Drop severity=note findings from PR posting by default #6

New issue

Closed

opened 2026-05-01 13:58:36 -07:00 by auto-review · 0 comments

auto-review commented

2026-05-01 13:58:36 -07:00

Collaborator

Surfaced from PR #4 (reviews 709–711, 16 noisy comments).

The note tier seems structurally useful as LLM reasoning scratchpad — letting the model externalise observations about the diff appears to make the review pass more thorough. But those notes are pure noise once posted to the PR ("💡 Note: Switching from find() to match_indices() ensures all occurrences are checked").

Proposed change: default AR_SEVERITY_FLOOR from note to warning. Notes stay in the schema (the LLM can still emit them) but the post step drops them before anything reaches Forgejo. Operators who want notes posted can opt back in via env.

Touches: crates/ar-review/src/post.rs (or wherever the floor filter lives), docs/OPERATIONS.md, and the env-var defaults in deploy/systemd/auto_review.env.example + Helm values.

Repro: any small fix-and-tests PR with current defaults — the bot will produce 5–6 LGTM-style notes per review cycle.

Surfaced from PR #4 (reviews 709–711, 16 noisy comments). The note tier seems structurally useful as LLM reasoning scratchpad — letting the model externalise observations about the diff appears to make the review pass more thorough. But those notes are pure noise once posted to the PR ("💡 **Note:** Switching from `find()` to `match_indices()` ensures all occurrences are checked"). **Proposed change:** default `AR_SEVERITY_FLOOR` from `note` to `warning`. Notes stay in the schema (the LLM can still emit them) but the post step drops them before anything reaches Forgejo. Operators who want notes posted can opt back in via env. Touches: `crates/ar-review/src/post.rs` (or wherever the floor filter lives), `docs/OPERATIONS.md`, and the env-var defaults in `deploy/systemd/auto_review.env.example` + Helm values. Repro: any small fix-and-tests PR with current defaults — the bot will produce 5–6 LGTM-style notes per review cycle.