autosearch:context-retention-policy

Context Retention Policy — Session-Level Context Governance

A research session can generate more evidence / tool results than any reasonable context window. This skill tells the runtime AI what to keep, what to offload, and when to compact.

Policy Parameters (defaults)

policy:
  keep_last_k_tool_results: 12         # inline, full
  keep_all_citations_index: true       # citation_index stays in context always
  keep_all_rubrics: true               # rubrics stay in context always
  offload_trigger_token_ratio: 0.7     # compact when context ≥ 70% full
  offload_target_token_ratio: 0.4      # after compact, aim for 40% full
  offload_archive_path: "session/<id>/offloaded/<ts>.jsonl"
  prefer_compact_over_drop: true       # summarize instead of silently drop
  never_compact:
    - clarify_result
    - reflective_loop_state
    - graph_plan
    - citation_index
    - rubrics

Compaction Procedure

When current_tokens / max_tokens >= offload_trigger_token_ratio:

1. Sort evidence / tool_results by age (oldest first).
2. Identify candidates — everything NOT in never_compact and older than the last K results.
3. For each candidate batch (every ~3 evidence items):
   • If prefer_compact_over_drop: use a Fast-tier LLM to summarize the batch into a single "digest" item (5-10 lines max, preserves URLs + key specifics verbatim).
   • Else: drop the batch but write full content to offload_archive_path for later recovery.
4. Replace the original items with the digest. Keep URLs in the citation_index so citations still resolve.
5. Recheck token ratio. If still above offload_target_token_ratio, iterate.

Preservation Rules (never compact these)

• The current reflective-search-loop state (gaps, visited, bad_urls).
• The current graph-search-plan graph structure.
• The citation-index entries (they're short and referenced by all sections).
• The rubrics from run_clarify.
• The original query + clarify verification message.
• The last 3 tool results regardless of K.