AgentSkillsCN

fp-pattern-finder-signals

端到端工作流：从标签中识别FP模式，深入探究典型案例，并生成调优方案。

SKILL.md
--- frontmatter
name: fp-pattern-finder-signals
description: 'End-to-end workflow: detect FP patterns from labels, investigate examples,
  and generate a tuning proposal.'
user-invocable: true
allowed-tools:
- Bash
- Read

fp-pattern-finder-signals

When to use

  • •You want to reduce FP rate and need concrete, data-driven tuning ideas.
  • •You have new labels (manual or outcome backfill) and want to mine for patterns.

Inputs

  • •days window and thresholds for pattern detection.
  • •output paths for patterns JSON and proposal YAML.

Workflow

  1. •
    1. •Ensure labels exist: run outcome backfill and/or manual labeling.
  2. •
    1. •Detect patterns: python -m ops.cli quality find-patterns --days 30 --out /tmp/patterns.json.
  3. •
    1. •Generate tuning proposal: python -m ops.cli quality propose-tuning --patterns /tmp/patterns.json --out /tmp/proposal.yaml.
  4. •
    1. •Review proposal notes + actions, then (optionally) apply safe patches with quality apply-tuning.

Outputs

  • •patterns JSON report + tuning proposal YAML.

Guardrails

  • •Do not auto-apply tuning without review; most actions are suggestions, not deterministic fixes.
  • •Validate impact with quality stats after deploying changes.

References

  • •references/reference.md
  • •docs/QUALITY_OPS_ARCHITECTURE.md