Termix/packages/plugin-sdk/ARCHITECTURE.md

Termix plugin architecture

This is the contract between Termix core and a plugin. It is the reference for anyone writing a plugin and the specification the runtime in src/backend/plugins/ implements.

Status: the redesign lands across several steps. Everything marked A1 is built today. Later markers name the step that delivers the rest.

---

The shape of it

1. One tier

Every plugin, official or community, runs in-process in the Termix server and in the browser. There is no first-party allowlist, no worker tier, and no special case keyed on a plugin id anywhere in core.

This is not a sandbox. See What this protects.

2. SDK only

A plugin imports @termix/plugin-sdk, its own files and its own npm dependencies. Entry points:

Entry	Contents
@termix/plugin-sdk/backend	PluginContext, definePlugin(), PluginCapabilityError
@termix/plugin-sdk/db	defineTable(), the column builders, the legacy-table map
@termix/plugin-sdk/ddl	The per-dialect DDL emitter, shared with the CLI
@termix/plugin-sdk/frontend	The app object types (stubs until A7)
@termix/plugin-sdk/manifest	Manifest types, validation, the JSON schema
@termix/plugin-sdk/capabilities	The capability catalog
@termix/plugin-sdk/settings	Settings field types and the shared value validator
@termix/plugin-sdk/testing	Test helpers, incl. createFakeContext()

Plugins never import from src/ or @/, and core never imports from plugins/. Anything a plugin needs from core becomes a typed, documented SDK API. Enforced by lint, with the exceptions recorded under legacy core imports.

The SDK holds the contract: types, the catalog, and pure helpers. The ctx implementation lives in core (src/backend/plugins/ctx.ts) and is injected at activation, so a plugin cannot reach around it.

3. Capabilities are what a plugin's code may do

Declared in the manifest's capabilities array using domain:action names from one catalog in the SDK. Every privileged SDK call checks the capability and writes an audit line naming the plugin and the acting user.

Two things must both be true, and a grant for something the manifest never declared is ignored rather than honoured:

1. the manifest declares it, and
2. it is granted in plugin_permission_grants.

That means widening a plugin's reach always requires a new manifest the user can see. In 2.9.0 bundled plugins are granted everything they declare automatically, with source: "bundled".

Adding a capability is one catalog entry (id, risk, i18n title and consequence text) plus the SDK method that checks it.

4. User permissions are what a user may do

These stay in core RBAC and keep their dotted spelling (hosts.view), so they can never be confused with colon-separated capabilities.

A5 delivered this. A plugin declares short names in contributes.permissions, and core registers each as <pluginId>.<name>. The plugin never writes the prefix, so it cannot claim a core group or another plugin's namespace: parseManifest refuses a name starting with hosts, snippets, credentials, admin or the plugin's own id, and the runtime re-checks against every namespace already in the catalog, because the validator cannot know which other plugins exist.

The ids are unchanged by the move. ai + use is ai.use, which is exactly what the core catalog held before, so no role had to be migrated when ai.* and automations.* left PERMISSION_CATALOG and SYSTEM_ROLE_DEFAULTS.

A permission outlives its plugin. Core records every id it has ever registered in rbac_known_permissions, loaded at boot before any request is served and backfilled from what roles already hold. Disabling a plugin now only marks its group enabled: false: the group stays in the catalog so the role editor can grey it, and PUT /rbac/roles/:id keeps accepting it. Previously the group was unregistered outright, so an admin could not save any change to a role that still held one of these while the plugin was off. A permission nobody has ever registered is still rejected.

Defaults apply once. defaultRoles names the seeded system roles (admin, user) and nothing else. Each application is recorded in rbac_applied_defaults, a core table with no foreign key to plugins, so an admin who revokes a default does not get it handed back on the next boot or by reinstalling the plugin. That ledger used to live in plugin_storage, which cascades with the plugin row.

Both tables are deliberately unlinked from plugins: the whole point is surviving the plugin being gone.

ctx.rbac is the plugin-side view: has(permission) for the current actor, hasFor(userId, permission), and require(permission) as route middleware. All three resolve a bare short name against this plugin's prefix and take a full id belonging to a core group or another plugin as given, which is what makes a cross-plugin check expressible. require additionally refuses anything the plugin does not declare itself, at registration time, and denies with the same 401/403 bodies requirePermission() sends.

5. Data

Plugins own their tables. A plugin declares them with defineTable() from @termix/plugin-sdk/db and ships migrations for sqlite, postgres and mysql under migrations/<dialect>/NNNN_name.sql. Table names are prefixed p_<id with - as _>_, which defineTable adds. ctx.kv is for small key/value state. Core schema.ts ends up holding only core tables.

One table object, three DDL emitters. The sqlite-core definitions already encode correctly on every engine at query time, which is why core's own 44 repositories use schema.ts on all three. Only DDL genuinely differs, and that mapping lives in @termix/plugin-sdk/ddl so termix-plugin migrations and the server's runner cannot drift into disagreeing.

MySQL cannot index TEXT. Any column carrying a key becomes varchar(255), and defineTable refuses an index on text() or json() outright rather than silently truncating it.

Migrations are applied at activation, before activate() runs, so a plugin's first query never meets a missing table. Applied migrations are recorded in plugin_migrations with a checksum, which makes them immutable: editing one that already ran blocks that plugin rather than leaving two installs with quietly different schemas. A migration that throws marks its plugin failed and core keeps booting.

Adoption is how a feature keeps its rows when it moves out of core. The legacy table is renamed into the plugin's namespace (fleets -> p_fleets_fleets) rather than copied, so nothing is duplicated and no row is migrated one at a time. LEGACY_TABLE_OWNERS maps each legacy table to the one plugin allowed to adopt it, enforced by the CLI at build time and by the runner at activation, because a plugin installed from a tarball never ran the CLI.

Anything DDL-only in db/index.ts is SQLite-only. alert_rules, alert_rule_channels and alert_firings were created only there and never declared in schema.ts, so they never existed on Postgres or MySQL at all. Automations replaced them and A3 drops them. A core table needs all three artifacts in lockstep: the schema.ts declaration, the index.ts DDL, and the regenerated schema.pg.ts/schema.mysql.ts plus drizzle migrations.

Sync entities are registered, not hardcoded. ctx.sync.registerEntity adds an entity to remote sync; core registers its own ten the same way in database/routes/sync-entities.ts. The wire names are unchanged, because existing tombstones and rows on the far side match on those strings. The Electron client asks the server for the ordered list and falls back to its frozen array for a server that predates the endpoint.

6. HTTP and WebSockets

Plugin routes live under /plugin-api/<id>/ and sockets under /plugin-ws/<id>/<path>, served by the main backend with core auth. No plugin gets its own port or nginx block.

A4 delivered this. A plugin calls ctx.http.router(options) and gets an Express Router core mounts at /plugin-api/<id>/, or ctx.ws.route(path, handler) for a socket. Core runs the same middleware in front of every plugin:

1. Auth, unless the path is listed in options.public. A public path is audited when the router is registered and logged on every request, because "this plugin opened a hole in auth" should be findable later. It matches the full path, with :param segments allowed, so declaring /webhook/:token public cannot open /webhook/:token/anything.
2. The actor, from the user core authenticated, never from the request body.
3. An enabled check returning 503 while the plugin is disabled. That is a different fact from 404 ("no such plugin") and a caller can act on it.
4. Body limits, defaulting to the 2mb core itself accepts. database.ts skips its own global parser for /plugin-api so a plugin's bodyLimit actually applies; rawBody turns the parsers off for a router that parses its own (multer, a raw body for signature checks).
5. An error wrapper that logs against the plugin, never leaks a stack, and feeds the error budget.

ctx.rbac.require(permission) adds a per-route RBAC gate, and a plugin may only require a permission it declares itself.

Both surfaces need network:serve. The declaration is checked when the router or route is created; the grant is checked per request and per upgrade, because it lives in the database and router() has to be synchronous. A revoked grant therefore takes effect on the next request without a restart: HTTP answers 403, an upgrade is refused 403.

Sockets are served from the main server's upgrade event, on both the HTTP server and the direct HTTPS one. Auth is the same utils/ws-auth.ts path every other Termix socket uses, and a token still awaiting TOTP is refused exactly as it is for HTTP. Binary frames and backpressure are untouched. Everything a plugin registered is disposed on deactivate, and live sockets are closed, not merely unrouted: a terminal session must not outlive the plugin that owns it.

ctx.ws.upgrade(path, handler) hands over the raw upgrade, after core has authenticated it, for a library that insists on owning its own WebSocketServer. guacamole-lite is why it exists. Three routes are public because they authenticate themselves in a way core cannot: the terminal (a share-link guest arrives with a share token), the Docker console (per-message auth), and the Guacamole display (a single-use encrypted connection token in the query, minted by an authenticated route).

7. Settings

Plugins declare settings fields in the manifest with scope admin, user or host. Core renders them from the schema, stores values in plugin_settings, and encrypts secret fields.

A6 delivered this. A field is { key, type, labelKey, ... }, with types boolean, string, number, select, multiselect, secret, textarea, json and custom. requires names a boolean in the same scope that gates it, group is a section heading, and permission is who may write it: admin fields fall back to admin.plugins.manage, and a short name resolves against the plugin's own permissions exactly as ctx.rbac does.

type: "custom" names a component the frontend registered, for the few things a schema cannot express (a device browser, a provider list). It is deliberately narrow: a plugin supplies data everywhere else, so it cannot ship its own form styling and drift from the rest of the app, and a page disappears cleanly when its plugin does.

One service, two callers. src/backend/plugins/settings.ts is behind both ctx.settings and the HTTP routes, so validation, encryption and redaction are decided once. Two implementations would drift, and the one that drifted would be the one writing to the database.

The manifest is the schema. A read is driven by the declared fields, not by the stored rows: an unwritten field comes back with its default, and a row for a field the plugin has since dropped is ignored rather than handed to the UI. A write to a key the manifest never declared is refused, which is what stops a PUT from filling a plugin's namespace with arbitrary keys.

Secrets never leave the server. A secret field is encrypted with encryptSystemSecret (the installation key, not a user DEK: these have to be readable before any user is unlocked) and every HTTP read replaces it with { set: boolean }. Sending that marker back is a no-op, so saving a form cannot clear a key it was never shown. Clearing one writes null rather than an encrypted empty string, so it reads back as unset.

Storage. plugin_settings(plugin_id, scope, scope_id, key, value, encrypted, updated_at), unique on the first four. scope_id is null for admin, a user id for user and a host id as text for host. It is polymorphic, so it carries no foreign key and the engine cannot cascade it: UserRepository and HostRepository delete these rows explicitly, with tests that fail if those calls ever go. The foreign key to plugins is real, so uninstalling leaves nothing.

Routes, all with core auth: GET|PUT /plugins/:id/settings/admin (admin.plugins.manage, or the field's own permission), /settings/user (always the caller, never a user id from the body) and /settings/host/:hostId (needs edit access to that host). A rejected PUT returns 400 { errors: { <key>: message } } per field, so a form shows every problem at once.

Host payloads carry them. Host list and get responses include a pluginSettings map, enabled plugins only and secrets redacted, built in one query for the whole list rather than one request per plugin per host. It is attached after the sanitizers, because the connect-level projection of a shared host reduces it to an allowlist.

ctx.settings is get/set, getUser/setUser, getHost/setHost, getAll, onChange (disposed with the plugin) and readCore. Reading and writing a plugin's own settings needs no capability: the manifest already declares every field, and a plugin that had to ask permission to read its own configuration would be useless. Only readCore, which reaches outside the plugin's namespace, is gated on settings:read-core, and it serves a short allowlist (CORE_SETTINGS_ALLOWLIST) rather than the whole settings table.

8. UI

A plugin frontend is a built ESM bundle loaded at runtime by the core loader. It registers everything through the app object: rail items, panels, tabs, host editor sections, host actions and badges, context menu items, palette entries, dashboard cards, homepage widgets, settings components, action slots, login/2FA UI, SSH auth editors. The shell contains no plugin ids. A7.

9. Cross-plugin use

Through ctx.services (typed, versioned, RBAC-checked per call), ctx.events, and frontend actions and slots. Dependencies are declared in the manifest: dependencies are hard, optionalDependencies are soft. The loader activates in dependency order. A plugin must keep working when an optional dependency is missing.

10. Lifecycle

activate(ctx) creates all state. Everything registered through ctx or app is disposed automatically on deactivate. See Lifecycle rules.

11. Tests

Each plugin has plugins/<id>/tests/backend (node) and plugins/<id>/tests/frontend (jsdom), run by its own vitest config:

// plugins/<id>/vitest.config.ts
import { pluginVitestConfig } from "@termix/plugin-sdk/vitest-preset";

export default pluginVitestConfig(import.meta.url);

The preset supplies both projects, the shared setup file (@termix/plugin-sdk/testing/setup, which core's vitest.setup.ts re-exports so there is one copy) and the @/ alias the frontends still use.

npm run test runs core only, npm run test:plugins runs every plugin and npm run test:all runs both. Core tests (src/backend/tests, src/ui/tests) never test plugin code; they may test the plugin runtime with fixture plugins, which is why src/backend/tests/plugins/ still holds the loader, manifest, capability, registry, service, secret and shutdown suites.

@termix/plugin-sdk/testing provides createMockCtx(), which enforces capabilities the way the runtime does, createFakeContext() for tests that do not care about the gates, and renderWithApp() for frontend tests (A7 finishes it).

12. Auth

Core keeps password login, sessions, API keys, trusted proxy login, Electron auto-session and trusted devices, plus the base SSH auth types password, key, stored credential, agent and none. Every other login method, second factor and SSH auth method is a plugin through ctx.auth. A8.

13. Where plugins live

Official plugins live in plugins/<id>/ as npm workspace packages in this monorepo for all of 2.9.0, and get split into their own repos later. Every plugin is enabled by default in 2.9.0; there is no install UI until 3.0.0.

The layout is fixed, and the directory name must equal the manifest id:

plugins/<id>/
  manifest.json            id, capabilities, contributions
  package.json             @termix-plugin/<id>, private, scripts call the CLI
  tsconfig.json            extends @termix/plugin-sdk/tsconfig.plugin.json
  vitest.config.ts         one line, the SDK preset
  src/backend/index.ts     exports activate(ctx) and deactivate()
  src/frontend/index.tsx   exports activate(app) and deactivate() (A7)
  locales/en.json          English strings; Crowdin translates the rest
  migrations/{sqlite,postgres,mysql}/  owned tables, one .sql per dialect
  migrations/snapshot.json        diff basis for the generator
  src/backend/tables.ts           defineTable() definitions
  tests/backend/           vitest, node
  tests/frontend/          vitest, jsdom
  README.md  CHANGELOG.md
  dist/                    build output, gitignored

14. Upgrades must be lossless

Every move of data, settings or host columns into a plugin comes with a migration for all three dialects and a test proving existing data survives.

Moving a host column into a plugin

A6 built the host-scope settings path but moved no columns; each Phase B step moves its own. ssh_data still holds enable_docker, stats_config, web_ui_config, enable_proxmox, enable_rdp, mac_address and the rest, and the hardcoded host editor tabs still read them. The order matters, because dropping a column before its data has moved loses it:

1. Declare the fields in the plugin's contributes.settings.host, keeping the key names the form already uses so the editor tab can be deleted rather than rewritten.
2. Write a boot migration that copies each row's column into host-scope plugin_settings for that plugin, keyed by the host id as text. Make it idempotent by checking for an existing row first, the way tailscale-settings-migration.ts does, and cover it with a test that runs it twice.
3. Switch the plugin's backend to ctx.settings.getHost and delete the hardcoded editor tab; the schema-driven Plugins group replaces it.
4. Only then drop the column, in lockstep across all four artifacts: the schema.ts declaration, the db/index.ts bootstrap DDL, a drizzle migration per dialect, and npm run schema:generate for the pg and mysql variants. The two bootstrap guard tests fail if these disagree.

A column that is still read anywhere in core is not ready to be dropped. Ship the copy and the switch first, and drop it in a later step if that is safer.

---

Manifest v2 reference

plugins/<id>/manifest.json. The directory name must equal id. Unknown fields are rejected at every level: a typo'd contribute used to validate clean and be silently dropped, which made a manifest describe something it did not do.

{
  "id": "example", // ^[a-z][a-z0-9-]{1,39}$, equals the directory name
  "name": "Example",
  "version": "1.0.0", // semver
  "description": "What it does.",
  "author": { "name": "You", "url": "https://..." },
  "license": "MIT",
  "repository": "https://...", // optional
  "category": "Productivity", // see the enum below
  "icon": "Sparkles", // optional: Lucide name or a bundled icon.svg

  "engine": {
    "termix": ">=2.9.0", // display only, never enforced
    "api": "1", // the real compatibility gate
  },

  "capabilities": ["kv:own", "ui:surface"],

  "dependencies": { "other-plugin": "^1.0.0" }, // missing one blocks activation
  "optionalDependencies": { "maybe": "^1.0.0" }, // missing one is normal

  "provides": [
    {
      "service": "example.thing",
      "version": "1.0.0",
      "permission": "example.use",
    },
  ],
  "requires": [
    { "service": "other.thing", "versionRange": "^1.0.0", "optional": true },
  ],
  "providesSecret": [
    { "key": "api-key", "permission": "example.secrets.share" },
  ],
  "requiresSecret": [
    { "plugin": "other-plugin", "key": "api-key", "optional": true },
  ],

  "contributes": {
    "tabs": [
      {
        "id": "example",
        "titleKey": "nav.example",
        "icon": "Sparkles",
        "openFrom": ["rail"],
      },
    ],
    "actions": [
      {
        "id": "example.open",
        "titleKey": "k",
        "handler": "open",
        "permission": "example.use",
      },
    ],
    "actionSlots": [{ "id": "example.toolbar", "accepts": ["button"] }],
    "permissions": [
      {
        "name": "use", // registered as "example.use"
        "titleKey": "permissions.use.title", // plugin-relative
        "descriptionKey": "permissions.use.description",
        "defaultRoles": ["user"], // system roles only, applied once
      },
    ],
    "settings": {
      "admin": [
        {
          "key": "apiKey",
          "type": "secret", // encrypted at rest, never sent to a browser
          "labelKey": "settings.apiKey.label",
          "descriptionKey": "settings.apiKey.description",
          "placeholderKey": "settings.apiKey.placeholder",
          "group": "settings.group.api", // section heading
        },
        {
          "key": "retries",
          "type": "number",
          "labelKey": "k",
          "min": 0,
          "max": 5,
          "default": 3,
        },
        {
          "key": "browser",
          "type": "custom", // a registered component, for what a schema cannot express
          "component": "devices",
        },
      ],
      "user": [
        {
          "key": "advanced",
          "type": "boolean",
          "labelKey": "k",
          "permission": "tweak", // one of this plugin's own permissions
        },
      ],
      "host": {
        "enableKey": "enableExample", // boolean rendered first, gating the rest
        "enableLabelKey": "k",
        "fields": [
          {
            "key": "port",
            "type": "number",
            "labelKey": "k",
            "requires": "enableExample",
          },
        ],
      },
    },

    "hostCapability": {
      "key": "enableExample",
      "labelKey": "k",
      "editorTab": "general",
    },
  },

  "backend": "dist/backend.js", // default
  "frontend": "dist/frontend.js", // default
  "locales": "locales", // default
  "platforms": ["linux", "win32", "darwin"], // optional
}

category is one of: Terminal, Files & Transfer, Infrastructure, Monitoring, Networking, Access & Security, Productivity.

openFrom is one of: rail, host-context-menu, palette.

Field types are boolean, string, number, select, multiselect, secret, textarea, json and custom. labelKey is required for every type but custom, which needs component instead; options is required and non-empty for select and multiselect and rejected elsewhere; min/max belong to number alone; and requires must name a boolean field in the same scope, which includes the host section's enableKey.

Cross-field rules the JSON schema cannot express, checked by parseManifest:

• engine.api must match the SDK major version this build implements.
• Every provides[].permission, providesSecret[].permission and contributes.actions[].permission must appear in the plugin's own contributes.permissions, compared against the qualified <pluginId>.<name> form. A permission the catalog never sees is one no admin can grant, so the surface would be invisible rather than denied.
• requiresSecret[].plugin and both dependency maps must not name the plugin itself, and a plugin cannot be in both dependencies and optionalDependencies.
• Every contributes.permissions[].name must be unique, must not start with a core group (hosts, snippets, credentials, admin) or with the plugin's own id, and every defaultRoles entry must be admin or user.
• A contributes.settings field permission written as a bare short name must appear in the plugin's own contributes.permissions. A dotted id is taken as given, because the validator cannot know which other plugins exist.

contributes.hostCapability, provides, requires, providesSecret and requiresSecret are carried forward from v1 unchanged because the service and secret registries read them at activation. A6 and A7 reshape them.

Removed in A5: contributes.permissionGroup and its defaultForRole, replaced by contributes.permissions. A plugin no longer picks its own group name.

Removed in v2: permissions (replaced by capabilities), sidecars, the capabilities.{backend,frontend,electron,platforms} object (replaced by the backend/frontend/platforms fields), contributes.dashboardCards, contributes.settingsPanel, contributes.apiPrefix, and process:transport-owner (the marker for the tier that no longer exists).

---

Capability catalog

Colon-separated, ordered worst first. The i18n keys resolve under plugins.capabilities.<id>.title and .consequence in src/ui/locales/en.json, and the copy describes the consequence to the person deciding, not the mechanism.

Capability	Risk	Meaning
credentials:read	critical	See plaintext secrets for any reachable host
ssh:connect	high	Run commands on hosts it is connected to
process:spawn	high	Run programs on the Termix server
users:write	high	Create and change user accounts
auth:provide	high	Add a login method or second factor
system:tls	high	Request and replace the server certificate
hosts:write	medium	Create and change hosts
credentials:use	medium	Connect using a host's stored credentials, without seeing them
network:outbound	medium	Make outbound requests
network:serve	medium	Open a listening port
users:read	medium	See usernames and roles, never hashes or 2FA state
events:core	medium	Subscribe to and emit core events carrying other users' data
notify:send	medium	Send notifications through configured channels
audit:read	medium	Read the audit log
desktop:window	medium	Open desktop windows (Electron only)
hosts:read	low	See the host list it can already see
db:own	low	Own tables
kv:own	low	Small key/value state
secrets:own	low	Its own encrypted secrets
settings:read-core	low	Read core server settings
ui:surface	low	Contribute tabs, panels and settings

Rules

• One catalog, in packages/plugin-sdk/src/capabilities.ts. Nothing else defines a capability.
• Adding one is an entry here plus the SDK method that checks it. A capability nothing checks is decoration, and v1 shipped nineteen of those.
• A capability is checked where the privileged thing happens, not at activation. Activation-time checks say what a plugin might do; call-time checks say what it did.
• events:core is the exception that proves the shape: without it a plugin may only emit under plugin.<its id>., because otherwise it could publish host.status and drive the automations engine as though core had.

---

The ctx surface

Built per plugin in src/backend/plugins/ctx.ts and passed to activate.

Member	Capability	Status
ctx.pluginId, ctx.manifest	none	A1
ctx.log.*	none	A1
ctx.events.emit / .on	events:core for core topics	A1
ctx.kv.get/set/delete/list	kv:own	A1
ctx.registry.*	none	A1
ctx.services.provide / .get	per-service RBAC	A1
ctx.secrets.offer / .withdraw / .getShared	per-secret RBAC	A1
ctx.disposables.add	none	A1
ctx.asUser(userId, fn)	none, always audited	A1
ctx.currentActor()	none	A1
ctx.db.define / .client / .refs	db:own	A3
ctx.sync.registerEntity	none	A3
ctx.http.router / ctx.ws.route / .upgrade	network:serve	A4
ctx.rbac.has / .hasFor / .require	own permissions only	A5
ctx.hosts.*	hosts:read / hosts:write	B
ctx.ssh.*	ssh:connect, credentials:use	B
ctx.settings.*	settings:read-core (readCore only)	A6
ctx.notify.*	notify:send	A6
ctx.auth.*	auth:provide	A8
ctx.fetch	network:outbound	B

The actor

Never comes from plugin code. It is held in AsyncLocalStorage and set two ways, with no third:

• A request. A4's HTTP and WebSocket middleware runs the handler as the user core already authenticated.
• ctx.asUser(userId, fn) for background work with no request behind it. Always audited, and every core API inside still applies that user's RBAC.

A plugin cannot pass a user id to a guarded method and have it believed. This is what closes the forged-caller bug the old worker broker had, where a worker could set callerUserId on its own messages.

---

Lifecycle rules

1. Everything is created inside activate. No servers, timers, listeners or connections at module scope. ESM caches modules, so disable-then-enable re-runs activate against the same module object: anything created at module scope is already torn down by the second call and never comes back.
2. Everything is released on deactivate. Registrations made through ctx are tracked automatically. Anything the plugin creates itself goes in ctx.disposables.add(fn). A plugin that keeps a port after deactivate makes "disabled" a lie and makes re-enabling fail on a port it still holds.
3. Disposal does not depend on the plugin behaving. The runtime empties the disposable bag even when the plugin's deactivate() throws or does not exist, in reverse registration order, isolating each disposer.
4. A failed activate is still cleaned up, and the plugin's own deactivate() is not called: it never finished starting, so it has no state to tear down.
5. Plugins never touch process signals or process.exit. Core calls deactivate on shutdown: gracefulShutdown in src/backend/starter.ts calls shutdownPlugins(), which stops every plugin in reverse activation order. A plugin that installs its own SIGTERM handler and exits skips the rest of that sequence, including other plugins and the database flush.
6. Enable, disable, enable must work without a restart.

States

enabled, disabled, blocked (a hard dependency is missing, recoverable), failed (activation threw, a dependency cycle, or the error budget tripped). lastError carries the reason.

Error budget

Core wraps plugin callbacks it invokes. Five failures inside 60 seconds (both configurable) marks the plugin failed and deactivates it, because a plugin throwing on every call is worse than one that is off. POST /plugins/:id/retry clears the budget and starts it again.

---

What this protects, and what it does not

Be precise about this, because it is easy to overclaim and the previous architecture did.

What the capability system gives you

• Every privileged thing a plugin can do through ctx is declared in a manifest you can read before installing it.
• Every privileged call is checked against what was granted, and refused if it was not.
• Every privileged call, allowed or refused, leaves an audit line naming the plugin and the acting user. You can answer "what did this plugin do, and on whose behalf".
• A plugin cannot name a user and be believed, so it cannot act as someone it was not invoked by.
• Removing a capability from a manifest revokes the grant on the next boot.

What it does not give you

• Containment. A plugin runs in the server process. It can import("node:fs") and read whatever the server user can read, open its own sockets, spawn processes, and reach core modules directly. Guarding ctx does not change any of that.
• Protection from malicious code. A plugin that wants to bypass the gate can. Nothing here stops it.

ctx.db is a clear example. It hands back a handle scoped to the plugin's own tables, and the p_<id>_ prefix, the CLI check and the runner's check all say a plugin may only touch its own. None of that is enforced by the engine: a plugin holding that handle can write any table in the database, and ctx.db.refs hands it users and ssh_data outright. The prefix, the capability, the audit line, lint and review are the contract. The engine is not.

So the honest boundary is: capabilities stop accidental and casual overreach, and make every privileged call auditable. What stops malicious code is signing, review and the kill list - the same trust model as any other software you choose to install on a server.

The old worker tier claimed more than this, and the claim was not true either: worker_threads shares the process, so a worker plugin could already reach anything core could. The honest options for real containment are a child process with --permission, or a WASM isolate. Neither is on the roadmap, and until one is, this document does not pretend otherwise.

---

Building a plugin

termix-plugin, the CLI the SDK ships, is the only build a plugin needs. It runs from the plugin's own directory, so a plugin that moves to its own repo keeps working unchanged.

Command	What it does
termix-plugin build	esbuild bundles src/backend/index.ts to dist/backend.js (ESM, node22) and src/frontend/index.tsx to dist/frontend.js (ESM, browser), then copies locales/ and migrations/ into dist/. Imported CSS lands in dist/frontend.css.
termix-plugin validate	Runs the SDK's parseManifest, checks the files the manifest names exist, and checks that migrations only touch this plugin's tables and exist for every dialect.
termix-plugin migrations	Diffs src/backend/tables.ts against migrations/snapshot.json and writes one .sql per dialect. --check fails when a definition changed without a migration. A rename or a type change is refused rather than guessed at, because that is a data decision.
termix-plugin test	Runs the plugin's vitest suite.
termix-plugin pack	Writes a .tgz of the manifest, dist/, locales, migrations and README. D3 adds signing.

npm run build:plugins builds every plugin and stages the result in dist/plugins/<id>/, which is what getBundledPluginsDir() resolves to and what Docker and electron-builder package. npm run build, build:backend and dev:backend all call it.

TypeScript only type-checks (tsconfig.plugins.json, noEmit); esbuild does every emit. Nothing is written next to a source file.

Host-provided packages

These ship with the Termix server and are never bundled into a plugin, because a second copy of express or React is a bug, and a second ssh2 means a second set of native bindings. A plugin may depend on anything outside these lists, and that is bundled into its output.

Backend (plus every node builtin): @termix/plugin-sdk, express, ssh2, ws, multer, cookie-parser, axios, jszip, guacamole-lite, @anthropic-ai/sdk, drizzle-orm.

Frontend: react, react-dom, react/jsx-runtime, @termix/plugin-sdk, i18next, react-i18next, and the UI libraries core provides today (lucide-react, sonner, axios, cytoscape, react-cytoscapejs, guacamole-common-js, react-xtermjs, @xterm/*). A7 serves these through the SDK ui entry.

The lists live in packages/plugin-sdk/cli/lib/externals.mjs.

---

Legacy core imports: the debt D1 removes

The twelve bundled plugins predate the SDK. They still reach core by relative path (../../../../src/backend/...), which an esbuild plugin, packages/plugin-sdk/cli/lib/legacy-core-imports.mjs, keeps out of the bundle and rewrites to the compiled output path (../../../backend/backend/..., or ../../../backend/types/... for shared types). The rewrite is output-relative, so a source file at any nesting collapses to the same prefix. D1 deletes that file.

This is expected and temporary. The SDK does not yet expose what they need: the SSH connection pool, host resolution, the repository layer, the shared component library. D1 removes the debt once A3 through A8 have given them supported APIs to move onto.

What the lint fence enforces today, in eslint.config.mjs:

Direction	Severity	Offenders	Emptied by
Core importing a plugin backend	Error	0	-
A plugin backend importing frontend code or @/	Error	0	-
The shell importing plugin components	Warning	16 files	A7
A plugin importing core by relative path	Warning	70 files	D1
A plugin importing another plugin's source	Warning	3 files	B18

A warning does not fail a build, so the counts are held by scripts/check-plugin-boundaries.cjs, run by npm run lint. Every offender is listed in scripts/plugin-boundary-allowlist.json: a new one fails, and so does an entry that stopped being an offender, so the list shrinks as each step lands rather than rotting. --write regenerates it. D1 empties it and the two plugin-side warnings become errors.

The two plugin-backend directions use different rule names (no-restricted-imports and @typescript-eslint/no-restricted-imports) because flat config replaces a rule's options rather than merging them, so one rule cannot carry both severities on the same files.

Known specifics:

• TODO(B18): plugins/ai/src/backend/tools/executor.ts imports automations' routes.js directly. It is declared as a hard dependencies entry so the loader starts automations first and marks ai blocked if automations is disabled. In B18 ai calls automations through an automations service as an optionalDependency and the direct import goes away.
• TODO(B18): plugins/host-metrics/src/backend/ reaches automations twice, for a MetricsSnapshot type and a fire-and-forget import("...headless-viewer.js") that registers a viewer bridge. Both degrade silently when automations is absent, so host-metrics declares it as an optionalDependency (added in A2; it previously declared nothing, and the loader had no reason to order the two).
• Core feature servers that are not plugins yet still own ports: tunnel 30003, file-manager 30004, dashboard 30006, tmux 30010, serial 30011 and homepage 30012. Each keeps its nginx block until its own Phase B step. No plugin owns a port any more (A4).
• src/ui/shell/pluginLoader.ts still hardcodes plugin ids in BUILT_IN_TABS_BY_PLUGIN and applyFirstPartyActions. A7 removes both.