The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on executing local Python scripts to manage its workflow, specifically stitch_harness.py, setup_harness_runtime.py, and a local MCP proxy stitch_mcp_proxy.py.
Evidence: SKILL.md and references/workflow.md provide specific command lines for running these scripts from the plugin root.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external sources, creating a surface for indirect prompt injection.
Ingestion points: SKILL.md (reads .stitch/specs/<page-id>.json), references/workflow.md (processes OCR recognition results and ImageGen art artifacts), and references/evidence-contracts.md (receives external tool envelopes).
Boundary markers: The skill mentions NFKC and whitespace normalization for OCR and data-purpose mapping for HTML reconstruction, but lacks robust delimiters or explicit instruction filtering for all external inputs.
Capability inventory: The skill can execute shell commands via stitch_harness.py, write files to the local artifacts/ directory, and interact with a remote Stitch MCP via a local proxy.
Sanitization: Instructions explicitly forbid saving credentials (API keys, cookies, tokens) in evidence logs and mandate hash verification for artifacts, but do not describe sanitization to prevent the agent from obeying instructions embedded in design specs or OCR text.

stitch-design-harness