fix: keep sub-agents live while a message is still streaming

src/collect/subagents.ts

@@ -69,14 +69,21 @@ export async function agentContext(path: string) {
   // caller drives it as a floating promise, so a throw here would take the
   // whole TUI down over one bad line. Whatever was read stays usable.
   try {
-    // mid-flight: a tool call was issued (awaiting its result) or a result
-    // just arrived (awaiting the next turn). A final text-only assistant turn
-    // means the agent finished, so the mtime window alone governs it. Derived
-    // before the message contents below, which walk arbitrary blocks: a
-    // malformed one there must not cost a live agent its row.
+    // mid-flight: a tool call was issued (awaiting its result), a result just
+    // arrived (awaiting the next turn), or the message is still streaming.
+    // Claude Code writes each content block as its own entry and only the
+    // last carries a stop_reason; the earlier ones record an explicit null.
+    // So a text block with stop_reason null is mid-message — typically
+    // narration while the model generates a long tool call (a big Write),
+    // which can take minutes. A final text-only turn (end_turn) means the
+    // agent finished, so the mtime window alone governs it. Derived before
+    // the message contents below, which walk arbitrary blocks: a malformed
+    // one there must not cost a live agent its row.
     out.running =
       !!last &&
-      ((last.type === "assistant" && hasBlock(last.message, "tool_use")) ||
+      ((last.type === "assistant" &&
+        (hasBlock(last.message, "tool_use") ||
+          last.message?.stop_reason === null)) ||
         (last.type === "user" && hasBlock(last.message, "tool_result")));
     if (lastUsage) {
       out.model = lastUsage.model;

test/collect.test.ts

@@ -629,6 +629,31 @@ describe("transcript file scanning", () => {
     expect(ctx.running).toBe(false);
   });
 
+  // Claude Code writes each content block as its own entry; only the final
+  // one carries a stop_reason, the rest an explicit null. Narration streamed
+  // ahead of a long tool call must keep the agent live; end_turn ends it.
+  test("treats a mid-message text block as running", async () => {
+    const block = (text: string, stop_reason: string | null) => ({
+      type: "assistant",
+      message: {
+        model: "claude-opus-5",
+        usage: { input_tokens: 8 },
+        content: [{ type: "text", text }],
+        stop_reason,
+      },
+    });
+    const streaming = write(
+      "agent-3.jsonl",
+      jsonl([block("Now the recorder script.", null)]),
+    );
+    expect((await __test.agentContext(streaming)).running).toBe(true);
+    const finished = write(
+      "agent-4.jsonl",
+      jsonl([block("all done", "end_turn")]),
+    );
+    expect((await __test.agentContext(finished)).running).toBe(false);
+  });
+
   // collectRows runs from a floating promise, so a throw out of here would
   // take the whole TUI down over one malformed line in one agent transcript
   test("survives a malformed entry instead of throwing", async () => {