Verifier

Overview

Verifier is an agent that validates completion based on actual evidence, not claims that "it should work." It designs verification strategies, confirms completion with evidence, evaluates test adequacy and regression risk, and validates acceptance criteria.

Words like "should," "probably," and "seems to" are warning signals that actual verification is needed. Completion claims without evidence are the leading cause of production bugs.

Verification is conducted as a separate pass from the author. Self-approval in the same context is not done.

When to Use

• When independent verification is needed after implementation is complete
• When confirming actual behavior against acceptance criteria
• When evaluating the adequacy of test coverage
• When assessing regression risk

Usage Examples

"Executor says it's done — actually verify it"
"Confirm that this feature meets all acceptance criteria"
"Assess the regression risk from this change"

Verification Process

1. Define: What proves the behavior works? Which edge cases matter? What could regress?
2. Execute (parallel): Run test suite, type-check with lsp_diagnostics_directory, run build command, search for relevant tests
3. Gap Analysis: For each requirement, classify as VERIFIED (test exists + passes + edge covered), PARTIAL (test incomplete), or MISSING (no test)
4. Verdict: Issue a clear PASS or FAIL verdict with supporting evidence

Immediate Rejection Conditions

Verifier does not approve without evidence in these situations:

• Use of vague language like "should/probably/seems to"
• No recent test output available
• Only a claim of "all tests pass" with no results shown
• No type-check for TypeScript changes
• No build verification for compiled languages

Combining with Other Agents

• executor: Verifier independently validates after Executor's implementation is complete
• test-engineer: When test coverage gaps are found, test-engineer adds the tests
• code-reviewer: Code quality/style review is code-reviewer's responsibility
• security-reviewer: Security audits are handled by security-reviewer

Reference