Plugin Authoring Guide

A comprehensive guide to creating Fusion plugins that extend the task board with custom tools, routes, and lifecycle hooks.

Table of Contents

1. Getting Started
2. Plugin Manifest Reference
3. Plugin Settings Schema
4. Available Hooks and Signatures
5. Registering Tools
6. Registering Routes
7. Registering UI Slots
8. Registering Top-Level Dashboard Views
   • Theming & Overlay Layering for Dashboard Views
9. Registering Agent Runtimes
10. Plugin Context API Reference
11. Plugin Lifecycle States
12. Testing Plugins
13. Publishing Plugins
14. Example Plugins
15. Registering Skills
16. Registering Workflow Steps
17. Contributing Prompt Modifications
18. Plugin Binary Setup Hooks

---

See also: External Plugin Authoring guide

1. Getting Started

What Are Fusion Plugins?

Fusion plugins extend the task board with custom functionality:

• Lifecycle Hooks: React to task creation, movement, completion, and errors
• AI Agent Tools: Add custom tools that AI agents can use during task execution
• Custom API Routes: Create dashboard API endpoints for frontend integration
• Settings: Accept user configuration via typed settings schemas

Prerequisites

• Node.js 18+
• TypeScript familiarity
• A Fusion project with the plugin system installed

Quick Start

External authors should use the standalone scaffold and dev loop (see the External Plugin Authoring guide):

fn plugin new my-first-plugin
cd my-first-plugin
pnpm install
fn plugin dev .
pnpm test

The legacy fn plugin create scaffold remains available for workspace-bound examples.

Optional AI Security Scan (Opt-in)

Plugin installs now support an opt-in aiScanOnLoad flag. When enabled, Fusion runs an AI security review before loading plugin code.

• Opt-in: disabled by default (aiScanOnLoad: false)
• When it runs: on plugin load/reload and explicit rescan
• Scan inputs (deterministic order): manifest.json, optional package.json, optional README.md, entry module, then prioritized source files
• Boundaries: excludes node_modules, dist, lockfiles, binary assets, files over 20 KB each, and enforces a 120 KB total raw-content cap

Scan Verdicts

• clean — no concerning patterns found
• warning — suspicious patterns found; plugin may still load
• blocked — dangerous patterns found; plugin is blocked before import
• error — scan failed to produce a valid decision
• unavailable — AI scan service unavailable

When a plugin is blocked (blocked/error/unavailable), Fusion does not execute plugin code for that load attempt and stores the scan result on plugin metadata (lastSecurityScan) for operator visibility.

Author Guidance for Blocked Plugins

If your plugin is blocked:

• remove dynamic execution patterns (eval, shell-outs, hidden network exfiltration behavior)
• keep behavior explicit in source and manifest
• document external calls and sensitive operations in README
• ask operators to run fn plugin rescan <id> after publishing fixes

Signature Verification and Publisher Trust

Fusion supports deterministic detached-signature verification for plugin provenance before plugin code is loaded.

Expected plugin files at the plugin root:

• manifest.json
• plugin-publisher.json (publisher identity + public key metadata)
• plugin-signature.json (detached signature over canonical payload)

Canonical payload inputs (sorted, deterministic):

• normalized manifest.json
• publisher metadata from plugin-publisher.json
• declared file digest map (sorted by relative path)

Verification statuses:

• trusted-local — bundled/in-repo plugin path trusted by local policy exception
• verified-trusted — signature verifies and publisher key is trusted
• verified-untrusted — signature verifies but publisher/key is not trusted yet
• unsigned — no signature bundle present
• invalid — signature or digest validation failed (tampered/corrupt)

Trust decisions are explicit and keyed by publisher ID + key fingerprint. Manifest author/homepage strings are informational only and are never used as trust proof.

Plugin Author Checklist for Signed Releases

1. Produce deterministic file digests for distributed plugin files
2. Publish plugin-publisher.json with stable publisher ID and key fingerprint
3. Sign the canonical payload and ship plugin-signature.json
4. Keep digest/signature files in source control and release artifacts
5. In release notes, include publisher ID + key fingerprint so operators can verify trust prompts

Plugin Project Structure

my-plugin/
├── package.json          # Plugin metadata + "fusion-plugin" keyword
├── tsconfig.json         # TypeScript configuration
├── vitest.config.ts      # Test configuration
├── src/
│   ├── index.ts         # Plugin entry point (exports default FusionPlugin)
│   └── __tests__/
│       └── index.test.ts # Plugin tests
└── README.md            # Plugin documentation

---

2. Plugin Manifest Reference

The manifest defines your plugin's metadata and capabilities:

import type { PluginManifest } from "@fusion/plugin-sdk";

const manifest: PluginManifest = {
  id: "my-custom-plugin",           // Unique identifier (kebab-case)
  name: "My Custom Plugin",          // Human-readable name
  version: "1.0.0",                  // Semver version
  description: "Does something useful",
  author: "Your Name",
  homepage: "https://github.com/you/plugin",
  fusionVersion: ">=1.0.0",         // Optional: minimum Fusion version
  dependencies: [],                   // Optional: plugin IDs this depends on
  settingsSchema: { /* ... */ },     // Optional: configuration schema
  runtime: {                         // Optional: agent runtime metadata
    runtimeId: "code-interpreter",
    name: "Code Interpreter",
    description: "Executes code in a sandbox",
  },
};

Fields

Field	Type	Required	Description
id	string	Yes	Unique identifier (kebab-case, validated)
name	string	Yes	Human-readable display name
version	string	Yes	Semver version (e.g., "1.0.0")
description	string	No	Short description
author	string	No	Author name or organization
homepage	string	No	URL to documentation or repository
fusionVersion	string	No	Minimum Fusion version required
dependencies	string[]	No	IDs of plugins this depends on
settingsSchema	Record<string, PluginSettingSchema>	No	Configuration schema
runtime	PluginRuntimeManifestMetadata	No	Agent runtime metadata for discovery

---

3. Plugin Settings Schema

Settings allow users to configure your plugin through the dashboard:

import type { PluginSettingSchema } from "@fusion/plugin-sdk";

const settingsSchema: Record<string, PluginSettingSchema> = {
  webhookUrl: {
    type: "string",
    label: "Webhook URL",
    description: "URL to send notifications to",
    required: true,
  },
  maxRetries: {
    type: "number",
    label: "Max Retries",
    description: "Maximum number of retry attempts",
    defaultValue: 3,
  },
  enabled: {
    type: "boolean",
    label: "Enable Feature",
    description: "Toggle the feature on/off",
    defaultValue: true,
  },
  severity: {
    type: "enum",
    label: "Log Severity",
    description: "Minimum severity level to log",
    enumValues: ["debug", "info", "warn", "error"],
    defaultValue: "info",
  },
};

Setting Types

Common optional fields on all setting types:

• group?: string — Optional heading used by the dashboard to render settings in grouped sections (for example: "General", "Browser", "Prompt Contributions", "Skills"). Ungrouped settings still render first in their existing flat order, then grouped sections render under stable headings.
• description?: string — Helper text shown below the setting label.
• required?: boolean — Marks the field as required.
• defaultValue?: unknown — Default value used when no user value is provided.

Type	Description	Extra Fields
"string"	Text input	multiline?: boolean (renders textarea)
"number"	Numeric input	—
"boolean"	Toggle switch	—
"enum"	Dropdown select	enumValues: string[]
"password"	Password input (hidden)	—
"array"	Dynamic list with add/remove	itemType: "string" | "number"

Example: All Setting Types

const settingsSchema: Record<string, PluginSettingSchema> = {
  // Simple string input
  username: {
    type: "string",
    label: "Username",
    description: "Your username",
  },
  
  // Multiline text area
  message: {
    type: "string",
    label: "Message",
    description: "Multi-line message",
    multiline: true,
    defaultValue: "Hello!",
  },
  
  // Password input (hidden)
  apiSecret: {
    type: "password",
    label: "API Secret",
    description: "Your secret key",
  },
  
  // Number input
  maxRetries: {
    type: "number",
    label: "Max Retries",
    defaultValue: 3,
  },
  
  // Boolean toggle
  enabled: {
    type: "boolean",
    label: "Enable Feature",
    group: "General",
    defaultValue: true,
  },
  
  // Dropdown select
  severity: {
    type: "enum",
    label: "Severity",
    enumValues: ["debug", "info", "warn", "error"],
    defaultValue: "info",
  },
  
  // Array of strings
  tags: {
    type: "array",
    label: "Tags",
    description: "Tags to track",
    group: "Skills",
    itemType: "string",
    defaultValue: ["bug", "feature"],
  },
  
  // Array of numbers
  thresholds: {
    type: "array",
    label: "Thresholds",
    itemType: "number",
    defaultValue: [10, 20, 30],
  },
};

Accessing Settings

Settings are available in hooks via ctx.settings:

hooks: {
  onLoad: (ctx) => {
    const webhookUrl = ctx.settings.webhookUrl as string;
    if (!webhookUrl) {
      ctx.logger.warn("No webhook URL configured");
    }
  },
},

---

4. Available Hooks and Signatures

Hooks let your plugin react to events in the Fusion system:

import type { FusionPlugin, PluginContext } from "@fusion/plugin-sdk";

const plugin: FusionPlugin = {
  manifest: { /* ... */ },
  state: "installed",
  hooks: {
    onLoad: async (ctx) => {
      ctx.logger.info("Plugin loaded!");
    },
    onTaskCreated: async (task, ctx) => {
      ctx.logger.info(`New task: ${task.title}`);
    },
    // ... other hooks
  },
};

Hook Reference

Hook	Signature	When It Fires
onLoad	(ctx: PluginContext) => Promise<void> | void	Plugin first loaded and started
onUnload	(ctx: PluginContext) => Promise<void> | void	Plugin stopped/shutdown
onTaskCreated	(task: Task, ctx: PluginContext) => Promise<void> | void	New task created
onTaskMoved	(task: Task, fromColumn: string, toColumn: string, ctx: PluginContext) => Promise<void> | void	Task moved between columns
onTaskCompleted	(task: Task, ctx: PluginContext) => Promise<void> | void	Task reached "done"
onError	(error: Error, ctx: PluginContext) => Promise<void> | void	Error occurred in plugin execution
onPostgresSchemaInit	() => PluginPostgresSchemaDefinition	Before onLoad; Fusion validates and applies the declarative plan with a short-lived migration connection
onSchemaInit (legacy)	(db: Database) => Promise<void> | void	SQLite-only compatibility declaration; unsupported for third-party plugins in the PostgreSQL runtime
executorRuntimeEnv	(taskCtx: ExecutorRuntimeTaskContext, ctx: PluginContext) => Promise<ExecutorRuntimeEnvContribution> | ExecutorRuntimeEnvContribution	Before executor-spawned task commands run, to contribute task-scoped env and PATH prepends

Hook Behavior

• Single-load lifecycle: For one project in one Fusion process, Fusion invokes onLoad exactly once for each intentional load lifecycle, including when the host and engine bootstrap concurrently. Plugin authors do not need process-local locking to defend against an accidental second host/engine startup load. Explicit enable→load and reloadPlugin are new lifecycles and can invoke onLoad again after unload; request-scoped temporary loaders for another project root may also load and then stop a plugin while discovering skills. Keep registration idempotent where inexpensive so these intentional lifecycles remain safe.
• Context parity: onUnload receives the same PluginContext shape as onLoad.
• Timeout: 5 seconds per invocation (logged and skipped if exceeded)
• Error Isolation: Hook failures never block other hooks or abort startup
• Optional: Only define the hooks you need
• Schema preflight: onPostgresSchemaInit is evaluated and validated before the plugin is marked started or onLoad runs. An invalid or SQLite-only third-party schema fails without leaving onLoad subscriptions or timers behind.
• No privileged handle: The hook returns data and receives no database object. Fusion alone opens the short-lived migration connection; ordinary plugin hooks and routes continue to use the project-bound forced-RLS runtime role.
• Allowed DDL: Return one idempotent CREATE TABLE IF NOT EXISTS, CREATE [UNIQUE] INDEX IF NOT EXISTS, or ALTER TABLE statement per array item. Every object must live in the project schema and every referenced table must begin with the declared tablePrefix. Semicolon-separated batches and data-changing/admin statements are rejected.
• Required ownership: Every created table declares project_id text NOT NULL and a project_id-leading composite primary key. Fusion installs the column default, ownership trigger, forced RLS policy, and runtime grants after creation. Composite foreign keys should include project_id on both sides.
• Schema evolution: Increment version when the declarative plan changes. Statements remain idempotent and must safely rerun during restart or hot reload. Avoid data backfills or long-running logic.
• Legacy cutover: Third-party onSchemaInit(Database) hooks are no longer executable in the PostgreSQL runtime. Bundled plugins retain host-owned PostgreSQL equivalents during migration, but external plugins must provide onPostgresSchemaInit.

Example: Schema initialization hook

hooks: {
  onPostgresSchemaInit: () => ({
    version: 1,
    tablePrefix: "acme_",
    statements: [
      `CREATE TABLE IF NOT EXISTS project.acme_roadmaps (
        project_id text NOT NULL,
        id text NOT NULL,
        title text NOT NULL,
        created_at text NOT NULL,
        PRIMARY KEY (project_id, id)
      )`,
      `CREATE INDEX IF NOT EXISTS idx_acme_roadmaps_created_at
       ON project.acme_roadmaps(project_id, created_at)`,
    ],
  }),
},

executorRuntimeEnv: task-scoped executor subprocess environment

Use executorRuntimeEnv when your plugin needs to provide runtime environment values for executor-spawned user commands tied to a specific task.

This hook runs when Fusion prepares subprocess environments for executor command surfaces such as configured commands, verification commands, and step-session subprocesses.

It does not apply to internal git plumbing subprocesses used by worktree/branch management.

import type {
  ExecutorRuntimeEnvContribution,
  ExecutorRuntimeTaskContext,
  PluginContext,
} from "@fusion/plugin-sdk";

function executorRuntimeEnv(
  taskCtx: ExecutorRuntimeTaskContext,
  ctx: PluginContext,
): ExecutorRuntimeEnvContribution {
  ctx.logger.info(`Preparing env for task ${taskCtx.taskId}`);
  return {
    pathPrepend: ["/absolute/path/to/tools"],
    env: {
      MY_PLUGIN_TASK_ID: taskCtx.taskId,
    },
  };
}

ExecutorRuntimeTaskContext fields:

• taskId: Fusion task ID
• worktreePath: absolute path to the task worktree
• rootDir: project root directory
• branch?: task branch name when available

ExecutorRuntimeEnvContribution fields:

• pathPrepend?: array of absolute path strings prepended to PATH for executor-spawned commands
• env?: key/value string map merged into the subprocess environment
• description?: optional human-readable note for debugging/telemetry

Validation and merge behavior (from engine runtime collection):

• pathPrepend must be an array of absolute path strings; non-absolute entries are rejected.
• env values must be strings.
• env must not include PATH; use pathPrepend instead.
• When multiple plugins set the same env key, later plugins override earlier values and the engine logs a warning.
• pathPrepend entries from later plugins are placed earlier in the final prepend list.

Example: prepend a generated tool directory for executor commands

import { definePlugin } from "@fusion/plugin-sdk";
import path from "node:path";

export default definePlugin({
  manifest: {
    id: "fusion-plugin-tooling-example",
    name: "Tooling Example",
    version: "1.0.0",
  },
  state: "installed",
  hooks: {},
  executorRuntimeEnv: (taskCtx) => {
    const toolDir = path.resolve(taskCtx.rootDir, ".fusion/tools/my-plugin/bin");

    return {
      pathPrepend: [toolDir],
      env: {
        MY_PLUGIN_TOOL_HOME: toolDir,
      },
    };
  },
});

With this hook enabled, executor-spawned commands (for example verification commands or bash tool subprocesses in the task session) can resolve binaries from toolDir without mutating global process environment.

Example: Notification on Task Completion

hooks: {
  onTaskCompleted: async (task, ctx) => {
    const webhookUrl = ctx.settings.webhookUrl as string;
    if (!webhookUrl) return;

    await fetch(webhookUrl, {
      method: "POST",
      headers: { "Content-Type": "application/json" },
      body: JSON.stringify({
        text: `✅ Task completed: ${task.title || task.id}`,
      }),
    });
  },
},

---

5. Registering Tools

Tools extend AI agents with custom capabilities:

import type { FusionPlugin, PluginToolDefinition, PluginToolResult } from "@fusion/plugin-sdk";

const myTool: PluginToolDefinition = {
  name: "my_custom_tool",
  description: "Does something useful with input text",
  parameters: {
    type: "object",
    properties: {
      input: {
        type: "string",
        description: "The text to process",
      },
    },
    required: ["input"],
  },
  execute: async (params, ctx) => {
    const input = params.input as string;

    // Do something useful...
    const result = input.toUpperCase();

    return {
      content: [{ type: "text", text: result }],
    };
  },
};

const plugin: FusionPlugin = {
  manifest: { /* ... */ },
  state: "installed",
  tools: [myTool],
};

Tool Naming

• Use a unique name prefixed with your plugin ID (e.g., my-plugin_action)
• Avoid conflicts with built-in tools

Tool Result Format

interface PluginToolResult {
  content: Array<{ type: "text"; text: string }>;
  isError?: boolean;
  details?: Record<string, unknown>;
}

---

6. Registering Routes

Routes create custom API endpoints in the dashboard:

import type { FusionPlugin, PluginRouteDefinition } from "@fusion/plugin-sdk";

const routes: PluginRouteDefinition[] = [
  {
    method: "GET",
    path: "/status",
    description: "Get plugin status",
    handler: async (req, ctx) => {
      return { status: "ok", uptime: process.uptime() };
    },
  },
  {
    method: "POST",
    path: "/action",
    description: "Perform an action",
    handler: async (req, ctx) => {
      // Access request body
      const body = req as { action?: string };
      ctx.logger.info(`Action: ${body.action}`);
      return { success: true };
    },
  },
];

const plugin: FusionPlugin = {
  manifest: { /* ... */ },
  state: "installed",
  routes,
};

Route handlers may return either plain JSON values or a PluginRouteResponse envelope. PluginRouteResponse now supports optional headers and contentType fields so plugins can serve non-JSON payloads like downloadable HTML. Example: return { status: 200, body: html, contentType: "text/html; charset=utf-8", headers: { "Content-Disposition": "attachment; filename=\"report.html\"" } } to send an attachment response directly from a plugin route.

Route Mounting

Routes are mounted at /api/plugins/{pluginId}/{path}. Route handlers receive the same loader-built PluginContext used by hooks/tools, including real taskStore, plugin settings, logger, emitEvent, and engine-injected createAiSession (when available):

• Example roadmap plugin route: path: "/roadmaps" in plugin fusion-plugin-roadmap resolves to /api/plugins/fusion-plugin-roadmap/roadmaps

• Roadmap suggestion endpoints follow the same namespace (for example /api/plugins/fusion-plugin-roadmap/roadmaps/:roadmapId/suggestions/milestones)

• Do not document or depend on legacy host-owned /api/roadmaps routes unless your current source still ships them

• Plugin ID: fusion-plugin-notification

• Route path: /status

• Full URL: /api/plugins/fusion-plugin-notification/status

UI metadata endpoints

Frontend integrations can enumerate the host UI before presenting navigation or settings commands. Both endpoints inherit the dashboard's standard /api authentication, are read-only, return static project-independent metadata, and do not require a project id.

GET /api/views returns every registered built-in view id in dashboard order. This is the full registry, not a live navigation menu: it includes ids that are flag-gated or experimental (for example graph, todos, secrets), and internal, non-navigable destinations (see internal below). Whether any given id is currently reachable in the UI depends on feature flags and plugins the endpoint does not evaluate, so treat the list as the set of known view ids rather than a guaranteed set of visible navigation entries.

{
  "views": [
    { "id": "board", "label": "Board", "labelKey": "nav.board" },
    { "id": "command-center", "label": "Dashboard", "labelKey": "nav.commandCenter" },
    {
      "id": "dev-server",
      "label": "Dev Server",
      "labelKey": "nav.devServer",
      "aliases": ["devserver"]
    },
    {
      "id": "task-detail",
      "label": "Task Detail",
      "internal": true
    }
  ]
}

aliases lists accepted legacy ids; use the canonical id for new links. internal: true identifies a programmatic destination that is not a normal navigation entry. Optional fields are omitted when they do not apply.

label is the guaranteed English display string. labelKey is present only for views whose title the dashboard itself renders through that translation key, so a few ids (for example graph, whose label comes from a plugin manifest, and the internal task-detail) carry no labelKey at all — fall back to label when it is absent. Note also that some keys are not yet present in the shipped translation catalogs because the dashboard supplies their English text inline; resolving such a key yields nothing, so treat labelKey as best-effort localization support rather than a guaranteed lookup.

GET /api/settings/sections returns selectable Settings sections; non-selectable group-header rows are excluded:

{
  "sections": [
    {
      "id": "appearance",
      "label": "Appearance",
      "labelKey": "settings.nav.appearance",
      "scope": "global",
      "group": "Preferences",
      "keywords": ["theme", "color", "sidebar"],
      "searchableKeys": [],
      "advanced": false
    },
    {
      "id": "authentication",
      "label": "Authentication",
      "labelKey": "settings.nav.authentication",
      "scope": null,
      "group": "AI & Models",
      "keywords": ["login", "OAuth", "API key"],
      "searchableKeys": [],
      "advanced": false
    }
  ]
}

scope is global, project, or null for sections backed by a dedicated subsystem rather than a settings blob. group is the Settings navigation group label, keywords and searchableKeys support command/search matching, and advanced indicates whether the dashboard hides the section until Advanced settings are enabled.

keywords and searchableKeys are best-effort, non-contractual search hints, not a stable API surface. searchableKeys in particular exposes raw i18n translation-key strings that back a section's searchable copy; their exact values, ordering, and presence may change between releases as the dashboard's internal translation keys evolve. Use them to widen local search matching, but do not treat any specific key string as a stable identifier or depend on it programmatically.

Supported Methods

• GET
• POST
• PUT
• DELETE

---

7. Registering UI Slots

UI slots are mount points in the Fusion dashboard where plugins can inject custom UI components. Each slot is identified by a unique slotId that corresponds to a specific location in the dashboard UI.

How UI Slots Work

Plugins declare uiSlots in their FusionPlugin definition. The dashboard discovers all registered UI slots via GET /api/plugins/ui-slots and renders matching components at each mount point.

Available Slot IDs

Slot ID	Location	Description	Status
task-detail-tab	Task detail modal	Tab added to the task detail view	Available
header-action	Dashboard header	Action button in the header toolbar	Available
settings-section	Settings modal	Section added to the settings panel	Available
settings-provider-card	Settings → Authentication	Provider card contribution in Authentication section	Available
onboarding-provider-card	Onboarding modal → AI setup	Provider card content rendered before host fallback cards	Available
onboarding-setup-help	Onboarding modal → AI setup	Additional setup-help content rendered below provider sections	Available
post-onboarding-recommendation	Dashboard post-onboarding card	Recommendation item rendered in host-owned next-steps container	Available
settings-integration-card	Legacy structured name	Compatibility alias now normalized to settings-config-section in structured API	Compatibility only
onboarding-recommendation-card	Legacy structured name	Compatibility alias now normalized to onboarding-provider-recommendation in structured API	Compatibility only
task-card-badge	Task card on the board	Small badge displayed on task cards (e.g., CI status indicator)	Planned
board-column-footer	Board column	Footer area below the last card in a column	Planned

Defining UI Slots

Add uiSlots to your FusionPlugin definition:

import type { FusionPlugin, PluginUiSlotDefinition } from "@fusion/plugin-sdk";

const uiSlots: PluginUiSlotDefinition[] = [
  {
    slotId: "task-detail-tab",
    label: "CI History",
    icon: "history",
    componentPath: "./components/ci-tab.js",
  },
  {
    slotId: "task-card-badge",
    label: "CI Status",
    icon: "circle-check",
    componentPath: "./components/ci-badge.js",
  },
];

const plugin: FusionPlugin = {
  manifest: { /* ... */ },
  state: "installed",
  uiSlots,
  hooks: { /* ... */ },
};

PluginUiSlotDefinition Fields

Field	Type	Required	Description
slotId	string	Yes	One of the known slot IDs (e.g., "task-detail-tab", "header-action")
label	string	Yes	Human-readable label for the slot
icon	string	No	Lucide icon name for visual identification
componentPath	string	Yes	Path to the JS module exporting the component, relative to the plugin root

Component Module Format and Host Resolution

componentPath is part of the plugin contract, but dashboard rendering is intentionally host-resolved through a static slot registry (pluginId + slotId + componentPath).

Important implications:

• The dashboard does not load arbitrary plugin modules at runtime from componentPath.
• To render in dashboard flows, your slot entry must match a host-registered mapping.
• Unknown/unmapped entries degrade safely to a visible “missing component” shell (or render nothing when the host sets renderPlaceholder={false}).
• Host flows still own modal structure, navigation, callbacks, and fallback content when no slot entry exists.

For this reason, plugin authors should still provide stable componentPath values in manifests, but coordinate with dashboard host maintainers when adding new UI surfaces or module paths that need mapping.

Structured UI Contributions (data-only, parallel to uiSlots)

For Settings/onboarding/post-onboarding flows, use structured uiContributions instead of legacy placeholder slots.

• Discovery API: GET /api/plugins/ui-contributions
• Type surface: PluginUiContributionDefinition (@fusion/plugin-sdk)
• Structured surfaces:
  • settings-provider-card
  • settings-config-section
  • onboarding-provider-card
  • onboarding-setup-help
  • onboarding-provider-recommendation
  • post-onboarding-recommendation

Rules:

• Structured contributions are data-only JSON payloads.
• Do not include componentPath.
• Do not send live callbacks/functions through REST.
• Use actions: PluginUiActionDescriptor[] so host-owned renderers bind behavior.

Compatibility normalization:

• settings-integration-card → settings-config-section
• onboarding-recommendation-card → onboarding-provider-recommendation

The API only returns normalized surface names.

---

8. Registering Top-Level Dashboard Views

Top-level views are a sibling contribution type to uiSlots.

• uiSlots are embedded surfaces (task detail tab, header action, etc.)
• dashboardViews is the shipped top-level plugin field for full-screen dashboard destinations
• Earlier planning language may say views; the implemented API in FusionPlugin is dashboardViews

Register dashboardViews on the plugin definition:

import type { PluginDashboardViewDefinition } from "@fusion/plugin-sdk";

const dashboardViews: PluginDashboardViewDefinition[] = [
  {
    viewId: "graph",
    label: "Graph",
    componentPath: "./src/DependencyGraphView.tsx",
    icon: "Network",
    order: 40,
    placement: "overflow",
    description: "Explore task dependency links",
  },
];

PluginDashboardViewDefinition fields

Field	Type	Required	Description
viewId	string	Yes	Unique slug-like ID within your plugin namespace
label	string	Yes	Human-readable nav label
componentPath	string	Yes	Module path for the view component, relative to plugin root
icon	string	No	Lucide icon name
order	number	No	Lower values appear earlier in nav
placement	"primary" | "overflow" | "more"	No	Navigation placement hint (default is host-defined overflow behavior)
description	string	No	Short summary for nav/help surfaces

Current host constraints:

• Discovery API: GET /api/plugins/dashboard-views
• The dashboard does not eval or filesystem-load plugin code in-browser
• componentPath is stored for authoring symmetry/future expansion, but render resolution is currently done through a host-side static registry (pluginId + viewId)
• Use stable IDs; runtime view key format is plugin:${pluginId}:${viewId}

Static host registry model

Dashboard view components are resolved from a host-side registry and must be explicitly registered:

import { lazy } from "react";
import { registerPluginView } from "../app/plugins/pluginViewRegistry";

registerPluginView(
  "fusion-plugin-dependency-graph",
  "graph",
  lazy(() => import("@fusion-plugin-examples/dependency-graph/dashboard-view")),
);

The host then renders plugin views via PluginDashboardViewHost using the composite ID.

Bundled workspace plugin pattern:

• Keep plugin package under plugins/ (for example plugins/fusion-plugin-roadmap)
• Export backend/plugin entry from src/index.ts and keep dashboard view exports in the plugin package (for example ./dashboard-view)
• Do not re-export dashboard view entrypoints (including dashboard-view, manage-view, or other -view modules) from src/index.ts; the fusion/no-plugin-view-reexport ESLint guard catches violations because server-side plugin entries must not transitively load CSS
• Register the lazy dashboard component in host code (currently packages/dashboard/app/plugins/registerBundledPluginViews.ts)
• CLI bundling inlines backend plugin code from workspace packages; dashboard view modules are imported by the dashboard build via the host registry

Bundled bundled.js outputs must be self-contained at runtime. Do not leave private workspace package imports such as @fusion/core in emitted bundled plugin code; @fusion/plugin-sdk core runtime re-exports are resolved through the CLI runtime shim during packaging.

If a bundled plugin reads package-local files at runtime, stage those assets explicitly during CLI packaging. bundlePluginEntry() copies any committed src/skills/ directory into dist/plugins/<id>/skills/; use the same pattern for similar runtime-read assets instead of assuming esbuild will include files that are never imported.

Bundled plugin build-freshness guard

Bundled plugins shipped in @runfusion/fusion are tracked by the staged bundled-plugin set in packages/cli/src/plugins/staged-bundled-plugin-ids.ts; the default auto-install subset remains BUNDLED_PLUGIN_IDS in packages/cli/src/plugins/bundled-plugin-install.ts. The CLI build asserts every staged bundled plugin has a loadable entry under packages/cli/dist/plugins/<id>/, and the freshness test checks any per-plugin plugins/<id>/dist/index.js that exists against the newest src/** mtime.

This catches stale dist/ drift: resolvePluginEntryPath prefers bundled.js and compiled dist/index.js before falling back to src/index.ts, while per-plugin dist/ is gitignored and can lag behind source edits. If bundled-plugin-freshness reports dist is stale relative to src, run pnpm build from the workspace root to regenerate plugin dist/ outputs and staged CLI plugin artifacts before rerunning tests.

Bundled-plugin auto-install: CLI and desktop

The install/update/fail-soft-load logic for BUNDLED_PLUGIN_IDS (Dependency Graph, Hermes, OpenClaw, Paperclip, Cursor, Grok, CLI Printing Press, Compound Engineering, Linear Import, Reports, WhatsApp Chat, Roadmap) lives once in packages/core/src/plugins/bundled-plugin-install.ts as ensureBundledPluginInstalled(pluginStore, pluginLoader, pluginId, getCandidatePluginDirs). The only host-specific input is getCandidatePluginDirs — the ordered list of directories to probe for a plugin's manifest.json — because the CLI and desktop stage bundled plugins differently:

• CLI (packages/cli/src/plugins/bundled-plugin-install.ts): resolves <cli>/dist/plugins/<manifest-id>/ (plus source/dev fallbacks) from its own import.meta.url. The CLI module is now a thin adapter that supplies this resolver and re-exports the same public surface (ensureBundledPluginInstalled, ensureBundledDependencyGraphPluginInstalled, ensureBundledCursorRuntimePluginInstalled, isBundledPluginId, BUNDLED_PLUGIN_IDS, resolvePluginEntryPath) that dashboard.ts, serve.ts, and daemon.ts already depend on — no behavior change for CLI hosts.
• Desktop (packages/desktop/src/bundled-plugin-dirs.ts): resolves each manifest id (fusion-plugin-<short-name>) to its staged @fusion-plugin-examples/<short-name> npm package directory via import.meta.resolve against the package's "." export, which works whether node_modules is flat/hoisted (the packaged pnpm deploy closure) or nested (workspace dev). These plugins are workspace dependencies of @fusion/dashboard, so packages/desktop/scripts/workspace-tools.ts#stageDesktopDeploy already materializes their manifest.json + dist/index.js into the desktop closure — no desktop-specific build/staging changes were needed. A bundled id desktop does not depend on (currently reports, whatsapp-chat, linear-import) resolves to no candidate directories and is correctly reported as missing-bundle, matching the CLI's "not found in this build" behavior for an unstaged plugin.

Both packages/desktop/src/local-runtime.ts (createDashboardServerDefault) and packages/desktop/src/local-server.ts (DesktopLocalServerManager.start) wire this identically to the CLI dashboard command: auto-install the bundled Dependency Graph plugin before loadAllPlugins(), and pass an ensureBundledPluginInstalled callback into createServer(...) options so PUT /api/plugins/:id/settings can lazy-install Hermes/OpenClaw/Paperclip/etc. on first Settings save. Both paths run this inside the existing fail-soft try/catch (see the desktop pluginStore/pluginLoader wiring, FN-7623) so a plugin auto-install failure logs (via the FUSION_STARTUP_TRACE strace helper in local-runtime.ts) but never crashes embedded startup. packages/desktop still does not depend on the CLI package (@runfusion/fusion) — the shared helper lives entirely in @fusion/core.

Runtime host context contract:

• Registered views receive a context object from the dashboard host (PluginDashboardViewContext).
• Context includes the active projectId, current visible tasks, optional workflowSteps, openTaskDetail for launching the native task detail flow, and openFile(path, options?) for opening project-relative files in the dashboard's built-in file viewer.
• Keep view-specific UI behavior in the plugin; treat host context as service/data injection only.

Placement guidance:

• primary: top-level nav tab (host may limit count on mobile)
• overflow: desktop header overflow menu
• more: mobile More sheet / secondary nav surfaces

Project-scoped UI state guidance:

• Persist plugin view layout/state in browser storage using a plugin-owned base key and the shared project-scoped pattern (kb:${projectId}:${baseKey}).
• For dependency graph layout, the canonical base key is fusion-plugin-dependency-graph:positions.
• Do not persist plugin UI state in task metadata or server-side task records.

Theming & Overlay Layering for Dashboard Views

Use only the stable theme token contract for plugin UI. It provides supported surface, text, spacing, status, motion, and layering variables; internal CSS names may change without notice.

.my-plugin-panel {
  background: var(--surface);
  color: var(--text);
}

.my-plugin-owned-overlay {
  position: fixed;
  inset: 0;
  z-index: calc(var(--fusion-max-z) + 1);
  pointer-events: auto;
}

For overlays that should share Fusion's root stacking context, append the plugin element to the supported #plugin-overlay-root mount point:

const overlayRoot = document.querySelector("#plugin-overlay-root");
overlayRoot?.append(pluginOverlayElement);

The mount point is fixed and click-through; set pointer-events: auto on interactive plugin children. Its layer follows --fusion-max-z as Fusion's monotonic floating-window stack rises, so plugins do not need to track dashboard window focus in JavaScript.

---

9. Registering Agent Runtimes

Plugins can provide custom agent runtime implementations that extend the Fusion engine's ability to execute agent sessions. Runtimes are discovered through the plugin discovery pipeline and can be used by the engine to route agent session creation.

How Runtimes Work

A plugin runtime consists of two parts:

1. Runtime Metadata (in manifest): Declares the runtime's identity for discovery
2. Runtime Factory (in plugin instance): Creates the runtime instance when needed

Runtime Manifest Metadata

Declare runtime metadata in your plugin's manifest:

import type { PluginManifest } from "@fusion/plugin-sdk";

const manifest: PluginManifest = {
  id: "my-runtime-plugin",
  name: "My Runtime Plugin",
  version: "1.0.0",
  runtime: {
    runtimeId: "code-interpreter",
    name: "Code Interpreter",
    description: "Executes code in a sandboxed environment",
    version: "1.0.0",
  },
};

PluginRuntimeManifestMetadata Fields

Field	Type	Required	Description
runtimeId	string	Yes	Unique runtime identifier within the plugin (kebab-case slug)
name	string	Yes	Human-readable name for the runtime
description	string	No	Short description of what the runtime provides
version	string	No	Semantic version of the runtime implementation

Runtime Factory

The runtime factory is a function that creates the runtime instance:

import type { FusionPlugin, PluginContext } from "@fusion/plugin-sdk";

const plugin: FusionPlugin = {
  manifest: { /* ... */ },
  state: "installed",
  hooks: {},
  runtime: {
    metadata: {
      runtimeId: "code-interpreter",
      name: "Code Interpreter",
      description: "Executes code in a sandboxed environment",
    },
    factory: async (ctx: PluginContext) => {
      // Initialize the runtime with plugin context
      const apiKey = ctx.settings.apiKey as string;
      
      return {
        name: "code-interpreter",
        version: "1.0.0",
        async execute(code: string) {
          // Execute code in sandbox
          const result = await runSandbox(code, { apiKey });
          return result;
        },
      };
    },
  },
};

PluginRuntimeFactory Signature

type PluginRuntimeFactory = (ctx: PluginContext) => Promise<unknown> | unknown;

The factory receives PluginContext (same as hooks) and should return the runtime instance. The returned instance's structure depends on the runtime's purpose.

PluginRuntimeRegistration Structure

interface PluginRuntimeRegistration {
  metadata: PluginRuntimeManifestMetadata;
  factory: PluginRuntimeFactory;
}

Discovery Pipeline

Runtimes are discovered through the plugin discovery pipeline:

1. PluginLoader aggregates runtime registrations from all loaded plugins via getPluginRuntimes()
2. PluginRunner caches runtime registrations and exposes them via getPluginRuntimes()

Both components follow the same pattern as tools, routes, and UI slots.

Backwards Compatibility

Runtime registration is entirely optional. Plugins that don't declare a runtime field:

• Continue to work unchanged
• Pass manifest validation
• Don't affect the runtime discovery pipeline

Example: Complete Runtime Plugin

import { definePlugin } from "@fusion/plugin-sdk";

export default definePlugin({
  manifest: {
    id: "fusion-plugin-code-interpreter",
    name: "Code Interpreter Plugin",
    version: "1.0.0",
    description: "Provides a sandboxed code execution runtime",
    runtime: {
      runtimeId: "code-interpreter",
      name: "Code Interpreter",
      description: "Executes code in a sandboxed environment",
      version: "1.0.0",
    },
  },
  state: "installed",
  hooks: {
    onLoad: async (ctx) => {
      ctx.logger.info("Code interpreter runtime plugin loaded");
    },
  },
  runtime: {
    metadata: {
      runtimeId: "code-interpreter",
      name: "Code Interpreter",
      description: "Executes code in a sandboxed environment",
      version: "1.0.0",
    },
    factory: async (ctx) => {
      // Runtime initialization
      return {
        name: "code-interpreter",
        version: "1.0.0",
        async execute(code: string, language: string) {
          // Sandbox execution logic
          return { output: `Executed ${language} code`, result: "success" };
        },
      };
    },
  },
} satisfies FusionPlugin);

---

10. Plugin Context API Reference

The context object is passed to hooks, tools, and route handlers:

interface PluginContext {
  pluginId: string;
  taskStore: TaskStore;
  settings: Record<string, unknown>;
  logger: PluginLogger;
  emitEvent: (event: string, data: unknown) => void;
  createAiSession?: CreateAiSessionFactory;
}

Properties

Property	Type	Description
pluginId	string	Your plugin's unique ID
taskStore	TaskStore	Access to task data (read-only)
settings	Record<string, unknown>	User configuration (merged with defaults)
logger	PluginLogger	Structured logging
emitEvent	(event, data) => void	Emit custom events
createAiSession	CreateAiSessionFactory | undefined	Engine-injected AI session factory (undefined when engine isn't loaded)

Durable data access: PostgreSQL / AsyncDataLayer

Production Fusion hosts are PostgreSQL-only. Plugin routes, hooks, and dashboard-backed feature paths must not call ctx.taskStore.getDatabase(): it is the legacy synchronous SQLite accessor and throws in backend mode. Use a project-bound AsyncDataLayer and make store methods asynchronous instead.

function getWidgetStore(ctx: PluginContext): AsyncWidgetStore {
  const asyncLayer = ctx.taskStore.getAsyncLayer();
  if (!asyncLayer) {
    throw new Error("Widget plugin requires PostgreSQL AsyncDataLayer");
  }
  return new AsyncWidgetStore(asyncLayer);
}

const widget = await getWidgetStore(ctx).getWidget(widgetId);

Use a direct drizzle-orm dependency for plugin-owned PostgreSQL tables and scope every query by asyncLayer.projectId. See Reports' getReportStore / ReportStore async siblings and Quality's AsyncQualityStore for production patterns. SQLite/DatabaseSync is permitted only inside intentional unit harnesses, never as a production fallback.

The repository gate scripts/check-no-getdatabase.mjs scans tracked plugin, dashboard, engine, and core paths. A legitimate transitional exception must be reviewed in scripts/lib/getdatabase-allowlist.json and pin the exact file, one-based line, and trimmed snippet; there is no file-level or inline exemption. New plugin feature code is not eligible for that allowlist.

Logger Methods

interface PluginLogger {
  info(message: string, ...args: unknown[]): void;
  warn(message: string, ...args: unknown[]): void;
  error(message: string, ...args: unknown[]): void;
  debug(message: string, ...args: unknown[]): void;
}

createAiSession API

interface CreateAiSessionOptions {
  cwd: string;
  systemPrompt: string;
  tools?: "coding" | "readonly";
  defaultProvider?: string;
  defaultModelId?: string;
}

interface AiSessionResult {
  session: {
    prompt(text: string): Promise<void>;
    state: { messages: Array<{ role: string; content?: unknown }> };
  };
  sessionFile?: string;
}

type CreateAiSessionFactory = (
  options: CreateAiSessionOptions,
) => Promise<AiSessionResult>;

The factory is dependency-injected by the engine at runtime. In test-only or core-only environments where the engine module is not loaded, ctx.createAiSession is undefined, so guard before calling it.

Example: Using ctx.createAiSession()

Use this context factory for plugin AI features (for example roadmap milestone/feature suggestion generation). Avoid direct @fusion/engine imports from plugin code; engine wiring is injected by the host through PluginContext.

hooks: {
  onLoad: async (ctx) => {
    if (!ctx.createAiSession) {
      ctx.logger.warn("AI session factory unavailable; engine not loaded");
      return;
    }

    const { session } = await ctx.createAiSession({
      cwd: process.cwd(),
      systemPrompt: "You are a release assistant for this plugin.",
      tools: "readonly",
    });

    await session.prompt("Summarize what this plugin contributes.");
    const latest = session.state.messages.at(-1);
    ctx.logger.info("AI summary generated", latest);
  },
},

Example: Using the Context

hooks: {
  onLoad: (ctx) => {
    ctx.logger.info("Plugin starting...");

    // Access settings
    const apiKey = ctx.settings.apiKey as string;

    // Emit custom event
    ctx.emitEvent("my-plugin:ready", { timestamp: Date.now() });
  },
},

---

11. Plugin Lifecycle States

Plugins transition through these states:

┌────────────┐
│ installed  │ (registered, not loaded)
└─────┬──────┘
      │ enable
      ▼
┌────────────┐
│  started   │ ←─────┐ (loaded, hooks active)
└─────┬──────┘       │
      │              │ load
      │ stop         │
      ▼              │
┌────────────┐       │
│  stopped   │ ──────┘
└────────────┘

Any state can transition to:
┌────────────┐
│   error    │ (load failure or runtime error)
└────────────┘

State Descriptions

State	Description
installed	Plugin registered but not yet loaded
started	Plugin loaded and hooks active
stopped	Plugin shut down gracefully
error	Plugin failed during load or execution

Updating path-registered plugins

For plugins installed from a filesystem path, the normal update loop is now:

1. Pull or edit the plugin source.
2. Rebuild the plugin entrypoint if your plugin uses a build step.
3. Restart Fusion, or disable and re-enable the plugin.

On the next load/reload, Fusion re-imports the plugin module and refreshes the persisted manifest version and settingsSchema from the rebuilt manifest. Existing per-project enablement and saved setting values are preserved, so you do not need to unregister and re-register a path-based plugin just to expose a new version or settings field.

---

12. Testing Plugins

Use Vitest for unit testing your plugins:

Test Structure

// src/__tests__/index.test.ts
import { describe, it, expect, vi, beforeEach } from "vitest";
import plugin from "../index.js";

describe("my plugin", () => {
  beforeEach(() => {
    vi.clearAllMocks();
  });

  it("should export a valid plugin", () => {
    expect(plugin.manifest.id).toBe("my-plugin");
    expect(plugin.manifest.name).toBeDefined();
  });

  it("should call onLoad hook", async () => {
    const mockCtx = {
      pluginId: "my-plugin",
      settings: {},
      logger: {
        info: vi.fn(),
        warn: vi.fn(),
        error: vi.fn(),
        debug: vi.fn(),
      },
      emitEvent: vi.fn(),
      taskStore: {},
    };

    await plugin.hooks.onLoad?.(mockCtx as any);
    expect(mockCtx.logger.info).toHaveBeenCalled();
  });
});

Testing Tools

it("should return correct result from tool", async () => {
  const tool = plugin.tools![0];
  const mockCtx = { /* ... */ };

  const result = await tool.execute({ input: "hello" }, mockCtx as any);

  expect(result.content[0].text).toBe("HELLO");
});

Testing Routes

it("should return status from GET /status", async () => {
  const route = plugin.routes!.find(r => r.path === "/status");
  const req = { params: {}, method: "GET", url: "/status" };
  const ctx = { /* ... */ };

  const result = await route.handler(req as any, ctx as any);

  expect(result).toHaveProperty("status");
});

Running Tests

pnpm test

Fusion Host Regression Coverage (Plugin Discovery / Load / Registration)

When a plugin changes host integration contracts, add or update host-side regression tests in this repository:

• Core loader pipeline (packages/core/src/__tests__/plugin-loader.test.ts): verify PluginStore.registerPlugin() → PluginLoader.loadAllPlugins() / loadPlugin() and assert started state transitions, manifest validation failures, disabled-plugin skip behavior, missing entrypoint failures, and onLoad error handling.
• Dashboard API aggregation (packages/dashboard/src/__tests__/plugin-routes.test.ts, packages/dashboard/src/__tests__/plugin-routes.routes.test.ts): verify plugin visibility via GET /api/plugins, GET /api/plugins/ui-slots, and GET /api/plugins/runtimes using standard loader/store aggregation (no plugin-specific route branches).
• Dashboard slot consumers (packages/dashboard/app/components/__tests__/PluginSlot.test.tsx, packages/dashboard/app/hooks/__tests__/usePluginUiSlots.test.ts): cover slot filtering, ordering, and rendering behavior for host slot IDs used by your plugin.

Keep this layer focused on discovery/load/registration plumbing. Deeper feature-flow regressions (Settings UX, onboarding UX, runtime execution/provider behavior) belong in dedicated follow-up suites, not in these plumbing tests.

Runtime/Provider Migration Regression Placement

For runtime-provider migrations (like Droid), use layered regression suites instead of duplicating the same matrix everywhere:

• Engine runtime execution + fallback: packages/engine/src/__tests__/droid-runtime-e2e.test.ts (patterned after Hermes/OpenClaw/Paperclip E2E suites) verifies plugin runtime resolution + default pi fallback when missing.
• Runtime hint matrix guardrail: packages/engine/src/__tests__/runtime-selection-regression.test.ts keeps a lightweight hint-to-runtime routing assertion.
• Dashboard provider/auth routes: packages/dashboard/src/__tests__/routes-auth.test.ts covers POST /api/auth/droid-cli, GET /api/providers/droid-cli/status, and /api/auth/status readiness/authenticated surfacing.
• Dashboard model filtering + settings hook: packages/dashboard/src/__tests__/register-model-routes-droid-cli.test.ts and packages/dashboard/src/__tests__/register-settings-droid-cli.test.ts guard useDroidCli routing/filter behavior.
• Compatibility shim boundaries: if packages/droid-cli remains, keep tests there focused on delegation to plugin-owned implementations (not a second behavior matrix).

This keeps regressions durable while preserving clear ownership boundaries across engine, dashboard, plugin, and shim layers.

---

13. Publishing Plugins

For end-to-end standalone packaging, pnpm pack, and installing on another machine, follow the External Plugin Authoring guide. Run fn plugin publish --dry-run . before packing to validate the manifest, compiled entrypoint, lifecycle hook shape, and optional version bump without installing, uploading, or tagging anything.

Package Requirements

{
  "name": "fusion-plugin-my-plugin",
  "version": "1.0.0",
  "keywords": ["fusion-plugin"],
  "exports": {
    ".": {
      "types": "./dist/index.d.ts",
      "import": "./dist/index.js"
    }
  },
  "dependencies": {
    "@runfusion/fusion": "^0.48.0"
  }
}

Publishing Steps

1. Update package.json:

   • Set name to fusion-plugin-* or @scope/fusion-plugin-*
   • Add "keywords": ["fusion-plugin"]
   • Set "private": false

2. Build the plugin:

   pnpm build
   

3. Run the non-mutating publish preflight:

   fn plugin publish --dry-run . --previous-version 0.9.0
   

4. Publish to npm:

   npm publish --access public
   

Signed plugin recommendation

For production distribution, publish signed artifacts (plugin-publisher.json + plugin-signature.json) alongside your compiled plugin output so operators can verify provenance under warn/enforce trust policy modes.

Installation

Users can install your plugin via CLI:

fn plugin install fusion-plugin-my-plugin
# or
fn plugin install @scope/fusion-plugin-my-plugin

Or by copying to the plugins directory:

cp -r fusion-plugin-my-plugin ~/.fusion/plugins/

Dashboard registry manifest

The dashboard Browse Registry surface is backed by the static manifest at packages/dashboard/src/registry-manifest.json. Each entry describes a plugin that may appear in Settings → Plugins before it is installed.

Manifest entries use this shape:

{
  "id": "fusion-plugin-my-plugin",
  "name": "My Plugin",
  "description": "What the plugin adds to Fusion.",
  "category": "Runtime",
  "version": "0.1.0",
  "author": "Fusion Team",
  "homepage": "https://example.com/plugin",
  "path": "plugins/fusion-plugin-my-plugin/dist/index.js"
}

Required fields are id, name, description, and category. Optional metadata (version, author, homepage) is displayed when present. path is optional: entries with a path are installable from the registry; entries without one are discovery-only and render as Coming Soon (canInstall: false) until a loadable plugin entry path is available.

When adding a bundled plugin, mirror the existing bundled-plugin registration surfaces and then add a registry manifest entry whose path points at a concrete loadable file, not a package directory. Use stable ids because installed-state annotation matches registry entries to installed plugins by id.

---

14. Example Plugins

Explore these reference implementations:

Notification Plugin

Sends webhook notifications on task lifecycle events (Slack, Discord, generic HTTP).

• Demonstrates: onLoad, onTaskCompleted, onTaskMoved, onError hooks
• Features: Settings schema, webhook formatting, event filtering

Auto-Label Plugin

Automatically labels tasks based on description content using keyword matching.

• Demonstrates: onTaskCreated hook, AI agent tools
• Features: Text classification, event emission, tool registration

CI Status Plugin

Polls CI status for branches and provides custom API endpoints.

• Demonstrates: Custom routes, periodic background work, route handlers, UI slot registration
• Features: onLoad/onUnload lifecycle, setInterval polling, REST API, UI slots for task cards and task detail tabs

Roadmap Planner Plugin

Standalone roadmap planning plugin extracted from dashboard host code.

• Demonstrates: the bundled legacy hooks.onSchemaInit plus its host-owned PostgreSQL schema during the cutover; new external plugins use hooks.onPostgresSchemaInit
• Demonstrates: plugin-scoped route namespace under /api/plugins/fusion-plugin-roadmap/*
• Demonstrates: top-level navigation registration through dashboardViews (viewId: "roadmaps") and host static view registration
• Demonstrates: AI suggestion flows that consume ctx.createAiSession through plugin route handlers

Linear Import Plugin

Bundled integration plugin that imports Linear issues into Fusion tasks without adding host-owned Linear routes or settings.

• Demonstrates: password plugin settings, plugin-scoped HTTP routes under /api/plugins/fusion-plugin-linear-import/*, agent tools, top-level dashboardViews, host static dashboard view registration, and task creation through PluginContext.taskStore.
• Demonstrates: external SaaS API evidence in plugin docs, bounded GraphQL pagination, duplicate detection with durable source provenance, and safe error responses that do not leak API keys.

Droid Runtime Plugin

Reference runtime plugin that migrates a CLI-backed provider into the plugin system.

• Demonstrates: runtime adapter pattern (runtime-adapter.ts) and plugin-owned streaming/provider orchestration (provider.ts, process-manager.ts)
• Demonstrates structured contribution registration for settings-provider-card, settings-config-section, onboarding-provider-card, onboarding-setup-help, onboarding-provider-recommendation, and post-onboarding-recommendation
• Demonstrates dashboard probe delegation through plugin-owned probeDroidBinary
• Preserves provider id droid-cli via @fusion/droid-cli compatibility shim

Settings Demo Plugin

Example plugin demonstrating settings schema and runtime configuration with all four setting types.

• Demonstrates: Settings schema (string, number, boolean, enum), hooks that read settings, tools with settings-driven output
• Features: Configurable greeting message, tag limit, logging toggle, log level selector
• Install from Settings: Designed to be installed via the dashboard Settings → Plugins UI

Even Realities Glasses Plugin

Task-focused card bridge plugin for Even Realities glasses companion flows.

• Features: quick capture text into new tasks via the plugin route
• Features: polling-based task transition notifications on configured columns (default in-review)
• Features: board/task card endpoints (GET /board/cards, GET /board, GET /tasks/:id/cards)
• Features: agent actions for start work (in-progress) and request review (in-review), gated by enableAgentActions
• Features: webhook transport bridge with companion action ingestion (POST /transport/actions) and reconnect/status endpoints
• Demonstrates: settings schema for fusionApiBaseUrl, fusionApiToken, apiKey, glassesDeviceId, companionWebhookUrl, pollingIntervalSeconds, notifyOnColumns, quickCaptureDefaultColumn, and enableAgentActions
• Demonstrates FN-3737-aligned display limits: EVEN_CARD_MAX_CHARS_PER_LINE = 28, EVEN_CARD_MAX_LINES_PER_CARD = 8, EVEN_CARD_MAX_DECK_SIZE = 12

Installing Example Plugins from Settings

All example plugins can be installed via the dashboard Settings → Plugins UI:

1. Open Fusion dashboard and navigate to Settings (gear icon in header)
2. Click Plugins in the sidebar
3. Click the Install button
4. Enter the absolute path to the plugin directory (e.g., /path/to/fusion/plugins/examples/fusion-plugin-settings-demo)
5. Click Install to register the plugin
6. Enable the plugin using the toggle switch
7. Configure settings via the settings (gear) icon
8. The plugin will reload automatically with new settings

---

Quick Reference

Minimal Plugin

import { definePlugin } from "@fusion/plugin-sdk";

export default definePlugin({
  manifest: {
    id: "my-plugin",
    name: "My Plugin",
    version: "1.0.0",
  },
  state: "installed",
  hooks: {
    onLoad: (ctx) => {
      ctx.logger.info("Hello from my plugin!");
    },
  },
});

Full Plugin Example

import { definePlugin } from "@fusion/plugin-sdk";
import type { FusionPlugin, PluginContext, PluginUiSlotDefinition } from "@fusion/plugin-sdk";

// UI slots for custom dashboard components
const uiSlots: PluginUiSlotDefinition[] = [
  {
    slotId: "task-card-badge",
    label: "CI Status",
    icon: "circle-check",
    componentPath: "./components/ci-badge.js",
  },
  {
    slotId: "task-detail-tab",
    label: "CI History",
    icon: "history",
    componentPath: "./components/ci-history-tab.js",
  },
];

export default definePlugin({
  manifest: {
    id: "my-full-plugin",
    name: "My Full Plugin",
    version: "1.0.0",
    description: "A complete example with hooks, tools, routes, UI slots, and runtimes",
    settingsSchema: {
      apiKey: {
        type: "string",
        label: "API Key",
        required: true,
      },
    },
  },
  state: "installed",
  tools: [
    {
      name: "my_tool",
      description: "Does something useful",
      parameters: {
        type: "object",
        properties: {
          input: { type: "string" },
        },
        required: ["input"],
      },
      execute: async (params, ctx) => {
        const result = process(params.input as string);
        return { content: [{ type: "text", text: result }] };
      },
    },
  ],
  routes: [
    {
      method: "GET",
      path: "/status",
      handler: async () => ({ status: "ok" }),
    },
  ],
  uiSlots,
  hooks: {
    onLoad: (ctx) => ctx.logger.info("Loaded!"),
    onTaskCreated: (task, ctx) => {
      ctx.logger.info(`Task created: ${task.id}`);
    },
    onUnload: (ctx) => {
      // Cleanup with the same context shape passed to onLoad
      ctx.logger.info("Shutting down plugin");
    },
  },
} satisfies FusionPlugin);

---

For more information, see the Plugin SDK Reference.

---

15. Registering Skills

Plugins can contribute reusable skills that are surfaced in agent sessions through Fusion's skill-selection flow.

import type { PluginSkillContribution } from "@fusion/plugin-sdk";

const skills: PluginSkillContribution[] = [
  {
    skillId: "web-research",
    name: "Web Research",
    description: "Finds and summarizes web sources for a task",
    skillFiles: ["skills/research/web-research/SKILL.md"],
    enabled: true,
    triggerPatterns: ["research", "search the web", "find sources"],
  },
];

skillFiles are relative to the plugin root. The first entry, skillFiles[0], is the authoritative body file that Fusion resolves for the skill, so plugins can organize skill bodies in category subdirectories such as skills/research/web-research/SKILL.md while keeping a short skillId. When skillFiles is omitted or empty, Fusion falls back to the compatibility path skills/<name>/SKILL.md. skillId must be kebab-case.

Plugin skills are discovered per requesting project: the Skills view and workflow editor surface plugin:<id> skills only when that plugin is enabled for that project's plugin state, even if the daemon was started from a different directory.

Enabled plugin skills are delivered to agent sessions from their plugin-package SKILL.md files. Fusion resolves the first skillFiles entry (or the compatibility fallback) through the plugin root, adds the skill body directory to the session's skill discovery paths, and keeps the requested skill name in the same selection filter used for native and installed skills. The Skills view also reads the resolved SKILL.md plus sibling reference files from disk, so users can inspect the exact guidance agents receive.

16. Registering Workflow Steps

Plugins can ship workflow step templates that users can enable like built-in quality gates.

import type { PluginWorkflowStepContribution } from "@fusion/plugin-sdk";

const workflowSteps: PluginWorkflowStepContribution[] = [
  {
    stepId: "strict-review",
    name: "Strict Review",
    description: "Run an AI review with strict failure criteria",
    mode: "prompt",
    phase: "pre-merge",
    prompt: "Review this task for correctness, regressions, and missing tests.",
    toolMode: "readonly",
    defaultOn: true,
  },
  {
    stepId: "smoke-build",
    name: "Smoke Build",
    description: "Build package before merge",
    mode: "script",
    scriptName: "build",
    toolMode: "coding",
  },
];

Use mode: "prompt" | "script" and toolMode: "readonly" | "coding".

Plugin-contributed workflow steps are materialized through core resolvePluginWorkflowStep(...); mode, phase, scriptName, toolMode, defaultOn, modelProvider, and modelId are preserved from your contribution (with defaults when omitted).

16.5. Contributing Column Traits

Requires the experimentalFeatures.workflowColumns flag. Traits are the composable building blocks of workflow-defined columns (declarative flags + lifecycle hooks). Plugins contribute traits the same way they contribute workflow steps: declare them on the plugin object and the engine aggregates, caches, and invalidates them through the PluginRunner (mirroring workflowSteps).

A plugin trait is registered into the core trait registry under a plugin-namespaced id plugin:<pluginId>:<traitId>, so it can never collide with a built-in trait or another plugin's trait, and it resolves through the same registry lookup as the 14 built-in traits.

import type { PluginTraitContribution } from "@fusion/plugin-sdk";

const traits: PluginTraitContribution[] = [
  {
    traitId: "security-approval",
    name: "Security Approval Gate",
    description: "Holds a card until a security review prompt passes.",
    schemaVersion: 1,
    flags: { gate: true },
    hooks: {
      gate: {
        mode: "prompt",
        prompt: "Approve this change for security-sensitive paths?",
        gateMode: "blocking",
      },
    },
  },
  {
    traitId: "slack-notify",
    name: "Slack Notify",
    description: "Posts to Slack when a card enters/leaves the column.",
    schemaVersion: 1,
    flags: { notify: true },
    hooks: {
      onEnter: { mode: "script", scriptName: "slack-notify-enter" },
      onExit: { mode: "script", scriptName: "slack-notify-exit" },
    },
  },
];

export default definePlugin({
  manifest: { id: "my-plugin", name: "My Plugin", version: "1.0.0" },
  state: "installed",
  hooks: {},
  traits,
});

Contribution shape

Field	Required	Notes
traitId	yes	kebab-case slug, unique within the plugin
name	yes	display name
description	no	UI description
schemaVersion	yes	must be 1 — the versioned hook-descriptor contract (see below)
flags	no	declarative flags (restricted flags rejected, see below)
configSchema	no	declarative config fields ({ fields: [...] })
hooks	no	async hook descriptors (see below)

Hook points (async only)

Plugin traits get async hook points only:

• gate — evaluated before a card moves into the column (pre-move, outside the task lock). The verdict is recorded and re-checked cheaply when the move commits.
• onEnter / onExit — post-commit, async, idempotent effects.
• releaseCondition — evaluated by the hold/release sweep for hold columns.

The synchronous guard hook point is built-in-only and is rejected at validation. Sync guards run inside the task lock and must be fast and pure — a plugin hook there could wedge the lock, so plugins use the async gate surface instead.

Each hook descriptor mirrors the workflow-step shape:

{ mode: "prompt" | "script", prompt?: string, scriptName?: string, gateMode?: "blocking" | "advisory" }

Hooks execute through the same prompt-session / script / verdict machinery contributed workflow steps use — plugin trait code never runs raw in-process.

Gate semantics

• gateMode: "blocking" (default for gates) fails closed: a non-pass verdict — or no recorded verdict at move time — rejects the move with a typed TransitionRejection.
• gateMode: "advisory" records and allows: the verdict is logged but the move proceeds.
• Engine-sourced and recovery moves bypass gates entirely (they carry bypassGuards), so self-healing is never blocked by a plugin gate.

Restricted flags

A plugin trait may not declare these flags (rejected at validation, and as a backstop at registry registration):

• complete — a terminal-success column that silently satisfies dependencies.
• archived — globally hidden column semantics.

A plugin needing those semantics composes its trait alongside the built-in complete / archived trait on the same column.

Versioned hook-descriptor schema

schemaVersion: 1 is required. It pins the hook-descriptor contract so the built-in trait vocabulary can grow additively (new flags, hook points, config fields) without breaking already-published plugin traits. Validate your contribution with validatePluginTraitContribution(...) from @fusion/plugin-sdk.

Disabling a plugin with live dependents

If a card is currently sitting in a column that uses one of your plugin's traits, disabling/uninstalling the plugin is blocked with a typed error listing the dependent tasks (mirroring the built-in-workflow deletion block).

A force path degrades the affected columns to passive: the trait's hooks become no-ops (the registry resolves them to a no-op plus an audit warning), a single audit event is emitted, and the cards remain fully movable. A degraded gate column never blocks a card.

16.6. Contributing Workflow Extensions

Workflow extensions let plugins participate in engine and workflow decisions without replacing the base engine. Each extension is registered under plugin:<pluginId>:<extensionId> and may be referenced from workflow IR extensions metadata on columns or nodes.

Use WORKFLOW_EXTENSION_SCHEMA_VERSION from @fusion/plugin-sdk and declare extensions on the plugin object:

import {
  WORKFLOW_EXTENSION_SCHEMA_VERSION,
  type WorkflowExtensionContribution,
} from "@fusion/plugin-sdk";

const workflowExtensions: WorkflowExtensionContribution[] = [
  {
    extensionId: "security-move-policy",
    name: "Security Move Policy",
    kind: "move-policy",
    schemaVersion: WORKFLOW_EXTENSION_SCHEMA_VERSION,
    fallback: "failClosed",
    evaluate: async ({ task, fromColumn, toColumn, actor }) => {
      if (toColumn === "done" && actor?.kind !== "human") {
        return {
          allowed: false,
          reason: "human approval required",
          message: `Cannot move ${task.id} to done without human approval`,
        };
      }
      return { allowed: true };
    },
  },
];

export default definePlugin({
  manifest: { id: "my-plugin", name: "My Plugin", version: "1.0.0" },
  state: "installed",
  workflowExtensions,
});

Supported kinds:

Kind	Purpose	Binding
column-metadata	Typed metadata schema for workflow columns	Column extensions metadata
move-policy	Async pre-move policy that can allow or reject a valid transition	Global evaluator
work-engine	Claims execution for a column before the built-in executor starts	Column extensions metadata
node-handler	Handles an extension-marked workflow node before the built-in handler	Node extensions metadata
verdict-provider	Adds async task-completion verdicts before fn_task_done can finish	Global evaluator
merge-fact-provider	Adds route/fact inputs to auto-merge request initialization	Global evaluator

Fallback behavior controls what happens when an extension handler fails:

• degradeToDefault — continue through the built-in behavior.
• parkNeedsAttention — block or park the action with a retryable/manual signal.
• failClosed — block the action with a fail-closed diagnostic.

Workflow IR extension metadata is preserved only on v2 workflows. Keys must use the plugin:<pluginId>:<extensionId> form, and values must be JSON objects. If the extension is registered and declares a configSchema, the engine validates the metadata fields during IR validation.

16.7. Plugin-Gated Built-in Workflows

Plugin-gated built-in workflows are read-only workflow definitions that ship in Fusion's BUILTIN_WORKFLOWS array but stay hidden until their associated plugin is installed. For example, builtin:compound-engineering is bundled with core so it can be selected like any other built-in workflow, but it references Compound Engineering plugin skills that do not exist unless the plugin is active.

Runtime gating is declared in packages/core/src/builtin-workflows.ts. The PLUGIN_GATED_BUILTIN_WORKFLOWS map links each gated built-in workflow id to the plugin id that unlocks it:

// In packages/core/src/builtin-workflows.ts
const PLUGIN_GATED_BUILTIN_WORKFLOWS: ReadonlyMap<string, string> = new Map([
  ["builtin:compound-engineering", "fusion-plugin-compound-engineering"],
]);

When the dashboard or API lists workflow definitions, listWorkflowDefinitions() checks each built-in with isBuiltinWorkflowPluginGated() and then calls isPluginInstalled() for the required plugin before including it. Direct lookup uses the same rule: getWorkflowDefinition() returns undefined for a gated built-in until getRequiredPluginIdForBuiltinWorkflow() resolves to an installed plugin.

The workflow-definition cache is invalidated on plugin register/unregister events, so installing or removing a plugin updates the visible workflow list immediately. Gated built-ins are also excluded from defaultEnabledBuiltinWorkflowIds(): they are never enabled by default and only surface when the gating plugin is present.

Use this pattern when a first-party built-in workflow references plugin-provided skills, tools, workflow handlers, or other runtime capabilities that would be missing without the plugin. The gate prevents users from selecting a workflow that cannot run in their project.

The current first-party example is builtin:compound-engineering, which is gated on fusion-plugin-compound-engineering. It does not appear in the workflow picker or task workflow selection until the Compound Engineering plugin is installed and enabled; if the plugin is uninstalled, the workflow is hidden again.

Plugin author note: plugin-gated built-in workflows are currently for first-party, bundled workflows because the gating map lives in @fusion/core. Plugin authors who want to add their own workflow behavior should use the workflow step contribution API in §16 or workflow extensions in §16.6 instead.

17. Contributing Prompt Modifications

Prompt contributions let a plugin inject additional instructions into specific prompt surfaces.

Supported surfaces:

• executor-system
• executor-task
• triage
• reviewer
• heartbeat

Each contribution uses the PluginPromptContribution shape:

• surface: one of the five supported surfaces
• content: prompt text to inject
• position?: "append" (default) or "prepend"
• condition?: optional host-enforced gate evaluated against this plugin's per-project effective settings

condition supports a deliberately small, injection-safe grammar:

• settings["key"] === "value"
• settings["key"] !== "value"

Single or double quotes are accepted for both the setting key and string literal, and whitespace around settings, brackets, and operators is ignored. Fusion resolves effective settings by applying each settingsSchema defaultValue first, then overlaying stored per-project plugin settings. A missing or whitespace-only condition is treated as absent and includes the contribution; malformed or unsupported conditions fail closed and exclude the contribution. Comparisons are string-strict: non-string or missing setting values are not equal to a string literal, so === is false and !== is true.

import type { PluginPromptContributions } from "@fusion/plugin-sdk";

const promptContributions: PluginPromptContributions = {
  enabledByDefault: false,
  contributions: [
    {
      surface: "executor-system",
      position: "append",
      content: "Prefer .NET minimal APIs for route examples.",
      condition: 'settings["api-style"] === "minimal-apis"',
    },
  ],
};

Use enabledByDefault: false when contributions should require explicit opt-in.

18. Plugin Binary Setup Hooks

Plugins can expose setup metadata and lifecycle hooks for optional binaries or runtimes.

import type { PluginSetupCheckResult, PluginSetupHooks, PluginSetupManifest } from "@fusion/plugin-sdk";

const setupManifest: PluginSetupManifest = {
  binaryName: "agent-browser",
  description: "Headless browser runtime for web-enabled agents",
  channel: "stable",
  defaultTimeoutMs: 120_000,
};

const setupHooks: PluginSetupHooks = {
  async checkSetup(ctx): Promise<PluginSetupCheckResult> {
    return { status: "not-installed" };
  },
  async install(ctx) {
    // Use async process execution with timeout; never use execSync.
  },
  async uninstall(ctx) {
    // Remove managed binary/runtime artifacts.
  },
};

checkSetup is required. install and uninstall are optional.

Declarative MCP servers

Plugins may declare MCP servers with mcpServers. Declarations are active only in projects where the plugin is enabled; Fusion does not install the referenced binary.

mcpServers: [{
  name: "roslyn-navigator",
  transport: "stdio",
  command: "cwm-roslyn-navigator",
  args: [],
  env: { TOKEN: { secretRef: "roslyn-token", scope: "project" } },
  enabledByDefault: true,
}]

enabledByDefault defaults to true. Plugin declarations cannot set enabled: project settings own enablement. Effective precedence is global settings, enabled plugin declarations, then project settings by name. A project definition overrides a plugin declaration and a same-named project enabled:false entry tombstones it. Use Fusion secret references for sensitive env or headers; never ship plaintext credentials. Missing commands retain normal per-server MCP spawn-failure isolation.