OpenCode Plan Manager

AI-native implementation planning for agentic workflows.

Stop losing context mid-feature or with cross-session work. OpenCode Plan Manager gives your AI agents a structured way to plan, track, and execute complex work — from a single idea to a fully shipped feature.

Why Plan Manager?

Agentic coding workflows break down when context gets too large. Agents start hallucinating, lose track of tasks, and repeat work. Plan Manager solves this with four principles:

• Selective context loading — agents read only what they need: summary (stats), spec (requirements), or plan (task list).
• Zero-hallucination schemas — strict Zod validation prevents malformed plans and invalid state transitions.
• Filesystem Kanban — plan state lives in pending/, in_progress/, and done/ folders. Atomic, human-readable, no hidden database.
• Cross-session continuity — plans are plain files committed to your repo. Pick up exactly where you left off in any new session, on any machine, with any agent — the spec and task progress are always there.

Designed for the Planner → Builder pattern: a Plan Agent architects the spec, a Build Agent executes it with no ambiguity.

Installation

Add the plugin to your OpenCode configuration (~/.config/opencode/opencode.json):

{
	"$schema": "https://opencode.ai/config.json",
	"plugin": ["[email protected]"],
}

Pinning the version improves OpenCode startup time.

Configuration

Config is loaded with the following precedence (highest → lowest):

1. Project: <project-root>/.opencode/plan-manager.json
2. User: ~/.config/opencode/plan-manager.json
3. Defaults: built-in fallback

{
	// "markdown" (default), "json", or "toon" (https://github.com/toon-format/toon)
	"outputFormat": "markdown",
}

Plan Manager writes to .opencode/plans/*. For plan_create and plan_update to work, the agents that call them must have ask or allow permission on that path.

Example config for the Plan agent:

{
	"agent": {
		"plan": {
			"permission": {
				"edit": {
					"*": "deny",
					".opencode/plans/*": "ask",
				},
			},
		},
	},
}

See opencode.ai/docs/permissions for details.

Agentic Workflow

Plan Manager is optimized for a two-agent hierarchy, with prompts in src/prompts/:

Uses OpenCode's built-in Plan and Build agents (docs) with injected system prompts. Your own custom prompts always take priority.

Filesystem Layout

.opencode/plans/
├── pending/
│   └── feature_auth/
│       ├── metadata.json
│       ├── specifications.md
│       └── implementation.md
├── in_progress/
└── done/

Each plan is an isolated folder. Status moves atomically between pending/, in_progress/, and done/ — no database, no sync issues.

API

plan_create

plan_create({
	metadata: {
		title: "JWT Authentication",
		type: "feature",
		description: "Secure auth flow with refresh tokens",
	},
	specifications: {
		description: "Implement secure JWT-based authentication",
		functionals: ["User login", "Token refresh"],
		nonFunctionals: ["Passwords hashed with bcrypt"],
		acceptanceCriterias: ["Successful login returns a valid JWT"],
		outOfScope: ["Social OAuth"],
	},
	implementation: {
		description: "Phased rollout",
		phases: [
			{
				name: "Phase 1: Database",
				tasks: [
					{ content: "Create users table", status: "pending" },
					{ content: "Create sessions table", status: "pending" },
				],
			},
		],
	},
});

plan_read — selective views

plan_read({ id: "feature_auth", view: "summary" }); // metadata + progress stats
plan_read({ id: "feature_auth", view: "spec" }); // requirements only
plan_read({ id: "feature_auth", view: "plan" }); // task list only
plan_read({ id: "feature_auth", view: "full" }); // everything

plan_update — batch task updates

plan_update({
	id: "feature_auth",
	status: "in_progress",
	taskUpdates: [
		{ content: "Create users table", status: "done" },
		{ content: "Create sessions table", status: "in_progress" },
	],
});

Development

bun install # Install dependencies
bun test    # Run test suite
bun build   # Build for production

License

MIT © 2026 Yuri Maciel