neural-learn

name	neural-learn
description	Harvest knowledge from archived features into .neural/knowledge/ — run after archiving a feature, or invoke manually to rebuild the project knowledge base

Neural Learn — Harvest and Synthesize

Harvest knowledge from .neural/archive/ and synthesize it into .neural/knowledge/ — a shared context base that future interviews and plans draw from automatically.

Steps

1. Check the archive. List all feature directories in .neural/archive/. If none exist, stop: "No archived features yet. Run /neural:neural-archive after completing a feature."

2. Load existing knowledge. Read .neural/knowledge/ in full if it exists — merge, never overwrite.

3. Harvest from every archived feature.

   For each directory in .neural/archive/*/, read:

   • CONTEXT.md — Language section (domain terms) and feature decisions
   • PLAN.md — tech stack signals: test runner command, frameworks, file structure patterns
   • REVIEW.md — every CRITICAL and WARNING finding
   • docs/adr/*.md — architectural decisions

4. Synthesize into .neural/knowledge/. Write or update all four files. Every archived feature must be reflected before this step is complete.

   PROJECT-CONTEXT.md — project-level facts every feature inherits:

   • Tech stack: language, framework, test runner with exact command
   • Structural conventions: naming patterns, folder layout, import style derived from the archive
   • Architectural patterns that recur across 2+ features

   GLOSSARY.md — unified domain vocabulary:

   • One entry per term, merged from every Language section in every CONTEXT.md
   • Conflicts surface explicitly: ⚠️ Conflict: [feature-a] defines as X; [feature-b] defines as Y — resolved to Z
   • Strict glossary — no decisions, no implementation details

   DECISIONS.md — cross-feature architectural decisions:

   • Only decisions with impact beyond the originating feature
   • Each entry: decision summary, rationale, provenance ([feature-name])
   • Feature-specific ADRs that have no cross-cutting impact are excluded

   ANTIPATTERNS.md — patterns that blocked or warned in review:

   • Only findings that appear in 2+ REVIEW.md files — single occurrences are noise
   • Grouped by category (test quality, architecture, scope creep, code style)
   • Each entry: pattern, severity (CRITICAL / WARNING), provenance ([feature-a, feature-b])

5. Report.

   Knowledge base updated: .neural/knowledge/
   — PROJECT-CONTEXT.md  (stack, conventions, patterns)
   — GLOSSARY.md         (<N> terms)
   — DECISIONS.md        (<N> cross-feature decisions)
   — ANTIPATTERNS.md     (<N> recurring patterns)
   
   Loaded automatically on next neural-interview and neural-plan.