Codex works best with an LLM Wiki when it treats the wiki as durable reviewed knowledge, not as scratch space. The agent reads current guidance, stages source-backed edits, runs local checks, and stops before publication, security, privacy, legal, production, or unsupported-claim boundaries.
Support boundary
This page is practical LlmWikis guidance. It does not claim an official Codex adapter, hosted import pipeline, automatic publication workflow, public MCP server, certification, or public write API.
Role split
| Part | Job | Approval boundary |
|---|---|---|
| LLM Wiki | Durable, reviewed, source-linked organizational memory. | Do not rewrite authoritative knowledge without source review. |
| Codex | Execution agent for repo changes, draft synthesis, checks, and local implementation work. | Do not publish, deploy, widen public claims, or expose private material without approval. |
| Raw sources | Original evidence that supports compiled wiki pages. | Keep separate from public synthesis and do not sanitize by overwriting originals. |
| Staged drafts | Reviewable proposed public copy or documentation changes. | Promote only after source, privacy, authority, and verification checks. |
Codex startup path
- Read the front door. Start with project
AGENTS.md, thenreadme.human, then the listed current state files. - Refresh intake. Inspect active handoff buckets and summarize each file that still needs review.
- Find the authoritative pages. Use README, INDEX, TRUST_MODEL, GOVERNANCE, source policy, and route maps before broad edits.
- Preserve uncertainty. Keep contradictions, stale claims, missing evidence, and open questions visible instead of flattening them into confident copy.
- Draft narrowly. Convert source material into specific page edits, handoff records, roadmap entries, or archived evidence.
- Run targeted checks. Verify the affected files, routes, discovery records, and build paths that the change touched.
Good Codex outputs
- Source-backed page copy that cites existing wiki pages, original sources, or canonical external authority.
- Clear distinctions between durable LLM Wiki knowledge, compact AI Memory, Project Handoff, and execution-agent work.
- Public guidance that says what exists now and marks planned support before readers can mistake it for current capability.
- Completion reports that list files changed, checks run, failures, unresolved review questions, and recommended next commit or pull request title.
- Archived intake files after their disposition is recorded in durable state.
What Codex should not do
| Do not | Why | Safer replacement |
|---|---|---|
| Treat dropped reports as public truth. | Reports can be duplicative, speculative, stale, or internally contradictory. | Extract recommendations, reconcile them, and stage public copy for review. |
| Publish raw private transcripts or source files. | They may contain secrets, regulated data, customer details, or privileged material. | Use reviewed summaries with redaction and source status. |
| Claim automatic sync, certification, public MCP, or official adapters. | Those are support claims that need implemented public evidence. | Say planned, candidate, internal, or not current until evidence exists. |
| Collapse UAI/UAIX authority into LlmWikis. | UAIX owns UAI-1, AI Memory, Project Handoff, schemas, registries, and validator behavior. | Link to canonical UAIX pages and label LlmWikis material non-normative. |