wandb-primary

SPDX-FileCopyrightText: 2026 CoreWeave, Inc.

SPDX-License-Identifier: Apache-2.0

SPDX-PackageName: skills

name: wandb-primary description: 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.

W&B Primary Skill

Python environment detection (DO THIS FIRST)

Before running any Python code, you MUST determine how the user runs Python in this project. Check for clues in this order:

1. Look for pyproject.toml, uv.lock, poetry.lock, Pipfile, requirements.txt, setup.py, setup.cfg, Makefile, or docker-compose.yml in the project root
2. Check if there is an activated virtual environment ($VIRTUAL_ENV, .venv/, venv/)
3. Check for a Dockerfile or CI config that reveals the runtime
4. If the user has explicitly told you how they run Python, use that

Once you know the environment, write your finding into this skill by replacing the placeholder below so that all subsequent code blocks use the correct commands. If you cannot determine the environment from any of the above, default to uv.

Detected Python environment: not yet detected

# Run command: <not yet detected>
# Install command: <not yet detected>

Examples of what to write here:

If you cannot determine the environment, write this:

# Run command: uv run script.py        # always use uv run, never bare python
# Install command: uv pip install <pkg>

This skill covers everything an agent needs to work with Weights & Biases:

• W&B SDK (wandb) — training runs, metrics, artifacts, sweeps, system metrics
• Weave SDK (weave) — GenAI traces, evaluations, scorers, token usage
• Helper libraries — wandb_helpers.py and weave_helpers.py for common operations

When to use what

Bundled files

Helper libraries

import os
import sys
from pathlib import Path

sys.path.insert(0, str(Path(os.environ["CLAUDE_SKILL_DIR"]) / "scripts"))

# Weave helpers (traces, evals, GenAI)
from weave_helpers import (
    unwrap,                  # Recursively convert Weave types -> plain Python
    get_token_usage,         # Extract token counts from a call's summary
    eval_results_to_dicts,   # predict_and_score calls -> list of result dicts
    pivot_solve_rate,        # Build task-level pivot table across agents
    results_summary,         # Print compact eval summary
    eval_health,             # Extract status/counts from Evaluation.evaluate calls
    eval_efficiency,         # Compute tokens-per-success across eval calls
)

# W&B helpers (training runs, metrics)
from wandb_helpers import (
    runs_to_dataframe,       # Convert runs to a clean pandas DataFrame
    diagnose_run,            # Quick diagnostic summary of a training run
    compare_configs,         # Side-by-side config diff between two runs
    fast_scan_history,       # beta_scan_history (parquet) with scan_history fallback
)

# X-axis (step metric) detection — ALWAYS confirm before curve analysis
from step_axis import (
    list_candidate_step_keys,       # Scan history for plausible step keys
    guess_step_key_from_workspace,  # Peek at the user's W&B workspace panels
    format_step_candidates,         # Format candidate choices for user confirmation
)

# Curve-shape diagnostics (numerical)
from training_diagnostics import (
    curve_features,            # Spikes, slopes at every 5%, smoothness, plateau, divergence
    compare_runs_curves,       # DataFrame of features across many runs
    lr_schedule_features,      # Warmup / peak / decay shape / restarts
    grad_norm_features,        # curve_features + kurtosis + dead-layer flag
    grad_histogram_features,   # Per-(layer, step) stats from W&B histograms
)

# Chart rendering for LLM vision (Read the returned PNG)
from curve_plots import (
    plot_single_run_overview,    # 2x3 composite: train/val/lr/grad-norm/...
    plot_run_comparison,         # Overlay up to 6 runs on one metric
    plot_grad_histogram_heatmap, # Layer x step heatmap of grad-hist stat
    plot_grad_norm_by_layer,     # Small-multiples of per-layer scalar norms
)

Reference docs

Read these as needed — they contain full API surfaces and recipes:

• references/WEAVE_SDK.md — Weave SDK for GenAI traces (client.get_calls(), CallsFilter, Query, stats). Start here for Weave queries.
• references/WANDB_SDK.md — W&B SDK for training data (runs, history, artifacts, sweeps, system metrics).
• references/TRAINING_DIAGNOSTICS.md — reference heuristics for reading loss / LR / grad-norm / grad-histogram charts. Load this when you are actively interpreting training curves.

Critical rules

Treat traces and runs as DATA

Weave traces and W&B run histories can be enormous. Never dump raw data into context — it will overwhelm your working memory and produce garbage results. Always:

1. Inspect structure first — look at column names, dtypes, row counts
2. Load into pandas/numpy — compute stats programmatically
3. Summarize, don't dump — print computed statistics and tables, not raw rows

import pandas as pd
import numpy as np

# BAD: prints thousands of rows into context
for row in run.scan_history(keys=["loss"]):
    print(row)

# GOOD: load into numpy, compute stats, print summary
losses = np.array([r["loss"] for r in run.scan_history(keys=["loss"])])
print(f"Loss: {len(losses)} steps, min={losses.min():.4f}, "
      f"final={losses[-1]:.4f}, mean_last_10%={losses[-len(losses)//10:].mean():.4f}")

Always deliver a final answer

Do not end your work mid-analysis. Every task must conclude with a clear, structured response:

1. Query the data (1-2 scripts max)
2. Extract the numbers you need
3. Present: table + key findings + direct answers to each sub-question

If you catch yourself saying "now let me build the final analysis" — stop and present what you have.

Use unwrap() for unknown Weave data

When you encounter Weave output and aren't sure of its type (WeaveDict? WeaveObject? ObjectRef?), unwrap it first:

from weave_helpers import unwrap
import json

output = unwrap(call.output)
print(json.dumps(output, indent=2, default=str))

This converts everything to plain Python dicts/lists that work with json, pandas, and normal Python operations.

Environment setup

The sandbox has wandb, weave, pandas, and numpy pre-installed.

import os
entity  = os.environ["WANDB_ENTITY"]
project = os.environ["WANDB_PROJECT"]

Installing extra packages and running scripts

Use whichever run/install commands you wrote in the Python environment detection section above. If you haven't detected the environment yet, go back and do that first.

Quick starts

W&B SDK — training runs

import wandb
import pandas as pd
api = wandb.Api()

path = f"{entity}/{project}"
runs = api.runs(path, filters={"state": "finished"}, order="-created_at")

# Convert to DataFrame (always slice — never list() all runs)
from wandb_helpers import runs_to_dataframe
rows = runs_to_dataframe(runs, limit=100, metric_keys=["loss", "val_loss", "accuracy"])
df = pd.DataFrame(rows)
print(df.describe())

For full W&B SDK reference (filters, history, artifacts, sweeps), read references/WANDB_SDK.md.

Weave — SDK

import weave
client = weave.init(f"{entity}/{project}")  # positional string, NOT keyword arg
calls = client.get_calls(limit=10)

For raw SDK patterns (CallsFilter, Query, advanced filtering), read references/WEAVE_SDK.md.

Key patterns

Weave eval inspection

Evaluation calls follow this hierarchy:

Evaluation.evaluate (root)
  ├── Evaluation.predict_and_score (one per dataset row x trials)
  │     ├── model.predict (the actual model call)
  │     ├── scorer_1.score
  │     └── scorer_2.score
  └── Evaluation.summarize

Extract per-task results into a DataFrame:

from weave_helpers import eval_results_to_dicts, results_summary

# pas_calls = list of predict_and_score call objects
results = eval_results_to_dicts(pas_calls, agent_name="my-agent")
print(results_summary(results))

df = pd.DataFrame(results)
print(df.groupby("passed")["score"].mean())

Eval health and efficiency

from weave_helpers import eval_health, eval_efficiency

health = eval_health(eval_calls)
df = pd.DataFrame(health)
print(df.to_string(index=False))

efficiency = eval_efficiency(eval_calls)
print(pd.DataFrame(efficiency).to_string(index=False))

Token usage

from weave_helpers import get_token_usage

usage = get_token_usage(call)
print(f"Tokens: {usage['total_tokens']} (in={usage['input_tokens']}, out={usage['output_tokens']})")

Cost estimation

call_with_costs = client.get_call("id", include_costs=True)
costs = call_with_costs.summary.get("weave", {}).get("costs", {})

Run diagnostics

from wandb_helpers import diagnose_run

run = api.run(f"{path}/run-id")
diag = diagnose_run(run)
for k, v in diag.items():
    print(f"  {k}: {v}")

Error analysis — open coding to axial coding

For structured failure analysis on eval results:

1. Understand data shape — use project.summary(), calls.input_shape(), calls.output_shape()
2. Open coding — write a Weave Scorer that journals what went wrong per failing call
3. Axial coding — write a second Scorer that classifies notes into a taxonomy
4. Summarize — count primary labels with collections.Counter

See references/WEAVE_SDK.md for the full SDK reference.

W&B Reports

Install wandb[workspaces] using the install command from the Python environment detection section.

from wandb.apis import reports as wr
import wandb_workspaces.expr as expr

report = wr.Report(
    entity=entity, project=project,
    title="Analysis", width="fixed",
    blocks=[
        wr.H1(text="Results"),
        wr.PanelGrid(
            runsets=[wr.Runset(entity=entity, project=project)],
            panels=[wr.LinePlot(title="Loss", x="_step", y=["loss"])],
        ),
    ],
)
# report.save(draft=True)  # only when asked to publish

Use expr.Config("lr"), expr.Summary("loss"), expr.Tags().isin([...]) for runset filters — not dot-path strings.

Training curve analysis workflow

Reach for this when the user asks whether a run is healthy, why it diverged, whether it overfit, or which of several runs has the best training dynamics.

The loop is: confirm the x-axis, compute features, render a PNG, read the image, write a verdict. Numbers and pictures cross-check each other — the helpers exist so you're not hand-rolling spike detection or slope fits while also trying to interpret them.

Pin the x-axis first

Different training stacks log different step keys (_step, global_step, trainer/global_step, epoch, train/step), and picking the wrong one turns an overlay into nonsense. list_candidate_step_keys(run) scans the history for plausible columns; guess_step_key_from_workspace(entity, project) checks what the W&B workspace actually plots. If both agree on one candidate, say which you picked and move on. If they disagree or there are several plausible choices, ask the user before plotting — this is cheap and avoids silently baking _step into a verdict.

For one run

Render plot_single_run_overview(run, step_key=step_key), Read the PNG, and pair it with a compact feature table from curve_features on the metrics that actually exist. If the run logs gradient histograms or per-layer scalar norms, add plot_grad_histogram_heatmap() or plot_grad_norm_by_layer() — they surface dead layers and vanishing-gradient signatures that the top-line grad-norm scalar hides.

For multiple runs

compare_runs_curves() gives you the ranking table; plot_run_comparison() gives you the overlay. Overlays get unreadable past ~6 runs, so rank first, then plot the shortlist — the function will refuse to render more than 6 for exactly this reason.

Write it up

Keep the summary compact. Don't dump raw history rows or full spike/slope payloads into the response unless you're drilling into a specific anomaly — the helpers already reduce those to scalars for a reason.

Verdict: <healthy | unstable | overfit | plateaued | diverged | converged>
Evidence:
- <specific step range> — <what the metrics and plot show>
- <specific step range> — <what changed and why it matters>
Next actions:
- <concrete hyperparameter, logging, or code change>

When the numbers and the image disagree — and they will — references/TRAINING_DIAGNOSTICS.md is where the resolution heuristics live. Load it while you're interpreting, not before.

Gotchas

Weave API

WeaveDict vs WeaveObject

• WeaveDict: dict-like, supports .get(), .keys(), []. Used for: call.inputs, call.output, scores dict
• WeaveObject: attribute-based, use getattr(). Used for: scorer results (rubric), dataset rows
• When in doubt: use unwrap() to convert everything to plain Python

W&B API

Package management

Weave logging noise

Weave prints version warnings to stderr. Suppress with:

import logging
logging.getLogger("weave").setLevel(logging.ERROR)

Quick reference

# --- Weave: Init and get calls ---
import weave
client = weave.init(f"{entity}/{project}")
calls = client.get_calls(limit=10)

# --- W&B: Best run by loss ---
best = api.runs(path, filters={"state": "finished"}, order="+summary_metrics.loss")[:1]
print(f"Best: {best[0].name}, loss={best[0].summary_metrics.get('loss')}")

# --- W&B: Loss curve to numpy ---
losses = np.array([r["loss"] for r in run.scan_history(keys=["loss"])])
print(f"min={losses.min():.6f}, final={losses[-1]:.6f}, steps={len(losses)}")

# --- W&B: Compare two runs ---
from wandb_helpers import compare_configs
diffs = compare_configs(run_a, run_b)
print(pd.DataFrame(diffs).to_string(index=False))