Motivation

• Prevent unbounded context growth by keeping full-fidelity execution payloads in storage while sending only compact, high-signal information back into the rolling scratchpad used for prompt construction.
• Honor tool-provided metadata (e.g., per-part save: false) and enforce a central policy so ephemeral/large tool outputs are not repeatedly injected into LLM calls.

Description

• Add ExecutionResult::compact_for_history() which truncates long text, summarizes/compacts large JSON, preserves artifact references, and replaces inline images with lightweight placeholders to reduce prompt bloat.
• Ensure ToolResponse.parts_metadata (save: false) is applied during compaction so ephemeral parts are removed before re-injection into history.
• Wire compaction into ExecutorContext::store_execution_result() so all execution results persisted to the scratchpad are the compacted versions.
• Add a unit test test_compact_for_history_filters_save_false_and_truncates_large_parts validating truncation, JSON compaction, and save:false filtering.

Testing

• Ran unit tests for the new behavior with cargo test -p distri-types compact_for_history -- --nocapture, which passed.
• Attempted cargo check -p distri-core but it failed in this environment due to an external rusty_v8 archive download error (build script returned HTTP 403), not related to the compaction change.

Codex Task