| name | verify |
| description | Manually verify review findings one by one by running code, writing adhoc scripts, and comparing actual behavior instead of implementing blindly. |
| allowed-tools | Bash, Glob, Grep, Read, Write, Edit, Agent, AskUserQuestion |
| disable-model-invocation | true |
Verify review findings from the current conversation by actually testing each one.
Arguments: $ARGUMENTS
Context
The conversation contains review findings (e.g. from /review, /consult, or a manual code review). These are claims about bugs, behavioral differences, missing features, or other issues. Your job is to verify each finding empirically before anyone implements fixes.
Phase 1: Extract Findings
- Look at the conversation context for a list of review findings, issues, or behavioral differences
- If
$ARGUMENTS specifies a subset (e.g. "items 1-5", "critical only"), filter accordingly
- Number each finding for tracking
Phase 2: Verify Each Finding
Mode: If $ARGUMENTS contains --parallel, verify all findings concurrently using parallel Agent calls (one agent per finding). Otherwise, go through findings one at a time, in order.
For each finding:
2a. Understand the Claim
Read the relevant source files to understand what the finding is actually claiming.
2b. Test It
Choose the most appropriate verification method:
- Read and compare code: When the claim is about code structure, missing features, or API differences — read both sides and confirm
- Run the program: When feasible, run the actual program with inputs that would expose the issue
- Write an adhoc script: When the claim involves runtime behavior (env var handling, string formatting, edge cases), write a small throwaway script that isolates and tests the specific behavior
- Put scripts in
/tmp/verify-*.{sh,ts,py,rs} or similar
- Keep them minimal — test exactly the claim, nothing more
- Check documentation: When the claim is about library behavior (e.g. "SDK provides automatic retries"), verify against actual docs or source
- Compare outputs: When the claim is about differing output formats, run both implementations and diff
2c. Report Verdict
For each finding, report:
- CONFIRMED: The finding is accurate and is a real issue
- NOT AN ISSUE: The finding is technically correct but doesn't matter in practice (explain why)
- INCORRECT: The finding is wrong (show evidence)
- UNABLE TO VERIFY: Cannot test this without external resources/setup (explain what's needed)
Show the evidence (command output, code snippets, script results) that supports your verdict.
2d. Continue
After reporting the verdict for one finding, immediately proceed to the next. Do NOT ask the user between findings — verify all of them in one pass.
Phase 3: Summary
After all findings are verified, present a summary table:
# Verified | # Not an Issue | # Incorrect | # Unable to Verify
List the confirmed issues and ask the user which ones they want to fix.
Guidelines
- Be empirical, not theoretical — run code, don't just read it
- Keep adhoc scripts minimal and focused
- Clean up temp scripts when done
- If a finding is ambiguous, err on the side of testing it
- When comparing two implementations, test BOTH — don't assume one is correct
- With
--parallel, launch one Agent per finding simultaneously — each agent verifies independently and returns its verdict. Collect all results before producing the summary.