autosearch-context-retention-policy

name	autosearch:context-retention-policy
description	Session-level policy for keeping the runtime AI's context window healthy across long research — keep-last-k tool results, offload older evidence to disk, trigger compaction at thresholds. Borrows MiroThinker's keep_tool_result, deepagents' summarization middleware, and deer-flow's SummarizationEvent pattern. Orthogonal to assemble-context (which is per-synthesis-step); this is per-session.
version	0.1.0
layer	meta
domains	["workflow","context-management"]
scenarios	["long-session","token-budget","context-overflow","compaction-trigger"]
trigger_keywords	["keep last k","compact context","offload evidence","context policy"]
model_tier	Fast
auth_required	false
cost	free
experience_digest	experience.md

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.

Compacting these breaks the session's ability to finish correctly.

Offload Archive Format

Each compacted batch writes a JSONL record:

{
  "ts": "2026-04-22T03:10:00+08:00",
  "session_id": "...",
  "batch_id": "offload_0007",
  "reason": "token_budget_exceeded",
  "items": [
    {"kind": "tool_result", "tool_name": "run_channel", "args": {...}, "result": {...}},
    ...
  ],
  "digest_replacing_inline": "... 5-10 line summary ...",
  "original_token_count": 4200,
  "digest_token_count": 210
}

Archive is session-scoped; persist across process restarts in session/<id>/offloaded/. Used by trace-harvest and user-initiated "show me what got dropped" commands.

When to Use

• Every long research session (>= 10 tool calls expected).
• When runtime AI detects context bloat.
• When user asks "what do you still have in context?" — this skill's state answers.

When NOT to Use

• Short one-shot research (single decompose → 3 channel calls → synthesize). Overhead not justified.
• When the underlying runtime has its own compaction (check discover-environment first).

Cost

Fast-tier LLM per compaction batch (low per-call). Triggered only when token budget crosses threshold, typically 1-2× per long session.

Interactions

• Reads ← current session state (tool_results, evidence, citation_index, rubrics, loop state).
• Writes → digest items back into context + offload archive on disk.
• Complements ← assemble-context (per-synthesis-step context preparation).
• Triggered by → any tool-call that measurably grows the context beyond the trigger ratio.

Boss Rule Alignment

• Digest entries must preserve specifics verbatim (numbers, error codes, issue numbers, URLs, version strings, benchmark scores) — same rule that drove the m3 deprecation. A digest that loses specifics is worse than dropping the batch entirely.
• Context retention is a policy, not a pipeline — runtime AI enforces it when useful, skips it when not.

Quality Bar

• Evidence items have non-empty title and url.
• No crash on empty or malformed API response.
• Source channel field matches the channel name.