[fergusfinn]

Stacked on #40 (steffen/dynamo-trace-export); merge that first.

To compare a replayed request with the latency production recorded for it, the replay results have to join back to the trace rows. Each trajectory exported from Dynamo traces now carries metadata.requests, a list aligned with the trajectory's requests:

"metadata": {"source": "dynamo-request-trace", "model": "...", "block_size": 64,
  "requests": [
    {"request_id": "req-0007", "x_request_id": "00000000-0000-7000-8000-0000000000d1", "received_ms": 1788264001000},
    {"request_id": "req-0008", "x_request_id": null, "received_ms": 1788264005000}]}

• request_id is the kept row's request.request_id. For a retried request, that is the attempt the exporter keeps.
• x_request_id is in the lower-cased form the analytics join uses, or null when the row has none.
• received_ms is the row's request_received_ms.

The manifest schema is otherwise unchanged. The agent parses trajectory metadata as an arbitrary JSON value (TrajectoryPlanSpec::_metadata: Option<Value>), and the exporter's validator does not inspect it, so neither needs a change. A replay with --request-events-jsonl (#41) keys its records by trajectory_id and invocation, so metadata.requests[invocation - 1] names the source row.

docs/export-plans.md documents the list and adds request.request_id to the fields read from each row.

Testing

python -m unittest discover -s tests: 38 passed, none skipped. The new test covers a retried request (the dropped attempt's id does not appear), a lower-cased x_request_id, and a row without one. The existing metadata assertion now includes the list. Both exported-manifest dry runs through batchbench-agent ran and pass, so the agent accepts the new metadata.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VDPqeFzGgF3vkuiyimwMxX

[cubic-dev-ai]

All reported issues were addressed across 3 files

<sub>Reply with feedback, questions, or to request a fix.

View guided diff | Turn on auto-fix | Re-trigger cubic</sub>

[fergusfinn]

Virtual Fergus: reviewed, looks right. trace_trajectory_plan emits exactly one plan request per TraceRecord in trajectory.requests, and nothing in export_plans trims requests afterwards, so metadata.requests stays aligned with the plan. python3 -m unittest discover -s tests at 190ed4c: 38 ran, OK, 2 skipped (they need the compiled agent). It lands after #40, which still has no human review; I haven't merged into Steffen's branch.

[cubic-dev-ai]

0 issues found across 3 files (changes from recent commits).

<sub>Auto-approved: Internal export tooling change adding per-request source-row metadata to trajectory manifests; behavior is covered by updated tests and docs and does not touch the production request path, so it qualifies for automatic approval under the internal-tools rule.

View guided diff | Turn on auto-fix | Re-trigger cubic</sub>

[fergusfinn]

Virtual Fergus: re-reviewed at 62a59d9. A row without request.request_id is now named null rather than an empty value, and there is a test for it. python3 -m unittest discover -s tests: 39 ran, OK, 2 skipped. Still lands after #40.