genbench
Skill618 starsChanged 3 months agoApache-2.0View on GitHub
Run Vendo's buy-vs-build generation benchmark — generation time + accuracy of the real Vendo pipeline vs raw-Claude baselines (diy, claude-code) on JSON-defined worlds. Use when asked to benchmark generation, compare Vendo vs raw models, measure generation speed/cost, or add genbench worlds/cases.
- Reads credentials
The file
Files
genbench
One command, from the repo root (ANTHROPIC_API_KEY in the environment):
pnpm build # genbench reads the built @vendoai/* dists
pnpm genbench run --prompt <case-id> # one case, all contenders, opens preview.html
pnpm genbench run # every case in one world
pnpm genbench run --models opus,sonnet,haiku # expand the harness x model matrix
pnpm genbench run --world maple # choose world (default: maple)
pnpm genbench run --world all # every world into ONE run folderThe corpus is 14 worlds, 196 cases — 15 per world, except buildlog and
fieldops at 10 and logistics, observability, product-analytics and
trades-accounting at 14. Case ids in maple: spend-overview, spend-chart,
pending-transfers, account-balances, no-pending-transfers,
transfer-receipt, cancel-both-pending, transfer-activity-feed,
room-for-dana, rent-check, bills-calendar, money-dashboard,
dining-budget-cap, category-groups, savings-goals.
- Contenders:
vendo(real pipeline, this working tree) ·diy(one rawstreamTextcall) ·claude-code(stock Agent SDK in a scratch dir). Byte-identical world info per contender: ONE serializer,worldBlockingenbench/src/vendo.ts, enforced by the fairness test ingenbench/tests/diy.test.tsagainst what the vendo driver really receives. - Worlds:
genbench/worlds/<name>/{world.json, cases.json}plus an optionalfont.woff2the harness injects into every contender's page.world.json= tools + canned data + theme + style rubric;cases.json= prompt +passlines. Conventions: money in cents; a tool withdatais a read,takes-only is a write. - Output:
genbench/runs/<run>/<contender>/<case>/{artifact.tsx (vendo only), page.html, screenshot.png, result.json}+summary.json(the run's ONE aggregate, per column) +preview.html(live embedded screens, world-data panel, live tool-call feed).runs/is gitignored.<contender>is the column slug<harness>-<model>—vendo-sonnet,diy-opus,claude-code-haiku. Under--world allthe case folder is<world>/<case>, because two worlds ship the same case id. - Floor checks are deterministic (delivered / renders / valid / honestData /
wiredActions via click-probe). A check with nothing in front of it is VACUOUS
and one whose grader was unreachable is DEGRADED; neither scores, in either
direction. The
passlines are the rubric a pinned judge (versioned rubric contract) grades on every run — any edit to its prompt bumpsrubricVersionand resets comparability. - Exit code: any floor failure exits 1; a judge outage or a failed rubric
line does not. The last stdout line says which —
floor failures: 2 (exit 1). Throughpnpmthat is followed by pnpm's ownELIFECYCLEline, which is not a second failure. - Judge spend is reported separately —
judged.costinresult.jsonand one line under the preview's run header. It is NEVER folded into a contender'scost, which is only what that contender spent building its screen. - Budgets are per contender: five minutes for
vendoanddiy, twelve forclaude-code, which runs its own ten-minute wall clock inside the driver. - Rough cost: one case ≈ 1-4 min ≈ $0.30-$0.50 + judge; one world is 10-15x
that and
--world allis 196x;--modelsmultiplies by the model count. Prices insrc/meter.tsare as of 2026-08-08 (Sonnet 5 is on intro pricing through 2026-08-31) — token counts are the durable number, dollars are not. --promptruns openpreview.htmlon macOS; full runs just print the path.CIorGENBENCH_NO_OPEN=1suppresses the window.- Gotchas: test genbench with
pnpm --filter @vendoai/genbench test(itsvitest.config.tscaps the pool at 1-2 workers) — never the full repo suite, and not a barenpx vitest, which resolves a different vitest that rejects this repo's worker flags. The two money-spending tests (judge smoke, claude-code driver) needGENBENCH_LIVE=1andANTHROPIC_API_KEY; both stay skipped otherwise.--lane buildis deferred and says so.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

