test: make evals judge outcomes and self-explain failures

evals/edit-models-precisely.eval.ts

@@ -146,13 +146,16 @@ it.for(trials)(
 			`On the "product" type, add a "size" dropdown with the options S, M, and L, and an "author" link that can only point to "author" documents.`,
 		);
 
-		expect(result).toHaveRun("prismic", ["field", "add", "select"]);
-		expect(result).toHaveRun("prismic", ["field", "add", "content-relationship"]);
+		// Both `field add content-relationship` and `field add link --allow document`
+		// model a constrained relationship; judge the resulting model, not the command.
+		expect(result).toHaveRun("prismic", ["field", "add"]);
 		const model = await readLocalCustomType(project, product.id);
 		expect(model.json.Main.size.type).toBe("Select");
 		expect((model.json.Main.size.config as { options: string[] }).options).toEqual(["S", "M", "L"]);
 		expect(model.json.Main.author.type).toBe("Link");
-		expect(JSON.stringify(model.json.Main.author.config)).toContain("author");
+		const authorConfig = model.json.Main.author.config as { select?: string };
+		expect(authorConfig.select).toBe("document");
+		expect(JSON.stringify(authorConfig)).toContain("author");
 	},
 );
 

evals/it.ts

@@ -122,7 +122,8 @@ expect.extend({
 				const wanted = [bin, ...positionals].join(" ");
 				if (pass) return `expected no command matching \`${wanted}\`, but one ran`;
 				const seen = result.commands.map((c) => `  ${c}`).join("\n") || "  (no commands ran)";
-				return `expected a command matching \`${wanted}\`, but saw:\n${seen}`;
+				const final = "result" in result.result ? result.result.result : "";
+				return `expected a command matching \`${wanted}\`, but saw:\n${seen}\n\nagent's final message:\n${final}`;
 			},
 		};
 	},

evals/sync-models.eval.ts

@@ -43,7 +43,12 @@ it.for(trials)(
 		await exec("git", ["init"]);
 		await exec("git", ["add", "-A"]);
 		await exec("git", ["commit", "-m", "Initial commit"]);
-		const article = buildCustomType({ id: "article", label: "Article" });
+		// A field so the type looks finished; the agent balks at pushing an empty type.
+		const article = buildCustomType({
+			id: "article",
+			label: "Article",
+			json: { Main: { title: { type: "Text", config: { label: "Title" } } } },
+		});
 		await writeLocalCustomType(project, article);
 
 		const result = await agent(