Limits & Quotas
This page summarizes the validation and runtime limits enforced by the Culvii platform and @culvii/kit.
Agent skills
Skills are configured in metadata.runtimeDetails.skills on the primary agent. Understanding how these limits apply depends on three scopes:
- Configured skill source (per entry in
metadata.runtimeDetails.skills): File count (200files) and bundle size (8 MiB) apply independently to each entry in theskillsarray. If a single entry points to a parent folder containing multiple skills (for examplesource: 'skills'), all skills in that folder share one200-file and8 MiBbudget. If you list them as separate entries (for examplesource: 'skills/pdf'andsource: 'skills/xlsx'), each entry gets its own200-file and8 MiBbudget, even when both come from the same repository and merge into/skills/<repo-name>. - Individual skill (
SKILL.md): EachSKILL.mdfile inside a selected source defines one skill via its YAML frontmatter (nameanddescription). - Session skill catalog (across the entire session): Full
SKILL.mdfiles are not inlined into the system prompt on every turn. Instead, the platform injects a compact JSON catalog (skill_id,name,description) for all installed skills into the primary agent's system prompt, and the agent loads individual skill files on demand usingread_skill.
Validation happens in two stages: configuration-shape validation runs during culvii dev and culvii deploy (checking url, ref, credentialId, source, fallback_paths, and repository mount collisions without contacting GitHub), while session-start validation runs when a new session begins (resolving ref to a 40-character commit SHA, fetching the recursive GitHub tree, downloading and verifying blobs, and parsing SKILL.md frontmatter).
| Limit | Value | Details |
|---|---|---|
| Files per configured source | 200 files | Maximum number of regular Git files (100644 or 100755) imported by a single entry in metadata.runtimeDetails.skills. Counts SKILL.md and all files in subfolders under the resolved source (or first matching fallback_paths entry); excludes directory entries and repository files outside that path. |
| Size per file | 1 MiB | Maximum raw (unencoded) byte size of any single file in a skill source. |
| Total size per configured source | 8 MiB | Maximum combined raw (unencoded) byte size of all selected files within one configured skills entry. Separate entries in metadata.runtimeDetails.skills each get their own 200-file and 8 MiB allowance. |
| Combined catalog prompt size | 1 MiB | Maximum UTF-8 byte length of the session-wide skill catalog prompt (the JSON list of { skill_id, name, description } across all installed skills plus fixed usage instructions) injected into the primary agent's system prompt. |
| Max skill path length | 500 characters | Maximum character length for source, each fallback_paths entry, and every repository-relative file path within a skill (measured from the repository root, e.g. skills/pdf/references/checklist.md). Paths cannot start with /, contain \, control characters, or . / .. segments. |
Max fallback_paths per source | 100 entries | When fallback_paths is provided on a skill source, it must be a non-empty array of 1–100 ordered alternative repository-relative paths checked if source contains no skill instructions (only the first matching path is imported). |
Skill name format | 1–64 characters | Required name field in SKILL.md YAML frontmatter. Must match ^[a-z0-9][a-z0-9-]{0,63}$, start with a lowercase ASCII letter or digit, contain only lowercase letters, digits, and hyphens, and be unique across all skills selected by that source. |
Skill description length | 1–1,024 characters | Required non-empty description field in SKILL.md YAML frontmatter (no control characters other than \n, \r, or \t). The YAML frontmatter block itself must have unique keys and allows at most 10 YAML aliases. |
| GitHub repository URL length | Max 500 characters | Applies to url on each github_repository source. Must be an HTTPS https://github.com/<owner>/<repo> URL with no custom port, query string, fragment, or embedded credentials. |
GitHub ref length | 1–200 characters | Applies to the optional ref field (branch, tag, or commit SHA). Defaults to 'HEAD', cannot contain whitespace or control characters, and is resolved to an exact 40-character commit SHA at session start. |
GitHub credentialId format | 1–128 characters | Matches ^[A-Za-z0-9_-]{1,128}$. References a GitHub credential from the Console's Integrations tab and is required for both public and private repositories so skill downloads use authenticated GitHub API rate limits. |
| Max repository Git tree entries | 10,000 entries | Counts all files and directories across the entire GitHub repository's recursive Git tree at the target commit, not just the source folder. Repositories whose recursive tree exceeds 10,000 entries are rejected. |
| Max GitHub API response size | 10 MiB | Maximum JSON payload size accepted from GitHub commit and recursive tree API calls. |
| GitHub skill acquisition timeouts | 20 seconds / 5 minutes | Each individual GitHub API HTTP request times out after 20 seconds, and downloading an entire configured skill source (commit lookup, tree fetch, and all blob downloads) must finish within 5 minutes. |
Agent runtime and sandbox
| Limit | Value | Details |
|---|---|---|
Default maxIterations | 50 steps | Maximum number of model reasoning and tool-loop steps an agent can take in a single run before producing a final answer. Configured on the primary agent and applies to the entire session turn, including all secondary agent delegations. |
Default maxToolCalls | 100 calls | Maximum total number of individual tool executions allowed in a single agent run. Configured on the primary agent and applies to the entire session turn, including all secondary agent delegations. |
| Context window | Per-model catalog | Context-window capacity and maximum output token limits are configured per model in the platform model catalog (there is no default context window). The platform includes the current user prompt when deciding whether older session turns must be compacted. |
| Sandbox file read / attachment size | 10 MiB | Maximum file size when reading whole files from the sandbox for attach_output_file artifacts and read_skill file loads. |
read_file excerpt size | 2,000 lines / 50 KiB | Maximum UTF-8 text excerpt returned by a single read_file tool call (paginated via offset and limit). |
Sandbox command timeout (exec_process) | 60 seconds (max 300 seconds) | Default wall-clock timeout per exec_process command when timeoutMs is omitted; individual calls can request timeoutMs up to 300,000 ms (5 minutes). |
| Sandbox command / tool output size | 50 KiB | Maximum stdout / stderr or response output returned per stream by exec_process, list_files, search_text, git, and web_fetch before truncation. |
| Sandbox pod ready timeout | 120 seconds | Maximum time the platform waits for a session sandbox environment to become ready during creation or restoration. |
| Sandbox idle TTL | 30 minutes | Inactivity duration before an idle session sandbox is snapshotted and its pod is stopped. |
sandboxProfile name | 'default' | metadata.runtimeDetails.sandboxProfile defaults to 'default', and only 'default' is supported on the platform's single default sandbox image. |
networking.allowed_hosts hostname | Max 253 characters | Configured in metadata.runtimeDetails.networking.allowed_hosts on the primary agent when type: 'limited'. Must be exact public bare hostnames (63 characters max per DNS label); rejects schemes, ports, paths, wildcards (*), IP addresses, and localhost/.local/.internal names. |
Workflows and composition
| Limit | Value | Details |
|---|---|---|
| Max loop items | 500 items | Maximum items a core.loop step can iterate over in a single execution, enforced at culvii dev / culvii deploy for literal items arrays and at runtime for expression-resolved arrays (both sequential and parallel loops). Nested core.loop steps are rejected. |
| Max agent/workflow composition depth | 3 nodes | Maximum call-chain depth across workflows and agents. Valid 3-node chains are workflow → agent → workflow (where the leaf workflow contains no core.agent step) and agent → workflow → agent (where the leaf agent declares no workflow tools). Chains of 4 or more nodes fail validation on culvii dev and culvii deploy. |
| Workflow binary file size | 100 MiB | Maximum size of a single binary file processed within a workflow (e.g. via S3 Get or Manual Trigger). |
| Platform HTTP request body size | 150 MiB | Maximum size of the HTTP request body accepted by the platform (e.g. for workflow execution starts). This is the default whole-request HTTP body cap (accommodating base64 overhead on a 100 MiB binary file). |
| Single resource definition size | 5 MiB | Maximum UTF-8 JSON byte size of a single agent, workflow, or surface definition during culvii deploy (enforced by both the CLI and the platform deploy service). |
| Deploy plan request body size | 20 MiB | Maximum HTTP request body size for deploy plans (including decompressed gzip/deflate payloads). culvii deploy automatically gzip-compresses deploy plan payloads over 512 KiB. |
| Agent definition route body size | 10 MiB | Maximum HTTP request body size on direct agent creation and update API routes. |
culvii dev inactivity timeout | 30 minutes | Closes the culvii dev draft-sync WebSocket after 30 minutes without developer activity (saving a file to trigger a sync, or starting a draft workflow/agent run from the Console). Protocol-level 15-second WebSocket ping/pong heartbeats keep the connection alive but do not reset this inactivity timer. |
Dynamic UI surfaces
| Limit | Value | Details |
|---|---|---|
Surface id format | 1–63 characters | Applies to Surface.id. Must match ^[a-z0-9][a-z0-9-]{0,62}$ (kebab-case). Individual component id values inside definition.components allow 1–128 characters matching ^[A-Za-z0-9][A-Za-z0-9._-]{0,127}$. |
| Max serialized definition size | 256 KiB (262,144 bytes) | Maximum UTF-8 byte length of JSON.stringify(definition), including all components, props, actions, and wait-surface form fields. |
| Max components per surface | 300 components | definition.components must contain 1–300 components. All components must have unique IDs within the surface and be reachable by traversing children from definition.root. |
| Max component tree depth | 12 levels | Maximum nesting depth of the acyclic component graph starting at definition.root (depth 1). Cycles in children references are rejected. |
Wait-surface file input maxFiles | 1–5 files | Applies to { type: 'file' } fields in definition.form.fields on a wait surface. Caps how many files can be uploaded for that field when resuming a paused step (defaults to 5 when multiple: true, or 1 when multiple is omitted/false). |
Wait-surface file input maxBytes | Max 10 MiB (10,485,760 bytes) | Applies to { type: 'file' } fields in definition.form.fields on a wait surface. Positive integer up to 10 * 1024 * 1024 bytes capping the size of each individual file uploaded to that field. |