kgbench

A command-line operator for the kgbench coding-v1 benchmark, which tests how well retrieval systems and agents answer questions about a code repository. It can run, resume, grade, regrade, verify, and report benchmark results.

In plain words
What is it for?
Use it to run benchmark matrices, compare retrieval arms, reapply grading, validate evidence references, repair older run data, and generate results reports and charts.
Why use it?
It gathers retrieval quality, token use, and latency results while helping diagnose which model actually answers each question.

Command for Claude Code

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add commands/fwornle/coding/kgbench
Clone the repo
git clone --depth 1 https://github.com/fwornle/coding

Made for: Claude Code.

Per session 60 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 4,573 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 1 finding. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00060 $0.04573
Opus 5 $0.00030 $0.02286
Sonnet 5 $0.00012 $0.00915
Haiku 4.5 $0.00006 $0.00457

Measured yesterday against content hash 87a00b3df7cd, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

kgbench scanned grade A with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Makes network callslowCapability

Not a fault in itself. Listed so you know the mod talks to something, and to what.

curl -s localhost:3032/api/kgbench/active-run # {"runId":null} when the coast is clear
.claude/commands/kgbench.md · 323 lines

How it starts

The opening of the file, as written. The whole thing — 323 lines — stays where its author put it; the contents beside it link to each section on GitHub.

kgbench (/kgbench) — run, resume, grade and diagnose the retrieval benchmark

Operator front-end for the coding-v1 code-retrieval benchmark: which retrieval strategy answers questions about this repository best, at what token and latency cost.

This skill is a THIN WRAPPER. It shells to existing CLIs and reimplements no runner, grader, judge or report logic:

Script Role
scripts/kgbench-supervise.sh detached, self-resuming matrix supervisor
scripts/kgbench-run.mjs the matrix itself (--set --arms --only --reps --run-id)
scripts/kgbench-regrade.mjs re-apply graders offline; --rejudge re-runs the judge
scripts/kgbench-report.mjs render RESULTS.md + report.json (refuses to clobber hand-written files)
scripts/kgbench-charts.mjs render the SVG figures
scripts/kgbench-verify-questions.mjs validate questions and every evidence file:line
scripts/kgbench-backfill-tokens.mjs re-resolve token fields offline; refuses a window narrower than the cell
scripts/kgbench-repair-attempt-windows.mjs transitional: reconstruct per-attempt windows on runs recorded before the runner wrote them

Read docs/benchmarks/measurement-lessons.md before changing a question, a matcher or an answer key. It records defects that cost a full investigation each, and one blind spot that makes the obvious diagnosis wrong.

Two front-ends, one benchmark

There is also a Performance → Benchmarks sub-tab at localhost:3032/performance. It is a second front-end over the same scripts, not a second implementation: its launcher shells to kgbench-supervise.sh on the host, and its results view renders what lib/kgbench/report.mjs produces — the function kgbench-report.mjs calls. A number shown there and a number in the published report can differ only because the data changed.

CLI (/kgbench) Dashboard
run / resume ✅ (Launch, and a resume offer when the id has cells)
watch a run status ✅ live monitor + supervisor log tail
cancel a run Ctrl-C, or kill the group Cancel
results reportRESULTS.md ✅ live aggregate, and the published artefacts
regrade / rejudge ❌ — CLI only, deliberately (Step 4)
publish ❌ — CLI only (Step 3)
doctor partial: a model probe, not the judge probe (Step 5)

Read the full file on GitHub · 323 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. yesterday First seen · 323 lines · 60 tokens per session scan A 87a00b3df7cd

Subscribe to this mod's changes

kgbench is a command published in the GitHub repository fwornle/coding (2 stars, last pushed yesterday), licensed MIT. It adds 60 tokens to every session and 4,573 once invoked, about $0.0003 per session on Opus 5. A static security scan graded it A with 1 finding (makes network calls). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.