Changelog

What's new in each release of the command-line tool, newest first. The web app at nehemvix.aelsha.com is always on the latest release.

Upgrade with pipx upgrade aelsha-nehemvix, then check with nehemvix --version. Your saved jobs and remembered file summaries (~/.nehemvix) carry over, so the next run on a repository reuses them.

2.1.0, 10 November 2026

Nehemvix becomes a tool your own AI agents can use. Everything is additive: commands, options and exit codes from 2.0 work as before. How to use it.

New

• nehemvix finish <job-id>: when a guide finished with sections missing (its cap ran out while writing, or a model declined a section), write only those, on the same or any other model. Nothing finished is read or paid for again, and --cap covers the whole job. The web app offers the same as Finish N missing sections on the guide page. Details.
• --json on run, resume, finish, ask and doctor: exactly one JSON result on stdout, on every exit, with the status, the job, scores, cost, the estimate, the files written, the error and what to do next. Human messages go to stderr. The JSON Schema is published (version 1.0).
• Exit code 4, with --json: stopped before spending, a person must approve a cap (or give a price). Without --json that stop is still 1.
• --events jsonl on run, resume and finish: progress as JSON Lines on stderr. --estimate-only: estimate and stop, spending nothing.
• Read commands: nehemvix list, status, outline, section, search and readiness, over your data folder, the same tools the hosted MCP server offers.
• nehemvix mcp: a local MCP server over stdio for your agents. Read-only unless you turn spending on with a provider, a key variable and a ceiling (--max-cap); the agent must then give a cap within it, and the estimate comes first.
• A Python API for your own code and framework adapters: from nehemvix.agentcli.api import Nehemvix. Estimate, start with a cap, wait, pause, resume and read, with typed errors. Details.
• Tested with seven agent frameworks: LangChain, LangGraph, Google ADK, OpenAI Agents SDK, Claude Agent SDK, Microsoft Agent Framework and CrewAI.

Fixed

• Two runs started at the same moment on a new data folder no longer fail while it is being set up.

2.0.0, 10 November 2026

The tool has a new name: AELSHA™ NEHEMVIX. The command is nehemvix, the package is aelsha-nehemvix, and the address is nehemvix.aelsha.com. If you installed the previous package, remove it with pipx uninstall and run pipx install aelsha-nehemvix. Your progress and remembered file summaries move to ~/.nehemvix the first time you run the new command. RL_ settings and existing access tokens are unchanged.

1.1.0, 3 November 2026

1.1.0.post1 changes only the text of the PyPI page (this release's notes and a link here); the code is the same.

New

• Ask a question: nehemvix ask <job-id> "<question>" answers from a finished guide, with its sources and any names it couldn't find in the code. It's the web app's Ask, in a terminal. How to use it.
• Exports: --format html,pdf,docx,pptx,drawio on run and resume writes the same files as the web app's Export button.
• Private repositories: --git-token-env NAME reads a read-only access token from a variable you name (add --git-user if your Git host needs a user name). The token is used for one fetch and never stored. Details.
• nehemvix doctor: checks Python, git (2.50.1 or newer), Graphviz and PDF export, and lists which key variables are set.
• A readable estimate: estimate: $0.018 to $0.038 (likely $0.024), about 19 model calls, about 2 min. When a run stops before spending, it says which --cap is enough.

Fixed

• Wrong input prints one error: line and exits with code 2, instead of a Python traceback. That covers an invalid or http:// link, a private network address, an unknown provider or audience, a temperature out of range, and --price-in without --price-out.
• An --out folder that can't be written is found before anything is fetched or paid for (exit code 1). Before, the error came after the guide was written.
• Ctrl-C prints how to continue and exits with code 130; nehemvix resume picks up where it stopped.
• nehemvix resume on a job that already finished leaves it finished.
• A repository that doesn't exist is reported as "not found, or private".

Changed

• Dependencies moved to their newest stable releases that are at least two weeks old. Known vulnerabilities: none (pip-audit).
• New guide: Integrations, with GitHub Actions, GitLab CI, Docker, make and a grounding-score gate.

1.0.1 was prepared but never published; its fixes are in 1.1.0.

1.0.0, 30 October 2026

First release on PyPI.

• nehemvix run writes a guide for beginners, senior engineers or architects from any Git repository, with your own key: Gemini, Vertex AI, Claude, OpenAI, AWS Bedrock, Microsoft Foundry or any OpenAI-compatible service.
• Every guide has architecture diagrams and citations to exact lines, and every file and code name in it is checked against the code, with a grounding score.
• The run continues only if the estimate is within your --cap; otherwise it stops before spending anything.
• nehemvix resume continues a paused job on the same or another model, without paying twice for finished work.
• Outputs: the guide in Markdown, a readiness checklist and a machine-readable report with the score and the cost.
• Secrets in the repository are removed before anything is sent to a model. Security.
• With some models, a long batch of files could come back cut off, leaving those files without a summary. The published 1.0.0 already reads them again in smaller batches, so every file is covered.