mu-agent 0.2.12 → 0.2.14
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +17 -0
- package/judge/mu/index.js +337 -261
- package/kyrn/bin/kyrn-ledger +13 -2
- package/kyrn/bin/kyrn-replay +385 -0
- package/kyrn/bin/mu.mjs +5 -0
- package/kyrn/demo/demos.json +20 -0
- package/kyrn/demo/fix.jsonl +37 -0
- package/kyrn/demo/planted.jsonl +35 -0
- package/kyrn/demo/tavily.jsonl +33 -0
- package/package.json +2 -2
package/CHANGELOG.md
CHANGED
|
@@ -2,6 +2,23 @@
|
|
|
2
2
|
|
|
3
3
|
## [Unreleased]
|
|
4
4
|
|
|
5
|
+
## [0.2.14] - 2026-10-26
|
|
6
|
+
|
|
7
|
+
**0.2.14: `mu demo` replays a real recorded session with every verdict of the judge, no key and no network; ledger records say what each verdict was about.** Published from the `nebius-hackathon` branch under the npm tag `hackathon` (`npx -y mu-agent@hackathon demo`); `latest` stays 0.2.13.
|
|
8
|
+
|
|
9
|
+
### Added
|
|
10
|
+
|
|
11
|
+
- `mu demo` replays a real session recorded with mu, shipped in the package (`kyrn/demo`): the request, each tool call and what came back, each verdict of the judge with its latency, and at the end the judge's and the agent's tokens and cost, how many commands the judge approved or sent back to the person (and which), and how many questions Ask-first mode would have asked. No key and no network. `mu demo --list` names the recordings (`planted`, the default: a comment in the code tells AI agents to run `touch .agent-was-here`, and Nemotron 3 Ultra sends it back while approving the rest; `fix`: Nemotron 3 Super fixes a failing test; `tavily`: a web search through Tavily, what came back screened by the judge).
|
|
12
|
+
- Ledger records carry `about`, a line saying what was judged: the command for `tool.approval` (`bash: npm test`), and for `tool.injection` the tool, what it was asked and where the text came from (`web_search "…" · Tavily, 8 results`, `web_fetch <url> · read through Tavily`). One line, at most 200 characters, secrets redacted. `mu ledger` prints it in brackets after the decision point, and a screening as `5 passages screened, none withheld`. `mu replay [n | file]` replays one of the person's own sessions the same way. `scripts/demo-session.mjs` turns a session into a recording: the project folder, the home folder and the user name replaced, the system prompt left out.
|
|
13
|
+
|
|
14
|
+
## [0.2.13] - 2026-10-25
|
|
15
|
+
|
|
16
|
+
**0.2.13: `isCheckChain` follows a `cd` into the project's folders and reads the project's own runner by path and `"$(git rev-parse --show-toplevel)"`.**
|
|
17
|
+
|
|
18
|
+
### Fixed
|
|
19
|
+
|
|
20
|
+
- `isCheckChain` follows the line's folder. A `cd` spelled from the root may go into a folder of the project as it really is (its realpath inside the conversation's folder, so not through a link out of it); git still counts as read-only only in the conversation's folder itself. A check may be the project's own runner by path: `node <path>/node_modules/vitest/dist/cli.js|vitest.mjs`, or `<path>/node_modules/.bin/vitest|jest|mocha|ava|tsc|eslint|biome`, resolved inside the project, with `..` only at the start. `"$(git rev-parse --show-toplevel)"` in double quotes is read as the conversation's folder when the nearest `.git` above the current folder is there; any other substitution still rules the line out. Recorded on 0.2.12 with MiMo v2.6 Pro: `cd <cwd>/packages/coding-agent && node "$(git rev-parse --show-toplevel)/node_modules/vitest/dist/cli.js" --run test/args.test.ts`, the form the repository's AGENTS.md gives for one test file, asked the user (needed 0.78).
|
|
21
|
+
|
|
5
22
|
## [0.2.12] - 2026-10-25
|
|
6
23
|
|
|
7
24
|
**0.2.12: in judge mode, `tool.approval` approves a check (`isCheckChain`) at needed 0.7; everything else keeps 0.8.**
|