Qwen Code 0.25.0 SDK file: anonymized native sample
The official CLI called a fixed local llama.cpp/Qwen2.5 model in Podman with networking
disabled. provenance.json records source and version; original SHA-256
matches the container source record.
Retain SDK _rawAttributes arrays, _spanContext, timestamps, kind and permitted attributes.
Replace session and trace/span IDs consistently; exclude prompts, response bodies, URLs,
accounts and host information. Metrics retain only structure needed to distinguish them
from calls. The file contains 25 consecutive multiline JSON objects, rather than JSONL.
| Source | Count | Input (including cache) | Output | Cache read | thoughts | Total |
|---|---|---|---|---|---|---|
| interaction main loop | 1 | 10,226 | 2 | 0 | 0 | 10,228 |
| standalone managed-auto-memory-extractor | 1 | 5,639 | 556 | 3 | 0 | 6,195 |
| All CLI requests | 2 | 15,865 | 558 | 3 | 0 | 16,423 |
native.jsonl retains the first user record’s shape from this same native session, without message.parts contents, and its only assistant usage record: the main-loop 10,228 tokens. The first record tests manual-root routing. cli-stats.json contains permitted final CLI statistics. The API-response log and llm_request span observe the same request: select the span. Do not add HTTP, interaction or metrics values. Background spans have no parentSpanContext; a shared session.id does not establish a parent/child relationship.
Unreturned cache write/provider remain unknown. thoughts=0 does not verify inclusion rules for positive reasoning or other providers/versions. Select one native/exported representation per verified host/user/session/local-day partition. Do not infer call IDs from equal timestamps or tokens. Coverage before export was enabled is limited. Boundary, mixed-version, error and rollback cases are synthetic variants, without native SDK verification.