[zlum]

Problem

desloppify autofix <tool>-violation never runs the tool for generic (tool-backed) fixers, yet reports success.

cmd_autofix calls fixer.fix(entries, dry_run=dry_run) without a path. make_generic_fixer treats a missing path the same as a dry run (if dry_run or not path:), so it returns one {file, line} entry per issue without running fix_cmd. The existing tests always pass path= directly, so they never exercise the CLI call.

Seen on a Kotlin project (ktlint, 6390 violations in ~100 files):

  Fixed 6390 Fix ktlint issues across 6390 files

  app/src/main/java/.../MainActivity.kt  →  fixed
  app/src/main/java/.../MainActivity.kt  →  fixed
  ... and 6360 more files

  Auto-resolved 0 issues in state
  Fixed 6390/6390 (0 skipped, 0 issues resolved)

Nothing was changed on disk. Related problems in the same path:

• Once the tool does run, success is len(entries) - len(remaining), and entries[:fixed_count] are reported as "the fixed ones", which is an arbitrary slice. If the re-check fails (the tool errors and detect returns []), every entry counts as fixed.
• The summary and the query payload (files_fixed, files_would_fix) use len(results) as a file count. Generic fixers return one result per issue, hence "across 6390 files".
• --dry-run prints an empty "Sample changes (before → after)" section for generic fixers. The preview only shows entries whose name is in removed, and generic results have neither. fix: handle generic fixers that return entries without 'removed' key #484 fixed the crash on this shape but not the empty preview.

Fix

• FixerConfig.takes_scan_path (default False). make_generic_fixer sets it, and cmd_autofix passes path= only to fixers that set it, so the TypeScript and Rust fixer signatures are unchanged. A generic fixer called without a path outside dry-run now raises instead of silently doing nothing.
• The generic fixer groups results by file: {file, fixed_count, summary}. After running fix_cmd, it re-runs the check through run_tool_result and counts fixes per file (before − after). If the re-check errors, it claims nothing and reports recheck_failed. The fix command's exit code is still not treated as failure, because formatters such as ktlint exit non-zero when unfixable violations remain.
• _result_item_count() is shared by the command, the cascade and the preview: len(removed), else fixed_count, else 1.
• The summary and the query count distinct files.
• For tool-backed results, the dry-run preview shows the flagged lines with the tool's message under "Sample flagged lines". The tool decides the rewrite, so no "after" exists to show.

Tests

• 8 new tests in tests/lang/common/test_generic_plugin_integrations.py and tests/commands/test_cmd_autofix.py. 7 of them fail on main; the eighth (fixers without the flag still get no path) guards against a regression and passes on both.
• Full suite: 5636 passed, 35 failed. The same 35 tests fail on unmodified main on Windows (review runner and tree-sitter import tests), so none are related.
• ruff check . --select E9,F63,F7,F82, mypy (configured files) and lint-imports all pass.

Not in scope

The ktlint JSON parser is #680 / #778, and #778 also adds a Kotlin-specific fixer. This PR fixes the shared generic path that every other tool-backed language still uses.

🤖 Generated with Claude Code