wandb

Create or improve Senpai target-repository onboarding files: program.md plus instructions/prompt-advisor.md and instructions/prompt-student.md. Use this skill whenever the user wants to point Senpai at a fresh ML or research target repository, define the research objective, primary metric, benchmark contract, allowed edit boundaries, W&B reporting contract, advisor/student prompts, or prepare a repo for autonomous advisor/student experiment loops.

GitHub CLI primitives for the senpai research workflow — label swaps, send-back, close, mark-review, issue checks, PR queries. Use this skill whenever you need to manipulate PR labels, send a PR back to a student, close a dead-end experiment, mark a PR for review, or query the current state of PRs and issues. Also triggers for: "swap labels", "send back to student", "close this PR", "mark for review", "check human issues", "list review-ready PRs", "idle students".

Survey all experiment PRs on a branch and return a structured status report: which students are idle, which PRs await review, which are WIP. This is the heartbeat query — use it to understand the current state of the research track. Triggers for: "survey state", "check PR status", "who's idle", "any PRs ready for review", "what's the current state".

Squash-merge a winning experiment PR and update the baseline. Handles the merge, BASELINE.md update, commit, push, and branch pull. Also handles merge conflicts by sending the PR back for rebase. Use this skill to: merge a winning PR, update baseline, squash merge experiment. Triggers for: "merge winner", "merge this PR", "update baseline after merge", "squash merge".

Submit experiment results for advisor review. Commits changes, pushes the branch, marks the PR as ready, and swaps the status label from wip to review. Use this skill when you've finished running experiments and posted your results comment. Triggers for: "submit for review", "mark PR ready", "send results to advisor", "submit experiment results".

Comprehensive primary skill for agents working with Weights & Biases. Covers both the W&B SDK (training runs, metrics, artifacts, sweeps) and the Weave SDK (GenAI traces, evaluations, scorers). Includes helper libraries, gotcha tables, and data analysis patterns. Use this skill whenever the user asks about W&B runs, Weave traces, evaluations, training metrics, loss curves, model comparisons, or any Weights & Biases data — even if they don't say "W&B" explicitly. Also trigger on training-curve diagnostics questions — run health, divergence, overfit/convergence/plateau, spikes, LR-schedule/grad-norm/grad-histogram reading, dead layers, step-axis choice, and run comparisons.

Poll for assigned experiment PRs for a student. Use this skill to: check for assignments, poll for work, see if there's a PR assigned to me. Triggers for: "any work for me?", "check for assignments", "poll for PRs".

Produce a fresh status report for the senpai ML experiment fleet. Use when the user asks for an experiment status, final status, PR/W&B/pod health check, stale student triage, training shutdown harvest, advisor-state audit, or a "what is really happening right now?" report. The report must prioritize paper-facing test metrics over validation metrics and compare test results to dataset benchmarks or targets.

This skill should be used when the user wants to "configure weave", "set weave project", "change weave project", "set wandb api key", "update weave settings", "show weave config", "change weave configuration", or needs to read or update any Weave Claude Code plugin settings.

This skill should be used when the user wants to "install the weave plugin", "set up weave", "install weave-claude-plugin", "configure weave for the first time", "get started with weave tracing", or needs to complete the initial setup of the Weave Claude Code plugin including dependency installation and project configuration.

This skill should be used when the user wants to "check weave status", "verify the weave plugin is running", "see if weave is set up correctly", "check weave configuration", "is weave working", or needs to diagnose why Claude Code sessions are not appearing in Weave.

Bump the Weave Python SDK version for release. Use when preparing a new release.

Build and publish the Weave Python SDK to PyPI. Use when releasing a new version.

Analyze one or more Horangi model configs against (1) HuggingFace model card claims, (2) the W&B leaderboard rankings, and (3) category-level peer comparison. Invoke when the user asks to analyze/compare model performance for a config (e.g. "analyze <config>", "compare X and Y", "이 모델 성능 분석해줘").

Deep-dive error pattern analysis for a single (model, benchmark) pair using Weave traces. Surfaces how/why the model is getting answers wrong — answer bias, format violations, language mixing, and 3-5 representative failure samples. Invoke when the user asks to analyze wrong answers / failure patterns for a specific benchmark (e.g. "analyze errors in <bench>", "<model>의 <benchmark> 오답 패턴 분석", "틀린 문제 경향"). Commonly invoked as a follow-up to `horangi-analyze` when that skill flags a weak category.

Primary skill for querying, analyzing, and launching W&B projects. Covers W&B SDK (training runs, metrics, artifacts, sweeps, reports), Weave SDK (GenAI traces, evaluations, scorers, monitors), and W&B Launch (reproducing runs, submitting training jobs to compute, queue management). Use this for: project overviews, run analysis, trace inspection, eval summaries, report creation, monitor setup, launching/relaunching runs, and any broad 'what's going on' questions.

Comprehensive skill for adding W&B Weave to existing applications. Covers trace-first instrumentation, evaluation only after trace verification, documentation-first implementation, CLI-based Weave data access, and validation workflows. Activate this skill only when the user explicitly mentions the skill by name, such as `weave-integration`, `@skills weave-integration`, or another direct skill reference. Do not auto-trigger from generic Weave or W&B questions.