The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes PostToolUse hooks to automatically trigger a suite of external formatting and linting tools including ruff, biome, shfmt, taplo, hadolint, and shellcheck upon every file edit.
[COMMAND_EXECUTION]: The skill implements PreToolUse hooks to intercept and block standard system commands (e.g., pip, npm, yarn, pnpm), essentially overriding the user's expected environment behavior.
[DYNAMIC_EXECUTION]: The skill dynamically generates and executes subprocesses using the claude -p command to automate code repairs. This execution is gated by internal logic that selects different model tiers (Haiku, Sonnet, Opus) and time-outs based on the content of the file edits.
[INDIRECT_PROMPT_INJECTION]: The skill creates a feedback loop that is vulnerable to indirect prompt injection. It ingests untrusted code from file modifications, extracts violations, and passes that data to automated LLM subprocesses. An attacker could craft code that, when processed by the linters, triggers specific malicious behaviors in the fix-delegation phase.
Ingestion points: Hook runs on every file edit; content is parsed by multi_linter.sh.
Boundary markers: None specified for the violation JSON passed to subprocesses.
Capability inventory: Full file-system write access (via claude -p subprocesses) and shell command execution.
Sanitization: No evidence of sanitization for linter error messages before they are used as prompts for the delegated models.

plankton-code-quality