# Limits and Errors Limits on actions, executions, and calls, and the errors the actions API returns. ## Limits | Limit | Value | | ----------------------------------------------------------- | ---------------------------------- | | Triggers per action | 16 | | Trigger expression size | 16 KiB | | Expressions in a trigger `view` | 64 | | `session.schedule` entries | 8 | | `every` in a `session.schedule` entry | Whole minutes, at least 1m | | `turns.window` | 24h | | A view as the agent reads it | 4 KiB | | Configuration size | 352 KiB | | Event and handler output size | 1 MiB | | Samples per `validateAction` call | 100 | | Page size (`first`) | 1 to 100 | | `maxInvocations` | 24 | | `timeout` | 24h | | `retryDelay` and `maxRetryDelay` | 12h | | `responseTimeout` | 10m | | `completionTimeout` | 24h | | Tools an MCP server references | 128 | | Tool levels below a first call | 8 | | Calls beneath one call | 1,024 | | Agents in one call chain | 3 | | Output uses per invocation | 256 | | Transcript events per invocation | 4,096 | | Starlark source | 64 KiB | | Starlark steps per invocation | 1,000,000 | | `graphql`, `cel`, and `tools` calls per Starlark invocation | 64 each | | A `cel` expression from Starlark | 16 KiB; aggregate CEL cost 100,000 | ## Errors Errors follow the [API error format](https://www.twisp.com/docs/reference/api/errors.md). Reads of an identifier that was never created return null rather than an error. A request refused because it conflicts with what is recorded, such as rerunning a replay or an unfinished execution, replaying a replay, or answering an approval request twice, has code `BAD_REQUEST` and is not retriable: the same request is refused the same way. | Situation | Result | | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `updateAction`, `deleteAction`, `cancelExecution`, `rerunExecution`, or `replayExecution` on an identifier that was never created | Error with code `NOT_FOUND`. | | Invalid trigger, configuration, policy, or execution setting on create, update, or validation | Error whose message names the rejected setting, for example `retry delay cannot exceed the maximum retry delay`. | | More than 16 triggers, or no trigger, `tool` block, kind `skill`, or kind `mcp`, on create, update, or validation | Error `an action accepts at most 16 triggers`, or `an action requires a trigger, a tool block, kind skill, or kind mcp`. | | A name already used by another action, or a `tools` or `skills` entry that cannot be resolved | Error; the write is rejected. | | An invalid `schedule`, `cron`, `timezone`, `every`, or `session.ends`, or a schedule trigger without a name or resuming by `input` | Error whose message names the rejected setting, for example `session.schedule[0].every: every 1m30s: must be whole minutes`. | | `session.ends`, `session.schedule`, or a schedule trigger in a deployment without a scheduler | Error `this deployment has no scheduler: session.ends, session.schedule, and schedule triggers are unavailable`. | | A session turn that reports `COMPLETED` before the session's end, or `OPEN` in its final turn | The turn fails with `action_output_invalid`. An agent is offered only the allowed status. | | A skill whose name is also a tool name the action's model sees | Error `tool and skill names must be distinct within one action`. | | An agent's call to a tool or skill version it does not pin | Permission error carrying the code `tool_not_declared`. | | A call to a tool whose session is closed | Error with code `INVALID_STATE_TRANSITION` carrying `session_closed`; nothing runs or is recorded. | | A skill load whose input is not an empty object | Error `a skill load takes an empty input object`. | | `invokeAction` on a disabled, deleted, or unpublished tool, a [native tool](https://www.twisp.com/docs/reference/actions/tools.md#native-tools) that acts for its caller, or without `INVOKE` permission | Error that is not retriable, with code `NOT_SUPPORTED` or `ACCESS_DENIED`, whose `actionError` extension carries the reason's `code`, such as `action_not_callable` or `action_access_denied`, with a `message` and `hint`; nothing runs or is recorded. | | `startAction` on a disabled, deleted, or unpublished action, one without a `tool` block, a tool with a session, from inside an execution, without `INVOKE` permission, or with input that fails the schema | Error that is not retriable, with code `NOT_SUPPORTED`, `ACCESS_DENIED`, or `BAD_REQUEST`, whose `actionError` extension carries the reason's `code`, such as `action_not_callable`, `start_not_supported`, `action_access_denied`, or `action_input_invalid`, with a `message` and `hint`; nothing is queued or recorded. In gRPC, the status carries them as an `ErrorInfo` detail: `reason` is the code, and `metadata.hint` the hint. From inside an MCP request, run the action with [`invoke`](https://www.twisp.com/docs/reference/actions/mcp.md#invoke) instead, on a server that lists it. | | `startAction` of a native tool that acts for its caller, such as `program`, `invoke`, `read_turn_input`, or a harness tool, or `rerunExecution` of a call to one | Error that is not retriable, with code `NOT_SUPPORTED` and `actionError` code `start_not_supported`: such a tool runs only as its caller's call. Nothing is queued or recorded. | | `startAction` with an `executionId` already used with a different principal, action, version, or input | Error with code `ALREADY_EXISTS`, not retriable; an execution identifier names one start. | | An invocation of a started execution after its starter was revoked, may no longer invoke the action, or its token expired | The execution fails with `action_access_denied`; nothing of the invocation runs. | | `invokeAction` with input that fails the tool's schema, or a call that fails | A `FAILED` execution whose `invocation.error` carries the code and a hint; the caller can still commit. | | A write conflict, passed deadline, or interrupted transaction during a call | Retryable error; roll back and retry in a new transaction. | | Kind `workflow.starlark`, a `tool` block on a kind other than `tool`, `agent`, or a native tool's, or kind `tool` without one | Error; kind `tool` with config `{runtime, src}` replaces `workflow.starlark`. | | An agent whose schemas do not combine into its turn outcome, whose limits could derive more than 256 uses, or whose references read `tool.input` while it has triggers | Error whose message names the rule. | | An action named after a [native tool](https://www.twisp.com/docs/reference/actions/tools.md#native-tools), such as `graphql`, `cel`, or `read_turn_input` | Error; the names are reserved for native tools. | | Update or delete of a deleted action | Error; the action is immutable after `pending: DELETED`. | | `rerunExecution` of an intake record whose action is disabled or deleted, at a version without a `session`, or at a version whose triggers are no longer published | Error whose message names the reason; nothing is created. | | `rerunExecution` on a non-terminal source, or `executionId` reuse with different values | Error; the source must be terminal and an execution identifier names one rerun. | | `replayExecution` on an unfinished execution, a call another execution made, a replay, an intake record, an action that is not kind `agent` or `mcp`, or with a version whose session does not match the source's | Error whose message names the reason; nothing is created. | | `rerunExecution` of an MCP server's request | Error; a request runs only when its endpoint serves it. Replay it instead. | | `rerunExecution` or `replayExecution` of a draft's preview, an execution with a `draft` | Error; no version holds the draft it ran. Preview the draft again. | | `rerunExecution` or `replayExecution` without `INVOKE` on the action, or `rerunExecution` from inside an execution | Error; nothing is queued or recorded. From inside an MCP request, run the action with [`invoke`](https://www.twisp.com/docs/reference/actions/mcp.md#invoke) instead, on a server that lists it. | | `rerunExecution` of a call to a tool with a session, or of a call or start whose action is no longer callable | Error whose message names the reason; nothing is queued. Call the tool again instead. | | `replayExecution` with `executionId` reuse with different values | Error; an execution identifier names one replay. | | `cancelExecution` on an execution processing a session turn | Error; close the session instead. | | Explicit `null` on an update field outside `execution` | Error; omit the field to preserve it. | | An execution listing with `intake: true` and, in the gRPC `ListActionExecutions`, a `session_id` or `parent_execution_id` | Error; intake records run no turn and make no call. | | Cursor from a different filter, or `first` outside 1 to 100 | Error. | Because operations inside one `actions` block share a transaction, an error in any of them aborts all of them.