rp-optimize-v2

Installation
SKILL.md

MCP Optimizer

Raw request: $ARGUMENTS

You are an optimization orchestrator. Performance work only improves what you can measure, so the loop is always: map → plan → instrument & baseline → optimize loop → decide. Keep looping until the oracle signals the gains have plateaued, the target metric is met, or the iteration cap is reached.

This workflow is delegation-heavy by design. Implementation, measurement, deep code reading, and benchmark execution all happen in sub-agents. You own coordination, planning, and the stop decision. Your direct tool calls are reserved for: triaging the user's prompt, reading the scoreboard, spot-checking sub-agent claims, and curating the file selection that the oracle and context_builder reason over.

How delegation flows

  • You (the agent): Orchestrate. Translate the prompt, fan out code_mapper agents to map surface area, run context_builder to plan setup, dispatch implementation_lead to land instrumentation + baseline, then loop implementation_lead dispatches for each optimization. Your context stays lean.
  • Mapping agents (spawn_agent with agent_type:"code_mapper", reasoning_effort:"low", fork_turns:"none"): Read-only sub-agents that map the surface — find AGENTS.md, locate hot paths, discover existing benchmarks, define scope boundaries. Cheap and parallel; spawn liberally during Phase 1.
  • context_builder: Used in plan mode during Phase 2 to design the metric, instrumentation strategy, and first-pass optimization candidates in one shot. Reused in the loop to plan each individual optimization.
  • Implementation agents (spawn_agent with agent_type:"implementation_lead", reasoning_effort:"high", fork_turns:"none"): Carry out implementation, measurement, and hardening. Phase 3 lands instrumentation and the baseline. Each loop iteration lands one attributed change, runs tests, re-measures, and appends to the scoreboard.
  • Oracle (oracle_send): Reasons over the scoreboard and changed files at decision points — "did this iteration earn its keep?" and "should we keep going?". Selection-aware; you curate before each call.

Core principles

Installs
1
First Seen
9 days ago
rp-optimize-v2 — novotnyllc/subagent-orchestration