feat(agent): add controlled task execution

2026-09-24 04:19:02 +08:00
parent d018bcae67
commit b31033e9ce
48 changed files with 14498 additions and 86 deletions
@@ -0,0 +1,94 @@
+import { describe, expect, it } from "bun:test";
+import { createExecutionPlanSnapshot, executionPlanNeedsTaskPublish, executionPlanSchema, readExecutionPlans } from "../../src/routes/executionPlan.js";
+import { streamPromptResponse } from "../../src/routes/chatStream.js";
+import type { OpencodeRuntimeAdapter } from "../../src/runtime/opencode.js";
+
+const plan = {
+  title: "水质调查", reason: "收到投诉",
+  nodes: [
+    { id: "upstream", title: "上游监测", status: "completed", owner: "agent", detail: "合成数据", evidence: ["场景输入：上游指标正常"] },
+    { id: "sample", title: "现场采样", status: "waiting", owner: "human", detail: "等待现场记录", evidence: [] },
+  ],
+  edges: [{ from: "upstream", to: "sample", label: "上游正常" }],
+};
+
+describe("execution plans", () => {
+  it("rejects missing dependencies, cycles, duplicate IDs and unsupported completion", () => {
+    for (const invalid of [
+      { ...plan, edges: [{ from: "unknown", to: "sample", label: "" }] },
+      { ...plan, edges: [...plan.edges, { from: "sample", to: "upstream", label: "" }] },
+      { ...plan, nodes: [plan.nodes[0], plan.nodes[0]] },
+      { ...plan, nodes: [{ ...plan.nodes[0], evidence: [] }, plan.nodes[1]] },
+      {
+        ...plan,
+        nodes: [
+          { ...plan.nodes[0], status: "waiting", detail: "等待数据", evidence: [] },
+          { ...plan.nodes[1], status: "completed", evidence: ["evidence-2"] },
+        ],
+      },
+    ]) expect(executionPlanSchema.safeParse(invalid).success).toBe(false);
+  });
+
+  it("preserves completed evidence and versions across persisted UI messages and active snapshots", () => {
+    const first = createExecutionPlanSnapshot(plan, [], "session-1");
+    const messages = [{ parts: [{ type: "data-execution_plan", data: first }] }, { executionPlans: [first] }];
+    expect(readExecutionPlans(messages)).toHaveLength(1);
+    const next = createExecutionPlanSnapshot({ ...plan, reason: "等待人工补充" }, messages, "session-1");
+    expect(next.revision).toBe(2);
+    expect(next.nodes[1].status).toBe("waiting");
+    expect(() => createExecutionPlanSnapshot({ ...plan, nodes: [plan.nodes[1]], edges: [] }, messages, "session-1")).toThrow("Keep previous nodes");
+    expect(() => createExecutionPlanSnapshot({ ...plan, nodes: [{ ...plan.nodes[0], status: "running" }, plan.nodes[1]] }, messages, "session-1")).toThrow("Keep completed evidence");
+  });
+
+  it("publishes only successful validated tool completions, once per call", async () => {
+    const part = (status: string) => ({ type: "message.part.updated", properties: { sessionID: "runtime", part: {
+      id: "plan-call", messageID: "assistant", sessionID: "runtime", type: "tool", tool: "publish_execution_plan",
+      callID: "call-1", state: { status, input: plan, output: JSON.stringify({ plan_revision: 1, report_revision: null }), title: "plan", time: { start: 1, end: 2 } }
+    } } });
+    const events = [part("running"), part("completed"), part("completed"), { type: "session.idle", properties: { sessionID: "runtime" } }];
+    const runtime = {
+      subscribeEvents: async () => ({ async *[Symbol.asyncIterator]() { for (const event of events) yield event; } }),
+      prompt: async () => undefined, messages: async () => []
+    } as unknown as OpencodeRuntimeAdapter;
+    const published: unknown[] = [];
+    await streamPromptResponse({ runtime, sessionId: "runtime", clientSessionId: "session-1", message: "investigate", write(event, data) { if (event === "execution_plan") published.push(data); } });
+    expect(published).toHaveLength(1);
+    expect(published[0]).toMatchObject({ session_id: "session-1", server_published: true, nodes: plan.nodes });
+  });
+
+  it("falls back to route publication for a completed legacy local-only plan tool", async () => {
+    const completed = { type: "message.part.updated", properties: { sessionID: "runtime", part: {
+      id: "plan-call", messageID: "assistant", sessionID: "runtime", type: "tool", tool: "publish_execution_plan",
+      callID: "call-1", state: { status: "completed", input: plan, output: JSON.stringify(plan), title: "plan", time: { start: 1, end: 2 } }
+    } } };
+    const runtime = {
+      subscribeEvents: async () => ({ async *[Symbol.asyncIterator]() { yield completed; yield { type: "session.idle", properties: { sessionID: "runtime" } }; } }),
+      prompt: async () => undefined, messages: async () => []
+    } as unknown as OpencodeRuntimeAdapter;
+    const published: Record<string, unknown>[] = [];
+
+    await streamPromptResponse({ runtime, sessionId: "runtime", clientSessionId: "session-1", message: "investigate", write(event, data) { if (event === "execution_plan") published.push(data); } });
+
+    expect(published).toHaveLength(1);
+    expect(published[0]?.server_published).toBeUndefined();
+    expect(executionPlanNeedsTaskPublish(published[0])).toBe(true);
+  });
+
+  it("does not publish a failed tool plan or duplicate an already persisted plan", async () => {
+    const failed = { type: "message.part.updated", properties: { sessionID: "runtime", part: {
+      id: "plan-call", messageID: "assistant", sessionID: "runtime", type: "tool", tool: "publish_execution_plan",
+      callID: "call-1", state: { status: "error", input: plan, error: "server rejected plan", time: { start: 1, end: 2 } }
+    } } };
+    const runtime = {
+      subscribeEvents: async () => ({ async *[Symbol.asyncIterator]() { yield failed; yield { type: "session.idle", properties: { sessionID: "runtime" } }; } }),
+      prompt: async () => undefined, messages: async () => []
+    } as unknown as OpencodeRuntimeAdapter;
+    const published: unknown[] = [];
+
+    await streamPromptResponse({ runtime, sessionId: "runtime", clientSessionId: "session-1", message: "investigate", write(event, data) { if (event === "execution_plan") published.push(data); } });
+
+    expect(published).toEqual([]);
+    expect(executionPlanNeedsTaskPublish({ server_published: true })).toBe(false);
+    expect(executionPlanNeedsTaskPublish(plan)).toBe(true);
+  });
+});