The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external sources, specifically GitHub issues/PRs through gitcrawl and Discord support messages through discrawl. This untrusted data is processed and incorporated into a 'semantic map' and an 'agent handoff packet' used by the AI agent.
Ingestion points: Data enters through evidence files such as gitcrawl-evidence.json and discrawl-evidence.json as described in scripts/semantic-map.mjs and references/workflow.md.
Boundary markers: The generated visual board utilizes HTML escaping for safety. However, the machine-readable JSON 'agent handoff packet' provided to downstream agents lacks explicit boundary markers or 'ignore embedded instructions' warnings.
Capability inventory: The skill can execute shell commands via git and other tool CLIs, and it can read/write local files.
Sanitization: While escapeHtml is used for the visual HTML report, the text content within the JSON handoff packet is not sanitized, leaving it open to instructions intended to influence the agent's behavior.
[EXTERNAL_DOWNLOADS]: The skill's setup instructions in references/workflow.md direct the user or agent to clone and build external repositories, namely github.com/openclaw/clawpatch and github.com/vercel-labs/deepsec.
vercel-labs is a well-known and trusted organization.
openclaw is the primary target project ecosystem for which this skill is designed.
[COMMAND_EXECUTION]: The scripts/semantic-map.mjs script uses spawnSync to execute local git commands (e.g., ls-files, log) to gather repository metadata. Additionally, the skill workflow involves running the CLI tools of clawpatch, deepsec, gitcrawl, and discrawl. These operations are transparent and consistent with the skill's stated analytical purpose.
[DYNAMIC_EXECUTION]: The script scripts/semantic-map.mjs dynamically generates an HTML report that includes a static JavaScript block (uiScript()) for UI filtering and search functionality. This is a common pattern for local development tools and does not incorporate untrusted executable code.

semantic-slicing