The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local TypeScript script using the Bun runtime to handle search logic.
[COMMAND_EXECUTION]: The ContextSearch.ts tool spawns ripgrep (rg) as a subprocess to search through local directories. The arguments are passed as an array to Bun.spawn, and the search tokens are sanitized to include only alphanumeric characters, effectively preventing command injection.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface area by searching and displaying snippets from past conversation logs and work documents.
Ingestion points: Markdown work documents (ISA.md) and JSONL conversation logs found in the ~/.claude directory.
Boundary markers: Search results are presented within visual text boxes (═══), but individual snippets do not utilize specific 'ignore embedded instructions' warnings.
Capability inventory: The skill performs file reads and writes within the ~/.claude path and executes subprocesses via Bun.
Sanitization: User input is tokenized to strip non-alphanumeric characters, and extracted snippets are truncated to a maximum of 200 characters to mitigate the risk of long-form injection payloads.
[SAFE]: Data access is strictly scoped to the ~/.claude project directory. The code explicitly validates that search paths remain within the designated project slug directory before performing operations.

ContextSearch