memex/peripheral-vision-plan.md

Implementation Plan: Component IV - Peripheral Vision Extraction

Objective

Implement a sophisticated "Foveal-Peripheral" context model. This ensures the agent has high-resolution focus on the current task (Foveal) while maintaining a low-resolution "skeletal" awareness of the broader Memex structure (Peripheral), optimized for token efficiency and reasoning accuracy.

Key Files & Context

• Target: projects/org-agent/literate/context.org (Source of src/context.lisp)
• Core Concept: Deep pruning of the Org AST based on semantic distance and structural hierarchy.

Implementation Steps

1. Identify Foveal Focus

• Extend the SIGNAL structure processing to identify a target-id (the current headline being operated on).

2. Implement Tree Pruning (context-extract-peripheral-vision)

• Create a recursive function that walks the Object Store starting from the root (or active projects).
• Rule A (Foveal): If the node matches target-id, include it and its immediate children in Full Resolution (Content + Attributes).
• Rule B (Peripheral): For ancestors and siblings of the target, include only Title and ID.
• Rule C (Background): For unrelated nodes, omit entirely or include only at Level 1.

3. AST to Org Renderer (context-render-to-org)

• Implement a serializer that transforms our org-object structures back into valid Org-mode strings.
• This allows the LLM to "see" the Memex in its native habitat.

4. Integrate with context-assemble-global-awareness

• Update this function to use the new extraction and rendering logic.
• Ensure it respects a maximum token/character budget to prevent context overflow.

Phase E: Chaos (Verification)

• Structural Test: Verify that ancestors are rendered as "skeletons" (no body text).
• Foveal Test: Verify that the target node is rendered with its full body text.
• Budget Test: Verify that the output string stays within defined limits even for large Memex structures.