Module compaction

Expand description

Compaction: summarize older context when the window fills, keeping a recent tail. Also branch summarization support for /tree navigation.

Structs§

estimate_context_tokens: Estimated context size: prefer the last assistant usage (input+output+ cache) and add estimates for everything after it.
estimate_message_tokens
estimate_tokens_text: chars/4 heuristic, matching pi’s estimator for messages without usage.
extract_file_ops: Extract read/modified file paths from tool calls in the messages.
file_ops_details: Details payload stored on compaction/branch-summary entries.
generate_summary: Generate a structured summary with the LLM. One-off prompt: fresh session id and no cache writes wanted, so it’s a plain request.
plan_compaction: Choose the cut point: walk back from the newest message accumulating estimates until keep_recent_tokens, then snap to a turn boundary.
serialize_agent_messages: Serialize agent messages for summarization (convert then serialize).
serialize_conversation: Serialize a conversation to labeled text for the summary prompt. Tool results are capped so summarization stays cheap.
should_compact: Whether auto-compaction should trigger.