diff --git a/mcp/sicore-a2a-adapter/src/server.ts b/mcp/sicore-a2a-adapter/src/server.ts
index b65ccd66..562a8d78 100644
--- a/mcp/sicore-a2a-adapter/src/server.ts
+++ b/mcp/sicore-a2a-adapter/src/server.ts
@@ -6,7 +6,10 @@ import { createToolHandler, TOOL_DEFINITIONS } from "./tools.js";
 
 export const SERVER_INSTRUCTIONS = [
   "Use siclaw_investigate for operational questions that require the configured Siclaw SRE agent.",
-  "When it returns a non-terminal task, keep the current turn open and call siclaw_wait_task with the same task_id until terminal, unless the user requests fire-and-forget, asks to stop, or the overall investigation deadline is exhausted.",
+  "Independent questions or hypotheses can run concurrently: submit each as its own task with a distinct (or omitted) context_id instead of investigating them serially.",
+  "When unsure what the agent can reach, first ask it to list its bound clusters and data sources.",
+  "For diagnostic questions, ask it to cite the data sources it consulted and its confidence in the conclusion.",
+  "When a call returns a non-terminal task, keep the current turn open and call siclaw_wait_task with the same task_id until terminal, unless the user requests fire-and-forget, asks to stop, or the overall investigation deadline is exhausted.",
   "Never resubmit the same question merely because a task is still working.",
   "Use siclaw_list_tasks to recover server-side tasks after a client restart.",
 ].join(" ");
diff --git a/mcp/sicore-a2a-adapter/src/tools.test.ts b/mcp/sicore-a2a-adapter/src/tools.test.ts
index e0e4fcdf..d5c29c56 100644
--- a/mcp/sicore-a2a-adapter/src/tools.test.ts
+++ b/mcp/sicore-a2a-adapter/src/tools.test.ts
@@ -80,23 +80,63 @@ describe("tool contract", () => {
     expect(result.structuredContent).toMatchObject({ state: "completed", result: "root cause" });
   });
 
-  it("keeps a working response compact until the terminal result", async () => {
+  it("keeps a working response compact but surfaces a progress tail", async () => {
     const api = fakeApi();
     vi.mocked(api.getTask).mockResolvedValue({
       ...task(),
-      result: "partial evidence that should not be repeated",
+      result: "partial evidence worth relaying to the caller",
       updated_at: "2026-07-18T00:00:00.000Z",
     });
     const result = await createToolHandler(api)("siclaw_get_task", { task_id: "task-1" });
-    expect(result.content[0].text).not.toContain("partial evidence that should not be repeated");
+    expect(result.content[0].text).toContain("progress_tail (latest excerpt):");
+    expect(result.content[0].text).toContain("partial evidence worth relaying to the caller");
     expect(result.content[0].text).toContain("siclaw_wait_task");
     expect(result.structuredContent).toMatchObject({
       state: "working",
-      progress_chars: 44,
+      progress_chars: 45,
+      progress_tail: "partial evidence worth relaying to the caller",
     });
     expect(result.structuredContent).not.toHaveProperty("result");
   });
 
+  it("bounds the progress tail to the last 400 runes of a long partial report", async () => {
+    const api = fakeApi();
+    const partial = "x".repeat(700) + "the latest narrative sentence";
+    vi.mocked(api.getTask).mockResolvedValue({
+      ...task(),
+      result: partial,
+      updated_at: "2026-07-18T00:00:00.000Z",
+    });
+    const result = await createToolHandler(api)("siclaw_get_task", { task_id: "task-1" });
+    const tail = (result.structuredContent as { progress_tail: string }).progress_tail;
+    expect(tail.startsWith("…")).toBe(true);
+    expect(tail.endsWith("the latest narrative sentence")).toBe(true);
+    expect(Array.from(tail).length).toBe(401);
+    expect(result.structuredContent).toMatchObject({ progress_chars: 729 });
+    expect(result.content[0].text).not.toContain(partial);
+  });
+
+  it("omits the progress tail from terminal responses", async () => {
+    const result = await createToolHandler(fakeApi())("siclaw_get_task", { task_id: "task-1" });
+    expect(result.structuredContent).not.toHaveProperty("progress_tail");
+    expect(result.structuredContent).toMatchObject({ state: "completed", result: "root cause" });
+  });
+
+  it("withholds the progress tail from task listings", async () => {
+    const api = fakeApi();
+    vi.mocked(api.listTasks).mockResolvedValue({
+      tasks: [{ ...task(), result: "partial investigation narrative" }],
+      total_size: 1,
+      page_size: 20,
+      next_page_token: null,
+    });
+    const result = await createToolHandler(api)("siclaw_list_tasks", {});
+    const row = (result.structuredContent as { tasks: Record<string, unknown>[] }).tasks[0];
+    expect(row).not.toHaveProperty("progress_tail");
+    expect(row).not.toHaveProperty("result");
+    expect(result.content[0].text).not.toContain("partial investigation narrative");
+  });
+
   it("keeps the submitted task_id when the bounded wait fails after submission", async () => {
     const api = fakeApi();
     vi.mocked(api.waitForTask).mockRejectedValue(new Error("Sicore A2A request timed out"));
diff --git a/mcp/sicore-a2a-adapter/src/tools.ts b/mcp/sicore-a2a-adapter/src/tools.ts
index 4ea8677e..2870fac6 100644
--- a/mcp/sicore-a2a-adapter/src/tools.ts
+++ b/mcp/sicore-a2a-adapter/src/tools.ts
@@ -12,7 +12,7 @@ const STATUS_TO_A2A: Record<string, string> = {
 export const TOOL_DEFINITIONS = [
   {
     name: "siclaw_investigate",
-    description: "Ask the configured Siclaw SRE agent to investigate an operational question. This creates an asynchronous Sicore A2A task. Reuse context_id to continue the same investigation. If the returned task is not terminal, do not submit it again: call siclaw_wait_task until it finishes unless the user explicitly requested fire-and-forget. The configured A2A key fixes which Siclaw agent is used.",
+    description: "Ask the configured Siclaw SRE agent to investigate an operational question. This creates an asynchronous Sicore A2A task. Reuse context_id to continue the same investigation; tasks with distinct context_ids run in parallel server-side, so submit independent hypotheses concurrently. If the returned task is not terminal, do not submit it again: call siclaw_wait_task until it finishes unless the user explicitly requested fire-and-forget. The configured A2A key fixes which Siclaw agent is used.",
     inputSchema: {
       type: "object",
       properties: {
@@ -41,7 +41,7 @@ export const TOOL_DEFINITIONS = [
   },
   {
     name: "siclaw_wait_task",
-    description: "Wait for an existing Siclaw investigation without creating another task. Use this as the same-turn watchdog: call it repeatedly while the task is non-terminal, unless the user asks to stop or the overall investigation deadline is exhausted. Working responses are compact; the full report is returned once the task reaches a terminal state.",
+    description: "Wait for an existing Siclaw investigation without creating another task. Use this as the same-turn watchdog: call it repeatedly while the task is non-terminal, unless the user asks to stop or the overall investigation deadline is exhausted. Working responses are compact — a short progress_tail excerpt of the investigation narrative plus counters; the full report is returned once the task reaches a terminal state.",
     inputSchema: {
       type: "object",
       properties: {
@@ -110,15 +110,32 @@ type ToolResult = {
 type TaskToolView = Omit<SiclawTask, "result"> & {
   result?: string | null;
   progress_chars: number;
+  progress_tail?: string;
   wait_error?: string;
 };
 
-function taskView(task: SiclawTask, includeTerminalResult: boolean): TaskToolView {
+// Working responses withhold the partial report for compactness, but expose a
+// bounded tail so pollers can relay what the investigation is doing right now.
+const PROGRESS_TAIL_MAX_RUNES = 400;
+
+function progressTail(text: string | null | undefined): string | undefined {
+  if (!text) return undefined;
+  const runes = Array.from(text);
+  if (runes.length <= PROGRESS_TAIL_MAX_RUNES) return text;
+  return "…" + runes.slice(-PROGRESS_TAIL_MAX_RUNES).join("");
+}
+
+// includeReport gates all report content — the full result on terminal tasks
+// AND the progress_tail excerpt on working ones. Listings pass false: they
+// exist to recover task IDs and must never carry investigation text.
+function taskView(task: SiclawTask, includeReport: boolean): TaskToolView {
   const { result, ...summary } = task;
+  const tail = includeReport && !task.is_terminal ? progressTail(result) : undefined;
   return {
     ...summary,
-    progress_chars: result?.length ?? 0,
-    ...(includeTerminalResult && task.is_terminal ? { result } : {}),
+    progress_chars: result ? Array.from(result).length : 0,
+    ...(tail !== undefined ? { progress_tail: tail } : {}),
+    ...(includeReport && task.is_terminal ? { result } : {}),
   };
 }
 
@@ -165,8 +182,11 @@ function taskText(task: SiclawTask, waitError?: string): string {
   if (waitError) lines.push(`wait_error: ${waitError} (status polling failed, but the task was already created)`);
   if (task.is_terminal && task.result) lines.push("", task.result);
   if (!task.is_terminal) {
-    const progressChars = task.result?.length ?? 0;
-    if (progressChars > 0) lines.push(`progress_chars: ${progressChars} (partial report withheld until terminal)`);
+    const progressChars = task.result ? Array.from(task.result).length : 0;
+    if (progressChars > 0) {
+      lines.push(`progress_chars: ${progressChars} (full report is withheld until terminal)`);
+      lines.push("progress_tail (latest excerpt):", progressTail(task.result)!);
+    }
     lines.push("", "The investigation is still running. Call siclaw_wait_task with the task_id; do not submit the same investigation again.");
   }
   return lines.join("\n");