Skip to content

CHANGELOG.md

Latest commit

 

History

History
105 lines (85 loc) · 5.03 KB

CHANGELOG.md

File metadata and controls

105 lines (85 loc) · 5.03 KB

Changelog

All notable changes to M-flow will be documented in this file.

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

[0.3.4] - 2026-04-13

Fixed

  • Critical: Remove max_tokens parameter from LLM calls — incompatible with GPT-5 series, caused all LLM-dependent operations to fail with "Connection error" after 120s retry
  • Critical: Fix session/conversation history crash — 3 retrievers called compress_text(str) but function requires list[ContentFragment]; replaced with correct summarize_text(str)
  • Fix LLMGateway.transcribe_audio calling non-existent create_transcript method on adapter
  • Fix LLMGateway.describe_image calling non-existent transcribe_image method on adapter
  • Fix UUID serialization in search results, cache adapters, and graph property encoding
  • Fix _hash_sensitive dead code — unreachable return obj caused telemetry data corruption
  • Fix CLI _ToggleDebug/_LaunchUi argparse dest keyword conflict
  • Fix ImageLoader JPEG extension mismatch (.jpe/.jpeg had leading dots)
  • Fix Bedrock adapter show_prompt calling non-existent LLMService.read_query_prompt
  • Fix OpenAIAdapter.describe_image blocking event loop (sync litellm.completion → async acompletion)
  • Fix OpenAIAdapter.transcribe_audio blocking event loop (litellm.transcription → atranscription)
  • Fix retry decorator across all 14 LLM/embedding adapters — now excludes BadRequestError and AuthenticationError to prevent 120-second retry storms on deterministic client errors
  • Fix dedup _build_modern_id tenant_id normalization (None → empty string)
  • Fix graph relationship ledger UUID collision on batch insert
  • Fix PostgreSQL delete_database missing cache_clear() (SQLite branch already had it)
  • Remove WARNING-level debug log that printed user input on every structured extraction call

Changed

  • Replace Concept_name with Entity_name in vector collection references (class renamed during migration)
  • Update ruff configuration and format entire codebase (311+ files)

Security

  • Pin all GitHub Actions to exact commit SHA (81 GitHub-owned + 21 third-party)
  • Add permissions: {contents: read} to 7 workflow files lacking token restrictions
  • Add .github/dependabot.yml for automated dependency updates
  • Fix OpenSSF Scorecard SARIF upload (codeql-action SHA dereference)

CI

  • Configure GitHub Secrets for LLM/embedding API keys
  • Fix workflow YAML syntax errors (matrix refs in workflow_call job-level conditions)
  • Fix uv.lock consistency
  • Update test assertions to match current API signatures and edge type names
  • Remove 6 phantom workflow jobs referencing non-existent test scripts
  • Skip S3/Modal tests requiring external paid services

[0.3.3] - 2026-04-12

Added

  • Episode naming in precise mode when content routing is disabled
  • New LLM prompt precise_name_single_content.txt for single-content Episode naming

Fixed

  • Generate meaningful Episode names instead of hardcoded "Content" in precise mode

[0.3.2] - 2026-04-05

Added

  • Precise summarization mode (precise_mode): preserves all original factual information (dates, numbers, names, constraints) with lower compression ratio — RAG context will be longer but more accurate. Uses two-step pipeline: JSON topic routing + per-section concurrent compression with anchor verification
  • Configurable via API parameter, environment variable (MFLOW_PRECISE_MODE), or frontend toggle
  • KuzuDB adapter: entry-level deduplication and broadened error recovery for batch operations

Fixed

  • KuzuDB "duplicated primary key" crash in UNWIND+MERGE operations
  • Edge deduplication key collision in deduplicate_nodes_and_edges

[0.3.1] - 2026-03-28

Added

  • Procedural memory extraction and retrieval
  • Model Context Protocol (MCP) server support
  • Frontend Knowledge Graph visualization with procedural subgraph
  • Structured entry (manual ingest) with display text support
  • Episodic retriever with adaptive bundle search
  • Multi-dataset search with access control
  • Docker Compose profiles for flexible deployment

Changed

  • Improved episodic retrieval scoring with path-cost model
  • Enhanced content routing with atomic mode for short inputs
  • Structured JSON output for procedural search results
  • Unified triplet search with configurable vector collections

Fixed

  • Episode naming using dedicated summarization prompt
  • Content inflation from prompt injection in short inputs
  • WebSocket authentication token handling
  • Pipeline status tracking and stale detection
  • Manual ingest graph creation and embedding generation

[0.3.0] - 2026-02-15

Added

  • Multi-user access control with dataset isolation
  • Episodic memory architecture (Episodes, Facets, FacetPoints, Entities)
  • Frontend dashboard with real-time pipeline monitoring
  • LanceDB vector storage integration
  • KuzuDB graph database adapter

For detailed release notes, see GitHub Releases.