Kirrito-k423

Enforce AutoResearch loop cadence, attempt, token, time, sub-agent, remote-compute, and kill-switch limits.

Load and enforce AutoResearch loop constraints before triage, edits, remote work, or external writes.

Read AutoResearch run evidence and loop state, then propose one bounded next experiment or escalate. L1 is report-only.

从 Verl examples 脚本和示例 YAML 中维护“可控参数即特性”的中文情报库与 Excel 台账。Use when Codex needs to scan Verl examples/run scripts, extract all user-adjustable parameters and Hydra overrides, create or update per-parameter markdown explanation files under docs/verl/features/lists, use subagents to research unclear parameters from local explanation files, official Verl docs, and repo docs, and produce docs/verl/features/verl-example-parameters.xlsx with parameter name, category, Chinese explanation, common values, performance impact, accuracy impact, and example count.

Generate RMB task cost reports from Codex/OpenAI-style token usage, including total input tokens, cached input tokens, uncached input tokens, output tokens, and GPT plus DeepSeek API cost estimates. Use when the user asks to统计 token 消耗, API 费用, 人民币成本, cache hit 成本, GPT/DeepSeek 对比, or to produce an `RMB-Cost.md` report for a task, goal, session, or experiment. Use by default for any goal, experiment, remote run, long command, or automation whose expected or actual runtime exceeds 20 minutes.

Verl GRPO formal case adapter for AutoResearch. Use when building, running, or diagnosing autoresearch run verl-case; preparing Qwen/geo3k GRPO matrices; wiring Verl containers to model/data assets, W&B, Prometheus, reports, provenance, and numbered evidence bundles; or explaining val-only versus real GRPO training boundaries.

Generate, validate, inspect, and safely handle AutoResearch customer configuration files. Use when working on config init/show/validate commands, Pydantic config schema behavior, keyring or env secret placeholders, redacted display, or config/config.yaml templates.

Start, stop, and health-check local AutoResearch services including Archon, W&B, Prometheus, and Grafana. Use when working on autoresearch services commands, service compose files, local healthz checks, fixed ports, or local service orchestration.

Classify GitHub issues with configurable labels and bounded processed state. Use when Codex needs to process new or unrecorded issues in batches, assign labels such as bug, Ascend, Hardware, question, or enhancement, optionally apply labels only after explicit apply=true, and maintain a reusable processed issue list for repositories such as verl-project/verl.

用中文维护 GitHub issue 自动关闭评估与有界审计状态。Use when Codex needs to inspect oldest unrecorded GitHub issues, describe each issue's current state in Chinese, decide whether automatic closure is safe, draft a respectful close comment, close only after explicit apply=true, and maintain reusable Chinese close records for repositories such as verl-project/verl.

Draft and optionally post academic bilingual GitHub issue replies with bounded reply metadata. Use when Codex needs to inspect oldest unrecorded issues, decide whether a reply is appropriate, store each English-then-Chinese reply body in a separate file, optionally comment only after explicit apply=true, and maintain a reusable reply table for repositories such as verl-project/verl.

Publish bounded GitHub issue automation summaries to a tracking issue. Use when Codex needs to read close, classify, and reply process files, create or update an official automation tracking issue, replace a managed report block, and summarize auto-close reasons plus aggregate classification and reply counts for repositories such as verl-project/verl.

并行研究一个或多个 GitHub、GitCode 或本地代码仓，基于源码、CodeGraph、仓内文档、模型脚本、已合入 PR/MR 和官方网站构建全量能力证据库，并为每个仓生成模型、算法、优化特性地图、三张 PNG 一图流与需求洞察报告。用于代码仓能力盘点、跨仓对比、模型与算法支持矩阵、模型专项优化与已合入 PR 分析、竞争力分析、需求洞察或可大规模推广特性识别。

对两个及以上代码仓进行全量能力支持度联评：复用或自动补齐 report/仓名/map.md，按模型、算法、优化特性取并集，以红黄绿三态生成每仓一列的支持矩阵，并评估模型时效性、算法流行度、资源/吞吐收益及迁移代码量和人力。用于多仓选型、能力差距分析、特性移植规划、竞品矩阵、模型/算法/优化覆盖对比；当任一仓缺少能力地图时先调用 repo-capability-insight 生成。

Operate PJLab pod web terminal via Chrome DevTools MCP: connect, authenticate, simulate keyboard input via CDP, redirect output to mounted dirs, read from IDE. Invoke when user needs to run commands on PJLab remote pod.

每日从 AI HOT 公共只读接口和 DeepSpeed 官方教程、Release、Latest News、Examples 发现并筛选 SOTA 模型、训练与推理技术、AI 芯片和集群硬件突破，回到论文、官方仓库、模型卡、官方博客或规格页深入核验，再按 AIHisTrendFuture 的证据契约更新研究数据、时间线、比较页、GitHub 仓库和在线 Sites 站点。用户要求执行或补跑每日 AI 前沿扫描、核验某日热点是否值得入库、更新技术时空图谱、发布研究更新时使用；也适用于无人值守的日常自动化。