The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes Bash scripts (context-scan.sh, context-guard.sh) to perform discovery, snapshotting, and restoration operations across the project and global root (~/.claude). These scripts rely on standard system binaries and provide built-in safeguards to prevent pruning outside designated backup directories.
[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes various project files, memory, and rules to identify candidate text for elision. This activity creates a potential surface where malicious instructions embedded in project files could attempt to manipulate the compression and deduplication process.
Ingestion points: The scripts/context-scan.sh script recursively discovers and reads all markdown and configuration files within the specified scope (project and global tiers).
Boundary markers: The orchestration logic relies on mechanical verification gates (crit_tokens and crit_tokens_ext) defined in references/keep-catalog.md to ensure that critical tokens like paths, versions, and negations are preserved byte-exact.
Capability inventory: The skill possesses Bash, Edit, Write, and Agent (for spawning sub-agents) tools, allowing it to modify permanent agent instructions.
Sanitization: The skill employs regex-based validation and a multi-step arbitration lattice to ensure that only generic or redundant instructions are removed, refusing acceptance of the whole run if mechanical verification fails.

context-slim