返回资讯中心
外部精选
软件工程
#开发者实时流

Why my AI agents keep lie-ing? How do we trust?

Open source · MIT · New in 0.12: Talk to Blu →

Hacker News 最新Rikin_shah2026/10/30 · 18 天前12 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

Open source · MIT · New in 0.12: Talk to Blu →

dotpals is a small floating pal that watches Claude Code, Codex or any coding agent and checks its work: what it changed, what it ran, and whether the tests really passed. When Claude breaks the tests, dotpals sends it back to fix them. And any AI tool can ask dotpals what really happened.

One command on Windows, macOS or Linux. Free, and your agents’ work stays on your computer. Or straight from GitHub: npx --allow-git=all github:rikinshah787/dotpals setup

Press Ctrl+Shift+Space and talk. Blu looks at your screen when you ask about it, answers out loud, draws on it as it explains, does things on your computer and works in your apps, with the AI you already have.

✓ Explained in 3 steps · writing as it talked

Create a page called Weekly plan in Notion?

✓ Volume 40% → 30% · “turn the music down”

A report card for your coding agents, read from what they did, not what they said. And when a request goes wrong, replay it, or rewind it.

Agents say “done” in one fluent line. dotpals reads what the tests actually printed, and when Claude Code broke them, it sends Claude back to fix them before it finishes, commits or pushes.

The test-output rules build on claude-referee by Ismail Daşcı (MIT). Laya is by Convai Innovations; Jev is by TypeSafe.

✕ Tests failed: expected 3, got -1 (test/math.test.js:5)

Then the tests passed, but only because Claude changed what the test expects:

test('sum adds', () => {

  • assert.equal(sum(1, 2), 3);
  • assert.equal(sum(1, 2), -1);

Sent back “Put the test back and fix the code, or tell the user why the test was wrong.”

✓ Fixed: tests pass now · 3 passed, after the last change

dotpals mcp is an MCP server. Claude Code, Codex, Cursor and other assistants ask dotpals what your agents did, and get the evidence instead of the agent’s word.

Codex, Cursor and others: command npx, arguments dotpals mcp. It only reads, from dotpals on your computer. All the tools →

⚠ Not ready to merge · on branch feat/login

Cursor: Not yet. The tests only pass because auth.test.js was changed to expect the new result, and src/login.ts isn’t committed.

A single request can read dozens of files, edit a handful, run the tests, retry and push. dotpals folds all of it into a few chapters you can read in five seconds, next to what the agent says it did.

And it flags what deserves a second look:

Every step is still one click away, with the exact command, its output and the diff.

You asked Add a dark mode toggle to the settings page

Claude: Added a Dark mode toggle under Settings → Display. The choice is saved and applied on load, with a new dark.css. All tests pass.

A slim island shows every running agent and its live plan, like 2/4 · Detecting the system setting, plus your usage limits: Claude and Codex, 5-hour and weekly, with the time each one resets.

Setup connects Claude Code and Codex for you. Anything else connects with one JSON POST. Sessions never mix, and every Claude Code session shows up, even ones that started before dotpals was installed.

Hooks for live state, including permission prompts, plus each session's transcript, so the history is complete even if the pal opened late.

Follows Codex's session logs, so the CLI, the IDE extension and the app all show up. Nothing to install on the Codex side.

POST JSON to 127.0.0.1:5175/event from your loop, a hook script or a wrapper. Try it on the orange pal:

The same event from a script, with a session id so it gets its own pal and tab:

curl -s localhost:5175/event -d '{ "session": "run-42", "harness": "my-agent", "state": "working", "text": "Running tests" }'

04 Make your own pal

The same registerCustom() the dashboard uses, running live on this page. Pick, roll, then watch it think, work and celebrate.

Every session with its requests, files and full log. Search it, copy a recap for a PR, or export to Markdown or JSON.

Every request, what you asked and what the agent said it did.

Every tool call with its command, output and diff.

Requests per day, files changed, commands and time by project.

Your pal, sounds, notifications and how long to keep history.

A small local server, the bridge, turns what your agents report into one activity feed. It listens only on 127.0.0.1, only answers your own computer, and sends nothing anywhere unless you turn on the cloud test check or give Blu a cloud AI.

bridge127.0.0.1:5175

pal · dashboard · MCP

Give your agent a pal.

One command installs the pal, adds the Claude Code plugin if you have it, picks up Codex, and opens the dashboard.

Or straight from GitHub: npx --allow-git=all github:rikinshah787/dotpals setup

Only want the Claude Code plugin? /plugin marketplace add rikinshah787/dotpals then /plugin install dotpals@dotpals.

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →