The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Bash script (log.sh) and external CLI tools (atlas-review, plannotator) to maintain the audit trail and perform implementation reviews. The logging script includes a sanitization mechanism designed to prevent spreadsheet-based formula injection by escaping special characters.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection because it ingests untrusted data from tickets, specifications, and conversation history to generate implementation records. While output sanitization exists for logs, there are no instructions to isolate or ignore instructions embedded in the input data.
Ingestion points: External tickets, specifications, prototypes, and conversation history as defined in the Activation and Implementation Record sections.
Boundary markers: None identified.
Capability inventory: Writing to local files (HTML, Markdown, TSV) and executing shell commands (log.sh, atlas-review, plannotator).
Sanitization: Basic escaping for spreadsheet-sensitive characters is implemented in the log.sh script, but semantic sanitization of input instructions is absent.
[EXTERNAL_DOWNLOADS]: The skill manifest identifies external resources from GitHub repositories, including humanlayer/skills and backnotprop/pstack, which provide the source content for several reference files and scripts bundled with the skill.

tracking-implementation