Skip to main content
MergeWatch treats disagreement as data. Every signal below is recorded against the finding it concerns and rolled up into the Accuracy view, so a reviewer telling the bot it was wrong is not just a comment thread — it is the input that stops the same false positive recurring.

Comment commands

/mergewatch reject is the strongest negative signal available. Unlike the disposition counters below — which are inferred from what happened to a finding — a reject is timestamped at the moment you issue it, and is windowed by that timestamp rather than by when the finding was last seen. It says this was wrong, not this went away.

Inline replies

Replying to an inline finding is a conversation, not just a comment. MergeWatch reads the reply, responds in the thread, and where the reply resolves the concern, resolves the review thread. Resolutions are persisted, which matters more than it sounds: a finding you have already argued down stays down. The convergence guard uses those resolutions so a later review does not re-raise a point that was already rebutted — the whack-a-mole problem where fixing one thing makes an old finding reappear.

Reactions

👍 / 👎 on the review summary comment feed the satisfaction signal in the engagement rollup. This is the lowest-friction feedback in the system — no command to remember, no thread to write. It measures whether the review as a whole was worth reading, which is a different question from whether any individual finding was correct.

What each signal feeds

Signals aggregate into two rollup blocks: Finding dispositions — per-finding outcomes, counted as: Engagement — whether humans interacted with the review: reviews delivered, re-reviews requested, /mergewatch reject commands, helpful votes, and NPS responses.
Rates distinguish “no signal” from a real zero. A rate is null — not 0 — when nothing landed in the window, so a quiet week reads as no data rather than as nothing was useful. Do not treat a blank as a bad score.

Why it is worth using

Feedback is not a courtesy to the tool. Dispute rates by agent and by category feed directly back into review behavior: an agent whose findings are consistently rejected in your repository gets down-weighted there. Teams that reject bad findings instead of ignoring them get measurably quieter reviews over time; teams that silently close them keep seeing the same thing. If a finding is wrong, say so — /mergewatch reject costs a few seconds and changes future reviews.

Next steps

Accuracy

Where these signals surface — dispute rates, the false-positive funnel, themes.

Review behavior

How disputes change what later reviews report.