The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes production telemetry data (via the TELEMETRY_DATA parameter or docs/telemetry/production/latest.json) which is considered untrusted external input. This data is interpolated directly into the prompts for three core sub-agents (qe-metrics-optimizer, qe-defect-predictor, and qe-root-cause-analyzer) in steps/02-core-agents.md. While the prompts use clear delimiters, there is no evidence of sanitization, escaping, or filtering of the content before interpolation, which allows for potentially malicious instructions embedded in telemetry logs to influence agent behavior.
Ingestion points: TELEMETRY_DATA input parameter and the docs/telemetry/production/latest.json file monitored in Step 1.
Boundary markers: The skill uses delimiters such as === PRODUCTION TELEMETRY START === and === DORA METRICS DATA START === to separate data from instructions.
Capability inventory: The swarm has significant capabilities including spawning multiple sub-agents via the Task tool, writing reports to the filesystem using the Write tool, and persisting state via memory tools.
Sanitization: No explicit sanitization or instruction to escape special characters in the telemetry data is provided.
[COMMAND_EXECUTION]: The skill instructions require the execution of a CLI tool named ruflo via npx for memory search and storage operations.
Evidence: steps/01-flag-detection.md and steps/07-learning-persistence.md contain commands like npx --no-install ruflo memory search and npx --no-install ruflo memory store. While the use of --no-install prevents the download of new packages, ruflo is a non-standard third-party utility not identified as a well-known service or trusted vendor tool.

qcsd-production-swarm