Add mu

[qybaihe]

Projects

• qybaihe/mu: a coding agent (CLI and desktop app) built on Pi. It asks Jev short typed questions at decision points in its loop, for example which chunks of a long tool output enter the context, whether a rule-flagged command was asked for, and whether a web page, search result or MCP output carries instructions aimed at the model. Code acts on each answer. Category: developer_tools.

Jev integration evidence (optional)

• packages/kyrn-judge/src/providers/typesafe.ts: the System One request (https://api.typesafe.ai/v1/systemone, model jev-latest) and answer parsing; yes/no questions go out as noul.
• packages/kyrn-judge/src/decisions/tool-admission.ts: the per-chunk question for long tool output.
• packages/kyrn-judge/src/decisions/tool-risk.ts: the questions about a rule-flagged command.
• packages/kyrn-judge/src/decisions/tool-injection.ts: the passage-by-passage prompt-injection question.
• README.md: setup (npm i -g mu-agent, Node 22.19+) and which keys each Jev provider needs.

The same paths are in the entry's evidence field.

Affiliation

I maintain mu. It is developed with heavy use of AI coding agents, and this pull request was drafted by an AI agent at my request; I reviewed it before opening.

Checklist

• This PR changes 1–10 entry JSON files and no other files.
• Each repository is public and source code is licensed. (MIT for packages/ and kyrn/; desktop/ is Apache-2.0.)
• Each description is factual and source files support its claims.
• I ran npm run check (and npm test), both pass.
• I understand that unresolved entries must be resolved, removed, or split before the whole PR is merged.

Generated with Claude Code

[github-actions]

Jev review

Review could not be completed

Follow-up:

• ai-gateway.vercel.sh returned HTTP 400

PR commit: 429e5cbbd8e5638e356e2d4c5fcad0552a78f9ce
Review run and JSON report

Model judgments use Jev only; this comment is generated from a template. Probabilities are model judgments, not verified accuracy. A maintainer decides whether to merge.

[fatwang2]

Maintainer review of 429e5cb: approved after manual source review.

At qybaihe/mu@47c51b0 I inspected the TypeSafe System One provider and its registry wiring, the typed tool.admission/tool.risk/tool.injection decisions, and their registered tool_call/tool_result handlers. They support the submitted description: classifying long tool-output chunks, checking whether a rule-flagged command was requested, and screening external text for instructions directed at the model. The README documents CLI/desktop setup, judge credentials and modes. Root MIT and desktop Apache-2.0 licenses cover the public implementation. developer_tools fits; no effectiveness or safety guarantee is asserted by the entry.

Catalog checks passed, as did local validation and README generation; all four combined catalog tests passed. The advisory Jev review failed with an infra-classified ai-gateway.vercel.sh HTTP 400 and failed again after one rerun. That is not a successful automated review; the approval rests on the manual evidence above. I found no concrete author-side blocker requiring changes.