Clear old tool results after processing

Goal

Reduce repeated context cost from large tool outputs in long tool workflows.

Steps

1. Confirm the target result is safely re-fetchable and is not a recent result needed for the current decision. [toolclear]
2. Configure context_management.edits with type="clear_tool_uses_20250919", an input_tokens trigger, recent tool_uses to keep, and clear_at_least. [toolclear]
3. Inspect context_management.applied_edits for cleared calls and tokens freed; let the agent reread a tool if the original content is needed later. [toolclear]

How to verify

Compare token trajectories on a multi-turn synthetic file-read task; confirm tool-use pairing stays intact, recent results remain, and the final synthesis is correct.

Troubleshooting

If clearing does not trigger, check the beta header, model support, and trigger threshold; if an old result is needed, call its source tool again rather than assuming it can be restored.