onesie

module

v0.3.0 Latest Latest Go to latest Published: Sep 26, 2026 License: MIT

Details

Valid go.mod file
Redistributable license
Tagged version
Stable version
Learn more about best practices

Repository

github.com/frodi-karlsson/onesie

Links

Open Source Insights

README ¶

onesie

Probably the most robust System One CLI.

Pipe in some text, ask a question, get an answer you can script against. Asks TypeSafe's Jev, directly or through OpenRouter, or Berget's own System One models.

echo 'EVERYTHING IS DOWN, CALL ME NOW' | onesie 'does this convey urgency' -r
# 0.99

Gate an agent's shell commands. The exit code says what to do:

onesie -f shell-safety -q --state 'git status'
# exit 0: run it
onesie -f shell-safety -q --state 'rm -rf ~'
# exit 1: block it
onesie -f shell-safety -q --state 'echo hello > notes.txt'
# exit 7: not sure, ask a person

See Thresholds from evidence for where that gate's numbers come from.

Install

Homebrew, on macOS or Linux:

brew install frodi-karlsson/tap/onesie

The install script checks the download against the release checksums, and its build provenance when gh is logged in:

curl -fsSL https://raw.githubusercontent.com/frodi-karlsson/onesie/main/install.sh | sh

With Go:

go install github.com/frodi-karlsson/onesie/cmd/onesie@latest

Then store a key with onesie auth set.

In Claude Code, the plugin can come first and walk you through the rest:

/plugin marketplace add frodi-karlsson/onesie
/plugin install onesie@onesie

Thresholds from evidence

onesie calibrate runs your questions over records you have already labelled, and shows what each cut catches and what it flags by mistake. Here's the destroys question from shell-safety over 40 labelled commands:

onesie calibrate -f shell-safety -i jsonl --map .command --id .id -m jev-1.13.0 \
    --label destroys=.destroys --label secrets=.secrets --label network=.network \
    --cuts 0.25,0.5,0.65 --out answers.jsonl --resume < commands.jsonl

destroys, yes/no: labelled 40, 16 yes, 24 no, 0 failed. AUC 1.00
flagged means destroys.value >= cut

  cut   flagged  catches             false alarms   right when flagged
  0.25       18  16/16 100% 81-100%  2/24 8% 2-26%  16/18  89% 67-97%
  0.50       15  15/16  94% 72-99%   0/24 0% 0-14%  15/15 100% 80-100%
  0.65        7   7/16  44% 23-67%   0/24 0% 0-14%   7/7  100% 65-100%

worst misses
  echo-overwrite  labelled yes  answered 0.37

For destroys, shell-safety blocks at 0.5 and passes below 0.25. Anything in between goes to a person, like the worst miss above.

Commit answers.jsonl, and CI runs the same command with --offline and requirements, which needs no key:

onesie calibrate ... --offline \
    --require 'destroys.catches >= 1' --require 'destroys.false_alarms <= 0 at abstain'

Exit 1 means a requirement no longer holds. Exit 2 means the questions changed, so ask again.

To test a script that uses the gate, ONESIE_MOCK answers every onesie call from a file:

echo '{"destroys": 0.93, "secrets": 0.02, "network": 0.1}' > danger.json
ONESIE_MOCK=danger.json ./gate.sh   # takes the blocked branch

Asking

A yes or no question answers with how likely the yes is. --pick chooses one of your options, and --rate places the text on a scale you list from lowest to highest. -r prints only the answer.

echo 'I was charged twice for my order' | onesie 'who should handle this' --pick billing,shipping,technical -r
# billing
echo 'Third time asking. Fix it or I am leaving.' | onesie 'how frustrated is the customer' --rate calm,annoyed,furious -r
# furious

Give each question a name with --ask, and they all go out in one request:

onesie --ask urgent='is this urgent' \
       --ask team='who should handle this' --pick billing,shipping,technical \
       -o values < ticket.txt
# {"urgent":0.69,"team":"billing"}

--print-questions > .onesie/questions/NAME.yaml saves the questions to a file, and -f NAME loads them from anywhere in the repository.

Recipes

Triage tickets as they come in.

tail -f tickets.jsonl | onesie 'is this urgent' -i jsonl --map .body --id .id -o json

Wire the gate into a script. Exit 0 runs the command, 1 blocks it and 7 asks a person. Any other exit code means there was no answer.

onesie -f shell-safety -q --state "$cmd"
case $? in
    0) eval "$cmd" ;;
    1) echo blocked ;;
    7) ask_the_user ;;
    *) echo 'no answer, blocked' ;;
esac

Triage a spreadsheet. The same rows come back with a column per question.

onesie --ask urgent='does `body` convey urgency' -i csv -o csv --merge < tickets.csv > triaged.csv

Run a big batch you can resume. --map sends only the body, --id names each answer, and --resume skips every id the file already answers and asks the rest, failed ones included.

onesie 'is this urgent' -i jsonl -j 8 --map '.body' --id '.id' \
    --out answers.jsonl --resume < tickets.jsonl
jq -c 'select(.answer.value > 0.8)' answers.jsonl

Comment on a pull request. The output starts with the gate result as a GitHub alert, followed by a table of the answers.

onesie 'does this explain why the change is needed' --assert 'answer.value > 0.6' \
    -o markdown --state "$PR_BODY" | gh pr comment "$PR" --body-file -

In GitHub Actions, a failing gate fails the step only under pipefail, so name shell: bash or run set -o pipefail first. A comment holds at most 65536 characters, so send a large run to $GITHUB_STEP_SUMMARY instead.

Rank records by several questions. One --ask per dimension scores every record, and jq weighs the answers and sorts by the sum. .value is the yes/no probability, and .norm is the rate's position between 0 and 1. The weighting lives in jq, since --assert has no arithmetic, and select(has("error") | not) drops the records that failed.

onesie --ask impact='does this help many users' \
    --ask effort='how much work is this' --rate trivial,small,medium,large \
    -i jsonl --map '.body' --id '.id' -o json < ideas.jsonl |
    jq -s 'map(select(has("error") | not)) | map({id, score: (0.7 * .impact.value + 0.3 * (1 - .effort.norm))}) | sort_by(-.score)'

Starter gates. shell-safety, prompt-injection, personal-data and moderation are built in and calibrated. onesie questions lists them, and examples/README.md has their tables.

Also

Providers. TypeSafe is the default, and --provider openrouter or --provider berget asks through OpenRouter or Berget instead.
Caching. --cache or ONESIE_CACHE=1 answers a repeated request from disk, and a stream asks records whose request is identical once, giving each its own line.
Dry runs. --print-request prints the exact request without sending it, and needs no key.

Why no MCP server

onesie ships agent skills instead of an MCP server:

Validation. The CLI already checks every question, flag and assertion before any request, and an agent reads the same messages and exit codes a person does.
Documentation. onesie --help, -V and the error messages come from the binary, so they stay in step with it.
Discoverability. Skills have room for more prose than a tool description, such as worked recipes and the mistakes to avoid, and CI checks their examples against the binary.

Reference

REFERENCE.md covers asking, gating, testing a gate, calibrating, streams and resume, caching, keys and providers, exit codes and the rest of the flags.

Contributing

See CONTRIBUTING.md.

License

MIT. See LICENSE.

Directories ¶

?	: This menu
/	: Search site
f or F	: Jump to
y or Y	: Canonical URL

Path	Synopsis
cmd
examplesanswers command Command examplesanswers regenerates the starter sets' committed answers files from the live API, with the pinned model.	Command examplesanswers regenerates the starter sets' committed answers files from the live API, with the pinned model.
formulagen command Command formulagen writes the Homebrew formula for a release from the release's checksums.txt.	Command formulagen writes the Homebrew formula for a release from the release's checksums.txt.
onesie command Command onesie is the command line interface for TypeSafe Jev.	Command onesie is the command line interface for TypeSafe Jev.
proseblocks command Command proseblocks reads a unified diff on stdin and writes the prose blocks of the Go and Markdown files that overlap an added line as JSON lines of {"id":"path:line","text":"..."}.	Command proseblocks reads a unified diff on stdin and writes the prose blocks of the Go and Markdown files that overlap an added line as JSON lines of {"id":"path:line","text":"..."}.
releaseprep command Command releaseprep checks a release tag against the existing ones and bumps the plugin manifests to it.	Command releaseprep checks a release tag against the existing ones and bumps the plugin manifests to it.
skillcheck command Command skillcheck dry runs every rule's bad and good example under skills/, and checks the flags and commands the prose names against onesie --help.	Command skillcheck dry runs every rule's bad and good example under skills/, and checks the flags and commands the prose names against onesie --help.
skilleval command Command skilleval dry runs every onesie command an agent wrote during the last claude plugin eval run.	Command skilleval dry runs every onesie command an agent wrote during the last claude plugin eval run.
skillgen command Command skillgen generates the per client skill files from skills/.	Command skillgen generates the per client skill files from skills/.
examples Package examples describes the starter question sets, so the tests and the tool that regenerates their answers read each set the same way.	Package examples describes the starter question sets, so the tests and the tool that regenerates their answers read each set the same way.
questions Package questions holds the starter question sets built into onesie, so -f finds them by name with no copy step.	Package questions holds the starter question sets built into onesie, so -f finds them by name with no copy step.
internal
answer Package answer normalizes API answers into the shape the CLI prints and applies the per question policy.	Package answer normalizes API answers into the shape the CLI prints and applies the per question policy.
argv Package argv records group local flags in the order pflag encounters them, so the question groups in a command line can be reconstructed after parsing.	Package argv records group local flags in the order pflag encounters them, so the question groups in a command line can be reconstructed after parsing.
assert Package assert evaluates a boolean expression over a normalized record.	Package assert evaluates a boolean expression over a normalized record.
cache Package cache stores API responses on disk under a sha256 key, with private modes, atomic writes, a lifetime for model aliases and a size cap enforced by least recently used eviction.	Package cache stores API responses on disk under a sha256 key, with private modes, atomic writes, a lifetime for model aliases and a size cap enforced by least recently used eviction.
calibrate Package calibrate reads the labels a calibration run compares answers with, and scores and reports how well the answers agree with them.	Package calibrate reads the labels a calibration run compares answers with, and scores and reports how well the answers agree with them.
cli Package cli assembles the onesie command tree.	Package cli assembles the onesie command tree.
creds Package creds resolves, reads and writes the credential file that holds an API key when neither a flag nor the environment supplies one.	Package creds resolves, reads and writes the credential file that holds an API key when neither a flag nor the environment supplies one.
engine Package engine runs one evaluation per record the source yields, bounded by concurrency and ordered by input position.	Package engine runs one evaluation per record the source yields, bounded by concurrency and ordered by input position.
formula Package formula renders the Homebrew formula for a onesie release from the release's checksums.	Package formula renders the Homebrew formula for a onesie release from the release's checksums.
input Package input resolves where a state comes from and reads it, so the detection rule lives in one place rather than being rediscovered at every call site.	Package input resolves where a state comes from and reads it, so the detection rule lives in one place rather than being rediscovered at every call site.
interrupt Package interrupt lets a caller stop waiting on a read that has no deadline.	Package interrupt lets a caller stop waiting on a read that has no deadline.
jev Package jev is a client for the TypeSafe System One API, ported from the MIT licensed @typesafe-ai/sdk JavaScript client.	Package jev is a client for the TypeSafe System One API, ported from the MIT licensed @typesafe-ai/sdk JavaScript client.
jq Package jq compiles and runs the jq expressions that --map and --id take, through gojq.	Package jq compiles and runs the jq expressions that --map and --id take, through gojq.
limits Package limits holds every API limit onesie enforces locally and every constant it defaults to, so a binary built against stale numbers is diagnosable through --version.	Package limits holds every API limit onesie enforces locally and every constant it defaults to, so a binary built against stale numbers is diagnosable through --version.
mock Package mock reads a file of canned answers and checks it against the questions a run asks, so the run can answer from it with no key and no network.	Package mock reads a file of canned answers and checks it against the questions a run asks, so the run can answer from it with no key and no network.
output Package output encodes a normalized record in each of the modes the CLI offers.	Package output encodes a normalized record in each of the modes the CLI offers.
plan Package plan turns a recorded command line into a validated invocation, ready to run.	Package plan turns a recorded command line into a validated invocation, ready to run.
proseblocks Package proseblocks finds the pieces of prose a change touches in Go and Markdown files: one block per comment group, paragraph or list item.	Package proseblocks finds the pieces of prose a change touches in Go and Markdown files: one block per comment group, paragraph or list item.
qfile Package qfile loads a question file or a raw API request body into the ordered question slice the planner already consumes.	Package qfile loads a question file or a raw API request body into the ordered question slice the planner already consumes.
release Package release checks a release tag against the existing ones and bumps the plugin manifests to it.	Package release checks a release tag against the existing ones and bumps the plugin manifests to it.
skillcheck Package skillcheck dry runs every rule's bad and good example in a skill.	Package skillcheck dry runs every rule's bad and good example in a skill.
skilleval Package skilleval dry runs every onesie command an agent wrote during a claude plugin eval run, since the eval has no grader that can run code.	Package skilleval dry runs every onesie command an agent wrote during a claude plugin eval run, since the eval has no grader that can run code.
skillgen Package skillgen decodes and validates skill.json, the schema a generated skill is built from.	Package skillgen decodes and validates skill.json, the schema a generated skill is built from.