SKILLWARE.md

Skillware Integration

Navigation: Documentation hub · Settings & preflight (skill env_vars in .env)

Rooms supports optional Skillware-based tool use for agents.

Design Principles

• Skills are assigned per agent, not globally.
• Skillware is loaded lazily only when an agent has configured skills.
• Agent replies remain human-readable even when tools are used.
• Skill execution metadata is logged as structured session events.

Install

Skillware is included in requirements.txt and installed with the core package:

pip install -r requirements.txt

Skill manifests are still loaded lazily only when an agent has configured skills.

Skills CLI Commands

python cli.py skills list
python cli.py skills inspect finance/wallet_screening
python cli.py skills suggest --expertise finance,risk,compliance

Wizard Assignment Flow

When creating a custom agent, you can optionally:

• assign one or more skill IDs
• add per-skill override settings (key=value pairs)
• skip skills entirely

Assigned skills are stored in agent config:

• skills
• skill_settings

Runtime Behavior

• Tools are exposed to the model only for agents with configured skills.
• Skill calls are guarded by per-turn and per-session limits.
• Tool results are passed back to the model for second-pass synthesis.
• Final user-visible output remains a normal agent message, not raw JSON.

Session Logging and Transcripts

Skill calls are recorded as structured role: skill events in session history.

Saved transcripts include those events with:

• agent name
• tool name
• arguments
• status
• result payload