AI workflows

Start with AI integration setup. The adapter is included with SaveTether, but the core review panel works without AI, Python, or Toolset Registry.

Compare a declared release

1. Ask the user for the exact local declaration JSON and, when comparing releases, the exact previous manifest. Use only those paths; SaveTether does not search a project or open .sav files.
2. For a comparison, call SaveTetherTools.review_with_baseline(declarations_path, previous_manifest_path, max_records=10000). For an initial declaration set with no previous release, call review_initial_declarations(declarations_path, max_records=10000) instead. A returned “accepted” response means the review started, not that it passed.
3. Poll get_review_status() until the review is terminal. Inspect coverage and validation reasons, then read get_review_page("changes", 0, 10), get_review_page("mappings", 0, 10), or get_review_page("findings", 0, 10). Continue from the returned offset when more rows exist.

The included valid example compares A/B as conserved, C as removed, and D as added. The cycle example rejects the A → B → A mapping cycle. These results describe declared identifiers only; they do not establish that an existing save can load.

Write only after approval

Export only when the user explicitly asks and supplies the exact existing output folder: export_report(output_directory, allow_partial=false). Partial or cancelled coverage may be exported only after explaining the limits and obtaining explicit approval to use allow_partial=true.

Capture a new release baseline only after complete current coverage and no rejected mappings: capture_release(output_directory). This writes an identity manifest, not a save-game migration. Poll status to completion and verify the returned output path; never overwrite earlier release history.

Example request: “Compare these two declaration files and show additions, removals, mappings, and coverage. Do not export a report or capture a release until I approve.”