openchamber/packages/ui/src/components/chat/composer/DOCUMENTATION.md

Composer

The chat composer: the prompt language, the editor that renders it, and everything between typing and sending.

ChatInput.tsx (one directory up) is the orchestrator. It holds the composer's own state and wires these modules together; it should not grow logic that belongs to one of them.

ChatContainer.tsx keeps one ChatInput mounted while a new-session draft becomes its first session. Draft-only UI first fades for 120ms while the editor stays in place. The parent then moves the editor to its final session position with a 180ms transform-only FLIP animation. Reduced-motion mode skips these transitions. session-ui-store.ts marks sessions materialized from a submitted draft, so selecting an existing session while a draft is open switches without animation. Do not restore separate draft and session composer branches: remounting the editor loses focus and interrupts the transition. Keep the existing mobile fixed-position rules unchanged.

Layers

Directory	Owns
language/	What the text means: @ references, / and # tokens, markdown, and which picker a caret asks for
editor/	The CodeMirror view that renders the language and owns the caret
state/	Composer-local lifecycle state: ArrowUp/ArrowDown browsing, draft stash/restore, mobile shell, popup placement, draft targeting
submit/	Turning what the user has into what gets sent
attachments/	Files: paths, drop payloads
ui/	Presentation
text.ts	How inserted text meets the text already there
largeTextPaste.ts	Detect large plain-text pastes and build virtual .txt files
largeTextPasteOffer.ts	Ask-toast offer id begin/resolve (supersede + double-apply guards)

ChatInput.handlePaste owns paste orchestration: URL-over-selection markdown links, clipboard images (attach + citation), and large plain-text pastes. Large pastes (about 2,000 characters or 25 lines) follow the composer setting largeTextPasteBehavior (ask / attach / inline). Attaching creates an in-memory text/plain file named pasted-context-N.txt, inserts a bracket citation, and sends it through the same attachment pipeline as a manually picked .txt file. Ask-toast actions read live composer/attachment state so typing or other attaches between paste and choice stay consistent. Short text, images, and URL wraps keep their existing paths.

The prompt language

language/ is the single source of truth for composer syntax. Everything that needs to know what a token means — highlighting, send-time resolution, and the autocomplete triggers — goes through it.

This is the invariant that matters most in this module. Before it existed, the @ rule was written four times with divergent cleanup and the / rule three times with different valid character sets, so a token could be painted as a reference and then not resolve as one. Adding a construct meant finding every copy.

• mentions.ts — @ references. The start..end span is the reference itself and is what gets highlighted; in see @a/b.ts, the comma is sentence punctuation, not part of the file being referenced. Mentions are plain editable text: deleting a character edits the token and reopens the mention picker, the same way /skill tokens behave — not an atomic delete.
• prefixTokens.ts — /command, /skill, #snippet. Scanning is deliberately generous; membership in the command, skill or snippet registry is the authority, not the pattern. An unknown /token stays plain prose.
• triggers.ts — which picker a caret position asks for. Exactly one can be active, with precedence command > skill > snippet > mention.
• tokenize.ts — one pass producing every highlight range. Adding a construct to the language means adding it here, once.

The editor

editor/ wraps CodeMirror. The document is a plain string: getValue() is exactly what gets sent, so nothing downstream serializes a rich document model back into a prompt.

The document is not, however, the string it was given: CodeMirror normalizes line endings, so a \r\n pair becomes one break and the document ends up shorter than the inserted string. Never derive a caret position from the length of text you are inserting — a caret past the end makes dispatch throw, the transaction never applies, and the un-normalized text stays in React state to crash again on the next restore. Every edit that moves the caret goes through replaceWithCaret (editor/documentEdits.ts), which measures the change instead of the string.

The composer previously painted a transparent <textarea> over a mirror <div>. That restricted highlighting to styles which do not change glyph advance width — colour, background, underline — because anything else made the mirror drift out from under the caret. Bold and italic were impossible, and the overlay was disabled outright on mobile, where wrapped text drifted anyway. Those constraints are gone; adding a width-affecting style is now a question of design, not of feasibility.

Selection rendering: every device runs CodeMirror's drawSelection() — it keeps typing on the drawn-selection code path, and removing it makes CodeMirror enforce cursor association on the native selection, which iOS answers with severe input lag. That much is not platform-specific and must not be undone. What differs is who paints the selection, and composerSelectionExtension (editor/theme.ts) picks that per platform.

When CodeMirror 6.43.9's iOS predicate does not match, composerNativeSelectionExtension layers over drawSelection(): it re-shows the native selection, and — only while a range is selected — the native caret, hiding the painted layers those replace. The native selection is the one that shows for two reasons: the painted layer sits behind the content, so tokens with their own background (inline code, fences) cover it completely; and the platform's selection drag handles attach to the visible native selection and take their colour from the caret, so a transparent caret means invisible handles. The range-only caret scoping is load-bearing — a native caret visible while typing makes the browser re-render its caret UI after every keystroke, felt as severe input lag.

When CodeMirror 6.43.9's exact iOS predicate matches, composerIOSSelectionExtension leaves selection-handle geometry and appearance to CodeMirror. CodeMirror puts the handles in .cm-selectionLayer, normally at z-index: -1; the extension raises that layer above the content so opaque token backgrounds cannot cover them, and leaves it transparent to touch. The handle dots extend 8px past their range; matching scroller padding and negative margin expand the clip area without moving the text or changing the composer height. iOS still paints its taller system selection overlay even when CSS makes ::selection transparent. The extension therefore suppresses CodeMirror's synthetic selection rectangles on iOS while leaving its handles, cursor path and nativeSelectionHidden facet active. Otherwise the grey system highlight and themed rectangle overlap with visibly different heights. Do not add a second custom layer or custom handles here: overlapping translucent rectangles make selection darker at their seams and imitated handles drift from the geometry WebKit actually manipulates. What iOS avoids is installing the native-selection workaround above: explicitly restoring native paint and caret makes WebKit re-measure them after every decoration redraw, and the composer rebuilds every decoration on every keystroke. That cost is felt worst during IME composition.

The non-iOS native selection tint comes from --primary, not the selection token: themes define --interactive-selection with its own alpha, so mixing it with transparent again is nearly invisible. The iOS system overlay owns its visible selection fill.

The content element keeps the existing correction policy: on in the mobile UI, off elsewhere. CodeMirror also reads the attribute and reverts Apple and Android's insert-period-on-double-space only when its value is exactly off. editor/autocorrect.ts uses the HTML standard's ASCII case-insensitive autocorrect keywords to keep desktop word correction off while avoiding that CodeMirror-only revert. Its platform checks deliberately match CodeMirror's own browser flags.

composerLanguage.ts retokenizes the whole document on every change. The composer holds a prompt, not a source file: it is short enough that a full pass is cheaper and far simpler than incremental mapping, and it keeps the editor and the send path reading the same grammar.

Ordering rules worth knowing

• editor/ComposerEditor.tsx forwards a click on the composer's padding by focusing the view before setting the selection: CodeMirror reveals its drawn caret through a class it only writes while applying an update, so the selection has to be the update that follows the focus.
• submit/buildOutgoingMessage.ts flattens queued messages, the composer text, context drafts and linked references into OpenCode's one-primary-plus-parts shape. The oldest queued message becomes primary. Every attached context item (inline comments, terminal selections, browser annotations, PR context, linked issue/PR) becomes its own synthetic text part carrying structured metadata built by lib/messages/contextParts.ts; the timeline reads that metadata back to render context blocks. PR instructions precede the PR diff. The same module's buildComposerContext captures that context when a message is queued instead of sent: the chips leave the composer with the message (as QueuedContextParts on the queue item), the server or the VS Code auto-send delivers them through queuedContextToParts, and editing the queued message puts them back. A queued message is placed as captured — its mention, file mentions, and skill instruction were resolved when it was queued, never at delivery — and its context follows it before the next queued message.
• Local slash commands are planned by submit/slashCommands.ts before any attached context is consumed. Commands that act on session or UI state (/undo, /redo, /compact, /timeline, /handoff-review) take only their command text and leave comments, files, and linked context attached; commands that produce a prompt (/btw and the magic prompts) send that context with the prompt they produce. Session actions are planned only when a session exists, so typing one into a new-session draft stays on the normal send path. A local command is never queued as text: queueing runs it instead. A failed prompt command restores everything it consumed: text, confirmed mentions, files, comment drafts, and pending synthetic context.
• state/useComposerDraft.ts — a draft belongs to a (runtime, directory, session) identity. Writes are debounced while typing but forced at every edge where the page may stop running, because a pending timer is not a saved draft. Two orderings are load-bearing: the debounced write is skipped once while a draft is being restored, and a deleted draft's empty signature is recorded before a queued write could resurrect it. Fork replay text and files arrive in input-store.pendingComposerRestore, addressed to the fork's runtime, directory, and session. The hook consumes them after loading that identity's draft. Selection alone is not enough: the deferred chat column can still show the source composer. Ordinary pending text insertions keep their existing path in ChatInput.
• state/useDraftTarget.ts — the draft can target a directory that does not exist yet (a worktree being created). It must survive not appearing in the branch list, or the selector snaps back to the project root mid-creation. It also owns the advisory dirty state for the selected directory, clearing it as soon as the target changes so a warning never names a previous branch.
• ui/DraftTargetSelectors.tsx owns the controlled project/worktree picker state and registers its application shortcuts locally. The desktop project picker is a searchable popup: it ranks the current projects with rankByQuery over display label and path, keeps the query and the active result as transient local state that resets on every close, and commits through the existing project-change flow only on explicit activation. Filtering changes the result area below the anchored input without moving the search field. The worktree Select and the mobile bottom sheets are unchanged. The selectors only consume their shared prefix while the draft target UI is mounted.

Input recall ownership

Prompt recall has two owners on purpose.

• packages/ui/src/stores/useInputHistoryStore.ts owns the persisted source of truth. It keeps the runtime-scoped global bucket and the runtime + directory
  • session bucket, each capped by the configurable input-history limit. That setting defaults to 40 entries. Recall reads the current session's bucket by default; the Chat setting can widen it to every project on the runtime.
• state/useMessageHistory.ts owns only keyboard traversal through whichever bucket the composer was given. Moving away from a position stores the composer's current text and attachments as an overlay for that position, so the live draft and any edit made to a recalled prompt survive a round trip through history. Overlays never rewrite stored history; sending resets them.
• ChatInput.tsx applies the recalled text and attachments to the composer and places the caret.

In session scope the composer merges two sources, oldest first: the visible transcript's user prompts (useUserMessageHistory in sync-context.tsx), so sessions that predate the persisted store still recall, and the persisted session bucket, which adds attachments and keeps prompts a revert hid from the timeline. A prompt present in both collapses to the persisted entry. Global scope reads the persisted runtime bucket only.

Mobile

state/useMobileComposerShell.ts and state/useMobileViewportPin.ts are mostly not state machines but corrections for specific platform behaviors: mobile browsers dismissing the keyboard before a tap's click lands, iOS refusing programmatic focus outside a gesture, WebKit leaving the layout viewport panned after the keyboard hides, overlay chains handing off through a frame where nothing is open.

Every timeout and flushSync in them has a reason recorded next to it, and none of them is verifiable outside a real device. Change them only against hardware.

Testing

The package has no DOM test environment, so coverage stops at the state and logic layers: the language, the submit assembly, path and drop handling, text splicing, large-paste detection, paste-offer invalidation, input-history traversal, and the CodeMirror language extension at the EditorState level.

Rendering, focus, keyboard behavior, IME and WKWebView are not covered by tests and are verified by hand. That includes ArrowUp and ArrowDown recall, caret placement after recall, restored drafts, and any edited-entry overlay. Do not report a change to them as validated on the strength of type-check and unit tests.

Run tests per file (bun test <path>): mock.module is process-global, so suites that install module mocks are order-dependent.

Enter preference

keyboardPolicy.ts owns the submission decision. Until the Chat setting is changed, desktop Enter sends, mobile and focus mode require Ctrl/Cmd+Enter, and Shift-modified Enter does not send. An explicit choice applies across shared composers; Ctrl/Cmd+Enter sends in either configured mode.

CodeMirror's deferred mobile Enter loses modifier information. Untouched settings restore Shift to keep the original policy. Once configured, with mobile autocapitalization enabled, the editor cannot distinguish its Shift flag from an intentional Shift press and does not restore Shift. Consequently, deferred Shift+Enter can send when Enter-to-send is enabled and cannot serve as the send shortcut when it is disabled. Ctrl/Cmd+Enter remains the supported modified send shortcut on this path.