obsidian-livesync/docs/design_docs/service_feature_and_legacy_module_boundaries.md

date, commonlib-version, self-hosted-livesync-version, status

date	commonlib-version	self-hosted-livesync-version	status
2026-08-30	0.1.19	1.0.21	accepted

Service feature and legacy Module boundaries

Purpose

This document guides new Self-hosted LiveSync composition and bounded refactoring of existing application Modules. It supplements Commonlib's service feature composition guide with the risks and migration boundaries specific to AbstractModule and AbstractObsidianModule.

Existing Modules remain supported application structures. This guidance does not require mechanical conversion of working code. It defines why a new feature should normally use an existing Service handler or a serviceFeature, and when retaining a Module is still appropriate.

Default decision

For new behaviour:

1. add a handler to an existing Service when that Service already owns the result, priority, and lifecycle;
2. use a serviceFeature when the work composes several Services, ServiceModules, lifecycle events, commands, or host effects;
3. keep feature-local state in a private context, with functions which receive that context;
4. use a ServiceModule only when several consumers need the same long-lived operational capability or resource lifetime; and
5. use a focused class when stable identity, polymorphism, serialised ownership, replacement, abort(), close(), or dispose() is part of the contract.

Do not select AbstractModule or AbstractObsidianModule merely to obtain convenient access to LiveSyncBaseCore, settings, Services, or Obsidian APIs.

How the legacy Module layer works

LiveSyncBaseCore currently composes the application in this order:

1. retain the constructed Service Hub;
2. construct the ServiceModules record;
3. construct and register built-in and host-supplied Modules;
4. compose the built-in Commonlib serviceFeatures;
5. compose host-supplied serviceFeatures;
6. construct add-ons;
7. compose the late core serviceFeatures whose handlers must follow host features and add-ons; and
8. call onBindFunction() for each registered Module.

The Module constructor therefore runs before its handler bindings, while the complete Service Hub and ServiceModules already exist. bindModuleFunctions() then invokes every onBindFunction() and runs __$checkInstanceBinding(). That diagnostic compares underscore-prefixed prototype methods with method references found in the source text of onBindFunction().

This is a compatibility lifecycle. A serviceFeature does not need to wait for Module binding. It can consume the already constructed Services and ServiceModules directly.

Why new code should avoid AbstractModule

Dependencies are broader than the type signature

An AbstractModule constructor receives LiveSyncBaseCore. Through that one object, a subclass can reach:

• the complete Service Hub;
• every ServiceModule;
• the active local database;
• settings and setting persistence;
• application commands, views, ribbon icons, and protocol handlers; and
• path, readiness, logging, and test helpers.

A reader cannot determine the real dependency set from the constructor or class declaration. A serviceFeature using NecessaryServices makes that set visible and compiler-checked.

Initialisation is split across construction and binding

Module fields can dereference this.services during class field initialisation, while public behaviour is registered later in onBindFunction(). Correctness consequently depends on both the host construction order and a second binding phase.

This permits states which are difficult to express in a type:

• the class exists but its handlers are not registered;
• a field has captured a Service before the intended lifecycle point;
• a method passed as a callback has lost its receiver; or
• a test invokes onBindFunction() against a partial object which could not occur through ordinary composition.

Callback safety is checked at runtime

Legacy Modules commonly register this.method.bind(this). __$checkInstanceBinding() can report an underscore-prefixed method which is not referenced by onBindFunction(), but it does not type-check the registration or prove that a callback retains its receiver. A module-level function receiving an explicit context does not have a receiver to lose.

Registry and ordering dependencies remain implicit

Modules are stored in one runtime list. Construction order, binding order, getModule(), and subclass identity can become hidden dependencies. A serviceFeature is called at the composition root and returns only an intentionally retained view, so its consumers do not need a general Module locator.

Resource ownership is not part of the base contract

AbstractModule has no standard replacement, cancellation, or disposal contract. Individual Modules can register onUnload handlers, but accepting the core does not state which object owns a queue, remote handle, room, timer, or in-flight operation.

Use a focused owner when the resource lifetime is meaningful, then compose that owner through a serviceFeature. The owner should expose the smallest necessary abort(), close(), dispose(), or view contract.

Tests inherit unrelated application structure

Current Module tests sometimes call a prototype method with manually assembled objects:

ModuleReplicator.prototype.onBindFunction.call(module, {} as never, services as never);

Other tests construct a broad fake core so that the base class can expose one or two collaborators. These tests can verify behaviour, but the fixture cost obscures the actual interaction contract and makes unrelated Service changes more likely to affect them.

When a focused London School test requires a broad core fixture, repeated as never, deep mock chains, or manual prototype invocation, treat that friction as a design-review signal.

Why AbstractObsidianModule is a more restrictive boundary

AbstractObsidianModule adds direct access to the plug-in and app on top of the complete core. This is useful for existing Obsidian-owned integration, but it combines platform policy, application composition, and domain behaviour in one inheritance boundary.

For new behaviour, keep Obsidian-specific presentation or registration in an Obsidian-owned serviceFeature. Pass host-neutral operations or focused views into that feature. This permits the CLI, WebApp, WebPeer, and unit tests to reuse the operation without constructing an Obsidian plug-in.

Current examples

A small serviceFeature: language initialisation

src/serviceFeatures/onLayoutReady/enablei18n.ts declares only setting, API, and appLifecycle:

export const enableI18nFeature = createServiceFeature(async ({ services: { setting, API, appLifecycle } }) => {
    // Apply the language, persist a change, and register unload clean-up.
});

The local ObsidianLanguageAppliedNotice class is still appropriate. It owns one replaceable Obsidian Notice and has an explicit clear() lifetime operation. The class is not used as a service locator, and the serviceFeature owns its construction and host binding.

Operation and composition: database preparation

Commonlib's prepareDatabaseForUse() is independently callable and receives explicit collaborators. usePrepareDatabaseForUse() constructs the error manager and registers the operation with databaseEvents.initialiseDatabase.

This split allows tests to verify:

• database opening before scanning;
• short-circuiting after a failed step;
• completion handlers before pending-event commitment;
• readiness only after every required step; and
• registration of the composed operation.

The operation does not need an application Module identity.

Ordered start-up composition and registration-only features

Configured Vault admission and the checks which follow database preparation are composed by src/serviceFeatures/startupLifecycle/. The directory keeps onboarding admission, compromised-chunk inspection, incomplete-document repair, Config Doctor, and the obsolete bulk-send setting migration as separate operations. One feature composer owns their order and receives the compatibility-review wait operation explicitly; an individual operation does not call the composer.

The layout-ready admission handler uses priority 1. This preserves ordinary priority-0 host integration before admission, while keeping an unconfigured Vault outside the flag-file recovery handlers at priorities 5, 10, and 20, and the compatibility review at priority 30. Admission belongs to one plug-in process: an initially unconfigured process remains inert until setup restarts it, and declining the requested restart does not trigger an in-process reconfiguration. Changing an admitted process back to unconfigured retires its Config Doctor and incomplete-document repair request handlers. The handlers also recheck the current configured state and database readiness when invoked, so a pending restart cannot expose partially initialised or retired state. The first-initialise handler rechecks admission before retaining the established order after the file watcher has been started: database readiness, compromised chunks, incomplete documents, compatibility review, Config Doctor, and the bulk-send setting migration.

Command and ribbon registration are serviceFeatures for the same dependency-visibility reason, but they are not start-up migrations. The basic commands remain a host-neutral feature composed by LiveSyncBaseCore, while the replication ribbon remains an Obsidian-only feature composed by the Obsidian host. Both retain onInitialise registration so moving them out of the Module list does not make their effects run during construction.

Private state and ordered handlers: target filters

Commonlib's targetFilter.ts keeps each cache or readiness gate in the factory which owns one predicate. useTargetFilters() constructs those predicates and registers them in their required order.

The state remains private to the composed feature. It does not become a LiveSyncBaseCore property or a ServiceModule merely because it persists across calls.

Legacy example to improve when touched: conflict checking

ModuleConflictChecker currently combines:

• conflict policy decisions;
• two QueueProcessor owners;
• cancellation signalling;
• access to settings and active-file state; and
• registration into the conflict Service.

Its queues are class fields which dereference this.services during field initialisation, and its public handlers are bound later in onBindFunction().

A bounded change to this area should prefer a shape such as:

interface ConflictCheckContext {
    readonly checkQueue: QueueProcessor<FilePathWithPrefix, unknown>;
    readonly resolveQueue: QueueProcessor<FilePathWithPrefix, unknown>;
}

interface ConflictCheckDependencies {
    readonly conflict: ConflictCapability;
    readonly currentSettings: () => ConflictSettings;
    readonly getActiveFilePath: () => FilePathWithPrefix | undefined;
    readonly log: LogFunction;
}

function queueConflictCheck(
    context: ConflictCheckContext,
    dependencies: ConflictCheckDependencies,
    path: FilePathWithPrefix
): Promise<void> {
    // Make the decision and enqueue through explicit collaborators.
}

export function useConflictChecking(host: ConflictCheckingHost): void {
    const context = createConflictCheckContext(host);
    host.services.conflict.queueCheckFor.setHandler((path) => queueConflictCheck(context, dependencies, path));
}

The exact extraction should be made only when conflict-checking behaviour changes. The example describes the intended ownership boundary; it is not a request to convert the Module in an unrelated documentation change.

Interaction-based testing

Test a serviceFeature at two levels.

First, test the operation or state owner with narrow collaborators:

it("does not enqueue after an optional resolver completes the conflict", async () => {
    const enqueue = vi.fn();
    const resolveOptionally = vi.fn(async () => true);

    await queueConflictCheck(contextWith({ enqueue }), dependenciesWith({ resolveOptionally }), path);

    expect(resolveOptionally).toHaveBeenCalledWith(path);
    expect(enqueue).not.toHaveBeenCalled();
});

Second, test the composition:

it("registers conflict checking with the conflict Service", () => {
    const setHandler = vi.fn();

    useConflictChecking(makeHost({ setHandler }));

    expect(setHandler).toHaveBeenCalledOnce();
    expect(setHandler).toHaveBeenCalledWith(expect.any(Function));
});

The test should make the interaction contract legible: which collaborator is called, in which order, what result is returned, and what must not run after a failure.

Do not expose a private constructor, publish a broad mock, or attach a context to LiveSyncBaseCore solely to make a test possible. If the narrow test cannot be written cleanly, reconsider the responsibility split.

When retaining a Module is appropriate

Retain or extend an existing Module when the current change depends on its established:

• Module identity or getModule() lookup;
• binding order with neighbouring legacy Modules;
• Obsidian plug-in lifecycle integration;
• user interface object lifetime; or
• compatibility behaviour whose extraction would materially expand the change.

Even then, new domain operations can receive explicit dependencies instead of accepting the Module or complete core. Improve the affected ownership boundary without converting unrelated neighbours.

Migration approach

When a Module is already in scope:

1. name the behaviour being changed and the state or resource which owns it;
2. identify the smallest operation which can accept explicit dependencies;
3. add a focused regression or interaction test around that operation;
4. keep host-specific registration in the Module initially, if that is the smallest safe step;
5. move registration to a serviceFeature only when the current integration can do so without changing ordering or lifetime; and
6. remove the legacy Module only when no identity, lookup, ordering, or compatibility consumer remains.

This is an incremental boundary change, not an inheritance-removal campaign.

Review checklist

Before adding or changing application composition, confirm that:

• dependencies are visible in a function, context, or constructor type;
• mutable state has one named owner;
• shared state is not promoted to a ServiceModule without multiple consumers;
• external resources have explicit replacement and disposal semantics;
• host-specific UI remains outside host-neutral operations;
• a consumer receives a focused view rather than the complete core;
• handler ordering and failure short-circuiting are tested; and
• retaining a legacy Module is an explicit compatibility decision.