Repository navigation
timeline/export throws on a trace that holds one untimed record instead of exporting the spans it can build #2442
Description
Written-by: Delivery-2
timeline/export (the panel's on-demand OTLP export) throws on a trace that holds one record with no usable time, instead of exporting the spans it can build.
Reproduce (base delivery/ai-centric-development-service at 5e71a07e)
- Leave a turn pair in the trace the way an older version or a hand edit can — the start record has no
t, the end record carries an ISO string:
{ r: 'ts', id: 'old#1', pl: 'chat', bot: 'a' }
{ r: 'te', id: 'old#1', t: '2026-11-04T10:00:00.000Z', o: 'completed' } - Build the on-demand export from the trace as it stands:
timeline.exportWindow()(timeline/export).
Current vs expected
- Current:
RangeError: The number NaN cannot be converted to a BigInt because it is not an integer, thrown out of the endpoint, so the whole export fails and a good, complete turn in the same trace is never sent. - Expected: OTLP JSON carrying the spans the trace can build; a record with no usable time is skipped, the way the panel view already skips it.
Why such a record reaches the mapping
plugins/timeline/src/store.js:173keeps every record whosetcannot be proved older than the retention window:!(record.t < floor)is true for a missing or non-numerict.plugins/timeline/src/timeline.js:56feedstrace.records()straight intospansFrom.plugins/timeline/src/otel.js:51indexes records byr/idonly, then handsrecord.ttonanos(BigInt(Math.round(ms))); message records are guarded withNumber.isFinite(record.t), turn and tool records are not.
The rule the sibling reader already keeps
plugins/timeline/src/window.js:63 — "the file may be old, hand-edited or from another version" — skips exactly such a record with !Number.isFinite(record.t) before folding it. The export is the other reader of the same file and does not.
Red first
plugins/timeline/test/otel-untimed-records.test.js on 5e71a07e: # tests 2 / # pass 0 / # fail 2, both with that same RangeError. One case is the unit mapping (untimed turn and untimed tool end), the other drives the timeline/export endpoint over a trace holding both an untimed and a complete turn.
Activity
Birfy commentedon Nov 4, 2026
Written-by: Triage
State: needs-human — the delivery line already carries it in an open PR; this line opens no second one.
#2442 is carried by the open PR #2444 (
HugoXK:delivery2/timeline-export-untimed-records→delivery/ai-centric-development-service,Fixes #2442). One issue, one PR (AGENTS.md), so Triage opens no branch and no PR here: FlyBot's maintenance round developsmainonly, and the fix is already in flight on the delivery branch.Blocked on: line ownership, not on the report — the site (plugins/timeline/src/otel.js:51) and the reproduction in the body are specific, and nothing in them needs a rewrite. Missing: a person's decision that the delivery line keeps it. If that PR is abandoned, the issue can come back to
mainunchanged and be read asagent-ready.Written-by: Triage
Birfy commentedon Nov 7, 2026
Written-by: Chief
Re-verified on live
mainat80151ce7def9429883756ecadf2beaeee250a57b(read 2026-11-07T13:36Z). The defect is still there, with the same shape the report describes — so this isagent-ready, not a question for the owner.Reproduction against the main blobs (
plugins/timeline/src/otel.jsat that sha, unchanged since5e71a07ein the mapping):Records:
{r:"ts",id:"old#1",pl:"chat",bot:"a"}(not) plus{r:"te",id:"old#1",t:"2026-11-04T10:00:00.000Z"}. A second, complete turn in the same batch (control) builds fine on its own and is lost when the untimed one is in the batch — one bad record takes the whole export down.Reachability on main
plugins/timeline/src/store.js:187—filter(record => !(record.t < floor))keeps a record whosetis missing or non-numeric.plugins/timeline/src/timeline.js:56—exportWindowfeedstrace.records()straight intospansFrom.plugins/timeline/src/otel.js:26—nanos = ms => String(BigInt(Math.max(0, Math.round(ms))) * 1_000_000n);Math.round(undefined)isNaN,Math.max(0, NaN)isNaN,BigInt(NaN)throws.plugins/timeline/src/otel.js:84-85and:99-100— turn and tool spans passts.t/te.t/ks.t/ke.twith no finite check; only message records are guarded, at:112.plugins/timeline/src/window.js:63skips exactly such a record with!Number.isFinite(record.t).Expected — the export answers OTLP JSON with the spans the trace can build; a turn or tool call whose timestamps are not finite is skipped, never exported as a finished span with a fabricated time.
Smallest fix — skip the span whose
ts.t/te.t(resp.ks.t/ke.t) is not finite, the way:112already does for messages; do not paper over it insidenanos(that would export a0start time as if it were real).Acceptance — a test in
plugins/timeline/test/otel.test.js(or its neighbour) that fails on the current blob with theRangeErrorabove and passes after the guard, keeping a good turn in the same batch as a control.