Glossary

Definitions of the terms you meet in the OpenInspect app and documentation, in alphabetical order.

Agent harness The program that runs the model inside the sandbox. OpenInspect offers OpenCode (the default, any model family) and Claude Agent (Anthropic models only). The harness is fixed when a session is created and child sessions inherit it. See Agent harnesses.

Artifact An output the session records beside its messages: a pull request, a branch, a screenshot, a video, or a preview. Pull requests and branches appear in the sidebar; screenshots and videos appear under Media. See Reviewing changes.

Automation A saved instruction that starts sessions without a person present, fired by a schedule, an inbound webhook, a GitHub event, a Slack message, or a Sentry alert. See Automations.

Boot budget The 30-minute default allowance from sandbox creation until the sandbox is ready, covering cloning, setup.sh, start.sh, skills, and agent start. When it runs out the sandbox fails and the waiting prompt fails with the phase that was running. See Lifecycle and statuses.

Child session A session an agent spawns with spawn-child to work on part of a task in its own sandbox and branch. It inherits the parent's harness, provider auth, and skills. Limits come from the sandbox settings. See Child sessions.

Environment A named, reusable ordered set of up to 10 repositories with its own secrets scope and optional prebuilt image, managed under Settings › Environments. Sessions snapshot the environment at creation, so later edits do not change existing sessions. See Environments.

Fingerprint (prebuilt image) The identity of the repository set an image was built for: the ordered repositories and their base branches. A session uses an image only when its own set has the same fingerprint, and editing an environment retires the old image. See Prebuilt images.

Follow-up Any prompt sent to a session after its first one. Follow-ups queue behind the running turn and are processed in order. See Follow-ups and stopping.

Inbox The sidebar view of your sessions grouped into Needs attention, In progress, and Recent, with an Unread badge per viewer. See Inbox and search.

Managed skill Reusable instructions and supporting files stored in OpenInspect and installed into a session before the agent starts. Skills are assigned globally or to repositories and environments, and the revisions are pinned when the session is created. See Managed skills.

MCP server A Model Context Protocol server the agent can call as a tool source, configured under Settings › MCP Servers as a local command or a remote URL, for all repositories or selected ones. See MCP servers.

Multi-repository session A session whose sandbox holds several repositories cloned side by side under /workspace, either an ad-hoc "Multiple repositories" pick or an environment. Pushes and pull requests are per repository. See Core concepts.

Participant A user who has joined a session. Participants see the same event stream, and each prompt is attributed to the participant who sent it. Participant labels record who contributed; they are not an access list. See Multiplayer and sharing.

Prebuilt image A provider image built ahead of time for a repository or environment with the clone and setup.sh already done. Sessions that match its fingerprint boot from it and skip setup. Rebuilt on a 30-minute schedule when commits land. See Prebuilt images.

Primary repository The first repository in a multi-repository session or environment. It drives which settings apply and wins secret key collisions in ad-hoc sets. See Environments.

Provider account A connected OpenAI, xAI (SuperGrok), or Claude subscription stored installation-wide under Settings › Accounts. Sessions bound to an account receive short-lived access tokens instead of an API key. See Provider accounts.

Reasoning effort A per-session or per-message setting that selects how much reasoning a model applies. Each model lists the efforts it supports and a default. See Choosing a model.

Routing rule (Slack) A keyword-to-repository mapping under Settings › Integrations › Slack › Routing rules. A whole-word, case-insensitive match routes a Slack request straight to that repository, ahead of the channel's default. See Slack.

Run (automation) One session started by an automation firing. A multi-repository firing has one run per repository, and the firing row summarises them with statuses such as Completed, Failed, or Partial failure. See Automations.

Sandbox The isolated development container a session's agent works in. A session has many sandboxes over its lifetime; the sandbox status (Ready, Stopped, and so on) is separate from the session status. See Lifecycle and statuses.

Sandbox provider The backend that runs sandboxes for a deployment: Modal, Vercel, OpenComputer, E2B, or Daytona. Providers differ in how they resume sessions and which resource settings they honour. See Sandbox providers.

Session The unit of work in OpenInspect: a persistent, multiplayer conversation with an agent, tied to a session target, that holds messages, events, artifacts, participants, and a reference to its current sandbox. See Core concepts.

Session branch The branch a session commits to when it has no named branch: open-inspect/<session id>, derived from the session ID. See Your first pull request.

Session target What a session works on, chosen when it is created: a single repository, an ad-hoc set of up to 10 repositories, an environment, or no repository. The target decides which secrets and settings the session receives. See Core concepts.

Skill profile A personal saved subset of shared managed skills, selectable in the new-session skill picker. Entries that are disabled or not assigned to the chosen target are ignored. See Managed skills.

Snapshot A saved copy of the sandbox filesystem taken after a successful turn and before an inactivity shutdown. A later prompt restores it, keeping installed dependencies and uncommitted work, so setup.sh does not run again. Failed boots never produce one. See Lifecycle and statuses.

Spend limit The maximum reported cost in USD a session may accumulate, set as the sandbox setting maxSessionCostUsd (blank means no limit) or overridden for one session from the Budget sidebar section. See Spend limits.

Tunnel A public HTTPS URL for a port inside the sandbox, enabled through the tunnelPorts sandbox setting (up to 10). URLs appear in the Tunnel URLs sidebar section and in /workspace/.tunnels.env. See Sandbox tools.

Unattended mode A per-provider policy that decides whether launches with no person choosing (Slack, GitHub, Linear, and unpinned automation runs) use the provider's default account or API-key mode. See Provider accounts.

Workspace Two meanings. In the sandbox, /workspace is the directory holding the session's clones. In administration, the workspace is the OpenInspect installation whose members hold the Owner, Administrator, Member, or Viewer role. See Workspace access.

Next steps

• Core concepts
• Status reference
• FAQ