返回 Skill 列表
extension
分类: AI Agent 能力无需 API Key

auto-balancer

在明确的约束条件下,自动调整游戏/系统参数以达到目标指标。当迭代配置权重、赔付表、触发率或其他平衡杠杆时使用;根据模拟输出运行平衡循环;验证容差门限;以及准备通过/失败平衡签核文件。

person作者: jakexiaohubgithub

Auto Balancer

Use this skill to run controlled parameter tuning loops with deterministic validation gates.

Workflow

  1. Define balancing contract.
  • Declare target metrics, tolerances, hard constraints, and stop conditions.
  • Declare which parameters are allowed to move and their bounds.
  1. Establish baseline and iteration plan.
  • Record baseline metrics before tuning.
  • Apply small, traceable parameter changes per iteration.
  • Track config hash/version for each run.
  1. Run balance loop.
  • Execute simulation/evaluation runs.
  • Compare observed metrics to targets and compute deltas.
  • Keep only changes that improve objective without violating hard constraints.
  1. Validate gate conditions.
  • Check each metric against tolerance range.
  • Fail immediately on hard-constraint breaches.
  • Require minimum run count before final pass.
  1. Prepare sign-off handoff.
  • Return final parameter set, metric table, and failed/passed gates.
  • Include patch plan and exact verification commands.

Commands

python3 scripts/validate_balance_runs.py \
  --input <path/to/balance_runs.json> \
  --spec <path/to/target_spec.json>

Treat non-zero exits as blocker results.

Output Contract

Return:

  1. Target Contract: metrics, tolerances, and constraints.
  2. Run Summary: baseline, best run, and final run deltas.
  3. Gate Results: pass/fail per metric and per hard constraint.
  4. Patch Plan: exact files/params to update.
  5. Residual Risks: unresolved drift or instability concerns.

References

  • references/workflow.md: balancing process and iteration order.
  • references/metric-rules.md: tolerance and hard-constraint rules.
  • references/signoff-template.md: balancing sign-off template.

Execution Rules

  • Keep balancing changes bounded and reversible.
  • Keep hard constraints non-negotiable.
  • Keep baseline comparison in every report.
  • Flag non-convergent loops as blockers.