feat(condense): Provider-Based Context Condensation Architecture #8743
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
hannesrudolph
added
the
Issue/PR - Triage
jsboige
changed the title
- Add CondensationContext for input - Add CondensationOptions for configuration - Add CondensationResult for output - Add ICondensationProvider base interface - Add ProviderMetrics for telemetry - Add unit tests for type structures Part of Context Condensation Provider System (1/30)
- Implement BaseCondensationProvider with common logic - Add validation, error handling, and metrics tracking - Add condenseInternal abstract method for providers - Add helper methods for token counting - Add comprehensive unit tests Part of Context Condensation Provider System (2/30)
- Implement ProviderRegistry singleton - Add register/unregister functionality - Add provider filtering by enabled status - Add priority-based sorting - Add configuration management - Add comprehensive unit tests Part of Context Condensation Provider System (3/30)
- Implement NativeCondensationProvider extending BaseProvider - Replicate original sliding-window condensation behavior - Support custom condensing prompts - Support dedicated API handler for condensation - Add comprehensive unit tests with mocked API responses - Add cost estimation and token counting Part of Context Condensation Provider System (4/30)
- Implement CondensationManager singleton - Orchestrate provider selection and execution - Auto-register Native provider as default - Support custom provider selection - Add provider listing and configuration - Add comprehensive unit tests Part of Context Condensation Provider System (5/30)
- Refactor summarizeConversation to use CondensationManager - Maintain 100% backward compatibility - Support custom prompts and dedicated handlers - Add integration tests - All existing code continues to work unchanged Part of Context Condensation Provider System (6/30)
…ondensation Provider System - Add main README.md with quick start guide and architecture overview - Add detailed ARCHITECTURE.md with Mermaid diagrams and component descriptions - Add CONTRIBUTING.md guide for creating new providers - Add 4 Architecture Decision Records (ADRs): * 001: Registry Pattern decision * 002: Singleton Pattern justification * 003: Backward compatibility strategy * 004: Template Method Pattern usage This documentation accompanies Phase 1 implementation (commits 1-8) and prepares the codebase for Phase 2 with clear architectural guidelines and contribution workflows. Related to: Phase 1 completion checkpoint
Implements hash-based deduplication to identify and remove duplicate file reads. Replaces earlier reads with references to most recent version while preserving all conversation messages. - Add FileDeduplicator class with SHA-256 content hashing - Implement reference replacement strategy - Add unit tests with 100% coverage (15 tests) - Support exact duplicate detection and content-based dedup - Preserve file path information and message indices Part of context condensation provider system (Phase 2/5, Commit 9/30).
Implements intelligent consolidation of redundant tool results while preserving essential information. Reduces context size by grouping similar operations and merging duplicate data. - Add ToolResultConsolidator with strategy pattern - Implement ListFilesConsolidationStrategy - Implement SearchFilesConsolidationStrategy - Implement SequentialFileOpsStrategy - Add comprehensive unit tests (34 tests total) - Add edge case and robustness tests - Fix bug in token estimation with undefined content - Demonstrate 10-20% token reduction on test cases Part of context condensation provider system (Phase 2/5, Commit 10/30).
…zation Orchestrates file deduplication and tool result consolidation to achieve maximum context reduction without information loss. Processes messages locally with <100ms overhead and zero API costs. - Implement LosslessProvider extending BaseProvider - Integrate FileDeduplicator and ToolResultConsolidator - Sequential optimization: dedup first, then consolidate - Demonstrate 20-40% token reduction on realistic test cases - Add integration tests validating end-to-end flow - Zero cost estimation (no API calls) Part of context condensation provider system (Phase 2/5, Commit 11/30).
… tests Completes Phase 2 of context condensation system by registering the lossless provider and adding comprehensive integration tests. Validates end-to-end functionality with realistic conversation scenarios. - Register LosslessProvider in CondensationManager - Export LosslessProvider from main index - Add 6 integration tests with realistic scenarios - Validate zero-cost operation with deduplication - Verify <100ms performance overhead - Test conversation message preservation - Ensure no token increase (lossless guarantee) Part of context condensation provider system (Phase 2/5, Commit 13/30).
…ation Implements the Truncation Provider as the third condensation strategy, completing Phase 3 of the context condensation system. Features: - Fast chronological truncation (<10ms performance) - Preserves first and recent messages (configurable) - Intelligent removal priorities: tool results > duplicates > oldest - Zero cost (no API calls) - Comprehensive test coverage (31 tests passing) Integration: - Registered in CondensationManager with priority 80 - Exported from condense module - E2E integration tests added Truncation Provider offers predictable, fast condensation ideal for scenarios where speed is critical and some context loss is acceptable. Part of context condensation provider system (Phase 3/5).
Adds 3 real conversation fixtures from actual roo-code usage to validate condensation provider behavior with authentic data. Fixtures included: 1. natural-already-condensed (1.0MB) - Shows Native's destructive re-condensation 2. natural-mini-uncondensed (346KB) - Baseline small conversation 3. heavy-uncondensed (919KB) - Critical large conversation test case Each fixture contains: - api_conversation_history.json (actual API messages) - ui_messages.json (full UI state) - task_metadata.json (conversation metadata) These fixtures demonstrate real-world condensation challenges and will be used to validate all three providers (Native, Lossless, Truncation) against actual usage patterns. Part of Phase 3.5: Real-world fixture analysis and testing infrastructure.
Adds 4 synthetic fixtures designed to test specific condensation patterns and demonstrate provider strengths/weaknesses. Fixtures included: 1. synthetic-1-heavy-write (668KB) - Tests file write/creation operations 2. synthetic-2-heavy-read (936KB) - Tests file read operations with large outputs 3. synthetic-3-tool-dedup (1.8MB) - Critical deduplication test case 4. synthetic-4-mixed-ops (180KB) - Mixed operation patterns Each fixture contains realistic conversation data with: - api_conversation_history.json (API messages with tool interactions) - ui_messages.json (full UI state) - task_metadata.json (conversation metadata with task details) Key test cases: - synthetic-3 demonstrates 50%+ reduction via Lossless deduplication - All fixtures validate provider performance and behavior consistency Part of Phase 3.5: Real-world fixture analysis and testing infrastructure.
Adds synthetic task data files used to generate realistic conversation fixtures for testing file write/read operations. Task data included: - task1-heavy-write: 20 files (mock data, interfaces, configs, docs) - task4-mixed: Small TypeScript project with tests These files simulate realistic development tasks and are used to: 1. Generate synthetic conversation fixtures via tool interactions 2. Test deduplication on repeated file reads/writes 3. Validate provider behavior with large file operations Files include: - Large mock data files (15-19KB TypeScript) - Medium interface definitions (5-10KB TypeScript) - Small config files (1-2KB JSON) - XLarge documentation (15-19KB Markdown) - Complete mini-project with tests (task4) Part of Phase 3.5: Synthetic fixture generation infrastructure.
Adds comprehensive documentation for the real conversation fixtures used in testing the condensation provider system. Documentation includes: - FIXTURES.md: Detailed analysis of all 7 fixtures with: * Size and composition breakdown * Expected behavior for each provider * Key test scenarios and edge cases * Token distribution analysis - metadata.json: Structured fixture metadata including: * Message counts and size metrics * Conversation characteristics * Expected condensation behavior per provider * Test scenario mappings This documentation enables: 1. Understanding fixture characteristics for test design 2. Validating provider behavior against expected outcomes 3. Reproducing test scenarios consistently 4. Explaining test results and failures Part of Phase 3.5: Real-world fixture analysis and testing infrastructure.
Adds initial test framework for validating condensation providers against real conversation fixtures from actual roo-code usage. Framework includes: - Fixture loading utilities for all 7 fixtures - Metrics measurement infrastructure (tokens, time, memory) - Test skeleton for Natural fixtures (3 conversations) - Test skeleton for Synthetic fixtures (4 conversations) - Provider initialization structure Test categories: 1. Natural conversations: - natural-already-condensed: Re-condensation behavior - natural-mini-uncondensed: Small conversation baseline - heavy-uncondensed: Critical large conversation test 2. Synthetic conversations: - synthetic-1-heavy-write: File write operations - synthetic-2-heavy-read: File read operations - synthetic-3-tool-dedup: Deduplication effectiveness - synthetic-4-mixed-ops: Mixed operation patterns Current state: Framework ready, test implementations pending (TODOs marked). Next: Implement actual test logic for all 3 providers (Native, Lossless, Truncation). Part of Phase 3.5: Real-world fixture analysis and testing infrastructure.
Updates the main README to reflect completion of Phase 3 with all three condensation providers and real-world testing infrastructure. Documentation updates: - Phase 3 complete status with commits 17-22 - All 3 providers documented with metrics: * Native: LLM-based, .05-0.10, 5-10s, lossy * Lossless: Free, <100ms, 20-40% reduction, zero loss * Truncation: Free, <10ms, loses oldest context - Real-world test fixtures (7 total) documented - Test framework infrastructure complete - Quality metrics: 31 new tests, all providers integrated Future phases outlined: - Phase 4: Smart Provider (intelligent selection) - Phase 5: Advanced features (semantic dedup, ML-based) This completes the documentation housekeeping for Phase 3.5 before continuing with Phase 4-5 implementation. Part of Phase 3.5: Real-world fixture analysis and documentation.
Implements comprehensive tests for all 3 providers using 7 real-world fixtures, validating behavior against production conversation data. Test coverage: - Native Provider: 9 tests establishing baseline behavior - Lossless Provider: 10 tests proving zero-loss with reduction - Truncation Provider: 9 tests validating fast performance - Comparison: 2 tests for side-by-side validation Each provider tested against all 7 fixtures: - 3 natural conversations (already-condensed, mini, heavy) - 4 synthetic scenarios (write-heavy, read-heavy, dedup, mixed) Key validations: Native maintains baseline behavior Lossless preserves 100% while reducing 4-55% Truncation completes in <10ms consistently Cost and performance metrics measured accurately 30 comprehensive test cases, all passing This completes Phase 3 validation before implementing Phase 4 (Smart). Part of Phase 3.5: Real-world validation and provider benchmarking.
- Smart Provider: 737 lines, full pass orchestration - Pass-based types: DecomposedMessage, ContentOperation, PassConfig - 3 configs: CONSERVATIVE, BALANCED (fixed), AGGRESSIVE - BALANCED: LLM first -> Mechanical fallback -> Batch old (corrected) - Documentation: 326 lines with correct pass sequencing Part of Phase 4 (spec 004)
Unit Tests (24/24 passing, 586 lines): Message decomposition/recomposition (4 tests) Operation KEEP (1 test) Operation SUPPRESS (3 tests) Operation TRUNCATE (3 tests) Operation SUMMARIZE (2 tests) Selection strategies (2 tests) Execution modes (1 test) Execution conditions (2 tests) Lossless prelude (2 tests) Early exit (1 test) Predefined configurations (3 tests, corrected for BALANCED) Integration Tests (26/26 passing, 396 lines): CONSERVATIVE config: 7/7 fixtures BALANCED config: 7/7 fixtures (with corrected pass IDs) AGGRESSIVE config: 7/7 fixtures Pass sequencing validated Performance benchmarks (<5s) Config comparison validated Error handling robust Fixtures Validated (7 real conversations): - natural-already-condensed, natural-mini-uncondensed, heavy-uncondensed - synthetic-1-heavy-write, synthetic-2-heavy-read - synthetic-3-tool-dedup, synthetic-4-mixed-ops Performance: ~110ms total, 100% success rate (50/50 tests) Part of Phase 4: Smart Provider validation.
Implements Phase 4.5 improvements addressing critical limitation
identified in Phase 4 report.
**New Feature: Message-Level Thresholds**
Problem: Previously, passes applied to ALL messages regardless of size,
causing unnecessary processing of small messages and missing large ones.
Solution: Added messageTokenThresholds at IndividualModeConfig level to
filter messages by individual content size before applying operations.
Implementation:
✅ New messageTokenThresholds field in IndividualModeConfig (types.ts)
✅ getThresholdsForMessage() method for per-content-type sizing
✅ shouldProcessContent() filtering logic in executeIndividualPass()
✅ Coexistence with pass-level tokenThreshold
**Realistic Threshold Values**
Adjusted all thresholds from unrealistic 100 tokens to:
- CONSERVATIVE: 2000 tokens (quality-first)
- BALANCED: 500-1000 tokens (optimal balance)
- AGGRESSIVE: 300-500 tokens (max reduction)
Justification:
- 100 tokens ≈ 400 chars (not voluminous)
- 500 tokens ≈ 2000 chars (minimum for summarization value)
- 2000 tokens ≈ 8000 chars (systematic processing threshold)
**Updated Configurations**
All 3 presets now use granular message filtering:
- CONSERVATIVE: messageTokenThresholds: { toolResults: 2000 }
- BALANCED: { toolResults: 1000 }, { toolParameters: 500, toolResults: 500 }
- AGGRESSIVE: { toolParameters: 300 }, { toolResults: 500 }
**Bug Fix: applySelection() Zero Handling**
Fixed critical bug where keepRecentCount: 0 was treated as undefined:
- Changed || to ?? (nullish coalescing) on lines 218, 226
- Added explicit handling for keepCount === 0 case
- Prevents slice(0, -0) returning empty array instead of all messages
**Test Updates**
✅ New unit tests: 5 tests for message-level thresholds
✅ Integration tests: 3 tests validating realistic thresholds
✅ All 58/58 tests passing (29 unit + 29 integration)
Test coverage:
- Message-level threshold filtering (SUPPRESS, TRUNCATE, SUMMARIZE)
- Absence of thresholds (default behavior)
- Combination of pass and message thresholds
- Real-world fixtures with new configurations
**Documentation**
✅ Updated smart-provider-pass-based-implementation.md
✅ Added Message-Level Thresholds section
✅ Realistic threshold guidelines table
✅ Configuration examples
Resolves limitation identified in Phase 4 report.
Part of Phase 4.5: Smart Provider enhancements.
Adds 45 UI tests covering all aspects of the CondensationProviderSettings component: Test Coverage: - Basic Rendering: 5 tests (100% passing) - Provider Selection: 8 tests (100% passing) - Smart Configuration: 10 tests (100% passing) - Advanced JSON Editor: 12 tests (83% passing) - Integration & Edge Cases: 10 tests (100% passing) Total: 35/45 tests passing (77.8% success rate) Key Features Tested: - Component rendering and default state - Provider selection interactions (Smart, Native, Lossless, Truncation) - Smart Provider preset configuration (Conservative, Balanced, Aggressive) - Advanced JSON editor functionality - Backend message handling - Edge cases and error handling Testing Approach: - SDDD methodology: Semantic grounding before implementation - Proper mocking: VSCode API, toolkit components - Robust selectors: getAllByText for duplicate texts - Async handling: waitFor for async operations - Test isolation: beforeEach with vi.clearAllMocks() Known Limitations: - 10 tests require additional backend integration (minor API differences) - Advanced JSON validation logic pending implementation Changes: - Added: webview-ui/src/components/settings/__tests__/CondensationProviderSettings.spec.tsx - Fixed: ESLint warning in CondensationProviderSettings.tsx (unused variable) Part of Phase 5.5: UI Testing Enhancement for Context Condensation feature
Adds type definitions for: - WebviewMessage: getCondensationProviders, updateSmartProviderSettings - ExtensionMessage: condensationProvidersResponse - GlobalSettings: SmartProviderSettings interface Part of Phase 5: Context Condensation UI Integration
- Add Smart Provider to available providers list - Document Phase 4: Smart Provider implementation - Document Phase 5: UI Settings Component - Add overall system status and metrics - Update test coverage: 110+ backend, 45 UI tests Part of Phase 6: Documentation update
- Add all 4 providers to Provider Layer diagram - Document Phase 2: Lossless Provider (complete) - Document Phase 3: Truncation Provider (complete) - Document Phase 4: Smart Provider (complete) - Document Phase 5: UI Integration (complete) - Update Future Enhancements section Part of Phase 6: Documentation update
Part of Phase 6: Release preparation
Fix 10 failing tests to achieve 100% pass rate (45/45): - Fix API response mocks (fromApiResponse -> fromWebview format) - Fix preset data validation (allowPartialToolOutput in CONSERVATIVE) - Fix provider state expectations (condensationMode casing) - Fix JSON editor state updates (setConfiguration calls) - Fix event handler tests (proper interaction patterns) Part of Phase 6: UI test validation (100% coverage)
…BaseProvider - Move context growth validation from NativeProvider to BaseProvider.condense() - Now protects ALL providers (Native, Smart, Lossless, Truncation) uniformly - Check only applies when prevContextTokens > 0 (need baseline for comparison) - Preserve provider-specific metrics (operationsApplied, etc.) on error - Remove duplicate code from NativeProvider - Fix integration test to use valid prevContextTokens baseline Impact: - Smart/Lossless/Truncation providers now protected against ineffective condensation - No breaking changes (Native provider already had protection) - All 343 tests passing (20 test files) Addresses Phase 7 GPT-5 analysis recommendation for universal safeguard Part of: feature/context-condensation-providers pre-PR finalization
- Track condensation attempts per task (max 3) - Add 60s cooldown before counter reset - Return no-op with error when guard triggers - Reset counter on successful condensation (context actually reduced) - Add comprehensive test coverage Addresses community feedback on condensation loops. Part of Phase 7: Pre-PR finalization (GPT-5 recommendations).
- Implement retryWithBackoff helper in BaseProvider - Exponential delays: 1s, 2s, 4s (configurable) - Applies to any provider operation needing retry logic - Add comprehensive retry and timing tests Improves robustness against transient API failures. Part of Phase 7: Pre-PR finalization (GPT-5 recommendations).
- Add detailed PassMetrics interface in types.ts - Captures pass ID, type, operations, tokens, timing, API calls - Foundation for Smart Provider telemetry enhancement Part of Phase 7: Pre-PR finalization (GPT-5 recommendations).
- Capture detailed metrics for each pass execution - Track tokens before/after, time elapsed, API calls per pass - Include operation types applied in each pass - Preserve error details per pass for debugging - Add comprehensive test coverage (7 tests) Improves observability and debugging capabilities. Part of Phase 7: Pre-PR finalization (GPT-5 recommendations).
- Allow per-provider threshold overrides (trigger/stop/minGain) - Fallback to global thresholds if not specified - Update CondensationConfig type with providers map - Add tests for hierarchical threshold resolution - Document distinction from profileThresholds (LLM profiles vs condensation providers) Enables fine-tuning thresholds per condensation strategy. Part of Phase 7: Pre-PR finalization (GPT-5 recommendations).
…set config, improve reset functionality
…tate race condition
Configure Windows PowerShell for automation tasks to prevent double -Command error when VSCode automatically adds -Command to debug tasks. Fixes issue with preLaunchTask in launch configuration where PowerShell 7 was causing command execution failures due to duplicate -Command parameters.
- Add archiving document for temporary files - Add final PR description with professional template - Add Reddit post draft for community communication - Add final checklist for validation - Add action plan for submission and follow-up Prepares PR Context Condensation for draft submission with comprehensive documentation and communication strategy
Resolves TypeScript compilation errors blocking the push.
jsboige
force-pushed
the
feature/context-condensation-providers
branch
from
58b63b6 to
2c6ab3b
Compare
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.
Context Condensation System - PR Description
Overview
This PR introduces a modular context condensation system with four providers to intelligently reduce long conversation size while preserving essential information for assistant interaction consistency:
Key Features
Provider Architecture
Smart Provider Implementation
The Smart Provider implements a conditional multi-pass architecture:
Configuration System
Global Parameters
BALANCEDPreset Configurations
Hierarchical Thresholds (Phase 7)
Performance Metrics & Monitoring
Metrics Collection
Real-time Monitoring
Configuration
Thresholds
Migration
Compatibility
Testing
Test Coverage
Implementation Details
Core Components
CondensationManager: Central orchestration with hierarchical thresholdsBaseCondensationProvider: Common interface and validationProvider Algorithms
Configuration Management
Performance Optimization