Caliper Harness: open-source coding harness disciplines small local models with plans, checks, reviews
Saraozte01 · reddit · 2026-11-19
A developer released Caliper Harness, an open-source terminal coding harness built for small local models, betting that tiny models rush and claim false completion unless a harness enforces discipline.
- PAT workflow: models must produce a structured Plan → Approach → Task plan stored in a GraphQL state graph, refined level by level and staying coherent; users approve before any code is written, and the model is pushed to ask questions on real decisions instead of guessing.
- Step-checked execution: only one task is claimed at a time; each task has its own check command, and completion requires both the check and an independent review to pass, judged against what was written during planning.
- Adversarial review: a fresh-context reviewer pass audits finished work — in testing, with every test green, it still caught a real bug (a station situation rendered in a broken color and invisible on the map) that tests missed; the task was sent back and only passed after a proper fix plus regression tests.
- Positioning: unlike efficiency-focused harnesses like OpenCode built around large models, Caliper is deliberately slow and deliberate.
It is local-only for now, pointing at a local model server. The author tests it by having a model build a browser game, Space Transporter, from scratch, then stress-testing with larger follow-up features.
More from coding & agent
- Gradio and HF launch hackathon: reproduce NeurIPS 2026 papers with AI agents — Gradio · 2026-11-19
- Ramp SWE-Bench v2: top four coding models within points, but spend varies wildly — charliermarsh · 2026-11-19
- LangChain's deepagents hits 50M downloads, now the industry-standard agent harness — LangChain · 2026-11-19
- NVIDIA open-sources SkillEvaluator to vet AI agent skills before deployment — anthara_ai · 2026-11-19
- Resolve.ai publishes SRE-1 paper on AI agents for production incident investigations — iceman84 · 2026-11-19
- GPT-6.2 in Codex: split reviews hinge on your first line, says thread with 5 prompts — alexcovo_eth · 2026-11-19