ops
Codex Obsidian LLM wiki loop
This is the concrete local adaptation of the LLM wiki pattern for Shanepedia. Obsidian is the browsing IDE. Codex is the maintainer and operator. The markdown wiki is the durable compiled knowledge layer between raw inputs and future questions.
Core adaptation
The workflow is not "upload files and ask a question." It is:
- collect or ingest raw source material
- compile it into source summaries, topic pages, entity pages, project pages, maps, or dashboards
- browse the results in Obsidian
- ask new questions from the improved wiki
- write the best answers back into the wiki so the system compounds
That makes Codex closer to a programmer maintaining a codebase than a chatbot answering from scratch each time.
Roles in this stack
raw/is immutable evidencewiki/is the maintained synthesis layer- Obsidian is the human browsing and navigation surface
- Codex is the primary editor and maintainer
- OpenClaw, Gemini, Ralph, and child-lane loops are feeder runtimes that should improve the parent wiki instead of bypassing it
Why this is a fit for us
This pattern matches how Shanepedia already wants to work:
- Shane explores and asks better questions
- Codex does the summarizing, linking, filing, and maintenance work
- Obsidian makes the results legible in real time through links, graph, dashboards, and canvases
- ElSalKB and future lanes become child evidence/execution contexts, not competing roots
Difference from ordinary RAG
Ordinary RAG re-discovers fragments at query time. This loop compiles durable understanding forward. Cross-links, contradictions, reusable answers, and project hypotheses should already exist in the wiki before the next question arrives.
Codex operating contract
When Codex is working against Shanepedia:
- search the parent wiki first
- read the relevant pages before touching raw source
- when a result is likely to matter again, write it back as a page or update
- prefer updating an existing high-signal page over leaving the answer in chat
- create new focused pages when a subtopic is getting too large for one article
- after meaningful edits, rebuild index, lint, and rerun the relevant repo checks
Typical local loop
cd /Users/bestrobot/.openclaw/workspace/projects/shanepedia
open -a Obsidian "$PWD"
python3 scripts/wiki/query_wiki.py "query text"
scripts/wiki/grep_wiki.sh term
python3 scripts/wiki/update_index.py
python3 scripts/wiki/lint_wiki.py
python3 scripts/wiki/build_site.py
Good durable outputs
Questions should often resolve into one of these:
- a source summary page
- a topic or entity page update
- a new project hypothesis
- a dashboard or canvas improvement
- an ops page that makes future agent work easier
- a markdown brief that should be promoted into the wiki
Product angle
For Shanepedia Studio, this is not only an internal method. It is part of the offer. The deliverable is a maintained founder/operator wiki that compounds context over time and becomes an operating system for projects, research, travel, contacts, and execution.
Short rule
Obsidian is the IDE. Codex is the programmer. The wiki is the codebase. Raw sources are the dependency graph.