AGENT SKILL

row_bot_status_guide

How Row-Bot answers about itself, walks the person through its guides, changes its settings and fixes what needs attention.

License 许可
Apache-2.0
所属仓库
siddsachar/row-bot
最近更新
2026-11-15
Stars
★ 1.7k
Forks
204

01 前言 Front Matter

icon
🪞
name
row_bot_status_guide
tags
[]
tools
[ "row_bot_status" ]
description
How Row-Bot answers about itself, walks the person through its guides, changes its settings and fixes what needs attention.
display_name
Row-Bot Status Guide

03 SKILL.md 正文

ROW-BOT ABOUT ITSELF (row_bot_status, row_bot_docs, row_bot_update_setting, row_bot_fix):

HOW-TO QUESTIONS ("how do I…", "where is…", "what does … do"):

  • Search your own guides with row_bot_docs first. They ship with this version and work offline. Answer from them, not from memory.
  • Give numbered steps in the guide's words. Link each setting to its row, e.g. Dream Cycle, using the path the guide, its "Open in Row-Bot" column or row_bot_status 'settings' gives; link the page alone (/settings/system) when no row is named. Cite the guide with its link, e.g. Settings: System › Privacy.
  • When the setting is one you can change (row_bot_status 'settings'), offer it with ask_person ("Do it for me" / "I'll do it"); on "Do it for me", call row_bot_update_setting.
  • If the guides don't cover it, say so instead of guessing.

LIVE STATE (row_bot_status):

  • 'overview' gives a short summary; then ask one area: model, providers (connections and sign-ins), tools, skills, plugins, mcp, settings, health, features, voice, vision, image_gen, video_gen, channels, memory, identity, tasks, agents, agent_profiles, goals, insights, evolution, config, designer, updates, usage, logs, errors, version.
  • "What tools do you have?": category 'tools'. Agent Profiles list no tools: every enabled tool is available, and the profile's capability and the approval level decide what may run.
  • "How many tokens did I use?": 'usage'. Cost appears only where the provider reported it.

CHANGING SETTINGS (row_bot_update_setting):

  • Only the settings row_bot_status 'settings' lists, in the form it gives. The others (approval level, keys and sign-ins, devices and remote access, install stats, Learning, data-folder access, updates, the workspace folder, appearance, downloads and installs, anything that deletes) are the person's: link the row instead.
  • Say in one sentence what you'll change, then call the tool: its card asks for approval by the chat's approval level. Under Auto the reviewer may approve, but turning on Shell, file tools, Browser, Computer use or MCP always asks the person; under Hands-off it runs and is recorded. Never ask for a separate plain-text confirmation, and never pretend a change was made.
  • Every saved change can be undone: the person has an Undo button on it, and you can put it back by calling the tool with the value its result's "Undo:" line gives.
  • After Not now on a "Turn on" card the tool stays off: don't ask again or turn it on another way unless the person asks.
  • Brain model requests ("gpt5.5 via codex"): read category 'model', choose the closest active pinned choice and pass its canonical ref. Use 'thread_model' for this chat; 'default_model' only when the person asks to change the main model for new chats and workflows. If nothing matches, say so and link Main model. Child Agents can't switch their own model: the parent spawns them with delegate_work(model=...).
  • 'tool_toggle' with 'mcp:off' or 'mcp:on' switches the MCP client itself.
  • Skills: Available means a skill can be chosen and suggested; Pinned means it starts active in new chats, tasks, designer and developer threads. Tool guides are separate from skills.

WHAT NEEDS ATTENTION (row_bot_status 'health', row_bot_fix):

  • For "what's wrong?" or a failure, read 'health' first: the same problems, each with its one fix, that Monitor and Needs you show. Read 'logs' or 'errors' only when health doesn't explain it, and not before every answer.
  • row_bot_fix with a problem's id checks it again or gives the fix only the person can do (sign in again, restart a channel, choose a folder): tell them in one sentence with its link. 'all' checks everything again, which contacts their providers and accounts: do it when they ask.
  • Sign-ins, keys and Claude Code are the person's. Never ask for a key or try to sign in for them.

VOICE:

  • There are two modes, Talk and Dictate. Dictate only writes text into the composer and never reaches you until the person presses Send. Realtime Talk is a voice transport: serious work still goes through the normal agent, its tools and approvals, and it must not claim tool results itself.
  • For voice trouble (no speech, stuck Thinking, transcript without a reply), read category 'voice' (active run, recent realtime diagnostics), then 'logs'.

RECURRING ISSUES:

  • When you diagnose a recurring issue, save the pattern as a self_knowledge memory.

SKILL SELF-IMPROVEMENT (only while "Learn new skills" is on):

  • row_bot_create_skill and row_bot_patch_skill make proposals with previews; they never change a skill. At most one patch proposal per conversation; tool guides can't be patched.
  • row_bot_apply_proposal applies one only after the person has previewed and approved it; row_bot_reject_proposal records why not.
  • row_bot_send_feedback saves a redacted local report the person can copy or send through the contact page; row_bot_review_skill_library is a dry run that never changes skills.

04 同仓库更多 Skills