fix(nooa-bench): always persist traces and dump a local trajectory - #69

Problem

A Harbor run could finish with no trajectory stored anywhere on disk, which blocks post-hoc failure analysis. This was hit in practice — reported as:

"it appears that nooa does not have locally stored full trajectory? I could not find it in, say, nooa-base_ultra_on/astropy__astropy-14598__7oCi6qh/agent"

Two independent causes.

1. The exporter choice was mutually exclusive

_setup_tracing() installed either the streaming journal exporter or the JSONL file exporter:

if probe_otlp_endpoint(endpoint):
    enable_tracing(exporters=[nemo_exporters.journal(endpoint=endpoint)])
else:
    enable_tracing(exporters=[nemo_exporters.jsonl(TRACES_DIR)])

Benchmark configs set OTLP_ENDPOINT, so those runs streamed spans to the viewer and wrote nothing locally. If the viewer was down, restarted, or the network blipped, the trajectory was simply gone.

2. Nothing wrote per-turn history where anyone looks for it

The per-task agent/ directory held only nooa_bench.log and result.json — and result.json carries just the final response. So the directory named in the report above was genuinely empty of trajectory data under either branch, since JSONL goes to /logs/artifacts/traces/.

Changes

• The JSONL file exporter is now unconditional; the journal exporter is added alongside it when OTLP_ENDPOINT is reachable. Live viewing no longer costs the durable copy.
• Add _write_trajectory(), dumping the agent's full event history to agent/trajectory.json.
• extra_resource_attrs (eval.model, eval.agent_type) now apply in both cases, not just the streaming one, so file-only runs stay attributable.

Notes

Trajectory serialisation failures are logged and swallowed — a debug artifact must never fail a task. The OTLP spans under /logs/artifacts/traces/ remain the canonical record; trajectory.json is the greppable companion.

🤖 Generated with Claude Code