Bug Echo

The bug you just fixed is rarely the only one of its kind. Ordinary review checks whether code is wrong on its own terms — it has no reason to connect one file's defect to the same mistake three files away. This skill does exactly that connection, and nothing else.

Off the linear pipeline (.specify/memory/constitution.md) — a reflex to run right after a fix lands. It never edits; to fix what it finds, hand the list to /resolve-review or just work it directly.

When to use this

• Right after fixing a bug that could plausibly be a class — a missing guard, a wrong argument form, an absent platform affordance, a copy-paste that drifted.
• Skip for a genuinely one-off bug (a typo, a wrong constant used in exactly one place).
• Not a general audit — for that use /code-review or /improve.

Procedure

1. Characterize the fixed bug as a pattern. Write one sentence naming the wrong shape, not the symptom: "a modal opened without a dismiss control", "a helper called with a positional arg that later became keyword-only", "a derived table read as evidence without a uniform-result check". If you can't state the shape, there's no pattern to echo — stop.

2. Derive search terms for the shape. The anti-pattern's tokens, the call that should be present but isn't, the construct that's misused. Plan several Grep queries — the bare name, the qualified form, the surrounding idiom — not one.

3. Sweep the codebase in one parallel batch. Grep/Glob for structurally similar sites. Cast wider than you think you need; step 4 is the filter.

4. Read every hit. A match is only a sibling if the same defect is actually present — local context (a guard one frame up, a different code path, a caller that can't hit it) often saves a look-alike. Grep counts are not findings.

5. Rank what survives. Per site: file:line, one line on why it's the same bug, and a tag:

   • confirmed — the defect is present, same mechanism, same consequence.
   • likely — same shape, couldn't fully rule the path in or out from reading.
   • needs-eyes — structurally similar, genuine doubt, worth a human glance. Note blast radius (user-facing? data? a hot path?) where it changes priority.

6. Present the ranked list and stop. Lead with the confirmed sites. If the sweep found nothing, say so plainly and say what you searched for — don't manufacture hits.

Constraints

• Report only. No edits, no new tasks, no review artifact.
• One pattern per run. If fixing the bug revealed a second unrelated class, note it and run again — don't blur two sweeps.
• Don't downgrade "I couldn't tell" to confirmed to look decisive, or up to needs-eyes to look thorough.
• "No siblings" is a valid, useful result — state the searches that produced it.

Done criteria

• The fixed bug is stated as a one-sentence pattern.
• Every structural match was read, not just counted.
• Output is a confidence-ranked file:line list (or a stated-empty result with the searches shown).