airflow.providers.common.ai.sandbox.output¶

Bounding helpers for sandbox output that the model reads.

Two independent caps apply everywhere: a line count and a byte budget, whichever is reached first. Truncation never emits a partial line, so the model never has to guess whether it is looking at a whole record.

The direction differs by tool, and that difference is the point:

Command output keeps the tail. A traceback, a failing assertion and the exit status all live at the end; the head is usually build noise.
File reads keep the head and report the next offset, so the model can page forward through a file it is reading top-down.

Functions¶

`format_size`(num_bytes)	Render a byte count the way the truncation notes show it to the model.
`truncate_output`(text, *, max_lines, max_bytes[, ...])	Cap free-form command output, keeping the tail and marking anything dropped.
`render_file_window`(data, *, offset, limit, max_lines, ...)	Render a window of a text file, head-first, with a continuation offset.

Module Contents¶

airflow.providers.common.ai.sandbox.output.format_size(num_bytes)[source]¶

Render a byte count the way the truncation notes show it to the model.

airflow.providers.common.ai.sandbox.output.truncate_output(text, *, max_lines, max_bytes, already_truncated=False)[source]¶

Cap free-form command output, keeping the tail and marking anything dropped.

already_truncated says the backend dropped bytes while reading the stream, before this function saw text. The cut is then marked even when what survives happens to fit, so the model is never told it has the whole output when it does not.

airflow.providers.common.ai.sandbox.output.render_file_window(data, *, offset, limit, max_lines, max_bytes)[source]¶

Render a window of a text file, head-first, with a continuation offset.

offset is a 1-indexed line number and limit a line count, both as the model supplies them. The safety caps apply on top, so a file with very long lines is bounded even when the model asks for few lines.