passepartout/docs/PRD_dynamic_model_switching.org

PRD: Skill-Based Dynamic Model Switching

• Overview
• Mandates
• Requirements
• Acceptance Criteria

Overview

The `org-agent` currently relies on hardcoded LLM model strings within its pluggable provider skills. To match the flexibility of the legacy Openclaw system while maintaining a Minimalist Core, we will implement a "Homoiconic Configuration" model using the Skill Graph. Configuration settings will live as standard Org-mode properties in the user's memex, and a dedicated configuration skill will provide lookup services to other skills.

Mandates

• The Org Mandate: Configuration MUST be human-editable and machine-readable within `.org` files.
• Minimalist Core: No new global variables or state-management registries shall be added to the kernel (`src/*.lisp`).
• Skill Graph Sovereignty: Capabilities MUST build upon one another via dependencies.

Requirements

1. Property-Based Configuration: The system MUST allow defining LLM models via Org-mode properties (e.g., `:LLM_MODEL_OPENAI: gpt-4o`) anywhere in the loaded memex files.
2. Configuration Skill: A new skill (`skill-environment-config.org`) MUST be created to handle property lookups within the persistent `*object-store*`.
3. Skill Integration: Provider skills (OpenAI, Anthropic, OpenRouter) MUST declare a dependency on `skill-environment-config` and query it for their model string at runtime.
4. Late-Binding Updates: Because the kernel updates the `*object-store*` on every buffer save, changing a model string in an Org file results in an immediate, hot-swapped change to the agent's behavior.

Acceptance Criteria

• A user can add `:LLM_MODEL_OPENAI: gpt-4o` to a "Settings" headline in their memex.
• The OpenAI provider skill successfully fetches this value via the configuration skill.
• No modifications are made to the core Lisp daemon files.