Releases·Tool Release
Show HN: Onesie – An expressive CLI for System One models like Jev
Frodi Karlsson released onesie, an MIT-licensed CLI for using System One models (such as TypeSafe's Jev) in shell pipelines, scripts, and local agentic workflows. It adds caching, batching, deduplication, streaming, multi-question requests, and calibration of classifier thresholds against labelled data. It targets TypeSafe's Jev, Jev via OpenRouter, Berget's System One models, and any Jev-compatible API.
3 engagement1 sourceMon, Sep 28, 2026, 11:37 AM
AI Overview
- Onesie is a MIT-licensed Unix-style CLI for piping shell data to TypeSafe Jev and other System One models, with caching, batching, dedup, and streaming.
- Its onesie calibrate command tunes thresholds against labelled JSONL data, scoring e.g. AUC 1.00 on a 40-example shell-safety set at 0.25/0.5 cuts.
- Single requests can ask multiple questions --ask (boolean, --pick multi-choice, --rate scale), returning values like {"urgent":0.91,"team":"billing"}.
- Script gates exit 1 when blocked, 7 when uncertain, letting scripts eval safe commands or escalate to the user.
- Author skipped an MCP server and cost estimates, pointing users to shaharia-lab/jev-cli for those; install via brew or go install.
Related
Releases·Tool Release
Jeview: unofficial local visualizer for Jev (TypeSafe) proxies calls via 127.0.0.1:4777, logs to SQLite, maps live
An independent developer published Jeview on GitHub, a local proxy-visualizer for TypeSafe's Jev (System One) that logs every call to SQLite and draws them on a live map. It runs as a gateway between user code and api.typesafe.ai, needs only Node 24+, and has zero dependencies. It is not affiliated with or endorsed by TypeSafe AI.
An independent developer published Jeview on GitHub, a local proxy-visualizer for TypeSafe's Jev (System One) that logs every call to SQLite and draws them on a live map. It runs as a gateway between user code and api.typesafe.ai, needs only Node 24+, and has zero dependencies. It is not affiliated with or endorsed by TypeSafe AI.
57 engagement1 sourcegithubMon, Sep 21, 2026, 02:03 PM
Social·Community
I ran a $0.63 classifier over 150 intents + 1,000 out-of-scope utterances. The confidence number held up; the fallback bucket didn't.
An independent developer ran TypeSafe's Jev structured-output model over the full CLINC150 test split (4,500 in-scope plus 1,000 out-of-scope utterances) with a 151-way choice question. The returned confidence was well calibrated (ECE 0.025) and out-of-scope detection was solid, but 54 off-topic utterances were still assigned specific intents at confidence 0.9 or higher, so neither a fallback bucket nor a confidence gate fully catches misrouted input. The results include an escalation table showing the accuracy/human-review tradeoffs for confidence thresholds.
An independent developer ran TypeSafe's Jev structured-output model over the full CLINC150 test split (4,500 in-scope plus 1,000 out-of-scope utterances) with a 151-way choice question. The returned confidence was well calibrated (ECE 0.025) and out-of-scope detection was solid, but 54 off-topic utterances were still assigned specific intents at confidence 0.9 or higher, so neither a fallback bucket nor a confidence gate fully catches misrouted input. The results include an escalation table showing the accuracy/human-review tradeoffs for confidence thresholds.
4 engagement1 sourceredditSun, Sep 20, 2026, 05:28 PM
Social·Community
System 1 Models: Faster and Cheaper Proxy to Frontier Models — With a Working Router
Two Medium posts examine TypeSafe's System One family of decision-focused models — Jev, Laya, Kev, and Nimble — covering their architecture, typed decisions, parallel inference, and calibration. A related post demonstrates routing requests through a LiteLLM proxy to pick the appropriate model. The posts position these open-weights models as faster, cheaper proxies to frontier models for decision workloads rather than conversation.
Two Medium posts examine TypeSafe's System One family of decision-focused models — Jev, Laya, Kev, and Nimble — covering their architecture, typed decisions, parallel inference, and calibration. A related post demonstrates routing requests through a LiteLLM proxy to pick the appropriate model. The posts position these open-weights models as faster, cheaper proxies to frontier models for decision workloads rather than conversation.
0300 engagement3 sourcesrssThu, Sep 24, 2026, 12:00 AM
Releases·Tool Release
IronBee built a browser agent on TypeSafe's Jev classifier instead of an LLM — costs, failure modes, replay results
The IronBee team wrote up IronBee Express, a browser agent where each step is a single classifier call to TypeSafe's Jev, with an LLM only writing free text, unsticking runs, or explaining failures. The post offers concrete lessons for agent builders: put guidance in the offered options rather than the prompt, batch one request per step (~300 ms per step), and expect most bad choices to stem from missing context. Replays run with zero model calls at $0.00054 per replayed checkout versus $0.0042 for a 20-action explored run.
The IronBee team wrote up IronBee Express, a browser agent where each step is a single classifier call to TypeSafe's Jev, with an LLM only writing free text, unsticking runs, or explaining failures. The post offers concrete lessons for agent builders: put guidance in the offered options rather than the prompt, batch one request per step (~300 ms per step), and expect most bad choices to stem from missing context. Replays run with zero model calls at $0.00054 per replayed checkout versus $0.0042 for a 20-action explored run.
1 engagement1 sourceredditMon, Oct 5, 2026, 02:45 PM
Releases·Tool Release
Auto mode plugin for the Pi coding agent that uses Kev/Laya (running locally) or Jev to classify commands
A developer published pi-automode-classifier, an open-source plugin for the Pi coding agent that checks every shell command before execution in auto mode. It applies built-in rules first, then sends unknown commands to a small classifier model — hosted Jev or locally run Kev-0.8B/Laya — which returns a risk probability; above a threshold the user gets a confirm prompt. The model never blocks a command by itself.
A developer published pi-automode-classifier, an open-source plugin for the Pi coding agent that checks every shell command before execution in auto mode. It applies built-in rules first, then sends unknown commands to a small classifier model — hosted Jev or locally run Kev-0.8B/Laya — which returns a risk probability; above a threshold the user gets a confirm prompt. The model never blocks a command by itself.
2 engagement2 sourcesredditThu, Oct 8, 2026, 12:44 PM