-e2e-runner · internal-harness-optimizer · internal-gitmaster · internal-docs · internal-coder-helper

Explorers

internal-explorer-low · internal-explorer-medium · internal-research (web-enabled)

Config v2 reference

Config file locations (one of):

• ~/.config/opencode/open-multi-agent.jsonc
• .opencode/open-multi-agent.jsonc

Minimal v2 config:

{
  "$schema": "https://opencode.ai/config.json",
  "version": 2,
  "plugin": ["open-multi-agent@latest"],
}

Key v2 sections:

{
  "version": 2,
  "agent": {
    "lead": { "promptExtension": { "prepend": "...", "append": "..." }, "disabledTools": ["..."] },
    "cto": { "model": "claude-sonnet" },
  },
  "workflow": {
    "hookProfile": "standard",
    "orchestration": {
      "maxParallelAgents": 3,
      "timeoutMs": 300000,
      "retryPolicy": "exponential",
    },
  },
  "memory": { "enabled": true, "directory": ".opencode/.memory" },
  "github": { "remote": "origin" },
  "mcp": { "context7": { "enabled": true } },
  "disabledTools": { "global": ["..."], "perAgent": { "internal-coder-small": ["..."] } },
  "commandSurface": {
    "allowDangerousCommands": false,
    "restrictedCommands": ["rm -rf"],
    "requireConfirmation": ["git push --force"],
  },
}

See the full schema at schemas/open-multi-agent.schema.json and migration notes at docs/adr/ADR-008-config-schema-v2.md.

Quick start

1) Register the plugin

{
  "$schema": "https://opencode.ai/config.json",
  "plugin": ["open-multi-agent@latest"]
}

For local development in this repository, use .opencode/plugins/open-multi-agent.ts:

export { plugin as OpenMultiAgentPlugin } from '../../src/index.ts';

2) Add plugin config

Create ~/.config/opencode/open-multi-agent.jsonc or .opencode/open-multi-agent.jsonc with version: 2. Template: templates/open-multi-agent.jsonc

3) Verify plugin load

opencode debug config

Look for the plugin entry in the resolved plugin list.

4) Run locally

bun install
mise run build

Docs hub

• Getting started
• Architecture
• Commands and plan flow
• Configuration
• ADR: Actor hierarchy
• ADR: Config schema v2
• Contributing
• Release notes

Development checks

mise run build
mise run test
mise run typecheck
mise run lint
mise run format

Direct equivalents:

bun build ./src/index.ts --outdir dist --target bun
bun test
tsc --noEmit
eslint .
prettier --check .