Multi-Provider Architecture

Agents run on Claude, OpenAI (Codex), or Gemini. Provider selection is per-agent in the registry.

Provider Distribution

• Claude (51 agents): Primary. CLI: @anthropic-ai/claude-code
• OpenAI/Codex (4): margot, benjamin, jouke, dinand. CLI: @openai/codex
• Gemini (1): felice. CLI: @google/gemini-cli

Source of truth for agent→provider mapping: ge-ops/master/AGENT-REGISTRY.json (provider + providerModel fields)

Config Files

File	Controls	Read By
config/providers/claude.yaml	Claude CLI execution profile	Executor (provider_config.py)
config/providers/openai.yaml	Codex CLI execution profile	Executor (provider_config.py)
config/providers/gemini.yaml	Gemini CLI execution profile	Executor (provider_config.py)
config/provider-instructions/AGENTS.md	Codex native instructions	Injected into Codex sessions
config/provider-instructions/GEMINI.md	Gemini native instructions	Injected into Gemini sessions
config/provider-settings/codex-config.toml	Codex CLI config	Codex CLI reads at startup
config/provider-settings/gemini-settings.json	Gemini CLI config	Gemini CLI reads at startup

How It Works

1. Executor receives trigger for agent
2. Fetches agent config from admin-ui API (includes provider + model)
3. provider_config.py loads the provider YAML profile
4. Profile defines: CLI command template, native config paths, constitution handling
5. Prompt builder adapts: Claude gets full constitution; OpenAI/Gemini get universal principles only (quality gate in native instructions file)
6. CLI invoked via PTY with provider-specific flags

Constitution Handling

• Claude: Full constitution in prompt (constitution_handling: full)
• OpenAI/Gemini: Universal principles (1-5) in prompt; quality gate + turn budget in native instructions file (constitution_handling: universal_only)

Changing an Agent's Provider

1. Update provider and providerModel in AGENT-REGISTRY.json
2. Update via admin-ui: PATCH /api/agents/{id}/provider
3. Executor picks up the change on next trigger (no restart needed)