Skip untimed trace records in the OTLP export instead of throwing

changelog.d/2442-timeline-export-untimed-records.fixed.md

@@ -0,0 +1 @@
+The timeline's on-demand OpenTelemetry export now skips a trace record that carries no usable timestamp instead of throwing `RangeError: The number NaN cannot be converted to a BigInt`, so one stale or hand-edited record no longer costs the export of every complete turn beside it ([#2442](https://github.com/ThinkFlowLab/FlyBot/issues/2442)).

plugins/timeline/src/otel.js

@@ -42,14 +42,18 @@ const partyAttrs = (prefix, party) => {
  * Records -> OTLP span objects. Only CLOSED spans are built: a turn needs its `te` and a tool
  * call its `ke`. An open turn is never exported as a finished one — the exporter carries it
  * with `pendingRecords` until its end arrives. Deliveries are instant spans.
+ *
+ * A record whose `t` is not a finite number is skipped: a segment can come from an older version
+ * or a hand edit, and `fold` in src/window.js keeps the same rule. Without it `nanos` would turn
+ * the missing time into `BigInt(NaN)` and one stale record would throw the whole export.
  */
 export function spansFrom(records, { names = {} } = {}) {
   const starts = new Map()
   const ends = new Map()
   const toolStarts = new Map()
   const toolEnds = new Map()
   for (const record of records) {
-    if (!record || typeof record !== 'object') continue
+    if (!record || typeof record !== 'object' || !Number.isFinite(record.t)) continue
     if (record.r === 'ts') starts.set(record.id, record)
     else if (record.r === 'te') ends.set(record.id, record)
     else if (record.r === 'ks') toolStarts.set(record.id, record)

plugins/timeline/test/otel-untimed-records.test.js

@@ -0,0 +1,48 @@
+// A trace segment outlives the version that wrote it: the store keeps every record whose `t` is
+// not provably older than the retention window, and `fold` in src/window.js checks
+// `Number.isFinite(record.t)` before it trusts one. The OTLP mapping reads the same records, so it
+// has to keep the same rule: an untimed turn or tool call is skipped, not converted to a span.
+import assert from 'node:assert/strict'
+import { test } from 'node:test'
+import { spansFrom } from '../src/otel.js'
+import { createTimeline } from '../src/timeline.js'
+import { mockAdapter } from './mock-adapter.js'
+
+// The shape a segment written by an earlier version (or edited by hand) can carry: an ISO string
+// for the end, nothing at all for the start.
+const untimed = [
+  { r: 'ts', id: 'old#1', pl: 'chat', bot: 'a' },
+  { r: 'te', id: 'old#1', t: '2026-11-04T10:00:00.000Z', o: 'completed' },
+]
+const timed = [
+  { r: 'ts', t: 1_000, id: 'new#1', pl: 'chat', bot: 'a' },
+  { r: 'te', t: 1_600, id: 'new#1', o: 'completed' },
+]
+
+test('a turn or tool record with no usable time is skipped, never converted to a span', () => {
+  assert.deepEqual(spansFrom(untimed), [])
+  assert.deepEqual(spansFrom([
+    { r: 'ks', t: 1_100, id: 'c1', tid: 'new#1', nm: 'web_search', bot: 'a' },
+    { r: 'ke', t: '2026-11-04T10:00:01.000Z', id: 'c1', tid: 'new#1', ok: true },
+  ]), [])
+  const spans = spansFrom([...untimed, ...timed])
+  assert.deepEqual(spans.map(span => span.name), ['invoke_agent'])
+  assert.equal(spans[0].startTimeUnixNano, '1000000000')
+})
+
+test('timeline/export still answers OTLP JSON when the trace holds an untimed record', () => {
+  const adapter = mockAdapter()
+  const timeline = createTimeline(adapter, { persist: false, now: () => 5_000 })
+  adapter.feed('turn.start', { botId: 'a', place: 'chat', sessionId: 's1', turnId: 's1#1', turn: 1, time: 4_100 })
+  adapter.feed('turn.end', { botId: 'a', place: 'chat', sessionId: 's1', turnId: 's1#1', turn: 1, time: 4_900, outcome: 'completed' })
+  for (const record of untimed) timeline.trace.append(record)
+  // The store keeps the untimed records (their `t` cannot be proved older than the window); the
+  // export is the reader that has to cope with them.
+  const kept = timeline.trace.records()
+  assert.ok(kept.some(record => record.r === 'ts' && record.id === 'old#1'))
+  assert.ok(kept.some(record => record.r === 'te' && record.id === 'old#1'))
+  const spans = timeline.exportWindow().resourceSpans[0].scopeSpans[0].spans
+  assert.equal(spans.length, 1)
+  assert.equal(spans[0].name, 'invoke_agent')
+  timeline.dispose()
+})