feat(skill-runtime): add per-agent Skillware integration with lazy loading and guardrails

[rosspeili]

Problem Statement

Problem Statement

Issue #23 defines the high-level direction to turn Rooms into an operational control room with external tool usage via Skillware. Right now, Rooms agents can only produce text; they cannot call assigned tools. We need a safe, minimal runtime integration that preserves current behavior when no skills are configured.

We also need strict lazy loading: skill manifests/requirements should not be loaded unless an agent actually has skills assigned and attempts to use them.

Proposed Solution

Implement the core runtime plumbing for per-agent skills:

1. Extend AgentConfig with:

   • skills: List[str] = [] (skill IDs like compliance/tos_evaluator)
   • optional skill_settings: Dict[str, Dict[str, Any]] = {} for per-skill runtime config/env overrides

2. Add a new runtime module (example: rooms/skills_runtime.py) that:

   • lazily loads skills only when agent.skills is non-empty
   • uses SkillLoader.load_skill(skill_id) from Skillware
   • instantiates skill classes on-demand and caches them per session/agent
   • normalizes execution responses/errors into a consistent internal payload

3. Add tool execution guardrails:

   • per-agent allowlist only (no global unrestricted tool access)
   • max tool calls per turn/session
   • execution timeout
   • failure-safe handling (skill errors should not crash the room)

4. Integrate into Agent.generate_response():

   • no behavior change for agents with no skills
   • if skills are present, run provider-compatible tool loop and return synthesized agent answer

5. Dependency model:

   • make Skillware optional (feature-gated)
   • if skills are configured but Skillware is not installed, return clear actionable error guidance

6. Tests:

   • no-skill regression tests (existing flow unchanged)
   • lazy loading tests (SkillLoader not called unless skills assigned)
   • tool-call execution path tests (mocked skill execute)
   • guardrail tests (timeout/max calls/failure path)

Alternatives Considered

• Monolithic implementation with runtime + CLI + transcript UX in one PR.
  • Rejected: too risky and hard to review.
• Mandatory Skillware dependency for all Rooms users.
  • Rejected: increases install/runtime surface for users who only want think-tank mode.
• Global skill pool available to all agents.
  • Rejected: breaks least-privilege model and increases security risk.

Priority

High 🔴

Additional Context

Parent tracker: #23

This issue is intentionally runtime-only (no major CLI UX changes yet).
Follow-up issues will handle session UX/log rendering and CLI wrappers.

Acceptance criteria:

• Runtime support for per-agent skills exists and is behind explicit config.
• Lazy load behavior is enforced and tested.
• Guardrails are implemented and tested.
• Existing sessions (no skills) behave exactly as before.

[rosspeili]

Closed by #58