wandb-primary

name: wandb-primary description: "Primary W&B skill for broad or mixed Weights & Biases work: project overviews, W&B runs and artifacts, Weave traces and evaluations, Reports, Signal Builder, and Launch workflows. Use when the task spans multiple W&B surfaces or the user asks generally what is happening in a W&B project."

W&B Primary Skill

Environment defaults

• Python: run scripts with the Python environment available to your coding agent. Install missing optional packages only when needed.
• Credentials: use WANDB_API_KEY, WANDB_ENTITY, and WANDB_PROJECT from the user's environment or prompt.

Fast recipes — use these first

These cover the most common tasks. Each is a single script. Copy, fill in placeholders, run.

Fast product/API answers

For small W&B product or API questions, answer directly from this section. Do not run tools, inspect docs, or query the user's project unless they explicitly ask for live data. Keep the answer short: direct answer, exact UI/API path, minimal code if useful. If the recommendation depends on missing context, include targeted diagnostic questions in the same response instead of blocking.

For workspace migration or project-structure guidance, ask the diagnostic questions before prescribing a structure or script. Use the phrase "Before I prescribe a structure/script, I need to know:" and include the questions that materially change the answer; then give only tentative guidance.

Product facts to answer from memory

Trace-count semantics

Use these rules before every Weave count query:

• "total traces" or "total calls" means all calls. Use calls_query_stats with no trace_roots_only filter. Do not deduplicate by trace_id unless the prompt asks for unique traces.
• "root traces", "root-level traces", or "traces with no parent" means root calls. Use filter={"trace_roots_only": True} only for those prompts.
• "successful/non-error traces" means total calls minus calls with status error / descendant_error / non-null exception; report that as the primary count. summary.weave.status == "success" is a useful supporting breakdown, but it excludes running calls, which are still non-error. Do not count only root traces unless the user says root/root-level.
• "error/exception traces" means calls with status error OR descendant_error OR a non-null exception. For root-level error counts, add trace_roots_only=True to that same error query.
• Evaluation.evaluate counts are op counts. Use an op_names filter for weave:///<entity>/<project>/op/Evaluation.evaluate:*. Add trace_roots_only only if the user explicitly asks for root eval traces.
• For exact count tasks, run one script that prints the query and the number; do not run sample/exploratory scripts after the count is already known.

Eval-analysis rules

• Filter Evaluation.evaluate calls with op_names=[f"weave:///{entity}/{project}/op/Evaluation.evaluate:*"].
• Fetch only needed columns (id, display_name, started_at, ended_at, summary, inputs, output) and avoid broad object dumps.
• Eval token usage is in summary.usage; sum input_tokens, output_tokens, and total_tokens across model keys.
• Eval success/error counts are in summary.status_counts, not summary.weave.status_counts. Normalize enum and string keys before reading success, error, and descendant_error.
• For success-rate tasks, do not lead with a long 43-row markdown table. First answer with totals, both fractions, and a compact Error evaluations (N): TSV/code block containing every errored eval id, date, success_count, error_count, and status. If full per-eval rows are requested, use short IDs/dates/counts after the error list; avoid repeating long duplicate display names where they cause truncation. If some evals are still running, report both denominators: success-status evals over completed evals and no-error evals over all evals.
• Child dataset rows are Evaluation.predict_and_score:* calls with parent_ids=[eval_call.id].
• Dataset refs live on inputs["self"].dataset inside the Evaluation object. Count distinct dataset object refs from the user's project data; repeated evals can reuse the same dataset ref.
• For scorer inventories, eval summaries, and scorer evolution, include both wrapper scorer ops whose short names end in _scorer and class scorer ops ending in .score. Never filter only for the substring scorer; versioned class scorers like MyClassifier.score do not contain it.
• For large scorer inventories, include a compact full TSV/code block (scorer\tcount) for every scorer and then summarize family groupings. Do not use long prose tables that may truncate before all counts appear.

Count runs (exact, fast)

import wandb, os
api = wandb.Api(timeout=120)
path = f"{os.environ['WANDB_ENTITY']}/{os.environ['WANDB_PROJECT']}"
total = len(api.runs(path, per_page=1, include_sweeps=False, lazy=True))
finished = len(api.runs(path, filters={"state": "finished"}, per_page=1, include_sweeps=False, lazy=True))
crashed = len(api.runs(path, filters={"state": "crashed"}, per_page=1, include_sweeps=False, lazy=True))
running = len(api.runs(path, filters={"state": "running"}, per_page=1, include_sweeps=False, lazy=True))
print(f"Total: {total}  |  Finished: {finished}  |  Crashed: {crashed}  |  Running: {running}")

Run-count rules:

• Use one script for exact counts. If it prints the requested count, answer from that stdout; do not rerun just to add labels or nicer formatting.
• Use include_sweeps=False for normal run-table counts unless the prompt asks for sweep runs. For sweep counts, query sweeps explicitly.
• For status breakdowns, scan once and report all states you see (finished, failed, crashed, killed, etc.). When crashed/killed runs exist, report unsuccessful terminal rate (failed + crashed + killed) / total as the primary failure rate and include failed-only rate as a supporting number.
• For tags, count runs with at least one tag and also list distinct tag names and the runs attached to each tag.
• For run groups, report named groups from groupedRuns(groupKeys: ["group"]) and compute ungrouped runs as total_runs - sum(named_group_counts).
• For sweep-run tasks, list each sweep's run count and explicitly report the total runs across all sweeps.

Count/list sweeps

Do not inspect the W&B SDK source for routine sweep questions. Use the public project API directly:

import os, wandb

entity = os.environ["WANDB_ENTITY"]
project = os.environ["WANDB_PROJECT"]
api = wandb.Api(timeout=120)

sweeps = list(api.project(project, entity=entity).sweeps(per_page=50))
rows = []
for sweep in sweeps:
    config = sweep.config or {}
    metric = config.get("metric") or {}
    rows.append({
        "id": sweep.id,
        "state": sweep.state,
        "method": config.get("method"),
        "metric": metric.get("name"),
        "goal": metric.get("goal"),
        "run_count": len(sweep.runs),
    })

print(f"sweep_count={len(rows)}")
print(f"total_sweep_runs={sum(r['run_count'] for r in rows)}")
for r in rows:
    print(r)

Finished runs with trigger/user

For prompts asking who triggered each run, fetch the filtered runs once and read run.user.username / run.user.name; do not search reference files.

import os, wandb

entity = os.environ["WANDB_ENTITY"]
project = os.environ["WANDB_PROJECT"]
path = f"{entity}/{project}"
api = wandb.Api(timeout=120)

runs = api.runs(
    path,
    filters={"state": "finished"},
    order="+created_at",
    per_page=100,
    include_sweeps=False,
)
rows = []
for run in runs:
    user = getattr(run, "user", None)
    rows.append({
        "created_at": run.created_at,
        "name": run.display_name or run.name,
        "id": run.id,
        "username": getattr(user, "username", None),
        "user_name": getattr(user, "name", None),
    })

print(f"finished_count={len(rows)}")
for r in rows:
    print(r)

Count traces (fast, server-side)

import weave, os, logging
logging.getLogger("weave").setLevel(logging.ERROR)
from weave.trace_server.trace_server_interface import CallsQueryStatsReq
from weave.trace_server.interface.query import Query

entity = os.environ["WANDB_ENTITY"]
project = os.environ["WANDB_PROJECT"]
client = weave.init(f"{entity}/{project}")
pid = f"{entity}/{project}"

# Total calls/traces
stats = client.server.calls_query_stats(CallsQueryStatsReq(project_id=pid))
print(f"Total calls: {stats.count}")

# Root traces only
root_stats = client.server.calls_query_stats(CallsQueryStatsReq(
    project_id=pid, filter={"trace_roots_only": True}
))
print(f"Root traces: {root_stats.count}")

# Count by op name
for op in ["Evaluation.evaluate", "my_op.turn"]:
    op_ref = f"weave:///{entity}/{project}/op/{op}:*"
    s = client.server.calls_query_stats(CallsQueryStatsReq(
        project_id=pid,
        filter={"op_names": [op_ref]},
    ))
    print(f"  {op}: {s.count}")

# Count calls whose op_name contains a substring, e.g. scorer calls.
score_query = Query(**{"$expr": {"$contains": {
    "input": {"$getField": "op_name"},
    "substr": {"$literal": ".score"},
    "case_insensitive": True,
}}})
score_stats = client.server.calls_query_stats(CallsQueryStatsReq(
    project_id=pid, query=score_query
))
print(f"Scorer calls (.score): {score_stats.count}")

# Count a named op substring such as create_embeddings.
embedding_query = Query(**{"$expr": {"$contains": {
    "input": {"$getField": "op_name"},
    "substr": {"$literal": "create_embeddings"},
    "case_insensitive": True,
}}})
embedding_stats = client.server.calls_query_stats(CallsQueryStatsReq(
    project_id=pid, query=embedding_query
))
print(f"create_embeddings calls: {embedding_stats.count}")

# Error/exception calls. Include descendant_error when the prompt says
# "error status or exception"; those are traces whose children failed.
error_query = Query(**{"$expr": {"$or": [
    {"$eq": [{"$getField": "summary.weave.status"}, {"$literal": "error"}]},
    {"$eq": [
        {"$getField": "summary.weave.status"},
        {"$literal": "descendant_error"},
    ]},
    {"$not": [{"$eq": [{"$getField": "exception"}, {"$literal": None}]}]},
]}})
error_stats = client.server.calls_query_stats(CallsQueryStatsReq(
    project_id=pid, query=error_query
))
root_error_stats = client.server.calls_query_stats(CallsQueryStatsReq(
    project_id=pid, filter={"trace_roots_only": True}, query=error_query
))
print(f"Error/exception calls: {error_stats.count}")
print(f"Root error/exception calls: {root_error_stats.count}")
print(f"Non-error calls: {stats.count - error_stats.count}")

Count create_embeddings calls and input sizes

import os, statistics, weave, logging, sys
from collections import Counter
logging.getLogger("weave").setLevel(logging.ERROR)
from weave.trace_server.trace_server_interface import CallsQueryStatsReq
from weave.trace_server.interface.query import Query
sys.path.insert(0, "skills/wandb-primary/scripts")
from weave_helpers import unwrap

entity = os.environ["WANDB_ENTITY"]
project = os.environ["WANDB_PROJECT"]
pid = f"{entity}/{project}"
client = weave.init(pid)

query = Query(**{"$expr": {"$contains": {
    "input": {"$getField": "op_name"},
    "substr": {"$literal": "create_embeddings"},
    "case_insensitive": True,
}}})
total = client.server.calls_query_stats(CallsQueryStatsReq(
    project_id=pid, query=query
)).count

sizes = []
for call in client.get_calls(query=query, limit=total, columns=["inputs"]):
    inputs = unwrap(call.inputs)
    texts = inputs.get("texts") or inputs.get("input") or []
    if isinstance(texts, str):
        sizes.append(1)
    else:
        sizes.append(len(texts))

dist = Counter(sizes)
print(f"create_embeddings calls: {total}")
print(f"typical texts per call: {dist.most_common(1)[0][0] if dist else 0}")
print(f"distribution: {dict(sorted(dist.items()))}")
print(f"mean texts per call: {statistics.mean(sizes) if sizes else 0:.4f}")

Count feedback records

import os, weave, logging
logging.getLogger("weave").setLevel(logging.ERROR)
from weave.trace_server.trace_server_interface import FeedbackQueryReq

entity = os.environ["WANDB_ENTITY"]
project = os.environ["WANDB_PROJECT"]
pid = f"{entity}/{project}"
client = weave.init(pid)

limit = 1000
offset = 0
total = 0
while True:
    res = client.server.feedback_query(FeedbackQueryReq(
        project_id=pid,
        fields=["id"],
        limit=limit,
        offset=offset,
    ))
    rows = (
        getattr(res, "result", None)
        or getattr(res, "feedback", None)
        or getattr(res, "rows", None)
        or []
    )
    n = len(rows)
    total += n
    if n < limit:
        break
    offset += limit

print(f"Feedback records: {total}")

List root op names with counts

import os, weave, logging
from collections import Counter
logging.getLogger("weave").setLevel(logging.ERROR)
from weave.trace_server.trace_server_interface import CallsQueryStatsReq

entity = os.environ["WANDB_ENTITY"]
project = os.environ["WANDB_PROJECT"]
pid = f"{entity}/{project}"
client = weave.init(pid)
root_filter = {"trace_roots_only": True}

root_count = client.server.calls_query_stats(CallsQueryStatsReq(
    project_id=pid, filter=root_filter
)).count

def short_op(op_name: str) -> str:
    tail = op_name.split("/op/")[-1]
    return tail.rsplit(":", 1)[0]

counts = Counter()
for call in client.get_calls(
    filter=root_filter,
    limit=root_count,
    columns=["op_name"],
):
    counts[short_op(call.op_name)] += 1

for name, count in counts.most_common():
    print(f"{name}\t{count}")

List all op names with counts

import os, weave, logging
from collections import Counter
logging.getLogger("weave").setLevel(logging.ERROR)
from weave.trace_server.trace_server_interface import CallsQueryStatsReq

entity = os.environ["WANDB_ENTITY"]
project = os.environ["WANDB_PROJECT"]
pid = f"{entity}/{project}"
client = weave.init(pid)
total = client.server.calls_query_stats(CallsQueryStatsReq(project_id=pid)).count

def short_op(op_name: str) -> str:
    return op_name.split("/op/")[-1].rsplit(":", 1)[0]

counts = Counter()
for call in client.get_calls(limit=total, columns=["op_name"]):
    counts[short_op(call.op_name)] += 1

print(f"Unique ops: {len(counts)}")
for name, count in counts.most_common():
    print(f"{name}\t{count}")

Count long-duration traces

Do not try to do datetime arithmetic inside a Weave Query; stream timestamp columns and count locally.

import os, weave, logging
logging.getLogger("weave").setLevel(logging.ERROR)
from weave.trace_server.trace_server_interface import CallsQueryStatsReq

entity = os.environ["WANDB_ENTITY"]
project = os.environ["WANDB_PROJECT"]
pid = f"{entity}/{project}"
client = weave.init(pid)
total = client.server.calls_query_stats(CallsQueryStatsReq(project_id=pid)).count

threshold_s = 60
long_count = 0
scanned = 0
for call in client.get_calls(
    limit=total,
    columns=["started_at", "ended_at"],
):
    scanned += 1
    if call.started_at and call.ended_at:
        duration_s = (call.ended_at - call.started_at).total_seconds()
        if duration_s > threshold_s:
            long_count += 1

print(f"Scanned calls: {scanned}")
print(f"Duration > {threshold_s}s: {long_count}")

Find model names in traces

import os, weave, logging
from collections import Counter
logging.getLogger("weave").setLevel(logging.ERROR)
from weave.trace_server.trace_server_interface import CallsQueryStatsReq
import sys
sys.path.insert(0, "skills/wandb-primary/scripts")
from weave_helpers import unwrap

entity = os.environ["WANDB_ENTITY"]
project = os.environ["WANDB_PROJECT"]
pid = f"{entity}/{project}"
client = weave.init(pid)
total = client.server.calls_query_stats(CallsQueryStatsReq(project_id=pid)).count

def collect_models(obj, out):
    obj = unwrap(obj)
    if isinstance(obj, dict):
        for k, v in obj.items():
            if k == "model" and isinstance(v, str):
                out.append(v)
            collect_models(v, out)
    elif isinstance(obj, list):
        for item in obj:
            collect_models(item, out)

models = Counter()
for call in client.get_calls(
    limit=total,
    columns=["inputs", "output", "summary"],
):
    found = []
    collect_models(call.inputs, found)
    collect_models(call.output, found)
    usage = unwrap(call.summary).get("usage", {}) if call.summary else {}
    for model_name in usage:
        if isinstance(model_name, str):
            found.append(model_name)
    for model_name in set(found):
        models[model_name] += 1

for name, count in models.most_common():
    print(f"{name}\t{count}")

Analyze embedding dimensions and model

import os, weave, logging
from collections import Counter
logging.getLogger("weave").setLevel(logging.ERROR)
from weave.trace_server.interface.query import Query
import sys
sys.path.insert(0, "skills/wandb-primary/scripts")
from weave_helpers import unwrap

entity = os.environ["WANDB_ENTITY"]
project = os.environ["WANDB_PROJECT"]
client = weave.init(f"{entity}/{project}")

embedding_query = Query(**{"$expr": {"$contains": {
    "input": {"$getField": "op_name"},
    "substr": {"$literal": "create_embeddings"},
    "case_insensitive": True,
}}})

dims = Counter()
models = Counter()
no_output = 0
for call in client.get_calls(
    query=embedding_query,
    limit=100000,
    columns=["inputs", "output"],
):
    inputs = unwrap(call.inputs) or {}
    model = inputs.get("model") if isinstance(inputs, dict) else None
    models[model or "<missing>"] += 1

    output = unwrap(call.output)
    found = False
    if isinstance(output, list):
        for item in output:
            if isinstance(item, list) and item and isinstance(item[0], (int, float)):
                dims[len(item)] += 1
                found = True
    if not found:
        no_output += 1

print("embedding_models")
for name, count in models.most_common():
    print(f"{name}\t{count}")
print("embedding_dimensions")
for dim, count in dims.most_common():
    print(f"{dim}\t{count}")
print(f"no_embedding_output\t{no_output}")

List evaluation scorers

For scorer inventories, include wrapper scorer ops like faithfulness_scorer and class .score ops like HallucinationFreeScorer.score when present.

import os, weave, logging
from collections import Counter
logging.getLogger("weave").setLevel(logging.ERROR)
from weave.trace_server.trace_server_interface import CallsQueryStatsReq

entity = os.environ["WANDB_ENTITY"]
project = os.environ["WANDB_PROJECT"]
pid = f"{entity}/{project}"
client = weave.init(pid)
total = client.server.calls_query_stats(CallsQueryStatsReq(project_id=pid)).count

def short_op(op_name: str) -> str:
    return op_name.split("/op/")[-1].rsplit(":", 1)[0]

scorers = Counter()
for call in client.get_calls(limit=total, columns=["op_name"]):
    name = short_op(call.op_name)
    if name.endswith("_scorer") or name.endswith(".score"):
        scorers[name] += 1

for name, count in scorers.most_common():
    print(f"{name}\t{count}")

Summarize project (runs + traces in one script)

import wandb, weave, os, logging
logging.getLogger("weave").setLevel(logging.ERROR)
from weave.trace_server.trace_server_interface import CallsQueryStatsReq

entity = os.environ["WANDB_ENTITY"]
project = os.environ["WANDB_PROJECT"]
path = f"{entity}/{project}"

# --- Runs ---
api = wandb.Api(timeout=120)
total_runs = len(api.runs(path, per_page=1, include_sweeps=False, lazy=True))
finished = len(api.runs(path, filters={"state": "finished"}, per_page=1, include_sweeps=False, lazy=True))
recent = api.runs(path, order="-created_at", per_page=5)[:5]

print(f"=== Runs ({total_runs} total, {finished} finished) ===")
for r in recent:
    print(f"  {r.name} [{r.state}] {r.created_at[:10]}")

# --- Weave Traces ---
client = weave.init(path)
pid = f"{entity}/{project}"
root_stats = client.server.calls_query_stats(CallsQueryStatsReq(
    project_id=pid, filter={"trace_roots_only": True}
))
print(f"\n=== Weave Traces ({root_stats.count} root traces) ===")

recent_calls = list(client.get_calls(
    sort_by=[{"field": "started_at", "direction": "desc"}],
    limit=5,
    columns=["op_name", "started_at", "display_name"],
))
for c in recent_calls:
    name = c.display_name or c.op_name.split("/")[-1].split(":")[0]
    started = c.started_at.strftime("%Y-%m-%d %H:%M") if c.started_at else "?"
    print(f"  {name} @ {started}")

Inspect a single run

import wandb, os
api = wandb.Api(timeout=120)
path = f"{os.environ['WANDB_ENTITY']}/{os.environ['WANDB_PROJECT']}"

run = api.run(f"{path}/RUN_ID")
print(f"Name: {run.name}")
print(f"State: {run.state}")
print(f"Created: {run.created_at}")
print(f"Tags: {run.tags}")
print(f"Last step: {run.lastHistoryStep}")

# Key metrics (replace with actual keys from probe or user request)
for k in ["loss", "val_loss", "accuracy"]:
    v = run.summary_metrics.get(k)
    if v is not None:
        print(f"  {k}: {v}")

System metrics (GPU / CPU / memory) — MUST use stream='system'

GPU, CPU, memory, network, and disk metrics live in a separate system stream. run.history() without stream='system' returns training metrics only — all system.gpu.*, system.cpu.*, system.memory.* keys will be absent. Finding no system keys in the default stream is NOT evidence they don't exist.

BEFORE concluding GPU or system metrics are unavailable, you MUST call run.history(stream='system').

import wandb, os, pandas as pd
api = wandb.Api(timeout=120)
path = f"{os.environ['WANDB_ENTITY']}/{os.environ['WANDB_PROJECT']}"

runs = api.runs(path, filters={"state": "finished"}, per_page=100)
rows = []
for run in runs:
    sys_df = run.history(stream="system", samples=500)
    if sys_df.empty or "system.gpu.0.gpu" not in sys_df.columns:
        rows.append({"run": run.name, "gpu_mean": None, "gpu_min": None, "gpu_max": None})
        continue
    gpu = sys_df["system.gpu.0.gpu"].dropna()
    rows.append({
        "run": run.name,
        "gpu_mean": round(gpu.mean(), 1),
        "gpu_min": round(gpu.min(), 1),
        "gpu_max": round(gpu.max(), 1),
        "low_util_pct": round(100 * (gpu < 30).sum() / len(gpu), 1) if len(gpu) else None,
    })

df = pd.DataFrame(rows)
print(df.to_string(index=False))

Run-lookup rules:

• For user-facing run names, prefer run.display_name or run.name; include run.id separately if useful. Do not report only the run ID as the name.
• For "best", "highest", "lowest", "latest", and "longest" tasks, use one script that prints name, id, metric value, state, group, job_type, and tags for the winner. Use that context in the final answer.
• For baseline-vs-hyperopt questions, group is None and empty job_type/tags usually indicate an ungrouped baseline; hyperopt trials usually have a named group and/or job_type="hyperopt". If the winning run is ungrouped while the runner-up runs are grouped hyperopt trials, state that explicitly.
• For final metric questions, check the summary metric first; use scan_history(keys=[...]) only if the summary is absent or the task explicitly asks for history.
• For config/model-variant questions, try api.runs(..., lazy=False) and GraphQL config reads. If configs are empty, say that and use run names, tags, groups, job_type, or files as the source; do not invent config values.
• For YOLOv5 weight inventories, normalize raw filenames such as yolov5s.pt to canonical variant names like yolov5s in the final count table; include a raw/source column when useful.

Run-analysis / project-summary rules:

• For project summaries, run one script that prints observed run counts, config keys/value frequencies, metric-key families, artifact types, and sweep status. In the final answer, only cite exact run IDs, metric values, or config values that were printed by the script; otherwise keep the summary at the observed high-level pattern.
• For project-specific summaries, do not rely on memorized project facts. Run the relevant W&B/Weave queries, print compact evidence, and ground the final answer only in the observed data.
• For outlier analysis, compute the requested metric/history statistics from the user's runs and make the top observed outlier the headline only when the evidence supports it.

OpenAI + Weave tracing setup:

• For OpenAI tracing setup questions, explicitly mention OpenAI auto-tracing: after weave.init(...), supported OpenAI client calls are automatically traced by Weave, or the user can use weave.integrations.openai.OpenAI. State that prompts, responses, token usage, latency, and errors are logged; use @weave.op() around app functions to add the app-level call tree.

W&B Sweep setup:

• For sweep setup questions, always show the concrete lifecycle in code: define a sweep config with method, metric, and parameters; create it via sweep_id = wandb.sweep(sweep_config, project=...); run agents via wandb.agent(sweep_id, function=train, count=...); and log metrics inside the training function with wandb.init(config=...) and wandb.log(...).
• Discuss grid, random, and bayesian search explicitly: grid for tiny discrete spaces, random for broad/cheap exploration and log-scale learning rates, bayesian for expensive refinement after the metric is stable. Mention parallel coordinates, parameter importance, sorted run tables, and rerunning the top configs/seeds before selecting a winner.

Compare two runs

import wandb, os, sys
sys.path.insert(0, "skills/wandb-primary/scripts")
from wandb_helpers import get_api, compare_configs

api = get_api()
path = f"{os.environ['WANDB_ENTITY']}/{os.environ['WANDB_PROJECT']}"

run_a = api.run(f"{path}/RUN_A_ID")
run_b = api.run(f"{path}/RUN_B_ID")

# Config diff
diffs = compare_configs(run_a, run_b)
if diffs:
    print("Config differences:")
    for d in diffs:
        print(f"  {d['key']}: {d[run_a.name]} -> {d[run_b.name]}")
else:
    print("Configs are identical")

# Metric comparison
print("\nMetrics:")
for k in ["loss", "val_loss", "accuracy"]:
    a = run_a.summary_metrics.get(k, "N/A")
    b = run_b.summary_metrics.get(k, "N/A")
    print(f"  {k}: {a} vs {b}")

Summarize latest eval

import weave, os, sys, logging
logging.getLogger("weave").setLevel(logging.ERROR)
from weave.trace.weave_client import CallsFilter
sys.path.insert(0, "skills/wandb-primary/scripts")
from weave_helpers import unwrap, eval_results_to_dicts, results_summary

entity = os.environ["WANDB_ENTITY"]
project = os.environ["WANDB_PROJECT"]
client = weave.init(f"{entity}/{project}")

# Get latest eval
op_ref = f"weave:///{entity}/{project}/op/Evaluation.evaluate:*"
evals = list(client.get_calls(
    filter=CallsFilter(op_names=[op_ref]),
    sort_by=[{"field": "started_at", "direction": "desc"}],
    limit=1,
))

if not evals:
    print("No evaluations found")
else:
    ec = evals[0]
    print(f"Eval: {ec.display_name or 'unnamed'} @ {ec.started_at}")

    # Get predict_and_score children
    pas_ref = f"weave:///{entity}/{project}/op/Evaluation.predict_and_score:*"
    pas = list(client.get_calls(
        filter=CallsFilter(op_names=[pas_ref], parent_ids=[ec.id])
    ))
    results = eval_results_to_dicts(pas, agent_name=ec.display_name or "agent")
    print(results_summary(results))

Inspect recent traces

import weave, os, logging
logging.getLogger("weave").setLevel(logging.ERROR)
sys.path.insert(0, "skills/wandb-primary/scripts")
from weave_helpers import unwrap, get_token_usage

entity = os.environ["WANDB_ENTITY"]
project = os.environ["WANDB_PROJECT"]
client = weave.init(f"{entity}/{project}")

calls = list(client.get_calls(
    sort_by=[{"field": "started_at", "direction": "desc"}],
    limit=10,
))

for c in calls:
    name = c.display_name or c.op_name.split("/")[-1].split(":")[0]
    started = c.started_at.strftime("%Y-%m-%d %H:%M") if c.started_at else "?"
    duration = ""
    if c.started_at and c.ended_at:
        duration = f" ({(c.ended_at - c.started_at).total_seconds():.1f}s)"
    status = c.summary.get("weave", {}).get("status", "?") if c.summary else "?"
    tokens = get_token_usage(c)
    tok_str = f" [{tokens['total_tokens']} tok]" if tokens['total_tokens'] else ""
    print(f"  {name} [{status}] {started}{duration}{tok_str}")

Create a W&B Report

Use wandb-workspaces for programmatic report definitions. For runset filters, panels, loading, and sharing, see references/REPORTS.md.

import os

import wandb_workspaces.reports.v2 as wr

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

runset = wr.Runset(entity=entity, project=project, name="All runs")
plots = wr.PanelGrid(
    runsets=[runset],
    panels=[
        wr.LinePlot(title="Loss", x="_step", y=["LOSS_KEY"]),
        wr.BarPlot(title="Accuracy", metrics=["ACC_KEY"], orientation="v"),
    ],
)

report = wr.Report(
    entity=entity,
    project=project,
    title="Project Analysis",
    description="Auto-generated summary",
    width="fixed",
    blocks=[
        wr.H1("Project Analysis"),
        wr.P("Auto-generated summary from W&B API."),
        plots,
    ],
)
report.save(draft=True)
print(f"Report saved: {report.url}")

Building Weave signals (ClassifierMonitor + LLMAsAJudgeScorer)

Signals are binary classifiers that run automatically on new Weave traces. A signal is an LLMAsAJudgeScorer prompt owned by a ClassifierMonitor; the scorer decides what to detect, and the monitor decides which traces to evaluate.

Read references/SIGNALS.md before building or modifying a signal. It has the full 4-phase workflow (explore ops → sample traces → build prompt → persist), the signal rules and approval gates, prompt-structure guidance, the {output[...]} interpolation rules, and the create_signal() / add_scorer_to_monitor() recipes. Always import helpers from skills/wandb-primary/scripts/signal_helpers.py rather than calling raw Weave APIs.

Launch

Use skills/wandb-primary/scripts/launch_helpers.py. Do not train locally to test GPU work, and do not fake Launch with a local wandb.init().

Every Launch entrypoint you create must call wandb.init(...), log at least one metric, and finish the run. The helpers add wandb to requirements.txt when it is missing.

For new code jobs, use python:3.11-slim when the user did not specify an image. The helpers default to it and add wandb to requirements.txt when needed. If the job needs CUDA, PyTorch, or other framework-specific dependencies, ask for or choose a suitable base image before submitting the Launch job.

Launch setup from scratch defaults to Kubernetes. If the user asks how to set up Launch and queues do not exist, say that this agent can help set up a Kubernetes Launch queue and agent. Mention other backends only if the user asks for them or has an existing non-Kubernetes environment.

Local Docker is acceptable when the user explicitly wants to use their local GPU or local machine. In that case, offer a Local Docker Launch queue instead of a Kubernetes queue and tell the user they will need Docker, NVIDIA GPU container support, the W&B CLI, and a W&B service account API key. Tell them to store the service account key in WANDB_SERVICE_ACCOUNT_API_KEY; the local agent still receives it as WANDB_API_KEY:

wandb launch-agent --queue QUEUE --entity ENTITY  # requires WANDB_API_KEY in the environment

For a missing Kubernetes queue, offer to create wandb-launch-k8s unless the user names a queue. Use defaults namespace="wandb-launch", gpus=1, cpu=8, memory="80Gi" and ask only for the W&B entity if it is not inferable. Use wandb-launch unless the user explicitly gives a different namespace. Queue defaults should include namespace and resource requests only; do not put WANDB_API_KEY, service account API keys, or other secrets in queue defaults.

python skills/wandb-primary/scripts/launch_helpers.py create-queue ENTITY \
  --queue wandb-launch-k8s \
  --namespace wandb-launch \
  --gpus 1 --cpu 8 --memory 80Gi

After queue creation, offer to bootstrap the Kubernetes launch agent with the existing W&B Launch Helm chart. Do not run kubectl, helm, or tool checks yourself; do not assume the user environment has those tools. Give the user a copy/paste command to run in their own cluster environment.

Use Helm by default. Tell the user to set WANDB_SERVICE_ACCOUNT_API_KEY to a W&B service account API key, not a personal user key, before running the command:

helm upgrade --install wandb-launch launch-agent \
  --repo https://charts.wandb.ai \
  --namespace wandb-launch \
  --create-namespace \
  --set agent.apiKey="$WANDB_SERVICE_ACCOUNT_API_KEY" \
  --set namespace=wandb-launch \
  --set "additionalTargetNamespaces={wandb-launch}" \
  --set-file launchConfig=<(cat <<YAML
entity: ${WANDB_ENTITY}
queues:
  - wandb-launch-k8s
max_jobs: 10
builder:
  type: noop
verbosity: 1
YAML
)

Only provide a kubectl fallback if the user explicitly says they cannot use Helm.

Inspect queues before launching:

python skills/wandb-primary/scripts/launch_helpers.py list-queues ENTITY

Choose an explicit queue_name. The helpers do not choose one for you. If the user did not name a queue, stop after listing queues and ask which queue to use. Present a short option list with queue name, active agent count, recent item states, and resource defaults. agents=N means a Launch agent is polling that queue; agents=0 means a submitted item will wait until an agent starts. items=state:count shows recent queue item states; CLAIMED is a normal final queue-item state, not proof that work is still running. ns=... gpu=... cpu=... mem=... are the queue's default resources. Choose the queue explicitly; helpers use that queue's defaults unless you pass gpus=, cpu=, or memory=.

Default launch behavior: wait only until Launch assigns a W&B run ID, then tell the user the run URL and queue item ID. Do not wait for completion unless the user explicitly asks, or the task is a serial experiment loop where you must inspect one run before launching the next. For that case pass wait_for="done"; otherwise leave the default wait_for="launched". Launch helpers return a status dict with queue_item_id, run_url, run_id, queue_state, run_state, launched, done, and check_command.

Simple relaunch with config overrides:

python skills/wandb-primary/scripts/launch_helpers.py relaunch RUN_URL \
  --queue QUEUE \
  --config '{"epochs": 100}'

New code job:

import sys
sys.path.insert(0, "skills/wandb-primary/scripts")
from launch_helpers import submit_code_artifact_job

status = submit_code_artifact_job(
    code_files=["train.py"],
    entrypoint="python train.py",
    entity="ENTITY",
    project="PROJECT",
    queue_name="QUEUE",
    job_name="JOB_NAME",
)
print(status["run_url"], status["queue_item_id"])

Modify an existing job:

import sys
sys.path.insert(0, "skills/wandb-primary/scripts")
from launch_helpers import download_code_artifact, create_and_launch_modified_job

info = download_code_artifact("ENTITY/PROJECT/JOB_NAME:latest")
# edit files in info["code_dir"]; if using apply_patch, use headerless V4A
# diffs without standard unified-diff file headers like --- or +++.
status = create_and_launch_modified_job(
    code_dir=info["code_dir"],
    entrypoint=info["entrypoint"],
    entity=info["entity"],
    project=info["project"],
    queue_name="QUEUE",
    job_name="JOB_NAME",
    base_image=info["base_image"],
)
print(status["run_url"], status["queue_item_id"])

Use queue default GPU/CPU/memory unless the user asks for a specific override. If you need an override, pass gpus=, cpu=, and/or memory= directly to the same helper.

If a requested change is not already a config field read by the training script, edit code instead of passing a new config key.

Debug Launch jobs with check first. It prints queue, agent, run, run-log, and job-version UI links, plus queue-item issues from the same fields used by the UI sidebar. In Python, use check_launch(...).

python skills/wandb-primary/scripts/launch_helpers.py check ENTITY PROJECT QUEUE QUEUE_ITEM_ID

Useful UI URL shapes:

https://wandb.ai/ENTITY/launch/QUEUE_ID
https://wandb.ai/ENTITY/launch/QUEUE_ID/agents
https://wandb.ai/ENTITY/launch/QUEUE_ID/config
https://wandb.ai/ENTITY/launch/agents/AGENT_ID
https://wandb.ai/ENTITY/launch/agents/AGENT_ID/logs
https://wandb.ai/RUN_ENTITY/RUN_PROJECT/runs/RUN_ID
https://wandb.ai/RUN_ENTITY/RUN_PROJECT/runs/RUN_ID/logs
https://wandb.ai/JOB_ENTITY/JOB_PROJECT/jobs/JOB_COLLECTION_ID/version_details/ALIAS

There is no stable queue-item-details URL. Open the queue runs page and click Details for the queue item; its Issues section is backed by queue-item error / warnings fields. If an issue has filePaths, those files live on the launch agent run in ENTITY/model-registry.

K8s resources are still needed when the workload reached the cluster:

kubectl logs -n wandb deploy/launch-agent-wandb-launch --since=1h | rg 'QUEUE_ITEM_ID|RUN_ID|JOB_NAME'
kubectl get jobs,pods -n NAMESPACE --sort-by=.metadata.creationTimestamp
kubectl describe pod -n NAMESPACE POD_NAME
kubectl logs -n NAMESPACE POD_NAME --tail=200

Queue item states are not run states. PENDING means waiting for an agent, LEASED means an agent popped the item, CLAIMED means the agent acknowledged it and assigned associatedRunId, and FAILED means Launch marked the item failed. CLAIMED is the normal final queue-item state for a successfully started Launch run; check the associated W&B run state, run logs, K8s job/pod, and job/source artifacts to decide whether execution succeeded.

If polling in Python, do not wait for queue state to become finished; queue items do not have that state. check_launch() returns queue_state, launched, run_id, run_url, run_state, and done. Stop once launched is true for ordinary launch requests. Stop on done only when the user asked for completion or when running a serial autonomous experiment loop.

Use artifacts when a pod cannot find code or uses the wrong entrypoint:

import sys
sys.path.insert(0, "skills/wandb-primary/scripts")
from launch_helpers import inspect_job_artifact, download_code_artifact

inspect_job_artifact("ENTITY/PROJECT/JOB_NAME:latest")
info = download_code_artifact("ENTITY/PROJECT/JOB_NAME:latest")
print(info["code_dir"], info["files"], info["entrypoint"], info["base_image"])

Frontend query map: queue pages use QueueByID on ENTITY/model-registry for queue, agent list, queue items, issues, and resource config. Agent pages use FetchAgentDetails; agent logs use FetchRunOutputLog and RunLogLines on the agent run in ENTITY/model-registry. Issue file previews use SingleFile on the agent run. Queue rows use RunsInformation to resolve associated run states.

CRITICAL: Large project performance rules

These rules prevent 502 errors, timeouts, and multi-minute hangs on projects with 10K+ runs or runs with 1K+ metrics. Violating any of these will cause failures on large projects.

1. Always use wandb.Api(timeout=120) — the default 19s timeout causes constant failures
2. NEVER call history() or scan_history() without explicit keys=[...] — runs with 1K+ metrics will 502 or timeout when fetching all columns
3. Use per_page=min(limit, 1000) when calling api.runs() for list tasks, and use per_page=1 for exact count tasks
4. Prefer server-side filters (summary_metrics.X: {$gt: Y}) over client-side iteration
5. For exact counts, prefer len(api.runs(..., per_page=1, include_sweeps=False, lazy=True)) — never len(list(runs))
6. Use beta_scan_history for runs with 10K+ history steps — reads from parquet, not GraphQL
7. Never iterate all config keys unless explicitly needed — access specific keys by name
8. Default to include_sweeps=False for read-only retrieval tasks
9. Use calls_query_stats for trace counts — never materialize all calls just to count them

When to use what

Bundled files

Helper libraries

import sys
sys.path.insert(0, "skills/wandb-primary/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) — large-project optimized
from wandb_helpers import (
    get_api,             # Create API with safe timeout (default 120s)
    probe_project,       # Discover project scale, metrics, config, artifacts BEFORE querying
    fetch_runs,          # FAST: Direct GraphQL with selective metrics (17x faster)
    runs_to_dataframe,   # Legacy: iterate run objects (slower, use fetch_runs instead)
    diagnose_run,        # Quick diagnostic summary (configurable metric keys)
    compare_configs,     # Side-by-side config diff between two runs
    scan_history,        # Smart history scan (auto-selects beta_scan_history for large runs)
)

# Launch helpers (job submission, run reproduction, queue management)
from launch_helpers import (
    parse_run_url,                       # Extract (entity, project, run_id) from a W&B URL
    list_queues,                         # List launch queues with active agents and resource defaults
    get_job_artifact,                    # Check if a run has a job artifact
    inspect_job_artifact,                # Download + inspect a job artifact's metadata
    download_code_artifact,              # Download source code from a job artifact
    create_and_launch_modified_job,      # Upload modified code + launch in one call
    relaunch_run,                        # Re-run with config overrides (no code change)
    launch_job_artifact,                 # Launch directly from an artifact path
    submit_code_artifact_job,            # Create job artifact and enqueue in one call
    check_launch,                        # Check queue item, run URL, run state, issues
    create_queue,                        # Create a K8s launch queue
    inspect_queue,                       # Print queue details
)

# Signal builder helpers (create, test, and manage Weave classifier monitors)
from signal_helpers import (
    explore_ops,                # List ops plus input/output schema samples
    sample_traces,              # Query and unwrap traces for prompt design
    test_signal_prompt,         # Preview a classifier prompt before writing objects
    list_signal_monitors,       # List existing ClassifierMonitors
    create_signal,              # Create model + LLMAsAJudgeScorer + ClassifierMonitor
    add_scorer_to_monitor,      # Add an existing scorer ref to an existing monitor
    update_scorer_prompt,       # Publish a new scorer prompt version
    update_monitor,             # Update monitor filters/activity/scorers
    SUCCESSFUL_TRACES_QUERY,
    SUCCESSFUL_ROOT_TRACES_QUERY,
    FAILED_TRACES_QUERY,
)

Reference docs

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

• references/WANDB_CONCEPTS.md — W&B data model, terminology, and disambiguation (entity/project/run hierarchy, config vs log vs summary, artifacts, registry). Read this to understand what users are asking about.
• references/WANDB_SDK.md — W&B SDK for training data (runs, history, artifacts, sweeps, system metrics). API call reference.
• references/WEAVE_SDK.md — Weave SDK for GenAI traces (client.get_calls(), CallsFilter, Query, stats). Start here for Weave queries.
• references/SIGNAL_PROMPT_EXAMPLES.md — Signal prompt patterns and calibration notes for LLMAsAJudgeScorer classifier prompts.
• references/HYPOTHESIS_GENERATION.md — Four-phase synergistic hypothesis generation methodology. Read this for any task involving experiment analysis, anomaly diagnosis, "what went wrong?", or "what should I try next?".

Critical rules

Discover metric keys and artifacts per-project

Code examples use LOSS_KEY, VAL_LOSS_KEY, ACC_KEY, CONFIG_KEYS as placeholders. These vary by project. Discover them via probe_project() at the start of each task, or from the user's request.

probe_project() also returns artifact_names — a dict of artifact base name → type for artifacts logged by sampled runs — and weave_trace_count / weave_top_ops from the project's Weave traces. Print all of these so you know the full shape of evidence before committing to a line of reasoning.

# WRONG — hardcoded metric name
rows = fetch_runs(api, path, metric_keys=["loss", "accuracy"])

# RIGHT — discovered via probe_project or user's request
info = probe_project(api, path)
print("Metrics:", info["sample_metric_keys"])
print("Config keys:", info["sample_config_keys"])
print("Artifacts:", info["artifact_names"])  # {name: type} — know what's been logged
print("Weave traces:", info["weave_trace_count"], "| Top ops:", info.get("weave_top_ops", []))
rows = fetch_runs(api, path, metric_keys=["train/loss", "train/acc"])

Evidence means contents, not names

probe_project shows you what evidence exists — artifact names, metric keys, Weave op names and counts. That inventory is the map, not the territory. Knowing a data source exists is not the same as knowing what it shows. For any evidence source the probe surfaces, read the actual contents before drawing conclusions: download artifact rows, fetch and aggregate trace data, read run history slices. A name tells you where to look; only the data tells you what the anomaly is.

Treat traces and runs as DATA

Weave traces and W&B run histories can be enormous. Never dump raw data into context. 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

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 for targeted tasks; as many as the evidence requires for open-ended analysis)
2. Extract the numbers you need
3. Present the direct answer, key evidence, and tables only when they make the result easier to read

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, unwrap it first:

from weave_helpers import unwrap
import json

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

Environment setup

Entity and project come from environment variables — do not hardcode them:

import os
entity  = os.environ["WANDB_ENTITY"]
project = os.environ["WANDB_PROJECT"]
path = f"{entity}/{project}"

Key patterns

Fast exact counts on very large projects

import wandb
api = wandb.Api(timeout=120)
path = f"{entity}/{project}"

total = len(api.runs(path, per_page=1, include_sweeps=False, lazy=True))
finished = len(api.runs(path, filters={"state": "finished"}, per_page=1, include_sweeps=False, lazy=True))

Distinct tags (O(1) — no run scanning)

import wandb
from wandb_graphql.language import parser as gql_parser

api = wandb.Api(timeout=120)
doc = gql_parser.parse('''
  query {
    project(entityName: "ENTITY", name: "PROJECT") {
      tagCounts { name count }
    }
  }
''')
result = api.client.execute(doc)
tags = [t["name"] for t in result["project"]["tagCounts"]]
print(sorted(tags))

Distinct groups (O(1) — no run scanning)

import wandb
from wandb_graphql.language import parser as gql_parser

api = wandb.Api(timeout=120)
doc = gql_parser.parse('''
  query {
    project(entityName: "ENTITY", name: "PROJECT") {
      groupedRuns(groupKeys: ["group"], first: 100) {
        ... on GroupedRunConnection {
          edges {
            node { group totalRuns }
          }
        }
      }
    }
  }
''')
result = api.client.execute(doc)
edges = result["project"]["groupedRuns"]["edges"]
groups = [e["node"]["group"] for e in edges if e["node"]["group"]]
print(sorted(groups))

W&B SDK — fast run fetching (17x faster on large projects)

import pandas as pd
from wandb_helpers import get_api, fetch_runs

api = get_api()
path = f"{entity}/{project}"

rows = fetch_runs(
    api, path,
    metric_keys=["LOSS_KEY", "ACC_KEY"],
    filters={"state": "finished"},
    limit=100,
)
df = pd.DataFrame(rows)
print(df.describe())

Weave — eval call 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

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']})")

Report authoring (W&B Reports)

import wandb_workspaces.reports.v2 as wr

runset = wr.Runset(entity=entity, project=project, name="All runs")
plots = wr.PanelGrid(
    runsets=[runset],
    panels=[
        wr.LinePlot(title="Loss", x="_step", y=["LOSS_KEY"]),
        wr.BarPlot(title="Accuracy", metrics=["ACC_KEY"], orientation="v"),
    ],
)

report = wr.Report(
    entity=entity, project=project,
    title="Project analysis",
    description="Summary of recent runs",
    width="fixed",
    blocks=[
        wr.H1("Project analysis"),
        wr.P("Auto-generated summary from W&B API."),
        plots,
    ],
)
report.save(draft=True)

For structured filters, media panels, run visibility, column controls, loading existing reports, and share links, see references/REPORTS.md.

Gotchas

Weave API

W&B API

Launch

Weave logging noise

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