Query SwanLab experiment metrics and run status. Use when checking training/producer metrics, comparing runs, or investigating experiment health. Keywords: swanlab, metrics, experiment, loss, reward, training progress, wandb.
Launch a background agent to monitor a long-running remote job (K8s, SLURM, CI). Use when submitting jobs that need async monitoring. Keywords: monitor, watch job, background agent, submit and watch, check job status.
Generate a weekly report summarizing commits and PRs across the main repo and all external/ submodules. Keywords: weekly report, 周报, work summary, 工作总结.
Capture reusable knowledge after milestones. Systematically review what was learned and persist it to the right layer. Keywords: sediment, crystallize, capture, persist, remember, learned, milestone, retrospective.
Hypothesis-first bug diagnosis. Verify the user's diagnosis before proposing alternatives. Keywords: bug, diagnose, debug, fix, hypothesis, root cause, crash, error, timeout, traceback.
Build a first-failure timeline for distributed incidents. Use when debugging crashes, hangs, watchdog timeouts, collective mismatches, or cascading errors where the last error is often not the root cause.
Convert local debugging results into high-quality upstream issues or RFC comments. Use when filing bugs to external projects with clear reproduction, first-failure evidence, root-cause hypothesis, and scoped fix suggestions.
Use when the user asks or you need codex'suggestion to run Codex CLI (codex exec, codex resume) or references OpenAI Codex for code analysis, refactoring, or automated editing