Release v0.4.4 · YunTaiHua/illusion-agent

Highlights

Memory system overhaul — Rebuilt to align with Claude Code's architecture: typed memory directory, manual mode support, per-subagent model configuration, and cross-environment API client isolation
Risk-tiered permission system — High-risk operations can no longer be session-level exempted; whitelist supports full directive exemption with automatic sandbox degradation
Cron permission isolation — Cron jobs now specify permission mode via environment variable without polluting global settings

Memory system refactor — Type-based directory structure (user/, feedback/, project/, reference/), Claude Code-compatible extract/dream subagents, configurable extract/dream models, cross-env independent API client
Risk-tiered permissions — Three-tier risk classification (low/medium/high), high-risk operations require explicit confirmation per action, complete directive exemption via whitelist
Sandbox auto-degradation — Removed enabled/fail_if_unavailable config; sandbox automatically degrades to permission-only mode when unavailable
Cron permission modes — Three scenarios (yolo/channel/inherited session) resolved via ILLUSION_PERMISSION_MODE env var, not persisted to settings.json
Session rename — New /rename command to rename conversations
Miscellaneous improvements — Memory, session, and CLI enhancements

Thinking indicator flicker — Suppressed residual thinking metrics after assistant's final reply; assistant_complete now carries tool_chain_follows to exit busy state immediately
Daemon log dual-writer — Fixed duplicate log writer in daemon process, synced session display names, added log size fallback cleanup
PermissionPrompt type signature — Fixed inconsistent high_risk parameter typing and unbound os import
Memory logger CI failure — get_memory_logger now always cleans up stale handlers before rebuilding, eliminating test failures from global logger singleton pollution

Web message memoization — Message components memoized; historical messages no longer re-parse markdown during streaming output

Synced docs — Updated permission, sandbox, and CLI documentation to reflect new configurations

pip install illusion-agent==0.4.4

记忆系统重构 — 基于类型的目录结构（user/、feedback/、project/、reference/），兼容 Claude Code 的 extract/dream 子代理，可配置 extract/dream 模型，跨 env 独立 API client
风险分级权限 — 三级风险分类（低/中/高），高危操作每次需显式确认，白名单支持完整指令豁免
沙箱自动降级 — 移除 enabled/fail_if_unavailable 配置；沙箱不可用时自动降级为纯权限模式
Cron 权限模式 — 三种场景（yolo/渠道端/继承会话）通过 ILLUSION_PERMISSION_MODE 环境变量解析，不持久化到 settings.json
会话重命名 — 新增 /rename 命令重命名对话
杂项改进 — 记忆、会话与命令行功能增强

思考指标闪烁 — 抑制 assistant 最终回复后的残留思考指标；assistant_complete 携带 tool_chain_follows 使前端立即退出 busy 状态
守护进程日志双写者 — 修复守护进程日志重复写入，同步会话显示名称，增加日志体积兜底清理
PermissionPrompt 类型签名 — 修复 high_risk 参数类型不一致和 os 未绑定问题
记忆日志器 CI 失败 — get_memory_logger 现在始终清理旧 handler 后重建，消除全局 logger 单例污染导致的测试失败

pip install illusion-agent==0.4.4