delegation-bridge_skill

Overview

This skill implements a universal delegation bridge that routes file I/O, deployments, and routine tasks to Gemini Flash 3.0 via an MCP server for dramatic cost savings. It preserves AI-agnostic workflows so any model (Claude, Gemini, Grok, etc.) can keep logic and reasoning local while offloading expensive I/O to the cheapest execution layer. The result is consistent project state, cross-IDE continuity, and near-zero I/O costs.

How this skill works

The current AI calls an MCP server named delegation_bridge using the tool call_gemini_flash with a small JSON payload (context, optional code, action). Gemini Flash performs file reads/writes, dual deployments to project and NinjaTrader bin, docs updates, state tracking, and integrity checks. The skill updates .agent/state/* files so any AI can resume work without losing context. Fallbacks to Claude Haiku or the current AI are used if Gemini Flash or MCP is unavailable.

When to use it

• Always delegate file operations: create, read, save, list, and CRLF fixes
• Delegate documentation updates: CHANGELOG, README, and state files
• Delegate routine analysis: directory traversal, git status parsing, backups
• Delegate deployments: save to repo and NinjaTrader bin, verify integrity
• Delegate context and usage tracking: session_state.json and cost_tracking.json

Best practices

• Keep reasoning, algorithm design, and interactive Q&A in the current AI; do not delegate
• Before delegation update .agent/state/session_state.json with progress and use ${PROJECT_ROOT} paths
• Record each delegation in .agent/state/cost_tracking.json for audit and analytics
• Prefer snippet-or-file separation: have the model produce logic, then call call_gemini_flash to integrate and deploy
• Implement exponential backoff and log errors to .agent/state/errors.log for any delegation failures

Example use cases

• Generate a full strategy file with the current model and delegate deployment to Gemini Flash for repo + NinjaTrader bin
• Ask the bridge to list UniversalORStrategy*.cs files and return sorted version numbers
• Update CHANGELOG.md and .agent/state/current_version.txt after a release via a single deploy action
• Save session progress when a model hits usage limits and hand off state to another AI
• Run small routine checks: CRLF normalization, file size verification, and post-deploy integrity checks