Expand description
Content-addressed deduplication of tool-result blocks.
Agentic loops re-read the same file and re-run the same searches many times. Each duplicate tool output costs full input tokens on every subsequent turn. This module detects exact content duplicates via SHA-256 and replaces later copies with a short back-reference that points at the first occurrence.
§Safety
No information is lost — the model can always ask the agent to
re-run the original tool call. The back-reference preserves the
tool_call_id of the first sighting so the model (or a human
auditing the transcript) can correlate the two.
Only ContentPart::ToolResult blocks are considered; text,
assistant messages, tool-call arguments, images, and thinking blocks
are left untouched.
§Threshold
Tool outputs smaller than MIN_DEDUP_BYTES are left alone — the
marker itself would be nearly as long as the content, and small
outputs typically carry disambiguating structure (e.g. "ok" vs
"error").
Constants§
- MIN_
DEDUP_ BYTES - Tool outputs shorter than this byte count are never deduplicated. Chosen so the back-reference marker is always shorter than the content it replaces.
Functions§
- dedup_
tool_ outputs - Replace duplicate tool-result contents with a back-reference marker.