[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-81545":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":9,"language":10,"languages":9,"totalLinesOfCode":9,"stars":11,"forks":12,"watchers":13,"openIssues":14,"contributorsCount":15,"subscribersCount":15,"size":15,"stars1d":16,"stars7d":17,"stars30d":18,"stars90d":15,"forks30d":15,"starsTrendScore":19,"compositeScore":20,"rankGlobal":9,"rankLanguage":9,"license":21,"archived":22,"fork":22,"defaultBranch":23,"hasWiki":24,"hasPages":22,"topics":25,"createdAt":9,"pushedAt":9,"updatedAt":26,"readmeContent":27,"aiSummary":28,"trendingCount":15,"starSnapshotCount":15,"syncStatus":12,"lastSyncTime":29,"discoverSource":30},81545,"deputy-agent","SomeoneKong\u002Fdeputy-agent","SomeoneKong","a self-supervising master–worker agent framework that auto-scaffolds a task-tailored harness for long, autonomous delivery.",null,"TypeScript",32,2,26,1,0,3,5,6,9,1.43,"Apache License 2.0",false,"main",true,[],"2026-06-12 02:04:16","# Deputy Agent\n\n[English](README.md) | [中文](README.zh.md)\n\n> Deputy — a self-supervising master–worker agent framework that auto-scaffolds a\n> task-tailored harness for long, autonomous delivery.\n\nDeputy is a TypeScript framework for long-running, autonomous tasks. You hand it a task\ndescription; it generates a task-tailored harness and drives the work through a structured\nlifecycle to completion, syncing with you only at key checkpoints.\n\nIt uses a **master–worker** design rather than a peer multi-agent system: a master (the Meta\nrole) plans, prepares the harness, and arbitrates outcomes, while a worker executes the task —\nwith a Watcher (live observation) and a Reviewer (verdicts at stage gates) acting as review\nagents that audit and correct the worker. At heart it is a master–worker (2-agent)\narchitecture. Deputy is exposed both as a command-line tool and as a local Web GUI, and is\nmulti-provider: Claude and Codex are supported behind a common adapter layer.\n\n## Highlights\n\n- **Long, unattended tasks.** Built for jobs that run roughly **1 hour to ~2 days** without a\n  human in the loop, not short interactive turns.\n- **General (non-coding) focus.** Aimed at everyday white-collar \u002F knowledge work; it is\n  deliberately *not* specialized for coding.\n- **On-demand harness.** A task-tailored harness (methodology \u002F SOP \u002F tools \u002F completion\n  checks) is auto-generated per task, instead of one fixed harness for everything.\n- **Built-in review.** A master (the Meta role) drives a Worker, while a Watcher (live\n  observation of the worker's output) and a Reviewer (verdicts at stage gates) audit and\n  correct the work.\n- **File-system, task-level memory.** Agents coordinate through workspace files and messages;\n  the worker's multi-session state persists on disk.\n- **Per-role provider\u002Fmodel.** Each role can run on a different provider and model — e.g. a\n  cheaper model for the Worker.\n- **Built on Claude Code \u002F Codex.** Uses the Claude Code and Codex CLI agent kernels, in\n  TypeScript.\n\n## Project status\n\nThis is the **0.1.0** release, and a *reference implementation* of a higher-level design: the\nopen-source TypeScript here is essentially the compiled output of that higher-level spec. The\nspec and many of the detailed design principles are not part of this repository.\n\n- Because the published code is a compiled artifact, **future versions may change\n  substantially** — keep this in mind if you fork and intend to merge later releases.\n- Behavior is currently tuned primarily for **Claude**; **Codex \u002F GPT models perform less\n  well** today, since harness behavior is model-dependent.\n- It is **not yet production-grade** by the author's quality bar and needs further polishing,\n  though it may already exceed the tuning of some shipped products.\n- Cross-task memory and external know-how injection are **not implemented in 0.1.0** (only\n  task-level, file-system memory exists today). See\n  [docs\u002FLIMITATIONS.md](docs\u002FLIMITATIONS.md).\n\n## Requirements\n\n- Node.js >= 22\n- Provider credentials for at least one supported provider:\n  - Claude (via the Claude Agent SDK)\n  - Codex (optional, for roles bound to the `codex` provider)\n\n## Quickstart\n\n```bash\nnpm install\nnpm run build       # tsc + copy web static assets\nnpm run typecheck   # type-check only\n```\n\nAfter building, the **recommended** way to use Deputy is the local Web GUI:\n\n```bash\nnode dist\u002Fcli\u002Fbin.js web    # then open the printed URL (default http:\u002F\u002F127.0.0.1:4319)\n```\n\nThe Web GUI is the easiest way to submit a task and watch live progress, the conversation,\nand the agent output streams, and to drive the whole task lifecycle from one place.\n\nThe CLI exposes the same operations and is better suited to scripting \u002F headless use:\n\n```bash\nnode dist\u002Fcli\u002Fbin.js submit \"Write a report on X\"\nnode dist\u002Fcli\u002Fbin.js list\nnode dist\u002Fcli\u002Fbin.js status \u003CtaskId>\n```\n\n## Usage\n\nThe local **Web GUI (`deputy web`) is the recommended way to drive tasks**. The CLI exposes\nthe same operations and is best for scripting \u002F headless use: write commands (`submit`, `run`,\n`answer`, `feedback`, `upload`, `pause`, `resume`, `done`, `cancel`, `rename`, `delete`), read\ncommands (`list`, `status`, `inspect`), and the `web` command itself. See\n[docs\u002FUSAGE.md](docs\u002FUSAGE.md) (中文: [docs\u002FUSAGE.zh.md](docs\u002FUSAGE.zh.md)) for the full\ncommand reference, the `deputy.config.json` format, and an end-to-end example.\n\n## Project layout\n\n```\nsrc\u002F\n  shared\u002F      task capsule layout, manifest (state machine), atomic IO, ids, paths\n  wrapper\u002F     provider-neutral AgentRuntime interface + capability model\n    adapters\u002F  claude \u002F codex adapters, plus a stub for offline use\n    types\u002F     runtime \u002F capability \u002F session \u002F event type contracts\n  messaging\u002F   envelope schema + per-channel inbox bus (message passing)\n  prompts\u002F     prompt asset assembly for the agent roles\n  host\u002F        the daemon: tick loop, agent orchestration, stage machine\n    tools\u002F     host-provided tools the agents call\n    watcher\u002F   worker-stream windowing + dispatch to the observer role\n    done_criteria\u002F  declarative completion checks gating task completion\n  cli\u002F         CLI entry, argument parsing, config, daemon launch\n  web\u002F         loopback-only HTTP + SSE Web GUI backend\n```\n\n## Documentation\n\nEach document has an English version and a Chinese (`*.zh.md`) version.\n\n- [docs\u002FARCHITECTURE.md](docs\u002FARCHITECTURE.md) — overview: subsystem map, component diagram,\n  task lifecycle, glossary, and links into the focused docs below\n  (中文: [docs\u002FARCHITECTURE.zh.md](docs\u002FARCHITECTURE.zh.md))\n- [docs\u002FRUNTIME.md](docs\u002FRUNTIME.md) — host daemon tick loop, agent roles, stage machine,\n  message bus, concurrency & recovery\n  (中文: [docs\u002FRUNTIME.zh.md](docs\u002FRUNTIME.zh.md))\n- [docs\u002FDATA_FORMATS.md](docs\u002FDATA_FORMATS.md) — on-disk schemas: capsule layout,\n  `manifest.yaml`, envelopes, `events.jsonl`, `done_criteria.yaml`, `deputy.config.json`\n  (中文: [docs\u002FDATA_FORMATS.zh.md](docs\u002FDATA_FORMATS.zh.md))\n- [docs\u002FPROVIDERS.md](docs\u002FPROVIDERS.md) — the `AgentRuntime` interface, capability model,\n  and the claude \u002F codex \u002F stub adapters\n  (中文: [docs\u002FPROVIDERS.zh.md](docs\u002FPROVIDERS.zh.md))\n- [docs\u002FWEB.md](docs\u002FWEB.md) — local Web GUI: REST endpoints, SSE events, security model\n  (中文: [docs\u002FWEB.zh.md](docs\u002FWEB.zh.md))\n- [docs\u002FUSAGE.md](docs\u002FUSAGE.md) — installation, configuration, CLI and Web GUI usage\n  (中文: [docs\u002FUSAGE.zh.md](docs\u002FUSAGE.zh.md))\n- [docs\u002FLIMITATIONS.md](docs\u002FLIMITATIONS.md) — known limitations of the current implementation\n  (中文: [docs\u002FLIMITATIONS.zh.md](docs\u002FLIMITATIONS.zh.md))\n\n## License\n\nLicensed under the Apache License, Version 2.0. Commercial use is permitted; you must\nretain the copyright and license notices (see the [LICENSE](LICENSE) and [NOTICE](NOTICE)\nfiles), and state any significant changes you make to the files. The Apache-2.0 license\nalso includes an explicit patent grant. See [LICENSE](LICENSE) for the full terms.\n\n## Contact\n\nReach out if you:\n\n- want the higher-level spec, development know-how, or customization guidance;\n- are building a similar general-purpose long-task agent framework and want to compare notes\n  or collaborate;\n- want to apply this design in your own product.\n","Deputy Agent 是一个用于长时间运行、自主任务的 TypeScript 框架。其核心功能包括自动生成针对特定任务定制的工作流程，并通过结构化的生命周期管理直至完成，仅在关键检查点与用户同步。采用主从设计模式，由主控（Meta 角色）负责规划、准备和仲裁结果，而工作节点执行具体任务，另有观察者和评审者角色对工作进行审计和修正。此框架适合需要持续大约1小时到2天无需人工干预的知识型或白领工作任务场景。它支持多种提供商（如Claude和Codex），并提供命令行工具及本地Web GUI两种交互方式。虽然目前版本为0.1.0且尚未达到生产级标准，但已具备一定的实用价值。","2026-06-11 04:05:28","CREATED_QUERY"]