The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes data from external, potentially untrusted web applications during both design specification generation and the verification process.
Ingestion points: Data enters the agent context through tools like get_page_content, screenshot, and act_get during the visual exploration and flow execution phases (specified in SKILL.md and references/spec_generation.md).
Boundary markers: The skill implements a structured reconciliation layer where observations from the live DOM are compared against a local design.md intentional layer. Specific Markdown headers and JSON schemas are used to separate assertions from data.
Capability inventory: The skill includes tools for browser-side JavaScript execution (evaluate_js), browser state manipulation (act, navigate, type_text), and local file system operations via shell commands for hashing and manifest management.
Sanitization: The agent is instructed to perform a best-effort LLM 'Audit' pass (described in references/verification.md) to reconcile live site rules against design intent, flagging 'ORPHAN' or 'DIVERGENT' rules that lack documented source backing.
[DYNAMIC_EXECUTION]: The skill relies on the evaluate_js tool to perform complex DOM assertions and to inject a reporting overlay on tested pages. This involves executing multi-line JavaScript snippets within the browser session context, as detailed in references/annotate_failures.md and references/spec_generation.md.
[COMMAND_EXECUTION]: The skill generates and executes Bash scripts locally to manage the .integrity.json ledger and sessions.json manifests. These scripts utilize standard system utilities including shasum, date, cat, and mv to calculate file hashes and perform atomic file operations, as documented in references/spec_sync.md and references/verification_report.md.

ui-verification