autosearch:reflective-search-loop

Reflective Search Loop — Explicit Loop State

Most autosearch leaf skills are one-shot: call, get evidence, done. This meta skill codifies a multi-round loop that the runtime AI drives when the task is deeper than one query round can settle.

Loop State

loop:
  round: int                         # 0, 1, 2, ...
  budget_remaining:
    rounds: int                      # hard cap, e.g. 5
    cost_usd: float
    latency_seconds: int
  context:
    query: str                       # original user query
    rubrics: list[str]               # from run_clarify
  progress:
    gaps: list[str]                  # outstanding sub-questions, grows/shrinks
    answered_gaps: list[str]         # resolved this round
    all_questions: list[str]         # every question asked across all rounds
    visited_urls: list[str]          # URLs already fetched
    bad_urls: list[str]              # URLs that 404'd / were blocked / returned junk
    evidence: list[dict]             # accumulated Evidence items
    evaluator_failures: list[str]    # rubric/quality failures from previous round
  decision:
    next_action: "search_more" | "fetch_more" | "finalize" | "escalate_budget" | "abort"
    next_queries: list[str] | null
    next_urls_to_fetch: list[str] | null
    stop_reason: str | null

Round Structure

Each loop iteration:

1. Reflect (Best tier LLM): given current loop state, decide next_action.
2. Act: call run_channel(...) or fetch-jina/fetch-crawl4ai for the chosen next step.
3. Integrate: update progress (visited URLs, evidence list, new gaps discovered).
4. Evaluate (if enabled): run check-rubrics or evaluate-delivery; record failures in evaluator_failures.
5. Loop or stop: check budget + stop conditions.

Stop Conditions

ANY of:

• round >= budget_remaining.rounds (hard cap).
• len(gaps) == 0 AND len(evaluator_failures) == 0 (task complete).
• len(new_evidence_this_round) == 0 for 2 consecutive rounds (stalled).
• cost_usd > allocated budget.
• User explicitly aborts.