FAQ
Short answers to the questions people ask most about OpenInspect sessions, configuration, models, integrations, and administration.
Short answers to common questions, with a link to the page that goes deeper.
Getting started
Does closing my browser stop the session?
No. Sessions are persistent: the agent keeps working in its sandbox, and the session's messages, events, and artifacts are stored server-side. Come back later and the session picks up where it was. See Core concepts.
Can I run several sessions at once?
Yes. Each session has its own sandbox. Cost limits apply per session. See Session page.
Sessions
Why did my session start from a snapshot?
On Modal, a snapshot may be taken after a completed turn, even if execution failed. Modal, modal-vm, and Vercel can restore saved state after confirmed shutdown; Vercel and modal-vm skip per-turn snapshots. Daytona, E2B, and Boat resume theirs in place. Saved work includes uncommitted changes. An incompatible or unknown-version old snapshot is held rather than restored or replaced with a fresh checkout. See Lifecycle and statuses.
Why is setup.sh not running?
setup.sh runs only for image builds and fresh sessions. It is skipped for prebuilt-image starts (it already ran during the build) and for snapshot restores. start.sh runs on every non-build boot. Check OPENINSPECT_BOOT_MODE in the script if you need to branch on this. See Lifecycle scripts.
What happens if I send a prompt while the agent is busy?
It queues. Prompts are processed in order, and a session may hold up to 50 unfinished prompts. To change course instead, use the stop button, which stops the current prompt while queued prompts continue. See Follow-ups and stopping.
Who can see my session?
It depends on visibility and your deployment's enforcement mode. Workspace-visible sessions are readable by admitted workspace users with session read permission. With Teams enforcement on, Team visibility limits reads to owning-team members and workspace Administrators or Owners. Private sessions are for the owning user and explicit collaborators, with audited break-glass read access for workspace Owners only; that exception does not grant prompting or sandbox access.
Team defaults are Team or Workspace, never Private, but you can explicitly choose Private for an individual session, including a team-owned one. Prompting requires collaboration permission and session access; private collaborators on team-owned sessions must retain current team membership.
The default enforcement mode is shadow, which does not isolate non-private Team reads. Private access and team-owned action checks apply in every mode. Except for collaborator self-removal, every non-read action on a team-owned session requires current owning-team membership, even for administrators. Visibility can change, but ownership cannot. See Multiplayer and sharing, Teams, and Deployment.
My prompt failed with a message naming a boot phase. What now?
The boot budget (30 minutes by default), the 4-minute connect watchdog, or the 90-second boot heartbeat tripped. Send the prompt again; it starts a new sandbox. If setup.sh is slow, shorten it or enable a prebuilt image. See Recover from a failed run.
Can the agent merge its own PR?
There is no dedicated merge tool; the pull-request tool creates and updates PRs. The agent can still run gh pr merge from the sandbox, normally with the shared GitHub App installation token rather than your OAuth identity. If the App has permission, only GitHub's repository rules prevent that merge. Review and merge outside the sandbox by default, and enforce branch protection and required reviews in GitHub. See Reviewing changes.
Configuration
Where do secrets go?
Add them under Settings › Secrets (global or per repository), a team's Secrets tab, or an environment's Secrets tab, subject to your management access. They are stored encrypted and injected into the sandbox as environment variables at spawn. Precedence is global, owning team (if any), then target repository or environment; later values win. Environment sessions do not inherit their member repositories' secrets. Values are never shown back in the browser. See Secrets.
Why is my session not using the prebuilt image?
For a repository, the image must be Ready and the session must be on the default branch. For an environment, the image must be Ready and the environment must not have been edited since the build. Ad-hoc "Multiple repositories" sessions never use prebuilt images; save the set as an environment instead. See Prebuilt images.
I edited a managed skill. Why does my session not have the change?
Skills are pinned to the revisions that existed when the session was created, and restarting or restoring the session keeps those revisions. Start a new session to pick up the edit. See Managed skills.
Models and accounts
Why is a model missing from the picker?
Check Settings › Models: opt-in groups (xAI/SuperGrok, OpenCode Zen, OpenCode Go, Kimi Coding Plan, Z.AI Coding Plan, DeepSeek) must be enabled. The Claude Agent harness lists Anthropic models only. The picker uses the shared model catalog and saved preferences, not the sandbox or prebuilt-image version. See Choosing a model.
Do I need an API key to use OpenAI or Grok models?
Not necessarily. Connect a ChatGPT or SuperGrok subscription under Settings › Accounts and the session fetches short-lived access tokens from the control plane instead of using a key. API-key mode (OPENAI_API_KEY or XAI_API_KEY as a secret) remains available. Claude account mode is different: it uses a static setup token in the claude process. See Provider accounts.
What is the difference between OpenCode, Claude Agent, and Codex?
OpenCode is the default harness and runs any catalog model. Claude Agent runs Anthropic models only, can use a connected Claude subscription, and reads the repository's CLAUDE.md natively. Codex runs the OpenAI models that support it, can use a connected ChatGPT account, and reads the repository's AGENTS.md natively. The harness is chosen at session creation and cannot be changed afterwards. Slack, GitHub, and Linear launches use the harness chosen in their integration settings (Slack users can also pick their own in App Home), and automations use the harness chosen on the automation. Both default to OpenCode. See Agent harnesses.
Integrations
Does a Slack thread reply start a new session?
In a DM, a reply inside the original thread within about 7 days goes to the same session. In a channel, an interactive thread follow-up must also @mention the bot; an ordinary reply without the mention is not dispatched as an interactive follow-up. A fresh top-level request or a reply after the mapping expires may start target selection and a new session. See Slack.
Why did the Slack bot ask which repository to use?
It could not decide from the message. Name the repository or environment in the request, ask an administrator to add a routing rule for the keyword, or choose from the dropdown within an hour. See Slack.
Why did a PR comment resume my session?
PR Feedback Autofix is enabled under Settings › Integrations › GitHub. A human review or comment on an open PR that a session created is queued into that session as a follow-up, and the timeline shows "Resumed by PR feedback". Attempts are capped at 30 per PR per 24 hours by default. See GitHub.
Administration
How do I stop an automation that keeps failing?
Pause it from the automations list or its detail page. An automation that fails 3 times in a row is paused automatically and shows Paused; fix the cause, use Trigger Now to verify while it is still paused, then Resume, which resets the failure counter. See Automations.
Who can change settings, secrets, and integrations?
Workspace Owners and Administrators manage workspace configuration. Team leads can manage team resources where their workspace permissions and resource-specific checks allow it, including team secrets; a team role does not replace a workspace role. Only Owners can transfer workspace ownership or change who holds the Owner role. See Workspace access and Teams.
Can I get an archived session back?
Yes, if you still have access. Open Settings › Data Controls, find the session in the archived list, and restore it. Restoring needs session lifecycle permission and, for a team-owned session, current owning-team membership even for administrators. See Data controls.