Skip to main content
← Back to list
01Issue
FeatureShippedSwamp CLI
Assigneesstack72

Relationships

#526 swamp-extension adversarial review skill needs mandatory mechanical verification checklist

Opened by webframp · 6/1/2026· Shipped 6/3/2026

Problem

The adversarial review reference in the swamp-extension skill provides excellent dimensional guidance (credentials, error handling, idempotency, etc.) but lacks a mandatory mechanical verification step that catches the most common class of bugs AI agents produce.

The Recurring Failure Pattern

When an AI agent writes extension code and then self-reviews using the adversarial review dimensions, it consistently misses:

  1. Schema/write mismatches: schema declares a field but the write stores incompatible data (e.g., schema says username but API returns assignee.login)
  2. Missing truncation fields: method paginates with a cap but output schema has no truncated field
  3. Placeholder values left in place: schema declares byUser with full structure but implementation writes byUser: [] and never implements it
  4. Wrong resource spec: method writes data to a spec whose schema doesn't match

These are NOT subtle logic errors. They are mechanical mismatches between declared schemas and actual writes.

Proposed Enhancement

Add a Mandatory Mechanical Checks section to the adversarial review reference that agents must execute BEFORE presenting findings:

  1. Schema conformance: For every writeResource call, enumerate every field in the spec schema. Verify each field appears in the data object with a non-hardcoded value.
  2. Truncation honesty: For every pagination loop, verify the output schema has truncated: boolean AND the write sets it.
  3. Instance name uniqueness: Verify no two methods write incompatible data to the same spec + instance.
  4. Schema field coverage: Read the schema. Read the write. Do they match 1:1?

Why This Matters

The CI adversarial review catches these every time because it reads every line equally. The agent self-review has confirmation bias — it reviews interesting code deeply and skims obvious code. The mechanical checklist forces equal attention on every method.

Evidence

This pattern occurred across multiple extensions: aws/inventory (instance name collision), aws/adopt (incomplete setupCommand), github-copilot-governance (SeatsSchema mismatch, byUser always empty, wrong spec, missing truncated). Each time the local adversarial review passed CLEAR and CI caught it.

02Bog Flow
OPENTRIAGEDIN PROGRESSSHIPPED+ 1 MOREASSIGNED+ 2 MOREREVIEW+ 3 MOREPR_MERGED+ 1 MORECONTRIBUTOR_NOTIFIED

Shipped

6/3/2026, 4:11:22 PM

Click a lifecycle step above to view its details.

03Sludge Pulse
stack72 assigned stack726/3/2026, 2:19:04 PM
Editable. Press Enter to edit.

stack72 commented 6/3/2026, 4:11:30 PM

Thanks @webframp for reporting this! The fix has been merged and a release is on its way. We appreciate your contribution to swamp.

Sign in to post a ripple.