Agent authority
Volli checks deterministic boundaries around Session tools and its app API. It does not ask a model to judge each tool call or ask you to approve ordinary file reads, writes, and commands.
Working copy and execution
Section titled “Working copy and execution”A Session starts in its recorded working copy: a ticket worktree or the main checkout. Desktop Sessions use host-native file and command tools. Their starting directory is not a filesystem sandbox.
Commands run as your macOS user and can access files and the network with that user’s permissions. A worktree separates source changes from other tickets; it does not isolate processes or protect the rest of your Mac. Review the repository, scripts, Skills, and MCP servers you make available to an agent.
Code Mode has a program sandbox, but nested tool calls retain the Session’s ordinary authority. It cannot turn an unrestricted shell into an isolated one.
Secrets
Section titled “Secrets”Enter credentials through a person-only secret field, not the chat composer. Secret requests keep values out of prompts, transcripts, and Session events. Volli redacts registered secret values from tool output before showing that output to the model or in the transcript. Redaction is best effort; encoded or transformed values can leak. Ordinary question answers cannot submit credentials.
See Session credentials for credential requests, scopes, and the project’s Secrets controls. Provider sign-ins and MCP credentials have their own Settings fields. Local data and network access covers storage and outbound connections.
Actor permissions
Section titled “Actor permissions”Volli derives the caller from its authenticated connection, not from an actor name in a request. Per-actor policy controls coordination actions, which Ticket and Session events can wake a watch, and how much of another Session’s transcript the caller may read. Unauthenticated socket callers are read-only by default and cannot read Session transcripts.
A Session’s tool list is a separate boundary. Instructions cannot grant a tool it does not hold. Existing tools keep their names and routes; additions arrive only at supported refresh boundaries. Changing Settings does not revoke tools from a running Session. See the agent operating model and capability changes.
Browser Tabs also have a hold: one Session or the person can drive a tab at a time. Reads do not require a hold. A refused hold does not authorize the agent to bypass the Browser tools.
Questions and limits
Section titled “Questions and limits”Agents can ask questions and present existing confirmations. When a Session reaches its delegation allowance, policy can refuse another helper or ask through a permission card. An unattended Run does not supply an answer for you.
Decision models and classify help with typed
decisions. They do not approve calls. Old review records and saved grants in
Session history do not authorize new tool calls.