Settings

Shibbo reads environment variables; there's no config file.

Variable	Used by	Default	Effect
SHIBBO_MODEL	check, rules, hooks	none	The model that writes, for model-ranked rules and phrase weights
SHIBBO_DISABLE	hooks	none	Tell ids to skip, comma-separated
SHIBBO_EDIT	hooks	on	off turns off the after-edit check
SHIBBO_STOP	hooks	block	warn notes reply hits without asking for a rewrite; off skips the reply check
SHIBBO_STOP_SCORE	hooks	50	Replies of 60+ words scoring under this, with 3+ hits, get one rewrite request
SHIBBO_JUDGE	reply hook	unset	A provider and model, such as openai:<model>, that reads each long reply for clarity
SHIBBO_LOG	hooks	unset	A file path; each hook check adds one JSON line to it
SHIBBO_LOG_TEXT	hooks	unset	1 adds the reply's text to its log line
NO_COLOR	check	unset	Plain output

Log each check

Set SHIBBO_LOG to a file path to measure the harness on your own work. The after-edit hook adds a line for each file it checks, and the reply hook adds a line for each reply of 60 words or more:

{"at":"2026-11-13T01:46:31.843Z","event":"edit","model":"gemini-4-argon","file":"/repo/docs/intro.md","words":412,"score":91,"hits":["x-matters","phrase:for anyone"]}
{"at":"2026-11-13T01:52:10.270Z","event":"reply","model":"gpt-6.1-sol","words":85,"score":15,"hits":["sycophantic-opener","contrastive-negation","delve"],"held":true}
{"at":"2026-11-13T02:30:23.999Z","event":"reply","model":null,"words":122,"score":100,"hits":[],"held":true,"judge":{"failed":["buried-point","report-buries-result"]}}

Field	Meaning
at	Time of the check, in UTC
event	edit or reply
model	The session's model as the harness reported it, or null
file	The edited file (edit lines only)
words, score	Word count and score of the whole file or reply
hits	Ids of the scored hits; a measured phrase reads phrase:<text>. For an edit, only hits in the new text
held	true when shibbo asked for a rewrite (reply lines only)
judge	With a judge set and a reply of 80+ words: {"failed": [ids]}, or null when the judge gave no verdict
judge_error	Why the judge gave no verdict, such as HTTP 404 and the start of the provider's answer
text	The reply, when SHIBBO_LOG_TEXT=1

Shibbo creates the file's folder if it's missing. A log that can't be written never stops a hook. The log holds file paths, and reply text if you ask for it, so keep it out of version control.

A judge for the final reply

The reply check reads wording. It can't see that a report retells the work and names its result last. Set SHIBBO_JUDGE and a small model answers the five clarity questions about each final reply.

Value	Key read from the environment
anthropic:<model>	ANTHROPIC_API_KEY
openai:<model>	OPENAI_API_KEY
openrouter:<model>	OPENROUTER_API_KEY

Shibbo ships no default model, so the judge stays off until you name one. It reads keys from the environment only and never opens a .env file. To use one, load it in the hook command:

set -a; [ -f .env ] && . ./.env; set +a; SHIBBO_JUDGE=openai:<model> npx shibbo hook stop

The judge runs when the wording check passes a reply of 80 words or more. Shibbo asks for one rewrite when the reply buries its result (report-buries-result), or when it fails two or more questions. The note quotes the words that show each fault, gives a fix, and offers a sentence to open with. SHIBBO_STOP=warn and off apply as they do to the wording check.

Each reply costs one request, sent straight to the provider you named. It holds a fixed prompt, the five questions and the reply text. Shibbo cuts the middle from a reply over 12,000 characters. It sends no file, path, session id or earlier message.

A 300-word reply comes to about 1,000 tokens in and 200 out. On a model priced at $1 and $5 per million tokens, that is about a fifth of a cent per reply.

A judge that fails never holds a reply. The call stops after 8 seconds. After a timeout, a missing key, a refused model id or an answer that isn't the JSON shibbo asked for, the hook lets the reply through and logs judge: null. A fail counts only when the judge quotes words that appear in the reply.

Turning off one tell

On the command line, --disable id,id. In hooks, SHIBBO_DISABLE. For a passage you're quoting, use an ignore block.

Model ids

Ids your harness reports work as they are. shibbo resolve <id> shows which measured model an id maps to: exact matches first, then known aliases, then the newest measured model in the same family.