Enforce mandatory context loading before AI code changes

the-honoured-one eliminates the most common AI coding failure - confident wrongness, where the AI proposes or implements something based on how it assumes the code is structured rather than how it actually is. It enforces four sequential phases before any action: a Context Audit declaring every relevant file and which of them have actually been read this session (identifying "blind spots"); a Mandatory Read Pass that reads every blind-spot file for real, transitively pulling in anything those files import; an Orientation Statement summarizing the actual architecture, what the task touches, and the conventions observed, plus any remaining unknowns; and a Confidence Gate that blocks proceeding until every unknown is resolved - by asking the user, reading another file, or explicitly stating an assumption and getting user confirmation.

Use it when modifying multiple files in an existing codebase, designing or changing a system component, adding a feature that integrates with existing code, debugging a component the AI hasn't yet read, or any task where the AI would otherwise need to assume how existing code is structured. It explicitly should not be used for isolated single-file tasks where the file has already been read - the overhead isn't warranted there.

The Context Audit and Orientation Statement are both structured, printed blocks - the audit lists relevant files with a read/unread status per file and ends with "Cannot proceed - reading blind spots now" if any are unread; the orientation statement lists files read, 2-3 sentences on the actual architecture, what the task touches, observed conventions to follow, and remaining unknowns (or "None, ready to proceed"). A four-question self-check gates every action: has every touched file been read, is the relevant pattern understood from reading rather than assumption, are the actually-observed conventions being followed, and are there any remaining blind spots.

THE HONOURED ONE - CONTEXT AUDIT
Relevant files identified: [...]
Files read this session: [...]
Unread but relevant (blind spots): [...]
Cannot proceed - reading blind spots now.

Its hard rules explicitly forbid "I assume this file does X," skipping a file because its name looks obvious (a file called utils.ts can contain anything), calling or importing code from unread files, and relying on general pattern familiarity as a substitute for reading this codebase's actual implementation of that pattern. It pairs with @brainstorming (used earlier, to figure out what should be built) and is explicitly distinguished from @not-a-vibe-coder, which is for entirely new projects rather than existing ones.

Developers who want an AI to earn the right to act on an existing codebase by demonstrably reading it first, rather than integrating with a module it never opened, using naming conventions inconsistent with the rest of the code, or calling functions that don't exist because it assumed they'd be there - at the cost of higher token usage and a slower initial response while the mandatory read pass runs.