·
102 commits
to main
since this release
Added
Memory & Neo4j
- Neo4j memory backend foundation: entities, facts, temporal attributes, graph schema, search, and backend abstraction
- Hybrid memory search with BM25, vector search, graph BFS, RRF, reranking, and all-scope search
- Bi-temporal memory search using
valid_atfilters and boosts - Fact/entity extraction pipeline with contradiction-based invalidation and entity resolution
- Community detection and graph context injection for Priming
- Memory migration CLI with backup, checkpoint, and rollback support
- LoCoMo benchmark adapter and evaluation tooling for memory quality comparisons
- Per-Anima memory backend switching
RAG Reliability
- RAG auto-repair and supervised repair lifecycle
- Worker-only ChromaDB access through the vector worker to avoid unsafe multi-process native access
- Recovery for missing ChromaDB collections and wiped vectordb
- CPU fallback when CUDA probing fails during embedding model load
- Serialized native vector operations to reduce ChromaDB HNSW corruption risk
Skills
- Enhanced
SKILL.mdloader/index compatible with Hermes-style skills - Skill usage tracking and metrics
- SkillIndex integration into the prompt builder and ToolHandler
- Skill hub importer, skill curator lifecycle, quarantine, trust gates, and promotion pipeline
- Explicit skill activation backend and activation rendering
- Procedure skill promotion pipeline and autonomous skill lifecycle
- Skill-backed cron context
- Legacy flat skill migration into trusted skill bundles
- Hermes and OpenClaw skill migration importers
TaskBoard & Work Management
- TaskBoard SQLite projection foundation
- TaskBoard API, dashboard UI, summary projection, stale housekeeping, and attention resolver
- Task execution gating with TaskBoard attention
- Persistent goal loop and goal judgment
- Workspace access grant tool
Workspace Replay
- Workspace replay foundation, semantic events, and replay narrative UI
- Semantic replay support for the activity timeline and workspace visualization
Execution & Runtime Isolation
- Runtime session isolation for chat, non-chat, consolidation, and execution lanes
- Agent session lane isolation and duplicate
submit_tasksguard - Auto-approval for AnimaWorks MCP tools in Mode C
- Action-aware priming for Mode S and action memory gate for Mode C
- PreCompact hook to block unsafe SDK auto-compaction
- Azure OpenAI support for Codex mode
- Config-level default model fallback and smarter credential/mode resolution
Supervisor & Bootstrap
- Bootstrap lifecycle repair and complete bootstrap repair command
- Better scheduler/cron health detection, including expected-fire gating and a 24h no-execution window
- Inbox TTL hygiene and failed-message retry/skip protection
- Board outbound sync whitelist
- Reverse-proxy sub-path deployment via
base_path
Community Contributions
- Merged external contributor fixes:
Changed
- Memory architecture now supports both legacy and Neo4j-backed operation
- Priming can now include graph/community context and memory pointers
- Skill handling moved toward explicit activation, trust, quarantine, promotion, and bundled
SKILL.mdstructures submit_tasksis gated to explicit background sessions to prevent duplicate work- Tool and action behavior rules are synchronized across behavior templates, action memory, and skill rules
- Cron health warnings are softer for low-frequency daily/weekly crons
- Documentation and README content updated for the current memory architecture
Fixed
Stability
- Fixed Claude/Codex subprocess cleanup and SDK drift handling
- Fixed chat cache directory configuration
- Fixed Agent SDK missing session ID guard
- Fixed consolidation credential resolution and session contamination
- Fixed startup idle compression blocking readiness
- Fixed heartbeat crash recovery polluting conversation state
- Fixed repeated chat session responses
- Fixed multiple silent
except: passpaths by logging debug details
Memory & RAG
- Fixed Neo4j schema inconsistencies, Cypher group scope, idempotent ingest, and query-aware retrieval
- Fixed vector collection recovery and ChromaDB access safety
- Fixed graph retrieval and all-scope search edge cases
- Fixed stable defaults for the legacy memory backend
- Fixed RAG repair recovery and vector worker isolation
Skills
- Restored flat skill compatibility
- Fixed skill usage event recording
- Fixed skill hub source fetching, staging bounds, HTTPS enforcement, and lock rollback
- Fixed skill curator and autonomous skill lifecycle review gaps
- Fixed trusted skill actor derivation in tool paths
- Fixed legacy scalar skill source metadata handling
Task & Supervisor
- Fixed TaskBoard execution gate review issues
- Fixed snoozed batch task suppression behavior
- Fixed task completion notification limits
- Fixed scheduler cron health false positives and no-execution false alarms
- Fixed pending task and attention handling regressions
UI / API
- Fixed base-path TaskBoard access
- Fixed workspace replay hardening issues
- Fixed setup wizard shared assets
- Fixed Slack send whitelist policy and outbound routing feedback
- Fixed media proxy and frontend test regressions
Documentation
- Updated memory docs in Japanese and synchronized foreign-language memory documentation
- Added implemented issue/review documents for major lifecycle changes
- Added common knowledge updates for action rules, task delegation, memory, and PPC handling guidance
- Updated README files and feature/API/CLI documentation
Internal
- Added CI workflow updates and broad test coverage for memory, skills, TaskBoard, replay, runtime isolation, and RAG repair
- Added benchmark artifacts and LoCoMo evaluation results
- Large refactors across memory, skills, prompt building, ToolHandler, supervisor, and server routes