@stigmer/runner 3.15.1 → 3.15.2-dev.20260913223433

package/src/activities/execute-deep-agent/__tests__/subagent-tracker.test.ts

@@ -1,7 +1,9 @@
 import { describe, it, expect, beforeEach } from "vitest";
 import { create } from "@bufbuild/protobuf";
 import { AgentExecutionStatusSchema } from "@stigmer/protos/ai/stigmer/agentic/agentexecution/v1/api_pb";
+import { SubAgentExecutionSchema } from "@stigmer/protos/ai/stigmer/agentic/agentexecution/v1/subagent_pb";
 import { SubAgentStatus, MessageType, ToolCallStatus } from "@stigmer/protos/ai/stigmer/agentic/agentexecution/v1/enum_pb";
+import { cancelInProgressSubAgentProtos } from "../../../shared/subagent-rows.js";
 import { V3StatusBuilder } from "../v3-status-builder.js";
 import { normalize } from "../v3-protocol-normalizer.js";
 import type { V3ProtocolEvent } from "../v3-event-recorder.js";
@@ -88,8 +90,6 @@ describe("SubAgentTracker (via V3StatusBuilder integration)", () => {
         makeMessageFinish("parent-1"),
         makeTaskToolStarted("call_sub_1", "researcher", "Research renewable energy"),
       ]);
-
-      sb.syncSubAgentExecutions();
       const subs = sb.currentStatus.subAgentExecutions;
       expect(subs).toHaveLength(1);
       expect(subs[0].id).toBe("call_sub_1");
@@ -106,8 +106,6 @@ describe("SubAgentTracker (via V3StatusBuilder integration)", () => {
         makeTaskToolStarted("call_sub_1", "researcher", "Research topic"),
         makeTaskToolFinished("call_sub_1", "Renewable energy summary here."),
       ]);
-
-      sb.syncSubAgentExecutions();
       const sub = sb.currentStatus.subAgentExecutions[0];
       expect(sub.status).toBe(SubAgentStatus.SUB_AGENT_COMPLETED);
       expect(sub.output).toBe("Renewable energy summary here.");
@@ -120,8 +118,6 @@ describe("SubAgentTracker (via V3StatusBuilder integration)", () => {
         makeTaskToolStarted("call_sub_1", "researcher", "Research topic"),
         makeTaskToolError("call_sub_1", "Context overflow in sub-agent"),
       ]);
-
-      sb.syncSubAgentExecutions();
       const sub = sb.currentStatus.subAgentExecutions[0];
       expect(sub.status).toBe(SubAgentStatus.SUB_AGENT_FAILED);
       expect(sub.error).toBe("Context overflow in sub-agent");
@@ -129,19 +125,19 @@ describe("SubAgentTracker (via V3StatusBuilder integration)", () => {
     });
   });
 
-  describe("lifecycle — cancellation", () => {
-    it("cancelSubAgents marks active sub-agents as CANCELLED", () => {
+  describe("lifecycle — a stopped turn (the shared cancel act + the tracker's streaming finalize)", () => {
+    it("the shared cancelInProgressSubAgentProtos marks the tracked rows CANCELLED in place", () => {
       const sb = makeBuilder();
       feedAll(sb, [
         makeTaskToolStarted("call_sub_1", "researcher", "Task 1"),
         makeTaskToolStarted("call_sub_2", "shell", "Task 2"),
       ]);
 
-      sb.cancelSubAgents();
+      cancelInProgressSubAgentProtos(sb.currentStatus.subAgentExecutions);
       const subs = sb.currentStatus.subAgentExecutions;
       expect(subs).toHaveLength(2);
       expect(subs[0].status).toBe(SubAgentStatus.SUB_AGENT_CANCELLED);
-      expect(subs[0].error).toContain("parent execution was cancelled");
+      expect(subs[0].completedAt).toBeTruthy();
       expect(subs[1].status).toBe(SubAgentStatus.SUB_AGENT_CANCELLED);
     });
 
@@ -153,11 +149,61 @@ describe("SubAgentTracker (via V3StatusBuilder integration)", () => {
         makeTaskToolStarted("call_sub_2", "shell", "Task 2"),
       ]);
 
-      sb.cancelSubAgents();
+      cancelInProgressSubAgentProtos(sb.currentStatus.subAgentExecutions);
       const subs = sb.currentStatus.subAgentExecutions;
       expect(subs[0].status).toBe(SubAgentStatus.SUB_AGENT_COMPLETED);
       expect(subs[1].status).toBe(SubAgentStatus.SUB_AGENT_CANCELLED);
     });
+
+    it("finalizeSubAgentStreaming clears the streaming flag on a sub-agent's in-flight message", () => {
+      const sb = makeBuilder();
+      feedAll(sb, [
+        makeTaskToolStarted("call_sub_1", "researcher", "Task 1"),
+        makeSubAgentEvent("call_sub_1", "messages", { event: "message-start", id: "msg_sub_1", run_id: "sub-run-1" }),
+        makeSubAgentEvent("call_sub_1", "messages", {
+          event: "content-block-delta", index: 0, delta: { type: "text-delta", text: "Half a sente" }, run_id: "sub-run-1",
+        }),
+      ]);
+      const msg = sb.currentStatus.subAgentExecutions[0].messages[0];
+      expect(msg.isStreaming, "the delta left the message streaming").toBe(true);
+
+      sb.finalizeSubAgentStreaming();
+      expect(msg.isStreaming).toBe(false);
+    });
+  });
+
+  describe("the status's array is built into by reference (S3 M1 deferred debt)", () => {
+    it("new rows are pushed onto the array the builder was handed, never a replacement", () => {
+      const status = create(AgentExecutionStatusSchema, {});
+      const rows = status.subAgentExecutions;
+      const sb = new V3StatusBuilder("exec-test", status);
+      feedAll(sb, [makeTaskToolStarted("call_sub_1", "researcher", "Task 1")]);
+      expect(status.subAgentExecutions, "the same array object").toBe(rows);
+      expect(rows).toHaveLength(1);
+    });
+
+    it("a seeded row re-announced by a replayed task start is reconciled, not duplicated", () => {
+      const seeded = create(SubAgentExecutionSchema, {
+        id: "call_sub_1",
+        name: "researcher",
+        status: SubAgentStatus.SUB_AGENT_IN_PROGRESS,
+      });
+      const status = create(AgentExecutionStatusSchema, { subAgentExecutions: [seeded] });
+      const sb = new V3StatusBuilder("exec-test", status);
+      feedAll(sb, [
+        makeTaskToolStarted("call_sub_1", "researcher", "Task 1"),
+        makeSubAgentEvent("call_sub_1", "messages", { event: "message-start", id: "msg_sub_1", run_id: "sub-run-1" }),
+        makeSubAgentEvent("call_sub_1", "messages", {
+          event: "content-block-delta", index: 0, delta: { type: "text-delta", text: "Resumed." }, run_id: "sub-run-1",
+        }),
+        makeSubAgentEvent("call_sub_1", "messages", { event: "message-finish", reason: "end_turn", run_id: "sub-run-1" }),
+        makeTaskToolFinished("call_sub_1", "done"),
+      ]);
+      expect(status.subAgentExecutions, "one row, the seeded one").toHaveLength(1);
+      expect(status.subAgentExecutions[0]).toBe(seeded);
+      expect(seeded.status).toBe(SubAgentStatus.SUB_AGENT_COMPLETED);
+      expect(seeded.messages.map((m) => m.content), "child events routed onto the seeded row").toEqual(["Resumed."]);
+    });
   });
 
   describe("namespace routing — sub-agent messages isolated from parent", () => {
@@ -182,8 +228,6 @@ describe("SubAgentTracker (via V3StatusBuilder integration)", () => {
         makeTaskToolFinished("call_sub_1", "Research complete."),
       ]);
 
-      sb.syncSubAgentExecutions();
-
       // Parent messages should NOT contain sub-agent text
       const parentMessages = sb.currentStatus.messages;
       expect(parentMessages).toHaveLength(1);
@@ -219,8 +263,6 @@ describe("SubAgentTracker (via V3StatusBuilder integration)", () => {
           event: "message-finish", reason: "end_turn", run_id: "sub-run-1",
         }),
       ]);
-
-      sb.syncSubAgentExecutions();
       const sub = sb.currentStatus.subAgentExecutions[0];
       expect(sub.messages).toHaveLength(2);
       expect(sub.messages[0].type).toBe(MessageType.MESSAGE_THINKING);
@@ -262,8 +304,6 @@ describe("SubAgentTracker (via V3StatusBuilder integration)", () => {
           output: { lc: 1, type: "constructor", id: ["langchain_core", "messages", "ToolMessage"], kwargs: { content: "Found 3 matches", status: "success", tool_call_id: subToolCallId, name: "grep" } },
         }, { namespace: [toolsNodeSegment(taskCallId), `tools:${subToolCallId}`] }),
       ]);
-
-      sb.syncSubAgentExecutions();
       const sub = sb.currentStatus.subAgentExecutions[0];
       expect(sub.messages).toHaveLength(1);
       expect(sub.messages[0].toolCalls).toHaveLength(1);
@@ -307,8 +347,6 @@ describe("SubAgentTracker (via V3StatusBuilder integration)", () => {
         makeTaskToolFinished("call_sub_1", "A done"),
         makeTaskToolFinished("call_sub_2", "B done"),
       ]);
-
-      sb.syncSubAgentExecutions();
       const subs = sb.currentStatus.subAgentExecutions;
       expect(subs).toHaveLength(2);
 
@@ -332,8 +370,6 @@ describe("SubAgentTracker (via V3StatusBuilder integration)", () => {
         makeTaskToolStarted("call_2", "researcher", "Research B"),
         makeTaskToolFinished("call_2", "Result B"),
       ]);
-
-      sb.syncSubAgentExecutions();
       const subs = sb.currentStatus.subAgentExecutions;
       expect(subs).toHaveLength(2);
       expect(subs[0].id).toBe("call_1");
@@ -356,8 +392,6 @@ describe("SubAgentTracker (via V3StatusBuilder integration)", () => {
         makeTextDelta("parent-2", "The research is complete."),
         makeMessageFinish("parent-2"),
       ]);
-
-      sb.syncSubAgentExecutions();
       const msgs = sb.currentStatus.messages;
 
       // First AI message has the task tool call
@@ -380,8 +414,6 @@ describe("SubAgentTracker (via V3StatusBuilder integration)", () => {
         makeTaskToolStarted("call_sub_1", "researcher", "Research"),
         makeTaskToolStarted("call_sub_1", "researcher", "Research"),
       ]);
-
-      sb.syncSubAgentExecutions();
       expect(sb.currentStatus.subAgentExecutions).toHaveLength(1);
     });
 
@@ -390,8 +422,6 @@ describe("SubAgentTracker (via V3StatusBuilder integration)", () => {
       feedAll(sb, [
         makeTaskToolFinished("unknown_call_id", "output"),
       ]);
-
-      sb.syncSubAgentExecutions();
       expect(sb.currentStatus.subAgentExecutions).toHaveLength(0);
     });
 
@@ -431,8 +461,6 @@ describe("SubAgentTracker (via V3StatusBuilder integration)", () => {
           event: "message-finish", reason: "end_turn", run_id: "sub-run",
         }, { namespace: [`tools:${callId}`, "model_request"], node: "model_request" }),
       ]);
-
-      sb.syncSubAgentExecutions();
       const subs = sb.currentStatus.subAgentExecutions;
       expect(subs).toHaveLength(1);
       expect(subs[0].name).toBe("researcher");
@@ -459,8 +487,6 @@ describe("SubAgentTracker (via V3StatusBuilder integration)", () => {
           event: "message-finish", reason: "end_turn", run_id: "parent-run",
         }, { namespace: [], node: "model_request" }),
       ]);
-
-      sb.syncSubAgentExecutions();
       // First message: AI message with task tool call (created by handleToolStarted)
       // Second message: parent text that follows
       expect(sb.currentStatus.messages).toHaveLength(2);

package/src/activities/execute-deep-agent/__tests__/subagent-transformer.test.ts

@@ -7,7 +7,6 @@ import {
   createBuiltinSubagents,
   transformSingleSubagent,
   filterMcpToolsForSubagent,
-  collectAllSkillRefs,
   resolveSubagentSkillPrompt,
   compileSubagents,
   transformAndCompileSubagents,
@@ -448,7 +447,7 @@ describe("transformAndCompileSubagents", () => {
     parentMcpTools: [] as StructuredTool[],
     parentMcpServerToolMap: new Map<string, StructuredTool[]>(),
     parentMcpUsages: [] as McpServerUsage[],
-    skillClient: {} as unknown,
+    skills: new Map(),
     workspaceBackend: mockWorkspaceBackend(),
     approvalGate: null,
     parentModelName: "claude-sonnet-4-6",
@@ -543,91 +542,32 @@ describe("transformAndCompileSubagents", () => {
   });
 });
 
-// =========================================================================
-// Tests: collectAllSkillRefs
-// =========================================================================
-
-describe("collectAllSkillRefs", () => {
-  it("returns empty for subagents without skills", () => {
-    const result = collectAllSkillRefs([
-      mockSubAgentProto({ skillRefs: [] }),
-      mockSubAgentProto({ skillRefs: [] }),
-    ] as unknown as import("@stigmer/protos/ai/stigmer/agentic/agent/v1/spec_pb").SubAgent[]);
-
-    expect(result).toEqual([]);
-  });
-
-  it("collects unique refs from multiple subagents", () => {
-    const result = collectAllSkillRefs([
-      mockSubAgentProto({ skillRefs: [{ slug: "org/skill-a" }] }),
-      mockSubAgentProto({ skillRefs: [{ slug: "org/skill-b" }] }),
-    ] as unknown as import("@stigmer/protos/ai/stigmer/agentic/agent/v1/spec_pb").SubAgent[]);
-
-    expect(result).toHaveLength(2);
-    expect(result.map((r) => r.slug)).toEqual(["org/skill-a", "org/skill-b"]);
-  });
-
-  it("deduplicates refs by slug", () => {
-    const result = collectAllSkillRefs([
-      mockSubAgentProto({ skillRefs: [{ slug: "org/same-skill" }] }),
-      mockSubAgentProto({ skillRefs: [{ slug: "org/same-skill" }] }),
-    ] as unknown as import("@stigmer/protos/ai/stigmer/agentic/agent/v1/spec_pb").SubAgent[]);
-
-    expect(result).toHaveLength(1);
-    expect(result[0].slug).toBe("org/same-skill");
-  });
-
-  it("skips refs without slugs", () => {
-    const result = collectAllSkillRefs([
-      mockSubAgentProto({ skillRefs: [{ slug: "" }, { slug: "org/valid" }] }),
-    ] as unknown as import("@stigmer/protos/ai/stigmer/agentic/agent/v1/spec_pb").SubAgent[]);
-
-    expect(result).toHaveLength(1);
-    expect(result[0].slug).toBe("org/valid");
-  });
-});
-
 // =========================================================================
 // Tests: resolveSubagentSkillPrompt
 // =========================================================================
 
 describe("resolveSubagentSkillPrompt", () => {
-  it("returns empty string for subagent without skill refs", () => {
-    const proto = mockSubAgentProto({ skillRefs: [] });
-    const result = resolveSubagentSkillPrompt(
-      proto as unknown as import("@stigmer/protos/ai/stigmer/agentic/agent/v1/spec_pb").SubAgent,
-      new Map(),
-    );
-    expect(result).toBe("");
+  const asSubAgent = (proto: unknown): import("@stigmer/protos/ai/stigmer/agentic/agent/v1/spec_pb").SubAgent =>
+    proto as import("@stigmer/protos/ai/stigmer/agentic/agent/v1/spec_pb").SubAgent;
+
+  it("returns empty string for a sub-agent the runtime mounted no skills for", () => {
+    const proto = mockSubAgentProto({ name: "researcher", skillRefs: [] });
+    expect(resolveSubagentSkillPrompt(asSubAgent(proto), new Map())).toBe("");
   });
 
-  it("returns empty string when no skills match", () => {
-    const proto = mockSubAgentProto({ skillRefs: [{ slug: "org/missing" }] });
-    const result = resolveSubagentSkillPrompt(
-      proto as unknown as import("@stigmer/protos/ai/stigmer/agentic/agent/v1/spec_pb").SubAgent,
-      new Map(),
-    );
-    expect(result).toBe("");
-  });
-
-  it("generates prompt section for matched skills", () => {
-    const proto = mockSubAgentProto({ skillRefs: [{ slug: "org/my-skill" }] });
-    const skillsBySlug = new Map([
-      ["org/my-skill", {
-        skill: {
-          metadata: { id: "sk-1", name: "my-skill", slug: "org/my-skill" },
-          spec: { name: "my-skill", description: "A test skill", skillMd: "# Skill\nDo things." },
-        },
-        path: ".stigmer/skills/my-skill",
-      }],
+  it("renders the skills section from the sub-agent's own mounted skills, keyed by its name", () => {
+    const proto = mockSubAgentProto({ name: "researcher", skillRefs: [{ slug: "org/my-skill" }] });
+    const skills = new Map([
+      ["researcher", [{ name: "my-skill", description: "A test skill", path: ".stigmer/skills/my-skill/SKILL.md" }]],
+      ["other", [{ name: "not-mine", description: "Another agent's", path: ".stigmer/skills/not-mine/SKILL.md" }]],
     ]);
 
-    const result = resolveSubagentSkillPrompt(
-      proto as unknown as import("@stigmer/protos/ai/stigmer/agentic/agent/v1/spec_pb").SubAgent,
-      skillsBySlug,
-    );
+    const result = resolveSubagentSkillPrompt(asSubAgent(proto), skills);
 
-    expect(result.length).toBeGreaterThan(0);
+    expect(result).toContain("## Skills");
+    expect(result).toContain("### my-skill");
+    expect(result).toContain("**Location**: `.stigmer/skills/my-skill/`");
+    expect(result).not.toContain("not-mine");
   });
 });
 
@@ -681,7 +621,7 @@ describe("subagent-transformer edge cases", () => {
       parentMcpTools: [],
       parentMcpServerToolMap: new Map(),
       parentMcpUsages: [],
-      skillClient: {} as unknown,
+      skills: new Map(),
       workspaceBackend: mockWorkspaceBackend({ rootDir: "" }),
       approvalGate: null,
       parentModelName: "claude-sonnet-4-6",
@@ -705,7 +645,7 @@ describe("subagent-transformer edge cases", () => {
       parentMcpTools: [],
       parentMcpServerToolMap: new Map(),
       parentMcpUsages: [],
-      skillClient: {} as unknown,
+      skills: new Map(),
       workspaceBackend: mockWorkspaceBackend(),
       approvalGate: null,
       parentModelName: "claude-sonnet-4-6",

package/src/activities/execute-deep-agent/__tests__/subagent-wiring.test.ts

@@ -1,10 +1,10 @@
 import { describe, it, expect, vi } from "vitest";
 import { buildSubAgentMiddleware } from "../subagent-wiring.js";
-import { createCostCapMiddleware } from "../../../middleware/cost-cap.js";
+import { createCostAdvisoryMiddleware } from "../../../middleware/cost-advisory.js";
 import * as approvalGateModule from "../../../middleware/approval-gate.js";
 
 describe("buildSubAgentMiddleware", () => {
-  it("returns the correct middleware order without cost cap", () => {
+  it("returns the correct middleware order without a cost advisory", () => {
     const stack = buildSubAgentMiddleware();
 
     expect(stack).toHaveLength(5);
@@ -15,8 +15,8 @@ describe("buildSubAgentMiddleware", () => {
     expect(stack[4].name).toBe("ErrorHintsMiddleware");
   });
 
-  it("includes cost cap view when parent cost cap is provided", () => {
-    const parentCostCap = createCostCapMiddleware({
+  it("includes the parent's cost advisory view when one is provided", () => {
+    const parentCostAdvisory = createCostAdvisoryMiddleware({
       maxCostUsd: 10,
       inputPricePerMillion: 3,
       outputPricePerMillion: 15,
@@ -24,15 +24,15 @@ describe("buildSubAgentMiddleware", () => {
       warningPct: 80,
     });
 
-    const stack = buildSubAgentMiddleware({ costCap: parentCostCap });
+    const stack = buildSubAgentMiddleware({ costAdvisory: parentCostAdvisory });
 
     expect(stack).toHaveLength(6);
-    expect(stack[4].name).toBe("CostCapSubAgentView");
+    expect(stack[4].name).toBe("CostAdvisorySubAgentView");
     expect(stack[5].name).toBe("ErrorHintsMiddleware");
   });
 
-  it("sub-agent cost cap view shares parent state", () => {
-    const parentCostCap = createCostCapMiddleware({
+  it("the sub-agent advisory view shares the parent's running total and has no tool hook", () => {
+    const parentCostAdvisory = createCostAdvisoryMiddleware({
       maxCostUsd: 1,
       inputPricePerMillion: 3,
       outputPricePerMillion: 15,
@@ -40,13 +40,13 @@ describe("buildSubAgentMiddleware", () => {
       warningPct: 80,
     });
 
-    const stack = buildSubAgentMiddleware({ costCap: parentCostCap });
+    const stack = buildSubAgentMiddleware({ costAdvisory: parentCostAdvisory });
 
-    expect(parentCostCap.runningCost).toBe(0);
+    expect(parentCostAdvisory.runningCost).toBe(0);
 
     const subView = stack[4];
     expect(subView.afterModel).toBeDefined();
-    expect(subView.wrapToolCall).toBeDefined();
+    expect(subView.wrapToolCall, "the advisory never blocks a tool; the runtime enforces the cap (Q-S3-3)").toBeUndefined();
     expect(subView.beforeAgent).toBeUndefined();
   });
 
@@ -92,7 +92,7 @@ describe("buildSubAgentMiddleware", () => {
   });
 
   it("orders the gate before the cost-cap view", () => {
-    const parentCostCap = createCostCapMiddleware({
+    const parentCostAdvisory = createCostAdvisoryMiddleware({
       maxCostUsd: 10,
       inputPricePerMillion: 3,
       outputPricePerMillion: 15,
@@ -101,16 +101,16 @@ describe("buildSubAgentMiddleware", () => {
     });
 
     const stack = buildSubAgentMiddleware({
-      costCap: parentCostCap,
+      costAdvisory: parentCostAdvisory,
       approvalGate: { policies: new Map(), toolServerMap: new Map() },
     });
 
-    // loop, budget, tool intent, truncation, approval gate, cost cap view,
+    // loop, budget, tool intent, truncation, approval gate, cost advisory view,
     // error hints. Hints AFTER the gate matches the parent nesting: the
     // gate's HITL interrupt stays outside the hints' try/catch (issue #255).
     expect(stack).toHaveLength(7);
     expect(stack[4].name).toBe("ApprovalGateMiddleware");
-    expect(stack[5].name).toBe("CostCapSubAgentView");
+    expect(stack[5].name).toBe("CostAdvisorySubAgentView");
     expect(stack[6].name).toBe("ErrorHintsMiddleware");
   });
 

package/src/activities/execute-deep-agent/__tests__/summarization-verification.test.ts

@@ -309,7 +309,7 @@ describe("middleware stack ordering in createDeepAgent", () => {
      *   const resolvedModel = request.model ?? await getChatModel();
      *
      * Since no model option is provided, the middleware uses request.model
-     * which is the same ChatAnthropic instance from setup.ts with the
+     * which is the same ChatAnthropic instance from turn-setup.ts with the
      * proxy baseURL. All summarization LLM calls route through the proxy.
      */