The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions and referenced modules (e.g., references/claim-verification.md) utilize common command-line tools for repository inspection and plan verification. These include grep, find, wc, git log, npm test, pytest, cargo test, tsc, eslint, and curl. This is standard behavior for an agent skill tasked with codebase analysis and runtime behavior verification.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external data including repository source code, documentation (ADRs, RFCs), and user-provided specifications. This creates a surface for indirect prompt injection if those files contain malicious instructions. However, the skill incorporates a structured 'Plan contract', a 'Plan Quality Rubric', and a 'Claim Verification' workflow to mitigate accidental obedience by grounding all claims in local evidence.
Ingestion points: Repository files (code, tests), project documentation (ADRs, RFCs, READMEs), and user prompts.
Boundary markers: Uses specific markdown headers and sections defined in the 'Plan contract' and 'handoff-plans.md' to delimit instructions from data.
Capability inventory: File system read access, shell command execution for testing and metrics, and network access via curl for API verification.
Sanitization: Emphasizes falsifiable hypotheses and raw command output verification over accepting text-based claims at face value.
[SAFE]: No hardcoded credentials, malicious persistence mechanisms, or obfuscated payloads were found across the skill files.

planning