jev-codes

CLI + agent commands that audit a git diff against editable standards packs, scored hunk-by-hunk with TypeSafe's Jev model. Jev never writes code — it answers typed questions (yes/no, score, choice) with a probability, and your agent fixes only the flagged hunks.

A 500-line diff audits in seconds for a fraction of a cent: one parallel Jev call per hunk, output tokens free.

Index

• Install
• Usage
• Supported harnesses
• How it works
• Standards packs
• Evals
• About Jev
• Data path & telemetry

Install

npx @kushwho/jev-codes

With no arguments it onboards inline: asks for your TypeSafe key, validates it, saves it to ~/.jev-codes/config.json (mode 600), and offers the plugin install to supported harnesses. Or be explicit:

npx @kushwho/jev-codes init [--key <k>] [--no-plugin]

TYPESAFE_API_KEY in env always beats the config file. pin with npx @kushwho/jev-codes@latest.

Usage

npx @kushwho/jev-codes audit                  # staged diff (default)
npx @kushwho/jev-codes audit --staged --json  # agent-parseable report
npx @kushwho/jev-codes audit --base main --pack core --fail-on high

Or via installed bins (jev-codes / jev). audit --json writes only JSON to stdout (humans to stderr) and exits 1 when findings reach --fail-on (default high in CI/--json, none for interactive TTY).

More: packs list|add|show (fetch community packs), update (pull newer packs with a question/threshold diff preview, never silent).

Supported harnesses

Every adapter runs the same audit --json loop — PATH binary first, npx fallback — and fixes only high/medium findings at the reported lines. init offers to install for any combination: pick 1–5, all, or none (or pass --harness claude,cursor non-interactively, --no-plugin to skip).

Harness	Entry point	Invoke
Claude Code	claude-plugin/commands/jev-audit.md	/jev-audit
Cursor	.cursor/commands/jev-audit.md	/jev-audit
Codex CLI	.codex/prompts/jev-audit.md	/jev-audit
opencode	.opencode/commands/jev-audit.md	/jev-audit
Antigravity	.agents/skills/jev-audit/SKILL.md	/jev-audit

How it works

git diff → parse hunks → filter (ignores, .env*, gitignored)
→ build state {file, language, hunk, context_before/after}
→ one Jev call per hunk (all pack questions, 8 in flight)
→ thresholds + min_confidence → JSON / Markdown / TTY report

Only added/changed lines are judged. Hunk + context is capped under 24k tokens (est. chars/3 for code) and split at blank lines when over. Results cache by hash of state + question text in ~/.jev-codes/cache/, so re-runs only re-score changed hunks; jev-latest entries expire after 24h so model rollovers can't serve stale answers.

Standards packs

Packs are YAML — the only thing that defines "good code"; the CLI has no opinions. Bundled core (pinned to the CLI release) covers leftover debug, duplicate logic, swallowed errors, comment noise, over-abstraction, plus a report-only change_kind label. Per-repo overrides live in .jev-codes/standards.yaml (extends: core, override by question key). noul questions skip the confidence gate by design — the threshold is the whole decision.

Evals

jev-codes-eval/ holds 15 hand-labeled diffs plus a runner (node jev-codes-eval/run.mjs) against live Jev. Latest result on jev-1.13.0 (17 calls, ~8s, $0.0006):

• HIGH gate: 6/6 → precision 1.000 (target ≥0.80)
• Overall precision 1.000, recall 0.909 — single miss is case 04 (duplicate_logic), a model-discrimination limit, not a tunable threshold
• Labels (change_kind): 14/17

That run is what tuned leftover_debug 0.70 → 0.75 (killed the only false positive with 0.2 margin). Rerun any time; --verbose prints per-question values, --case <prefix> scopes to one case.

About Jev

Jev is TypeSafe's System One model: you send state + typed questions, you get structured answers back — no text generation, no parsing. Three primitives, mixed in one call: noul (yes/no probability), score (position on your ordered levels), choice (one of ≤255 options + confidence). Every question is evaluated in parallel against the same state, so a tenth question costs tokens but almost no time. Model alias is jev-latest; reports always record the concrete response_model. Reference notes live in docs/jev.md.

Data path & telemetry

audit sends per-hunk state to https://api.typesafe.ai: file, language, unified-diff hunk, plus up to context_lines of surrounding code. Whole files are never uploaded; ignored/.env*/gitignored paths are never sent. Auth via TYPESAFE_API_KEY (env beats config file).

Telemetry: none.