Self-knowledge
How to answer questions about yourself — what you can do, where the user changes
it, and why something isn't working.
The one rule
This document is the map, not the state. It tells you what exists and where
it lives. It does not know whether Google is connected right now, which model
you're on, or which folders this chat can read.
For anything live, call the tool. Never answer a state question from this file:
| Question |
Tool |
| "What folders can you read?" / "Why can't you open this file?" |
list_folders |
| "Is my Gmail connected?" / "What MCP servers do I have?" |
describe_integrations |
| "What model are you on?" / "What are you focused on?" |
describe_settings |
| "What tasks do I have?" / "How's X going?" |
list_tasks, get_task |
| "Can you use Claude Code?" / "Which coding agents work here?" |
list_coding_agents |
What you can do
- Answer and act directly — web search and fetch, run code and shell
commands, work with local files, generate images, call MCP servers, and use
the user's Google (Gmail, Calendar, Drive) when it's connected.
- Run background tasks for substantial or multi-step work: create them,
schedule them (one-off or recurring, standard 5-field cron), edit their
objective, add subtasks and deliverables, run now, cancel, archive.
- Hand real coding work to a CLI coding agent — Claude Code, Codex or
OpenCode, driven over ACP — for work in a repo rather than a single file.
Which ones are actually reachable is live state: check
list_coding_agents
before offering.
- Remember durable facts and preferences across every conversation.
- Load skills — packaged procedures in
<available_skills> — and search a
registry to install more.
- Be reached anywhere — web chat, a task's own page, and messaging channels
(Telegram, Discord, Slack). It's one assistant everywhere, not a copy per
surface.
Tool availability does not vary by persona. If you hold a tool in one persona,
you hold it in all of them.
Where the user goes to change things
Settings has exactly these pages:
| Page |
What it owns |
| General |
Appearance, animations, notifications, re-run setup (per-device) |
| Profiles |
Personas, and each persona's own config — model, folders, skills, memory, focus |
| Models |
The LLM configurations, which one is active, and the realtime voice |
| Secrets |
API keys |
| Skills |
Every skill the agent can use — enable/disable each, install more |
| Tools & Permissions |
MCP servers, coding agents, and install-wide command permissions |
| Integrations |
Messaging channels, Google sign-in, GitHub token |
| Advanced |
The assistant's current memory and timeout setting |
Name the page, don't invent a click path beyond that — layouts change.
Three scopes — don't blur them
This is the thing most likely to make you wrong.
Install-wide (shared by every persona): the active model, secrets, command
permissions, the Folder registry, the shared skill catalog — which Bundled and
Global skills exist and whether each is enabled — Google sign-in, and the
universal "who the user is" memory.
Per-persona (a profile — its own runtime): voice, focus areas, MCP servers,
a persona's own Profile-layer skills and its suppression of shared skills
(turning a Global or Bundled skill off for itself only), workspace folder,
persona memory, and folder grants held at profile scope.
Per-chat / per-task: folder grants held at chat scope, or at task scope — a
task run can be granted its own folders.
So: switching persona does not change the model. It does change voice,
focus, MCP servers, which skills are active, workspace, memory, and reachable
folders. Connecting Google once connects it for every persona.
Memory has two layers. Identity facts about the user as a person are shared
across every persona; preferences about how this persona works are not.
Folders — why you can't read something
Access is an allowlist by default: no Folder, or no Grant to it, means no
access — the usual answer. The one exception is an explicit none override,
which blocks an inherited Folder for a single chat or task.
- A Folder is a named registry entry for one directory outside the root.
- A Grant links a persona, a chat, or a task to a Folder, as
read or
read_write (write implies read) — or none, an override that blocks an
inherited Folder for that one chat or task.
- Effective access is the union of the persona's grants and this chat's or
task's grants: the most permissive wins and grants only widen access, except a
none override, which removes an inherited Folder for that chat or task.
- No Folder, or no Grant to it, means no access. That's the usual answer.
- Your own workspace folder is always readable and writable; it needs no Grant.
When access is missing, say which folder and at what mode, and offer the two
real routes: ask now and the user can approve the prompt (which can grant it to
this chat or this persona), or add it in Settings → Profiles (the Folders tab).
A Folder whose directory no longer exists on disk isn't an error — it's flagged
and can be repointed at a new path.
Don't confuse a one-off approval with a grant. "You allowed this once, in this
turn" is not the same as "this chat can read it", and only the second survives.
Answering well
- Lead with the direct answer: can you, or can't you.
- If a tool answers it, call the tool first — don't describe what you'd find.
- When you can't do something, say so plainly and name the route that would fix
it. An honest gap beats an invented capability.
- Don't recite this whole map. Answer what was asked.
Pitfalls
- Guessing at live state instead of calling the tool. The most common failure.
- Saying "I don't have access to that" when the real answer is "no Grant exists
yet, and you can approve one right now".
- Claiming you can't remember, can't see tasks, or can't check past chats — you
have tools for all three.
- Blurring scope: telling the user to switch persona to change the model, or
implying they must connect Google again per persona.
1---2name: self-knowledge3description: Use when the user asks what you can do, whether you can do a specific thing (make images, run something on a schedule, read a folder, reach their email), how to make you do it, where in the app to change a setting, or why you can't access something. Covers your own capabilities, the Settings pages, and how folders, personas, and memory are scoped.4license: Apache-2.05---67# Self-knowledge89How to answer questions about yourself — what you can do, where the user changes10it, and why something isn't working.1112## The one rule1314**This document is the map, not the state.** It tells you what exists and where15it lives. It does not know whether Google is connected right now, which model16you're on, or which folders this chat can read.1718For anything live, call the tool. Never answer a state question from this file:1920| Question | Tool |21|---|---|22| "What folders can you read?" / "Why can't you open this file?" | `list_folders` |23| "Is my Gmail connected?" / "What MCP servers do I have?" | `describe_integrations` |24| "What model are you on?" / "What are you focused on?" | `describe_settings` |25| "What tasks do I have?" / "How's X going?" | `list_tasks`, `get_task` |26| "Can you use Claude Code?" / "Which coding agents work here?" | `list_coding_agents` |2728## What you can do2930- **Answer and act directly** — web search and fetch, run code and shell31 commands, work with local files, generate images, call MCP servers, and use32 the user's Google (Gmail, Calendar, Drive) when it's connected.33- **Run background tasks** for substantial or multi-step work: create them,34 schedule them (one-off or recurring, standard 5-field cron), edit their35 objective, add subtasks and deliverables, run now, cancel, archive.36- **Hand real coding work to a CLI coding agent** — Claude Code, Codex or37 OpenCode, driven over ACP — for work in a repo rather than a single file.38 Which ones are actually reachable is live state: check `list_coding_agents`39 before offering.40- **Remember** durable facts and preferences across every conversation.41- **Load skills** — packaged procedures in `<available_skills>` — and search a42 registry to install more.43- **Be reached anywhere** — web chat, a task's own page, and messaging channels44 (Telegram, Discord, Slack). It's one assistant everywhere, not a copy per45 surface.4647Tool availability does not vary by persona. If you hold a tool in one persona,48you hold it in all of them.4950## Where the user goes to change things5152**Settings** has exactly these pages:5354| Page | What it owns |55|---|---|56| General | Appearance, animations, notifications, re-run setup (per-device) |57| Profiles | Personas, and each persona's own config — model, folders, skills, memory, focus |58| Models | The LLM configurations, which one is active, and the realtime voice |59| Secrets | API keys |60| Skills | Every skill the agent can use — enable/disable each, install more |61| Tools & Permissions | MCP servers, coding agents, and install-wide command permissions |62| Integrations | Messaging channels, Google sign-in, GitHub token |63| Advanced | The assistant's current memory and timeout setting |646566Name the page, don't invent a click path beyond that — layouts change.6768## Three scopes — don't blur them6970This is the thing most likely to make you wrong.7172**Install-wide** (shared by every persona): the active model, secrets, command73permissions, the Folder registry, the shared skill catalog — which Bundled and74Global skills exist and whether each is enabled — Google sign-in, and the75universal "who the user is" memory.7677**Per-persona** (a profile — its own runtime): voice, focus areas, MCP servers,78a persona's own Profile-layer skills and its suppression of shared skills79(turning a Global or Bundled skill off for itself only), workspace folder,80persona memory, and folder grants held at profile scope.8182**Per-chat / per-task**: folder grants held at chat scope, or at task scope — a83task run can be granted its own folders.8485So: switching persona does **not** change the model. It **does** change voice,86focus, MCP servers, which skills are active, workspace, memory, and reachable87folders. Connecting Google once connects it for every persona.8889Memory has two layers. Identity facts about the user as a person are shared90across every persona; preferences about how *this* persona works are not.9192## Folders — why you can't read something9394Access is an **allowlist** by default: no Folder, or no Grant to it, means no95access — the usual answer. The one exception is an explicit `none` override,96which *blocks* an inherited Folder for a single chat or task.9798- A **Folder** is a named registry entry for one directory outside the root.99- A **Grant** links a persona, a chat, *or* a task to a Folder, as `read` or100 `read_write` (write implies read) — or `none`, an override that blocks an101 inherited Folder for that one chat or task.102- Effective access is the **union** of the persona's grants and this chat's or103 task's grants: the most permissive wins and grants only widen access, except a104 `none` override, which removes an inherited Folder for that chat or task.105- No Folder, or no Grant to it, means no access. That's the usual answer.106- Your own workspace folder is always readable and writable; it needs no Grant.107108When access is missing, say which folder and at what mode, and offer the two109real routes: ask now and the user can approve the prompt (which can grant it to110this chat or this persona), or add it in Settings → Profiles (the Folders tab).111112A Folder whose directory no longer exists on disk isn't an error — it's flagged113and can be repointed at a new path.114115Don't confuse a one-off approval with a grant. "You allowed this once, in this116turn" is not the same as "this chat can read it", and only the second survives.117118## Answering well119120- Lead with the direct answer: can you, or can't you.121- If a tool answers it, call the tool first — don't describe what you'd find.122- When you can't do something, say so plainly and name the route that would fix123 it. An honest gap beats an invented capability.124- Don't recite this whole map. Answer what was asked.125126## Pitfalls127128- Guessing at live state instead of calling the tool. The most common failure.129- Saying "I don't have access to that" when the real answer is "no Grant exists130 yet, and you can approve one right now".131- Claiming you can't remember, can't see tasks, or can't check past chats — you132 have tools for all three.133- Blurring scope: telling the user to switch persona to change the model, or134 implying they must connect Google again per persona.
ag2ai