diff --git a/devs.md b/devs.md index a8a3eb26..22a52cf8 100644 --- a/devs.md +++ b/devs.md @@ -129,16 +129,16 @@ Changes spanning both repositories must first produce a packed Commonlib artefac ## Architecture -### Service composition and legacy modules +### Service composition and legacy Modules -The application is composed from Services, ServiceModules, serviceFeatures, and a legacy module layer: +The application is composed from Services, ServiceModules, serviceFeatures, add-ons, and a legacy Module layer: -- **Service Hub**: the long-lived registry of service contracts. A simple extension, such as a check before replication, belongs in an existing Service handler. +- **Service Hub**: the long-lived registry of service contracts. Add a simple extension, such as a pre-replication check, to the handler owned by the relevant Service. - **ServiceModule**: a host-created, long-lived stateful or resource-owning capability shared through the typed `ServiceModules` record. Current examples include storage access, file handling, and database rebuilding. -- **serviceFeature**: a typed composition function which accepts only its declared Services and ServiceModules. It registers lifecycle handlers, commands, UI bindings, or other host glue, and may return a focused view. It is not a runtime registry entry. -- **AbstractModule** and **AbstractObsidianModule**: the legacy application module layer. Existing modules are loaded by the application and bound after the Service graph has been composed; this broad core access is not the preferred dependency boundary for new orchestration. +- **serviceFeature**: a typed composition function which accepts only its declared Services and ServiceModules. It registers lifecycle handlers, commands, user-interface bindings, or other host glue, and may return a focused view. It is not a runtime registry entry. +- **AbstractModule** and **AbstractObsidianModule**: the legacy application Module layer. Existing Modules are loaded by the application and bound after the Service graph has been composed; this broad core access is not the preferred dependency boundary for new orchestration. -The normal composition order is the Service Hub, replicator-provider registration, ServiceModules, serviceFeatures, add-ons, and finally legacy module binding. A serviceFeature may therefore consume an already constructed ServiceModule. Preferring a serviceFeature for new composition is a dependency-boundary rule, not an initialisation-order rule. +The normal composition order is the Service Hub, replicator-provider registration, ServiceModules, serviceFeatures, add-ons, and finally legacy Module binding. A serviceFeature may therefore consume an already constructed ServiceModule. Preferring a serviceFeature for new composition is a dependency-boundary rule, not an initialisation-order rule. Mutable state is permitted in a serviceFeature. State alone is not a reason to create a class, a ServiceModule, or retain an AbstractModule. Prefer one private context, with module-level functions which receive that context, when identity and polymorphism are not part of the contract. Separate the state, transitions, and invariants from the surrounding function which registers lifecycle handlers and connects downstream effects. Give the stateful boundary narrow collaborators rather than `LiveSyncBaseCore`. @@ -148,15 +148,17 @@ Several narrow views over one lifetime do not require several state owners or a When a core-owned serviceFeature returns a view needed by one host-specific consumer, pass that view through host composition instead of storing it as a public `LiveSyncBaseCore` property or promoting it to a ServiceModule. The receiving host should inject the view into the narrow command or application context which uses it. -Commonlib's `targetFilter.ts` and `prepareDatabaseForUse.ts` demonstrate the intended split: focused factories or operations own their private state and behaviour, while the corresponding `use...` function composes dependencies and registers handlers. The P2P composition follows the same direction at a larger scale by separating durable policy and room-session ownership from host lifecycle and UI wiring. Existing modules do not apply this boundary consistently; improve the affected boundary when changing their behaviour rather than performing an unrelated mechanical conversion. +Commonlib's `targetFilter.ts` and `prepareDatabaseForUse.ts` demonstrate the intended split: focused factories or operations own their private state and behaviour, while the corresponding `use...` function composes dependencies and registers handlers. The P2P composition follows the same direction at a larger scale by separating durable policy and room-session ownership from host lifecycle and user-interface wiring. Existing Modules do not apply this boundary consistently; improve the affected boundary when changing their behaviour rather than performing an unrelated mechanical conversion. Use interaction-based, London School unit tests for the composition boundary. Verify collaborator calls, ordering, failure short-circuiting, and handler registration, then test the focused state owner for its transitions and invariants. If a test needs a broad core fixture, a large class mock, deep mock chains, or unrelated Services, treat that friction as a design-review signal and consider a private context with narrower functions before adding more test machinery. -Legacy modules remain grouped by directory: +See [Service feature and legacy Module boundaries](docs/design_docs/service_feature_and_legacy_module_boundaries.md) for the selection criteria, current examples, reasons to avoid new `AbstractModule` subclasses, incremental migration guidance, and test shapes. Commonlib's [service feature composition guide](https://github.com/vrtmrz/livesync-commonlib/blob/main/docs/service-feature-composition.md) defines the shared host-neutral boundary. + +Legacy Modules remain grouped by directory: - `core/` contains platform-independent core behaviour; - `coreObsidian/` contains Obsidian-specific core behaviour; -- `essential/` contains required modules; +- `essential/` contains required Modules; - `features/` contains optional features; and - `extras/` contains development and testing tools. @@ -229,20 +231,21 @@ Commonlib owns the typed English fallback for messages requested by its services ## Common Patterns -### Module Implementation (Now not recommended for new features, use services instead) +### Service feature implementation ```typescript -export class ModuleExample extends AbstractObsidianModule { - async _everyOnloadStart(): Promise { - /* ... */ - } +type ExampleHost = NecessaryServices<"appLifecycle" | "API", never>; - onBindFunction(core: LiveSyncCore, services: typeof core.services): void { - services.appLifecycle.handleOnInitialise(this._everyOnloadStart.bind(this)); - } -} +export const useExampleFeature = createServiceFeature((host: ExampleHost) => { + host.services.appLifecycle.onLoaded.addHandler(async () => { + host.services.API.addLog("Example feature loaded"); + return true; + }); +}); ``` +Existing legacy Modules continue to register their handlers in `onBindFunction()`. Follow [Service feature and legacy Module boundaries](docs/design_docs/service_feature_and_legacy_module_boundaries.md) when new behaviour touches one of those Modules. + ### Settings Management - Settings are defined by Commonlib (`ObsidianLiveSyncSettings`) diff --git a/docs/design_docs/service_feature_and_legacy_module_boundaries.md b/docs/design_docs/service_feature_and_legacy_module_boundaries.md new file mode 100644 index 00000000..3b00a5e8 --- /dev/null +++ b/docs/design_docs/service_feature_and_legacy_module_boundaries.md @@ -0,0 +1,250 @@ +--- +date: 2026-08-30 +commonlib-version: "0.1.19" +self-hosted-livesync-version: "1.0.21" +status: 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](https://github.com/vrtmrz/livesync-commonlib/blob/main/docs/service-feature-composition.md) 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; and +7. 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: + +```typescript +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`: + +```typescript +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. + +### 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: + +```typescript +interface ConflictCheckContext { + readonly checkQueue: QueueProcessor; + readonly resolveQueue: QueueProcessor; +} + +interface ConflictCheckDependencies { + readonly conflict: ConflictCapability; + readonly currentSettings: () => ConflictSettings; + readonly getActiveFilePath: () => FilePathWithPrefix | undefined; + readonly log: LogFunction; +} + +function queueConflictCheck( + context: ConflictCheckContext, + dependencies: ConflictCheckDependencies, + path: FilePathWithPrefix +): Promise { + // 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: + +```typescript +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: + +```typescript +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.