🔮 hunch

Sort, filter and route text by what it means.

Pipe in logs, tickets, reviews or diffs, ask a question in plain English, and hunch answers it for every line: yes or no, which category, or how much. Each answer comes with a confidence score your next command can act on.

kubectl logs api | hunch "Is this an authentication failure?" --min 0.8 --raw

✨ Why hunch

• Powered by Jev. Each record is judged by Jev, TypeSafe's System One model, through TypeSafe, OpenRouter or OpenCode Zen.
• Calibrated answers. Every answer is a probability you can trust, so thresholds, sorting and routing are built in.
• Fast and cheap enough to run on everything. About 100 ms a record, and output tokens are free.
• A good Unix citizen. Lines, JSONL or whole files in; JSONL, TSV or the original lines out; exit codes that work in if and git hooks.

Say hunch labels every pull request in CI. A typical PR (title, description and a few hundred lines of diff) is about 2,000 tokens. At $0.042 per million input tokens, that's less than a hundredth of a cent per PR: one dollar labels about 12,000 PRs. Further questions about the same PR, such as how risky it is or whether it adds secrets, go into that same call and add little latency.

_{TypeSafe pricing as of October 19, 2026; OpenRouter and OpenCode Zen may differ.}

🚀 Install

brew tap yarolegovich/hunch https://github.com/yarolegovich/hunch
brew trust yarolegovich/hunch   # Homebrew asks you to trust third-party taps
brew install hunch
# or
go install github.com/yarolegovich/hunch@latest

Binaries are also on Releases. Then pick a provider and store its key:

hunch auth login

🧰 Commands

Command	What it asks
hunch yes <question>	a yes/no question (also plain hunch <question>)
hunch pick <question> -c a,b,c	which of several choices fits
hunch score <question> -l low,mid,high	where it falls on ordered levels
hunch ask -f questions.yaml	several questions at once, in one call
hunch auth	set up the provider, key and model
hunch models	list the provider's models

Input modes, output formats, filtering, exit codes and configuration are in docs/usage.md.

🍳 Cookbook

Real workflows, start to finish. Each one is a single page you can copy from:

1. CI auto-labeler: label every pull request by kind and risk in a GitHub Actions workflow.
2. Log explorer: get the shape of a huge log, zoom in on what matters and watch a live tail.
3. Issue triage bot: label new GitHub issues, ask for missing repro steps and close spam.
4. Pre-commit guard: stop secrets and leftover debug code before they're committed.
5. Flaky CI retry: when CI fails, re-run it only if the environment broke, not the code.

💡 One-liners

# judge one field of each JSON record, keep the rest for later steps
hunch "Is the customer asking for a refund?" --jsonpath '$.body' --min 0.8 < tickets.jsonl | jq -r .id

# route support tickets into a file per team, low-confidence ones aside
hunch pick "Which team should handle this?" -c billing,technical,sales \
  --jsonpath '$.body' --conf 0.6 -o 'out/{choice}.jsonl' < tickets.jsonl

# top 10 angriest reviews
hunch score "How frustrated is the customer?" -l calm,frustrated,furious --top 10 reviews.txt

# label a pull request in CI
gh pr diff "$PR" | hunch pick --in whole "What kind of change is this?" -c feature,fix,refactor,docs --tsv \
  | cut -f1 | xargs -I{} gh pr edit "$PR" --add-label {}

# gate a commit on the staged diff (exit 1 when the answer is likely yes)
git diff --cached | hunch --in whole "Does this diff add debug code or secrets?" --max 0.3 >/dev/null

📄 License

Apache-2.0