frontpage.
newsnewestaskshowjobs

Open Source @Github

fp.

Open in hackernews

Show HN: Caliper – pass@k reliability testing for Claude Code and Codex skills

https://github.com/edonadei/caliper
2•edonadei•2h ago
Skills for Claude Code and Codex are hard to test. What I mean by hard is that there's no standard way to do it. You evaluate the skill once on something, it looks like it works. You publish it. Then the new super model releases (GLM 5.2 anyone?), it will quietly break for some part, and you won't find out until your users complain.

I also faced the same problem, so I tried to build something lightweight to stop doing that. Caliper.

It's a local and lightweight harness that runs a skill k times in isolated environments and gives you a pass@k score (How much times it succeeded in these k times). As a non-deterministic technology, you can't just say "it worked once". You need to answer how much it passed in k times.

You define success in a YAML spec. I picked YAML to keep a schema and make it still readable for a human. You either use a LLM judge, a Python assertion, or both:

Here's an simple evaluation example with a JSON extraction, so you write this in a YAML file:

  tasks:
    - name: Extracts action items as clean JSON
      prompt: "Read /tmp/transcript.txt and write the
               action items to /tmp/actions.json."
      expect: "A valid JSON array where every item has
               owner, task, due. No markdown fences."
      assert: |
        import json
        items = json.load(open("/tmp/actions.json"))
        assert isinstance(items, list)
        assert all({"owner","task","due"} <= i.keys()
                   for i in items)
Then with the CLI, you'll run it:

caliper run extract-actions.eval.yaml --k 5 --baseline

What's cool about the --baseline flag is that it will re-runs everything without the skill, so you can see whether the skill is doing the work or the base agent was going to pass anyway:

  ID      Task                           k(5)  pass@k
  task-1  Extracts action items as JSON  5/5   100%  PASS
  With skill   100%
  No skill      60%
  Delta        +40%
Most models know how to get the JSON right most of the time (JSON extraction was solved by 2 years old already). But that's it, "most of the time" is the bug. That delta shows how the skill actually helped. (It's sometimes 0%, sometimes -100%!)

I also created two skills you can get started right away with your favorite harness, e.g. Claude Code, Codex or Pi:

- evaluate-skill: run and manage evals without leaving your workflow

- grill-skill: reads your SKILL.md, interviews you about what "good" looks like, writes a 3-task spec (happy path, edge case, adversarial), and runs it

You can install the skill with the command: npx skills@latest add edonadei/caliper

I for now support claude-code, codex, pi, claude-api, openai-api. You can run the agent and the judge as separate backends, so you can run a skill on one and judge with another.

GitHub: https://github.com/edonadei/caliper PyPI: https://pypi.org/project/caliper-eval/

Of course, it's a first step. I think the autorater layer can be vastly improved, more handholding to create and iterate on evaluation specs, supporting more harness, why not including this layer into a self-improvement bigger system?

If you're also building agentic evaluations, I'm genuinely interested to hear how you are handling that.

Comments

edonadei•1h ago
btw FYI, a really good article on evaluation, I vastly based my research and iteration from it https://www.anthropic.com/engineering/demystifying-evals-for...

Show HN: Zanagrams

https://zanagrams.com/
94•pompomsheep•4h ago•30 comments

Show HN: DRM-Free Books

https://frequal.com/Perspectives/DrmFreeAuthors.html
27•TeaVMFan•3h ago•9 comments

Show HN: Bash4LLM+ – A lightweight, dependency-free Bash wrapper for LLM APIs

https://github.com/kamaludu/bash4llm/
3•kamaludu•18m ago•3 comments

Show HN: NanoEuler – GPT-2 scale model in pure C/CUDA from scratch

https://github.com/JustVugg/nanoeuler
3•vforno•23m ago•1 comments

Show HN: Decomp Academy – Learn to decompile GameCube games into matching C

https://decomp-academy.dev
181•jackpriceburns•18h ago•69 comments

Show HN: Caliper – pass@k reliability testing for Claude Code and Codex skills

https://github.com/edonadei/caliper
2•edonadei•2h ago•1 comments

Show HN: Import the HN Home to a reading queue with clean reader view and TL;DR

https://readplace.com/import?mode=from-url
3•fagnerbrack•4h ago•1 comments

Show HN: Metaspec: The DpANS3R Common Lisp Spec in S-Expr and HTML Format

https://metaspec.dev/#
15•dlowe-net•4d ago•0 comments

Show HN: Adrafinil – keep a lid-closed Mac awake only while agents work

https://github.com/kageroumado/adrafinil
118•kageroumado•23h ago•75 comments

Show HN: FSM – an advanced system monitor for Linux

https://github.com/mskrasnov/FSM
21•mskrasnov•18h ago•7 comments

Show HN: Hacker News on a train station-style flip board

https://popflame.quickish.space/hn-flipboard/
110•PaybackTony•1d ago•21 comments

Show HN: DBOSify – Drop-in Temporal replacement built on Postgres

https://github.com/dbos-inc/dbosify-py
87•KraftyOne•4d ago•19 comments

Show HN: Starglyphs - A constellation puzzle game based on Euler paths

https://starglyphs.com
26•telman17•22h ago•7 comments

Show HN: A faithful MUMPS 76 anniversary implementation – the original NoSQL DB

https://github.com/rochus-keller/mumps/
3•Rochus•8h ago•1 comments

Show HN: Kiso, an open-source publishing engine for Open Knowledge Format

https://oak-invest.github.io/kiso/
18•straumat•22h ago•0 comments

Show HN: Hacker Times – HN Reader

https://times.hntrends.net/
7•bencevans•9h ago•8 comments

Show HN: role-model, a router for hybrid local/cloud AI

https://github.com/try-works/role-model
2•try-working•9h ago•1 comments

Show HN: Engye – transfer files between any two devices by scanning a QR code

https://engye.fuzzyworld.net/
12•psafronov•19h ago•2 comments

Show HN: WebBase-III – dBASE III rebuilt in the browser with its own interpreter

https://github.com/DDecoene/WebBaseIII
95•ddecoene•4d ago•27 comments

Show HN: Smart model routing directly in Claude, Codex and Cursor

https://github.com/workweave/router
209•adchurch•2d ago•111 comments

Show HN: Autofit2 – End-to-end pipeline for multilingual text classification

https://github.com/neospe/autofit2
28•leschak•3d ago•2 comments

Show HN: I made Google Trends for Hacker News by indexing 18 years of comments

https://hackernewstrends.com
804•ytkimirti•3d ago•155 comments

Show HN: Clanker TV

https://botflix.tv/
4•jshaivitz•5h ago•1 comments

Show HN: Turn native language audio into flashcards and shadowing practice

https://lingochunk.com/try
91•alder•3d ago•37 comments

Show HN: KV-psi, using Linux PSI to to trim an LLM KV cache

https://github.com/infiniteregrets/kv-psi
8•infiniteregrets•21h ago•0 comments

Show HN: Overfitted a 900KB Transformer to Compress a 100MB CSV into 7MB

111•spidy__•5d ago•67 comments

Show HN: OpenKnowledge – open source AI-first alternative to Obsidian/Notion

https://github.com/inkeep/open-knowledge
374•engomez•3d ago•170 comments

Show HN: Foveon – Bayer to Foveon X3, learned, Mac App using deep learning

https://code.intellios.ai/photo/
4•coolwulf•15h ago•2 comments

Show HN: Bible as RAG Database

https://www.crosscanon.com/
160•jacksonastone•3d ago•94 comments

Show HN: Shopify UCP is insanely powerful

https://stack412.com/
3•westche2222•18h ago•3 comments