Module github_text 

Posting gate shared by GitHub titles, descriptions and comments.

Structs

LanguageDecision

A model’s verdict on whether a title and a body’s prose are English.

Withheld

What the gate withheld. Categories only, never the text.

Enums

Reply

What the owner’s answer asks for.

Violation

Categories only: diagnostics must never repeat the offending secret.

Constants

ASK_NODE

Graph node of the question filed when the gate withholds a title or body.

ASK_SEAT

Seat recorded on that question.

NEUTRAL_BODY

Fixed fallback for a rejected description or comment.

NEUTRAL_TITLE

Fixed fallback for a rejected title.

POST_ORIGINAL

Choice: post the original text of every withheld field, unmodified, past the posting gate. Publishing cannot be taken back.

USE_FALLBACK

Choice: post the fixed neutral text for every withheld field.

USE_MY_TEXT

Choice: post the owner’s own replacement title (their latest message).

Functions

artifact_name

Artifact file name for the scrubbed full text of a withheld message.

artifact_text

What the artifact holds: the original text exactly as written and the text use fallback would post, for the withheld fields. It stays inside the run directory on the operator’s machine and is never sent anywhere.

category

Stable name of a violation category.

check

Pure checker. Language exemptions never exempt sensitive data.

check_with

check with an optional decision from judge_language.

deputy_brief

Brief for the deputy that serves this question (node ASK_NODE).

diagnose

Which part of the text each fired rule points at, as ready-made lines. Deterministic (heuristics only): the same text gives the same lines. Sensitive data yields field, line and kind only, never the value; a snippet is shown only when it passes the redaction rules untouched.

expired

Has the fixed asked_at + timeout deadline passed?

fallback_text

What use fallback posts, as (title, description): neutral text for a field withheld for language, the redacted text for one withheld only for sensitive data. The one place that decides it, for both the question’s CHANGED version and the posting itself (which still runs prepare_with over it, a no-op on already redacted text). title / body are the original texts; a field that is not withheld comes back redacted too.

fingerprint

Identity of a rejected text: FNV-1a over title and body, so one run asks at most once per text and the text itself is never stored.

judge_language

Ask [roles] language_judge whether title and body’s prose are English.

parse_decision

Read the judge’s reply: one JSON object with required bools, optionally in a code fence, else the last non-empty line (a wrapper may log before it).

prepare

Scrub local identity and pattern matches, then replace failing prose. Every intervention is recorded without copying the offending material.

prepare_with

prepare with the decision judge_language reached for this very text.

question_choices

The choices a question for w offers.

question_detail

Question body (English; it is also what a deputy reads). Names the rules that fired and shows, for each withheld field, the ORIGINAL text exactly as written next to the CHANGED text use fallback would post. The question lives only in the local UI, so the original is not redacted here; nothing of it goes to a notification, GitHub or the event log. rejected carries the categories of an owner-supplied title that failed the gate (a retry).

question_summary

One-line summary; the only line allowed to follow the configured language.

read_reply

Read the answer. use my text takes the latest operator message that is not newer than the answer; none means fallback.

shareable

May this text be shown to the owner? Only when the shared redaction rules leave it untouched.

vet_title

Vet an owner-supplied title with the same rules as a generated one. Err carries categories only. Sensitive data is rejected, never redacted into something the owner did not write.