hermes - 💡(How to fix) Fix Orchestrator brain won't reliably delegate coding to the claude-code skill — no supported tool-routing mechanism

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…
RAW_BUFFERClick to expand / collapse

Setup

Hermes v0.15.1 in an orchestrator/worker config: a Gemini brain (google provider) for coordination, with the built-in claude-code skill (Claude Code CLI, print mode) as the intended coding worker. The goal is exactly the pattern the claude-code SKILL.md itself describes — the orchestrator delegates coding to Claude Code rather than writing code itself.

Problem

The brain will not reliably delegate. Given a coding task it uses its own write_file / terminal tools inline instead of invoking claude -p via the skill — even with an explicit SOUL.md instruction (including the exact command) telling it to delegate all non-trivial coding to claude-code.

What I tried

  1. SOUL.md persona instruction — insufficient; the brain codes inline. Tested across gemini-3.5-flash and gemini-2.5-pro, ~4 coding asks, zero delegations.
  2. delegate_task — spawns child AIAgent instances on the same brain model, not the claude-code CLI. Wrong target for this.
  3. pre_tool_call blocking hook — appears to be the only lever, but looks fragile: the claude-code skill runs claude -p through the terminal tool, so naively blocking terminal/write_file risks deadlocking the very call that performs the delegation.

What seems missing

A supported, declarative way to route a task class to a specific skill/worker — e.g. "for coding tasks, prefer the claude-code skill over the built-in write_file/terminal tools," or a first-class "preferred coding agent" setting. Persona-steering is too soft, and delegate_task can't target an external CLI skill.

Questions

  1. Is there a supported way to make the brain prefer a skill (like claude-code) for a task type, short of a pre_tool_call hook?
  2. Is reliable skill-delegation expected to be model-dependent? gemini-2.5-pro was no better than gemini-3.5-flash here.
  3. Is a pre_tool_call hook redirecting coding to claude-code the recommended pattern — and if so, is there guidance for avoiding the terminal-tool deadlock?

Environment

  • Hermes v0.15.1, Ubuntu 22.04
  • google-provider brain (gemini-2.5-pro / gemini-3.5-flash)
  • Built-in claude-code skill enabled; skill_manage disabled

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING