Update Task tool config to show cleaner subagent information in the UI.
Simplifies the input display by showing only the prompt when no
optional fields are present, reducing visual clutter. Updates title
format to "Subagent / {type}" for better categorization. Enhances
result display to better handle complex agent response structures with
array content types, extracting text blocks for cleaner presentation.
Add visual distinction for the Task tool (subagent invocation) by
introducing a new 'agent' category with purple border styling. This
separates subagent tasks from regular task management tools
(TaskCreate, TaskUpdate, etc.) for clearer user feedback.
Also refactor terminal command layout in OneLineDisplay to properly
nest flex containers, fixing copy button alignment issues.
- fix top-pagination lockups by only locking when older messages are actually fetched
- make fetched older messages visible immediately by increasing `visibleMessageCount` on prepend
- prevent unintended auto-scroll-to-bottom during older-message loading and scroll restore
- replace state-based pagination offset with a ref to avoid stale offset/reload side effects
- ensure initial auto-scroll runs only after initial session load completes
- reset top-load lock/restore state and visible window when switching sessions
- loosen top-lock release near the top to avoid requiring a full down/up cycle
The chat UI could remain in a stale "Thinking/Processing" state when session IDs
did not line up exactly between view state (`currentSessionId`), selected route
session, pending session IDs, and provider lifecycle events. This was most visible
with Codex completion/abort flows, but the same mismatch risk existed in shared
handlers.
Unify lifecycle cleanup behavior in realtime handlers and make processing tracking
key off the active viewed session identity.
Changes:
- src/hooks/chat/useChatRealtimeHandlers.ts
- src/components/ChatInterface.tsx
- src/hooks/chat/useChatSessionState.ts
What changed:
- Added shared helpers in realtime handling:
- `collectSessionIds(...)` to normalize and dedupe candidate session IDs.
- `clearLoadingIndicators()` to consistently clear `isLoading`, abort UI, and status.
- `markSessionsAsCompleted(...)` to consistently notify inactive/not-processing state.
- Updated lifecycle branches to use shared cleanup logic:
- `cursor-result`
- `claude-complete`
- `codex-response` (`turn_complete` and `turn_failed`)
- `codex-complete`
- `session-aborted`
- Expanded completion/abort cleanup to include all relevant session IDs
(`latestMessage.sessionId`, `currentSessionId`, `selectedSession?.id`,
`pendingSessionId`, and Codex `actualSessionId` when present).
- Switched processing-session marking in `ChatInterface` to use
`selectedSession?.id || currentSessionId` instead of `currentSessionId` alone.
- Switched processing-session rehydration in `useChatSessionState` to use
the same active-view session identity fallback.
Result:
- Prevents stale loading indicators after completion/abort when IDs differ.
- Keeps processing session bookkeeping aligned with the currently viewed session.
- Reduces provider-specific drift by using one lifecycle cleanup pattern.
- This fixes an issue where the sidebar was showing 6+ even when there were only 5 sessions, due to the hasMore logic not accounting for the case where there are exactly 6 sessions.
It was also showing "Show more sessions" even where there were no more sessions to load.
- This was because `hasMore` was sometimes `undefined` and the logic checked for hasMore !== false, which treated undefined as true.
Now we explicitly check for hasMore === true to determine if there are more sessions to load.
When the “Frequently Used” section is visible, command clicks/hover could use a
UI-local index instead of the canonical `filteredCommands` index, causing the
wrong command to execute (e.g. clicking `/help` running `/clear`).
- map rendered menu items back to canonical command indices using a stable key
(`name + namespace/type + path`)
- use canonical index for hover/click selection callbacks
- deduplicate frequent commands from other grouped sections to avoid duplicate
rows and selection ambiguity
- keep and restore original inline comments, with clarifications where needed
The /cost command was listed as built-in but had no handler, causing execution to
fall through to custom command logic and return 400 ("command path is required").
- Add a built-in /cost handler in server/routes/commands.js
- Return the expected payload shape for the chat UI (`action: "cost"`, token usage,
estimated cost, model)
- Normalize token usage inputs and compute usage percentage
- Add provider-based default pricing for cost estimation
- Fix model selection in command execution context so codex uses `codexModel`
instead of `claudeModel`
- Improve frontend command error handling by parsing backend error responses and
showing meaningful error messages instead of a generic failure
Problem
Stop requests were unreliable because aborting depended on currentSessionId being set, Esc had no actual global abort binding, stale pending session ids could be reused, and abort failures were surfaced as successful interruptions. Codex sessions also used a soft abort flag without wiring SDK cancellation.
Changes
- Add global Escape key handler in chat while a run is loading/cancellable to trigger the same abort path as the Stop button.
- Harden abort session target selection in composer by resolving from multiple active session id sources (current, pending view, pending storage, cursor storage, selected session) and ignoring temporary new-session-* ids.
- Clear stale pendingSessionId when launching a brand-new session to prevent aborting an old run.
- Update realtime abort handling to respect backend success=false responses: keep loading state when abort fails and emit an explicit failure message instead of pretending interruption succeeded.
- Improve websocket send reliability by checking socket.readyState === WebSocket.OPEN directly before send.
- Implement real Codex cancellation via AbortController + runStreamed(..., { signal }), propagate aborted status, and suppress expected abort-error noise.
Impact
This makes both UI Stop and Esc-to-stop materially more reliable across Claude/Cursor/Codex flows, especially during early-session windows before currentSessionId is finalized, and prevents false-positive interrupted states when backend cancellation fails.
Validation
- npm run -s typecheck
- npm run -s build
- node --check server/openai-codex.js
The staged changes remove the main source of project-load latency by avoiding repeated full scans of ~/.codex/sessions for every project and by removing redundant client-side cursor session refetches.
Server changes (server/projects.js):\n- Add a per-request Codex index reference in getProjects so Codex metadata is built once and reused across all projects, including manually added ones.\n- Introduce normalizeComparablePath() to canonicalize project paths (including Windows long-path prefixes and case-insensitive matching on Windows).\n- Introduce findCodexJsonlFiles() + buildCodexSessionsIndex() to perform a single recursive Codex scan and group sessions by normalized cwd.\n- Update getCodexSessions() to accept indexRef and read from the prebuilt index, with fallback index construction when no ref is provided.\n- Preserve existing session limiting behavior (limit=5 default, limit=0 returns all).
Client changes (src/hooks/useProjectsState.ts):\n- Remove loadCursorSessionsForProjects(), which previously triggered one extra /api/cursor/sessions request per project after /api/projects.\n- Use /api/projects response directly during initial load and refresh.\n- Expand projectsHaveChanges() to treat both cursorSessions and codexSessions as external session deltas.\n- Keep refresh comparison aligned with external session updates by using includeExternalSessions=true in sidebar refresh path.
Impact:\n- Reduces backend work from roughly O(projects * codex_session_files) to O(codex_session_files + projects) for Codex discovery during a project load cycle.\n- Removes an additional client-side O(projects) network fan-out for Cursor session fetches.\n- Improves perceived and actual sidebar project-loading time, especially in large session datasets.
