Shears Fix

Phase 1: Load Selected Problems

For each selected problem ID:

1. Read the corresponding .md file from ${SESSION_DIR}/check/
2. Read the corresponding .json file to get structured data
3. If a file is missing, warn and skip that problem
4. If the JSON already has "status": "resolved", skip it silently — it was already fixed in a previous run

Build a list of selected problems with their full details.

Phase 2: Group by File and Detect Conflicts

1. Group problems by their target .tex file.
2. Within each file group, sort by line_start.
3. Detect overlapping line ranges: if fix A's [line_start, line_end] overlaps with fix B's range:
   • Keep the one with higher confidence score
   • Warn about the skipped fix
4. Report: "N problems selected across M files. K conflicts detected (resolved by confidence)."

Phase 3: Backup

BACKUP_DIR=".shears/backups/$(date +%Y-%m-%d_%H-%M-%S)"
mkdir -p "${BACKUP_DIR}"

For each affected .tex file, copy it into the backup directory preserving the relative path:

mkdir -p "${BACKUP_DIR}/$(dirname <relative_path>)"
cp <file> "${BACKUP_DIR}/<relative_path>"

Also backup the current PDF:

cp <pdf_path> "${BACKUP_DIR}/"

Tell the user: "Backup created at ${BACKUP_DIR}"

Phase 4: Apply Fixes

Load all selected problem descriptions (the .md files) into your context. These contain the full fix details.

Apply fixes file by file. For each affected file:

1. Read the current file content.
2. Collect all fixes targeting this file, sorted by line_start descending (bottom-to-top).
3. For each fix, use the Edit tool to apply the change:
   • Use the original_text from the JSON as old_string
   • Use the suggested_fix from the JSON as new_string
   • If original_text is not found exactly, search nearby lines. If still not found, mark as skipped (see below) and warn.
4. After each fix attempt, update the problem's JSON sidecar at ${SESSION_DIR}/check/<id>.json with the outcome:
   • Successful: add/set "status": "resolved" and "resolved_at": "<ISO8601 UTC timestamp>"
   • Skipped (not found / conflict): add/set "status": "skipped" and "skipped_reason": "<short reason>"
   • Preserve all existing fields in the JSON. Use Read → modify → Write (Write overwrites the file).
5. After applying all fixes to a file, move to the next file.

Working bottom-to-top ensures earlier line numbers remain valid after each edit. The web UI polls /api/problems and will pick up the new status field within a couple of seconds — resolved problems get a green strikethrough in the list.

Phase 5: Recompile

Recompile the paper:

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/compile_latex.py "${MAIN_TEX}"

If compilation fails:

• Report the errors.
• Ask the user: "Compilation failed after applying fixes. Would you like to restore from the backup at ${BACKUP_DIR}?"
• If yes, restore all files from backup.
• If no, leave the modified files for manual debugging.

Phase 6: Ask Diff Mode

Present the user with the comparison options:

1. latexdiff — Generate a diff-annotated PDF with red (removed) / blue (added) markup. Requires latexdiff to be installed.
2. Side-by-side PDF — Produce both the old PDF (from backup) and the new PDF for manual comparison.
3. Tex diff — Generate a unified diff (patch format) of all modified .tex files.
4. Skip — No comparison needed.

Phase 7: Generate Comparison

Based on the user's choice:

python3 ${CLAUDE_PLUGIN_ROOT}/scripts/generate_diffs.py \
  --backup-dir "${BACKUP_DIR}" \
  --paper-root "${PAPER_ROOT}" \
  --main-tex "${MAIN_TEX}" \
  --mode <chosen_mode> \
  --output-dir ".shears/diffs/"

Phase 8: Report

Summarize:

• How many fixes were applied successfully
• How many were skipped (conflicts, not found, etc.)
• Backup location: ${BACKUP_DIR}
• Comparison artifacts location: .shears/diffs/ (if generated)
• If latexdiff PDF was generated, offer to open it.