context-audit

Run a context audit to measure every always-loaded file, then apply the ranked recommendations: delete exact redundancies, compress verbose instructions, and extract workflow-specific content into on-demand skills. The audit reports estimated token savings per finding.

The audit reads the full always-loaded stack and flags conflicting rules, numbers, or policies across files with quoted evidence from each side. Contradictions are called out first in the report since they cost correctness, not just tokens.

No, the audit is strictly report-only and never edits any audited file, including zero-risk deletions. It produces a recommendation list that you or a separate skill apply deliberately.

No, those files are bootstrap-rendered and direct edits are overwritten by the next render. Recommendations target the interview answer bank or templates, followed by a re-render of the affected file.

context-audit handles token and structure hygiene: what the stack costs per turn and where files repeat or contradict each other. soul-audit handles identity content: what SOUL.md or USER.md should say, sourced from the user's interview answers.

The default judging pass uses a single cheap utility-tier model for one cycle, costing a few cents. A full multi-model panel across three providers runs only when you explicitly request a full or multi-model audit.