Reference

Everything you can set, and every file the plugin keeps, in one place.

Commands

/vibesenpai:insights	opens the pane: findings, fixes with Apply, and the health of the history pass and the model reads
/vibesenpai:judge	on or off for the model reads; with no argument, their status
/vibesenpai:report	writes your history as one HTML file and opens it

/plugin disable vibesenpai stops everything, hooks included.

Settings

None are required. Apart from /vibesenpai:judge on|off, settings live in the config file or in the environment Claude Code runs in.

If a setting is given in more than one place, the environment variable wins over the config file, and the config file over the default.

setting	environment	config file key	default
transcript sources	VIBESENPAI_SOURCES, separated by , or :	sources (a list)	~/.claude/projects (under CLAUDE_CONFIG_DIR when that is set) and ~/.codex/sessions
the store’s location	VIBESENPAI_ARCHIVE	archive	$XDG_DATA_HOME/vibesenpai, else ~/.local/share/vibesenpai
free-check preset	VIBESENPAI_PRESET	preset	core
muted free checks	VIBESENPAI_MUTE, separated by , or :	mute (a list)	none
config file	VIBESENPAI_CONFIG		$XDG_CONFIG_HOME/vibesenpai/config.json, else ~/.config/vibesenpai/config.json

Only Claude Code transcripts are read: there is no parser for files under ~/.codex/sessions.

The config file is JSON. To turn on Repeated call:

{ "preset": "standard" }

To keep a free check from running:

{ "mute": ["edit_retry"] }

An unknown preset, or a mute name that is not a free check, is reported as a warning. It is never silently ignored.

Presets and mute

A preset decides which free checks run. It does not affect the model reads.

check	core (default)	standard	all
Edit retry (edit_retry)	yes	yes	yes
Repeated call (redundant_call)		yes	yes

all is the same as standard today. Cache rebuilds are not part of a preset; they are shown whenever one happened.

mute takes free-check names (edit_retry, redundant_call). A muted check is not computed, counted or mentioned.

Environment only

variable	effect
VIBESENPAI_DISABLE=1	every hook returns before doing anything. Nothing runs
VIBESENPAI_COLOR	1 forces colour in text output, 0 turns it off
NO_COLOR	turns colour off, unless VIBESENPAI_COLOR=1

Without either, colour is on only when the output is a terminal.

Model-read settings

Kept in judge-settings.json in the store. Until a setting has been written, the defaults apply. /vibesenpai:judge on and off set enabled; the others have no slash command.

key	default	what
enabled	true	whether the model reads run
model	haiku	the model passed to claude -p --model
dailyTokens	2000000	tokens, in and out, the automatic reads may use per local day
catchUp	200	past sessions read once, newest first
cadence	50	new turns before a live session is read again
maxBatchTokens	60000	a batch larger than this is split
concurrency	3	batches sent at once
lookback	true	a misread intent, a constraint violation or an other finding must cite where the expectation was set before the agent acted

Files

Everything lives in the store (the archive setting), ~/.local/share/vibesenpai by default. Transcripts are never copied into it; they are read where Claude Code keeps them, under ~/.claude/projects.

file	holds	if you delete it
store.db	model findings and their quotes; how far each session has been read; tokens and dollars spent; your answers; when each fix was requested and first seen loaded	you lose the most. Model findings come back only by reading again, which spends tokens again, and only for transcripts Claude Code still keeps. The 200-session catch-up starts over. Hidden findings reappear. Fix history comes back only from transcripts still kept
standing.json	your history from the free checks	nothing: the next history pass rebuilds it
finding-fixes.json	each model finding’s own fix	nothing: the next history pass rebuilds it
judge-settings.json	the model-read settings above	the defaults return, including enabled: true: model reads you turned off come back on
disclosed	that the one-time notice was shown	the notice is shown again
last-session-start	when the last session started	the next session-start summary shows no “new since your last session” count
jobs/	each background job’s last run, failures and hold	the record starts fresh; a hold on the model reads is released
reports/	the HTML reports you wrote	those files
index.json	an optional transcript index	nothing: it is safe to delete

Claude Code deletes old transcripts after a while, and nothing here is a copy of one. So there is no history beyond what Claude Code still keeps.

The JSON files and reports are written to a temporary name and then renamed, so a crash leaves the old version whole.