| name | review-digest |
| description | Turn a PR's AI-reviewer comment pile (codex / cursor / claude / copilot) into a clean, line-grouped digest using a LOCAL model — each bot's own headline + stated severity + verdict, extracted (not re-judged), grouped by file:line. Free, offline, no cloud tokens. Use when someone says "triage the comments on PR N", "what did the bots flag on PR N", "digest the review comments", "summarize the PR feedback", or invokes /review-digest N. A pre-pass that feeds a review coordinator — advisory only, it never gates a merge. |
| argument-hint | <pr-number> [-repo owner/name] — e.g. /review-digest 181 |
| user_invocable | true |
Review-comment digest (local reviewer-triage)
Collapses the four-bot comment flood on a PR into a scannable, line-grouped list, using a local
model. For each inline bot comment it extracts the bot's own headline + stated severity — it does
not re-judge whether the finding is valid (small local models comprehend dense findings poorly but
extract reliably). One model call per comment; grouped by file:line so same-line findings sit
together.
Prerequisites
- A local reviewer-extraction CLI that pulls a PR's bot comments (via
gh) and prints one row per
finding. Set REVTRIAGE_BIN to its path, or edit the command below. A reference implementation is
a small Go tool that calls Ollama over HTTP — swap in whatever you have.
- Ollama running with a small instruct model pulled (e.g.
qwen2.5:7b).
gh authenticated.
How to run
No cwd needed — it uses gh + Ollama over HTTP. Pass the PR number (and the repo if it isn't the
current one):
"$REVTRIAGE_BIN" <pr-number>
"$REVTRIAGE_BIN" -repo owner/name <pr-number>
"$REVTRIAGE_BIN" -json -repo owner/name <pr-number>
It prints one row per bot finding: line · severity · verdict · confidence · bot · headline.
-json emits the same findings as a JSON array with comment_id/file/url for joining back to
the source comments — the form a review coordinator consumes as its ingest pre-pass.
What to relay
- The digest as-is, then a one-line read: which findings cluster on the same locus, and which look
like the real must-fixes (trust
severity over verdict — verdict is the soft field).
Rules
- Trust
severity over verdict. Severity is read off the bot's own text; verdict is a softer
inference. unknown severity is normal for Copilot (it doesn't emit a severity badge) — don't
invent one.
- This is a pre-pass, not a gate. It reduces noise before human/coordinator review; it never
decides merge/no-merge.
- Prereqs: Ollama running with a small model pulled;
gh authenticated. First call cold-loads the
model (~30-60s); subsequent comments are a few seconds each.
- If a headline comes out as a long sentence, the bot led with prose instead of a title — the
severity + line are still reliable.
