hermes - 💡(How to fix) Fix Feature: Add `skills` parameter to `delegate_task` (like `cronjob` already has) [1 participants]

Summary

The cronjob tool already supports a skills parameter that loads specified skills before executing a scheduled job. The delegate_task tool has no equivalent — subagents start with completely blank context, no access to the skill library. The parent must manually copy relevant skill guidance into the context field, which is error-prone, token-wasteful, and fragile.

This led to a real failure mode today: a research subagent given toolsets=["web"] fell into a 42-call search loop because it could only search, not navigate URLs. Our delegate-task-guide skill documents the toolset selection pitfall, but the subagent had no way to load it.

Current Behavior

Proposed Behavior

Add an optional skills parameter to delegate_task, mirroring the existing cronjob pattern:

delegate_task(
    goal="Analyze the source code at these URLs...",
    toolsets=["browser", "terminal", "file"],
    skills=["delegate-task-guide", "read-write-safety"],
    context="Fetch https://github.com/example/repo/blob/main/src/index.ts and summarize..."
)

• Skills are loaded into the subagent context before execution (same as cronjob)
• Empty list or omitted means no skills (preserves current default behavior)
• Explicit skills=[] could confirm "deliberately no skills" for clarity

Why Not Auto-Inherit All Parent Skills

That would bloat every subagent context with ~20+ skill schemas, most of which are unused for any given task. Explicit opt-in keeps subagent context lean while still providing access to critical skill knowledge. The cronjob approach is already opt-in and works well.

Real-World Failure (Today)

A delegate_task call was made to analyze a GitHub project:

delegate_task(
    description="Analyze A2A plugin source code",
    goal="Fetch and summarize the source code...",
    toolsets=["web"]  # ← Wrong toolset
)

The subagent could only use SearXNG search and Reddit tools (no browser, no terminal). It couldn't navigate to the GitHub URLs provided. Instead, it searched for them — 42 search API calls, zero source files actually fetched. It kept searching with different query variants, never finding what it needed because the tool didn't support URL navigation.

If the subagent had been able to load the delegate-task-guide skill, it would have had guidance on toolset selection and an explicit exit condition ("after 3 failed attempts, stop"). Instead, it had no access to that knowledge and looped until the parent interrupted it.

Benefits

1. Consistency — Both async (cronjob) and sync (delegate_task) subagents use the same skill-loading pattern
2. Token efficiency — Skills are loaded once, not manually duplicated into context
3. Freshness — Skills stay current (loaded at execution time), not stale (manually copied weeks ago)
4. Discoverability — Parent agents already know their skill names; passing them is natural
5. Pitfall prevention — Task-specific skills like delegate-task-guide, read-write-safety, and systematic-debugging can prevent entire categories of subagent failures

Environment

• Hermes Agent: v0.12.0 (2026-4-30), 39 commits behind latest
• OS: Ubuntu 25.10 (Questing Quokka), kernel 6.17.0-23-generic, x86_64
• CPU: 12th Gen Intel i7-12700H
• GPU: NVIDIA RTX 3060 Laptop 6GB, driver 595.58.03
• RAM: 30 GB
• Python: 3.13.7
• Model: glm-5.1:cloud via custom provider (Ollama)
• Enabled toolsets: web, browser, terminal, file, delegation, cronjob, skills, memory, session_search, todo, clarify, messaging, vision, image_gen, tts, homeassistant, reddit, robloxstudio, searxng
• Installed skills: 130
• Profile: default (Hazel)

Related Code

• cron/scheduler.py — existing skills parameter handling for cron jobs (reference implementation)
• tools/delegate_task.py — current delegate_task tool definition (needs the new parameter)
• agent/skill_commands.py — skill loading logic (could be reused)
• skills/ directory — skill discovery and loading infrastructure already exists