chat_compactor

Fold a chunk of chat history into a structured rolling summary that fits within token budgets.

Entry

compact

Final output

chat_summary — structured rolling summary with sections (topic_arc, decisions, pending, session_user_facts, artifacts_referenced) plus covers_through_seq derived deterministically by the postprocessor.

How it composes

Single-phase skill: compact finishes immediately after one LLM call. The LLM produces chat_summary_raw (sections + verbatim new_turn_seqs list); the skill postprocessor then computes covers_through_seq = max(new_turn_seqs) in pure Python and emits the caller-facing chat_summary artifact. This keeps arithmetic out of the LLM contract and prevents turn duplication or loss in ChatSession.history.

Caveats

• Invoked internally by ChatSession._maybe_compact() — not intended for direct CLI use. You can test it with reyn run chat_compactor '...' by constructing the history_chunk_to_compact input manually.
• Excluded from available_skills for the chat router (internal infrastructure).
• Section token caps are soft; minor overrun is expected and self-corrects on the next compaction pass.
• Requires python permission for postprocessor.py:compute_covers_through_seq.

Usage

Not invoked directly in normal use. ChatSession calls it automatically when the uncovered BODY region exceeds chat.compaction.trigger_total_tokens. For testing:

reyn run chat_compactor '{"type":"history_chunk_to_compact","data":{...}}'

Source

src/reyn/stdlib/skills/chat_compactor/skill.md