nooa-trace-viewer

name	nooa-trace-viewer
description	Run and use the NVIDIA OO Agents trace viewer — the web UI + OTLP receiver for browsing agent traces and eval results. Use when starting the viewer, importing/exporting/deleting traces, querying the viewer's REST API, or wiring an agent run so traces show up at localhost:5001.
compatibility	nooa package with the [viewer] extra (fastapi, uvicorn); nooa-cli for the `nooa` commands

Trace Viewer

The viewer (src/nooa/viewer/) is a single FastAPI app that is both the OTLP receiver agents send spans to and the web UI (React SPA) for browsing traces, LLM calls, and evaluation experiments. Traces persist in SQLite, so they survive restarts.

Start it

nooa start-dev                                  # http://localhost:5001
nooa start-dev --port 5002                      # custom port
nooa start-dev --db /path/to/other.db           # separate trace store
nooa start-dev --host 127.0.0.1                 # default host is 0.0.0.0

Stop with Ctrl-C. If deps are missing: uv sync --extra viewer. If the port is busy, the command reports the offending PID.

• Database resolution: --db > $NOOA_TRACE_DB > ~/.config/nooa/traces.db.
• Alternate entry point: python -m nooa.viewer (this path honors $NOOA_TRACE_VIEWER_PORT, default 5001; start-dev uses --port instead). DB default for the raw module is ./traces.db — prefer start-dev.

End-to-end workflow

nooa start-dev                                # terminal 1
uv run python my_agent.py                        # terminal 2 — auto-traced (see nooa-capturing-traces)
open http://localhost:5001/traces                # browse the session

Agents auto-detect the viewer at startup and stream spans; no code changes needed. If the agent ran on another host/port, set OTLP_ENDPOINT=http://<host>:<port>/v1/traces in the agent's environment.

Using the UI

Route	What you get
/traces	paginated session list — search, filter by experiment / batch ID, delete
/traces/view?session_id=…	trace detail: span-hierarchy event list, expand/collapse, text search
/eval	evaluation experiments with live status
/eval/experiment/{id}	experiment summary; drill into per-test traces

Trace-detail features worth knowing:

• Timeline — dual-canvas zoomable timeline of all events; filter sidebar — toggle event types, filter by span/agent/LLM/execution ID.
• LLM call view — full reconstructed conversations per call (/api/traces/{session_id}/calls under the hood).
• Annotations — score/label/comment/tags on a trace; quick 👍/👎.
• Playground — re-run an LLM turn with a different model/temperature and diff the output.
• Batch ID chip in the header (copyable) — groups imported/eval runs; filter the list with /traces?batch_id=….
• Keyboard: j/k navigate, / search, ? shows all shortcuts. URL captures view state, so links are shareable; add &embed=true to embed.

Import, export, delete

# Import OTLP .jsonl trace files (from exporters.jsonl(...) or a viewer export)
nooa import-traces ./traces/                          # dir (recursive) or single file
nooa import-traces run.jsonl --endpoint http://host:5001 --batch-id my-exp-v2

# Import a Harbor eval job directory
nooa import-harbor <job-dir>

# Export one session as .jsonl (round-trips through import-traces)
curl -o session.jsonl "http://localhost:5001/api/trace/export?session_id=<ID>"

# Delete
nooa delete-traces --batch-id my-exp-v2               # one batch via API
# DELETE /api/traces/{session_id}                        # one session
# DELETE /api/traces?confirm=true                        # everything

# Housekeeping for on-disk trace FILES (not the viewer DB)
nooa traces list
nooa traces delete --older-than 7

Import notes: only OTLP JSON lines ({"resourceSpans": [...]}) are accepted; the session ID is derived from the filename; already-imported sessions are skipped; omitted --batch-id auto-generates import_<timestamp>_<hex>.

REST API (for scripts and agents)

Base: http://localhost:5001. Most useful endpoints:

Endpoint	Purpose
GET /api/traces?page&limit&search&experiment&batch_id	paginated session list
GET /api/trace?session_id=&limit=&offset=	a session's raw OTLP spans (paginate with limit/offset)
GET /api/trace/export?session_id=	download session as .jsonl
GET /api/traces/{session_id}/calls	reconstructed LLM calls with full messages
GET /api/trace/resource?session_id=	resource attrs (incl. batch_id)
GET /api/experiments, GET /api/experiment/{id}/traces	experiment grouping
GET /api/eval/experiment/{id}/summary, .../tests	eval results
POST /v1/traces	OTLP JSON ingest (what agents/import use)
POST /v1/sync	block until the ingest queue is drained (use before reading back a just-finished run)
GET /api/explorer/*	structured analysis endpoints — use via TraceExplorerClient (see nooa-trace-explorer)

Ingestion is queued and written by a single background writer, so a GET immediately after a run may miss the tail — POST /v1/sync first when scripting.

Pitfalls

• nooa start-dev resolves the DB and sets NOOA_TRACE_DB in its own environment before starting uvicorn (same process); a separately launched python -m nooa.viewer without that var uses ./traces.db in the CWD — easy way to "lose" traces into a second DB.
• The viewer refuses to start (exit 1) if the SQLite DB is locked by another viewer instance — check for an already-running start-dev.
• Old docs mention a separate viewer package and a nooa viewer subcommand; the viewer now lives in src/nooa/viewer/ and nooa start-dev is the canonical command.

Related skills

• nooa-capturing-traces — how spans get produced and exported.
• nooa-trace-explorer — programmatic trace analysis (CLI + Python) on top of files or this viewer.