# jev-cli

> A Python CLI (jcli) that sends JSON, NDJSON, JSONC, and text logs to TypeSafe AI's Jev decision model and returns typed, confidence-scored answers you can gate CI on.

jev-cli is a Python command-line tool, invoked as `jcli`, that analyzes logs, audit trails, and other system artifacts with TypeSafe AI's Jev decision model. Instead of prose summaries, it returns typed answers, each tied to the source lines it came from and carrying a number that says how sure the model is. It is an early alpha project by Josh Long.

## What It Is

jcli reads JSON, NDJSON, JSONC, or plain text logs, groups records into windows, and sends each window to Jev along with a set of typed questions. Jev answers every question in one call, so the README notes that asking eight questions costs about what asking one costs. Answers come back as one of three types: `noul` (a probability that a statement is true), `choice` (a label plus its full distribution), or `score` (a position on an ordered rubric).

## Question Packs and Inline Questions

Questions are grouped into YAML packs. Five ship built in: `logs.triage`, `logs.anomaly`, `security.audit`, `security.finding`, and `agent.session`. Several packs can be combined in one run without adding requests. You can also define a one-off question with `--ask`, scaffold and validate your own packs, and override or drop individual questions per run with `--set` and `--drop`.

## Text Logs Through Transforms

Most logs are not JSON, so jcli uses transforms: small YAML files that describe where a record starts and which named regex groups become fields. Only `syslog` ships built in; custom transforms can be passed with `-t` or dropped into the config folder for auto-detection. Auto-detection scores every transform by parse rate, and if there is no clear winner the run stops and lists candidates instead of guessing. Lines that fail to parse are kept under `_unparsed` and counted in a warning.

## Confidence Handling

`choice` and `score` answers carry the API's own confidence. `noul` answers do not, so jcli reports a separate certainty computed from the distance to 0.5 and marks it in its own column rather than mixing the two. Flags let you flag weak answers (`--mark-uncertain`) or drop them (`--confident`), each with an explicit threshold.

## Using It as a CI Gate

`--fail-on` takes a boolean expression over question IDs and exits with code 6 when any window matches. Other exit codes separate auth failures, rate limiting, and partial failures, and a partial failure still emits the results from windows that succeeded. Filtering, field selection, and secret redaction all run before any data leaves the machine, and `--dry-run` prints the exact request bodies without an API key or network access.

## Features
- Analyze JSON, NDJSON, JSONC, and text log files
- Five built-in question packs for log triage, anomaly detection, security audit, security findings, and agent sessions
- Three typed answer types: noul, choice, and score
- All questions in a window evaluated in one API call
- Inline questions via --ask
- Custom question packs with init and validate commands
- Per-run question overrides with --set and --drop
- Text log transforms with parse-rate auto-detection
- Record-aware filtering with =, !=, ~, !~, >, >=, <, <= operators
- Field projection with --select
- Secret redaction before data leaves the process
- Configurable windows: N records, per record, or whole input
- Confidence and certainty thresholds with --mark-uncertain and --confident
- CI gate via --fail-on with distinct exit codes
- Output as table, JSON, NDJSON, CSV, or summary
- Dry-run mode that prints exact request bodies
- probe command to measure the API's state size limit
- Credential lookup from flag, file, env var, config file, or keyring
- Config file defaults in ~/.config/jcli/config.toml

## Integrations
TypeSafe AI Jev API, TypeSafe Python SDK, uv, keyring

## Platforms
CLI

## Pricing
Free

## Version
0.1.0

## Links
- Website: https://github.com/joshLong145/jev-cli
- Documentation: https://github.com/joshLong145/jev-cli#readme
- Repository: https://github.com/joshLong145/jev-cli
- EveryDev.ai: https://www.everydev.ai/tools/jev-cli