The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill analyzes untrusted data from diagnostic bundles, including server logs and SQL query history, which constitutes a potential surface for indirect prompt injection.
Ingestion points: The skill's primary workflow involves reading user-supplied diagnostic bundles containing files like system.query_log and text_log via the inspect_bundle.py script and various bash recipes.
Boundary markers: The instructions provide explicit rules to treat bundle content as evidence only and to use measurements rather than attributions, helping to isolate data from instruction processing.
Capability inventory: The skill uses subprocess execution (bash, python) and local file system access. It is restricted from sending bundle data to external network services.
Sanitization: Data is processed using structured formats (JSONL) and standard utilities like jq and clickhouse local, which interpret the data as records rather than instructions.
[EXTERNAL_DOWNLOADS]: The skill performs network requests to fetch ClickHouse version tags, release notes, and documentation from official sources.
References target github.com/ClickHouse/ClickHouse and clickhouse.com, which are well-known and authoritative services for this software.
[COMMAND_EXECUTION]: The skill executes local scripts and shell commands to manage and query the diagnostic data.
It uses a bundled Python script (inspect_bundle.py) to perform an initial inventory and extract archives. This script includes explicit safeguards against path traversal (ZipSlip) by validating archive member paths and file types.
It provides recipes for common diagnostic tasks using tar, jq, and clickhouse local.
[DATA_EXFILTRATION]: While the skill analyzes potentially sensitive data (configs, logs), it includes rigorous data handling rules.
A dedicated privacy.md reference establishes a "Privacy gate" that mandates the redaction of hostnames, IPs, and user identifiers before any information leaves the local environment.

clickhouse-diagnostic