Claude Skill

hermes-ephemeral-delegation

Trigger: broad exploration, multi-file reads, tests/builds, fresh review, or multi-step debug. Orchestrate complex work via delegate_task to protect context.

LLM Mart · 0 points · 8 0 0

Virus-scanned Reviewed automatically before listing.

Full trust report

Download gentleman-programming-gentle-ai-internal_assets_skills_hermes-ephemeral-delegation-d2e3443.zip · 2 KB

gentleman-programming/gentle-ai 6979 762 forks MIT Updated 7d ago

Part of gentleman-programming/gentle-ai — 36 skills

Install

skills CLI

npx skills add https://github.com/Gentleman-Programming/gentle-ai/tree/main/internal/assets/skills/hermes-ephemeral-delegation

Claude Code

claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install gentleman-programming-gentle-ai@llmmart

Git git clone https://github.com/Gentleman-Programming/gentle-ai.git

The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole gentleman-programming/gentle-ai collection as a plugin from our marketplace. Git is the plain clone.

Skill manifest

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_task for all complex work listed above. Do NOT execute it inline.
Workers are EPHEMERAL: each delegate_task call 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.md paths to load before work
Call delegate_task with 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_task configuration 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.

Files (gentle-ai)

references

tuning-knobs.md 1.3 KB

# Hermes Delegation Tuning Knobs

Configure these in `~/.hermes/config.yaml` under the `delegation` key, or pass per `delegate_task` call.

| Parameter | Default | Effect |
|-----------|---------|--------|
| `max_spawn_depth` | 2 | Maximum recursive delegation depth. Set to 1 to prevent workers from spawning their own workers. |
| `max_concurrent_children` | 4 | Maximum number of parallel workers active at once. |
| `max_iterations` | agent default | Iteration budget for each worker before it is forced to return. |
| `child_timeout_seconds` | agent default | Hard wall-clock timeout per worker. |
| `inherit_mcp_toolsets` | false | When true, workers inherit the parent's MCP toolsets automatically. When false (default), pass toolsets explicitly in the mission. |
| `subagent_auto_approve` | false | When true, workers auto-approve all tool calls. When false (default), workers prompt for approval on dangerous calls. |

## Mission Checklist

When `inherit_mcp_toolsets` is false (the default), every mission passed to `delegate_task` MUST explicitly list:

- Which toolsets the worker is allowed to use (file read, shell, browser, etc.)
- Which MCP servers the worker can call (engram, context7, etc.)
- Which `SKILL.md` paths the worker must load before starting work

Without this, workers start with no tools beyond their built-in defaults.

SKILL.md 3.3 KB

---
name: hermes-ephemeral-delegation
description: "Trigger: broad exploration, multi-file reads, tests/builds, fresh review, or multi-step debug. Orchestrate complex work via delegate_task to protect context."
license: Apache-2.0
metadata:
  author: gentleman-programming
  version: "1.0"
---

## 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_task` for all complex work listed above. Do NOT execute it inline.
- Workers are EPHEMERAL: each `delegate_task` call 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

1. Identify which gate applies. If none applies, skip delegation.
2. 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.md` paths to load before work
3. Call `delegate_task` with that mission.
4. Wait for the worker summary.
5. Verify the claimed output (check file existence, test result, side effect).
6. 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](references/tuning-knobs.md) — Full table of `delegate_task` configuration parameters and the explicit toolset/MCP/skill checklist for worker missions.
- [../../hermes/sdd-orchestrator.md](../../hermes/sdd-orchestrator.md) — SDD orchestrator protocol that uses this delegation standard for SDD phase work.