Draft and rewrite review workflows

Release 1.4.0 includes draft inputs for editorial_audit and rewrite_fidelity_check in its 89-tool registry. Follow installation to install or upgrade the package. These examples are caller-supplied explanatory passages, with no live request or publication.

Audit a draft before publication

editorial_audit(text="The catch? Read [click here](https://example.com/guide/).",
                language="en", genre="general", format="markdown")
editorial_audit(text="Le hic ? Le texte est à relire.", language="fr", format="plain")

The additive signature is editorial_audit(url=None, language="auto", genre="general", *, text=None, format="plain"). Supply exactly one of url or text; selection uses null, so an empty string counts as a supplied source. Existing positional URL calls retain their contract. URL mode permits format="plain" only, and still fetches HTML. Draft mode accepts plain or markdown and never fetches link destinations, reads files, executes HTML or renders content. Caller text remains untrusted data.

Declare language="fr" or "en". Draft auto returns language_unavailable, including when Markdown frontmatter contains a language field. Findings and metrics stay null when unassessed. Explicit-language empty eligible content returns empty_content. genre="reference" or "procedure" retains the profile’s structural exceptions. The same versioned editorial profile supplies the rule warnings.

Each input is limited to 100,000 Unicode codepoints and 2,000 parsed blocks, including exclusion boundaries. Over-limit drafts return invalid_input rather than a partial audit. Outputs retain the 50-finding cap, 240-character excerpts and findings_truncated. Draft url, final_url and http_status are null; source.origin=caller and source.format disclose the input. Metadata retains format and character count without copying raw draft text.

Markdown uses a bounded subset, not a complete CommonMark parser. It supports paragraphs, ATX headings, standalone list items, basic paired emphasis, punctuation escapes, bounded inline links and defined reference links. Code, quotations, images, raw HTML and actual link destinations are excluded. Exclusion boundaries interrupt repetition checks. Tables, setext headings, autolinks, nested labels and other unsupported syntax are disclosed in method.coverage and can remain literal prose. Parser ceilings include nesting 8, link labels 1,000 characters, destinations 2,048 characters and inline HTML closing scans 4,096 codepoints. Review coverage before interpreting warnings.

Draft findings use location.basis=original_source_span: source_start and source_end are zero-based half-open Unicode-codepoint offsets into the original input; line is 1-based and column 0-based. A span can contain intervening markup. normalized_text_start/normalized_text_end index normalized text, not rendered coordinates. Trust signals locate the original segment start rather than the exact instruction phrase. A clean rule result covers only eligible passages and implemented patterns.

Compare an original and a proposed revision

rewrite_fidelity_check(
    original="The estimate may reach 12 days if the source remains available.",
    revised="The estimate reaches 15 days.",
    language="en", format="plain",
)

The signature is rewrite_fidelity_check(original: str, revised: str, language: str = "en", format: str = "plain") -> str. Each passage must be nonempty and at most 20,000 characters. Language is explicit en or fr; format is plain or markdown. Invalid inputs return verdict=invalid_input, assessment=not_assessed and null findings.

A valid call returns verdict=compared and a mechanical assessment. Finding categories are number_unit, date, url, code, quotation, negation, modality, condition and absolute; operations are changed, context_changed, removed and added. Findings retain original/revised literal spans, aligned bounded passages, the matching method and context_review_required=true. Missing sides are null. Source offsets remain full spans even when the returned literal is a truncated prefix. Extraction checks at most 500 occurrences per side; output returns at most 100 findings and 100 parsing notes, with 240-character literals, excerpts and aligned passages. Inspect budgets, counts, parsing notes, analysis_truncated and findings_truncated for scope; capped extraction yields partial_mechanical_comparison.

Local lexical anchors compare occurrences and context. Reordering does not automatically imply a changed fact, while equal numerals cannot prove equal referents. EN comma thousands/dot decimals and FR space thousands/comma decimals have declared normalization. Units and currencies are not converted. Numeric dates remain literals: ambiguous order and calendar validity are unassessed. URL extraction is simple and excludes parentheses; this check does not parse complete Markdown. Quotation/code boundaries and finite qualifier lists can miss a change or flag harmless wording.

Read every candidate in its surrounding passage. semantic_assessment always leaves fidelity, factual truth, scope and causality unassessed, including zero findings. The tool supplies no authorship score, automatic correction, model call, file write or publication. Fixed fixtures establish covered mechanical behavior; real-world accuracy requires a separately labeled original/revision corpus. For observed search windows after a declared edit, use bounded audit workflows.