🤖 Live compaction #309
Merged
🤖 Live compaction #309
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Remove compact_summary tool and replace with direct text streaming: - Model now streams summary text directly instead of using a tool call - Removed compact_summary tool definition and service implementation - Updated StreamingMessageAggregator to extract text from compaction streams - Updated WorkspaceStore to handle compaction on stream-end instead of tool completion - Removed toolPolicy requirement forcing compact_summary tool usage - Updated tests and mock scenarios to reflect new flow Benefits: 1. Live feedback: Summary streams as it's being generated (visible to user) 2. Context efficiency: No duplicate text in tool calls and assistant message 3. Simpler protocol: Fewer moving parts, less code to maintain Generated with cmux
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
cmux/tests/e2e/scenarios/slashCommands.spec.ts
Lines 90 to 96 in 5df7ade
Update /compact e2e expectation after removing compaction tool
The commit removes the compact_summary tool entirely, so compaction now streams text directly. The /compact e2e test still asserts that a tool-call-start event for compact_summary occurs. That expectation will now always fail, breaking the e2e suite. The test should be updated to assert the new behavior (e.g., streamed summary text) instead of a tool call.
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
Add special styling for compaction messages as they stream: New components (isolated complexity): - CompactionBackground: Animated shimmer effect during compaction - CompactingMessageContent: Container with max-height (300px) and sticky scroll Changes: - Added --color-plan-mode-rgb CSS variable for alpha blending - Pass isCompacting prop through MessageRenderer to AssistantMessage - Wrap streaming compaction content in special container with animation - Auto-scroll to bottom as compaction text streams in - Custom scrollbar styling with plan-mode colors UX improvements: - Visual distinction for compaction streams - Constrained height prevents overwhelming UI - Smooth shimmer animation provides feedback - Sticky scroll keeps latest content visible Generated with cmux
Move CompactionBackground from content wrapper to MessageWindow level: - Added backgroundEffect prop to MessageWindow for flexible background effects - Set MessageBlock to position: relative for absolute background positioning - Added z-index to MessageHeader and MessageContent to layer above background - CompactionBackground now covers entire message area, not just text content - Removed CompactionBackground from CompactingMessageContent (now at window level) This fixes the uncomfortable rectangular border effect by extending the shimmer animation to cover the full message window. Generated with cmux
Replace horizontal shimmer with futuristic green laser scanning effect: - Vertical bar sweeps back and forth across the message window - Green color (rgba 34, 197, 94) with gradient fade - Box shadow for glow effect - 2.5s scan cycle with ease-in-out timing - Refactored to proper React component structure This creates a more futuristic, sci-fi feel that better communicates the active compaction process. Generated with cmux
Replace tacky laser scanning with elegant gradient wave: - Smooth gradient using plan-mode color variables (matches theme) - Single directional sweep (left to right, 3s cycle) - Subtle opacity (0.6) for understated effect - Uses existing CSS color variables for consistency - Wider gradient spread (30-70% instead of sharp bar) Creates a more polished, professional appearance while still providing visual feedback that compaction is in progress. Generated with cmux
Increase visibility of gradient wave effect: - Increase opacity from 0.6 to 1.0 - Use stronger color-mix for center (25% plan-mode color) - Tighter gradient spread (25-75% instead of 30-70%) Effect is now more noticeable while still being elegant. Generated with cmux
Add multi-layered gradient for more polished appearance: - BaseGlow: Subtle pulsing radial gradient (provides depth) - GradientWave: Main sweeping gradient with slight blur (smooth motion) - HighlightWave: Crisp highlight that follows the main wave (adds definition) Visual improvements: - Smoother cubic-bezier easing for more natural motion - Multiple gradient stops for better color distribution - Slight blur on main wave for softer appearance - Pulsing base glow synced to 2s cycle (complements 3s sweep) Creates a more premium, polished look with depth and dimension. Generated with cmux
Improve animation smoothness and consistency: - Wider gradient waves (150% width) so always visible during sweep - Adjusted translate range (-50% to 100%) for seamless loop - Changed to linear timing for consistent, predictable motion - Slowed to 4s cycle for more relaxed, satisfying pace - Gradient stops repositioned for continuous presence The wave now walks smoothly across without start/stop gaps, creating a more satisfying, meditative effect. Generated with cmux
Replace static gradient wave with more dynamic animation: - Animated multi-directional gradient (400% size, 8s cycle) - Fast-moving shimmer layer (3s horizontal sweep with white highlight) - 8 floating particles with varied timing for organic movement Technical approach: - Background gradient animates position for smooth color flow - Shimmer uses background-position animation (GPU-accelerated) - Particles use transform: translate for float effect - All animations staggered with different delays/durations Visual improvements over previous version: - More fluid motion (multiple animation cycles at different speeds) - Shimmer adds light-reflection feel - Particles add depth and activity - Less predictable/monotonous than single linear sweep Inspired by modern loading animations from uiverse.io, CodePen shimmer effects, and animated gradient patterns.
When compaction is interrupted (e.g. user presses Ctrl+C), append "[truncated]" marker on a new line to the partial summary before saving it. Implementation: - Added handleCompactionAbort() method parallel to handleCompactionCompletion() - Checks if aborted stream was a compaction request (same logic as completion) - Extracts partial summary text streamed before interruption - Appends "\n\n[truncated]" to indicate incomplete summary - Calls performCompaction() with truncated summary Flow: 1. User interrupts compaction stream (Ctrl+C) 2. StreamManager emits stream-abort event 3. WorkspaceStore.processStreamEvent() receives abort 4. handleCompactionAbort() detects it was compaction 5. Fetches partial summary via aggregator.getCompactionSummary() 6. Adds [truncated] sentinel 7. performCompaction() replaces history with truncated summary This makes it clear when compaction didn't complete, helping users understand the summary is partial.
Removed sparkle particles and reduced shimmer opacity from 0.15 to 0.06 for a more subtle compaction background effect. Changes: - Reduced ShimmerLayer white opacity: 0.15 → 0.06 - Removed Particle styled component and related float animation - Removed particle array generation and rendering The animated gradient and shimmer are now the only visual elements, providing a cleaner, less busy background during compaction.
Changed shimmer scan line from white (rgba(255,255,255,0.06)) to plan-mode blue (--color-plan-mode-alpha) so it blends better with the animated gradient background and appears less distinct as a separate scan line element.
Removed src/test-temp-zbpEr8/test-repo submodule that accidentally got committed during testing. Added .gitignore patterns to prevent future test temporary directories from being committed: - src/test-temp-*/ - tests/**/test-temp-*/ This prevents similar mistakes where test fixtures create git repos that get accidentally tracked as submodules.
- Remove overflow-y: auto and auto-scroll logic - Add CSS mask-image gradient for line-by-line fade from top - Implies content above without janky scroll behavior
1. Anchor content to bottom of viewport - Add flexbox with justify-content: flex-end - Keeps newest content visible at bottom as it streams - Older content fades at top when exceeding max-height 2. Disable all tools during compaction - Set toolPolicy: [] in applyCompactionOverrides - Ensures compaction never has tool access regardless of UI mode - Add test to verify tools disabled even with base policy
**Ctrl+C during compaction**: Cancel and restore command - Interrupts stream - Removes compaction request + partial summary from history - Restores original /compact command to input for re-editing **Ctrl+A during compaction**: Accept early with [truncated] - Uses existing handleCompactionAbort flow - Adds [truncated] sentinel to partial summary - Performs compaction with incomplete summary **Implementation**: - Created src/utils/compaction/handler.ts to consolidate logic - Added ACCEPT_EARLY_COMPACTION keybind (Ctrl+A) - Extended ChatInputAPI with restoreText() method - Updated useAIViewKeybinds to handle both cases - Updated UI hints in ChatInput, AIView, UserMessage **Architecture**: - Compaction logic grouped in dedicated handler module - Minimal indirection - direct calls to IPC layer - Uses percentage-based truncateHistory API for removal
**Problem**: Ctrl+C during compaction resulted in empty chat because both Ctrl+C (cancel) and Ctrl+A (accept early) called interruptStream, triggering handleCompactionAbort which would always perform compaction. **Solution**: Add cancellingCompaction flag to distinguish between the two flows. **Implementation**: - WorkspaceStore: Added cancellingCompaction Set and markCompactionCancelling() method - handleCompactionAbort: Check flag before compaction, skip if set (Ctrl+C path) - cancelCompaction: Set flag before interrupt, prevent compaction - useAIViewKeybinds: Pass workspaceStore to cancelCompaction **Two distinct flows**: - Ctrl+C: Set flag → interrupt → skip compaction → truncate → restore command - Ctrl+A: No flag → interrupt → perform compaction with [truncated] **Benefits**: - Simple: Single short-lived flag (~10ms), self-cleaning - Frontend-only: All compaction logic stays in frontend - Well-commented: Clear explanations for future Assistants All tests pass (574 unit tests). _Generated with `cmux`_
**Problem**: Previous fix (98bbcc2) used in-memory flag that was lost on reload, causing compaction to happen anyway after page refresh. **Root cause**: Flag stored in WorkspaceStore (in-memory) doesn't persist across reloads. StreamAbortEvent arrives, flag is gone, compaction happens. **Solution**: Use localStorage to persist cancellation intent across reloads. **Implementation**: - Store cancellation marker in localStorage BEFORE calling interruptStream - Include messageId for freshness verification - handleCompactionAbort checks localStorage and verifies messageId matches - Clean up marker after handling (prevents stale data) - No flag state, no backend changes needed **Flow**: 1. Ctrl+C: Store {messageId, timestamp} in localStorage 2. Interrupt stream → StreamAbortEvent 3. handleCompactionAbort checks localStorage, verifies messageId 4. If cancelled: skip compaction, clean up marker 5. cancelCompaction truncates history and restores command **Benefits**: - Reload-safe: localStorage persists across page refreshes - MessageId verification ensures freshness (no stale cancellations) - Simpler than backend changes - Frontend-only solution All tests pass (574 unit tests). _Generated with `cmux`_
The assistant message ID changes with each stream attempt, causing mismatches. The compaction-request user message ID is stable and correctly identifies which compaction is being cancelled.
**Problem**: cancelCompaction was trying to calculate percentage truncation
after committing partial, leading to incorrect truncation removing too much history.
**Solution**: Add abandonPartial flag to interruptStream that deletes the partial
instead of committing it to history.
**Implementation**:
- Add options parameter to interruptStream IPC (abandonPartial?: boolean)
- Backend: if abandonPartial is true, delete partial instead of committing
- AIService: check if partial exists before committing on stream-abort
- cancelCompaction: pass {abandonPartial: true} when interrupting
- Simplified truncation: now just remove compaction-request message
**Benefits**:
- No race conditions with partial commit timing
- Simple percentage calculation (no +1 guesswork)
- Backend explicitly handles the cleanup
- Clear separation: Ctrl+C abandons partial, Ctrl+A commits with [truncated]
All tests pass (574 unit tests).
_Generated with `cmux`_
No need to truncate - the compaction-request message is just a normal user message. Leaving it in history is fine and shows what the user tried to do. Simplifies the code by removing all truncation logic from cancelCompaction.
When user cancels compaction with Ctrl+C, drop them into edit mode on the compaction-request message with the original /compact command. This makes it easy to edit the command or delete the message.
Editing state lives in AIView, not ChatInput. Pass setEditingMessage to useAIViewKeybinds and call it directly from cancelCompaction callback.
- Added findCompactionRequestMessage() helper to eliminate duplicated message finding logic - Refactored WorkspaceStore to use isCompactingStream() and findCompactionRequestMessage() - Simplified cancelCompaction() to use the new helper - Fixed outdated comment in storage.ts about messageId tracking Net result: -13 lines of duplicated code across WorkspaceStore methods
ammario
approved these changes
Oct 18, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Fixes #244
Summary
Replaced the
compact_summarytool with direct text streaming for conversation compaction. The model now streams the summary directly instead of using a tool call to return it.Changes
compact_summarytool definition and service implementationStreamingMessageAggregatorto extract compaction summary from text partsWorkspaceStoreto handle compaction on stream-end instead of tool completioncompact_summarytool usageBenefits
Testing
src/utils/messages/compactionOptions.test.ts)Generated with
cmux