The Agent Skills Directory

[SAFE]: The skill implements a legitimate multi-agent orchestration workflow. It manages project state through a local file system and structured documentation.
[COMMAND_EXECUTION]: The bundled script scripts/flow.py is a utility for managing a task board in JSON format. It performs standard file operations (reading/writing board.json) and graph validation (dependency checks, cycle detection). The script does not utilize network resources or execute arbitrary external commands.
[PROMPT_INJECTION]: The skill exposes a surface area for indirect prompt injection, as it ingests and processes untrusted user data into agent mandates.
Ingestion points: SKILL.md (Phase 0: 'vague idea', Phase 4: 'user reports a bug').
Boundary markers: The mandate template in references/worker-protocol.md provides structural organization but lacks explicit delimiters (such as XML tags or dedicated 'ignore instructions' warnings) to isolate untrusted user input within the worker's context.
Capability inventory: scripts/flow.py (file system writes). The orchestrator agent is designed to manage and spawn other agents/workers.
Sanitization: The instructions do not specify sanitization or validation of user-provided product ideas or bug reports before they are interpolated into mandates.
Note: This reflects the inherent risk model of orchestration skills. The framework attempts to mitigate this through 'Honesty invariants' and alignment benchmarks called 'golden case cards'.

delivering-mvp-fleet