Agent Skills
› Gentleman-Programming/gentle-ai
› hermes-ephemeral-delegation
hermes-ephemeral-delegation
GitHub作为父编排器，通过delegate_task将复杂任务（如多文件操作、测试构建、深度调试）委派给临时子代理，以保护主上下文并高效协调工作流。
触发场景
需要探索多个文件或进行代码库映射
涉及多个非平凡文件的实现或修改
执行测试套件或构建流程
对代码差异进行独立审查
多步骤调试导致上下文过载
安装
npx skills add Gentleman-Programming/gentle-ai --skill hermes-ephemeral-delegation -g -y
SKILL.md
Frontmatter
{
"name": "hermes-ephemeral-delegation",
"license": "Apache-2.0",
"metadata": {
"author": "gentleman-programming",
"version": "1.0"
},
"description": "Trigger: broad exploration, multi-file reads, tests\/builds, fresh review, or multi-step debug. Orchestrate complex work via delegate_task to protect context."
}
Activation Contract
Load this skill when you are acting as the parent orchestrator and the work ahead falls into any of these categories:
- Broad exploration (4+ files to understand, codebase mapping, approach comparison)
- Multi-file implementation (touching 2+ non-trivial files)
- Test or build execution
- Fresh adversarial review (diffs, PR readiness, incident audit)
- Multi-step debugging that would flood the parent context
Do NOT load this skill if you are already inside a delegated child task — you are the executor, not the orchestrator.
Hard Rules
- Use
delegate_taskfor all complex work listed above. Do NOT execute it inline. - Workers are EPHEMERAL: each
delegate_taskcall creates a fresh context. Do NOT request persistent agent files or profiles. - Pass a self-contained mission. Workers have no memory of the parent conversation.
- Treat worker output as self-report: verify file writes, test pass/fail, URLs, and IDs before reporting success to the user.
- Batch parallel calls only for INDEPENDENT workstreams. Sequential dependencies must run sequentially.
Decision Gates
| Situation | Action |
|---|---|
| Need to read 4+ files to understand | Delegate a narrow exploration worker |
| Need to write 2+ non-trivial files | Delegate a single writer with the full mission |
| Need to run tests or builds | Delegate an executor; do not run inline |
| Need an adversarial review of a diff | Delegate a fresh-context reviewer |
| Multi-step debug that grows the context | Delegate a debug worker; feed results back inline |
| Simple 1-file edit you already understand | Do it inline; no delegation needed |
| Quick git/state check | Do it inline; no delegation needed |
Execution Steps
- Identify which gate applies. If none applies, skip delegation.
- Draft a self-contained mission for the worker — include:
- Exact goal (one sentence)
- File paths or targets to act on
- Relevant prior context the worker needs (decisions, conventions, prior findings)
- Constraints (style, test runner, budget)
- Expected evidence to return (e.g., file written, test output, URL found)
- Allowed toolsets/MCP/skills the worker should use
- Any
SKILL.mdpaths to load before work
- Call
delegate_taskwith that mission. - Wait for the worker summary.
- Verify the claimed output (check file existence, test result, side effect).
- Synthesize the verified result into your orchestrator reply.
Output Contract
After synthesizing worker results, return:
- What was delegated and to how many workers
- What each worker returned (verified, not just claimed)
- Any discrepancy between worker self-report and verified evidence
- Final answer or next step for the user
References
- references/tuning-knobs.md — Full table of
delegate_taskconfiguration parameters and the explicit toolset/MCP/skill checklist for worker missions. - ../../hermes/sdd-orchestrator.md — SDD orchestrator protocol that uses this delegation standard for SDD phase work.
版本历史
- e01b114 当前 2026-07-25 06:59
