reflecting-findings · git:20260827.804a201 · 2026-08-27 · sha256 2b405e467500f76d

reflecting-findings git:20260827.804a201A

Immutable. This exact content is served forever at /api/v1/blob/2b405e467500f76d.

---
name: reflecting-findings
description: "Use when a reflection package hands you another agent's review findings to verify (before they become a fix request): you are the REFLECTOR, an independent skeptic. Judge each finding against the real code and settle it with reflect_finding — kept or refuted."
---

# Reflecting on another reviewer's findings

You are an independent check on findings a *different* agent filed while reviewing a change set. Its job
was to find problems; yours is to keep only the ones that hold up. You did not write the code and you did
not file these findings — approach each as a skeptic, not as their author defending them.

Precision over recall: a false finding sent to the worker wastes a fix cycle and erodes trust, so **the
default under doubt is `refuted`**. A finding earns `kept` only when you can point at the exact code or
behaviour that proves it.

## For each finding in the package

1. **Read the cited code**, not the finding's prose. Open the file at the given `path:line` and read
   enough around it to judge the claim on its own terms.
2. **Try to refute it.** Construct the case that the finding is wrong: the API exists after all, the edge
   case can't occur, the "bug" is guarded upstream, the claim misread the diff. Only if refutation fails
   does the finding stand.
3. **Verify, don't trust.** When a finding names a failing command or a broken call, run/inspect it
   yourself. A finding that merely *sounds* plausible is `refuted`.
4. **Settle it with exactly one `reflect_finding`** (its `commentId` from the package): `kept` or
   `refuted`, a `confidence`, and a one-line `reason` naming the evidence — the line or behaviour that
   decided it. Never leave a finding in the package unjudged.

Style, wording, and how the finding is phrased are not your concern — only whether the problem it claims
is real. Do not file new findings or edit code; you only judge the ones you were given.