Trace Explorer

The trace explorer (src/nooa/trace_explorer/) is a CLI + Python library purpose-built for agent-driven root-cause analysis of traces. Where the viewer is a web UI for humans, the explorer emits LLM-friendly text (or JSON) and supports search, diffing, and experiment-wide aggregation. It is not itself an Agent — it's a tool you (the coding agent) drive.

Data sources: local OTLP .jsonl trace files, or a running viewer (per-session or per-experiment). It never opens traces.db directly — viewer data goes through the HTTP API.

Exploration strategy: progressive disclosure

Start broad, drill in. Overview → errors → session → turn → search.

uv run trace-explorer trace.jsonl                    # 1. overview: call graph, sessions, pass/fail
uv run trace-explorer trace.jsonl --errors           # 2. all errors (or --first-error)
uv run trace-explorer trace.jsonl -s 278a10          # 3. one session's turns (IDs can be 6-char prefixes)
uv run trace-explorer trace.jsonl -s 278a10 -t 0     # 4. one turn: context window → LLM output → execution result
uv run trace-explorer trace.jsonl --search "Timeout" # 5. regex/text search across everything

Output includes navigation hints for the next drill-down. Other useful flags: --timeline, --eval (evaluation context), --json (structured output), -v (full detail), -q (suppress parser warnings), --raw <span_id> (raw span JSON), --diff other.jsonl, --api-help (prints the Python API guide).

Against a running viewer

uv run trace-explorer --viewer http://localhost:5001 --session-id <ID>            # overview
uv run trace-explorer --viewer http://localhost:5001 --session-id <ID> --errors
uv run trace-explorer --viewer http://localhost:5001 --session-id <ID> -s <SID> -t 0
uv run trace-explorer --viewer http://localhost:5001 --session-id <ID> --span-id <SPAN>  # jump to a span

Experiment-level analysis (eval runs)

uv run trace-explorer --viewer <URL> --experiment <ID>            # pass/fail rates + drill-in commands
uv run trace-explorer --viewer <URL> --experiment <ID> --errors   # Python exceptions across failed sessions
uv run trace-explorer --viewer <URL> --experiment <ID> --failures # wrong-answer eval failures (no exception)
uv run trace-explorer --viewer <URL> --experiment <ID> --search "pattern"
uv run trace-explorer --viewer <URL> --experiment <ID> --json     # machine-readable: all tests + session_ids

--errors = crashes; --failures = wrong answers. For a full sweep: get --json first, then iterate sessions with --session-id.

Python API — full parse (TraceExplorer)

from nooa.trace_explorer import TraceExplorer

trace = await TraceExplorer.from_file("traces/<session_id>.jsonl")
# or: trace = await TraceExplorer.from_viewer("http://localhost:5001", "<session-id>")
# or: explorers = await TraceExplorer.load_experiment_sessions(url, "<experiment>")  # {session_id: TraceExplorer}

print(await trace.get_overview())          # text for LLM context
print(await trace.get_errors())
print(await trace.get_session("278a10"))
print(await trace.get_turn("278a10", 0))
print(await trace.search("ValidationError"))
print(await trace.get_timeline())
print(await trace.find_first_error())
print(await TraceExplorer.diff(trace, other))   # first divergence, call-graph diff

Every text method has a structured twin returning dataclasses with .todict(): getoverviewdata(), getsessiondata(), getturndata(), geterrorsdata(), searchdata(), gettimelinedata(), findfirsterrordata(), getevalcontextdata(), comparedata(). Stats helpers: .agentcount, .maxagentdepth, getmethodcounts(), getrecursionpattern().

Python API — thin client for huge traces (TraceExplorerClient)

For very large traces (>100k spans / GB-scale), don't download spans — delegate analysis to the viewer server (/api/explorer/*, results cached server-side):

from nooa.trace_explorer import TraceExplorerClient

client = TraceExplorerClient("http://localhost:5001", "<session-id>")
await client.get_summary()          # instant DB query: span count, duration, error count
await client.get_agent_spans()      # agent-tree skeleton only
await client.get_error_spans()
await client.search_fast("query")   # SQLite FTS5 full-text search
await client.get_session_fast("abc123", span_id="<full_span_id>")
await client.get_turn_fast("abc123", span_id="<full_span_id>", turn_index=0)
await client.get_overview()         # server builds + caches the full explorer

Prefer the thin client when the trace is huge, when exploring interactively (cache makes repeats instant), or when you only need one session's subtree.

Typical workflows

• Failing eval run: --experiment <ID> → note failed tests' session_ids → --session-id <ID> --errors → -s <SID> -t <N> to read the exact LLM turn that went wrong.
• Agent behaved oddly, no exception: overview → getsession to scan turns → getturn to read the context window the model actually saw (bad prompts are usually visible right there).
• Regression between two runs: trace-explorer a.jsonl --diff b.jsonl → shows first divergence point and call-graph differences.
• "Where did X come from?": --search "X" then --raw <span_id> for the exact span.

Tips

• Session IDs abbreviate to 6 characters everywhere.
• --install-skill copies the bundled skill to ~/.claude/skills/trace-explorer/ (there is an in-repo copy at src/nooa/trace_explorer/skill/SKILL.md).
• Entry points: trace-explorer console script or python -m nooa.trace_explorer — there is no nooa subcommand for it.
• Eval-artifact files are rejected by from_file (.006eval. by filename; *.noo-eval.jsonl fails span validation) — point it at the session trace, not the eval sidecar.
• Reasoning content is shown by default; hide with --no-reasoning.

Related skills

• nooa-capturing-traces — produce the .jsonl files this tool reads.
• nooa-trace-viewer — the web UI and the server backing --viewer / thin-client modes.