AutoResearch

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.

Probe remote GPU/NPU server hardware over SSH and parse npu-smi or nvidia-smi outputs. Use when collecting accelerator inventory, driver versions, memory/core utilization, process occupancy, partial hardware data, or hardware probe CLI behavior.

Check local and remote network access for AutoResearch, including baidu, Hugging Face, GitHub speed probes and SSH proxy/tunnel setup. Use when diagnosing external network reachability, 127.0.0.1:7890 proxy fallback, remote no-network paths, or network-check CLI behavior.

Verify that remote training servers can reach local AutoResearch services such as W&B, Prometheus, Grafana, and Pushgateway. Use when testing service URLs from a remote host, validating SSH reverse proxy behavior, or implementing reachability CLI results.

Inspect remote training stack health for conda envs, verl or veomni versions, and minimal one-step training smoke tests. Use when validating training dependencies, stack probe outputs, minimal runner behavior, or train stack readiness before data collection.

Collect AutoResearch experiment evidence from minimal or formal runs, including W&B offline sync, remote logs, Prometheus evidence, manifests, config locks, and local-first run bundles. Use when implementing collect commands, run evidence persistence, bundle-local fallback, or formal Verl case data collection.

Render AutoResearch experiment reports from manifests, logs, W&B data, Prometheus evidence, and formal Verl case outputs. Use when building or diagnosing report HTML, Chinese report copy, bundle-local artifact loading, validation matrix explanations, or missing metric diagnostics.

Manage AutoResearch model weights and dataset assets. Use when downloading, locating, staging, recording, auditing, or cleaning model/data caches such as Qwen3.5-2B or geo3k; when editing config/data.yaml; when deciding local versus remote storage; or when an adapter needs dataset/model paths, ModelScope/Hugging Face source priority, proxy behavior, or remote asset registry rules.

Clean AutoResearch or local W&B run directories that have no successful step evidence. Use when asked to clean wandb data, remove empty or failed W&B runs, delete W&B logs without step/validation/success metrics, or preserve only runs that contain at least one useful step signal such as output.log step metrics, wandb-summary metrics, passed_rows, val-core/val-aux metrics, or one-step probe summary values.