Module truncate 

Bounded tool output.

Tool results enter conversation history verbatim today, so one verbose command can fill the context window in a single iteration. This module truncates output once, at ingest, middle-out: the head and tail of the output are kept because that is where a command’s intent and its outcome live, and a warning header states what was removed so the model can ask for the rest.

The policy shape and the header text follow Codex CLI’s utils/output-truncation (Apache-2.0). The implementation is UAR’s own and counts tokens through TokenService so budgets agree with the rest of the runtime.

Enums

TruncationPolicy

How much of a tool result may enter history.

Constants

DEFAULT_OUTPUT_BYTE_BUDGET

Default byte budget for a tool result when the tool declares none.

WARNING_HEADER_PREFIX

Start of the header every truncated result carries.

Functions

formatted_truncate

Truncate content to policy if needed, prefixing a warning header that states the original token count and total line count. Output already within the policy is returned unchanged.

formatted_truncate_for_model

Truncate tool output against the resolved model’s tokenizer. An empty or unknown model uses TokenService’s documented cl100k_base fallback.

truncate_middle

Keep the head and tail of content within byte_budget bytes, replacing the middle with a marker that states how many bytes were removed. Cuts land on character boundaries so the result is always valid UTF-8.