Rewrite workflow

Rewrite asks the installed CLI to generate, judge, detector-score, and select candidates. The desktop app supplies controls and renders the returned RewriteResult; it does not run any pipeline stage itself.

Input and controls

Control	UI range/default	CLI argument
text editor	non-empty	stdin, final -
detectors	empty, free text	optional global --detectors
candidates per chunk	1–32, default 8	--candidates
max words per chunk	50–2,000 by 50, default 400	--max-words-per-chunk
temperature	0.1–1.5 by 0.1, default 0.9	--temperature formatted to two decimals
minimum quality	0.5–1.0 by 0.05, default 0.90	global --minimum-quality, formatted to two decimals
trained policy path	empty	optional --policy

Empty text prevents Run with Paste the text to rewrite.

Exact invocation shape

skryba [--detectors NAMES] --minimum-quality 0.90 rewrite --json --candidates 8 --max-words-per-chunk 400 --temperature 0.90 [--policy PATH] -

The app always requests --json because the view needs the complete candidate evidence.

Successful result

The top panel shows text and provides Copy. Then every chunks[] item shows:

• 1-based chunk number;
• selected strategy and AI probability;
• whether source fallback was used;
• selected quality score and total candidate count;
• selected per-detector rows and cache labels;
• quality judge issues when non-empty.

The view does not render every losing candidate, but the decoded CLI result contains them. Selection is never repeated in Swift.

Source fallback

When used_source_fallback is true, the detail reads:

No candidate passed the quality judge; the source text was kept.

This is a display of the CLI's decision, not a desktop fallback.