Knowledge Base
The Markdown articles under ~/.codex/knowledge/raw/ are canonical. Search indexes and the HTML browser are generated views.
Select one operation
- Ingest: add an operator-approved source as a new article.
- Refresh: update an existing article from its canonical primary source.
- Maintain: repair metadata, links, taxonomy, or generated views without adding research claims.
- Query: use the
workflow knowledgecommands; do not mutate the corpus.
Mutation procedure
- Read frontmatter schema and the narrow operation-specific reference.
- For external material, browse the canonical primary source using the current web capability. Do not rely on unavailable legacy tool names.
- Keep sourced facts, inference, and operator decisions distinct. Never store credentials, customer content, or raw session transcripts.
- Write or update one canonical Markdown article; do not edit the generated HTML browser.
- Run the corpus lint and
workflow knowledge reindex. - Rebuild the browser with
python3 scripts/build_browser.pyafter a content mutation. - Query the changed topic and inspect the generated article before declaring success.
Delegation
Use direct deterministic work for one or two articles. A bounded read-only sidecar is appropriate only for multiple independent primary sources; the main agent owns adoption decisions and writes.
Boundary
research-evaluator: decides whether external research should be adopted.knowledge-base: stores and retrieves accepted engineering knowledge.recall: reconstructs local session history.context-graph: stores project decision traces.
References and scripts
- frontmatter schema
- learning and evidence rules
- canonical surface URLs
- ingest gotchas
scripts/detect_updates.py— compare tracked upstream versions.scripts/build_browser.py— rebuild the derived HTML browser.