Suppress flaky / false-positive AI review comments

If you searched for "reduce false positives in AI code review", "suppress noisy AI review comments", or "stable AI review", this recipe is for you.

The problem

LLM-driven review is non-deterministic: the same diff can surface different comments on each run, and some comments contradict the actual implementation intent. When the same finding re-fires on every PR, reviewers lose trust and developers start ignoring the bot. Left unmanaged, that noise erodes the signal that made the review worth reading.

How River Review addresses it

River Review splits false-positive handling into two layers:

• Deterministic detectors (security-basic, logging-observability, test-existence, coverage-gap) are guarded by canary fixtures — a corpus of known false-positive patterns. Once a false positive is fixed, a regression test keeps it fixed. See the detector evaluation report.
• Semantic noise and intent mismatches are handled in the midstream phase by the suppression-feedback skill.

suppression-feedback does not silently delete findings. It classifies a finding so a reviewer (human or AI) can decide between suppressing and fixing it, then guides the river suppression add CLI that records the decision in Riverbed Memory:

• false_positive — records a comment that contradicts the implementation intent.
• accepted_risk — records a finding you knowingly keep after weighing the cost.
• duplicate — points at an existing fingerprint instead of repeating the comment.
• major / critical findings are not auto-suppressed unless the classification is accepted_risk with a rationale (the HIGH_SEVERITY guard); lower-severity minor / info noise is auto-suppressed for any classification.

Because the decision lives in Riverbed Memory, the next PR's review can recall it and stop re-firing the same noise.

Try it

There is no dedicated demo — the workflow runs through Riverbed Memory suppression. To silence a comment you have judged a false positive:

1. Run a review and note the fingerprint of the comment you want to suppress.
2. Decide whether it is a false positive or an accepted risk.
3. Record it with river suppression add:

river suppression add \
  --fingerprint <fp> \
  --feedback false_positive \
  --rationale "one or two sentences on why this is a false positive"

For the full storage flow into .river/memory/index.json, see Use Riverbed Memory. For the design rationale, see Riverbed Memory.

What it does NOT do

• It does not auto-generate suppression entries. Writing to the CLI and to Riverbed Memory requires human approval.
• It is human-in-the-loop, not auto-merge — River Review never merges around a finding on its own; a human makes the final call.
• It is not a replacement for static analysis or your custom linters. The deterministic layer is guarded by canary fixtures; suppression-feedback focuses only on the semantic judgment those tools cannot make.
• It does not blanket-suppress major / critical findings. Auto-suppression is allowed only when both accepted_risk and a rationale are present.