feat(service): toggle subagents with HANSI_SUBAGENTS

README.md

@@ -131,6 +131,7 @@ The service validates its environment on start (`apps/service/src/env.ts`) and e
 | `HANSI_VERIFY_MODEL`          | review model  | The verifier's model, as `provider:model` with the review model's provider.                       |
 | `HANSI_REVIEW_REASONING`      | provider's    | The reviewer's reasoning effort: `none`, `minimal`, `low`, `medium`, `high`, `xhigh` or `max`.    |
 | `HANSI_VERIFY_REASONING`      | review effort | The verifier's reasoning effort.                                                                  |
+| `HANSI_SUBAGENTS`             | `enabled`     | `disabled` stops the reviewer from handing parts of a review to subagents.                        |
 | `HANSI_AI_SECRET`             |               | The provider's key. Required except for `openai-compatible`. See [Models](#models).               |
 | `HANSI_AI_REGION`             |               | AWS region. Required for `bedrock`.                                                               |
 | `HANSI_AI_BASE_URL`           | provider's    | The provider's API, instead of its default. Required for `openai-compatible`.                     |

apps/service/src/env.test.ts

@@ -19,6 +19,12 @@ describe('parseEnv', () => {
 		expect(env.HANSI_JOBS_PORT).toBe(8081);
 		expect(env.HANSI_CONSUMER_COMMAND).toBeUndefined();
 		expect(env.HANSI_SANDBOX_RUNTIME_CLASS).toBe('gvisor');
+		expect(env.HANSI_SUBAGENTS).toBe('enabled');
+	});
+
+	test('turns subagents off, and takes enabled or disabled only', () => {
+		expect(parseEnv({ ...required, HANSI_SUBAGENTS: 'disabled' }).HANSI_SUBAGENTS).toBe('disabled');
+		expect(() => parseEnv({ ...required, HANSI_SUBAGENTS: 'off' })).toThrow(/HANSI_SUBAGENTS/);
 	});
 
 	test('takes a known sandbox isolation tier only', () => {

apps/service/src/env.ts

@@ -48,6 +48,11 @@ const schema = z
 		 * Appwrite takes over model choice once it serves keys and models.
 		 */
 		HANSI_REVIEW_MODEL: modelId,
+		/**
+		 * Whether the reviewer may hand parts of a review to subagents that run in parallel. Off, one
+		 * agent reviews the whole change: slower on large pull requests, but fewer model calls.
+		 */
+		HANSI_SUBAGENTS: z.enum(['enabled', 'disabled']).default('enabled'),
 		/** The verifier's model. Defaults to the reviewer's, and uses the same provider. */
 		HANSI_VERIFY_MODEL: modelId.optional(),
 		/** Reasoning effort per model. Unset leaves the provider's default. */

apps/service/src/index.ts

@@ -46,7 +46,12 @@ await clearWorkdir(undefined);
 const models = createModels(env);
 // Without a verify model the engine verifies with the review model; measured as verification.
 const verify = models.verify ?? models.review;
-const context = { github, sandboxes, results: sink };
+const context = {
+	github,
+	sandboxes,
+	results: sink,
+	subagents: env.HANSI_SUBAGENTS === 'enabled'
+};
 const handlers = createHandlers(
 	{
 		...context,