AgenticX v0.5.0
Published Python package: agenticx==0.5.0
Installation
pip install -U agenticx==0.5.0PyPI
Highlights
-
Core / Runtime
- ToolSearch projects only the tools needed each round (with MCP catalog adapter and adaptive threshold/hysteresis), cutting tool-schema bloat on large MCP setups.
- Stronger turn reliability: stuck-prevention self-healing, futile-resume circuit breaker, empty-tool-turn fallback, truncated-reply auto-retry, and SSE drop / auto-resume continuation.
- Human-in-the-loop primitives: blocking
request_clarification, generic action confirmation, and a canonical assistant-output protocol with terminal reply finalization. - Context efficiency: prompt-cache hit awareness, vision-history budget trimming, smarter autocompact driven by token-window usage, and persisted compaction notices across reload.
widget_flow_guardauto-retries text “flow diagrams” into proper widget rendering; reasoning chains persist across session reload.
-
Memory
- Graph search merges into
memory_search/ recall; chunk composite rerank and access reinforcement improve hit quality. - Subject-scoped workspace resolution with group-chat bootstrap;
memory_append/memory_searchroute by subject. - Turn-archive recall plus episode governance (retention / natural-language forget) and corrupt Kuzu DB auto-recovery for end users.
- Graph search merges into
-
Skills / Learning
- Skill self-evolution reaches Hermes-parity gates (GEPA + five quality gates) for safer auto skill creation.
- In-chat blocking confirm for
skill_manageproposals; new research skillscode-deep-researchandcode-module-summaries(custom layout + batch subagent workflow).
-
MCP
MCPServerConfiggainstransport/url/headers;MCPClientV2supports streamable-HTTP remote MCP servers.
-
Studio / Server
- Sub-agent run records + activity timeline with review REST API.
video_understandtool (adaptive frame extraction + optional audio transcription).- Background bash toolchain for interactive/long-running commands; session completion truth aligned with
turn_terminal.
-
Hooks
- New
hook_manageagent tool for conversation-driven hook creation, with safety gating for command-type hooks.
- New
Fixes & Improvements
-
Core / Runtime
- Hardened LLM stream watchdog / stop spinning; fixed repeated budget compress, same-turn double-compact, and post-compact context-chain breakage that could silently end tasks.
- Reasoning-only empty turns no longer strand users; streamed reasoning is preserved across fallbacks; GLM stream finalization and inline tool parse improved.
- Sub-agent completion reliability: fewer false
failedjudgments, output-path priority over generic path detection, and retry reuse of the originalagent_id. - Current date/time grounded on the local clock for Desktop and Enterprise runtimes.
-
Memory
- CJK substring search for workspace memory; sanitize long entries before writing
MEMORY.md. - Memory-graph stability: empty Qwen ingest handling, episode endpoint NameError, delete deadlocks, and Kuzu DELETE SIGSEGV workarounds via rebuild-style delete.
- CJK substring search for workspace memory; sanitize long entries before writing
-
Studio / Server
- Cold-start empty window fixed (SQLite lock contention + supervisor materialization); restored
GroupChatRegistryimport that could crashagx serve. - Session “false idle” while tool chains were still running; background bash job ownership isolation; document/image hydration before model turns.
- Meta default workspace unified to
~/.agenticx/workspaceand scoped per session where needed.
- Cold-start empty window fixed (SQLite lock contention + supervisor materialization); restored
-
LLMs
- Keyless intranet OpenAI-compatible gateways; context-window faults classified with compact notices; LiteLLM noise suppressed; MOMA slash model ids routed via
openai/for LiteLLM gateways; Kimi K3 temperature forced to1.
- Keyless intranet OpenAI-compatible gateways; context-window faults classified with compact notices; LiteLLM noise suppressed; MOMA slash model ids routed via
-
MCP / Skills / Extensions
load_mcp_configtolerates per-entry parse failures; default ClawHub market when no registry is configured; third-party skill provenance / source labeling.
-
Packaging
- Desktop packaging path hardened for ChromaDB collect + BailianEmbedding cross-loop session reuse (keeps local knowledge-base installs usable in bundled backends).
Stats
- Commits: 706 commits over 63 days (2026-06-06 → 2026-08-08)