merge-coauthor

name	merge-coauthor
description	Read what a co-author or supervisor changed in the file they sent back — a tracked-changes .docx, a revised .tex, a returned .md — as a structured report rather than a wall of coloured text: which sections moved, which citation keys were added or removed, which numbers changed or appeared with no run behind them, which claims moved up the strength ladder, and every Word comment they left. Nothing is merged automatically. Use when a draft comes back from a co-author; use /import-manuscript to bring in a manuscript the kit has never seen, and /respond-review for a journal's decision letter.
disable-model-invocation	true
argument-hint	<returned-file.docx|.tex|.md|.pdf>

/merge-coauthor — the three changes a read-through misses

Word's track-changes view answers which words moved. That is the one question whose answer is already obvious, and it is what hides the three that are not:

• a citation key disappeared while the sentence it supported stayed
• a number changed, or a new one appeared that no recorded run produced
• a claim moved up the ladder — "our results suggest" became "we prove"

All three look like the paper, which is why they survive being read. They are extracted mechanically here, and the prose diff is reported as counts underneath them.

Gate

drafts/ must hold the version you sent. That is the comparison baseline — if the co-author is returning something that was never in the kit, use /import-manuscript instead.

Step 1: run the comparison

python scripts/coauthor_diff.py --returned inbox/paper-supervisor.docx

.docx and .tex convert through pandoc, .pdf through pymupdf4llm — the same converters /import-manuscript uses, never a hand-written parser. A missing converter exits BLOCKED (3) naming what to install, rather than degrading into a regex.

Step 2: read the flags first

flags:
  citation_removed        Results   @kaya2023deep
  claim_strengthened      Results   "we prove"  (was: predictive)
  number_not_in_run_register        0.91  (as written: "0.91")

Each one is "this changed, look at it" — never "this is wrong":

Flag	What to do
citation_removed	Read the sentence that kept the claim. If the claim stands, it needs a source; if it does not, it needs cutting.
claim_strengthened	Compare against the evidence: python scripts/claim_check.py. A co-author's stronger sentence is sometimes the right one — but it has to be earned in the same revision.
number_not_in_run_register	python scripts/experiment_db.py trace <number>. Either a run was never registered, or the number was typed. Both are worth knowing before submission.

Step 3: read the comments

Word comments are pulled out of the .docx itself, because pandoc drops them and they are frequently the only place the supervisor said why:

comments:
  Prof. Yılmaz (2026-08-15) — "Bu iddia için kaynak gerekiyor."

Step 4: apply, by hand

Nothing was written into drafts/. Work through the report section by section and apply what you accept — this is the part with a person behind it, and a tool that merged for you would have made those calls silently.

After applying:

python scripts/cite_audit.py                 # keys still resolve
python scripts/consistency_check.py          # narrated vs presented numbers
python scripts/claim_check.py                # claim strength vs evidence

What is deliberately not compared

Prose. Two honest edits of one paragraph share no wording, so a similarity score over them measures nothing but the length of the paragraph. What is compared is citation keys, quantities, claim strength and word counts — four things with an unambiguous before and after.

Whether a change is an improvement. That is a judgement about the argument, and it is yours and your co-author's.

Reference

• Script: python scripts/coauthor_diff.py --returned <file> [--drafts drafts] [--runs analysis/runs.jsonl]
• Reads: drafts/*.md, the returned file, analysis/runs.jsonl
• Writes: nothing
• Related: /import-manuscript (a manuscript the kit has not seen), /respond-review (a journal's letter), /revise (/critique's own findings), /translate (whose quantity comparison this reuses)

---

The dangerous edit is never the one you can see in red. It is the sentence that still reads like yours.