CapyWorks
The chillest software factory: an LLM writes the code, Capy (TypeSafe's Jev decision model) judges every step for AI slop in ~150 ms.
About the project
CapyWorks is a software factory with a System 1. Coding agents are fast, but they ship AI slop: TODOs, leftover console.logs, any everywhere, tests that assert nothing. Tests don’t catch it: in our demo, the sloppy patch passes every test.
CapyWorks splits the work the way Thinking, Fast and Slow splits a brain:
- System 2 (slow): an LLM coding agent (Qwen3.8-27B on W&B Inference) writes the code.
- System 1 (fast): Capy the capybara, powered by TypeSafe’s Jev decision model, makes every judgment call by asking typed yes/no, choice and score questions and getting probabilities back in about 150 ms.
The line: Intake (Jev: is this ticket specific enough to build?) → Triage (Jev: kind of work + risk) → Build (LLM) → Tests (hard gate) → Capy Gate (one Jev call asks five anti-slop checks in parallel; failed checks become feedback for the next attempt) → Onsen (merged, and Capy gets a yuzu).
A real recorded shift: 4 tickets, 3 merged, 1 bounced for being too vague. The sloppy CAPY-2 patch passed all 5 tests, but Capy failed it on all 5 checks in 114 ms; the retry merged clean. 30 decisions in 11 Jev calls, 136 ms average, $0.00029 in total, against 45.6 s of LLM build time.
Built clean: strict TypeScript, zero runtime dependencies, 21 tests. Every check lives as data with an explicit threshold, agents can only write src/ and test/ files, every attempt gets a fresh workspace, and every decision is recorded in a Markdown ledger. A live factory-floor dashboard streams the shift over Server-Sent Events, and an 80-second demo video was made with HyperFrames and narrated with ElevenLabs.
Built with
AI tools
Prize Categories Entered
- Submitted
Best Demo Video Using HyperFrames
Our 80-second demo video is an HTML composition rendered to MP4 with HyperFrames: seven GSAP-timed scenes, captions timed word by word from Capy’s ElevenLabs narration, and every number pulled from the recorded shift’s run log by a build script. A new shift re-renders with one command (bash video/scripts/render.sh), with no new audio needed.
- Event prizeSubmitted
Best Overall Hack
CapyWorks is a capybara-run software factory: an LLM coding agent writes code, and Capy (TypeSafe’s Jev decision model) makes every judgment call, from intake and triage to an anti-slop gate after the tests. In a real recorded shift it failed a sloppy patch that had passed all its tests on 5 checks in 114 ms, then merged the clean retry. 30 decisions cost $0.00029, and a live dashboard and an 80-second video show it all.
- Event prizeSubmitted
Best Use of System One Models
Jev is the factory’s control plane, not a bolt-on. Every judgment is a typed System One question: yes/no for intake and the five anti-slop gate checks, choice for the kind of work, and score for risk, with explicit thresholds kept as data in src/checks.ts. One Jev call answers the whole gate checklist in parallel (136 ms average), so a real shift made 30 decisions for $0.00029: cheap and fast enough to judge every attempt, not a sample.
- Event prizeSubmitted
Cleanest Use of a Software Factory
CapyWorks fights AI slop at every station: vague tickets bounce at intake, tests are the hard gate, and Capy’s gate fails diffs with TODOs, debug logs, any, scope creep or tests that assert nothing, then feeds the failed checks back to the agent. Agents may only write src/ and test/ files, each attempt gets a fresh workspace, and every decision is recorded in a ledger (runs/demo.md). The repo itself: strict TypeScript, zero runtime dependencies, 21 tests.