openJiuwen agent-core v0.1.17

openJiuwen agent-core v0.1.17

Release Date: August 18, 2026

Welcome to v0.1.17! This release substantially strengthens the reliability of multi-agent teams, SwarmFlow, trajectory-driven evolution, and context and memory processing. It also introduces persistent session goals, third-party agent integration, OAuth account authentication, prefix caching, richer workflow operators, and multimodal Skill-Omni support. The release hardens filesystem and package boundaries, improves observability and model compatibility, and resolves a broad set of cancellation, concurrency, recovery, and tool-execution issues.

🚀 Features

Agent Teams & SwarmFlow

• Added third-party agent support for AgentTeam – Teams can now integrate external agents alongside native members (#342).
• Added persistent, session-scoped goals – DeepAgent sessions can keep a goal across reconnects and resumes, with lifecycle commands for setting, pausing, resuming, clearing, and assessing progress (#252).
• Expanded SwarmFlow operators and runtime control – Added human-in-the-loop nodes, stateful agents_session workers, budget constraints, and branch/operator-level suspend, resume, and cancel controls (#669).

Evolution, Skills

• Added Skill-Omni and multimodal Harness support – Harnesses can process image and video inputs when invoking Skill-Omni capabilities (#702).
• Added explicit and dynamic prefix caching hooks – Supports marking reusable model-request prefixes to reduce token cost and latency (#704).
• Added third-party OAuth account authentication – JiuwenSwarm can use OAuth-authorized model-service plans in addition to API keys (#695).
• Enabled multi-agent WebAgent execution – Team members can run isolated browser sessions concurrently, with headless execution and resource cleanup support (#703).

🐛 Bug Fixes

Agent Teams

• Fixed deleting a team before it has been built – Simple team-mode queries can now be cleaned up safely (#638).
• Fixed multiple @member mentions in group messages – Subsequent mentions are now parsed and routed instead of being treated as plain text (#634).
• Reduced excessive tool exposure in agent mode (#629).
• Fixed idle-triggered scheduling and task acceptance defaults – Scheduling mode and task acceptance now default to disabled, preventing unintended team.idle activation (#415).
• Fixed terminated DeepAgents returning to the queue or leaving termination state (#404).
• Initialized worktree sessions before task scheduling (#362).
• Prevented members from treating team_desc artifact requirements as executable instructions – Avoids concurrent overwrites of the leader's output (#358).
• Blocked leader messages addressed to the pseudo-recipient user (#248).
• Fixed human-member task assignment deadlocks – Assigned tasks no longer remain pending when a human member cannot call claim_task (#157).
• Fixed cancelled-task member shutdowns preventing team completion (#226).
• Fixed unread messages from shutdown members blocking team completion (#138).
• Fixed Feishu human-member routing and duplicate delivery – Point-to-point leader messages are no longer leaked to other members, and repeated requests no longer fail with unavailable-member errors (#259, #176, #173).
• Fixed team-member crash handling and shutdown recovery (#1394).
• Fixed temporary-team cleanup when no worktree or project_dir exists (#1318).
• Added adaptive multi-agent collaboration example – New example demonstrates adaptive coordination patterns for agent teams (#154).

Harness / Core

• Fixed SwarmFlow runs stopping without a final result (#480).
• Fixed unaccounted token consumption when a workflow exceeds its execution limit (#504).
• Fixed Windows concurrent worktree creation – Enables long-path-safe creation and serializes Git configuration updates to prevent path-length failures and .git/config.lock contention (#528).
• Fixed worktree reuse when the web client does not provide project_dir (#196).
• Moved .agent_history out of user project directories (#398).
• Persisted ReAct context and message history after interruption or cancellation – Resumed sessions now retain cleaned context and prior progress (#371, #223, #1396).
• Fixed branch-condition regular-expression evaluation (#696).
• Added fallback exception handling for runtime loops (#1387).
• Enabled collaborative multi-agent evolution in Context Evolving (#354).
• Introduced the refactored ContextEngine pipeline – New default processor stack includes message offloading, rule-based compression, reinjection, dumps, and active compression; the former pipeline remains available as an opt-in legacy fallback (#158, #494).
• Added AscendAffinity KV-cache affinity protocol and lifecycle management (#1372).
• Unified trajectory message reconstruction and online-evolution access – Trajectory now owns cross-span message rebuilding and OpenAI-compatible conversion; evolution rails can retrieve selected message fields through a shared entry point (#547).
• Consolidated canonical trajectory APIs – Isolated offline compatibility utilities, removed conflicting legacy online exports, and standardized runtime and persistence data models (#534, #460, #447).
• Made observability spans reliable for trajectory consumers – Canonical spans can now be consistently consumed by Evolution and Agent RL workflows (#492, #508, #519).
• Fixed evaluation aggregation for invalid model output – null values and malformed JSON no longer cause all completeness or accuracy aggregates to fail (#521).
• Fixed evaluator API inconsistencies after agent and skill migration (#434).
• Fixed model-anomaly detection rail behavior (#513).
• Fixed skill-spectrum build failures and false completion states – Build errors now surface when models are unavailable, and incremental/full builds no longer remain stuck at 3% (#421, #420).
• Hardened equivalent-Skill candidate completion – Invalid, missing, self-referential, or over-expanded candidates no longer fail large capability-tree builds (#545).
• Fixed Skill evolution trigger and write-contract mismatches (#306).
• Fixed early failures during Skill evolution review and interrupted-run recovery (#1378).
• Decoupled Skill creation from automatic evolution prompts (#1326).
• Persisted Skill rail session baselines and isolated runtime Skill mutations (#1368).
• Added structured output required for mode-based execution (#304).
• Matched tool-loop compaction using tool sets and arguments only (#557).
• Fixed long-running claw-bench runs stopping mid-execution (#552).
• Changed goal prompt injection to dynamic injection – Goal context no longer creates stale persistent prompt state (#383).
• Simplified redundant system-prompt logic (#346).
• Restored browser subagent tool-result windowing (#130).
• Fixed continuous same-tool calls triggering compression exceptions (#183).
• Added session-memory compression and compressed recall support (#159, #202).
• Fixed daily-memory cache reuse – Repeated same-day queries no longer unnecessarily re-read the database (#207).
• Fixed memory search/get availability and retrieval correctness (#1320, #1377).
• Fixed memory index rebuilding after embedding-dimension changes (#1319).
• Adapted jiuwen_memory_provider to the mem2.0 branch (#314).
• Updated the Chinese-query translation lexicon (#1321).
• Updated ContextEngine documentation (#312).
• Added missing AgentTeam documentation (#296).
• Supplemented online reinforcement-learning and Skill-evolution documentation – Includes updated self-evolution guidance and navigation entry points (#237, #230).

Time & LLM Parsing

• Fixed IntelliRouter usage conversion with third-party UsageMetadata – Missing reasoning_tokens no longer raises an attribute error (#348, #335).
• Added the Anthropic dependency and fixed incompatible parameter combinations – Anthropic requests no longer send top_p and temperature together (#326, #287).
• Updated model context-window token configuration (#281).
• Preserved OpenAI's original finish_reason (#195).
• Fixed HTTPS calls using only the system trust store (#654).
• Made OpenAI device-flow polling resilient to transient SSL and connection failures (#149).
• Improved model pooling and reduced call latency (#184).

Tools & Filesystem

• Fixed MCP tool-result parsing to retain all content blocks (#290).
• Fixed SSE MCP client teardown across task cancellation scopes (#321).
• Added reconnect handling after SSE MCP TCP resets (#180).
• Fixed excluded_commands split checking in JiuwenBox (#458).
• Added tool-call timeouts to prevent tasks from blocking indefinitely (#228).
• Clarified code-tool and input-parameter descriptions for models (#147, #142).
• Refined Todo prompts – Reduces unnecessary planning and requires incremental status updates (#185).
• Fixed parallel edit_file overwrite behavior and a read/write lock memory leak (#653, #1325).
• Optimized browser-agent core performance and removed obsolete manual browser startup (#276, #1393).
• Fixed browser shutdown to prevent lingering Chrome processes (#1383).
• Optimized image reads – Eliminates repeated PIL decoding and Base64 encoding for large images (#697).
• Corrected Base64 image payload token accounting (#271).
• Reduced session initialization delays caused by LSP startup (#257).
• Added Yuanrong sandbox support (#648).
• Added tool-call records to team observability and SwarmFlow support to TeamMonitor (#206, #133).
• Added session_id to OpenTelemetry span data (#136).
• Preserved structured model inputs and outputs in the OpenTelemetry rail – Handlers can format payloads without receiving prematurely stringified values (#642).
• Added TeamMonitor.get_task for task-detail retrieval (#1389).
• Fixed stale usage-token baseline invalidation (#300).
• Fixed hybrid publisher reconnection after idle disconnects (#536).
• Fixed PerStream startup failures (#153).
• Defaulted omitted wake_offset_seconds to 0 (#1397).
• Added pre-assignment support for non-leader members in autonomous create_task flows (#137).
• Separated artifact directories for the primary agent and subagents (#1381).
• Added checkpoint support to team_harness.dispose (#1380).

Security & Path Handling

• Made SecurityRail opt-in (#243).
• Restricted SkillTool reads to the selected directory boundary – Prevents sandbox-local file reads through path-boundary bypasses (#198, #1324).
• Hardened EvolutionStore Skill-package installation paths – Prevents directory traversal during package installation (#1323).

📚 Documentation

• Added sandbox documentation (#215).
• Added OpenAI Account OAuth authentication documentation (#143).

🙏 Acknowledgments

Thank you to all contributors who reported issues, supplied detailed reproduction paths, and improved the implementation and documentation for v0.1.17. This release reflects substantial community work across team collaboration, evolving agents, runtime reliability, tooling, and security.