The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes a user-provided task description which is interpolated into prompts for the planner, analyst, and worker subagents. A malicious task could contain instructions designed to bypass constraints or manipulate the output of these agents.
Ingestion points: The inp.task variable in workflow.py is the primary entry point for untrusted user data.
Boundary markers: The workflow lacks explicit boundary markers or instructions to the subagents to ignore potential instructions embedded within the task text.
Capability inventory: The system uses spawn_subagent (with tools_needed=True), write_artifact, and emit_finding activities.
Sanitization: No sanitization or escaping is performed on the user-provided task text before prompt interpolation.
Mitigation: The skill employs a multi-stage validation architecture where independent critic agents (e.g., plan-validator, critic, spec-compliance-reviewer) verify the output of each stage before the pipeline proceeds, which significantly mitigates the impact of potential injections.
[COMMAND_EXECUTION]: The workflow involves the generation and execution of shell commands. The analyst agent defines verification_command entries in the PRD, which are subsequently executed by worker agents to verify task completion.
Evidence: The analyst system prompt in workflow.py requires agents to provide verification_command for each acceptance criterion, which are then used by the worker agent.
[DYNAMIC_EXECUTION]: The skill is designed to orchestrate agents that write, test, and fix code based on user requirements. This involves the generation and execution of code within the agent's runtime environment.
Evidence: The worker and fix-worker agents in workflow.py are explicitly instructed to follow a Test-Driven Development (TDD) protocol, implementing and running code to satisfy acceptance criteria.

team-temporal