New advertisers: 50% off your first month with code FIRSTMONTHSee slots

Reads Claude Code session transcripts in place to flag misread intent, broken rules, and false claims.

version
0.1.0
GitHub stars
3.8K

What it does

  • Adds /vibesenpai:insights, /vibesenpai:judge, and /vibesenpai:report for a findings pane, model-read control, and HTML history report
  • Covers misread intent, broken constraints, misreported work, overreach, edit retries, repeated calls, and cache rebuilds, quoted in your own words
  • Enforces no transcript copies and no sends until you approve the exact text; model reads stop after /vibesenpai:judge off
  • Designed for developers who want to see where agents failed and apply a fix without the plugin writing files

Reads the Claude Code session transcripts already on disk and shows where the agent went wrong: the turns where it misread intent, broke a stated constraint, misreported its own work or overreached, quoted in the developer's own words, plus edit retries, repeated calls and cache rebuilds. A pane beside the chat lists the findings and offers a fix for each kind (a rule, a permission or a hook); Apply drafts it into the prompt box and nothing changes until the developer sends it. Transcripts are read in place and never copied. Model reads are off until the developer says yes after seeing the exact text they would send; they then send redacted session text to Claude Haiku on the user's own Claude Code account, within about 2M tokens a day, until /vibesenpai:judge off.

Add marketplace

/plugin marketplace add davepoon/buildwithclaude

Run it inside Claude Code once, then install the plugin below.

Install

/plugin install vibesenpai@buildwithclaude

Run it inside Claude Code after adding the marketplace.

Readme

vibesenpai

The mascot: an old sensei in a straw hat, pointing at turn 41 in a transcript, where an edit was refused because the file was not read first.

CI License: MIT Node ClaudePluginHub

Discover how your agents failed you. vibesenpai reads the sessions already on your disk and shows you where they went wrong: the exact turns, in your own words, and a fix you can apply.

It is a Claude Code plugin. It shows evidence, and a fix you can apply. It does not score you and does not tell you how to work.

Documentation · What it shows you · How your data is used

Demo: the developer says do not push, and the agent pushes to main anyway. Next session, the finding is waiting, with the developer's own words and the turn that broke them, and Apply drafts a fix that makes Claude Code ask before every git push.

Install

In Claude Code:

/plugin marketplace add vibesenpai/vibesenpai
/plugin install vibesenpai@vibesenpai

It needs Node 22.13 or later on your PATH, and Claude Code 2.1.299 or later.

Usage

  1. Keep working. The free checks read every session, on your machine, and send nothing.
  2. Open the pane. /vibesenpai:insights opens it beside the chat. Each finding leads with your own words and the turns where it went wrong. The model reads, which find misread requests, broken rules and false claims, are off until you say yes: Turn on… shows the exact text first, and Enter turns them on. Turned on, they keep running on Claude Haiku, in your own Claude Code account, after each turn ends, until /vibesenpai:judge off. They are cheap: usually a few cents a session at most, within 2M tokens a day.
  3. Apply a fix you agree with. Apply drafts the change into your prompt box. Your Enter is the yes, and the agent makes the change under Claude Code's own permission prompt.
CommandWhat it does
/vibesenpai:insightsthe pane: findings, their fixes, and Apply
/vibesenpai:judgeon shows what model reads would send, yes turns them on, off turns them off; no argument shows their status
/vibesenpai:reporta self-contained HTML report of your history

/vibesenpai:judge off stops the model reads and keeps the free checks. /plugin disable vibesenpai turns everything off.

To update to a new release: /plugin, then Installed, vibesenpai, Update now; or claude plugin update vibesenpai@vibesenpai from a terminal.

How your data is used

Your transcripts are read where Claude Code keeps them and are never copied. Nothing is sent until you turn the model reads on, after seeing exactly what they would send. Then a redacted, length-capped copy of session text goes to Claude Haiku on your own Claude Code account (plan or API key), and the send stops if redaction left anything behind. Session text includes what the agent's tools returned, such as files it read and command output, so code from a session can be part of it. vibesenpai never opens your repository or writes your files; a fix is made by the agent, after you press Enter. How your data is used, with packet examples.

Contributing

See CONTRIBUTING.md and the contributor guide. Briefly:

git clone https://github.com/vibesenpai/vibesenpai && cd vibesenpai
npm install
npm test                       # the whole suite, no network
npm run cli -- insights        # run against your own sessions

There are no runtime dependencies, on purpose. npm run verify:corpus fails if any published count changes or anything leaks; it fetches the public corpus it needs.

License

MIT. See LICENSE.

Details

Marketplace
davepoon/buildwithclaude
Source
vibesenpai/vibesenpai
Author
vibesenpai contributors
Category
Monitoring
License
MIT
Keywords
observability, transcripts, session-review, claude-code-mod, hooks

More from davepoon/buildwithclaude

All 237

Similar to vibesenpai

More DevOps & Cloud