The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill reads metadata and content from various markdown artifacts in the project to determine the current research state. This creates an attack surface where malicious content in processed files could influence the orchestrator's routing logic.
Ingestion points: The skill reads files from the research/, .agents/skill-artifacts/, and .agents/experience/ directories as specified in SKILL.md and scripts/manifest-sync.ts.
Boundary markers: The skill relies on standard markdown frontmatter and a structured JSON manifest (.agents/manifest.json) for data organization.
Capability inventory: The skill has access to shell execution (Bash) and file system operations via Bun-based TypeScript scripts.
Sanitization: The scripts/manifest-sync.ts file implements a custom regex-based frontmatter parser that performs basic extraction and coercion of metadata fields, providing some level of structural validation.
[COMMAND_EXECUTION]: The skill executes local TypeScript scripts using the bun runtime to perform project maintenance and metadata synchronization.
Evidence: SKILL.md contains instructions to execute bun scripts/manifest-sync.ts to regenerate the state index.
Evidence: scripts/scaffold-eval-loop.ts uses Bun.spawnSync to execute the manifest sync script during workspace setup.
[DYNAMIC_CONTEXT_INJECTION]: The skill utilizes dynamic context injection to provide the agent with immediate visibility into the repository state upon invocation.
Evidence: SKILL.md contains shell commands using the ! command`` syntax to list artifacts in .agents/skill-artifacts, check for canonical folders, and display the last five git commits.

orchestrate-research