Jev in Codex

A Codex plugin for batch labelling with Jev. It turns a JSONL file of text records into a complete labelled file, while Codex reviews uncertain decisions.

One tool: jev_label, with custom question/criteria policies and the feedback_theme preset. The previous capability-selection, context-search and output-triage tools have been removed because their tested workflows did not show a useful improvement. There is no lexical or local-classifier fallback.

Define your own labels for topics, sentiment, document collections or other text decisions. See custom labelling for a complete example, taxonomy design and limits. Each call assigns one label per record.

Why this capability is included

The 0.3 custom-policy benchmark passed separately on action-needed, sentiment, document routing and feedback: 21.5–31.9% less total Codex input and 8.4–23.5% less elapsed time across four paired repetitions per task. Both arms labelled all records correctly. These are 64-record synthetic workloads, not a guarantee for every taxonomy or batch size.

The original 0.2 release benchmark on 64 synthetic feedback messages measured 42.7% less total Codex input and 26.4% less elapsed time, with all labels correct across four paired runs. Tiny batches performed worse in the earlier sweep.

The benchmark report includes the shipped implementation check, prototype results, methods, complete artifacts and negative results. This is exploratory verification on known synthetic development data, not independent confirmation. It does not prove real-world accuracy, general labelling superiority, invoice savings or automatic desktop adoption.

New capabilities, built-in presets and workflow expansions require a passing, reviewable benchmark against plain Codex before being exposed. User-supplied taxonomies use the custom-policy interface; their accuracy and benefit still need validation on the actual data. See the admission policy.

Related MCP server: siftr

Use

Prepare an approved UTF-8 JSONL file inside the configured workspace:

{"id":"feedback-1","text":"Saved reports disappear after reopening. Please preserve them."}
{"id":"feedback-2","text":"Please offer a cheaper plan for occasional users."}

Call:

{"path":"feedback/items.jsonl","policy":"feedback_theme"}

The tool creates feedback/decisions.jsonl, containing every original ID and one label:

{"id":"feedback-1","label":"reliability"}
{"id":"feedback-2","label":"pricing"}

These examples explain the format; a two-record batch is not the benchmarked use case. Labels are reliability, usability, pricing and feature. The response includes the policy, counts, model, request count, output path, and original text for decisions below confidence 0.8. Codex can correct that file after review. Above-threshold confidence does not guarantee correctness.

Use output_path to choose another new relative file. Existing files are never overwritten. Both input and output stay inside the configured workspace; the output parent must exist. Input accepts 1–256 records, unique IDs of up to 100 letters/digits/underscore/hyphen, nonempty text up to 6,000 characters per record, and no extra fields. Files are limited to 1 MiB. These are operational limits, not validated performance ranges.

Install

Ask Codex:

Install https://github.com/teempai/jev-in-codex using docs/INSTALL.md, configure my workspace and TypeSafe key privately, and verify that only jev_label is available. Preserve existing instructions. Do not add global instructions.

See installation and migration. Node.js 22+ is required; ripgrep is no longer required. Build with npm ci --ignore-scripts && npm run check. Set TYPESAFE_API_KEY privately and pass --root or JEV_WORKSPACE_ROOT. The model is pinned to the benchmarked jev-1.13.0; the old JEV_MODEL override is no longer used.

For direct MCP configuration, resolve the Node and built-server paths on your machine:

[mcp_servers.jev]
command = "/absolute/path/to/node"
args = ["/absolute/path/to/jev-in-codex/dist/index.js", "--root", "/absolute/workspace"]
env_vars = ["TYPESAFE_API_KEY"]
tool_timeout_sec = 90

Prefer the plugin installation for its bundled skill. Do not register both copies. The tool writes a file, so normal host approval may apply. Installation does not change global approval policy or persistent global instructions.

Data and failure handling

Every input record and the selected policy are sent over HTTPS to https://api.typesafe.ai/v1/systemone. Jev makes the decisions; local code only validates, batches, serializes and selects uncertain records for review. Up to eight requests run concurrently, with at most eight questions per request and a 28,000-byte body budget. Requests have an eight-second timeout and a twenty-second total inference deadline.

Missing credentials, malformed responses or a failed batch produce an explicit error and no labelled output. No local labels replace Jev results. Existing outputs, traversal, escaping symlinks, common credential paths, invalid UTF-8, binary and oversized input are rejected. Provider error bodies are not exposed. The production server has no content logs, telemetry or persistent cache.

Filename exclusions are best-effort, not secret detection. Use only data authorized for TypeSafe. This local path boundary is not a sandbox against concurrent malicious filesystem changes. Labels are advisory and never authorize external actions. See the testing guide; the older security review covers the retired implementation, not this new writing capability.

Development

Contributing · Benchmark report and reproduction · Capability admission policy

MIT. Inspired by the TypeSafe Jev approach to bounded decisions and fast-jev-compaction; this repository does not include that project's compactor code.