Changelog

All notable changes to the Flowbaby extension will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.6.0] - 2025-12-11

Changed

• Documentation refresh: Updated extension README and marketplace-facing documentation for clarity around setup, features, and troubleshooting. No functional code changes compared to 0.5.23.

[0.5.23] - 2025-12-11

Added

• Per-Workspace Python Bridge Daemon (Plan 054): Introduced a long-lived Python bridge daemon that eliminates per-request Python startup overhead (~2-3 seconds) by keeping Cognee imported and warm. Retrieval and capture requests now respond in ~300ms instead of 2-3s.
• New Configuration Options:
  • Flowbaby.bridgeMode: Choose between daemon (default, faster) or spawn (legacy per-request mode) for bridge execution
  • Flowbaby.daemonIdleTimeoutMinutes: Configure how long the daemon stays alive after inactivity (default: 5 minutes)
• Daemon Lifecycle Management: The extension automatically starts, monitors, and restarts the daemon as needed. Graceful shutdown on workspace close.
• Automatic Fallback: If daemon mode fails repeatedly, the extension automatically falls back to spawn-per-request mode for reliability.

Changed

• Bridge Architecture: Retrieval and ingestion now route through the daemon manager when daemon mode is enabled, with transparent fallback to the existing spawn-per-request pattern.

Technical

• Added PythonBridgeDaemonManager TypeScript class for daemon lifecycle management
• Added daemon.py Python script as the long-lived bridge process entrypoint
• Added JSON-RPC 2.0 communication protocol over stdio for daemon IPC
• Added daemon health monitoring with automatic restart and exponential backoff

[0.5.22] - 2025-12-10

Fixed

• Data Integrity Check False Positives: Fixed initialization data-integrity check that was incorrectly reporting "Data mismatch: X SQLite entries but only Y vector tables" warnings. The previous implementation counted LanceDB directories (a fixed ~6 tables in Cognee 0.4.x) instead of actual row counts, producing false positives once SQLite rows exceeded the table count. The fix now compares SQLite document counts against actual LanceDB embedding row counts from the DocumentChunk_text table.

Added

• Shared Data Integrity Utilities: New data_integrity_utils.py module provides shared helper functions (get_sqlite_document_count, get_lancedb_embedding_count, evaluate_data_health) used by both init.py and recover_data.py for consistent data integrity checking.
• Improved Health Semantics: Data integrity warnings now use smarter thresholds—small workspaces (≤5 documents) with minor discrepancies are treated as healthy to avoid noise, while significant mismatches in larger workspaces trigger appropriate warnings.

Changed

• LanceDB Row Counting: The lancedb_count field in initialization JSON output now represents actual embedding row count rather than table directory count. This is a semantic change but maintains backward compatibility for consumers checking the healthy field.

[0.5.21] - 2025-12-10

Changed

• Retrieval Timeout Increased: Increased retrieval timeout from 15 seconds to 30 seconds to accommodate large workspaces and cold-start environments. Retrieval operations will now wait longer before timing out, reducing spurious failures in legitimate high-latency scenarios.
• Retrieval Diagnostics: Added phase markers (bridge call start/finish) and duration logging to retrieval operations for improved performance diagnostics and future bridge-daemon compatibility (Plan 054).

Added

• Capture Input Visibility: Added status bar message when capture input opens ($(edit) Flowbaby capture: input box open at top — Enter to save, Esc to cancel) to draw attention to the input UI.
• First-Time Capture Hint: One-time onboarding toast message for new users explaining that the capture input appears at the top of the window, stored in globalState to avoid repeated notifications.
• Improved Capture Prompt: Updated capture input prompt and placeholder text for better discoverability (Flowbaby capture: type in the top input box (Esc cancels)).

[0.5.20] - 2025-12-10

Fixed

• Tool Registration Guards: Fixed over-aggressive guard logic that incorrectly set setToolsRegistered(true) based on host inventory presence rather than actual registration outcomes. Guards now follow evidence-only semantics per architecture invariant 4.2.
• All-or-Nothing Tool Set Registration: Tool registration now treats both flowbaby_storeMemory and flowbaby_retrieveMemory as an atomic unit—the guard is only set if all tools register successfully or return known duplicate errors.
• Copilot Tool Availability: Resolved issue where Flowbaby tools would not appear in GitHub Copilot's tool list after non-initial activations because guards were being set prematurely.

Added

• Duplicate Error Classifiers: Added isKnownDuplicateToolError() and isKnownDuplicateParticipantError() helper functions to centralize error classification and provide forward-compatible matching for VS Code API changes.
• Host Tool Snapshots: Added createHostToolSnapshot() for compact diagnostic snapshots of registered tools (observational only, does not affect guard state per invariant 4.3.1).
• Plan 056 Test Suite: Added comprehensive tests for guard behavior, duplicate-error classification, tool/participant independence (invariant 4.5.4), and evidence-only semantics.

Changed

• Participant Registration: Updated registerFlowbabyParticipant() to use centralized isKnownDuplicateParticipantError() classifier and moved guard setting to immediately after successful registration.
• Guard Lifecycle: Registration guards are now completely independent—tool guard state does not affect participant guard state and vice versa, ensuring one subsystem's registration issues don't block the other.

[0.5.19] - 2025-12-07

Added

• Activation ledger now records workspace path, health/requirements state, registration guards, and deactivation details for every activation, enabling lifecycle reconstruction across test runs.
• Structured safePush diagnostics capture context fingerprints, host tool snapshots, and test-mode push probes so DisposableStore anomalies can be attributed to specific intents.
• Workspace initialization isolation from Plan 050 is preserved under repeated activations, with each workspace tracked independently so a broken workspace cannot contaminate others sharing the same window or host.
• Single-activation lifecycle model from Plan 051 is enforced via registration helpers and activation ledger state, ensuring that extension-owned registrations are routed through lifecycle-safe paths and never pushed into known-disposed contexts.

Changed

• Language model tool and chat participant registration skip when host inventory already contains Flowbaby assets, reducing duplicate registration noise in reused test hosts.
• Test-mode prompt handling now logs suppression reasons with activation metadata and avoids modal dialogs that block CI runs.
• Debug logging automatically enables in test mode to surface lifecycle and registration traces during automated suites.

[0.5.18] - 2025-12-07

Added

• Global Shortcut: Ctrl/Cmd+Alt+F now works across VS Code views without focus restrictions, still prompting for capture input when no editor selection is available.
• Session Toggle: New Flowbaby.sessionManagement.enabled setting (default on) lets users disable session IDs for stateless interactions.
• Dependency Mismatch Notice: Early hash check marks the workspace as needing setup and prompts to refresh when bridge dependencies are out of date.

Changed

• Graph Retriever Debug Gating: Cognee graph completion debug logs now obey the Flowbaby.debugLogging setting to reduce noisy output by default.
• Workspace Isolation: Initialization status is tracked per workspace to avoid a broken workspace contaminating others in the same window.

[0.5.17] - 2025-12-05

Added

• Dependency Enforcement: Added strict verification of Python environment dependencies. The extension now blocks operations and prompts for updates if requirements.txt hashes do not match the installed environment.
• Environment Guardrails: Added FlowbabySetupService checks to FlowbabyContextProvider and command handlers to prevent crashes from stale dependencies.
• Update Modals: Added user-friendly modals to guide users through updating managed or external environments.

Changed

• Verification Logic: FlowbabySetupService now tracks verification state and only allows operations when the environment is fully verified.

[0.5.16] - 2025-12-05

Fixed

• Ingestion Crash: Fixed a critical UnboundLocalError in the background ingestion process (cognify-only mode) that caused crashes when API keys were missing.
• Notification Throttling: Removed throttling for failure notifications to ensure users are always alerted to issues. Success notifications are now only suppressed by user setting, not by frequency.

[0.5.15] - 2025-12-05

Changed - Cognee Upgrade & Session Infrastructure

• Cognee 0.4.1 Upgrade: Upgraded managed Python environments to cognee==0.4.1 for improved stability and native session management support.
• Session-Aware Retrieval: Added infrastructure to pass session IDs to the memory engine when available, laying the foundation for conversational memory.
• Workspace Caching: Enhanced workspace-local caching configuration (CACHE_ROOT_DIRECTORY) to support future session features.
• Isolation: Maintained strict workspace isolation and backward compatibility for all ingestion/retrieval workflows.

[0.5.14] - 2025-12-05

Added - Unified Session Context

• Implemented unified session context for chat threads. Flowbaby now maintains conversational continuity by passing a consistent session_id to the memory engine for each chat thread.
• Updated Python bridge to accept and map __user_session_id to Cognee's session_id.

[0.5.13] - 2025-12-03

Changed - Maintenance Release

• Version bump only to align with latest internal planning and release numbering. No user-visible functionality changes beyond those already included in 0.5.12.

[0.5.12] - 2025-12-03

Fixed - Windows Installation and Stability

Windows Support Release - Resolves remaining Windows installation issues and makes the bridge fully reliable on Windows.

Improvements

• Background Ingestion on Windows: Switched background ingestion to use pythonw.exe when available, preventing stray console windows and improving stability for long-running operations.
• Initialization Race Fix: Resolved a fresh workspace initialization race on Windows, ensuring the bridge environment and Kuzu databases are fully ready before marking setup as complete.
• Kuzu and DLL Handling: Added targeted Windows Kuzu tests and fixed Python detection / Kuzu DLL load issues, resulting in the first fully working end-to-end Windows configuration.
• Platform Prerequisites: Updated documentation to call out the Visual C++ Redistributable requirement on Windows, reducing installation surprises.

[0.5.11] - 2025-12-01

Added - Plan 047: Enhanced Setup Diagnostics

Diagnostics Release - Improves troubleshooting for environment setup failures by exposing detailed command execution logs and system errors.

Improvements

• Command Transparency: Setup logs now show the exact command, arguments, and working directory for every subprocess execution.
• System Error Codes: Failures now report specific Node.js system error codes (e.g., ENOENT, EACCES) instead of generic messages.
• Full Stack Traces: Error logs include full stack traces to pinpoint the exact failure location in the code.
• Automatic Log Focus: The Flowbaby Output Channel is automatically revealed when a setup error occurs, ensuring immediate visibility of diagnostic information.
• Captured Output: Standard output from failed commands (which was previously swallowed) is now included in the error logs.

Windows Reliability Improvements

• Background Ingestion on Windows: Switched background ingestion to use pythonw.exe when available, preventing stray console windows and improving stability for long-running operations.
• Initialization Race Fix: Resolved a fresh workspace initialization race on Windows, ensuring the bridge environment and Kuzu databases are fully ready before marking setup as complete.
• Kuzu and DLL Handling: Added targeted Windows Kuzu tests and fixed Python detection / Kuzu DLL load issues, resulting in the first fully working end-to-end Windows configuration.
• Platform Prerequisites: Updated documentation to call out the Visual C++ Redistributable requirement on Windows, reducing installation surprises.

[0.5.10] - 2025-02-18

Fixed - Plan 046: Windows Spawn Fix

Hotfix Release - Resolves "Environment creation failed" error on Windows when workspace path contains spaces.

Issues Resolved

• Environment Creation Failure: Fixed a regression in v0.5.9 where Flowbaby: Initialize Workspace failed immediately on Windows if the path contained spaces.
• Root Cause: The manual quoting logic introduced in v0.5.9 conflicted with spawn's shell: true behavior on Windows, leading to malformed commands.
• Fix: Switched FlowbabySetupService.runCommand to use shell: false, allowing Node.js to handle argument escaping natively and robustly across all platforms.

Technical Changes

• FlowbabySetupService: Removed quoteIfNecessary helper and manual quoting.
• Spawn Options: Changed shell: true to shell: false for python and pip commands.
• Verification: Confirmed that FlowbabyClient already uses shell: false and was unaffected.

[0.5.9] - 2025-12-01

Fixed - Plan 022: Windows Path Spaces Fix

Critical Fix - Resolves environment creation failures on Windows when the workspace path contains spaces.

Issues Resolved

• Environment Creation Failure: Fixed a bug where Flowbaby: Initialize Workspace would fail or create stray folders if the workspace path contained spaces (e.g., C:\Users\My Name\Project).
• Root Cause: Node.js spawn with shell: true does not automatically quote arguments on Windows, causing paths with spaces to be split into multiple arguments.
• Fix: Implemented robust quoting logic in FlowbabySetupService to ensure both the command and arguments are correctly quoted when they contain spaces.

Technical Changes

• FlowbabySetupService: Added quoteIfNecessary helper in runCommand to wrap paths in double quotes if they contain spaces and aren't already quoted.
• Cross-Platform Safety: Verified that the quoting logic is safe for Windows (cmd.exe) and Unix-like systems (/bin/sh), ensuring no regressions on macOS/Linux.
• Test Coverage: Added unit tests to FlowbabySetupService.test.ts to verify that commands and arguments are correctly quoted before being passed to spawn.

[0.5.8] - 2025-11-30

Fixed - Hotfix: Fresh Workspace Memory Operations

Hotfix Release - Fixes issues discovered when using memory tools on fresh workspaces without stored data.

Issues Resolved

1. DatasetNotFoundError on fresh workspace retrieval: When querying @flowbaby or using flowbaby_retrieveMemory on a fresh workspace with no stored memories, the extension now gracefully returns empty results instead of throwing DatasetNotFoundError: No datasets found. This is the expected UX - the workspace simply has no memories yet.

2. BackgroundOperationManager not initialized error: When using flowbaby_storeMemory tool before full workspace setup, the extension now shows a helpful prompt to initialize the workspace instead of failing with a confusing "BackgroundOperationManager not initialized" error.

3. Misleading success log on staging failure: The "✅ Memory staged" log message now only appears when staging actually succeeds. Previously it showed success before checking the result, then showed failure notification.

Technical Changes

Python Bridge (retrieve.py)

• Added: Detection for DatasetNotFoundError in search exception handler
• Behavior: Fresh workspaces with no data return {"success": true, "results": [], "message": "No data has been ingested yet"}

Agent Commands (ingestForAgent.ts)

• Added: Try-catch wrapper around BackgroundOperationManager.getInstance() call
• Behavior: Returns helpful NOT_INITIALIZED error with "Initialize Now" action button
• UX: User sees clear path to fix instead of cryptic error

Store Memory Tool (storeMemoryTool.ts)

• Fixed: Success log message now appears after response check, not before
• Fixed: Failure log includes actual error message from response

[0.5.7] - 2025-11-30

Fixed - Hotfix: Smoke Test Issues from v0.5.6

Hotfix Release - Addresses five UX issues discovered during v0.5.6 smoke testing.

Issues Resolved

1. Ingestion failure notifications not showing: Failed memory ingestions via flowbaby_storeMemory tool now show warning toast with "View Logs" option. (Note: success notifications can be disabled via flowbaby.notifications.showIngestionSuccess, but failure notifications always show.)

2. Chat participant icon missing: @flowbaby icon now displays correctly in chat. Fixed icon path from non-existent icon.png to actual flowbaby-icon-tightcrop.png.

3. Missing API key causes infinite "working..." state: When API key is not configured, @flowbaby chat now shows helpful error message with link to configure instead of hanging indefinitely.

4. API key setup notice auto-dismissing: Post-initialization API key prompt now uses modal dialog that stays until user explicitly dismisses it.

5. Status bar not updating after API key config: After setting API key via command, status bar now updates from yellow "API Key Required" to green "Ready" state.

Technical Changes

• Added failure notification to StoreMemoryTool.invoke() for non-success responses
• Fixed participant.iconPath to use correct file path
• Added API key check in chat participant before retrieval operations
• Changed post-init prompt to modal dialog ({ modal: true })
• Added status bar update in Flowbaby.setApiKey command handler

[0.5.6] - 2025-11-30

Fixed - Plan 045 Hotfix: Smoke Test Issues

Hotfix Release - Addresses three issues discovered during v0.5.5 smoke testing.

Issues Resolved

1. Progress notification stuck: "Initializing Flowbaby databases..." progress notification no longer blocks waiting for user input. Post-init API key prompt now shows after progress completes.

2. Post-init prompt timing: API key setup prompt now appears after initialization fully completes, not while progress is still showing.

3. Configure API Key command not found: Flowbaby.configureApiKey command now works from Command Palette (was only Flowbaby.setApiKey before).

Technical Changes

• Moved post-init API key prompt outside withProgress callback to non-blocking .then() pattern
• Added Flowbaby.configureApiKey command as alias to Flowbaby.setApiKey
• Added getApiKeyState() method to FlowbabyClient for cached state access

[0.5.5] - 2025-11-30

Fixed - Plan 045: Fix Fresh Install API Key Blocking

Fresh Install Experience Release - Eliminates API key blocking issue that prevented new users from completing setup.

Problem Resolved

• Issue: Fresh installations failed with "LLM_API_KEY not found in environment" error
• Impact: New users couldn't complete Flowbaby setup without already having an API key configured
• Root Cause: init.py required API key before user had opportunity to set it

Changes

Python Bridge (init.py)

• Modified: API key validation now logs warning instead of returning error
• Added: api_key_configured and llm_ready fields to initialization response
• Behavior: Initialization completes successfully, directories and databases created without API key

TypeScript Client (flowbabyClient.ts)

• Added: ApiKeyState interface for centralized API key tracking
• Added: InitializeResult interface with success status and API key state
• Added: hasApiKey() public method for pre-checking API key availability
• Modified: initialize() now returns InitializeResult with API key state

Status Bar (FlowbabyStatusBar.ts)

• Added: NeedsApiKey status with distinct key icon and warning background
• Behavior: Shows $(key) icon when initialized but API key not configured

Extension Activation (extension.ts)

• Added: Post-initialization API key prompt when no key configured
• Added: Graceful handling of InitializeResult return type
• Behavior: Shows "Set your API key to enable LLM operations" with action button

Graceful Degradation

• Added: Pre-check in ingestForAgent and FlowbabyContextProvider.retrieveContext
• Behavior: If no API key, prompts user with "Set API Key" action before attempting operation
• Message: "Flowbaby memory operations require an API key."

User Experience

1. Fresh install now completes successfully
2. Status bar shows key icon indicating API key needed
3. User prompted to set API key after initialization
4. Memory operations prompt for API key if not configured
5. All existing functionality preserved when API key is set

[0.5.2] - 2025-11-29

Removed - Plan 044: Remove Non-Functional Auto-Ingest Setting

Configuration Cleanup Release - Removes experimental setting that never functioned.

Removed

• Setting: Flowbaby.autoIngestConversations removed from VS Code settings
• Code: Step 6 "feedback loop" code block removed from chat participant
• Rationale: Feature was introduced as experimental but never worked due to Cognee SDK limitations (file hashing bug). Keeping non-functional settings creates user confusion.

Impact

• Users who had this setting enabled will see no behavior change (it was already non-functional)
• Configuration surface is now cleaner and only exposes working features
• Any future auto-ingest or feedback-loop features will be redesigned under the current architecture (e.g., via structured summaries and Flowbaby tools) rather than reusing this removed pattern

[0.5.1] - 2025-11-28

Changed - Plan 043: Tool Descriptions and UX Improvements

Agent Integration & UX Polish Release - Enhances tool descriptions for better agent behavior and improves user experience.

Enhanced Tool Descriptions (Milestone 1)

• Updated: flowbaby_storeMemory and flowbaby_retrieveMemory tools now include comprehensive, structured instructions (~1,800 characters each) that encode the full Flowbaby memory contract
• Benefit: Agents automatically understand when, how, and what to store/retrieve without requiring custom .agent.md files
• Content: Descriptions include guidance on timing, query construction, what to capture, retrieval limits, and structure guidance

Configurable Success Notifications (Milestone 2)

• New Setting: flowbaby.notifications.showIngestionSuccess (default: true)
• Behavior: When false, suppresses success toast notifications ("Memory staged...", "Flowbaby finished...") while preserving all error notifications
• Use Case: Power users who find frequent success toasts distracting can disable them while still receiving critical error alerts

Keyboard Shortcut Change (Milestone 3)

• Changed: Default capture shortcut from Ctrl+Alt+C / Cmd+Alt+C to Ctrl+Alt+F / Cmd+Alt+F
• Rationale: Avoids potential conflicts with other extensions and common VS Code operations
• Migration: Users with custom keybindings should update to the new shortcut

Documentation

• Updated README.md with new keyboard shortcut references
• Added flowbaby.notifications.showIngestionSuccess to Configuration table
• Updated walkthrough step to reflect new shortcut

[0.5.0] - 2025-11-28

Changed - Plan 042: Comprehensive Lint Cleanup

Code Quality Release - Establishes consistent code quality standards across TypeScript, Python, and Markdown.

TypeScript ESLint (Milestone 1-5)

• Resolved: All 272 ESLint violations (12 errors, 260 warnings → 0 errors, 0 warnings)
• Updated: ESLint naming-convention rules to allow snake_case for Python bridge JSON contracts
• Fixed: 9 @typescript-eslint/no-explicit-any by adding proper type definitions
• Fixed: 9 @typescript-eslint/no-unused-vars by removing/prefixing unused variables
• Fixed: 3 @typescript-eslint/no-var-requires by converting to ES module imports
• Auto-fixed: 47 curly braces and useless escape warnings

Python Linting (Milestone 6)

• Added: ruff linter for Python bridge code (requirements-dev.txt)
• Created: extension/bridge/ruff.toml configuration
• Resolved: All Python lint errors (1322 → 0)
• Added: npm run lint:python script

Markdown Linting (Milestone 7)

• Updated: .markdownlint.json with appropriate rule exceptions
• Verified: Zero markdownlint errors on user-facing documentation
• Added: npm run lint:markdown script

Developer Experience

• New: npm run lint:all - Comprehensive lint check (TypeScript + Python + Markdown)
• New: npm run lint:python - Bridge code linting with ruff
• New: npm run lint:markdown - User-facing doc linting
• New: .github/workflows/lint.yml - CI lint enforcement on push/PR

[0.4.9]

Fixed - Plan 041: Capture Command UX Fixes

UX Improvement Release - Improves the Ctrl+Alt+C (Cmd+Alt+C on Mac) capture workflow for intuitive behavior.

Editor Selection Prioritization

• Fixed: Ctrl+Alt+C capture command now correctly uses highlighted text from the active editor
• Behavior: When text is selected in the editor, the input box pre-fills with that selection
• Benefit: Users can capture highlighted code or notes directly without manual copying

Cancellation Logic Fix

• Fixed: Pressing Escape now properly cancels the capture operation
• Previous: Escape would fall back to clipboard content, causing unintended captures
• Now: Escape shows "Capture cancelled" message and aborts without any ingestion

Clipboard Fallback (Power User Feature)

• Preserved: Submitting an empty input box (Enter with no text) still reads from clipboard
• Use Case: Power users can copy content to clipboard, press Ctrl+Alt+C, then Enter to capture

Content Source Transparency

• Improved: Output channel now logs the source of captured content:
  • "Capturing from Editor Selection (X chars)"
  • "Capturing from User Input (X chars)"
  • "Capturing from Clipboard (X chars)"
• Benefit: Users can verify what content was actually captured via Output > Flowbaby

Empty-All-Sources Handling

• Added: When no selection exists, input is empty, and clipboard is empty, shows "Nothing to capture" info message
• Benefit: Clear feedback instead of silent failure when there's nothing to capture

[0.4.8]

Fixed - Plan 040.1: First-Time Initialization Hotfix

Critical Hotfix - Resolves first-time workspace initialization failure introduced in v0.4.7 that required window reload.

FlowbabyClient Recreation After Environment Setup (M1)

• Bug: Fresh workspaces showed "No module named 'cognee'" error after clicking "Initialize" because FlowbabyClient was created at activation time with system Python before venv existed
• Root Cause: When createEnvironment() succeeded, the existing client retained its stale pythonPath: "python3" instead of detecting the newly-created .flowbaby/venv
• Fix: Recreate FlowbabyClient after successful environment creation to detect the new venv Python interpreter
• Benefit: Users can now use @flowbaby immediately after initialization without any window reload

Initialization Timeout Fix (M1.1)

• Bug: First-run database creation could timeout on slower machines because initialize() used the default 10-second timeout
• Fix: Increased init.py timeout from 10 seconds to 60 seconds
• Benefit: First-run initialization (SQLite, Kuzu, LanceDB setup) completes successfully on slower systems

Initialization Logging (M2, M3)

• Added detailed progress logging throughout the initialization flow:
  • [Plan 040] Starting workspace initialization...
  • [Setup] Virtual environment created successfully.
  • [Setup] Starting dependency installation...
  • [Setup] Dependency installation complete.
  • [Plan 040] ✅ Environment created successfully
  • [Plan 040] Recreating FlowbabyClient with new environment...
  • [Plan 040] ✅ Flowbaby client initialized successfully
• Benefit: Users and support can trace initialization issues via Output > Flowbaby channel

User Experience Improvements

• No-reload workflow: First-time initialization now completes in a single process
• Clear progress: Setup notifications show venv creation → dependency install → database init
• Error transparency: Detailed logging aids troubleshooting when issues occur

[0.4.7]

Fixed - Plan 040: Initialization JSON Errors and No-Reload Workflow

User Experience Release - Enables frictionless first-run experience by fixing JSON parse errors, eliminating window reloads, and providing clearer health status messaging.

M1: Fix init.py Stdout Pollution (Critical)

• Bug: The cognee SDK prints "User X has registered" and other messages to stdout during database initialization, corrupting the JSON output and causing JSON.parse() failures in the TypeScript client
• Root Cause: create_db_and_tables(), get_graph_engine(), and other cognee operations print directly to stdout
• Fix: Added suppress_stdout() context manager that captures all stdout/stderr during cognee operations and redirects to file logger
• Contract: init.py now emits exactly one JSON line to stdout and nothing else

M2: Chain flowbabyClient.initialize() After Environment Setup

• Bug: The "Initialize Workspace" command only created the venv but did not initialize the Flowbaby client, forcing users to reload the window
• Fix: After successful environment creation, the command now chains into flowbabyClient.initialize(), sets clientInitialized = true, and registers all providers/tools
• Benefit: Users can immediately use @flowbaby after clicking "Initialize" without any page reload
• Added: Structured error handling with user-facing notifications and Output channel logging for initialization failures

M3: Unify Health Check Logic for Fresh Workspaces

• Bug: Fresh workspaces with only .flowbaby/logs directory were incorrectly marked as "BROKEN" instead of "FRESH"
• Root Cause: checkWorkspaceHealth() returned BROKEN if .flowbaby existed but bridge-env.json was missing
• Fix: Missing bridge-env.json now correctly returns FRESH (setup required), not BROKEN (repair required)
• Benefit: Fresh workspaces show "Initialize" prompt instead of confusing "Repair Environment" prompt
• Updated: UX messaging now clearly distinguishes between "needs setup" (FRESH) and "needs repair" (BROKEN) states

M4: Increase Initialization Timeout

• Bug: First-run initialization could time out prematurely (15s) during database creation on slower machines or when downloading LanceDB embedding models
• Fix: Increased timeout from 15 seconds to 60 seconds for first-run initialization
• Benefit: Slow first-run initializations complete successfully without false timeout failures

User Experience Improvements

• No-reload workflow: Complete initialization from setup to ready state without any window reload
• Clear messaging: Distinct prompts for fresh workspaces ("Initialize") vs broken environments ("Repair")
• Error transparency: Initialization failures now show user-facing notifications with actionable guidance

[0.4.6]

Fixed - Plan 039: Initialization UX and Security Hardening

User Experience and Security Release - Resolves activation deadlock, improves initialization guidance, and implements security hardening from Plan 037 audit.

M1: Fix Activation Deadlock

• Bug: Extension activation could hang indefinitely when initialization failed, requiring VS Code restart
• Root Cause: Blocking await vscode.window.showWarningMessage() in failure path prevented activate() from returning
• Fix: Replaced blocking await with non-blocking .then() callback pattern, ensuring activation completes within 2 seconds

M2: Register Flowbaby.initializeWorkspace Command

• New: Added Flowbaby.initializeWorkspace command for workspace initialization
• Change: Flowbaby.setupEnvironment retained as backward-compatible alias
• Benefit: Consistent command naming within Flowbaby.* namespace

M3: Proactive Health Check

• New: checkWorkspaceHealth() method returns FRESH | BROKEN | VALID status
• Change: Activation flow now checks workspace health before attempting client initialization
• Benefit: Targeted UX guidance based on actual workspace state (fresh vs corrupted vs ready)
• Added: Migration marker validation (.migration-in-progress check) per architecture review R3

M4: Full Database Initialization

• Verified: init.py already performs complete database priming (SQLite, Kuzu, LanceDB) for fresh workspaces
• No changes required: Existing implementation meets acceptance criteria

M5: Remove .env API Key Support (Security - F2)

• Breaking Change: Removed workspace .env file API key loading
• Migration: Users must use Flowbaby: Set API Key command for secure SecretStorage
• Rationale: Plaintext API keys in .env files are credential exposure risk
• Files Updated: flowbabyClient.ts, init.py, ingest.py, retrieve.py, list_memories.py, validate_memories.py

M6: Audit Logging (Security - F7)

• New: AuditLogger class for security event tracking
• Format: JSON-lines structured logging to .flowbaby/logs/audit.jsonl
• Events Logged: API key changes, memory clear operations, environment initialization
• Benefit: Security audit trail for credential and data operations

M7: Safe Memory Clear with Soft-Delete (Security - F8)

• Change: clearMemory() now moves data to .flowbaby/.trash/{timestamp}/ instead of permanent deletion
• Benefit: Recovery possible for accidentally cleared memories
• New: purgeTrash() method for permanent trash cleanup
• Audit: All clear operations logged via AuditLogger

M8: Version and Release Artifacts

• Version: 0.4.5 → 0.4.6
• Changelog: Updated with Plan 039 changes

Breaking Changes

• API Key Source: Workspace .env files no longer used for API key resolution. Use Flowbaby: Set API Key command or set LLM_API_KEY environment variable.

Security Improvements

• Credentials now stored in VS Code SecretStorage (encrypted)
• Security-relevant operations logged to audit trail
• Data deletion uses soft-delete with recovery option

[0.4.5]

Fixed - Plan 034: Initialization Bugs and Background Notification Issues

Maintenance Release - Resolves 3 critical bugs affecting new workspace initialization and background processing notifications.

Database Initialization for Fresh Workspaces

• Bug: Fresh workspaces threw sqlite3.OperationalError: no such table: principals during user registration
• Root Cause: Database table creation was implicitly part of prune_system(). Fresh workspaces (where pruning is skipped) never called table creation
• Fix: Explicitly call create_db_and_tables() from cognee.infrastructure.databases.relational in the fresh workspace initialization path

Lance Namespace Import Error

• Bug: ModuleNotFoundError: No module named 'lance_namespace' prevented vector engine loading
• Root Cause: The lance-namespace==0.2.0 package on PyPI is broken/incomplete - missing the lance_namespace top-level module
• Fix: Pinned lance-namespace==0.0.21 in requirements.txt (verified working version)

Background Notification False Positives

• Bug: "Flowbaby processing failed" notifications appeared despite successful ingestion in logs
• Root Cause: ingest.py wrote status stubs to .cognee/background_ops/ (legacy path) while BackgroundOperationManager.ts looked for stubs in .flowbaby/background_ops/
• Fix: Updated ingest.py to write status stubs to .flowbaby/background_ops/

Technical Notes

• All fixes maintain backward compatibility with existing workspaces
• Legacy .cognee/background_ops stubs will be ignored (no functional impact)
• No migration required from v0.4.4

[0.4.4]

Fixed - Plan 033: Complete Environment Variable Hotfix

Hotfix - Completes the v0.4.3 environment variable fix that missed init.py, preventing "still initializing" hang and .cognee folder creation.

Root Cause

v0.4.3 correctly fixed ingest.py and retrieve.py to set SYSTEM_ROOT_DIRECTORY and DATA_ROOT_DIRECTORY before importing Cognee SDK. However, init.py was missed, causing:

1. Creation of orphan .cognee folder in workspace root
2. Missing .flowbaby/data directory
3. "Flowbaby is still initializing..." hang on first activation
4. Retrieval failures ("no relevant memories found")

Fix Applied

• Set SYSTEM_ROOT_DIRECTORY and DATA_ROOT_DIRECTORY environment variables before import cognee in init.py (lines 178-215)
• Changed legacy .cognee directory creation to .flowbaby for marker files (line 263)
• Conducted bridge-wide import audit - all 8 scripts now follow the "env vars before import" invariant
• Fixed additional non-compliant scripts: list_memories.py, validate_memories.py, recover_data.py

Bridge-Wide Invariant (Documented)

All bridge entrypoints must now follow this sequence:

1. Compute .flowbaby/system and .flowbaby/data paths
2. Create directories with Path.mkdir(parents=True, exist_ok=True)
3. Set os.environ['SYSTEM_ROOT_DIRECTORY'] and os.environ['DATA_ROOT_DIRECTORY']
4. Only then import cognee

This invariant is enforced because Cognee SDK uses pydantic-settings with @lru_cache, reading environment variables at import time.

Test Coverage Added

• test_sets_env_vars_before_cognee_import - verifies env var values
• test_creates_flowbaby_dirs_not_cognee_dirs - filesystem regression test
• test_env_vars_contain_flowbaby_path - path validation

[0.4.3]

Fixed - Hotfix: Cognee SDK Environment Variable Names

Critical Hotfix - Fixes storage directory misconfiguration causing FileNotFoundError: Storage directory does not exist: '.cognee_data'.

Root Cause

v0.4.2 used incorrect environment variable names (COGNEE_SYSTEM_ROOT_DIRECTORY, COGNEE_DATA_ROOT_DIRECTORY). The Cognee SDK's BaseConfig uses pydantic-settings which reads environment variables matching the exact field names—no COGNEE_ prefix.

Fix Applied

• Changed COGNEE_SYSTEM_ROOT_DIRECTORY → SYSTEM_ROOT_DIRECTORY in ingest.py and retrieve.py
• Changed COGNEE_DATA_ROOT_DIRECTORY → DATA_ROOT_DIRECTORY in ingest.py and retrieve.py

Impact

• Background cognify operations now correctly use workspace-local .flowbaby/ paths
• New data files appear in .flowbaby/data/ instead of ~/.cognee_data

[0.4.2]

Fixed - Plan 032: Activation and Background Processing Bugs

Maintenance Release - Resolves 6 bugs affecting extension activation, database generation, log rotation, and branding consistency.

Chat Agent Registration (#1)

• Bug: @flowbaby chat participant only appeared after extension fully initialized, requiring user to reload window
• Root Cause: registerFlowbabyParticipant() was called inside if (initialized) block, delaying UI registration
• Fix: Moved chat participant registration BEFORE flowbabyClient.initialize() with graceful degradation - handler shows "⏳ Flowbaby is still initializing..." message while backend completes setup

Cognee SDK Storage Pathing (#2)

• Bug: Database files created in ~/.cognee_data instead of workspace-local .flowbaby/data
• Root Cause: Cognee SDK may cache paths on import; environment variables were set after import in bridge scripts
• Fix:
  • Added explicit .flowbaby/data directory creation in init.py alongside .flowbaby/system
  • Restructured ingest.py and retrieve.py to set COGNEE_SYSTEM_ROOT_DIRECTORY and COGNEE_DATA_ROOT_DIRECTORY environment variables BEFORE import cognee
  • Safety check workspace_has_data() already existed and continues to prevent accidental data loss

Log Rotation Fix (#3)

• Bug: Python RotatingFileHandler writing to .log.1 instead of .log after rotation
• Root Cause: TypeScript BackgroundOperationManager was opening log file descriptor and passing it to spawned Python process; TypeScript holding the fd prevented Python from rotating properly
• Fix: Removed logFd file descriptor passing in spawnCognifyProcess(). Python bridge_logger.py now handles all log rotation independently using its own RotatingFileHandler

Branding Consistency (#4)

• Bug: Toast notifications showed "⚠️ Cognify failed" and "✅ Cognify finished" instead of Flowbaby branding
• Fix: Updated all user-facing toast messages in BackgroundOperationManager.ts:
  • '✅ Cognify finished' → '✅ Flowbaby processing finished'
  • '⚠️ Cognify failed' → '⚠️ Flowbaby processing failed'
  • Updated corresponding internal log messages for consistency

Technical Notes

• Plan 032 follows architectural guidance from Plan 031 and maintains backward compatibility
• No migration required from v0.4.1; users can update seamlessly
• Log file location unchanged: .flowbaby/logs/flowbaby.log

[0.4.1]

Fixed - Plan 031: Background API Key and Logging Consolidation

Critical Hotfix - Resolves v0.4.0 regressions affecting background ingestion and logging.

Background Ingestion API Key Fix

• Root Cause: BackgroundOperationManager.spawnCognifyProcess() was not injecting LLM_API_KEY from SecretStorage into the spawned process environment, causing MISSING_API_KEY errors for users who relied on the "Flowbaby: Set API Key" command.
• Fix: Added resolveApiKey() and getLLMEnvironment() methods to BackgroundOperationManager, mirroring the working implementation in FlowbabyClient. These methods now properly inject API keys from:
  1. Workspace .env file (highest priority)
  2. VS Code SecretStorage (flowbaby.llmApiKey)
  3. System environment variable (LLM_API_KEY)
• Scope: Applies to spawnCognifyProcess() (background cognify) and runPythonJson() (add-only retry).

Logging Consolidation

• Unified Log File: All Python bridge logs now write to .flowbaby/logs/flowbaby.log (was split between bridge.log and ingest.log).
• Updated Logger: bridge_logger.py updated to use .flowbaby/logs/flowbaby.log path and flowbaby.* logger names.
• View Logs Action: "View Logs" notification action now opens the consolidated log file.

Test Infrastructure

• rdflib Dependency: Added early detection for missing rdflib library in test environment with actionable error message in conftest.py.
• Updated Tests: test_logging_overhaul.py updated to use new log paths.

Documentation

• README.md: Fixed Windows path example to use .flowbaby instead of .cognee.
• extension.ts: Fixed incorrect pip install instruction - changed "pip install flowbaby" to "pip install cognee" (flowbaby is the VS Code extension name, not a Python package).

[0.4.0]

⚠️ BREAKING CHANGES

This release rebrands the extension from "RecallFlow Chat Memory" to "Flowbaby" with significant breaking changes:

Branding Changes

• Extension renamed: cognee-chat-memory → flowbaby
• Publisher changed: flowbaby → flowbaby
• Display name: RecallFlow Chat Memory → Flowbaby

Command Changes

• All cognee.* commands renamed to Flowbaby.*
• Example: cognee.captureMessage → Flowbaby.captureMessage

Settings Changes

• All cogneeMemory.* settings renamed to Flowbaby.*
• Example: cogneeMemory.enabled → Flowbaby.enabled
• Action Required: Update your settings.json if you have custom configurations

Chat Participant Changes

• @flowbaby-memory → @flowbaby

Tool Changes (for Agent Developers)

• flowbaby_storeMemory → flowbaby_storeMemory
• flowbaby_retrieveMemory → flowbaby_retrieveMemory
• #flowbabyStoreSummary → #flowbabyStoreSummary
• Action Required: Update your .agent.md files

Workspace Storage Changes

• All extension data now stored under .flowbaby/ (was .cognee/, .cognee_data/, .cognee_system/)
• Action Required: Re-set your API key via Flowbaby: Set API Key command
• Note: Previous workspace memories are not migrated

Logging Consolidation

• Single log file: .flowbaby/logs/flowbaby.log
• Removed redundant ingest.log

Migration Steps

1. Uninstall the old extension
2. Install Flowbaby v0.4.0
3. Run Flowbaby: Set API Key to configure your LLM API key
4. Update any custom keybindings from cognee.* to Flowbaby.*
5. Update any agent definitions to use new tool names

[0.3.17]

Fixed - Plan 028: Post-v0.3.16 Bug Fixes and Developer Experience

• Duplicate Output Channels: Fixed issue where multiple "Flowbaby Memory" output channels were created. Implemented singleton pattern in outputChannels.ts.
• Status Bar Unclickable: Fixed status bar command registration timing issue. Commands are now registered early in activation with try/catch error handling.

Added

• Debug Logging Channel: New opt-in debug output channel for troubleshooting. Enable via Flowbaby.debugLogging setting and use "Flowbaby: Show Debug Logs" command.
• Isolated Python Environment: venv moved from .venv to .flowbaby/venv to prevent conflicts with project virtual environments and Python language servers (e.g., Pylance/Jedi).
• venv Conflict Detection: When an existing .venv is detected, users are offered a choice between the isolated .flowbaby/venv (recommended) or using the existing .venv (advanced).
• Global API Key via SecretStorage: New "Flowbaby: Set API Key" command stores API keys securely via VS Code's SecretStorage API. Keys are shared across all workspaces.
• API Key Priority: Resolution order is now workspace .env > SecretStorage (global) > system environment variable.
• Global LLM Configuration: New settings Flowbaby.llm.provider, Flowbaby.llm.model, and Flowbaby.llm.endpoint for configuring LLM providers without per-workspace .env files.
• Clear API Key Command: New "Flowbaby: Clear API Key" command to remove stored SecretStorage key.

Changed

• Python Interpreter Detection: Priority order updated to: explicit setting > .flowbaby/venv > .venv (legacy) > system python3.
• Documentation: Updated README and SETUP docs to reflect new setup flow, removed global pip install requirement, documented new features and troubleshooting steps.

[0.3.16]

Fixed - Plan 027: Migration Marker Data Loss Bug

• Critical P0 Bug Fix: Fixed a bug that caused 96% data loss when the cognee Python package was reinstalled. The migration marker was being written to the volatile venv location instead of the workspace-local .flowbaby/system/ directory.
• Root Cause: get_relational_config() was called BEFORE cognee.config.system_root_directory(), returning the venv path instead of the configured workspace path.
• Reordered Initialization: Workspace directories are now configured FIRST, and the marker path is derived directly from the workspace path (never querying SDK config).
• Safety Check: Added workspace_has_data() function that refuses to prune if existing LanceDB or Kuzu data is detected, preventing accidental data loss.
• Prominent Warnings: Added visual warning separators (====) before any prune operation to make the action obvious in logs.
• Data Integrity Health Check: Added get_data_integrity_status() function to compare SQLite vs LanceDB counts for consistency monitoring.

Testing

• Added extension/bridge/tests/test_init_migration.py with 15 test cases covering marker location, safety checks, and integrity status.

[0.3.15]

Fixed - Plan 026: Path Canonicalization and Retrieval Scoring

• Path Canonicalization: All 5 Python bridge scripts (ingest.py, retrieve.py, init.py, list_memories.py, validate_memories.py) now strictly canonicalize workspace paths at entry. This fixes IngestionError 415 when opening workspaces with relative paths (e.g., code .).
• UI Initialization: Fixed a "chicken-and-egg" bug where the Status Bar and Setup Service were only created after successful initialization. They now appear immediately, ensuring the "Setup Required" indicator is visible in fresh workspaces.
• Synthesized Answer Scoring: Fixed confusing display of synthesized graph answers (which have a raw score of 0.00). The bridge now detects this sentinel, assigns a high ranking score, and attaches a confidenceLabel="synthesized_high" metadata field.
• Qualitative UI Labels: The UI now displays "High relevance (synthesized)" for graph completions instead of overloading the numeric score or showing 0.00.

[0.3.14]

Added - Plan 025: Simplified Python Environment Setup

• Managed Python Environment: Flowbaby now automatically creates and manages a dedicated .venv virtual environment in your workspace, eliminating manual setup steps.
• One-Click Setup: New "Initialize Workspace" command handles Python version checks (3.8+), virtual environment creation, and dependency installation in a single click.
• Status Bar Indicator: Added a status bar item showing environment health (Ready, Setup Required, Refreshing, Error) with quick access to repair actions.
• Dependency Management: New "Refresh Dependencies" command safely updates the bridge environment, coordinating with background operations to prevent conflicts.
• Smart Precedence: Automatically detects existing .venv or configured Python paths, but prioritizes the managed environment for reliability.
• Enhanced Walkthrough: The "Get Started" walkthrough now guides users through the automated setup process with context-aware completion steps.
• Robust Metadata: Tracks environment state and version compatibility in .flowbaby/bridge-env.json to ensure stability across updates.

[0.3.13]

Added - Plan 024: Configurable Search Parameters

• Configurable Search Depth (searchTopK): Added new Flowbaby.searchTopK setting (default 10, range 1–100) to control how many candidates the Flowbaby search engine returns before ranking.
• Higher Context Token Budget: Increased Flowbaby.maxContextTokens default from 2000 to 32000 and maximum to 100000 to better support large-context models.
• Bridge Safeguards: retrieve.py now clamps max_tokens into the [100, 100000] window, normalizes top_k so it is never lower than max_results, and hard-clamps top_k to 100 to prevent runaway latency.
• Normalization & Logging: When a user-provided top_k is raised to meet max_results or clamped to the ceiling, the bridge logs a structured warning with both requested and effective values for easier debugging.
• Client Logging & Truncation: CogneeClient logs the configured searchTopK value on each retrieval and passes it through to the bridge. Output-channel payloads remain truncated to keep the VS Code UI responsive.

[0.3.12]

Fixed - Plan 022: Retrieval Filtering and Truncation

• Retrieval Filtering: Fixed aggressive filtering of "no_relevant_context" responses. The bridge now uses an exact match check instead of a substring match, preventing valid answers that quote the sentinel value from being discarded.
• Client Truncation: Increased CogneeClient stdout buffer from 2KB to 1MB. This fixes JSON parse errors when retrieval results exceed the default buffer size.
• Error Logging: Improved error reporting to capture and log both stdout and stderr when JSON parsing fails, aiding in debugging bridge issues.
• Telemetry: Added filtered_count to retrieval telemetry to track how many items are filtered by the bridge.

[0.3.11]

Fixed - Plan 021: Retrieval Display for Synthesized Answers

• Retrieval Display Bug: Fixed a regression where synthesized answers (score 0.0) were silently filtered out by the client-side safety check.
• Sentinel-Aware Filter: Updated CogneeContextProvider to explicitly allow results with score === 0.0 (sentinel for synthesized answers) while maintaining strict filtering for low-confidence noise (e.g., 0.001 - 0.009).
• Logging: Added debug logging to indicate when synthesized answers are included in retrieval results.

[0.3.10]

Fixed - Plan 023: Retrieval Scoring Hotfix

• Retrieval Regression: Fixed an issue where valid GRAPH_COMPLETION answers (synthesized strings) were discarded because they lacked a semantic score and hit the strict score <= 0.01 filter introduced in v0.3.9.
• Logic Update: retrieve.py now bypasses the strict score filter for synthesized answers (which default to score 0.0) while maintaining the filter for NO_RELEVANT_CONTEXT responses.

[0.3.9]

Added - Plan 021: Logging Infrastructure Overhaul

Structured Bridge Logging - Python bridge scripts now use a shared logging utility that emits structured JSON-lines to both the VS Code Output Channel and a persistent workspace log file (.flowbaby/logs/bridge.log).

• Rotating Log File: bridge_logger.py writes untruncated diagnostics with 5MB rotation (3 backups) so retrieval and ingestion history is never lost.
• Detailed Retrieval Telemetry: retrieve.py logs candidate lists, semantic scores, recency/status multipliers, filtering reasons, and final selections for each query.
• Ingestion Metrics: ingest.py reports character counts and stage durations, making performance bottlenecks obvious.
• Consistent Formatting: All bridge scripts (init.py, validate_memories.py, list_memories.py) now produce structured JSON logs for easier parsing and UI rendering.
• VS Code Output Contract: Stderr handler keeps Output Channel entries lightweight (~2KB) while retaining full fidelity in file logs.

Testing

• Added extension/bridge/tests/test_logging_overhaul.py covering log file creation, ingestion metrics, retrieval scoring telemetry, and stderr formatting.

[0.3.8]

Added - Plan 021: Memory Visibility and Trust

Zero-Hallucination Retrieval - Implemented strict filtering to prevent irrelevant memories from polluting context:

• Strict Score Threshold: Retrieval results with distance scores > 0.01 are now filtered out by default.
• Explicit "No Context" Signal: Bridge now returns NO_RELEVANT_CONTEXT signal when no memories meet the threshold, preventing hallucinated "matches".
• Validation Tools: New commands for inspecting memory state:
  • Flowbaby: Validate Memories - Runs integrity checks on stored memories.
  • Flowbaby: List Memories - Dumps raw memory content for debugging.

Fixed:

• Integration Test Stability: Fixed path resolution errors in integration tests by correctly injecting extensionPath into the mock context. All 149 tests now pass.
• Branding Cleanup: Removed "formerly Cognee" references from extension description and UI strings for a cleaner "Flowbaby" identity.

[0.3.7]

Fixed - Plan 019: Flowbaby Rebranding Gaps

• Publisher Identity: Marketplace publisher ID now set to flowbaby with author updated to "Flowbaby Team" so the extension no longer appears under the legacy Cognee brand.
• Language Model Tools: Renamed tool identifiers and reference names to flowbaby_storeMemory / flowbaby_retrieveMemory so Configure Tools, chat autocomplete, and .agent.md definitions all reflect the Flowbaby brand.
• Documentation Alignment: README.md and AGENT_INTEGRATION.md now show the correct tool names and usage examples, preventing confusion for agent developers.
• Test Coverage: Integration tests updated to ensure the new tool identifiers register correctly with GitHub Copilot chat and that UI instructions stay in sync.

[0.3.6]

Added - Plan 019: Retrieval Fabrication and Silent Failure Fix

Rebranding to Flowbaby - The extension has been renamed to "Flowbaby" to better reflect its purpose as a memory and recall layer for AI agents.

• User-Facing Changes:
  • Extension name: "Flowbaby Chat Memory"
  • Commands: Flowbaby: Capture to Memory, Flowbaby: Toggle Memory, etc.
  • Output Channel: "Flowbaby Agent Activity"
  • Chat Participant: @flowbaby-memory
  • Tools: "Store Memory in Flowbaby", "Retrieve Flowbaby Memory"
  • Tool Renaming: flowbaby_storeMemory → flowbaby_storeMemory, flowbaby_retrieveMemory → flowbaby_retrieveMemory
• Backward Compatibility:
  • Internal configuration keys (Flowbaby.*) remain unchanged
  • File paths (.flowbaby/) remain unchanged
  • Python package (cognee) remains unchanged

Fixes & Improvements:

• Retrieval Fabrication Fix: Changed default semantic score for unscored results from 0.7 to 0.0. This prevents empty or irrelevant results from appearing with artificially high confidence scores.
• Silent Failure Detection:
  • Implemented log rotation for background processes (.flowbaby/logs/ingest.log)
  • Redirected stdout/stderr of background processes to log files to capture crashes
  • Added "zombie" detection for processes that exit without writing a status stub
• Proactive Prevention: Enforced 100k character limit on ingestion payloads to prevent bridge crashes and memory exhaustion.
• UX Enhancements: Added timestamps to Flowbaby: Show Background Operations quick pick items for better visibility.

[0.3.5]

Added - Plan 018: Metadata Infrastructure and Ranking

Intelligent Ranking & Filtering - Retrieval now prioritizes memories based on recency and status:

• Recency-Aware Ranking: New exponential decay algorithm prioritizes fresh memories. Configurable via Flowbaby.ranking.halfLifeDays (default: 7 days).
• Status Filtering: Memories can be marked as Active, Superseded, or DecisionRecord.
  • Superseded memories are hidden by default to reduce noise.
  • DecisionRecord memories receive a relevance boost.
• Metadata-Rich Storage: Summaries now embed structured metadata (topicId, status, timestamps) directly in the text, enabling robust tracking even without graph metadata support.

Technical Implementation:

• Enriched Text Fallback: Implemented "Enriched Text" pattern to store metadata within summary text, bypassing Cognee 0.3.4 DataPoint limitations.
• Migration Script: New migrate_summaries.py script automatically upgrades legacy memories to the new schema.
• Safe Migration: Uses file locking (.flowbaby/maintenance.lock) to pause background operations during migration.
• Updated Tool Definitions: flowbaby_storeMemory and flowbaby_retrieveMemory tools updated with clearer descriptions and privacy guarantees ("Data stays in this workspace").

Fixed

• Ranking Quality: Addressed issue where old, less relevant memories cluttered retrieval results.
• Status Visibility: Superseded plans and decisions no longer confuse agents.

[0.3.4]

Release

This release packages the async cognify() optimization (Plan 017) with no functional changes from v0.3.3. Version bumped to v0.3.4 for clean release tracking.

All features, improvements, and technical details remain as documented in v0.3.3 below. This is a packaging-only release to formalize the deployment.

[0.3.3]

Added - Plan 017: Async cognify() Optimization

Universal Async Memory Ingestion - ALL ingestion flows (agent tools, manual capture, headless commands) now return in <10 seconds:

• Staged Messaging: Every ingestion surface shows: "Memory staged – processing will finish in ~1–2 minutes. You'll get a notification when it's done."
• Background Processing: Knowledge graph construction (cognee.cognify()) runs in detached subprocess while agents continue working
• Completion Notifications:
  • Success (info): "✅ Cognify finished" with workspace name, summary digest, elapsed time, entity count, "View Status" action
  • Failure (warning): "⚠️ Cognify failed" with workspace name, summary digest, remediation guidance, "Retry"/"View Logs" actions
• Independent Throttling: Success and failure notifications throttled separately (≤1 per 5 min per workspace per outcome type)
• Background Status Command: New cognee.backgroundStatus command shows all in-flight/completed operations with quick-pick UI

Technical Implementation:

• Split ingest.py into 3 modes: --mode sync (diagnostic), --mode add-only (fast staging <10s), --mode cognify-only (background graph construction)
• BackgroundOperationManager service with:
  • Dual-ledger persistence (.flowbaby/background_ops.json + VS Code globalState)
  • Concurrency limits: max 2 concurrent + FIFO queue of 3 pending operations
  • Detached subprocess spawning with PID tracking
  • Activation reconciliation (reattach live PIDs, mark stale entries unknown)
  • Deactivation cleanup (SIGTERM with 5s grace + SIGKILL)
  • Atomic status stub writes prevent corruption on crashes
• Updated client methods: ingestSummaryAsync() and ingestAsync() for summary and conversation ingestion
• Comprehensive bridge test suite (15+ tests) covering all modes, error handling, backward compatibility

Performance Impact:

• Agent blocking time: 73s → <10s (86% reduction)
• Manual capture blocking time: 73s → <10s (86% reduction)
• Background processing: 60-90s (unchanged, runs asynchronously)
• Result: Agents can store 3 memories in 30s instead of 219s

User Experience Changes:

• Manual capture (Ctrl+Alt+C) now shows staged toast instead of blocking
• Agent tools (#cogneeStoreSummary) return immediately with operationId
• All users receive completion notification when background processing finishes
• Output channel shows full audit trail: [BACKGROUND] markers for start/success, [ERROR] for failures

Backward Compatibility:

• Sync mode (--mode sync) preserves legacy behavior for diagnostic/test use
• Conversation mode (positional args) continues to work with sync mode
• All existing tests pass unchanged

Fixed - Plan 016.1: Tool Lifecycle and Bridge Timeouts Hotfix

Bug Fixes:

• Tool Lifecycle UI Desync: Fixed issue where Configure Tools dialog showed stale enablement state when toggling tools on/off. Tools now register unconditionally at extension activation; VS Code's Configure Tools UI is the sole authorization mechanism (no redundant workspace setting).
• Bridge Timeout Opacity: Added comprehensive diagnostic logging to Python bridge scripts (ingest.py, retrieve.py) with [PROGRESS], [WARNING], and [ERROR] markers. Users now see step-by-step progress in Output channel when bridge operations are slow or fail.
• Redundant Authorization: Removed Flowbaby.agentAccess.* workspace settings entirely. Simplified authorization model: users enable/disable tools via Configure Tools UI only.

Breaking Changes:

• Removed settings: Flowbaby.agentAccess.enabled, Flowbaby.agentAccess.maxResultsDefault, Flowbaby.agentAccess.maxTokensDefault, Flowbaby.agentAccess.maxConcurrentRequests, Flowbaby.agentAccess.rateLimitPerMinute
• Status bar "Cognee Agent Access" indicator removed (Configure Tools dialog provides feedback)
• Error code ACCESS_DISABLED no longer returned (tools always registered)

Technical Details:

• Bridge scripts emit structured error payloads with error codes: LLM_API_ERROR, COGNEE_SDK_ERROR, PYTHON_ENV_ERROR, ONTOLOGY_LOAD_ERROR
• TypeScript client (cogneeClient.ts) parses stderr for diagnostic markers and surfaces at appropriate log levels (INFO/ERROR/WARN)
• Fixed sys module import shadowing bug in retrieve.py
• Test suite updated to reflect new authorization model

User Impact:

• Simpler authorization: single source of truth (Configure Tools)
• Better debugging: diagnostic logs visible in Output > "Flowbaby" channel
• Faster resolution of bridge issues: progress markers identify exactly where operations block

[0.3.2]

Fixed - Plan 016.1: Tool Lifecycle and Bridge Timeouts Hotfix

Bug Fixes:

• Tool Lifecycle UI Desync: Fixed issue where Configure Tools dialog showed stale enablement state when toggling tools on/off. Tools now register unconditionally at extension activation; VS Code's Configure Tools UI is the sole authorization mechanism (no redundant workspace setting).
• Bridge Timeout Opacity: Added comprehensive diagnostic logging to Python bridge scripts (ingest.py, retrieve.py) with [PROGRESS], [WARNING], and [ERROR] markers. Users now see step-by-step progress in Output channel when bridge operations are slow or fail.
• Redundant Authorization: Removed Flowbaby.agentAccess.* workspace settings entirely. Simplified authorization model: users enable/disable tools via Configure Tools UI only.

Breaking Changes:

• Removed settings: Flowbaby.agentAccess.enabled, Flowbaby.agentAccess.maxResultsDefault, Flowbaby.agentAccess.maxTokensDefault, Flowbaby.agentAccess.maxConcurrentRequests, Flowbaby.agentAccess.rateLimitPerMinute
• Status bar "Cognee Agent Access" indicator removed (Configure Tools dialog provides feedback)
• Error code ACCESS_DISABLED no longer returned (tools always registered)

Technical Details:

• Bridge scripts emit structured error payloads with error codes: LLM_API_ERROR, COGNEE_SDK_ERROR, PYTHON_ENV_ERROR, ONTOLOGY_LOAD_ERROR
• TypeScript client (cogneeClient.ts) parses stderr for diagnostic markers and surfaces at appropriate log levels (INFO/ERROR/WARN)
• Fixed sys module import shadowing bug in retrieve.py
• Test suite updated to reflect new authorization model

User Impact:

• Simpler authorization: single source of truth (Configure Tools)
• Better debugging: diagnostic logs visible in Output > "Flowbaby" channel
• Faster resolution of bridge issues: progress markers identify exactly where operations block

[0.3.1]

Added - Plan 015: Agent Ingestion Command

• Agent Ingestion API: Flowbaby.ingestForAgent command enables GitHub Copilot agents and third-party extensions to store structured summaries in Cognee
  • TypeScript schema validation with detailed error messages
  • Workspace-global access model with prominent privacy warnings
  • Structured JSON request/response with camelCase field naming
  • Auto-generation of missing IDs (topicId, timestamps)
  • Fast-fail validation before bridge invocation
• Agent Access Control: New Flowbaby.agentAccess.enabled setting (default: false)
  • Workspace-global trust model (all extensions granted access when enabled)
  • Prominent warning in settings UI about privacy implications
  • Access enforcement at command handler level
• Audit Logging: Comprehensive logging for all agent ingestion attempts
  • Real-time logs in Output > Cognee Agent Activity channel
  • Structured JSON audit log at .flowbaby/agent_audit.log
  • Privacy-preserving topic digests (8-char SHA-256 hash)
  • Tracks timestamp, agent name, result, error codes, duration
• Agent Integration Documentation: Complete API guide at extension/AGENT_INTEGRATION.md
  • TypeScript examples (minimal and full payloads)
  • Error handling patterns with error codes
  • Security model explanation
  • Best practices for topic ID generation and ingestion timing
  • Troubleshooting guide
• Bridge Contract Documentation: extension/bridge/INGEST_CONTRACT.md defines stable ingestion schema
  • JSON payload specification (required/optional fields)
  • Error codes reference (ACCESS_DISABLED, INVALID_PAYLOAD, etc.)
  • Example payloads (minimal, full, invalid)
  • Performance characteristics
  • Testing instructions
• Test Agent Extension: Reference implementation at test-agent/
  • Validates all ingestion scenarios (valid, invalid, access control)
  • Automated test suite with pass/fail reporting
  • Can be used as template for custom agent development

Added - Plan 016: Agent Retrieval and UI-Visible Extension Tools

• Agent Retrieval API: Flowbaby.retrieveForAgent command enables agents to query Cognee knowledge graph
  • Structured JSON request/response with CogneeContextRequest/CogneeContextResponse types
  • Returns metadata-rich entries (topic, topicId, planId, score, decisions, timestamps)
  • Concurrency limiting (max 2 in-flight requests, configurable up to 5)
  • Rate limiting (max 10 requests/minute, configurable up to 30)
  • Graceful degradation for legacy memories (null metadata fields)
• CogneeContextProvider Service: Centralized retrieval infrastructure with architectural guardrails
  • FIFO request queueing with concurrency enforcement
  • Per-minute rate limiting with sliding window
  • Structured error responses (ACCESS_DISABLED, RATE_LIMIT_EXCEEDED, QUEUE_FULL, BRIDGE_TIMEOUT, INVALID_REQUEST)
  • Settings clamping with safe upper bounds (prevents misconfiguration)
• UI-Visible Language Model Tools: Both tools appear in VS Code's "Configure Tools" dialog
  • flowbaby_storeMemory (#cogneeStoreSummary) - Store conversation summaries
  • flowbaby_retrieveMemory (#cogneeRetrieveMemory) - Retrieve relevant memories
  • Tools support # autocomplete in chat and .agent.md front-matter references
  • Atomic lifecycle: both tools register/unregister together when agentAccess.enabled toggles
  • Icon support for visual identity ($(database) and $(search))
• Custom Agent Integration: Full support for custom .agent.md files
  • Tools reference name format: tools: ['cogneeStoreSummary', 'cogneeRetrieveMemory']
  • Confirmation messages for transparency (optional, depends on user trust settings)
  • Retrieve tool returns BOTH narrative markdown AND verbatim JSON payload
• @flowbaby Participant Refactor: Now uses shared CogneeContextProvider
  • Consistent retrieval behavior across participant and tools
  • Leverages centralized concurrency/rate limiting
  • Enhanced metadata display (topicId, planId, score when available)
• Transparency Indicators: All agent activity is auditable
  • Output channel logs every retrieval/ingestion with timestamps, query hashes, result counts
  • Status bar indicator shows "Cognee Agent Access: Enabled" with spinner during operations
  • Click behavior opens Output channel for inspection
• Enhanced Documentation:
  • AGENT_INTEGRATION.md extended with retrieval examples, tool integration guide, error code reference
  • README.md includes "Using Cognee Tools with Custom Agents" section with .agent.md examples
  • Complete TypeScript interfaces in types/agentIntegration.ts
• Agent Access Settings: Additional configuration for retrieval behavior
  • Flowbaby.agentAccess.maxResultsDefault (default: 5)
  • Flowbaby.agentAccess.maxTokensDefault (default: 4000)
  • Flowbaby.agentAccess.maxConcurrentRequests (default: 2, max: 5)
  • Flowbaby.agentAccess.rateLimitPerMinute (default: 10, max: 30)

Changed

• Minimum VS Code Version: Requires VS Code 1.106+ for canBeReferencedInPrompt/toolReferenceName support
• Extension Activation: CogneeContextProvider initialization happens after CogneeClient setup
• Participant Behavior: @flowbaby now routes through CogneeContextProvider (no breaking changes to user experience)

Added - Test Coverage Enhancements (from previous release)

Added - Test Coverage Enhancements

• Summary Workflow Integration Tests: 14 new automated tests in summaryWorkflow.integration.test.ts
  • 7 workflow tests validate complete "summarize → generate → store" flow
  • 7 snapshot tests guard against README/template drift
  • Tests cover trigger detection, no-history errors, cancellation, large conversations
• Template Consistency Validation: Round-trip preservation tests ensure format → parse cycle stability
• Mixed-Mode Test Coverage: Validates enriched vs legacy memory handling in integration context
• Section Heading Stability: Tests enforce that headings match DATAPOINT_SCHEMA.md exactly

Fixed

• Manual Test Isolation: Added @pytest.mark.manual decorator to manual scripts to prevent CI failures
• Pytest Configuration: Registered manual marker in pytest.ini for proper test filtering
• QA Documentation: Updated with environment-specific test execution guidance

Improved

• Test Suite Reliability: Default pytest runs now succeed (37 passed, 1 skipped)
• VS Code Test Coverage: 77 total tests passing (up from 63), no regressions
• Documentation Accuracy: Snapshot tests catch template/README divergence automatically

[0.3.0]

Added - Plan 014: Structured Conversation Summaries

• Conversation Summary Generation: Create structured summaries via @flowbaby summarize this conversation
  • Adjustable turn count scope (default: last 15 turns)
  • Interactive scope adjustment before generation
  • LLM-powered summary extraction with Plan 014 schema
  • User confirmation required before storage
• Enriched Summary Schema: Structured format with Topic, Context, Decisions, Rationale, Open Questions, Next Steps, References, and Time Scope
• Metadata-Rich Retrieval: Summaries include status badges (📋 Status, 📅 Created, 🏷️ Plan ID), structured content sections, and temporal awareness
• Mixed-Mode Support: Seamless handling of both enriched summaries (Plan 014+) and legacy raw-text memories
• In-Chat Help: Type @flowbaby help or invoke with no query to see usage guide
• Bridge Contract Documentation: DATAPOINT_SCHEMA.md and RETRIEVE_CONTRACT.md define stable schemas for downstream consumers

Changed

• Retrieval Display: Now shows structured metadata badges and organized content sections for enriched summaries
• Python Bridge: retrieve.py returns structured JSON with parsed metadata instead of raw text
• CogneeClient: retrieve() method returns RetrievalResult[] with typed fields (topicId, status, createdAt, decisions, etc.)

Improved

• Summary Quality: Structured format improves retrieval precision by surfacing key decisions and rationale
• Temporal Context: Created timestamps enable recency-aware reasoning (foundation for Plan 015 ranking)
• Status Tracking: Active/Superseded status field enables decision lifecycle tracking (foundation for Plan 015 compaction)
• User Control: Explicit summary creation and confirmation flow ensures intentional memory capture
• Backward Compatibility: All legacy raw-text memories remain accessible; retrieval transparently handles both formats

Technical

• Enriched Text Storage: Summaries stored as markdown with embedded metadata per §4.4.1 (Cognee 0.3.4 constraint)
• Template Versioning: <!-- Template: v1.0 --> tag enables future format migrations
• Regex-Based Parsing: Deterministic section heading patterns for metadata extraction
• TypeScript Tests: 59/59 passing with updated RetrievalResult mocks
• Python Tests: 16/16 passing with enriched text formatting and contract validation

[0.2.3]

Fixed

• Display truncation in chat participant - users can now see full retrieved memory content (up to 2000 chars)
• Query logging truncation increased from 50 to 200+ characters for better debugging
• Ingestion timeout increased from 30s to 120s to reduce false-positive failures
• Error messages now distinguish timeout (may still succeed) from true ingestion failure
• Added comprehensive ingestion performance metrics (Python-side duration, step-level timing)
• Added process exit vs timeout logging for diagnosing bridge-level latency

Improved

• Transparent memory display with character count indicators aligns with discoverability goals (Epic 0.2.2.3)
• Clear truncation indication when memories exceed 2000 chars (shows "showing X of Y chars")
• Query length now shown in debug logs when queries exceed 200 chars
• Ingestion timeout errors now explain background processing may still succeed
• Extension logs include both Node-side and Python-side duration metrics for correlation
• Step-level metrics (load_env, init_cognee, config_llm, dataset_ontology, add, cognify) enable bottleneck identification

[0.2.2]

Fixed

• Display truncation in chat participant - users can now see full retrieved memory content (up to 2000 chars)
• Query logging truncation increased from 50 to 200+ characters for better debugging

Improved

• Transparent memory display with character count indicators aligns with discoverability goals (Epic 0.2.2.3)
• Clear truncation indication when memories exceed 2000 chars (shows "showing X of Y chars")
• Query length now shown in debug logs when queries exceed 200 chars

[0.2.1]

Breaking Changes

• Environment: LLM_API_KEY is now required in workspace .env. Previously supported OPENAI_API_KEY fallback removed to align with Cognee 0.4.0 conventions.

Fixed

• Ingestion failures and 30s timeouts caused by Cognee using site-packages storage instead of workspace-local directories
• Silent parameter fallback code that masked configuration errors
• File-not-found errors during ingestion due to cross-context storage mismatches

Changed

• All Cognee system and data directories now scoped to workspace (.flowbaby/system/, .flowbaby/data/)
• Removed fallback parameter retries; signature mismatches now surface as clear errors
• Enhanced error logging with structured details (exception type, parameters, context)

[0.2.0]

Added

• Debug Configuration: VS Code launch.json for F5 Extension Development Host debugging
• Visible Build Logging: esbuild now shows compilation progress and errors (logLevel: 'info')
• Developer Documentation: Enhanced SETUP.md and README.md with debugging workflow and troubleshooting guidance

Fixed

• Extension Activation Failure: Missing .vscode/launch.json prevented Extension Development Host from loading the extension
• Silent Build Errors: Build failures were hidden with logLevel: 'silent', now visible for rapid debugging

Implementation 008 Features (from v0.1.0 foundation)

Added

• Keyboard Shortcut Capture (Ctrl+Alt+C / Cmd+Alt+C): Selective conversation capture via keyboard shortcut + input box workflow
• @flowbaby Chat Participant: Explicit memory-augmented chat participant for context retrieval and informed responses
• Command Palette Capture: Alternative capture method via "Cognee: Capture to Memory" command
• Toggle Memory Command: Quick on/off toggle via "Cognee: Toggle Memory" command
• Clear Memory Command: Delete workspace memory via "Cognee: Clear Workspace Memory" command (with confirmation)
• OWL/Turtle Ontology: Chat-specific ontology file (ontology.ttl) with 8 classes and 12 object properties for grounded entity extraction
• Conversational Ingestion Format: Simplified natural prose format for better LLM extraction quality
• Step 6 Feedback Loop (Experimental): Optional automatic capture of @flowbaby conversations (disabled by default due to Cognee 0.4.0 bug)
• Graceful Degradation: Retrieval failures show warning but participant continues without context
• Configuration Setting: Flowbaby.autoIngestConversations for experimental feedback loop control

Changed

• User Workflow: Shifted from automatic global capture to selective, user-controlled keyboard shortcut capture
• Participant Model: @flowbaby requires explicit invocation; no passive injection into other participants
• Ontology Integration: Updated ingest.py to use ontology_file_path parameter with RDFLib validation and graceful fallback
• Ingestion Format: Changed from bracketed metadata format to conversational prose for improved extraction

Improved

• Automatic Python Interpreter Detection: Extension now auto-detects workspace .venv virtual environment, eliminating need for manual Flowbaby.pythonPath configuration in most cases
• Enhanced Error Messages: Python errors (missing packages, API key issues) now visible in Output Channel with actionable troubleshooting hints
• Workspace-Relative Execution: Bridge scripts run from workspace context for reliable path resolution
• Context Display: Retrieved memories formatted with clear markdown previews ("📚 Retrieved N memories")
• Performance Logging: Retrieval timing logged to Output Channel for monitoring (<1000ms P95 target)
• Error logs sanitize sensitive data (API keys, tokens) before display
• Configuration documentation clarifies when manual Python path setting is needed
• Structured error extraction from Python subprocess stdout
• Better diagnostics for missing dependencies and environment issues

Fixed

• Generic "exit code 1" errors replaced with specific failure reasons
• Missing cognee package now clearly diagnosed instead of silent failure
• Python interpreter mismatch no longer requires manual configuration for standard .venv setups
• Keyboard shortcut comment typo in code (Ctrl+Shift+M → Ctrl+Alt+C)

Known Issues

• Cognee 0.4.0 File Hashing Bug: Intermittent ingestion failures for repeated identical content affect Step 6 auto-ingestion; workaround via Flowbaby.autoIngestConversations=false (default)
• Manual Capture Workflow: Keyboard shortcut requires copy-paste; cannot extract message from chat UI directly (VS Code API limitation)
• Explicit Participant Invocation: Must type @flowbaby; cannot inject context into other participants (API limitation)

Technical Implementation

• 6-step participant flow: retrieval → format display → augment prompt → generate response → capture conversation (conditional)
• RDFLib ontology parsing with graceful fallback to no-ontology mode
• Asynchronous fire-and-forget ingestion (non-blocking)
• Comprehensive integration test suite (27 passing tests)
• VS Code Chat API v1.105+ compatibility

0.1.0

Added

• Initial release of Flowbaby extension
• Automatic capture of GitHub Copilot chat conversations
• Automatic context retrieval from Cognee memory before responses
• Workspace-isolated memory with separate knowledge graphs per workspace
• Chat-specific ontology defining User, Question, Answer, Topic, Concept, Problem, Solution, and Decision entities
• Hybrid graph-vector search combining relationship traversal with semantic similarity
• Recency and importance weighting for intelligent context relevance scoring
• Configurable settings:
  • Flowbaby.enabled - Toggle memory on/off
  • Flowbaby.maxContextResults - Maximum results to retrieve (1-10)
  • Flowbaby.maxContextTokens - Token budget for context (100-10000)
  • Flowbaby.recencyWeight - Prioritize recent conversations (0-1)
  • Flowbaby.importanceWeight - Prioritize marked conversations (0-1)
  • Flowbaby.pythonPath - Custom Python interpreter path
  • Flowbaby.logLevel - Debug verbosity (error/warn/info/debug)
• Comprehensive integration test suite with 6 test cases
• Python bridge scripts for Cognee library communication
• Output Channel logging for debugging and monitoring

Technical Implementation

• Uses Cognee 0.3.4 with dataset-based logical isolation
• SHA1-hashed workspace identifiers for stable and unique dataset naming
• Global marker file pattern prevents data loss during concurrent initialization
• Stateless ontology configuration applied per ingestion operation
• TypeScript extension with Python subprocess bridge architecture
• esbuild-based compilation for optimized bundle size

Known Limitations

• Requires workspace folder (doesn't work in single-file mode)
• Currently captures conversations through Chat Participant API
• Python and Cognee must be installed separately (not bundled)
• First conversation in new workspace has no context (memory starts empty)
• macOS and Linux tested; Windows support may require additional configuration