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.

[0.7.5] - 2026-01-26

Added

Evaluation Validity Hardening (Plan 113): KG-grounded validity guarantees for the benchmark harness ensuring trustworthy retrieval metrics:
- Multi-Positive Semantics: Benchmark explicitly uses filtered-style scoring where multiple correct answers per query are first-class—reports label metrics as relevance_semantics: filtered to distinguish from exact-match
- Label Shape Statistics: Score results now include positives-per-query distribution (min/max/mean/median/p95, percentage of 0/1/2+ positive queries) to detect unhealthy dataset characteristics
- Macro + Slice Reporting: Reports include both aggregate and macro-averaged metrics (per-query equal weighting) to prevent hub-dominance skew; slice breakdowns include query counts for context
- Split Discipline Enforcement: Dataset format supports train/validation/test splits with tooling that refuses invalid workflows—validate_split_discipline() raises SplitDisciplineError when test split is used for selection/tuning
- Relationship-Fidelity Queries: Golden dataset expanded with queries specifically testing relationship extraction (q013-q015), tagged for separate slice tracking
- Benchmark Documentation: Comprehensive README.md covering scorecard interpretation, macro vs aggregate guidance, positives-per-query guidance, and leakage prevention procedures

Changed

Golden Dataset Version: Bumped to 1.1.0 with new relationship-fidelity slice definition and three new multi-positive queries
ScoreResult Contract: Extended with relevance_semantics, label_shape_stats, and macro_metrics fields using stable field names for diffability
Topic Contract: Extended with optional split field (train/validation/test) for leakage discipline

[0.7.4] - 2026-01-24

Added

Daemon Startup & Recovery Hardening (Plan 116): Comprehensive improvements to daemon lifecycle reliability, diagnostics, and observability:
- Reliability Contract: Stable reason codes (DaemonUnavailableReason) for all daemon unavailability scenarios enabling programmatic error handling and telemetry aggregation
- Bounded Startup: All startup paths now complete within a bounded deadline (30s default) with deterministic cleanup—no more stuck "starting" states
- Recovery Loop: Automatic recovery attempts (3 max) with exponential backoff when daemon startup fails; degraded mode entered after budget exhaustion
- Daemon-Only Routing: Memory operations (retrieve, ingest, ingestSummary) now route exclusively through the daemon—no subprocess fallback eliminates reliability edge cases
- Background Cognify via Daemon: cognify() operations now use daemon-managed execution; subprocess auto-retry removed for simpler, more predictable behavior
- Diagnostics Command: New Flowbaby: Diagnose Daemon command provides comprehensive diagnostic report including current state, last failure details, recovery status, lock ownership, and remediation hints
- Normal vs Debug Observability: State transitions and reason codes logged in normal mode; stderr excerpts and detailed retry logs only in debug mode
- Startup Robustness Verification Matrix: Engineering test harnesses covering immediate exit, hang, protocol errors, stdio issues, and runtime crash scenarios

Changed

Error Messages: Daemon unavailability errors now include actionable guidance (e.g., "Run 'Flowbaby: Diagnose Daemon' for details")
Background Operation Manager: Removed AUTO_RETRY_COUNT, AUTO_RETRY_DELAY_MS, and forceSubprocess parameter—all cognify operations are daemon-only
Windows Subprocess Tests: Updated to reflect daemon-only routing (Plan 115 subprocess spawn tests replaced with daemon routing tests)
Windows Refresh Hardening & Diagnostics (Plan 115): Improved reliability and observability for dependency refresh operations on Windows:
- Correlation Fields: All daemon start/stop, lock acquisition/release, refresh, and preflight logs now include sessionId, extensionHostPid, and operationId for multi-window contention diagnosis
- Preflight Reason Codes: Preflight failures now emit a low-cardinality reasonCode (always-on) enabling triage without debug logs. Stderr excerpts logged only when debug logging is enabled.
- Windows Rename Hardening: Increased post-daemon-stop delay (300ms → 2000ms) and retry budget (maxAttempts=8, baseDelayMs=500, ~18s worst-case) to tolerate slow file-handle release and AV scanning
- Timing Context: Rename retry logs include timeSinceDaemonStop and daemon stop duration for race condition diagnosis
- Lock Contention Breadcrumbs: When lock acquisition fails, logs now include bounded snapshot of existing lock owner metadata (ownerPid, instanceId, lockAge, ownerPidAlive) to distinguish "another window owns the daemon" vs "stale lock"
Retrieval-First Benchmark Harness (Plan 112): Minimal, maintainable offline evaluation harness for Flowbaby's retrieval system:
- Deterministic Scorer: Multi-metric scorecard (Recall@K, Precision@K, MRR, MAP@K, nDCG@K) using ir_measures and pytrec_eval for reliable, offline-only evaluation.
- Golden Dataset: Initial 12-topic dataset covering mandatory risk slices (template-heavy, edge-sparse, technical, chat-shaped, relationship-heavy, no-context, short-factual, multi-hop).
- Stable ID Contract: Topic IDs mapped to canonical identifiers (uuid5-based) ensuring qrels remain valid across re-indexing and extraction churn.
- Run Exporter: Decoupled exporter step converts live Flowbaby retrieval results into benchmark-ready run files while filtering LLM-augmented content.
- Benchmark CLI: Command-line interface for scoring runs and generating markdown performance summaries.
- Artifact Provenance: Comprehensive run summaries tracking git SHA, dataset versions, and retrieval contract state.

[0.7.3] - 2026-01-15

Added

Cognee Import Preflight Verification (Plan 108): Fail-fast preflight check prevents cryptic "No module named 'cognee'" errors by verifying the Python environment before bridge operations:
- Metadata-First Interpreter Selection: Bridge operations now prioritize .flowbaby/bridge-env.json metadata over explicit config, eliminating interpreter drift that causes missing module errors
- Preflight Gate: All bridge calls pass through a preflight verification that checks cognee importability before spawning Python
- Actionable Remediation: Failures provide clear guidance (managed env → "Refresh Bridge Dependencies"; external env → pip install instructions)
- In-Memory Caching: 30-second cache prevents repeated verification during activation bursts; cache invalidates on interpreter/ownership/requirements changes
Diagnose Environment Command (Plan 108): New Flowbaby: Diagnose Environment command generates a shareable markdown report with:
- Interpreter selection details (path, selection reason, ownership)
- Preflight verification status (cognee importability, version)
- Recommended actions for unhealthy environments
- Technical details in structured JSON format
Command Reference Documentation (Plan 109): Auto-generated COMMANDS.md provides searchable reference of all Flowbaby commands:
- Generated from package.json contributions for guaranteed accuracy
- npm run generate:commands regenerates; npm run generate:commands:check verifies sync
- Includes command IDs, titles, categories, and keybindings

Changed

Interpreter Selection Priority (Plan 108): Interpreter resolution now follows metadata-first priority:
1. .flowbaby/bridge-env.json metadata (single source of truth when present)
2. .flowbaby/venv heuristic (managed environment)
3. Explicit Flowbaby.pythonPath setting
4. System Python fallback
Preflight Cache Invalidation (Plan 108): Preflight verification cache now invalidates on requirements hash changes (in addition to pythonPath and ownership changes), ensuring refresh completion correctly triggers re-verification
Managed Venv Defaults (Plan 109): Setup flow now defaults to managed .flowbaby/venv without showing a choice modal:
- Removed modal asking whether to use existing .venv or create .flowbaby/venv
- Always creates managed environment for consistent, maintainable setup
- External ownership migration path remains available via Flowbaby.diagnoseEnvironment
Cloud Auth Terminology (Plan 109): Updated logging from legacy "API key status" to "Cloud auth readiness" for clarity in cloud-only mode

Fixed

Dashboard Setup Button (Plan 109): Dashboard "Setup Environment" button now correctly triggers Flowbaby.initializeWorkspace instead of unregistered Flowbaby.setup command
Login Prompt Gating (Plan 109): Cloud login prompts are now gated on local environment health—when the Python environment is broken or requirements mismatched, users are offered "Refresh Dependencies" instead of "Login to Cloud"
Refresh Quiescence (Plan 108): Verified that dependency refresh correctly pauses background operations and stops the bridge daemon before venv mutation, preventing "import during install" races

[0.7.2] - 2026-01-15

Changed

Backup Safety and User Confirmation (Plan 107): Pre-upgrade backup now requires explicit user consent and includes hardened detection to eliminate false-positive triggers:
- User Confirmation Modal: Backup is gated behind an information modal with "Proceed" / "Ignore" actions; closing the modal without choosing aborts the backup (fail-closed semantics)
- Pre-Backup Revalidation: A second detectMigrationState() call occurs after user confirmation to detect race conditions (e.g., marker appearing during modal display)
- Early Marker Write: Embedding schema marker is now written at the start of environment initialization (before bridge-env.json) to prevent cross-window false positives
- Mandatory Forensic Logging: All backup events log to console.log (not debugLog) with [BACKUP-TRIGGER] prefix for production visibility regardless of debug settings
- Explicit Detection State Model: detectMigrationState() now returns one of three explicit states (NOT_LEGACY, LEGACY_CONFIRMED, UNKNOWN_IO_ERROR) to prevent silent failures from triggering backups

Fixed

VS Code Insiders OAuth Login (Plan 106): OAuth callback now uses runtime URI scheme; fixes login redirect for VS Code Insiders users who were previously redirected to VS Code stable:
- Dynamic Callback URI: Callback scheme is derived from vscode.env.uriScheme at runtime instead of hardcoded vscode://
- Runtime Extension ID: Callback authority is derived from the runtime extension ID to avoid hardcoded identifiers
- Scheme Allowlist: Only vscode and vscode-insiders schemes are supported; unsupported editors (Cursor, VSCodium, etc.) receive explicit error guidance
- Fail-Closed UX: Unsupported schemes fail immediately with actionable user messaging rather than silent fallback
Cloud Auth Session Persistence (Plan 104): Fixed unexpected logouts after VS Code restart or idle periods by improving session refresh handling:
- Side-Effect-Free Auth State Query: New getAuthState() method allows readiness evaluation without triggering logout on expired tokens
- Bounded Activation Refresh: On startup, attempts session refresh within a 2-second budget to avoid blocking activation
- TTL-Aware Refresh Threshold: Sessions within 2 minutes of expiry are now proactively refreshed (was 24 hours)
- Refresh Coordinator: Singleflight pattern prevents concurrent refresh requests; 30-second throttling after failures
- Fail-Closed Semantics: Invalid refresh tokens (revoked/reused) immediately log out the user rather than retrying
- Activation Diagnostics: One-time startup log captures session state, remaining TTL, and refresh outcome for debugging

[0.7.1] - 2026-01-13

Fixed

Configured Python Adoption Marker (Plan 103): Prevent unexpected .flowbaby backups when using a configured Python interpreter by ensuring the embedding schema marker is written during adoption.

Changed

Settings UX Reorganization (Plan 102): Flowbaby settings are now organized into two groups for clarity:
- Flowbaby: Core settings for everyday use (7 settings including enabled, cloud zone, synthesis model, session management, debug logging, and notification toggles)
- Flowbaby (Advanced): Tuning knobs for power users and support workflows (retrieval limits, ranking parameters, search tuning, Python path, log level)
- Settings are now ordered by importance with explicit order fields
- Advanced settings include guidance labels ("Tuning", "Support-only", "Power-user") in descriptions
Hidden-but-Supported Settings (Plan 102): Three settings are now hidden from the Settings UI but remain fully functional when set via settings.json:
- Flowbaby.bridgeMode: Bridge execution mode (daemon or spawn)
- Flowbaby.daemonIdleTimeoutMinutes: Daemon idle timeout (1–60 minutes)
- flowbaby.cloud.apiEndpoint: Cloud API endpoint override (developer use)
- This reduces misconfiguration risk while preserving advanced troubleshooting capabilities
Centralized Configuration Layer (Plan 102): All settings reads now flow through a single TypeScript module (src/config/flowbabyConfig.ts) to prevent configuration drift and ensure consistent behavior across the extension.

Removed

Agent Access Configuration (Plan 102): The undeclared Flowbaby.agentAccess.* configuration surface has been removed. Concurrency and rate limits are now hardcoded architectural constants (max 5 concurrent requests, 5 queue size, 30 requests/minute) per the architecture doc.

[0.7.0] - 2026-01-07

⚠️ Breaking Changes

API Contract v4.0.0 Cutover (Plan 098): This release requires @groupzer0/flowbaby-api-contract v4.0.0 or later. The Cloud backend must be running contract v4.0.0 for authentication to work.
- Auth requests now include clientType: 'extension' to identify the extension surface
- Refresh endpoint returns RefreshResponse (not AuthResponse) with surface-specific union typing
- New fail-closed behavior: if server returns unexpected surface variant (e.g., web response to extension client), extension forces logout and prompts re-auth
- New v4 error codes handled: INVALID_AUDIENCE, AUDIENCE_MISMATCH, REFRESH_TRANSPORT_INVALID, REFRESH_REUSED
Cloud-Only Mode (Plan 083): Flowbaby v0.7.0 is Cloud-only. Local API key configuration (Flowbaby: Set API Key) has been removed. Users must login to Flowbaby Cloud for LLM-powered memory operations.
- Removed commands: Flowbaby.configureApiKey, Flowbaby.setApiKey, Flowbaby.clearApiKey
- Removed settings: Flowbaby.llm.provider, Flowbaby.llm.model, Flowbaby.llm.endpoint
- Removed env var injection: LLM_API_KEY, LLM_PROVIDER, LLM_MODEL no longer injected into Python bridge
- Legacy API keys stored in SecretStorage are automatically migrated with a one-time info message
Zone-Based Region Selection (Plan 094): The flowbaby.cloud.preferredRegion setting has been replaced by flowbaby.cloud.preferredZone:
- New zones: us (Americas), eu (Europe), apac (Asia-Pacific)
- If you had preferredRegion set, you must reconfigure with the new preferredZone setting
- Backend selects the optimal AWS region within the chosen zone

Added

Daemon Lock Recovery & Observability (Plan 095): Automatic recovery from stale workspace daemon locks with improved diagnostics:
- Stale Lock Recovery: When daemon startup encounters an existing lock, automatically recovers if the lock owner process is dead or lock age exceeds 10 minutes (conservative threshold)
- Lock Owner Metadata: Writes owner.json inside daemon.lock/ with createdAt, extensionHostPid, instanceId, and workspaceIdentifier for safe staleness detection
- Fresh Lock Guard: Prevents race condition where cleanupStaleLock() could delete a freshly acquired lock before daemon PID file exists
- Failure Path Cleanup: Lock is properly released if daemon startup fails after acquisition
- Lock Lifecycle Logging: Structured [lock] breadcrumbs for acquire/release/recovery with decision reasons (e.g., owner_pid_dead, lock_age_exceeded) for postmortem diagnostics
- Security: Logs avoid absolute workspace paths and secrets; uses relative markers and PIDs only
Daemonized Graph Visualization (Plan 097): Graph visualization now runs in-process via daemon mode when enabled, eliminating visualize.py subprocess spawns:
- In-Process Visualization: New handle_visualize() daemon handler calls visualize.visualize_graph() directly, avoiding subprocess spawn overhead
- Daemon-Wide Stdout Protection: New stdout_to_stderr() context manager redirects all stdout to stderr during daemon handler execution, preventing third-party library prints from corrupting JSON-RPC framing
- Spawn Fallback: When daemon mode is disabled, visualization falls back to subprocess spawn with identical result contract
- Windows File-Lock Fix: Running visualization in-process eliminates file descriptor inheritance issues that caused "file in use" errors on Windows
Automatic Backup of Incompatible .flowbaby/ on Upgrade (Plan 101): Users upgrading from v0.6.2 to v0.7.0 no longer encounter confusing embedding dimension errors. The extension automatically detects pre-0.7.0 workspaces and backs up existing data:
- Automatic Detection: On activation, checks for pre-0.7.0 .flowbaby/ directories (missing system/EMBEDDING_SCHEMA_VERSION marker)
- Safe Backup: Renames .flowbaby to .flowbaby-pre-0.7.0-backup-{timestamp} with Windows-safe naming (no colons) and collision resistance
- Daemon Lock Safety: Stops daemon before rename to prevent Windows file lock failures
- Clear User Guidance: Info toast explains what happened, where backup is, and prompts reinitialization
- Fail-Closed Fallback: If backup fails (permissions/locks), shows manual remediation instructions instead of proceeding with incompatible storage
- Schema Marker: New v0.7.0 workspaces write EMBEDDING_SCHEMA_VERSION=2 marker to prevent repeated backup on subsequent activations
Dashboard View Provider Tests (Plan 097): Comprehensive test coverage for the sidebar dashboard:
- Tests validate view type registration (flowbaby.dashboardView), WebviewViewProvider interface compliance, and refresh command resilience
- Single dashboard surface: Legacy DashboardPanel (editor tab) removed; sidebar view is now canonical
Geographic Zone Selection (Plan 094): New flowbaby.cloud.preferredZone setting for zone-based region selection:
- Three zones: us (Americas), eu (Europe), apac (Asia-Pacific)
- Dropdown in VS Code Settings with "Backend default (recommended)" option
- Backend authoritative for actual AWS region within zone
- INVALID_ZONE errors surface as user-facing configuration problems with remediation guidance
- CI drift guard (npm run check:zones) validates enum matches contract at build time
- Supersedes Plan 091's preferredRegion dropdown with simpler zone abstraction
Preferred Region Dropdown (Plan 091): The Flowbaby Cloud: Preferred Region setting now renders as a dropdown in VS Code Settings instead of a freeform text input:
- Dropdown displays all supported Bedrock regions (us-east-1, us-west-2, eu-west-1, eu-central-1, ap-northeast-1, ap-northeast-2, ap-southeast-2, ap-south-1)
- First option "Backend default (recommended)" selects the backend's default region
- New npm run check:regions script validates enum matches API contract allowlist at build time
Accurate Credit Consumption (Plan 090): Credits are now charged per successful Bedrock operation rather than at vend time, ensuring users only pay for completed operations:
- New consume() method in FlowbabyCloudClient calls POST /usage/consume endpoint
- Metering wired after successful retrieve operations (both synthesis and legacy paths)
- Metering wired after successful ingest operations (sync summary, sync conversation, and async cognify paths)
- Idempotent billing via X-Idempotency-Key header prevents duplicate charges on retries
- Fire-and-forget pattern: consume failures logged but never block user operations
- DIP-compliant metering abstraction (IUsageMeter) keeps Cloud concerns isolated from core paths
Bedrock Health Check Module (Plan 088): New bedrock_health.py module provides deterministic Bedrock connectivity validation using configuration proven stable in Analysis 088:
- Uses TextOut Pydantic response model (not response_model=str which is unstable)
- Strict JSON instruction prompt achieving 20/20 clean responses in batch testing
- System-first message ordering for improved adherence
- max_completion_tokens=2048 to prevent Bedrock 400 validation errors
- Actionable error messages with remediation guidance for auth, model, and network failures
Cognee Probe Bypass (Plan 088): New cognee_probe_bypass.py module prevents add-only ingestion failures caused by Cognee's internal LLM probe:
- Sets _first_run_done=True on Cognee's setup module to skip test_llm_connection()
- Bridge-side health check (bedrock_health.py) serves as authoritative connectivity validation
- Embedding validation deferred to first-use with clear error handling
Cloud Readiness Service (Plan 087): Unified Cloud readiness state management with three-axis model (auth/vend/bridge) and throttled error display. Eliminates misleading login prompts when user is authenticated but vending fails.
- Structured readiness model: AuthReadinessState, VendReadinessState, BridgeReadinessState with CloudOverallStatus (ready/login_required/degraded/error)
- Error throttling: Vend failures surface via toast with 30s min interval and max 3 per 5 minutes to prevent notification spam
- Actionable remediation: getRemediation() provides context-specific guidance based on current readiness state
- Event-driven updates: onDidChangeReadiness event fires on state changes for UI responsiveness
Debug Logging for Cloud API Errors (Plan 087): When flowbaby.debug is enabled, Cloud client now logs HTTP status codes and error codes (without secrets) to the "Flowbaby Cloud" output channel for improved diagnosability.
Cloud API Endpoint Override Setting (Plan 084): New flowbaby.cloud.apiEndpoint setting allows developers and testers to override the Flowbaby Cloud API endpoint for testing against staging, development, or local environments without modifying code.
Flowbaby Cloud Credential Provider Wiring (Plan 081): Unified credential/provider architecture wired into extension activation and all Python bridge paths. Key changes:
- API Contract v2.1.0: Added BedrockRegion type with preferredRegion field on VendRequest for region preference passthrough.
- Region Preference Setting: New flowbaby.cloud.preferredRegion setting allows users to express region affinity (us-east-1, eu-west-1, me-south-1, ap-southeast-1). Backend remains authoritative for actual routing.
- Activation Wiring: Credentials manager and provider singleton now initialize during extension activation (after Cloud auth) with proper dependency ordering.
- Bridge Environment Injection: All three bridge code paths (spawn-per-request, daemon-mode, background operations) now inject Cloud credentials into Python subprocess environment when authenticated.
- Security Guardrails: Verified no secret values (AWS keys, tokens, JWTs) are logged; existing redaction patterns remain adequate.
Credential Single-Flight Guard (Plan 083 M1): Prevents concurrent /vend/credentials requests during high-frequency operations. Multiple callers share a single in-flight refresh, reducing API load and preventing credential thrashing.
Centralized Cloud Error → UX Mapping (Plan 083 M3): New errorMapping.ts module provides consistent error messaging across all Cloud error scenarios:
- Maps error codes to severity, actionable messages, and recovery actions
- Supports rate limiting, quota exceeded, session expiry, and network errors
- Provides showCloudError() helper for consistent toast notifications

Changed

API Contract Package Migration (Plan 089): Migrated from repo-local contract copies to the canonical @groupzer0/flowbaby-api-contract@3.1.0 npm package. Contract types, constants, and helpers are now imported from the versioned package, eliminating copy/sync drift and ensuring extension/backend stay aligned via explicit semver'd dependency. Retired .github/workflows/sync-contract.yml and extension/src/flowbaby-cloud/contract/ local copy.
Activation Prompt Gating (Plan 087): Post-init prompts now use CloudReadinessService.needsLogin() instead of legacy llmReady flag. Login prompts only appear when user genuinely needs to authenticate—not when already authenticated but vending is failing.
Status Bar Driven by Cloud Readiness (Plan 087): Status bar icon and text now reflect the unified CloudOverallStatus state. When vend fails for an authenticated user, shows "degraded" state with appropriate guidance instead of misleading "login required".
Vend Errors Now User-Visible (Plan 087): All three Python bridge code paths (FlowbabyClient, PythonBridgeDaemonManager, BackgroundOperationManager) now surface vend failures via throttled toast notifications with actionable remediation guidance.
Memory Storage Reliability Improvements (Plan 092): Enhanced resilience for memory ingestion operations with multiple fixes addressing transient failures:
- Credential Refresh Manager Integration: Daemon now restarts automatically when Cloud credentials are refreshed, preventing stale credential errors during long sessions
- Exclusive Daemon Locking: File-based lock mechanism prevents duplicate daemon instances, eliminating Kuzu DB lock contention errors
- Staging Auto-Retry: ingestSummaryAsync() now automatically retries on transient errors (ECONNRESET, ETIMEDOUT, EPIPE, network errors) with configurable retry count and delay
- Cognify Auto-Retry: Background cognify operations automatically retry via subprocess path when daemon fails, with retry count tracking and delay
Cloud API Endpoint Resolution (Plans 084, 099): Default Cloud API endpoint is now production (api.flowbaby.ai). Endpoint resolution follows clear precedence: VS Code setting > FLOWBABY_CLOUD_API_URL environment variable > built-in default. Staging endpoint (api-staging.flowbaby.ai) available via override setting for development/testing. This enables reliable Cloud connectivity for v0.7.0 launch.
Bootstrap Decoupled from Cloud Credentials (Plan 084): Python bridge initialization (init.py) no longer requires Cloud credentials to succeed. Extension bootstrap (environment creation, workspace indexing) now completes without Cloud authentication, with only LLM-powered operations (memory ingestion, retrieval synthesis) gated on Cloud login. This fixes activation failures where users saw "Cloud login required" before they could even log in.
Cloud Authentication Flow (Plans 077–081): The v0.7.0 Cloud integration now provides a complete OAuth → STS → Bridge credential flow, enabling managed AWS Bedrock inference without local API keys.
Status Bar UX (Plan 083 M6): Updated status bar for Cloud-only mode:
- Renamed NeedsApiKey status to NeedsCloudLogin (backward-compatible alias preserved)
- Status menu now shows "Flowbaby Cloud Status" instead of "Set API Key"
- Tooltip updated to "Cloud Login Required" instead of "LLM API Key Required"
Walkthrough Updated (Plan 083 M6): Getting Started walkthrough step changed from "Set Your API Key" to "Login to Flowbaby Cloud" with updated description.

Fixed

Graph Visualization "No Data" Despite Data Existing (Plan 091): Fixed issue where Flowbaby: Visualize Memory Graph would show "No graph data available" in the browser even when memories had been ingested. The visualization script now uses the shared environment wiring contract (bridge_env.apply_workspace_env()) to ensure it reads from the correct workspace-local database.
- Added fail-closed empty-graph detection: the command now returns a clear error instead of a misleading success toast when the generated graph is empty
- Distinguishes between "no data ingested" and "data exists but read from wrong store" scenarios
- TypeScript command updated to show warnings for empty-graph errors
Bedrock Add-Only Ingest Failures (Plan 088): Fixed add-only memory ingestion failures caused by Cognee's internal LLM probe using unstable Bedrock structured-output configuration. The probe used response_model=str which fails intermittently with InstructorRetryException. Now bypassed in favor of a deterministic bridge-side health check.
Missing boto3 Dependency (Plan 088): Added boto3 to bridge requirements and verification. Existing workspaces that were missing boto3 (which caused silent litellm failures) can now be detected and remediated via dependency refresh.
Misleading Login Prompts After Vend Failures (Plan 087): Fixed issue where users who were authenticated but experienced credential vending failures would be repeatedly prompted to log in. The extension now correctly distinguishes between "not logged in" and "logged in but vend failing" states, showing appropriate guidance for each.
Silent Credential Vending Failures (Plan 087): Fixed issue where credential vending failures were silently swallowed, leaving users confused about why operations were failing. Vend errors now surface with throttled toast notifications (max 3 per 5 minutes, 30s min interval) with context-specific error messages.
Cloud API Endpoint Hardcoding (Plan 084): Fixed hardcoded api.flowbaby.dev URL in types.ts that referenced a non-existent domain. Now correctly defaults to staging endpoint (api-staging.flowbaby.ai) with documented fallback to execute API gateway URL when needed.
Cloud Error Code Preservation (Plan 083 M2): Fixed error masking in flowbabyClient.ts, PythonBridgeDaemonManager.ts, and BackgroundOperationManager.ts where Cloud errors were being converted to generic "login required" errors. Now preserves original error codes (RATE_LIMITED, QUOTA_EXCEEDED, SESSION_INVALID, etc.) for accurate UX.
Cloud Login Commands Not Responding (Plan 085): Fixed UI entrypoints (first-run prompt buttons, status bar menu) that invoked unregistered command IDs (FlowbabyCloud.*) instead of canonical registered commands (flowbaby.cloud.*). Clicking "Login to Cloud" or "Flowbaby Cloud Status" now reliably triggers the intended action. Status bar now updates automatically after login/logout without requiring an extension reload.

[0.6.2] - 2025-12-20

Added

Copilot Synthesis Model Selection (Plan 075): New Flowbaby.synthesis.modelId setting allows users to choose which Copilot model performs memory retrieval synthesis. Options include GPT-5 mini (default), GPT-4o, and GPT-4.1. Synthesis fails with a clear, actionable error if the selected model is unavailable—no silent fallback to a different model.
Settings Clarity (Plan 075): Updated settings descriptions to clearly distinguish:
- Flowbaby.synthesis.modelId: Controls Copilot synthesis model for memory retrieval
- Flowbaby.llm.*: Controls Python bridge LLM for memory ingestion and graph operations (does NOT affect Copilot synthesis)
Rebuild Tool Batch Processing & Resilience (Plan 076): The developer/tester rebuild CLI tool (rebuild_workspace.py) now supports robust filesystem-based rebuild with:
- Batch ingestion: Configurable batch size (--batch-size) with rate limiting (--batch-delay) for API quota management
- Checkpointing & resumability: JSON checkpoints with SHA-256 input fingerprinting allow resuming interrupted operations (--resume) or starting fresh (--restart)
- Preflight/dry-run mode: --dry-run shows file counts, total size, and skip reasons without making changes
- Fail-closed semantics: Destructive reset-and-rebuild mode validates all inputs before resetting stores; aborts on encoding errors or permission issues
- Concurrent writer detection: Refuses to run when daemon mode is active to prevent data corruption
- File size limits: Configurable maximum file size (--max-file-size, default 10MB) with override (--allow-oversized)

Changed

Extension Activation Refactor (Plan 062): Decomposed the 2066 LOC extension.ts monolith into three cohesive modules: extension.ts (267 LOC orchestrator), activation/registrations.ts (1175 LOC for commands, LM tools, chat participant), and activation/init.ts (797 LOC for initialization, health checks, setup commands). No user-facing changes; architecture maintainability improvement only.

Fixed

Windows daemon I/O reliability (post-Plan 076): Improved bridge daemon stdio handling on Windows to prevent stuck requests and unreliable background operations.
Windows refresh tooling reliability: Improved Flowbaby: Refresh Bridge Dependencies behavior on Windows where .flowbaby\\venv can be locked (common EPERM rename failures). The refresh/health flows now behave more predictably when a Python process is still holding file handles.
Workspace health check robustness (Windows + general): Fixed setup health checks to better detect and report environment readiness, reducing false negatives during initialization.
Cross-platform VSIX verification: npm run verify:vsix now supports Windows environments without unzip by falling back to PowerShell Expand-Archive, preventing packaging regressions before publishing.
Notification Observability (Plan 075): Removed temporary diagnostic logging from toast notification code paths, keeping production logs clean while preserving notification functionality.
Daemon Idle-Timeout KuzuDB Locks (Plan 061): Fixed aggressive SIGKILL escalation during idle-timeout shutdown that left KuzuDB .pkl file locks orphaned. The daemon manager now uses a 3-phase graceful shutdown (5s graceful → 3s SIGTERM → SIGKILL as last resort), allowing Cognee/KuzuDB time to release locks cleanly. Added operational fallback: after 3 consecutive forced kills, daemon mode automatically suspends and falls back to spawn-per-request until manually re-enabled.
Memory Ingestion KuzuDB Lock Conflicts (Plan 062): Fixed flowbaby_storeMemory tool failures with "Could not set lock on file" errors. Memory ingestion now routes through the daemon process (which holds the single KuzuDB connection) instead of spawning a separate subprocess that would conflict with the daemon's lock. Both ingestSummary() and ingestSummaryAsync() now use daemon-based ingestion when daemon mode is enabled, with fallback to subprocess when daemon is unavailable.

Changed

Cognee Upgrade to 0.5.1 (Plan 059): Upgraded from Cognee 0.4.1 to 0.5.1 to enable filesystem-backed session caching. Cognee 0.5.0+ introduces CACHE_BACKEND=fs via diskcache (bundled), removing the implicit Redis dependency that caused connection failures in managed environments without Redis installed.
Filesystem Cache Backend Default: Managed environments now default to CACHING=true and CACHE_BACKEND=fs (filesystem session cache) instead of falling back to Redis. This eliminates Redis-related timeout errors and improves retrieval reliability out-of-the-box.
Environment Variable Precedence: Cache configuration now respects explicit user-provided environment variables (CACHING, CACHE_BACKEND). Flowbaby-managed defaults are only applied when values are not already set.
Daemon Default Idle Timeout: Increased default idle timeout from 5 minutes to 30 minutes to reduce daemon restart frequency during normal development sessions.

Added

Cache Configuration Logging: All bridge entry points (daemon, init, ingest, retrieve) now log cache configuration at startup for improved observability and troubleshooting.
Automatic Schema Migration (Plan 060): Databases created with Cognee 0.4.x are automatically migrated to 0.5.x schema on initialization. The migration adds required columns for multi-tenant support (datasets.tenant_id and 10 dataset_database columns) using additive-only ALTER TABLE operations. No data loss occurs.
- Fail-Closed Safety: If schema migration fails, initialization fails cleanly rather than proceeding with an incompatible database.
- Receipt Generation: Migration attempts are logged to .flowbaby/system/schema_migration_receipt.json for auditability.
- External Repair: Users with external Python environments can run python bridge/migrate_cognee_0_5_schema.py <workspace> --dry-run to check schema status and apply migrations manually.
Daemon Startup Hygiene (Plan 061): Added stale PID file cleanup and orphan process detection on daemon startup to prevent ghost daemon accumulation and ensure clean daemon lifecycle management.

[0.6.1] - 2025-12-13

Changed

Switched the project and extension licensing from PolyForm Strict to MIT and prepared the repository for public open-source use, based on user feedback.

[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

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.
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.
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

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

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

Fresh install now completes successfully
Status bar shows key icon indicating API key needed
User prompted to set API key after initialization
Memory operations prompt for API key if not configured
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:

Creation of orphan .cognee folder in workspace root
Missing .flowbaby/data directory
"Flowbaby is still initializing..." hang on first activation
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:

Compute .flowbaby/system and .flowbaby/data paths
Create directories with Path.mkdir(parents=True, exist_ok=True)
Set os.environ['SYSTEM_ROOT_DIRECTORY'] and os.environ['DATA_ROOT_DIRECTORY']
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

Uninstall the old extension
Install Flowbaby v0.4.0
Run Flowbaby: Set API Key to configure your LLM API key
Update any custom keybindings from cognee.* to Flowbaby.*
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:  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

[0.7.5] - 2026-01-26​

Added​

Changed​

[0.7.4] - 2026-01-24​

Added​

Changed​

[0.7.3] - 2026-01-15​

Added​

Changed​

Fixed​

[0.7.2] - 2026-01-15​

Changed​

Fixed​

[0.7.1] - 2026-01-13​

Fixed​

Changed​

Removed​

[0.7.0] - 2026-01-07​

⚠️ Breaking Changes​

Added​

Changed​

Fixed​

[0.6.2] - 2025-12-20​

Added​

Changed​

Fixed​

Changed​

Added​

[0.6.1] - 2025-12-13​

Changed​

[0.6.0] - 2025-12-11​

Changed​

[0.5.23] - 2025-12-11​

Added​

Changed​

Technical​

[0.5.22] - 2025-12-10​

Fixed​

Added​

Changed​

[0.5.21] - 2025-12-10​

Changed​

Added​

[0.5.20] - 2025-12-10​

Fixed​

Added​

Changed​

[0.5.19] - 2025-12-07​

Added​

Changed​

[0.5.18] - 2025-12-07​

Added​

Changed​

[0.5.17] - 2025-12-05​

Added​

Changed​

[0.5.16] - 2025-12-05​

Fixed​

[0.5.15] - 2025-12-05​

Changed - Cognee Upgrade & Session Infrastructure​

[0.5.14] - 2025-12-05​

Added - Unified Session Context​

[0.5.13] - 2025-12-03​

Changed - Maintenance Release​

[0.5.12] - 2025-12-03​

Fixed - Windows Installation and Stability​

Improvements​

[0.5.11] - 2025-12-01​

Added - Plan 047: Enhanced Setup Diagnostics​

Improvements​

Windows Reliability Improvements​

[0.5.10] - 2025-02-18​

Fixed - Plan 046: Windows Spawn Fix​

Issues Resolved​

Technical Changes​

[0.5.9] - 2025-12-01​

Fixed - Plan 022: Windows Path Spaces Fix​

Issues Resolved​

Technical Changes​

[0.5.8] - 2025-11-30​

[0.7.5] - 2026-01-26

Added

Changed

[0.7.4] - 2026-01-24

Added

Changed

[0.7.3] - 2026-01-15

Added

Changed

Fixed

[0.7.2] - 2026-01-15

Changed

Fixed

[0.7.1] - 2026-01-13

Fixed

Changed

Removed

[0.7.0] - 2026-01-07

⚠️ Breaking Changes

Added

Changed

Fixed

[0.6.2] - 2025-12-20

Added

Changed

Fixed

Changed

Added

[0.6.1] - 2025-12-13

Changed

[0.6.0] - 2025-12-11

Changed

[0.5.23] - 2025-12-11

Added

Changed

Technical

[0.5.22] - 2025-12-10

Fixed

Added

Changed

[0.5.21] - 2025-12-10

Changed

Added

[0.5.20] - 2025-12-10

Fixed

Added

Changed

[0.5.19] - 2025-12-07

Added

Changed

[0.5.18] - 2025-12-07

Added

Changed

[0.5.17] - 2025-12-05

Added

Changed

[0.5.16] - 2025-12-05

Fixed

[0.5.15] - 2025-12-05

Changed - Cognee Upgrade & Session Infrastructure

[0.5.14] - 2025-12-05

Added - Unified Session Context

[0.5.13] - 2025-12-03

Changed - Maintenance Release

[0.5.12] - 2025-12-03

Fixed - Windows Installation and Stability

Improvements

[0.5.11] - 2025-12-01

Added - Plan 047: Enhanced Setup Diagnostics

Improvements

Windows Reliability Improvements

[0.5.10] - 2025-02-18

Fixed - Plan 046: Windows Spawn Fix

Issues Resolved

Technical Changes

[0.5.9] - 2025-12-01

Fixed - Plan 022: Windows Path Spaces Fix

Issues Resolved

Technical Changes

[0.5.8] - 2025-11-30