Cleanup and Consolidation of the State-Machine Layer

Status

Proposal.

This plan follows plans/state-machines-hardening.md. The hardening work made the state-machine layer substantially safer: transition results are discriminated, operation identity is explicit, lifecycle cleanup is shared, selected controllers use transactional dispatch, and transition traces are more useful. It did not attempt to make the resulting architecture small or uniform.

The next problem is comprehensibility. Several workflows now have an authoritative state-machine snapshot and a second Jotai representation of the same lifecycle. Some machines use TransactionalDispatcher; others still implement their own dispatch loop. Some cross-machine signals travel through typed facades, while others travel through atoms that act as mailboxes or observable flags. Providers and managers repeat similar ownership plumbing.

This plan removes those transitional structures without weakening the race, identity, disposal, persistence, or observability guarantees established by the hardening work.

Executive summary

The target architecture is:

producer event
    |
    v
typed machine facade
    |
    v
TransactionalDispatcher -> committed immutable snapshot -> pure selectors
    |                                                       |
    v                                                       v
command adapter                                      React domain hook
    |                                                       |
    v                                                       v
IPC / Query / UI-only runtime stores                      component

The central rule is:

A lifecycle fact represented in a machine snapshot is not also stored in Jotai.

Jotai remains appropriate for client-only state that is not owned by a machine: edit buffers, navigation preferences, high-frequency console and stream content, transient selections, and independently sourced diagnostics. React Query remains authoritative for IPC-backed entities. Main-process machines may expose renderer read models across IPC, because that is a process boundary rather than a second same-process authority.

The cleanup is incremental. Each domain migrates its consumers first, deletes its compatibility projection in the same PR, and retains focused regression tests. There is no repository-wide flag day and no period in which two independent writers are accepted as a steady state.

Problem statement

1. Same lifecycle, multiple representations

The largest example is app_run.

RunState already contains:

• the lifecycle (idle, starting, ready, reloading, stopping, stopped, or errored);
• the operation and startedAt;
• the current invocation identity;
• the current URL;
• the exit code;
• the operation error.

The renderer also stores related values in:

• previewRunStateByAppIdAtom;
• appUrlByAppIdAtom;
• previewAppExitByAppIdAtom;
• previewErrorByAppIdAtom;
• previewReloadTokenByAppIdAtom.

Some of these are exact projections, some combine independent sources, and some are imperative UI epochs. Keeping them together in previewRuntimeAtoms.ts obscures ownership and makes one output event update the machine and atoms in separate commits.

chat_stream, first_prompt, and image_generation also publish same-process machine projections into Jotai:

• isStreamingByIdAtom;
• firstPromptSagaAtom;
• imageGenerationJobsAtom and its derived atoms.

Single-writer enforcement prevents the worst races, but it does not remove the second representation, projection lifecycle, cleanup ordering, or reviewer burden.

2. Atoms used as cross-machine protocols

Three current dependencies use Jotai as an event or status bus:

• preview_iframe watches previewRunStateByAppIdAtom to infer that app_run restarted;
• plan_handoff watches isStreamingByIdAtom to infer that chat_stream became idle;
• screenshot producers write pendingScreenshotAppIdsAtom, which the screenshot provider consumes as a mailbox.

These dependencies are difficult to discover from the machine types. They also lose domain information: a boolean edge or map mutation is weaker than a typed event carrying the relevant operation identity.

3. Controller mechanics remain inconsistent

There are thirteen machine domains:

• renderer workflows: app_run, chat_stream, first_prompt, github_ops, image_generation, plan_handoff, preview_iframe, screenshot, version_preview, and voice_to_text;
• main-process workflows: connection_flow, mcp_oauth, and user_input.

Only image_generation, screenshot, and voice_to_text currently use TransactionalDispatcher. The remaining runtimes use combinations of SnapshotStore, hand-written re-entrancy queues, direct observer calls, registry maps, timers, and domain-specific command drains.

Some deviations are legitimate, especially main-process registries owning external resources. The current code does not make the distinction obvious: custom mechanics and necessary domain policy are interleaved.

4. Aggregate UI state is sometimes stored instead of selected

Examples:

• activeCheckoutCounterAtom mirrors whether any version_preview controller is mutating;
• isStreamingByIdAtom is an aggregate index over per-chat snapshots;
• image-generation job arrays are copied from a manager projection store into Jotai and then selected again.

Aggregate views are useful, but they should be read-only external-store selectors over authoritative snapshots. They should not require another general-purpose state container.

5. Domain adapters know too much about Jotai

Machine command adapters legitimately perform UI effects, but several also read lifecycle flags or write lifecycle projections. This makes a pure machine look authoritative while its effective behavior still depends on atom state.

For example, plan_handoff reads streaming status from Jotai, and app_run commands separately write URL and error atoms. These should be explicit dependencies or machine events, not implicit access to a shared store.

Goals

1. Give each lifecycle fact exactly one authoritative owner.
2. Make renderer components read machine snapshots through domain hooks and pure selectors.
3. Remove same-process machine-to-Jotai lifecycle projections.
4. Replace atom mailboxes and cross-machine flag watching with typed facades or typed events wired at composition roots.
5. Migrate custom controller transaction mechanics to TransactionalDispatcher unless a documented resource-owning registry genuinely requires a different runtime.
6. Reduce provider, manager, and selector boilerplate without introducing a framework that owns domain policy.
7. Keep high-frequency and orthogonal UI state out of machine snapshots.
8. Preserve all operation-correlation, stale-event, disposal, hydration, queue-ownership, and cross-process guarantees from the hardening work.
9. Make ownership mechanically auditable in tests and repository boundaries.

Non-goals

• Replacing the pure TypeScript machines with XState or another statechart framework.
• Moving all renderer state into machines.
• Moving console logs, partial streamed text, form buffers, modal visibility, selected files, or other high-frequency/ephemeral UI state into machine snapshots.
• Replacing React Query with machine or Jotai caches.
• Rewriting stable domain transitions merely to normalize state names.
• Combining independent machines into one application-wide machine.
• Removing cross-process renderer read models when the authority lives in the main process.
• Changing visible product behavior as part of a mechanical cleanup.

Ownership model

Every stateful value touched by a machine migration must be classified before code changes begin.

Machine-owned lifecycle state

Examples:

• current phase;
• active operation identity;
• retry/cancellation status;
• machine-owned URL or selected version;
• machine-owned error and recovery state;
• capability flags derived from the snapshot.

Rules:

• stored only in the machine snapshot;
• read through a domain hook/facade;
• derived with pure selectors;
• never mirrored into a writable atom;
• never reconstructed by watching command side effects.

External entity data

Examples:

• apps, chats, versions, settings, files, reports, persisted plans.

Rules:

• owned by React Query or the main-process persistence layer;
• invalidated or refreshed by command adapters;
• not copied into machine state unless a stable operation snapshot is required for correctness.

UI/runtime state

Examples:

• chat edit buffers;
• partial streamed message text;
• console output buffers;
• terminal panel visibility;
• dismissed banners;
• visual-editor selections;
• local dialog/form state.

Rules:

• Jotai when it must survive unmounts or be shared across distant components;
• local React state when confined to one subtree;
• keyed by entity where applicable;
• not promoted into a machine solely to reduce atom count.

Cross-process projections

Examples:

• renderer-visible pending user-input requests whose authority is the main-process user_input registry.

Rules:

• explicitly named as read models or projections;
• one adapter owns hydration, ordering, and writes;
• public APIs are read-only;
• not described as duplicate renderer machine state;
• may remain Jotai-backed when Jotai composition is materially useful.

Derived indexes

Examples:

• “any version checkout in progress”;
• “which chats are currently streaming”;
• image-generation job lists.

Rules:

• computed from authoritative keyed snapshots;
• exposed by a manager as a reference-stable external-store snapshot only when a real cross-key consumer exists;
• never independently mutated;
• prefer per-key subscriptions so unrelated entities do not rerender;
• document retention and cleanup if the index includes terminal snapshots.

Current-state audit and intended disposition

Target renderer APIs

Keyed machine hook

Each keyed renderer machine exposes one domain hook:

interface AppRunView {
  state: RunState;
  projection: AppRunProjection;
  send(event: AppRunInput): void;
}

function useAppRun(appId: number | null): AppRunView;

The projection is a reference-stable, pure function of the snapshot:

function projectAppRun(state: RunState): AppRunProjection;

It can expose convenient values such as isLoading, url, operationError, and capabilities, but it does not store them elsewhere.

Imperative facade

Non-React consumers and other machines receive a narrow facade:

interface ChatStreamStatusFacade {
  getState(chatId: number): StreamState;
  subscribe(chatId: number, listener: () => void): () => void;
}

The facade is injected at a composition root. A machine does not import another machine's manager, controller, provider, or atom.

Aggregate external-store selector

Cross-key UI uses an aggregate manager snapshot only when per-key component subscriptions are impractical:

interface ImageGenerationProjectionSource {
  getSnapshot(): readonly ImageGenerationJobView[];
  subscribe(listener: () => void): () => void;
}

The manager owns reference stability and retention. React consumes it with useSyncExternalStore; it is not copied into Jotai.

One-shot domain events

Events such as stream completion or app-run restart remain subscriptions, not state:

interface AppRunLifecycleEvent {
  appId: number;
  invocationRef: AppRunInvocationRef;
  type: "restart-started" | "stopped";
}

Callbacks run after the committed snapshot is visible. Event APIs document whether they are lossless, replayable, or live-only.

Shared infrastructure changes

1. Selector-aware external-store bindings

Add small React helpers under src/state_machines/react.ts:

• useMachineSelector(controller, selector, isEqual?);
• useKeyedMachineSelector(manager, key, selector, isEqual?);
• useProjectionSource(source).

Requirements:

• no resubscription when only the selector closure changes;
• reference-stable server snapshot behavior;
• optional equality for small scalar/object projections;
• tests for unrelated-key updates and StrictMode replay;
• no dependency on Jotai.

Prefer React's supported selector shim if already available transitively; otherwise keep the helper small and tested rather than implementing a broad state library.

2. Standard manager facade

Keep KeyedControllerHost, but standardize the common renderer manager surface:

interface KeyedMachineManager<Key, State, Input> {
  getSnapshot(key: Key): State;
  subscribeKey(key: Key, listener: () => void): () => void;
  send(key: Key, input: Input): void;
  disposeKey(key: Key): void;
  dispose(): void;
}

Do not force promise-returning dispatch, recovery indexes, or specialized registrations into this base interface. Those remain domain extensions.

3. Projection-free provider convention

Providers own managers and lifecycle only. They do not copy snapshots to atoms.

The standard provider shape is:

1. construct the manager without external subscriptions;
2. start subscriptions after commit;
3. register entity disposal;
4. expose the manager through context;
5. stop synchronously and dispose with the existing StrictMode-safe lifecycle.

Domain hooks live beside the provider and return state/projection/actions.

4. Composition-root wiring

Create typed adapters at the nearest common owner for:

• app_run -> preview_iframe;
• chat_stream -> plan_handoff;
• producers -> screenshot;
• user_input -> chat_stream (preserve the existing facade direction).

Record the final dependency graph in module headers and in a focused architecture test. It must remain acyclic.

5. Ownership boundary test

Extend src/state_machines/boundaries.test.ts with enforceable rules:

• state.ts and transition.ts cannot import @/atoms, Jotai, React, IPC, or another machine;
• a machine directory cannot import another machine's controller, registry, manager, or provider;
• machine lifecycle projection modules cannot export writable Jotai atoms;
• approved cross-process projection modules are allowlisted with a reason;
• approved UI/runtime atom imports in command adapters are inventoried;
• new uses of registerAtomWriter or projectToAtom outside the cross-process allowlist fail the test.

This is intentionally stricter for new code than for the initial migration. Temporary exceptions carry an owner and deletion PR.

Domain cleanup details

A. app_run

This is the first and most important cleanup because it has the most confused ownership and feeds preview_iframe.

State and projection

Add app_run/projection.ts with a pure, cached projectAppRun selector. Expose:

• lifecycle state;
• isLoading;
• active operation and startedAt;
• current RunUrl | null;
• operation error;
• stopped/ready status;
• capabilities such as canStart, canRestart, canStop, and canReload.

Do not expose invocation identity to ordinary UI consumers unless required for diagnostics.

Split previewRuntimeAtoms

Delete machine-owned storage:

• previewRunStateByAppIdAtom;
• currentPreviewRunStateAtom;
• currentPreviewLoadingAtom;
• currentPreviewRunStartedAtAtom;
• appUrlByAppIdAtom;
• currentAppUrlAtom;
• the dyad-app portion of previewErrorByAppIdAtom.

Audit before deciding the fate of:

• previewAppExitByAppIdAtom: if the UI needs the last output event timestamp independently of current machine state, rename it to lastPreviewExitEventByAppIdAtom and document it as diagnostics history; otherwise derive exit information from RunState;
• previewReloadTokenByAppIdAtom: move iframe identity changes into PreviewIframeState.iframeEpoch and delete the token;
• previewErrorByAppIdAtom: split independent iframe/client/sync diagnostics into explicitly named keyed stores, then compose display priority in a pure preview selector;
• console entries and package-manager warnings: retain as runtime/UI state.

Commands and producer output

• applyUrl changes machine state; it must not separately write a URL atom.
• Operation failure changes machine state; it must not separately write a lifecycle error atom.
• APP_EXIT must have one authoritative admission/commit path. Independent diagnostic history, if retained, is written only after the machine admits the correlated event.
• HMR/manual reload emits a typed iframe reload/restart event through the composition adapter rather than incrementing an atom token.
• Console logging and warnings remain adapter effects.

Runtime

Migrate AppRunController to TransactionalDispatcher.

Preserve:

• non-blocking run/stop IPC settlement;
• per-app command scheduling policy;
• invocation registration and claims;
• dispatch waiter settlement;
• legacy ref-less producer compatibility;
• external agent lifecycle operations;
• disposal of pending waiters and late settlements.

Delete:

• the hand-written processing flag;
• the pendingEvents FIFO;
• duplicate snapshot/observer ordering code;
• projection-writer lifecycle from AppRunManager.

B. chat_stream

Remove the streaming atom

Delete isStreamingByIdAtom after migrating all consumers.

Provide pure selectors:

• selectIsStreamActive(state);
• selectCanSubmitImmediately(state);
• selectCanCancel(state);
• selectStreamError(state).

React components that render one chat use a keyed machine selector. Tab-list rows subscribe per chat rather than reading a global map. If measurement shows that this creates unacceptable subscription overhead, add a manager-owned read-only active-chat index; do not restore a writable atom.

Migrate imperative consumers

• resyncChat receives a ChatStreamStatusFacade;
• plan_handoff receives the same facade and subscribes to the target chat;
• queue dispatch consults the controller snapshot, not the streaming atom;
• tests and hybrid harnesses drive the manager/controller rather than writing isStreamingByIdAtom directly.

Clarify retained Jotai state

Retain, with documented ownership:

• chatMessagesByIdAtom for optimistic/streaming renderer messages;
• streamingPreviewByChatIdAtom for high-frequency partial content;
• editable/persisted queued message data if the queue remains a user-editable renderer model;
• queuePausedByIdAtom only if pause is intentionally queue policy outside the stream lifecycle;
• chatErrorByIdAtom only for errors not represented by StreamState.

For each retained value, add a short ownership comment explaining why it is not derivable from StreamState. Split mixed error state if necessary.

Runtime

Migrate the controller transaction loop to TransactionalDispatcher, while retaining the command scheduler that allows long-lived stream work without blocking event admission.

Preserve:

• globally unique invocation refs;
• registration/cancel races;
• finalize side effects;
• queue ownership and durable follow-up acknowledgement;
• quiescent controller release;
• post-commit streamFinished delivery;
• late transport cleanup.

C. first_prompt

• Expand the provider context to expose a reference-stable snapshot source.
• Add useFirstPrompt() returning { state, projection, send/resume }.
• Move projectFirstPromptState to a pure cached selector without Jotai.
• Migrate home.tsx, TitleBar, SetupBanner, and ProviderSettingsPage to the hook.
• Delete firstPromptSagaProjectionWriteAtom, firstPromptSagaAtom, its manual subscription, and disposal reset.
• Keep home input, attachments, selected app, and setup-dialog visibility in their existing UI owners.
• Migrate controller transaction mechanics after projection removal so any behavior regression is bisectable.

D. image_generation

• Keep the manager-owned aggregate SnapshotStore because it provides a real retained cross-job read model.
• Rename getProjection/subscribeProjection to getJobsSnapshot/subscribeJobs for clarity.
• Add hooks for all jobs, pending count, and chat-visible jobs using external store selectors.
• Migrate the progress dialog, progress button, chat strip/input, and toast orchestration to those hooks or direct manager subscriptions.
• Delete _imageGenerationJobsAtom, imageGenerationJobsAtom, setImageGenerationJobsProjectionAtom, pendingImageGenerationsCountAtom, and chatImageGenerationJobsAtom.
• Retain dismissedImageGenerationJobIdsAtom as independent UI state.
• Keep terminal retention policy in the manager and test its reference stability.

E. version_preview

• Add an aggregate manager selector for active mutations, derived from keyed snapshots.
• Migrate ChatHeader from isAnyCheckoutVersionInProgressAtom to the selector.
• Delete activeCheckoutCounterAtom and isAnyCheckoutVersionInProgressAtom.
• Remove counter increment/decrement from the command adapter.
• Ensure the aggregate includes every mutation phase that previously held the counter, including recovery/return flows where appropriate.
• Migrate the custom controller to TransactionalDispatcher in a separate PR with trace comparison and recovery tests.

F. preview_iframe

• Stop watching previewRunStateByAppIdAtom.
• Wire an app-run lifecycle facade at the provider composition root.
• Send RUNTIME_RESTARTED with sufficient identity to reject duplicate or stale notifications; do not deduplicate only by startedAt.
• Route URL changes from committed app-run snapshots/events.
• Make PreviewIframeState.iframeEpoch the only iframe replacement/reload identity.
• Split preview iframe errors from app-run and sync errors; compose them only at the display selector.
• Migrate the controller to TransactionalDispatcher.

G. plan_handoff

• Replace watch-stream-idle's Jotai subscription with the injected ChatStreamStatusFacade.
• The watcher reads the current committed chat snapshot before subscribing, subscribes by chat key, and rechecks after subscription to close the check/subscribe race.
• Preserve task-scope cleanup when leaving the waiting state.
• Keep plan content in its existing renderer owner; the handoff snapshot retains only operation facts required for correctness.
• Migrate the controller to TransactionalDispatcher and timer leases.

H. screenshot

• Define a narrow ScreenshotRequestFacade with requestCapture(appId, source).
• Inject it into useCommitChanges, chat-stream command dependencies, and other producers.
• Delete pendingScreenshotAppIdsAtom and the provider consumer effect.
• Decide request coalescing explicitly in ScreenshotManager: replacement, queueing, or ignore-by-state must be a transition policy, not an incidental Map overwrite.
• Preserve app-key disposal and selector-settle watchdog behavior.

I. github_ops

• Preserve useGithubOps and projectGithubOps as the reference projection-free public API.
• Migrate the controller to TransactionalDispatcher.
• Remove its hand-written event queue.
• Retain domain-specific command concurrency and conflict-resolution runner registration.
• Run capability consistency and branch-inventory integration tests unchanged.

J. Main-process registries

connection_flow

• Use TransactionalDispatcher for state commit, observers, and timer lease cancellation if the commandless derived-effect model can be expressed without changing public synchronous claim semantics.
• Represent timeout installation/cancellation with TimerLeaseScope.
• Keep provider-specific timeout policy in the registry.
• If synchronous start/claimReturn results prevent direct dispatcher use, extract a small dispatcher-backed core and document the facade boundary.

mcp_oauth

• Separate state transaction mechanics from resource ownership: listener handles, authorization callbacks, waiters, port-close barriers, and provider aborts remain in the registry.
• Route state changes through TransactionalDispatcher.
• Use task/timer scopes for listeners and timeouts.
• Preserve the port/flow identity registry and late listener-close barriers.
• Add disposal and stale-callback conformance tests before changing structure.

user_input

The main registry remains authoritative and the renderer remains a cross-process read model.

In the first cleanup pass:

• rename projection types and comments consistently as a renderer read model;
• retain the single-writer hydration/revision logic;
• retain Jotai if its derived composition materially simplifies consumers;
• exclude it explicitly from the no-projection boundary rule.

In an optional later pass, compare a service-owned SnapshotStore plus domain hooks against the current atoms. Migrate only if it reduces total adapters and consumer complexity. Atom count alone is not sufficient justification.

Rollout

Each PR must be independently reviewable and must delete the compatibility path it replaces. Do not land new hooks while leaving indefinite dual consumption.

PR 1 — Ownership inventory and boundary enforcement

• Add the ownership table to repository documentation.
• Add boundary-test rules for new machine lifecycle atoms, atom mailboxes, and machine-to-machine imports.
• Allowlist current violations with explicit deletion PR numbers/order.
• Add selector-aware React binding tests.
• No production behavior changes.

PR 2 — First-prompt projection removal

This is the smallest proof that direct machine hooks can replace a global atom.

• Add useFirstPrompt.
• Migrate four consumers.
• Delete both projection atoms and manual synchronization.
• Add provider/hook tests, including StrictMode and disposal.

PR 3 — Image-generation projection removal

• Add manager projection hooks/selectors.
• Migrate consumers and toast orchestration.
• Delete machine projection atoms.
• Retain and test independent dismissal state.

PR 4 — Screenshot typed ingress

• Add ScreenshotRequestFacade.
• Migrate all producers.
• Delete the mailbox atom and consumer effect.
• Test simultaneous per-app requests and same-app coalescing.

PR 5 — App-run ownership split

• Add projectAppRun and direct hook consumers.
• Split runtime diagnostics from lifecycle state.
• Remove loading, URL, lifecycle error, exit overlap, and reload-token projections as resolved by the audit.
• Wire typed app-run events to preview_iframe.
• Keep the existing controller runtime for this PR to isolate ownership changes.

PR 6 — App-run transactional migration

• Move AppRunController to TransactionalDispatcher.
• Remove custom FIFO/commit plumbing and projection writer.
• Run controller conformance plus stale-output, external lifecycle, and disposal tests.

PR 7 — Chat-stream direct status API

• Add per-chat selectors/facade.
• Migrate React and imperative consumers.
• Delete isStreamingByIdAtom.
• Update hybrid fixtures to drive the real manager boundary.
• Preserve the existing controller runtime for bisection.

PR 8 — Plan-handoff and chat-stream transactional migrations

• Replace stream-idle atom watching with the facade.
• Migrate both custom controller loops.
• Preserve durable follow-up acknowledgement and post-commit completion notification.
• Run co-simulation and integration suites for submit/cancel/queue/handoff races.

PR 9 — Version-preview aggregate cleanup

• Replace the global checkout counter with a manager selector.
• Delete the atoms and adapter counter writes.
• Test concurrent per-app mutations and controller disposal.

PR 10 — Remaining renderer controller migrations

Use separate commits, and split into multiple PRs if review size grows:

1. github_ops;
2. preview_iframe;
3. version_preview.

Each migration requires before/after trace comparison and controller conformance.

PR 11 — Main-process registry consolidation

Prefer separate PRs per registry:

1. connection_flow;
2. mcp_oauth;
3. user_input runtime mechanics only if the dispatcher fits its synchronous registry contract.

Resource ownership and cross-process protocols receive focused tests rather than a mechanical bulk conversion.

PR 12 — Remove compatibility infrastructure

After all same-process writers are gone:

• delete unused registerAtomWriter/projectToAtom code if only the allowlisted cross-process projection no longer needs it;
• otherwise move the helpers next to the cross-process projection and narrow their names;
• remove all temporary boundary-test allowlist entries;
• update rules/state-machines.md, rules/jotai-state.md, and docs/why-state-machines.md;
• add a repository test asserting that no lifecycle atom names from this plan return.

Verification strategy

Pure transition tests

Every domain continues to run:

• reachable state/event exploration;
• state and command inventory checks;
• ignored-event reference identity checks;
• capability/transition consistency where interactive actions exist;
• stale invocation tests.

Projection removal must not require transition changes unless the old projection exposed a missing domain fact. Such a change is isolated and reviewed as behavior, not folded into a mechanical consumer migration.

Controller conformance

Every migrated controller runs runControllerConformanceSuite, covering:

• observer/subscriber/command re-entrancy;
• commit-before-notify ordering;
• synchronous throws and async rejections;
• disposal during commands;
• late emissions after disposal;
• recreate-after-dispose with stale callbacks;
• final cleanup and idempotent disposal.

Renderer tests

For each deleted atom:

• test the replacement hook under a test-owned manager;
• assert unrelated entity transitions do not rerender keyed consumers;
• assert selectors update synchronously after committed transitions;
• test provider replacement and StrictMode replay;
• test deletion cleanup.

Do not mock Jotai to simulate machine lifecycle after the migration.

Integration tests

Required scenarios:

• app run/restart/stop updates toolbar, iframe, URL, errors, and exit display from one committed snapshot path;
• proxy URL before/after IPC settlement;
• stale proxy/exit output after controller replacement;
• chat double-submit queues rather than drops;
• cancel before and after stream registration;
• stream completion wakes plan handoff without an atom edge;
• screenshot requests from chat completion and explicit commit;
• simultaneous image-generation jobs and terminal retention;
• simultaneous version mutations across apps;
• provider setup resumes first prompt without projection atoms.

Use the renderer+IPC integration harness where possible. Use Playwright only for behavior requiring the real iframe, Electron output subscription, or browser interaction. Rebuild before E2E runs.

Performance checks

Projection removal must not replace one global atom rerender with broad provider rerenders.

Measure or assert:

• keyed subscribers are notified only for their entity;
• high-frequency stream chunks do not rerender lifecycle-only consumers;
• console appends do not rerender app-run controls;
• aggregate indexes reuse their reference when their selected value is unchanged;
• controller retention remains bounded.

Migration rules

1. Consumer migration precedes projection deletion within the same PR.
2. A compatibility projection has exactly one writer until deletion.
3. Do not add new writes to a projection scheduled for removal.
4. Do not make a formerly read-only projection writable to ease migration.
5. Cross-machine adapters are injected; machine modules never import each other's owners.
6. When replacing an atom edge with a subscription, close the read-before-subscribe race by checking before and immediately after subscription.
7. One-shot events run only after the authoritative snapshot commit.
8. Preserve operation identity through new facade events.
9. UI forms clear or close only on authoritative settlement.
10. Every removed keyed atom has an explicit entity-deletion and provider disposal replacement.
11. Do not mix controller-runtime migration with domain behavior changes unless the old mechanics make separation impossible.
12. Any retained lifecycle-like atom must be justified in the ownership table and boundary allowlist.

Risks and mitigations

Broader React subscriptions

Direct external-store subscriptions could rerender more components than fine-grained atoms.

Mitigation: selector-aware keyed hooks, scalar selectors, equality tests, and keeping high-frequency content outside lifecycle snapshots.

Lost aggregate visibility

Deleting global maps can make “any entity active?” queries harder.

Mitigation: add manager-owned, read-only aggregate indexes only for actual consumers. Test reference stability and cleanup.

Cross-machine timing changes

Replacing atom observation with direct events can change callback ordering.

Mitigation: specify post-commit delivery, carry invocation identity, and test re-entrant sends. Do not emit lifecycle events from command side effects when the transition itself is authoritative.

Mixed error sources

Splitting previewErrorByAppIdAtom may change precedence between app-run, iframe, client, and sync errors.

Mitigation: inventory current precedence and encode it in one pure display selector with table-driven tests before storage changes.

Controller migration regressions

Custom runtimes may contain undocumented scheduling behavior.

Mitigation: capture before/after traces, characterize scheduler concurrency, run conformance, and migrate one high-blast-radius controller per PR.

Test harness dependence on writable atoms

Several tests currently set lifecycle atoms directly.

Mitigation: introduce small test manager drivers and event fixtures. Tests should exercise the same authority boundary as production.

Success criteria

The cleanup is complete when:

• no same-process machine lifecycle is mirrored into Jotai;
• app_run UI URL/loading/error state comes from its committed snapshot;
• isStreamingByIdAtom, the first-prompt projection atoms, image-generation projection atoms, screenshot mailbox atom, and version checkout counter are deleted;
• cross-machine coordination uses typed injected facades/events;
• every renderer controller uses TransactionalDispatcher;
• every remaining custom main-process registry has a documented resource or synchronous-contract reason and uses shared transaction/lease primitives where possible;
• providers own lifecycle but do not synchronize machine snapshots into atoms;
• aggregate views are read-only manager projections with bounded retention;
• boundary tests prevent reintroduction of lifecycle atoms and atom mailboxes;
• no UI regression occurs in run/restart, chat queue/cancel, plan handoff, image generation, screenshot capture, version preview, OAuth, or user-input flows;
• rules/state-machines.md, rules/jotai-state.md, and docs/why-state-machines.md describe the implemented architecture rather than the transitional one.

Expected outcome

The number of pure transitions and domain states will not necessarily shrink; they encode real workflow complexity. The surrounding code should shrink materially:

• fewer writable atoms and setters;
• no projection-writer lifecycle in renderer managers;
• fewer manual subscriptions and edge-detection effects;
• fewer hand-written controller queues;
• clearer domain hooks;
• explicit cross-machine dependencies;
• one place to answer each lifecycle question.

The desired review experience is that a contributor can determine:

1. where a lifecycle fact is stored;
2. which transition changes it;
3. which selector exposes it;
4. which command performs its effects;
5. which typed facade connects another machine;

without searching for a second atom, counter, ref, or effect that must agree.