autosearch:experience-compact

experience-compact — Per-Skill Digest Updater

Reads a skill's experience/patterns.jsonl + existing experience.md, promotes recurring patterns to the digest, prunes stale rules, and rotates raw events if the file is getting large.

Trigger Conditions

Run compact when ANY of:

1. ≥10 new events added to patterns.jsonl since last compact.
2. patterns.jsonl > 64 KB.
3. Last event carried explicit user feedback (user_feedback: accepted or rejected).
4. Session ending and this skill was used at least once during the session.

Runtime AI (or a cron on long-running instances) calls this skill per trigger.

Promotion Rules

A candidate rule is extracted from patterns.jsonl when the same winning_pattern string or good_query template recurs. To promote it into the Active Rules section of experience.md:

• seen >= 3 (appeared in at least 3 distinct events)
• success >= 2 (at least 2 of those events have outcome: success)
• last_verified <= 30 days (most recent appearance within the last 30 days)

Each promoted rule carries a metadata footer:

seen=5, success=4, last_verified=2026-04-21

Anti-Pollution Rules

• Single success never promotes. Even if an event looks great, one instance is noise — requires 3 occurrences.
• User-rejected / rubric-failed / low-relevance events go only into Failure Modes, never Active Rules.
• User corrections (explicit "no, that's wrong") always promote to Failure Modes with high weight.
• Cold rules expire. Any Active Rule with last_verified > 30 days moves to a "pending revalidation" bucket; if not re-verified in another 30 days, removed entirely.

experience.md Template

# <skill-name> experience

## Active Rules        # ≤ 20 entries
- <rule> — seen=5, success=4, last_verified=2026-04-21
- <rule> — seen=3, success=3, last_verified=2026-04-19

## Failure Modes       # ≤ 15 entries
- <failure pattern> — seen=4, last_verified=2026-04-20
- <user correction> — seen=1, last_verified=2026-04-18

## Good Query Patterns # ≤ 20 templates
- `{brand} {feature} 翻车 {year}` — seen=7
- `{category} 避雷 小红书 最近` — seen=5

## Last Compacted
- 2026-04-22, from 37 events, promoted 3 rules, archived raw events before 2026-03-22.

Hard ceiling: 120 lines total across all sections. If exceeded after promotion, drop oldest Failure Modes first, then lowest-seen Active Rules, then lowest-seen Good Query Patterns.

Rotation

When patterns.jsonl > 1 MB:

1. Rename current file to experience/archive/YYYY-MM.jsonl (month of first event in the file).
2. Create fresh empty patterns.jsonl.
3. Update experience.md ## Last Compacted line with rotation note.

Archives are permanent — never delete. They serve as the lineage trail for AVO self-evolution audits.

AVO Safety Boundaries

This skill writes experience.md. That's the promotion layer. It does not:

• Modify the skill's SKILL.md body (that's auto-evolve's domain, which must commit via scripts/committer and be reversible via git revert).
• Modify judge.py or PROTOCOL.md (project-level fixed contracts; AVO must not change its own evaluation function).
• Delete or rewrite patterns.jsonl (append-only; only rotation is allowed).

Invocation Pattern

Runtime AI calls compact via a sync tool:

compact_experience("search-xiaohongshu")

Implementation (Standard tier LLM call):

1. Read autosearch/skills/channels/search-xiaohongshu/experience/patterns.jsonl.
2. Read existing autosearch/skills/channels/search-xiaohongshu/experience.md (if exists).
3. Group events by winning_pattern / good_query template; count seen and success.
4. Apply promotion thresholds; drop events older than 30 days for rule evaluation.
5. Build new experience.md body: keep old Active Rules that still pass threshold; add new ones that newly qualify; move expired ones to Failure Modes or drop; refresh ## Last Compacted line.
6. Write the new experience.md.
7. If patterns.jsonl > 1 MB, rotate.

Latency target: < 10 seconds, one LLM call for rule grouping if the patterns.jsonl is large.

What This Skill Is NOT

• Not a summarizer of the skill's purpose (that's the SKILL.md body).
• Not a decision-maker for which skills to call (that's the router + group-index layer).
• Not a replacement for AVO's auto-evolve (which can edit SKILL.md with human / automated review).

It is a safe, bounded, reversible digest writer that turns append-only event logs into a short, runtime-visible rule list.

Relationship to Other Skills

• Reads ← <leaf>/experience/patterns.jsonl (from experience-capture).
• Writes → <leaf>/experience.md.
• Feeds → runtime AI (which reads experience.md before calling the skill).
• Feeds → auto-evolve (which uses the rule history as input when proposing SKILL.md edits).

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.