[ancongui]

The bot lost the current human question before drafting its status reply, then repeated the product brief and internal acceptance terminology. Carry the exact question to the writer and reviewer as untrusted quoted context, and let the AI answer it with useful headings and bullets. The editor can request one correction for confusing language or poor structure.

Recorded final-check commands and output can now supply exact code-fenced excerpts selected by reference. The model cannot invent or rewrite their bytes. Current, retained and unknown provenance stay distinct; sensitive-looking fragments are withheld, local paths and destinations are removed, and credential/current-authority checks remain in place. A known configuration-admission failure gains a display-only explanation without becoming recovery authority.

Instructions use three additive YAML/Jinja prompt versions. All341 previous prompt entries and payloads are preserved. Architecture, model-profile and repository workflow documentation explain the behavior.

Validation:205 focused communication, question-flow and prompt regressions pass, including20 new cases. Full typing passes410 source files; Ruff and formatting pass. Strict MkDocs and73-page/10,330-link validation pass. Independent privacy/provenance review reproduced and checked the corrected cases. Controlled model responses are test evidence; the new wording still needs the qualified installed runtime and a real GitHub round trip before merge. The previous live bot response was explicitly rejected for its prose quality.

This topic is consolidated into PR75. Its exact branch head is an ancestor of PR75; no implementation or history is discarded. Review and delivery continue there. Closing this duplicate does not claim its functionality is released or that the CRM journey is complete.