unisound-pe-missing-positive
ReviewAudited by ClawScan on May 16, 2026.
Overview
This is a coherent medical-record quality-check skill, but users should know it sends selected medical text to the HiVoice model API using an app key and can optionally save preprocessed text locally.
Install/use this only if you are allowed to process the medical records with the configured HiVoice MaaS endpoint. De-identify patient text before use, keep the app key secret, avoid untrusted --base URLs, verify any shared preprocessing helper, and do not enable --save-prepared unless you are comfortable storing a local copy of the record text.
Findings (5)
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
Medical record content may leave the local machine and be processed by the configured HiVoice-compatible model endpoint.
The skill posts LLM messages containing extracted medical-record fields to the HiVoice MaaS API. This is disclosed and central to the skill, but it is still a sensitive clinical-data transfer.
DEFAULT_LLM_BASE = "https://maas-api.hivoice.cn/v1" ... resp = _http_post(url, payload, headers, timeout=timeout)
Use only de-identified records, confirm the endpoint and data-handling agreement are appropriate, and avoid overriding --base to an untrusted URL.
The app key grants access to the model service and should be treated as a secret.
The skill uses a user-provided app key as a bearer token for the LLM API. The behavior is expected for this integration, but the registry metadata declares no primary credential.
headers = {"Authorization": f"Bearer {appkey}"}Provide a scoped key only at runtime, do not store it in the repository, and rotate it if it may have been exposed.
Running the multi-format path depends on whatever _shared/doc-preprocess implementation is present in the local skills tree.
The multi-format entrypoint imports a shared preprocessing module outside the listed skill files. SKILL.md discloses this dependency, but that helper is not part of the provided manifest.
_preprocess_dir = SKILLS_ROOT / "_shared" / "doc-preprocess" / "scripts" ... from preprocess import
Use a trusted copy of the shared preprocessing helper, or inspect it before using run.py for PDF/Office/CSV/JSON inputs.
If --save-prepared is used, a local file may contain sensitive medical-record text.
Although the documentation says inputs and intermediate results are not persisted, this user-selected debugging option writes the preprocessed record text to disk. The option is disclosed, but users should understand the privacy exception.
parser.add_argument("--save-prepared", action="store_true", help="保存预处理后的文本，便于调试。") ... prep_path.write_text(record_text, encoding="utf-8")Do not use --save-prepared with sensitive records unless the output directory is protected, and delete the prepared file when it is no longer needed.
A crafted or unusual record could cause an inaccurate “无缺陷/有缺陷” result.
User-supplied record fields are interpolated directly into the same LLM user message as the task instructions, so instructions embedded in a record could influence the model's QC answer.
【病历】\n体格检查：{pe}\n诊断：{dx}\n【质控结果】Treat outputs as advisory, have clinicians review results, and consider stronger prompt separation that clearly marks record text as data only.