菜单
Change Ola PPO credit assignment, reward scale, greedy feedback, or log_prob. Use when the user mentions AI策略, PPO, evolution_pref, Hamming, log_prob, reward, or greedy feedback.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/045200/Ola --skill ppo-credit命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Diagnose or change how the Ola APK detects the Magisk daemon. Use when the user mentions 引擎核心, daemon not running, ola.sock, nc -U, DaemonClient, or DaemonStatusProvider.
Run Ola without the AdaptiveTuner APK. Use when the user mentions 卸载 APK, 移除 APK, 全AI接管, headless, or daemon-only policy.
Fix or change Ola launch-splash hang (freeze/cpuset of a newly opened app). Use when the user mentions 启动图, splash, 前台卡, 打开慢, freeze, apply_freeze, or launch grace.
正在显示 SKILL.md
| name | ppo-credit |
| description | Change Ola PPO credit assignment, reward scale, greedy feedback, or log_prob. Use when the user mentions AI策略, PPO, evolution_pref, Hamming, log_prob, reward, or greedy feedback. |
Online writes go through learner.step → decode_action → SafetyGuard.filter_action → apply_params_with_rollback. Do not expand ACTION_DIM.
encode_params_to_pref is an offset in [-0.2, 0.2] relative to 0.5. fuse_evolution_pref / apply_evolution_pref adds 0.1 * pref. Never mix pref as a [0,1] action (0.8 must not be pulled toward 0.2).apply_params_with_rollback returns (applied, rollback, ...). If applied is false (interval or unchanged params), do not push_transition and do not overwrite last_action.filter_action, re-encode with encode_params_to_action so buffer / last_action match the landed params.learner.sync_last_action and tuning::sync_applied_params so Hamming and LAST_PARAMS follow the verified set.cpu_usage is [0,1] (global_cpu_usage()/100). Idle gate is < 0.15. Power penalty is > 0.80. GPU busy is 0-100.compute_log_prob_scalar and compute_log_prob_batch both clamp (z-μ)/σ to ±10.E_{a~π_old}[log π_old - log π_new], not the reversed difference.0.9*old + 0.1*new (ema_alpha=0.1).Follow session-memory: /home/wlq/wlq/record.sh with WLQ_DIR=/home/wlq/wlq.