Self Learning Code Intelligence

Self Learning Code Intelligence

Core Rule

Treat this repository as a closed learning system. Capture useful interaction facts into project-local sources, query code intelligence before risky changes, refresh graph indexes after changes, and keep all generated memory inside this repository unless the user explicitly asks to export it.

Automatic Loop

1. Start with project-local context:
   • Run npm run graphify -- query "<task/question>" when graphify-out/graph.json exists.
   • Run npm run gitnexus:status; if stale, run npm run gitnexus:analyze.
   • Read AGENTS.md, docs/OPERATING_MODEL.md, or knowledge/README.md only when the task touches project boundaries or learning behavior.
2. During the task, prefer project-local wrappers:
   • LLM Wiki: npm run llmwiki -- <args>
   • GitNexus: npm run gitnexus:analyze, npm run gitnexus:status
   • Graphify: npm run graphify -- <args>
   • Imported global digest: npm run digest:imported
   • Obsidian inbox: npm run obsidian:sync
   • Obsidian intelligence publish: npm run obsidian:publish-intelligence
   • One-click digestion: npm run learn:all
   • Health check: npm run doctor
3. After meaningful work, capture the learning:
   • Run powershell -NoProfile -ExecutionPolicy Bypass -File skills/self-learning-code-intelligence/scripts/capture-learning.ps1 -Title "<short title>" -Summary "<what changed or was learned>"
   • Include decisions, corrections, and open questions when useful.
4. Refresh local intelligence:
   • Run npm run gitnexus:analyze
   • Run npm run graphify -- update .
   • Run npm run llmwiki:status
   • Run npm run llmwiki -- compile --review --lang zh-CN only when LLM Wiki provider credentials are available or the user asked for compilation.
   • For broad automatic ingestion or routine learning refreshes, prefer npm run learn:all so Obsidian, imported digests, LLM Wiki review, GitNexus, Graphify, Obsidian intelligence publishing, and learning capture stay in one closed loop.

Tool Roles

Capture Quality

Write source notes as facts and operational rules, not chat transcripts. Include:

• user intent and exact acceptance boundary,
• commands or tools that worked,
• commands or assumptions that failed,
• project-local decisions,
• portability or isolation implications,
• next trigger that should cause this skill to be used.

Isolation Rules

• Do not write global Codex, Claude, Cursor, or MCP config by default.
• Do not import memories or skills from other workspaces unless the user asks.
• When global project assets are imported under project-imports/global-new-project/, treat them as references and candidate skills. Prefer LLMWIKI-local skills, memory, .env, and AGENTS.md rules.
• Before using an imported global skill, read docs/IMPORTED_GLOBAL_DIGEST.md or the digest sources under sources/imported-global-new-project/digests/. Promote imported skills into the main skills/ directory only on explicit request or repeated local use.
• Do not run Graphify --global, GitNexus global setup, or plugin global installers unless explicitly requested.
• Keep API keys in project .env; never hardcode them in sources or skill files.
• Import only the Obsidian LLMWIKI inbox by default; do not scan the whole vault unless explicitly requested.
• Generated state is rebuildable. Portable logic belongs in tools/, skills/, knowledge/, and docs/.

Common Mistakes

References

• Read references/tool-contracts.md when choosing which tool to call.
• Use scripts/capture-learning.ps1 for deterministic learning capture.