slash-command-routing

name	slash-command-routing
description	Route user intent to Hermes slash commands before normal tool or skill execution. Use when the request is really a session-control, backgrounding, planning, queueing, resume, rollback, config, or gateway command trigger.

Slash Command Routing

Goal

• Treat Hermes slash commands as a first-class routing surface.
• When the user intent is really a slash command, prefer the slash command path over ad hoc tool behavior.

Use when

• The user asks for background execution, session control, queueing, plan mode, rollback, resume, status, model/provider toggles, or other Hermes command-surface actions.
• The user mentions words like 后台, slash, 斜杠命令, 用 /background, 先计划别执行, 恢复会话, 撤销刚才, 压缩上下文, 排队, 暂停提问, 快模式, reasoning, yolo, reload mcp, debug report.

Core rule

• If a built-in slash command matches the user intent well, route to that command surface first instead of approximating the behavior with ordinary tool calls.
• For long-running autonomous work that should continue outside the current turn, prefer /background over only starting a raw background process.

Primary mappings

1. Background / asynchronous work

Trigger phrases:

• 后台
• 后台跑
• 后台监控
• 继续盯
• 跑完回报
• 必要时介入
• 先后台
• /background

Route:

• /background <self-contained prompt>

Rules:

• The prompt must be self-contained: goal, workdir, commands, state files, expected output, failure handling.
• Do not use /background for trivial one-shot checks.
• If the user wants the current session free while work continues, /background is the default.

2. Plan mode instead of execution

Trigger phrases:

• 先计划
• 别执行
• 写方案
• plan mode
• /plan

Route:

• /plan [request]

3. Queue next prompt

Trigger phrases:

• 排队
• 下一轮再做
• 等当前完成后再问
• /queue

Route:

• /queue <prompt>

Important:

• Prefer /queue, not /q, because /q conflicts with /quit in practice.

4. Session reset / branch / resume / undo / retry

Trigger phrases and routes:

• 新开会话 / 重开 / 清空重来 -> /new or /clear
• 撤销刚才 -> /undo
• 重试上一条 -> /retry
• 恢复会话 / 继续之前那个 -> /resume [name]
• 分支试另一条路 / fork this -> /branch [name]
• 压缩上下文 / 对话太长 -> /compress [focus topic]
• 回滚 checkpoint / 恢复快照 -> /rollback or /snapshot

Guideline:

• /rollback is for filesystem checkpoints.
• /snapshot is for Hermes config/state snapshots.
• Distinguish them explicitly.

5. Status / usage / debug

Trigger phrases and routes:

• 状态 / 当前会话 -> /status
• token 用量 / 费用 / 成本 -> /usage
• debug report / 上传日志 -> /debug
• 近 30 天分析 -> /insights

6. Model / provider / speed / reasoning / personality

Trigger phrases and routes:

• 切模型 -> /model
• 看 provider -> /provider
• 快一点 / fast mode -> /fast
• reasoning 高一点/低一点 -> /reasoning
• 换人格 -> /personality

7. Tools / browser / MCP / skills / cron / reload

Trigger phrases and routes:

• 开关工具 -> /tools
• 看 toolsets -> /toolsets
• 接浏览器 -> /browser
• 技能搜索/安装/审计 -> /skills
• 定时任务 -> /cron
• 重载 MCP -> /reload-mcp
• 重载 env -> /reload
• 插件状态 -> /plugins

8. High-risk explicit override

Trigger phrases:

• 别再问审批
• 全自动
• YOLO
• /yolo

Route:

• /yolo

Safety:

• Only route here when the user explicitly asks.
• Do not infer it from impatience alone.

9. Messaging-only commands

When operating in chat/gateway contexts, these are first-class and should be recognized:

• /approve
• /deny
• /commands
• /update
• /restart
• /sethome

Do not suggest them in pure CLI-only contexts unless relevant.

Source of truth

• Slash command reference: ~/.hermes/hermes-agent/website/docs/reference/slash-commands.md
• Central registry: ~/.hermes/hermes-agent/hermes_cli/commands.py

Router discipline

1. First ask: is this really a slash-command intent?
2. If yes, route to the slash command before generic tool use.
3. If no, continue normal skill/tool routing.
4. For slash-command routing updates, keep this skill and .hermes/routing/slash-command-router.md in sync.

Pitfalls

• Do not approximate /background with only a raw background process when the user wants autonomous monitoring.
• Do not use /q; use /queue explicitly.
• Do not confuse /rollback with /snapshot.
• Do not route to /yolo without explicit user intent.