multimodel-orchestrator

Multimodel Orchestrator Protocol — v0.1

Specification for coordinating multiple AI coding agents on a single project. v0.1 is a protocol document — runtime implementation comes in v0.2+.

Purpose

When multiple AI tools work on the same codebase, they need:

1. Role clarity — who does what
2. File boundaries — who touches which files
3. Session logging — how to pass context between agents
4. Conflict prevention — avoid stepping on each other's changes

Agent Roles

Define agents in .ai/config.yaml:

orchestrator:
  agents:
    - name: "architect"
      tool: "claude"
      role: "Design and plan architecture"
      files: ["docs/**", "AGENTS.md", "MEMORY.md"]
      permissions: "read+write"

    - name: "implementer"
      tool: "cursor"
      role: "Write implementation code"
      files: ["src/**", "lib/**", "tests/**"]
      permissions: "read+write"

    - name: "reviewer"
      tool: "codex"
      role: "Review code and suggest fixes"
      files: ["**"]
      permissions: "read-only"

Execution Modes

Mode	Behavior
sequential	One agent at a time. Session log required between each.
parallel	Multiple agents work simultaneously on different file scopes.
supervised	Human reviews each agent's output before the next agent starts.

Session Log Protocol

When an agent finishes a task, it writes a session log to .ai/session-logs/:

Session Log Format

# Session: {agent_name} → {next_agent}

**Timestamp:** {ISO 8601}
**Agent:** {tool_name} ({role_name})
**Mode:** {execution_mode}

## Action Summary
{What was done in 2-3 sentences}

## Files Changed
- {path} ({created|modified|deleted})

## Next Steps
1. {Concrete next action}
2. {Another action}

## Blockers
- {Any blockers, or "None"}

## Context for Next Agent
{Any important context the next agent needs}

Naming Convention

.ai/session-logs/{YYYY-MM-DD}-{agent}-{summary}.md

Example: .ai/session-logs/2026-05-30-architect-auth-design.md