Limits reference

Every documented limit, cap, and default in OpenInspect, grouped by area, with where each one applies.

This page collects the limits and defaults that OpenInspect enforces. Values marked as settings can be changed by an administrator; the rest are fixed.

Sessions

Limit	Value	Where it applies
Repositories per session target	10	Ad-hoc multi-repository sets and environments
Prompt length	64,000 characters	Web prompts and child follow-up prompts sent with `send-child-prompt`
Unfinished prompts per session	50	Prompts that are pending or processing. A further prompt is rejected with "A session may have at most 50 unfinished prompts"
Images per prompt	6	Web and Slack prompts
Image size	10 MiB each	PNG, JPEG, WebP, or GIF
Concurrent child sessions per parent	5 by default	Sandbox setting `maxConcurrentChildSessions`
Total child sessions per parent	15 by default	Sandbox setting `maxTotalChildSessions`
Session cost limit	Blank by default (no limit); a positive USD amount when set	Sandbox setting `maxSessionCostUsd`, overridable per session from the Budget sidebar section

Boot and sandbox

Limit	Value	Where it applies
Connect watchdog	4 minutes	From sandbox creation until the runtime first connects. Covers the provider launching the container, not your scripts
Boot budget	30 minutes by default	From sandbox creation until ready, including `setup.sh` and `start.sh`. Operators change it with `SANDBOX_BOOT_TIMEOUT_MS`
Heartbeat timeout	90 seconds	While connected (booting or ready), a runtime that stops sending heartbeats for this long is marked `stale` and stopped
Heartbeat interval	30 seconds	How often the runtime reports in while connected
Tunnel ports	10	Sandbox setting `tunnelPorts`
Tunnel URL wait before `start.sh`	30 seconds by default	`TUNNEL_WAIT_TIMEOUT_SECONDS` in the sandbox; `start.sh` runs without fresh URLs if it expires
Final snapshot buffer	10 minutes by default, 5 minutes minimum	Sandbox setting `finalSnapshotBufferMs`: time reserved before provider expiry to save state
Sandbox lifetime	Blank: 2 hours on Modal and E2B, 45 minutes (the maximum) on Vercel, the provider's own default on OpenComputer. When set: whole seconds, at least 1 second	Sandbox setting `sandboxTimeoutMs`
Prompt execution timeout	The configured sandbox timeout; otherwise the operator's `EXECUTION_TIMEOUT_MS`; otherwise 2 hours	How long a message may stay processing before it is failed

Secrets

Limit	Value	Where it applies
Secrets per scope	50	Global, repository, and environment scopes
Key length	256 characters	Any scope
Key format	Letters, digits, and underscores; cannot start with a digit	Any scope
Value size	16 KB	One secret
Total value size per scope	64 KB	One scope
Combined size per session	128 KB	Global plus the session target's secrets after merging. A larger payload fails the spawn with a per-scope byte breakdown
Reserved keys	`PYTHONUNBUFFERED`, `SANDBOX_ID`, `CONTROL_PLANE_URL`, `SANDBOX_AUTH_TOKEN`, `REPO_OWNER`, `REPO_NAME`, `GITHUB_APP_TOKEN`, `SESSION_CONFIG`, `RESTORED_FROM_SNAPSHOT`, `OPENCODE_CONFIG_CONTENT`, `PATH`, `HOME`, `USER`, `SHELL`, `TERM`, `PWD`, `LANG`	Cannot be saved as secrets; the form shows a validation error

Managed skills

Limit	Value	Where it applies
Canonical name	64 characters	Must be unique. `agent-browser`, `record-video`, `upload-screenshot`, `visual-verification`, and `customize-opencode` are reserved
Description	1,024 characters	Skill metadata
License	200 characters	Skill metadata
Compatibility	500 characters	Skill metadata
Metadata key / value	100 / 500 characters	Skill metadata
Supporting files	99 per skill revision	Files beside the generated `SKILL.md`
Individual file	256 KiB	One supporting file
Complete skill revision	1 MiB	Instructions plus supporting files
Managed skill content in a session	5 MiB	All installed skills together
Supporting-file path	Relative, `/` separators, no `.` or `..` segments, at most 10 segments or 240 UTF-8 bytes, no control characters, not `SKILL.md`	Each supporting file

Automations

Limit	Value	Where it applies
Name length	200 characters	Any automation
Instructions length	15,000 characters	Any automation
Repositories and environments per automation	10 combined	Multi-select is available on schedule triggers only
Minimum schedule interval	15 minutes	Cron schedules
Webhook payload size	64 KB	Inbound webhook triggers
Concurrent runs	1 per automation	Scheduled and manual firings; a firing during an active run is recorded as Skipped
Consecutive failures before auto-pause	3	Timed-out runs count as failures; partial failures never reset the counter
Run execution timeout	The session's execution budget (the configured sandbox timeout for the run's scope, otherwise the operator's `EXECUTION_TIMEOUT_MS`, otherwise 2 hours), plus a 100-minute sweep grace	The session fails its own prompt at the budget. If that result never reaches the scheduler, a sweep fails the run with `execution_timeout` once the grace has also passed. Timed-out runs count toward auto-pause
Slack thread continuation	7 days	Replies in a triggering thread continue the same session
Slack thread context	20 earlier messages, 1,024 characters each	Passed to the agent when the triggering message is a reply
Automations list page	25 by default, 100 at most	`GET /automations`
Invocation history page	100 at most	`GET /automations/:id/invocations`

Prebuilt images

Limit	Value	Where it applies
Build timeout	30 minutes by default, 60 minutes at most	Sandbox setting `buildTimeoutSeconds`; covers clone and `setup.sh`
Vercel build cap	35 minutes	Vercel limits sandbox lifetime to 45 minutes, so execution is capped to keep the finalization reserve
Finalization reserve	10 minutes on top of the build timeout	Callback delivery and snapshot or checkpoint capture
Rebuild scheduler	Every 30 minutes	Checks each enabled scope for new commits, a missing image, or an outdated runtime
Builds per scope	1 at a time	A trigger while a build is in flight is a no-op

Slack

Limit	Value	Where it applies
Images per message	6, 10 MiB each	Attached images and images inside forwarded messages, PNG, JPEG, WebP, or GIF
Forwarded messages per request	10	Each shared message's text is truncated at 4,000 characters
Generated media per completion	5 files, 10 MiB per file, 25 MiB total	Media delivery, when the operator enables it
Target dropdown lifetime	1 hour	The original request is kept while you pick a repository or environment
Thread-to-session mapping	About 7 days	Replies after that may start target selection again
Thread messages included with a follow-up	Up to 10	Messages posted after the previous prompt and before the new request
Thread context for channel-message automations	Up to 20 messages, 1,024 characters each	Only when the triggering message is a reply
Session instructions	10,000 characters	Settings › Integrations › Slack › Session Instructions, appended to the first prompt of new Slack sessions
Routing rules	100 rules, 100-character keywords	Settings › Integrations › Slack › Routing rules

Linear

Limit	Value	Where it applies
Issue-to-session mapping	7 days	After it expires, a new agent request may start a new session

PR Feedback Autofix

Limit	Value	Where it applies
Attempts per pull request	30 per 24 hours by default	Settings › Integrations › GitHub › PR Feedback Autofix; "No Autofix attempt limit" removes the cap
Review comments per feedback item	100	One submitted review
Diff hunk per review comment	4,000 characters	Longer hunks are truncated and flagged
Prompt size	200,000 bytes	The follow-up prompt built from the feedback

Analytics

Limit	Value	Where it applies
Time range	7, 14, 30, or 90 days; 30 is the default	The Analytics page