Persist composer drafts per thread and reuse hidden draft threads#128

🟢 Low src/composerDraftStore.ts:433

If images contains duplicates sharing the same blob URL, the first accepted image's URL gets revoked when processing the second. Consider only revoking URLs for duplicates whose URL differs from any accepted image's URL.

✅ Resolved in 751d864218c4e5976bb100e21a74cd717b879348

Revoked blob URLs break image previews on send failure

Medium Severity

clearComposerDraftContent revokes blob preview URLs for all composer images, but composerImagesSnapshot (a shallow copy taken earlier at line 1635) holds references to the same image objects with those now-invalid blob URLs. When the send fails and the error handler restores images via addComposerImagesToDraft(composerImagesSnapshot), the restored images have revoked previewUrl values, causing broken image thumbnails in the composer.

Fix in Cursor Fix in Web

Duplicated storage key constant across two files

Low Severity

COMPOSER_DRAFT_STORAGE_KEY is defined independently in both ChatView.tsx and composerDraftStore.ts with the same string value "t3code:composer-drafts:v1". If one is updated without the other, the debug helper readComposerDraftStorageChars in ChatView.tsx would silently read from the wrong key.

Fix in Cursor Fix in Web

Draft thread reuse silently drops branch/worktree options

Medium Severity

When handleNewThread is called with options containing branch and worktreePath (via the "new chat" keyboard shortcut that inherits branch from the active thread), both the stored-draft reuse path and the active-draft reuse path return early without applying the branch/worktree options to the reused thread. The options parameter is only used when a brand-new thread is created.

Fix in Cursor Fix in Web

🟢 Low src/composerDraftStore.ts:124

Hydrated images get Date.now() for lastModified (from fileFromDataUrl), but the dedup key includes lastModified. This breaks deduplication for attachments restored from localStorage. Consider either persisting lastModified alongside other attachment data, or excluding it from the dedup key.

✅ Resolved in 4ae74fc4ea8b465eea666f2a3ef09ff30b99ddae

🟢 Low src/composerDraftStore.ts:234

Empty files (e.g., data:text/plain;base64,) have an empty string payload after split. The check !payload rejects empty strings, returning null and causing data loss. Consider using payload === undefined instead.

No longer relevant as of a5ed08c8f51ecd15e37708015adf1fd7c046cc6b

Blob URLs revoked too early break optimistic message images

Medium Severity

clearComposerContent revokes blob URLs via revokeObjectPreviewUrl before clearing images from the store. However, those same blob URLs are still referenced by the optimistic user message attachments (captured in composerImagesSnapshot and added to optimisticUserMessages moments earlier). After revocation, the <img src="blob:..."> elements in the optimistic message bubble will show broken images until the server message replaces them. The old code did not revoke blob URLs during send — only on thread change — so optimistic previews remained intact.

Fix in Cursor Fix in Web

Model stored permanently, preventing draft auto-removal

Low Severity

onModelSelect passes its value through resolveModelSlug, which never returns null (it falls back to DEFAULT_MODEL). The store's setModel then stores this always-non-null value. Unlike setEffort — which normalizes DEFAULT_REASONING to null — there's no equivalent normalization for the default model. Since clearComposerContent preserves model and shouldRemoveDraft requires model === null, any thread where the user ever selected a model (including re-selecting the default) creates a draft entry that never auto-removes. Over many threads these orphaned draft entries accumulate in localStorage.

Fix in Cursor Fix in Web

🟡 Medium src/composerDraftStore.ts:870

clearComposerContent should revoke blob URLs for images before clearing (like clearThreadDraft does on lines 899-901). Without this, blob memory leaks until page unload.

Kept clearComposerContent non-revoking intentionally because those blob URLs are also referenced by optimistic user-message previews during send. Revoking there breaks thumbnails until the server response lands. Instead, this update adds explicit blob URL cleanup when optimistic messages are removed (both on server reconciliation and send-failure rollback), so previews stay intact during send while object URLs are still released promptly after they are no longer shown.

Thread errors silently lost for local draft threads

Medium Severity

setThreadError dispatches a SET_ERROR action to state.threads, but local draft threads only exist in the zustand draft store, not in state.threads. The updateThread reducer maps over state.threads and silently finds no match. Meanwhile buildLocalDraftThread hardcodes error: null, so the ThreadErrorBanner never renders errors for draft threads. This affects image validation errors (wrong file type, size limit, attachment count) and onSend failure messages — the user gets no feedback.

Fix in Cursor Fix in Web

Store action isDraftThread is never consumed externally

Low Severity

The isDraftThread action is defined in the ComposerDraftStoreState interface and implemented in the store, but no component or module ever selects or calls it. It's dead code that adds unnecessary surface area to the store.

Fix in Cursor Fix in Web

Race condition when clearing draft before server sync

High Severity

clearDraftThread removes the draft from draftThreadsByThreadId immediately after turnStartSucceeded, but the newly created server thread (from thread.create) may not yet be synced to state.threads via the async event stream. This leaves a window where both threadExists and draftThreadExists are false in the route guard, causing routeThreadExists to be false and triggering an unwanted redirect to /. The old code had an explicit getSnapshot() sync after thread.create, which was removed in this refactor.

Fix in Cursor Fix in Web

🟢 Low src/composerDraftStore.ts:1000

Returning the shared mutable EMPTY_THREAD_DRAFT object risks state corruption if any consumer mutates it. Consider returning createEmptyThreadDraft() instead, or freeze EMPTY_THREAD_DRAFT with Object.freeze().

Applied. EMPTY_THREAD_DRAFT and its nested empty arrays are now frozen with Object.freeze(...) so the shared fallback cannot be mutated accidentally, while preserving referential stability for the selector fallback.

Unused variable serializationFailedImageIds is dead code

Low Severity

serializationFailedImageIds is declared and populated inside the image-persistence useEffect, but it is never read after the Promise.all completes. The non-persisted image tracking is handled separately by syncPersistedAttachments via a deferred localStorage check, making this array entirely dead code.

Fix in Cursor Fix in Web

🟡 Medium src/composerDraftStore.ts:493

When removing a draft thread mapping (line 493), the corresponding entry in draftsByThreadId is not deleted. This leaves orphaned draft content (potentially with large base64 images) in localStorage indefinitely. Consider also deleting nextDraftsByThreadId entries when removing orphaned thread mappings—this applies to similar cleanup in clearProjectDraftThreadId and clearProjectDraftThreadById.

✅ Resolved in cb830d562fde3876cbad1061989621c4c91ee807

🟢 Low src/composerDraftStore.ts:359

Using split(",", 2) truncates non-base64 payloads containing literal commas (e.g., data:text/plain,item1,item2). Consider splitting only on the first comma by using indexOf instead.

✅ Resolved in cb830d562fde3876cbad1061989621c4c91ee807

🟢 Low routes/_chat.$threadId.tsx:143

Consider using Object.hasOwn() instead of !== undefined to check if the draft thread exists. Direct property access on draftThreadsByThreadId returns inherited prototype properties (e.g., toString, constructor) as truthy values, bypassing the 404 redirect for those thread IDs.

✅ Resolved in cb830d562fde3876cbad1061989621c4c91ee807

Store strips DEFAULT_MODEL override for non-default base models

Medium Severity

The setModel store action unconditionally normalizes DEFAULT_MODEL ("gpt-5.3-codex") to null. The onModelSelect caller already passes null when the selected model equals the thread's base model. When a thread's base model differs from DEFAULT_MODEL (e.g., "gpt-5.2-codex") and the user explicitly picks DEFAULT_MODEL, the store's additional filtering discards this legitimate override. The composer falls back to the base model, ignoring the user's selection. For local draft threads with no server-side fallback, this results in the wrong model being used permanently.

Fix in Cursor Fix in Web

🟢 Low src/composerDraftStore.ts:461

Using "branch" in options with ?? null means { branch: undefined } clears the existing value instead of preserving it. Consider using options?.branch === undefined ? existingThread?.branch ?? null : options.branch ?? null to match setDraftThreadContext behavior.

✅ Resolved in 9e99b8efdc43870933fe114a94a82c2663471f2e

🟢 Low src/composerDraftStore.ts:124

The frozen EMPTY_PERSISTED_ATTACHMENTS array is typed as mutable PersistedComposerImageAttachment[], so consumers may call .push() or .sort() and get a runtime TypeError. Consider typing it as readonly or documenting that returned drafts should not be mutated in place.

Error dispatch bypasses local draft thread error handling

High Severity

When a local draft thread in worktree mode has no branch selected, onSend dispatches SET_ERROR directly via dispatch instead of using setThreadError. Since local draft threads don't exist in state.threads, the store's updateThread silently ignores the action, and the user sees no error feedback explaining why their message wasn't sent. The setThreadError helper at line 1514 was specifically designed to route errors to localDraftErrorsByThreadId for non-server threads.

Fix in Cursor Fix in Web

Worktree setup script skipped for draft threads

Medium Severity

The isServerThread guard at line 1775 prevents the project setup script from running for local draft threads in worktree mode. After the worktree is created but before thread.turn.start, thread.create is dispatched — but the setup script is never retried after that. The worktree environment may be uninitialized (e.g. missing npm install) when the first turn starts.

Fix in Cursor Fix in Web

BranchToolbar errors silently dropped for draft threads

Medium Severity

The setThreadError helper in BranchToolbar directly calls dispatch({ type: "SET_ERROR", ... }) without handling local draft threads. Since draft threads don't exist in state.threads, the reducer's updateThread silently no-ops. Any branch checkout or creation errors (e.g., from checkoutMutation or createBranchAndCheckoutMutation onError callbacks) are silently lost when the active thread is a local draft, leaving the user with no error feedback. Unlike ChatView, this component was not updated to route errors to the local draft error state.

Fix in Cursor Fix in Web