Require production-code verification

Author: halfaipg

src/agent/system-prompt.test.ts

@@ -46,6 +46,8 @@ describe("buildSystemPrompt", () => {
 		const out = buildSystemPrompt();
 		expect(out).toContain("# Verifying your work");
 		expect(out).toMatch(/never characterize unverified work as complete/i);
+		expect(out).toMatch(/execute the production code/i);
+		expect(out).toMatch(/audit the changed implementation against every requested behavior/i);
 	});
 
 	it("explains system-reminder semantics so the model knows they're from the runtime", () => {

src/agent/system-prompt.ts

@@ -91,6 +91,12 @@ export function buildSystemPrompt(opts: BuildSystemPromptOptions = {}): string {
 	lines.push(
 		"- If a check fails, fix the underlying cause rather than working around it (no --no-verify, no skipping tests, no commenting out asserts).",
 	);
+	lines.push(
+		"- Tests must execute the production code they claim to verify. Never copy or reimplement application logic inside a test just to produce a green check.",
+	);
+	lines.push(
+		"- After checks pass, audit the changed implementation against every requested behavior. A passing command is evidence, not proof that the tests covered the request; add a focused negative or interaction check when a requirement is otherwise untested.",
+	);
 	lines.push("");
 	lines.push("# Conversation conventions");
 	lines.push(

src/headless/reliable.ts

@@ -20,6 +20,8 @@ Rules:
 - Do not make verification commands pass by masking failures with fallbacks like "|| true", "|| echo", "|| :", or by appending "; echo $?".
 - If proving something is absent, use an unmasked negated check such as "! grep ..." in an "&&" chain, or write a tiny verify script that exits non-zero on failure.
 - If verification fails, fix the underlying issue and run verification again.
+- Verification must execute the changed production code. Do not duplicate implementation logic inside a test or verify only a test-side copy of the behavior.
+- After verification passes, inspect the changed implementation and audit every requested behavior against direct evidence. Add a focused negative, edge, or interaction check for requirements the existing command did not exercise.
 - For code or file changes, name the fresh passing verification command in the final answer.
 - For read-only or memory-only work, keep the task lifecycle auditable and state that no file-change verification was needed.`;