Files

Mike Swanson cd80f5e447 feat: add wiki knowledge layer (Phase 0 + Phase 1 seed)

Implements LLM-compiled wiki layer between raw session logs and live
CONTEXT.md, inspired by Karpathy's knowledge base workflow. Adds wiki/
directory structure, article templates, spec docs, and seeds first two
articles (Cascades of Tucson, GuruRMM) from 60+ session logs.

Updates CLAUDE.md to check wiki first on all context-loading triggers.
Captures verified ACG IP/hostname map and Neptune physical-location
clarification (Dataforth D2, subnet overlap TODO) in memory.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

2026-05-24 15:42:38 -07:00

3.2 KiB

Raw Blame History

Wiki Layer — References

Files This Feature Reads From

File / Path	Role
`session-logs/*.md`	Raw source — root general logs
`projects//session-logs/.md`	Raw source — project-specific logs
`clients//session-logs/.md`	Raw source — client logs
`projects/*/CONTEXT.md`	Cross-reference for live state facts
`.claude/memory/*.md`	Cross-reference for discrete facts
`credentials.md`	Cross-reference for access details

Files This Feature Writes To

File / Path	Role
`wiki/index.md`	LLM-maintained master index
`wiki/clients/<slug>.md`	Per-client synthesized knowledge
`wiki/projects/<slug>.md`	Per-project synthesized knowledge
`wiki/systems/<slug>.md`	Per-system synthesized knowledge
`wiki/patterns/<slug>.md`	Cross-cutting pattern articles

Commands This Feature Touches

Command	Change
`.claude/commands/context.md`	Add wiki-first search step (Phase 3)
`.claude/commands/save.md`	Add post-save wiki update prompt (Phase 5)
`.claude/commands/wiki-compile.md`	New — created in Phase 2
`.claude/commands/wiki-lint.md`	New — created in Phase 4

Integration Points

GrepAI

wiki/ is under the repo root — GrepAI watcher picks it up automatically
No config change needed
Wiki articles get full semantic search immediately after creation
CLI: D:/claudetools/grepai.exe search "cascades billing rate" --json -c -n 5

Ollama (Compilation Engine)

Model: qwen3:8b on DESKTOP-0O8A1RL (86 tok/s, fits in 12GB VRAM)
Model: qwen3:14b on other machines
Prompt style: structured-format, JSON frontmatter first, then article body
Endpoint: http://localhost:11434 (DESKTOP-0O8A1RL) | http://100.92.127.64:11434 (others)
See .claude/OLLAMA.md for full routing table

Git / Gitea Sync

wiki/ is tracked in Git — syncs via sync.sh on every /save and /sync
Cross-machine sync is automatic: Mike compiles on DESKTOP, Howard gets it on next pull
No special handling needed

Coord API

Wiki does NOT replace coord API — they serve different purposes
Coord API: real-time state (locks, messages, component state)
Wiki: synthesized historical knowledge
/wiki-compile may read coord API component state for currency checks but does not write to it

Existing Relevant Docs

Doc	Why Relevant
`.claude/CLAUDE.md`	Auto-context-loading section needs updating to reference wiki
`.claude/OLLAMA.md`	Compilation uses Ollama — routing table and examples
`.claude/FILE_PLACEMENT_GUIDE.md`	`wiki/` placement is new — add section
`.claude/memory/MEMORY.md`	Memory entries should link to wiki articles where applicable
`CONTEXT.md` (root)	Add wiki to "Where to Find Things" section

Karpathy Reference

Original post: https://x.com/karpathy/status/2039805659525644595 Posted: 2026-04-02 Key insight adopted: raw data → LLM-compiled wiki → Q&A against wiki → outputs filed back into wiki Key insight NOT adopted: Obsidian IDE, RAG pipelines, finetuning (unnecessary at our scale) Key divergence: We already have GrepAI (replaces his naive search engine) and coord API (real-time comms he lacks)

3.2 KiB Raw Blame History