| name | rale |
| description | Use when working with Reactive Agentic Loop Engineer MCP Server, including persisted prompt loops, agent-card registration, capacity discovery, capacity-fit master-plan decomposition, dispatch gates, pause/resume lifecycle, completion, and SQLite-backed audit state. |
RALE MCP Server
Use this skill when an agent needs to operate or extend the Reactive Agentic Loop Engineer MCP server.
Workflow
- Create a loop with
rale_create_loop and a conservative tokenLimit.
- Treat
tokenLimit as a hard prompt-length ceiling. RALE must never emit a goal whose Prompt.Length exceeds that limit.
- List or inspect state with
rale_get_loop and rale_list_goals.
- Claim a goal with
rale_claim_next_goal before executing it. Only one executor should succeed because claims use optimistic concurrency.
- Pause or resume active work with
rale_pause_goal and rale_resume_goal when work must stop before final completion.
- Complete work with
rale_complete_goal; include result metadata that helps reconstruct decisions and downstream dependencies.
Multi-Agent Orchestration Workflow
- Register agents with
rale_register_agent, including capabilities, supported task types, max concurrent goals, max token capacity, trust level, endpoint, and tool scopes.
- Use
rale_discover_agent_capacity before planning or when a task profile changes. RALE queries GET /agents/{id}/capacity?taskProfile=... when an endpoint is configured and falls back to cached/profile capacity when needed.
- Create capacity-fit plans with
rale_create_master_plan. Choose serial when generated subtasks must run in dependency order and parallel when they can run independently.
- Dispatch with
rale_assign_next_task. Assignment enforces dependencies, agent load, policy state, and human approval gates.
- Clear human gates with
rale_approve_goal; policy violations remain visible on each goal for audit.
- Record liveness/provenance with
rale_record_goal_heartbeat while agents iterate.
- Use
rale_resplit_goal when an agent reports capacity mismatch. Re-splitting is bounded by the goal iteration limit and records audit events.
Operational Rules
- Keep MCP transport on stdio and logs on stderr.
- Store intermediate outputs in
GoalResults; rely on LoopEvents for audit history.
- Split prompts early when accounting is uncertain.
- Sanitize prompts and restrict tool scope before delegating goal execution.
- Treat agent endpoints as open-world calls. Use cached capacity and human approval gates when discovery fails or trust/tool-scope checks are not satisfied.
- Prefer TUnit tests and TUnit assertions when changing decomposition, goal status transitions, or persistence behavior.