autosearch:trace-harvest

Trace Harvest — Distill Successful Sessions

MiroThinker's collect-trace + DeepResearchAgent's general_memory_system patterns adapted for autosearch. Analyzes a session's tool-call trace, identifies what worked, and turns it into promote-candidate patterns for experience-capture to store.

Input

A completed autosearch session's trace:

trace:
  session_id: str
  query: str
  clarify_result: ClarifyResult (from run_clarify)
  tool_calls: list[{name, args, result_summary, success, latency_ms, cost}]
  final_answer: str | null
  user_feedback: "accepted" | "rejected" | "unknown" | null
  rubrics_passed: list[str]
  rubrics_failed: list[str]

Selection Rules

Only harvest traces where ANY of:

1. user_feedback == "accepted" (explicit positive).
2. len(rubrics_passed) >= 0.6 * (rubrics_passed + rubrics_failed) (objective quality gate).
3. Session finished within budget AND generated >= 5 unique evidence citations.

Reject traces where:

• User rejected the final answer.
• Session hit a budget-exhausted error.
• More than 3 tool calls failed consecutively.

Harvest Output

Per successful trace, emit patterns keyed by leaf skill:

{
  "ts": "...",
  "session_id": "...",
  "skill": "search-xiaohongshu",
  "winning_pattern": "brand + pain_word + recent_date_window",
  "good_query": "品牌 痛点词 近30天",
  "context": {
    "task_domain": "product-research",
    "clarify_had_rubrics": 4,
    "subsequent_channels_used": ["search-douyin", "search-zhihu"]
  },
  "metrics": {
    "relevant_out_of_returned": "9/18",
    "led_to_user_acceptance": true
  },
  "promote_candidate": true,
  "trace_ref": "traces/<session_id>.jsonl"
}

Each emitted pattern is appended via experience-capture to the relevant skill's patterns.jsonl. Trace-harvest does NOT write experience.md directly — that's experience-compact's job.

When Run

• At session end (if session had user_feedback signal).
• Batch mode: process the trace archive nightly, harvest successful sessions from the last 24h.