The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill utilizes a multi-model pipeline where synthesis and adversarial roles reconcile outputs from various providers. This creates an attack surface for indirect prompt injection if a task contains malicious instructions designed to influence the final decision log or implementation.
Ingestion points: Task intake descriptions and responses from multiple configured model providers (GPT, Claude) as defined in assets/routing.toml.
Boundary markers: The skill explicitly requires one isolated context per task and role across all phases to prevent context leakage (SKILL.md and references/workflow_contract.md).
Capability inventory: The skill can write multiple artifacts to .ai-workflow/panel/delivery, execute shell commands for testing, and invoke external runner scripts.
Sanitization: No explicit content sanitization or instruction filtering is documented for the cross-model synthesis process, although the manual review phase serves as a human-in-the-loop checkpoint.
[DYNAMIC_EXECUTION]: The core workflow involves the dynamic generation and execution of code and tests. In the 'Red', 'Green', and 'Refactor' phases, the agent is instructed to create or modify test files and implementation code, then execute them to verify behavior.
Evidence: SKILL.md defines a workflow where the agent must "add or update the failing test", "make the smallest change that passes it", and "simplify only while tests stay green".
[COMMAND_EXECUTION]: The skill relies on executing local scripts and shell commands to manage the panel workflow and verify task completion.
Evidence: The skill configuration in assets/routing.toml specifies execution paths for model runners, such as .agents/skills/claude-runner/scripts/run_claude.py. It also utilizes helper scripts like shared/scripts/record_native_response.py and validate_artifacts.py to manage state and audit logs.

collaborative-delivery