All plugins

super-agent

Runs the Hermes chat agent like a secretary with a team of workers: keeps chat turns responsive (60s reminder, 120s edit/execute pause, 30s terminal auto-detach), briefs every subagent before it starts, stops subagents that repeat identical results, brings open decisions back to the main agent, has it ask the user after repeated failed subagents, never takes a subagent that ended without its report as done, tracks background results until they reach the user, and records every delegated task in a SQLite ledger across gateway restarts (hermes super-agent status shows it). Governance is a positive allowlist: only the platforms named in governed_platforms (default weixin) and the subagents they spawn are affected, and every other surface runs untouched. Disclosure — registers hooks for pre_llm_call, pre_api_request, pre_auxiliary_call, post_auxiliary_call, pre_tool_call, post_tool_call, transform_tool_result, subagent_start, subagent_stop, on_session_end, on_session_reset and on_session_finalize, a system-prompt section and a tool_execution middleware, so for governed sessions it can block or amend tool calls and results, LLM request context and subagent handoffs, and asks a governed foreground terminal command still running after 30 s to move to Hermes's background process registry (through core's internal yield flag, which a Hermes release may change); it copies your Hermes USER.md and MEMORY.md memory files in full into every governed delegate_task brief, can start an internal agent turn via inject_message to re-report an unseen subagent result, writes a local ledger.db under the plugin's Hermes data directory, and at startup reads the effective Hermes config through Hermes' own resolver to tell the main agent once about settings that weaken it (inline delegation, a high delegation.max_iterations); it also registers a read-only hermes super-agent status CLI command that prints its settings and the ledger's recent tasks; no network calls, no agent tools, no credentials, no shell commands.

JerryLiu3695Automationv0.2.2Updated 1 day ago

About

Runs the Hermes chat agent like a secretary with a team of workers: keeps chat turns responsive (60s reminder, 120s edit/execute pause, 30s terminal auto-detach), briefs every subagent before it starts, stops subagents that repeat identical results, brings open decisions back to the main agent, has it ask the user after repeated failed subagents, never takes a subagent that ended without its report as done, tracks background results until they reach the user, and records every delegated task in a SQLite ledger across gateway restarts (hermes super-agent status shows it). Governance is a positive allowlist: only the platforms named in governed_platforms (default weixin) and the subagents they spawn are affected, and every other surface runs untouched. Disclosure — registers hooks for pre_llm_call, pre_api_request, pre_auxiliary_call, post_auxiliary_call, pre_tool_call, post_tool_call, transform_tool_result, subagent_start, subagent_stop, on_session_end, on_session_reset and on_session_finalize, a system-prompt section and a tool_execution middleware, so for governed sessions it can block or amend tool calls and results, LLM request context and subagent handoffs, and asks a governed foreground terminal command still running after 30 s to move to Hermes's background process registry (through core's internal yield flag, which a Hermes release may change); it copies your Hermes USER.md and MEMORY.md memory files in full into every governed delegate_task brief, can start an internal agent turn via inject_message to re-report an unseen subagent result, writes a local ledger.db under the plugin's Hermes data directory, and at startup reads the effective Hermes config through Hermes' own resolver to tell the main agent once about settings that weaken it (inline delegation, a high delegation.max_iterations); it also registers a read-only hermes super-agent status CLI command that prints its settings and the ledger's recent tasks; no network calls, no agent tools, no credentials, no shell commands.

Hermes Super Agent

Runs your Hermes chat like a good secretary with a team of workers. The main agent stays reachable in the chat while subagents do the slow work in the background, and every delegated job ends with a result that is checked, reported to you and recorded. Plain Python: no model calls of its own, and it never writes into your chat.

Left, a chat in which the assistant hands a test run to a subagent, answers another question meanwhile, then reports the checked result. Right, the briefing attached to the subagent's task and the note the main agent receives when the subagent finishes.

Install

hermes plugins install JerryLiu369/hermes-super-agent/super-agent
hermes plugins enable super-agent

Optional, in ~/.hermes/config.yaml (restart the gateway after changing it):

plugins:
  entries:
    super-agent:
      allow_gateway_injection: true # lets the plugin wake the main agent to finish a report you interrupted

By default it governs weixin chats and the subagents they spawn; set governed_platforms to choose others.

What it does

  • The chat stays responsive. At 60 seconds in one turn the main agent is reminded to answer or delegate, and at 120 seconds editing and running commands pause for the rest of the turn. A terminal command still running after 30 seconds moves to the background.
  • Subagents start informed. A brief under 120 characters, or one naming a file another running subagent is changing, goes back once. Every accepted task gets your last 6 messages, USER.md and MEMORY.md, the chat's task record, working rules and how to report. When you write while subagents run, the main agent is reminded which ones, so a change reaches them with a steer.
  • Subagents do not spin or guess. The same step with the same result draws a warning the 3rd time and a stop the 5th. A choice that is not the subagent's to make comes back as a needs_decision question with options. After 2 subagents end without finishing since your last message, the next dispatch goes back once so the main agent asks you first.
  • Results are checked, reported and remembered. The main agent is told to check a result's evidence before passing it on (a subagent that ends without its report is not taken as done), a background result stays unreported until you have heard it, and every task is kept in a SQLite ledger that survives a gateway restart.

hermes super-agent status shows the settings in effect and the latest tasks. Settings, the full explanation and a FAQ are in the README (简体中文).

MIT licensed.

README from the reviewed commit 260b449; it updates when the author re-pins.

Developer Mode

Explore built-in Hermes toolsets and actions.

HermesAgent

Developer?

Create a Hermes extension for Desktop or agent workflows.

Developer Section