@electric-ax/agents-server-conformance-tests 0.1.9 → 0.1.11

package/dist/index.cjs

@@ -328,11 +328,8 @@ var ElectricAgentsScenario = class {
 		});
 		return this;
 	}
-	readStream(stream) {
-		this.steps.push({
-			kind: `readStream`,
-			stream
-		});
+	readStream() {
+		this.steps.push({ kind: `readStream` });
 		return this;
 	}
 	list(filter) {
@@ -671,7 +668,6 @@ async function executeStep(ctx, step) {
 			(0, vitest.expect)(entity, `Webhook payload must contain entity context`).toBeDefined();
 			(0, vitest.expect)(entity.url).toBeTruthy();
 			(0, vitest.expect)(entity.streams.main).toBeTruthy();
-			(0, vitest.expect)(entity.streams.error).toBeTruthy();
 			if (step.checks?.url) (0, vitest.expect)(entity.url).toBe(step.checks.url);
 			if (step.checks?.type) (0, vitest.expect)(entity.type).toBe(step.checks.type);
 			if (step.checks?.status) (0, vitest.expect)(entity.status).toBe(step.checks.status);
@@ -712,7 +708,7 @@ async function executeStep(ctx, step) {
 		}
 		case `readStream`: {
 			if (!ctx.currentEntityStreams) throw new Error(`No current entity streams`);
-			const streamPath = step.stream === `error` ? ctx.currentEntityStreams.error : ctx.currentEntityStreams.main;
+			const streamPath = ctx.currentEntityStreams.main;
 			const res = await fetch(appendPathToUrl(ctx.baseUrl, `${streamPath}?offset=0000000000000000_0000000000000000`));
 			if (res.status === 200) {
 				const text = await res.text();
@@ -1207,14 +1203,11 @@ function checkEntityContextOnWebhook(history) {
 	}
 }
 /**
-* Spec S5 — Structural: stream paths must match {entity.url}/main and {entity.url}/error.
+* Spec S5 — Structural: stream path must match {entity.url}/main.
 * Soundness: Sound | Completeness: Complete (within trace)
 */
 function checkStreamPathsMatchEntityUrl(history) {
-	for (const event of history) if (event.type === `entity_spawned`) {
-		(0, vitest.expect)(event.streams.main).toBe(`${event.entityUrl}/main`);
-		(0, vitest.expect)(event.streams.error).toBe(`${event.entityUrl}/error`);
-	}
+	for (const event of history) if (event.type === `entity_spawned`) (0, vitest.expect)(event.streams.main).toBe(`${event.entityUrl}/main`);
 }
 /**
 * Spec S4 — Safety: entity status transitions must be valid.
@@ -1276,8 +1269,8 @@ function checkAdditiveSchemaEvolution(history) {
 }
 /**
 * Spec R4 — Registry stream consistency: every entity_spawned must refer to an
-* entity that has not already been killed, and stream URLs must follow the
-* convention {entity.url}/main and {entity.url}/error.
+* entity that has not already been killed, and stream URL must follow the
+* convention {entity.url}/main.
 * Soundness: Sound | Completeness: Complete (within trace)
 */
 function checkRegistryStreamConsistency(history) {
@@ -1287,7 +1280,6 @@ function checkRegistryStreamConsistency(history) {
 		if (event.type === `entity_spawned`) {
 			(0, vitest.expect)(!killedEntities.has(event.entityUrl), `Registry consistency: entity_spawned for ${event.entityUrl} but entity was already killed`).toBe(true);
 			(0, vitest.expect)(event.streams.main, `Registry consistency: entity ${event.entityUrl} streams.main must be ${event.entityUrl}/main`).toBe(`${event.entityUrl}/main`);
-			(0, vitest.expect)(event.streams.error, `Registry consistency: entity ${event.entityUrl} streams.error must be ${event.entityUrl}/error`).toBe(`${event.entityUrl}/error`);
 		}
 	}
 }
@@ -1832,8 +1824,6 @@ function runElectricAgentsConformanceTests(config) {
 		}).spawn(`spawn-test-agent`, `entity-1`).expectStatus(`running`).custom(async (ctx) => {
 			const mainRes = await fetch(appendPathToUrl(ctx.baseUrl, ctx.currentEntityStreams.main), { method: `HEAD` });
 			(0, vitest.expect)(mainRes.status).toBe(200);
-			const errorRes = await fetch(appendPathToUrl(ctx.baseUrl, ctx.currentEntityStreams.error), { method: `HEAD` });
-			(0, vitest.expect)(errorRes.status).toBe(200);
 		}).run());
 		(0, vitest.test)(`spawn at unregistered type returns UNKNOWN_ENTITY_TYPE`, () => electricAgents(config.baseUrl).custom(async (ctx) => {
 			const res = await fetch(appendPathToUrl(ctx.baseUrl, routeControlPlanePath(`/unregistered/entity-1`)), {
@@ -2904,25 +2894,51 @@ function runElectricAgentsConformanceTests(config) {
 			const block = result.content[0];
 			return block?.type === `text` && block.text ? block.text : ``;
 		}
+		async function makeSandbox(workingDirectory) {
+			const { unrestrictedSandbox } = await import(`../../agents-runtime/src/sandbox/unrestricted`);
+			return unrestrictedSandbox({ workingDirectory });
+		}
 		(0, vitest.test)(`bash tool captures stdout and stderr`, async () => {
 			const { createBashTool } = await import(`../../agents-runtime/src/tools`);
-			const tool = createBashTool(`/tmp`);
-			const result = await tool.execute(`test-tc`, { command: `echo "hello" && echo "error" >&2` });
-			(0, vitest.expect)(firstText(result)).toContain(`hello`);
-			(0, vitest.expect)(firstText(result)).toContain(`error`);
-			(0, vitest.expect)(result.details.exitCode).toBe(0);
+			const sandbox = await makeSandbox(`/tmp`);
+			try {
+				const tool = createBashTool(sandbox);
+				const result = await tool.execute(`test-tc`, { command: `echo "hello" && echo "error" >&2` });
+				(0, vitest.expect)(firstText(result)).toContain(`hello`);
+				(0, vitest.expect)(firstText(result)).toContain(`error`);
+				(0, vitest.expect)(result.details.exitCode).toBe(0);
+			} finally {
+				await sandbox.dispose();
+			}
 		});
 		(0, vitest.test)(`bash tool enforces timeout`, async () => {
 			const { createBashTool } = await import(`../../agents-runtime/src/tools`);
-			const tool = createBashTool(`/tmp`);
-			const result = await tool.execute(`test-tc`, { command: `sleep 60` });
-			(0, vitest.expect)(result.details.timedOut).toBe(true);
+			const sandbox = await makeSandbox(`/tmp`);
+			try {
+				const tool = createBashTool(sandbox);
+				const result = await tool.execute(`test-tc`, { command: `sleep 60` });
+				(0, vitest.expect)(result.details.timedOut).toBe(true);
+			} finally {
+				await sandbox.dispose();
+			}
 		}, 35e3);
 		(0, vitest.test)(`read_file rejects paths outside working directory`, async () => {
 			const { createReadFileTool } = await import(`../../agents-runtime/src/tools`);
-			const tool = createReadFileTool(`/tmp/test-workdir`);
-			const result = await tool.execute(`test-tc`, { path: `../../etc/passwd` });
-			(0, vitest.expect)(firstText(result)).toContain(`outside the working directory`);
+			const fs = await import(`node:fs/promises`);
+			const dir = `/tmp/test-workdir-${Date.now()}`;
+			await fs.mkdir(dir, { recursive: true });
+			const sandbox = await makeSandbox(dir);
+			try {
+				const tool = createReadFileTool(sandbox);
+				const result = await tool.execute(`test-tc`, { path: `../../etc/passwd` });
+				(0, vitest.expect)(firstText(result)).toContain(`outside the working directory`);
+			} finally {
+				await sandbox.dispose();
+				await fs.rm(dir, {
+					recursive: true,
+					force: true
+				});
+			}
 		});
 		(0, vitest.test)(`read_file rejects binary files`, async () => {
 			const { createReadFileTool } = await import(`../../agents-runtime/src/tools`);
@@ -2937,10 +2953,15 @@ function runElectricAgentsConformanceTests(config) {
 				2,
 				255
 			]));
-			const tool = createReadFileTool(dir);
-			const result = await tool.execute(`test-tc`, { path: `test.bin` });
-			(0, vitest.expect)(firstText(result)).toContain(`binary file`);
-			await fs.rm(dir, { recursive: true });
+			const sandbox = await makeSandbox(dir);
+			try {
+				const tool = createReadFileTool(sandbox);
+				const result = await tool.execute(`test-tc`, { path: `test.bin` });
+				(0, vitest.expect)(firstText(result)).toContain(`binary file`);
+			} finally {
+				await sandbox.dispose();
+				await fs.rm(dir, { recursive: true });
+			}
 		});
 		(0, vitest.test)(`read_file rejects oversized files`, async () => {
 			const { createReadFileTool } = await import(`../../agents-runtime/src/tools`);
@@ -2950,10 +2971,15 @@ function runElectricAgentsConformanceTests(config) {
 			await fs.mkdir(dir, { recursive: true });
 			const bigPath = path.join(dir, `big.txt`);
 			await fs.writeFile(bigPath, `x`.repeat(600 * 1024));
-			const tool = createReadFileTool(dir);
-			const result = await tool.execute(`test-tc`, { path: `big.txt` });
-			(0, vitest.expect)(firstText(result)).toContain(`too large`);
-			await fs.rm(dir, { recursive: true });
+			const sandbox = await makeSandbox(dir);
+			try {
+				const tool = createReadFileTool(sandbox);
+				const result = await tool.execute(`test-tc`, { path: `big.txt` });
+				(0, vitest.expect)(firstText(result)).toContain(`too large`);
+			} finally {
+				await sandbox.dispose();
+				await fs.rm(dir, { recursive: true });
+			}
 		});
 		(0, vitest.test)(`web_search tool has correct interface`, async () => {
 			const { braveSearchTool } = await import(`../../agents-runtime/src/tools`);
@@ -2961,9 +2987,15 @@ function runElectricAgentsConformanceTests(config) {
 			(0, vitest.expect)(typeof braveSearchTool.execute).toBe(`function`);
 		});
 		(0, vitest.test)(`fetch_url tool has correct interface`, async () => {
-			const { fetchUrlTool } = await import(`../../agents-runtime/src/tools`);
-			(0, vitest.expect)(fetchUrlTool.name).toBe(`fetch_url`);
-			(0, vitest.expect)(typeof fetchUrlTool.execute).toBe(`function`);
+			const { createFetchUrlTool } = await import(`../../agents-runtime/src/tools`);
+			const sandbox = await makeSandbox(`/tmp`);
+			try {
+				const tool = createFetchUrlTool(sandbox);
+				(0, vitest.expect)(tool.name).toBe(`fetch_url`);
+				(0, vitest.expect)(typeof tool.execute).toBe(`function`);
+			} finally {
+				await sandbox.dispose();
+			}
 		});
 	});
 	(0, vitest.describe)(`Electric Agents Send — State Protocol Format`, () => {

package/dist/index.d.cts

@@ -64,7 +64,6 @@ type HistoryEvent = {
   status: string;
   streams: {
     main: string;
-    error: string;
   };
   parent?: string;
 } | {
@@ -184,7 +183,6 @@ interface WebhookEntityContext {
   url: string;
   streams: {
     main: string;
-    error: string;
   };
   tags?: Record<string, string>;
 }
@@ -273,7 +271,6 @@ interface RunContext {
   currentEntityUrl: string | null;
   currentEntityStreams: {
     main: string;
-    error: string;
   } | null;
   currentWriteToken: string | null;
   notification: WebhookNotification | null;
@@ -320,7 +317,7 @@ declare class ElectricAgentsScenario {
   expectEntityContext(checks?: EntityContextChecks): this;
   expectStatus(status: string): this;
   expectStreamContains(messageType: string): this;
-  readStream(stream?: `main` | `error`): this;
+  readStream(): this;
   list(filter?: {
     type?: string;
     status?: string;

package/dist/index.d.ts

@@ -64,7 +64,6 @@ type HistoryEvent = {
   status: string;
   streams: {
     main: string;
-    error: string;
   };
   parent?: string;
 } | {
@@ -184,7 +183,6 @@ interface WebhookEntityContext {
   url: string;
   streams: {
     main: string;
-    error: string;
   };
   tags?: Record<string, string>;
 }
@@ -273,7 +271,6 @@ interface RunContext {
   currentEntityUrl: string | null;
   currentEntityStreams: {
     main: string;
-    error: string;
   } | null;
   currentWriteToken: string | null;
   notification: WebhookNotification | null;
@@ -320,7 +317,7 @@ declare class ElectricAgentsScenario {
   expectEntityContext(checks?: EntityContextChecks): this;
   expectStatus(status: string): this;
   expectStreamContains(messageType: string): this;
-  readStream(stream?: `main` | `error`): this;
+  readStream(): this;
   list(filter?: {
     type?: string;
     status?: string;

package/dist/index.js

@@ -304,11 +304,8 @@ var ElectricAgentsScenario = class {
 		});
 		return this;
 	}
-	readStream(stream) {
-		this.steps.push({
-			kind: `readStream`,
-			stream
-		});
+	readStream() {
+		this.steps.push({ kind: `readStream` });
 		return this;
 	}
 	list(filter) {
@@ -647,7 +644,6 @@ async function executeStep(ctx, step) {
 			expect(entity, `Webhook payload must contain entity context`).toBeDefined();
 			expect(entity.url).toBeTruthy();
 			expect(entity.streams.main).toBeTruthy();
-			expect(entity.streams.error).toBeTruthy();
 			if (step.checks?.url) expect(entity.url).toBe(step.checks.url);
 			if (step.checks?.type) expect(entity.type).toBe(step.checks.type);
 			if (step.checks?.status) expect(entity.status).toBe(step.checks.status);
@@ -688,7 +684,7 @@ async function executeStep(ctx, step) {
 		}
 		case `readStream`: {
 			if (!ctx.currentEntityStreams) throw new Error(`No current entity streams`);
-			const streamPath = step.stream === `error` ? ctx.currentEntityStreams.error : ctx.currentEntityStreams.main;
+			const streamPath = ctx.currentEntityStreams.main;
 			const res = await fetch(appendPathToUrl(ctx.baseUrl, `${streamPath}?offset=0000000000000000_0000000000000000`));
 			if (res.status === 200) {
 				const text = await res.text();
@@ -1183,14 +1179,11 @@ function checkEntityContextOnWebhook(history) {
 	}
 }
 /**
-* Spec S5 — Structural: stream paths must match {entity.url}/main and {entity.url}/error.
+* Spec S5 — Structural: stream path must match {entity.url}/main.
 * Soundness: Sound | Completeness: Complete (within trace)
 */
 function checkStreamPathsMatchEntityUrl(history) {
-	for (const event of history) if (event.type === `entity_spawned`) {
-		expect(event.streams.main).toBe(`${event.entityUrl}/main`);
-		expect(event.streams.error).toBe(`${event.entityUrl}/error`);
-	}
+	for (const event of history) if (event.type === `entity_spawned`) expect(event.streams.main).toBe(`${event.entityUrl}/main`);
 }
 /**
 * Spec S4 — Safety: entity status transitions must be valid.
@@ -1252,8 +1245,8 @@ function checkAdditiveSchemaEvolution(history) {
 }
 /**
 * Spec R4 — Registry stream consistency: every entity_spawned must refer to an
-* entity that has not already been killed, and stream URLs must follow the
-* convention {entity.url}/main and {entity.url}/error.
+* entity that has not already been killed, and stream URL must follow the
+* convention {entity.url}/main.
 * Soundness: Sound | Completeness: Complete (within trace)
 */
 function checkRegistryStreamConsistency(history) {
@@ -1263,7 +1256,6 @@ function checkRegistryStreamConsistency(history) {
 		if (event.type === `entity_spawned`) {
 			expect(!killedEntities.has(event.entityUrl), `Registry consistency: entity_spawned for ${event.entityUrl} but entity was already killed`).toBe(true);
 			expect(event.streams.main, `Registry consistency: entity ${event.entityUrl} streams.main must be ${event.entityUrl}/main`).toBe(`${event.entityUrl}/main`);
-			expect(event.streams.error, `Registry consistency: entity ${event.entityUrl} streams.error must be ${event.entityUrl}/error`).toBe(`${event.entityUrl}/error`);
 		}
 	}
 }
@@ -1808,8 +1800,6 @@ function runElectricAgentsConformanceTests(config) {
 		}).spawn(`spawn-test-agent`, `entity-1`).expectStatus(`running`).custom(async (ctx) => {
 			const mainRes = await fetch(appendPathToUrl(ctx.baseUrl, ctx.currentEntityStreams.main), { method: `HEAD` });
 			expect(mainRes.status).toBe(200);
-			const errorRes = await fetch(appendPathToUrl(ctx.baseUrl, ctx.currentEntityStreams.error), { method: `HEAD` });
-			expect(errorRes.status).toBe(200);
 		}).run());
 		test(`spawn at unregistered type returns UNKNOWN_ENTITY_TYPE`, () => electricAgents(config.baseUrl).custom(async (ctx) => {
 			const res = await fetch(appendPathToUrl(ctx.baseUrl, routeControlPlanePath(`/unregistered/entity-1`)), {
@@ -2880,25 +2870,51 @@ function runElectricAgentsConformanceTests(config) {
 			const block = result.content[0];
 			return block?.type === `text` && block.text ? block.text : ``;
 		}
+		async function makeSandbox(workingDirectory) {
+			const { unrestrictedSandbox } = await import(`../../agents-runtime/src/sandbox/unrestricted`);
+			return unrestrictedSandbox({ workingDirectory });
+		}
 		test(`bash tool captures stdout and stderr`, async () => {
 			const { createBashTool } = await import(`../../agents-runtime/src/tools`);
-			const tool = createBashTool(`/tmp`);
-			const result = await tool.execute(`test-tc`, { command: `echo "hello" && echo "error" >&2` });
-			expect(firstText(result)).toContain(`hello`);
-			expect(firstText(result)).toContain(`error`);
-			expect(result.details.exitCode).toBe(0);
+			const sandbox = await makeSandbox(`/tmp`);
+			try {
+				const tool = createBashTool(sandbox);
+				const result = await tool.execute(`test-tc`, { command: `echo "hello" && echo "error" >&2` });
+				expect(firstText(result)).toContain(`hello`);
+				expect(firstText(result)).toContain(`error`);
+				expect(result.details.exitCode).toBe(0);
+			} finally {
+				await sandbox.dispose();
+			}
 		});
 		test(`bash tool enforces timeout`, async () => {
 			const { createBashTool } = await import(`../../agents-runtime/src/tools`);
-			const tool = createBashTool(`/tmp`);
-			const result = await tool.execute(`test-tc`, { command: `sleep 60` });
-			expect(result.details.timedOut).toBe(true);
+			const sandbox = await makeSandbox(`/tmp`);
+			try {
+				const tool = createBashTool(sandbox);
+				const result = await tool.execute(`test-tc`, { command: `sleep 60` });
+				expect(result.details.timedOut).toBe(true);
+			} finally {
+				await sandbox.dispose();
+			}
 		}, 35e3);
 		test(`read_file rejects paths outside working directory`, async () => {
 			const { createReadFileTool } = await import(`../../agents-runtime/src/tools`);
-			const tool = createReadFileTool(`/tmp/test-workdir`);
-			const result = await tool.execute(`test-tc`, { path: `../../etc/passwd` });
-			expect(firstText(result)).toContain(`outside the working directory`);
+			const fs = await import(`node:fs/promises`);
+			const dir = `/tmp/test-workdir-${Date.now()}`;
+			await fs.mkdir(dir, { recursive: true });
+			const sandbox = await makeSandbox(dir);
+			try {
+				const tool = createReadFileTool(sandbox);
+				const result = await tool.execute(`test-tc`, { path: `../../etc/passwd` });
+				expect(firstText(result)).toContain(`outside the working directory`);
+			} finally {
+				await sandbox.dispose();
+				await fs.rm(dir, {
+					recursive: true,
+					force: true
+				});
+			}
 		});
 		test(`read_file rejects binary files`, async () => {
 			const { createReadFileTool } = await import(`../../agents-runtime/src/tools`);
@@ -2913,10 +2929,15 @@ function runElectricAgentsConformanceTests(config) {
 				2,
 				255
 			]));
-			const tool = createReadFileTool(dir);
-			const result = await tool.execute(`test-tc`, { path: `test.bin` });
-			expect(firstText(result)).toContain(`binary file`);
-			await fs.rm(dir, { recursive: true });
+			const sandbox = await makeSandbox(dir);
+			try {
+				const tool = createReadFileTool(sandbox);
+				const result = await tool.execute(`test-tc`, { path: `test.bin` });
+				expect(firstText(result)).toContain(`binary file`);
+			} finally {
+				await sandbox.dispose();
+				await fs.rm(dir, { recursive: true });
+			}
 		});
 		test(`read_file rejects oversized files`, async () => {
 			const { createReadFileTool } = await import(`../../agents-runtime/src/tools`);
@@ -2926,10 +2947,15 @@ function runElectricAgentsConformanceTests(config) {
 			await fs.mkdir(dir, { recursive: true });
 			const bigPath = path.join(dir, `big.txt`);
 			await fs.writeFile(bigPath, `x`.repeat(600 * 1024));
-			const tool = createReadFileTool(dir);
-			const result = await tool.execute(`test-tc`, { path: `big.txt` });
-			expect(firstText(result)).toContain(`too large`);
-			await fs.rm(dir, { recursive: true });
+			const sandbox = await makeSandbox(dir);
+			try {
+				const tool = createReadFileTool(sandbox);
+				const result = await tool.execute(`test-tc`, { path: `big.txt` });
+				expect(firstText(result)).toContain(`too large`);
+			} finally {
+				await sandbox.dispose();
+				await fs.rm(dir, { recursive: true });
+			}
 		});
 		test(`web_search tool has correct interface`, async () => {
 			const { braveSearchTool } = await import(`../../agents-runtime/src/tools`);
@@ -2937,9 +2963,15 @@ function runElectricAgentsConformanceTests(config) {
 			expect(typeof braveSearchTool.execute).toBe(`function`);
 		});
 		test(`fetch_url tool has correct interface`, async () => {
-			const { fetchUrlTool } = await import(`../../agents-runtime/src/tools`);
-			expect(fetchUrlTool.name).toBe(`fetch_url`);
-			expect(typeof fetchUrlTool.execute).toBe(`function`);
+			const { createFetchUrlTool } = await import(`../../agents-runtime/src/tools`);
+			const sandbox = await makeSandbox(`/tmp`);
+			try {
+				const tool = createFetchUrlTool(sandbox);
+				expect(tool.name).toBe(`fetch_url`);
+				expect(typeof tool.execute).toBe(`function`);
+			} finally {
+				await sandbox.dispose();
+			}
 		});
 	});
 	describe(`Electric Agents Send — State Protocol Format`, () => {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@electric-ax/agents-server-conformance-tests",
-  "version": "0.1.9",
+  "version": "0.1.11",
   "description": "Conformance test suite for Electric Agents server implementations",
   "author": "Durable Stream contributors",
   "license": "Apache-2.0",
@@ -35,7 +35,7 @@
   ],
   "dependencies": {
     "@durable-streams/client": "^0.2.6",
-    "@electric-sql/client": "^1.5.19",
+    "@electric-sql/client": "^1.5.20",
     "fast-check": "^4.6.0",
     "vitest": "^4.1.0"
   },
@@ -43,7 +43,7 @@
     "@mariozechner/pi-agent-core": "^0.70.2",
     "tsdown": "^0.9.0",
     "typescript": "^5.0.0",
-    "@electric-ax/agents-server": "0.4.13"
+    "@electric-ax/agents-server": "0.4.16"
   },
   "engines": {
     "node": ">=18.0.0"