AI agent skipped adversarial review step before smoke testing extension code

Problem

When authoring extension model code, the swamp-extension-model skill explicitly requires an adversarial self-review (step 6 in the workflow) before running unit tests or smoke tests. The instruction reads:

After writing or significantly modifying extension code, and before running unit tests or smoke tests, read references/adversarial-review.md and self-review against all applicable dimensions. Present findings to the user before proceeding to testing.

During a recent session authoring the @hivemq/honeycomb extension (adding createBurnAlert/deleteBurnAlert methods), the AI agent skipped this step entirely and proceeded directly from code authoring to smoke testing against the live Honeycomb API. The adversarial review was never performed and findings were never presented to the user.

Proposed Solution

Add a hard gate in the skill workflow that forces the agent to pause and explicitly present adversarial review findings to the user before any testing step is allowed. This could be:

• A checklist item the agent must mark complete (with findings) before the smoke-test step
• A required swamp extension review CLI command that the agent must run and whose output it must present
• A prominent WARNING block in the skill markdown that is harder to overlook under momentum

Alternatives

• Keep the current instruction but make it a numbered checklist item with a checkbox the agent must output explicitly (lower friction, may still be skipped)
• Add a pre-smoke-test check to the CI workflow that surfaces a reminder if no review comment was recorded

Context

The skip occurred because the session had high momentum (iterating rapidly against a live API), and the adversarial review step sits between code authoring and testing with no hard stop enforcing it. The agent later acknowledged the miss when asked directly.