Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add PsChina/deepseek-as-subagent --skill delegate-to-deepseekgit clone --depth 1 https://github.com/PsChina/deepseek-as-subagentWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/pschina/deepseek-as-subagent/delegate-to-deepseek)<a href="https://agentmods.dev/skills/pschina/deepseek-as-subagent/delegate-to-deepseek"><img src="https://agentmods.dev/badge/skills/pschina/deepseek-as-subagent/delegate-to-deepseek/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/pschina/deepseek-as-subagent/delegate-to-deepseek"><img src="https://agentmods.dev/badge/skills/pschina/deepseek-as-subagent/delegate-to-deepseek.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector pass
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00125 | $0.01989 |
| Opus 5 | $0.00063 | $0.00994 |
| Sonnet 5 | $0.00025 | $0.00398 |
| Haiku 4.5 | $0.00013 | $0.00199 |
Grade A, and why
delegate-to-deepseek scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 10d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 126 lines — stays where its author put it; the contents beside it link to each section on GitHub.
delegate-to-deepseek — 主 Agent 派工准则
“主 Agent”指负责决策、整合与最终验收的上层 agent。
1. 不可突破的边界
- 用户显式要求派 / 不派、指定执行者或执行方式时，优先服从用户。
- 权限、安全、隐私、敏感信息和非授权写入边界不可绕过。
DEEPSEEK_MODE=off时不要派工。- DeepSeek 读取的文件内容会发送到配置的 API endpoint；敏感工作区不得派工。
- coding 能力固定为 Read / Write / Edit / Bash / Glob / Grep / NotebookEdit；readonly 固定为 Read / Glob / Grep。不得通过 task、steering 或其它参数提权。
- coding Bash 是受边界约束的 trusted-host Bash，不是操作系统级沙箱。
- 派工结果必须由主 Agent 独立验证，失败由主 Agent 收口。
- 出现文件 mutation、取消、断连或 MCP 重启后，先调用
get_deepseek_recovery()，核验实际文件，再用精确 transaction IDs 调用acknowledge_deepseek_mutations(...)；未确认前不要重试 mutation delegation。
2. API 选择
| 需求 | API |
|---|---|
| coding，直接等结果 | delegate_to_deepseek(task, context="", model="flash") |
| 纯静态文件分析，直接等结果 | delegate_to_deepseek_readonly(task, context="", model="flash") |
| coding，需要 steering / status / cancel | start_deepseek(task, context="", model="flash") |
| readonly，需要 steering / status / cancel | start_deepseek_readonly(task, context="", model="flash") |
| 查询后台任务 | get_deepseek_status(job_id) |
| 追加/修正后台指令 | send_deepseek_message(job_id, message) |
| 取消后台任务 | cancel_deepseek(job_id) |
| 获取最终结果 | get_deepseek_result(job_id) |
| 查询 mutation recovery | get_deepseek_recovery() |
| 确认已核验 mutation | acknowledge_deepseek_mutations(transaction_ids) |
只读、搜索、review 已存在文件且整个任务不执行命令/写文件时用 readonly；其余或不确定时用 coding。readonly job 后续需要 Bash 或写文件时，结束/取消后重新启动 coding job，不能 steering 提权。
3. 模型路由
model 只允许 flash 或 pro；不传时使用 flash。
- Flash：默认通用子代理，约 Sonnet / Terra 档。用于正常 coding、review、调查、重构、测试、批处理和常规多文件任务。
- Pro：困难任务子代理，约 Opus / Sol 档。用于复杂 debugging、架构级推理、困难多文件推理，或 Flash 已明显不足后的升级。
- 没有明确困难信号时保持 Flash；不要因为 Pro 可用就默认 Pro。
- 主 Agent 只选择
flash/pro，不要尝试控制 reasoning effort；thinking 档位由用户配置决定。 - background job 启动时冻结模型、reasoning effort 与能力。需要换模型时结束/取消当前 job，再新建 job。
4. 默认派工策略
默认派给 Flash：
- 脚本、测试、文档、CRUD、单组件 / 单 endpoint
- 批量修改、重命名、翻译、提取、ETL、日志扫描
- spec 清晰的 feature
- 单领域重构、常规多文件任务
- 静态代码/日志调查（优先 readonly）
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 10d ago First seen · 126 lines · 125 tokens per session scan A cdd5c6ccdc1f
delegate-to-deepseek is a skill published in the GitHub repository PsChina/deepseek-as-subagent (31 stars, last pushed 11d ago), licensed MIT. It adds 125 tokens to every session and 1,989 once invoked, about $0.0006 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
delegate
Delegate implementation to Stuntman's headless workers (Claude via proxy, opencode, Codex, Antigravity, or Muse). The current host plans, reviews the diff, verifies, and iterates. Use when asked to use Stuntman to delegate a task, send work to a stunt double, or run the delegate skill.
launch
Build a complete, research-backed product-launch plan. Fans out a multi-agent workflow that does cited web research on every competitor, sizes the market, finds the regulatory/trust tailwinds, ranks launch channels, then has strategists write the positioning, pricing, and a week-by-week launch playbook …
wiki
Scaffold an LLM-wiki "second brain" for the current folder in one shot — lays the vault skeleton, writes a note per project from its README/code, builds a graphify knowledge graph, and wires the graphify MCP so the current host can query prior work across projects. Auto-detects a single project vs a folder of many…
scaffold
Set up Stuntman project memory in AGENTS.md for Codex, CLAUDE.md for Claude Code, or both. Create and populate HANDOFF, STATUS, SPEC, and STRATEGY without clobbering existing content. Use when asked to scaffold project memory or set up handoff/status docs.
floor
Open a local live board of Claude Code sessions, Codex sessions, and Stuntman workers across projects, shown as Laghari Vexel avatars. Use when the user invokes /floor, asks which agents are running or need feedback, or wants to read agent conversations and prompt ready tmux sessions from one board.
usages
Show usage and limits for every stunt-double subscription on one board — Claude's 5-hour/weekly windows (live), Codex's 5-hour/weekly windows (from its last session snapshot), the DeepSeek balance behind opencode (live), and honest n/a rows for backends that expose nothing (agy, muse). Use when the user invokes…