fix(framework): keep Codex LLM spans non-empty during backgrounded commands

apps/framework/scripts/upload-braintrust.test.ts

@@ -278,6 +278,54 @@ describe('logTranscript', () => {
     });
   });
 
+  it('starts LLM spans during a backgrounded command at the previous edge', () => {
+    const root: RecordedSpan = { children: [] };
+    const message = (requestId: string, s: number) => ({
+      type: 'message' as const,
+      role: 'assistant' as const,
+      content: requestId,
+      ts: (100 + s) * 1000,
+      requestId,
+    });
+    logTranscript(recorder(root), {
+      prompt: 'go',
+      agentReport: '',
+      checks: [],
+      judgeCalls: [],
+      passed: true,
+      modelId: 'm',
+      startTime: 100,
+      toolLabels: [],
+      transcript: [
+        {
+          type: 'tool_call',
+          name: 'Bash',
+          input: {},
+          id: 't1',
+          ts: 104_000,
+          resultTs: 120_000,
+          requestId: 'r1',
+        },
+        message('r2', 8),
+        message('r3', 12),
+        message('r4', 25),
+      ],
+    });
+    expect(
+      root.children[0]?.children.map(({ name, start, end }) => [
+        name,
+        start,
+        end,
+      ])
+    ).toEqual([
+      ['m', 100, 104],
+      ['Bash', 104, 120],
+      ['m', 104, 108],
+      ['m', 108, 112],
+      ['m', 120, 125],
+    ]);
+  });
+
   it('ends setup at a leading non-assistant message and starts task there', () => {
     const root: RecordedSpan = { children: [] };
     logTranscript(recorder(root), {

apps/framework/scripts/upload-braintrust.ts

@@ -424,7 +424,7 @@ export interface SpanSink {
  * https://github.com/braintrustdata/braintrust-coding-agent-plugins/blob/c0346dcdb16ae9f136b3abf83efb6458191f9300/bt-daemon/src/translate/claude.rs#L1173-L1300
  * https://github.com/braintrustdata/braintrust-spec/blob/b068e39112e081e45b6070e035877f1e2e83f9b7/skills/instrumentation-spec/references/features/eval-spans.md#trace-shape
  *
- * An LLM span starts where the previous span ended, since the agent only
+ * An LLM span starts where the latest earlier span ended, since the agent only
  * writes a request's timestamps once its blocks finish generating. Braintrust's
  * Claude Code translator starts it at the first block instead, which leaves the
  * generation time as a gap. Their Codex translator does what we do.
@@ -498,6 +498,15 @@ export function logTranscript(
   });
   task.log({ input: row.prompt, output: row.agentReport });
   let latestTime = taskStart;
+  // A backgrounded command can end after later requests, so each request
+  // starts at the latest edge before it.
+  const edges = taskStart === undefined ? [] : [taskStart];
+  const mark = (time: number | undefined) => {
+    if (time !== undefined) {
+      edges.push(time);
+    }
+    latestTime = latest(latestTime, time);
+  };
   let toolIndex = 0;
   // Seeded with the prompt, which the transcript omits.
   const history: Record<string, unknown>[] = row.prompt
@@ -522,10 +531,8 @@ export function logTranscript(
       open.usage ??= part.usage;
       return open;
     }
-    const startTime =
-      latestTime === undefined || at === undefined
-        ? (latestTime ?? at)
-        : Math.min(latestTime, at);
+    const before = at === undefined ? [] : edges.filter((edge) => edge <= at);
+    const startTime = before.length ? Math.max(...before) : (at ?? latestTime);
     const request: Request = {
       id,
       span: task.startSpan({
@@ -565,7 +572,7 @@ export function logTranscript(
           ? `${message.content}\n${part.content}`
           : part.content;
       }
-      latestTime = latest(latestTime, at);
+      mark(at);
       continue;
     }
     if (part.requestId) {
@@ -586,7 +593,8 @@ export function logTranscript(
     }
     const label = row.toolLabels[toolIndex++];
     const endTime = sec(part.resultTs) ?? at;
-    latestTime = latest(latestTime, endTime);
+    mark(at);
+    mark(endTime);
     const span = task.startSpan({
       name: label ? `${part.name}: ${label}` : part.name,
       type: 'tool',