Timeouts in Oda

Slow work in DataMind Oda is bounded: a step that can stall carries a time limit, and when it passes that limit Oda receives a timeout in place of the step's result, then decides what to do next. This page is the reference for those limits: where each one comes from, what a timeout reports, and how a limit's value is checked. What a timeout looks like while you follow a task is on Repeated steps and tool timeouts.

The two layers of a limit

A tool's own budget. Each tool that can stall carries its own limit. Web search and page reading carry 30 seconds each by default, along with the other limits of those calls (Web search and page reading).
A backstop behind the tool. The page reader also keeps a second limit of its own behind the tool: a resource backstop, not the budget a call through the web tools is judged by.
Which layer reports depends on what fired first. A call judged by a tool budget reports the tool-call timeout, TOOL_TIMEOUT. A call that reaches the page reader with no tool budget above it reports the reader's own code instead: WEB_FETCH_TIMEOUT.
Shell steps bring their own limits. What those are, and what happens to the command when one fires, is on Shell commands.

What a timeout reports

A limit that fires ends the wait and reports a timeout. The report carries a short code that names the kind of timeout, and the text with it says how long the wait ran.

Code	What fired
`TOOL_TIMEOUT`	A tool's call budget. The text says it timed out and how long it waited, like `timed out after 50ms`.
`BASH_TIMEOUT`	A shell step's own time limit. The reason reads like `BASH_TIMEOUT after 100ms`: the kind of timeout, then the limit that fired.
`WEB_FETCH_TIMEOUT`	The page reader's own backstop, on a call that ran without a tool budget above it: `web fetch timed out`.
`LLM_STREAM_IDLE_TIMEOUT`	A streaming reply that went quiet past the route's Stream idle timeout (Provider route settings reference).

The first reason wins. When a stop and a limit fire at the same moment, the report follows whichever reason arrived first; a later reason does not replace it.

What a limit does, and does not do

It ends the wait; it does not end the work. A timeout is a notification. The step that was running stops only when it responds to the request, by stopping its own process, socket or task. Work that ignores the request can keep things waiting.
The idle watch is not a deadline for the whole reply. The stream idle watch runs only while a step waits for the next piece of a streaming reply. As long as pieces keep arriving the wait goes on: it fires on quiet, not on length.
File work is never cut off. Reading, writing and editing local files take no time limit and run untimed.

How a limit's value is checked

Left out means the default. A limit that is not set falls back to the tool's own default.
Too high is capped. A value above a limit's maximum is capped at the maximum, and no wait is longer than 2,147,483,647 ms, the largest delay that can be scheduled.
A refused value names itself. A limit that is not a positive, finite number is refused before any work starts, and the message names the value and says it must be a positive finite number.
Zero never switches a limit off. There is no zero setting that disables a timeout: zero and negative values are refused.
Limits you can set live where each one belongs: the command limit on Shell settings, the stream idle wait on Provider route settings reference, and a server's tool call limit on MCP servers.

Before a time zone is stored

A time zone value is checked when it arrives, before anything stores it: Oda accepts UTC or an IANA name in Area/Location form and answers the platform's canonical spelling, so an alias is never what gets stored.
Refused are blank and padded values, abbreviations such as CST or GMT+8, single-segment names such as Asia, lower-case utc, and names the platform does not support.
A refused zone shows as an invalid-time-zone error: session/invalid-time-zone where a message carried the value, and subagent/invalid-time-zone for a subagent continuation. How Oda reads time while it works is on Time and terminal context.

Questions, answered

Where does a call's time limit come from?

Each tool that can stall carries its own budget; web search and page reading carry 30 seconds each by default. Behind the page reader sits a second backstop, a resource guard rather than the budget the call is judged by: a call through the web tools is judged by the tool's budget, and a call that reaches the reader with no budget above it is judged by the backstop.

What do the timeout codes mean?

TOOL_TIMEOUT is a tool's call budget; BASH_TIMEOUT is a shell step's own time limit; WEB_FETCH_TIMEOUT is the page reader's backstop, seen when a call ran without a tool budget above it; and LLM_STREAM_IDLE_TIMEOUT is a streaming reply that went quiet. The code travels with the timeout text, for example timed out after 50ms.

Can I switch a time limit off?

No. There is no zero setting that disables a timeout: zero and negative values are refused. A limit is either left out, so the tool's own default applies, or set as a positive, finite number.

What happens if I leave a limit out or set it above its maximum?

Left out, the limit falls back to the tool's own default. Set above its maximum, it is capped at the maximum.

Is the stream idle watch a limit on the whole reply?

No. The stream idle watch runs only while a step waits for the next piece of a streaming reply: as long as pieces keep arriving, the wait goes on. It fires when the reply goes quiet, not when it runs long.

What happens when a stop and a timeout fire together?

The report follows whichever reason arrived first. A reason that arrives after the first one is already recorded does not replace it.

Which time zone names are accepted?

UTC or an IANA name in Area/Location form, such as Europe/Berlin; the accepted value is kept in the platform's canonical spelling. Blank or padded values, abbreviations such as CST or GMT+8, single-segment names such as Asia, lower-case utc and names the platform does not support are refused, and a refusal shows as an invalid-time-zone error.

How large can a single wait be?

No wait is longer than 2,147,483,647 ms, the largest delay that can be scheduled. A value above a limit's own maximum is capped at that maximum.