| Project | What it does |
|---|---|
| Agent Harness | |
| looongtime-agent | A harness for long-horizon coding tasks. Task state is durable and survives interruption; completion is judged on execution evidence, not the model's own report. 面向长程任务的 Coding Agent Harness。任务状态持久化，中断后可恢复；完成与否以执行证据判定，而非模型自述。 |
| AAA-Agent | A terminal coding agent built from scratch. Routes models by task, adapts execution policy to model capability, and verifies its output before handing it back. 从零构建的终端 Coding Agent。按任务路由模型，执行策略随模型能力调整，产出经验证后再交付。 |
| agent-merge | Multi-agent orchestration with version control for agent sessions. Workers run in isolated Git worktrees, patches are selected by test results, failed attempts go to a repair round, and every attempt is retained as a traceable timeline. 多 Agent 编排与会话版本控制。多个 Agent 在独立 Git worktree 中并行处理同一任务，以测试结果筛选补丁，失败的进入修复轮；全部尝试留存为可回溯的时间线。 |
| agent-learning-gate | Governs what an agent learns. Checks evidence and scope before anything is written to memory, rules or skills, and records rejected approaches so a new session does not retry a path already ruled out. 管控 Agent 的长期学习。写入记忆、规则或 Skill 前校验证据与作用域，避免把一次性反馈固化为长期指令；同时留存被否决的方案，防止新会话重复已淘汰的路径。 |
| Speech · 语音 | |
| TurnAlign | Streaming ASR orchestration with timestamps and speaker diarization. The underlying recognition model is replaceable. 流式 ASR 编排，输出带时间戳与说话人分离结果，底层识别模型可替换。 |
| DayAudio | Local transcription for day-long recordings. Resumable, speaker-aware, and every line of the summary is linked back to its position in the audio. 全天录音的本地转写。支持断点续跑与说话人识别，摘要中每条结论均关联到原始音频位置。 |
| Writing · 写作 | |
| guanzhengpm.github.io | Notes on agents, evaluation and LLM product work. 关于 Agent、评测与大模型产品的笔记。 |
