Errors and limits

Reference for what bounds a task run, what each failure means, and what a task config is refused for.

Budgets

Setting	Default	Range	Bounds
`max_steps`	25	1–50	Model → tools → model rounds per attempt, on a standalone run.
`timeout_seconds`	300	10–600	Wall-clock time for the whole try: reading its files and every attempt.

A step is a round, not a tool call. Four tools called in one reply are one step.
Each re-prompt restarts the step budget. A run without a result is re-prompted up to twice, so max_steps: 25 can mean 75 rounds. timeout_seconds is the real bound; max_steps is the runaway-loop guard.
Provider tools spend time, not steps. A web_search runs inside one round on the provider’s side, so a run that searches heavily fails as timeout with steps to spare, and raising max_steps does nothing.
Entered by an agent, max_steps is not used: the task runs on the calling agent’s own max_steps (null means 4 on voice and video, 25 on text). timeout_seconds still applies and is the only limit on how long the task holds the caller, so keep it short for voice. See agent tasks.

`error.type`

A failed standalone run has status: "failed" and error: {type, message}; a canceled one has status: "canceled" and error.type: "canceled". A task entered by an agent has no run record; it fails as an ordinary tool failure in that call.

`error.type`	Meaning	Fix
`invalid_input`	An input could not be used: a required one had no value (seen on email batch rows; `POST /v1/task-runs` refuses this with a `400`), a PDF could not be downloaded or read or is over a limit, or it has scanned pages and the model cannot read images. Caught before the model ran, so nothing was spent on it.	The message names the input and the reason. Fix the value or file, or pick a model with vision.
`no_output`	The model finished without calling `submit_result`, including after re-prompts.	Fix the prompt and `submit_result_description`. Do not raise `max_steps`. See writing a task prompt.
`step_limit`	An attempt used every step without a result.	Raise `max_steps`, or narrow the work. The prompt is not the problem.
`timeout`	`timeout_seconds` elapsed.	Raise it, or cut the work. Time slow tools on their own first.
`provider_error`	The model provider or an MCP server failed.	A background run is retried automatically when it is transient; otherwise retry. An `llm.fallback` covers the model side.
`configuration`	Something the run needs is missing: a provider key, a tool deleted or unpublished since the task was published, an integration whose credential no longer resolves.	The message names it. Add the provider key, republish the tool, or reconnect the integration.
`platform`	A failure on our side.	Retry; contact hello@talqing.com if it persists.
`canceled`	Somebody canceled the run, or a deploy stopped an inline or drafting run. Tokens already spent are still priced.	An email batch redrafts the row automatically; rerun a manual run. A background run stopped by a deploy is retried rather than canceled.

Retries

Only background runs are retried, at most 3 tries in all, and only when:

the worker running the try stopped (a deploy or a crash), whatever it had done; or
it failed with provider_error that the provider marks as transient (a 5xx, a rate limit, a dropped connection), or with platform, and no tool or on_enter hook ran in that try.

An error after a tool ran is never retried, but a worker stop is, from the start, so a side-effecting tool can run twice. Retries wait 30, then 60 seconds. A run that has had its tries fails with platform, the worker running this stopped; gave up after 3 tries, or with its last error. tries, usage and provider_cost on the run cover every try.

Refused at run time

POST /v1/task-runs and POST /v1/task-runs/batch refuse before anything runs, with no run recorded. A batch is refused whole, each problem prefixed runs[i]:.

Status	Cause
`400`	A value for an undeclared input: `this task does not declare an input named 'emial'`
`400`	A `required` input with no value and no default: `'email' is required and has no default`
`400`	A file input sent under `vars`, a text input under `files`, more than 5 files, a file over 20 MB, or invalid base64
`400`	No published version: `publish '<name>' before running it`
`422`	The resolved config is invalid: a `"draft"`, inline `task` or `task_override` that would not publish, listed error by error
`422`	A request field of the wrong shape: a `vars` value that is not a string (send `"007"`, not `7`), a non-`https` file URL, a `metadata` key outside `A-Za-z0-9_.-`
`409`	An `idempotency_key` already used for a different request
`413`	A request body over 50 MB
`429`	100,000 runs already queued in the workspace
`403`	Viewer role. Running is a write.
`404`	No such task, or no such `task_version`.

Refused at save

Every create and PATCH is checked; a failure is a 400 listing every problem, and nothing is stored.

Refusal
A tool that does not exist or has no published version
An integration that is not MCP-capable
Two attached MCP servers sharing a `tools_namespace`, or two tools exposing the same name
An unknown LLM provider or model, or a `reasoning_effort` that model does not offer
A `{{token}}` with no valid root (`{{name}}`, `{{customer.number}}`), or an unknown `{{system_vars.x}}`
`{{system_vars.now}}`, `.date` or `.time` with no `timezone`
`{{secrets.…}}`, `{{args.…}}` or `{{tooldata.…}}` in the prompt
An `untrusted` or `file` input read in the prompt, an `untrusted` one in a tool’s URL, query or headers, or a `file` one in any template
The same name twice in `output`, or a name that is both a variable and an output field
`409`: a task name, or an inline tool or MCP server name, already used in the workspace
`422`: a field a task does not have (`greeting`, `stt`, `channel`, `handoffs`…), more than 25 top-level `output` fields, output nested over 4 levels or over 100 fields, an output name outside `^[a-z][a-z0-9_]{0,39}$`, an empty description, `fields`/`items`/`enum` on the wrong type, a `default` on a `file` input, `max_steps` or `timeout_seconds` out of range, a non-IANA `timezone`

Refused at publish

POST /v1/tasks/{task_id}/validate runs these without publishing; POST /v1/tasks/{task_id}/publish refuses on them.

Refusal
No provider key for the model (or its fallback): `no OpenAI API key for this workspace - add one under Provider keys`
`output` is empty
A problem in an attached tool’s operation tree
An attached tool (or hook tool) containing a `handoff` operation
A drafting email batch whose CSV lacks a column for a variable this config makes required
A drafting email batch whose `field_map` names an output field this config removes

The last two also refuse a rollback to a version with the same problem.

Warnings

Returned by validate and publish; they do not block.

Warning	Why it matters
The prompt reads an undeclared `{{vars.x}}`	It resolves to nothing. Usually a typo.
The task reads untrusted input and can call tools	Text inside the input could steer which tools it calls. Keep them read-only.
The task has no prompt	The model only knows its tools’ descriptions.
The prompt reads `{{system_vars.human_phone_number}}`, `.agent_phone_number`, `.direction` or `.channel`	Empty on a standalone run.
An attached integration is not `active`	Its tools are left out of runs until it is fixed.
An attached FAQ has no questions	It contributes nothing.

Limits

Limit	Value
`output` fields	1–25 top-level to publish (a draft may have 0); 100 in all
Output nesting	4 levels
Output field name	`^[a-z][a-z0-9_]{0,39}$`
Output field types	`string` (optionally `enum`, 1–50 values), `boolean`, `integer`, `number`, `object`, `array`
Result size	64 KB serialized; larger is refused back to the model
PDF inputs	5 per run, 20 MB and 100 pages each, 10 scanned pages each, 200,000 extracted characters per run
Background runs at once	3 per workspace; the rest queue, oldest first
Queued runs	100,000 per workspace
Runs per batch request	1000
Tries per background run	3
`metadata`	16 keys; key `A-Za-z0-9_.-` up to 40, value up to 500 characters
Attached FAQs	10
`max_steps`	1–50, default 25
`timeout_seconds`	10–600, default 300
Re-prompts per run	2 (so `attempts` ≤ 3)
Trace entries per run	200
Trace field length	4,000 characters, then cut and flagged `truncated`
Runs per page	default 50, max 200
Tasks per page	default 200, max 200
Versions per task	Unlimited

Get started

Build

Guides

Errors and limits

Budgets

`error.type`

Retries

Refused at run time

Refused at save

Refused at publish

Warnings

Limits

​Budgets

​error.type

​Retries

​Refused at run time

​Refused at save

​Refused at publish

​Warnings

​Limits

Budgets

`error.type`

Retries

Refused at run time

Refused at save

Refused at publish

Warnings

Limits