2.2 KiB
2.2 KiB
Phase 2: The Delegator (Roadmap)
Overview
Phase 2 shifts focus from kernel architecture to capability expansion. Because the kernel (Phase 1) is a generalized Event Bus, all Phase 2 features are implemented entirely as Org-Native Skills. No modifications to the Lisp Core (`core.lisp`, `neuro.lisp`, `symbolic.lisp`) are required.
The Phase 2 Skill Roster
1. The LLM Router
- Skill File: `skills/skill-router.org`
- Role: Meta-cognition. Replaces complex Lisp trigger logic with LLM intent classification.
- Trigger: Catches general `:user-command` or ambiguous `:buffer-update` events.
- Action: Instead of editing a file, it emits an internal Lisp event `(:type :EVENT :payload (:sensor :delegation :target-skill <skill-name>))`, forcing the engine to re-loop and trigger the specific skill.
2. Deep Memory (Vector/Semantic Search)
- Skill File: `skills/skill-atomic-notes.org`
- Role: Context augmentation beyond the live RAM `*object-store*`.
- System 2 (Lisp): Wraps a local search tool (like `ripgrep` or a lightweight Lisp vector db).
- System 1 (LLM): Formats the retrieved text into the Context API so downstream skills can read it before making decisions.
3. The Shell Actuator
- Skill File: `skills/skill-shell-actuator.org`
- Role: Gives the agent hands outside of Emacs.
- Registration: Uses `(org-agent:register-actuator :shell #'execute-shell-safely)`.
- System 2 (Gatekeeper): A massive, paranoid Lisp function that whitelist-checks commands (e.g., allows `git status`, blocks `rm`).
4. The Cron Scheduler
- Skill File: `skills/skill-cron.org`
- Role: Autonomous temporal action.
- Trigger: `(eq sensor :heartbeat)`
- System 2 (Lisp): Queries the `*object-store*` for deadlines. If a deadline has passed, it creates an action.
- System 1 (LLM): Drafts a polite, contextual warning message to send to the user.
5. Web Research
- Skill File: `skills/skill-web-research.org`
- Role: Internet connectivity via headless text browsing.
- System 2 (Lisp): Wraps `lynx` and `curl` to fetch webpage content.
- System 1 (LLM): Parses the raw HTML/text and synthesizes a summary for the user.