orchestrate-project
Warn
Audited by Gen Agent Trust Hub on Oct 19, 2026
Risk Level: MEDIUMPERSISTENCEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [PERSISTENCE]: The skill establishes a persistence mechanism by using the
schedule_taskfunction to create a recurring 'heartbeat' (defined inPOLLING.md). This task is designed to restart the orchestrator's polling waiter if the session dies, allowing the agent to continue running unattended and performing git/GitHub operations indefinitely. - [INDIRECT_PROMPT_INJECTION]: The skill's primary workflow involves reading untrusted data from Notion tickets to drive the behavior of autonomous worker threads.
- Ingestion points: The orchestrator fetches ticket titles, descriptions, and acceptance criteria from Notion via SQL queries and API calls (documented in
NOTION-GRAPH.mdandscripts/poll-prs.mjs). - Boundary markers: None identified. Untrusted ticket content is directly interpolated into worker prompts in
WORKER.md. - Capability inventory: Worker threads have the authority to implement code changes, execute local tests and builds (via Docker), and push commits to GitHub repositories.
- Sanitization: Absent. The skill relies on the LLM to safely interpret external ticket content without explicit validation or escaping.
- [DYNAMIC_EXECUTION]: The skill dynamically generates multi-step shell commands for git operations, including rebasing and force-pushing, within
scripts/stack.mjs. These command strings are passed to worker threads to be executed sequentially, creating a risk if the underlying ledger data were to be manipulated. - [COMMAND_EXECUTION]: The skill performs extensive execution of local shell commands, git, the GitHub CLI (
gh), and Docker Compose (scripts/reap.sh) to manage the project lifecycle, worktrees, and stack coordination.
Audit Metadata