4 findingsStatic scanned
Evolves the planner, reviewer and retry policies around your model, testing each change in a sandbox.
347.3k installs/moMIT
Weight Eft
metaharness · Agent
Weight Eft is for teams running an AI coding agent with a model cascade, where a cheaper open-source model such as GLM, Qwen or DeepSeek tries first and hard problems go to a frontier model. It reads your agent's archive of successful and failed runs and builds LoRA fine-tuning data for the cheaper model: supervised examples from every resolved run, preference pairs from the cheaper model's own wins and failures, and a plan for training on a GPU host. The aim is for fewer tasks to need the frontier model.
Written by ooruby from the README of @metaharness/weight-eft 0.1.2.
Not claimed by its maker. Is this yours? Prove it and answer the findings
Why it's useful
Escalating to a frontier model costs money, and this aims to have your cheaper model handle more tasks by learning from runs your agent already solved.
What you could do with it
- Export your agent's resolved runs as OpenAI chat JSONL to fine-tune Qwen with LoRA on a GPU host.
- Hold out your evaluation tasks so the exporter refuses training data that overlaps them.
- Review the export report to see which runs were dropped for reward hacking or excessive length.
Written from its documentation: what it says it can do, not what our checks found.
Maintenance
Not measuredNo release date is on record for this package yet. The nightly publisher check records one as it reaches each npm package, about once every ten nights.
Release activity only: it says nothing about quality or safety, and a finished small package can be fine without releases. How it is measured
What this does not cover
- Everything. Treat it as you would software from anywhere else.
Receipt history: every signed receipt for every version of this listing, and what changed between them.
In its maker's words
Fine-tune cheap open-source LLMs (GLM, Qwen, DeepSeek) on your AI coding agent's successful runs with LoRA (SFT + DPO) so your model cascade escalates to expensive frontier models (GPT, Claude) less often — cutting cost-per-resolved. Turns run history int
Compare with
Compare all 4How far this has been checked
Static scanned
- What it establishes
- Its published package, or for a hosted server its source repository at a recorded commit, was read file by file, without running it, by every check in our current scan.
- What it does not
- How it behaves when it runs, or anything the scan does not read yet: several rubric checks, the repository's history, and compiled code in folders named dist or build. For a hosted server, that the endpoint runs the code that was read. The ooruby Index sets out exactly what was read.
- Exactly what was read
- Version 0.1.2, the file with digest sha512-2C6Ivgz1yJgRrklG0NqsNAjm+THyFsfX4DR2VyeqkLHJeR+LnVEOmzdUPQPp0yBJcXvoO9+du8R6HzZ4vDItrQ==, as the registry publishes it. A published version cannot be replaced, so this names the same bytes for anyone who checks.
- Authentication
- Runs locally. No hosted endpoint is listed for this package. You run it on your own machine, so there is no endpoint of ours to authenticate against.
Adding it
Run it with:
npx -y @metaharness/weight-eft@0.1.2Pinned to 0.1.2, which is the version the findings above were found in. Drop the version to take whatever is newest, and the report on this page stops describing what you installed.
Where this came from
- Registry
- npm
- Package
- @metaharness/weight-eft
- Version read
- 0.1.2
- Resolved
- 2026-10-09
No build provenance published. The repository above is the one the publisher declared, and nothing links it to the package you would install. That is not a mark against this listing, since most packages are published this way, but it is a check nobody can run.
The badge, if this is your listing
It renders the current rung (static scanned) and links back here, where what that does and does not establish is one click away. It updates itself as the evidence deepens.
[](https://ooruby.com/market/metaharness-weight-eft)About this listing
- Kind
- Agent
- Category
- Developer tools
- Pricing
- Free
- Sandbox
- No
- Hosted
- No
- Updated
- 2026-10-09
More from metaharness
See the storefrontSimilar agents
All agentsVerification records what our published tests found on a specific version at a specific date. It is not a warranty, and it does not certify that software is free of defects.