Knowledge Skill

Project Cortex - Structural Knowledge Injection and Long-Term Knowledge Retrieval.

Boundary note:

• This skill is for durable/reusable knowledge.
• Short-term operational memory belongs to Rust memory core and is exposed separately via memory skill facade.

Commands

Unified Search (primary)

Single entry point for all knowledge search. Use this instead of separate search/link_graph_hybrid_search to avoid tool ambiguity.

Parameter	Type	Default	Description
query	str	-	Search query (required)
mode	str	"hybrid"	hybrid (link_graph+vector), keyword (ripgrep), link_graph (links only), vector (semantic only)
max_results	int	10	Maximum results
scope	str	"all"	For mode=keyword only: docs, references, skills, harvested, all

Example:

@omni("knowledge.search", {"query": "UltraRAG research report"})
@omni("knowledge.search", {"query": "architecture", "mode": "keyword", "scope": "harvested"})

Documentation Commands

search_documentation

Search markdown documentation and references for specific topics.

Parameter	Type	Default	Description
query	str	-	Search term (required)

Example:

@omni("knowledge.search_documentation", {"query": "trinity architecture"})

search_standards

Search coding standards and engineering guidelines in docs/reference/.

Parameter	Type	Default	Description
topic	str	-	Engineering topic (required)

Example:

@omni("knowledge.search_standards", {"topic": "python linting"})

Semantic Search Commands

knowledge_search (alias: code_search)

Semantic search for code patterns and documentation in knowledge base.

Parameter	Type	Default	Description
query	str	-	Natural language query (required)
limit	int	5	Maximum results

Example:

@omni("knowledge.knowledge_search", {"query": "error handling patterns", "limit": 5})

code_context

Get LLM-ready context blocks for a query.

Parameter	Type	Default	Description
query	str	-	Query for context (required)
limit	int	3	Number of context blocks

Example:

@omni("knowledge.code_context", {"query": "how to handle errors", "limit": 3})

Knowledge Base Commands

update_knowledge_base

Save knowledge entry for future retrieval.

Parameter	Type	Default	Description
category	str	-	patterns/solutions/errors/techniques/notes (required)
title	str	-	Entry title (required)
content	str	-	Markdown content (required)
tags	list[str]	[]	Tags for categorization

Example:

@omni("knowledge.update_knowledge_base", {
    "category": "patterns",
    "title": "Error Handling Pattern",
    "content": "Use Result types instead of exceptions...",
    "tags": ["error", "python"]
})

search_notes

Search existing notes and knowledge entries.

Parameter	Type	Default	Description
query	str	-	Search query (required)
category	str	None	Filter by category
limit	int	10	Maximum results

Example:

@omni("knowledge.search_notes", {"query": "error handling", "category": "patterns"})

summarize_session

Summarize current session trajectory into structured markdown.

Parameter	Type	Default	Description
session_id	str	-	Unique session identifier (required)
trajectory	list[dict]	-	Execution steps (required)
include_failures	bool	true	Include failed approaches

Example:

@omni("knowledge.summarize_session", {
    "session_id": "session-123",
    "trajectory": [{"step": 1, "action": "search", "result": "found 5 files"}],
    "include_failures": true
})

Knowledge Ops Commands

ingest_knowledge

Ingest or update project knowledge base.

Parameter	Type	Default	Description
clean	bool	false	Full re-index if true

Example:

@omni("knowledge.ingest_knowledge", {"clean": false})

knowledge_status

Check knowledge base status.

Example:

@omni("knowledge.knowledge_status")

Link Graph and TOC (use unified search for link_graph-only)

For link-graph-only (link reasoning, no vector), use the unified search with mode="link_graph":

@omni("knowledge.search", {"query": "agent skills progressive disclosure", "mode": "link_graph", "max_results": 5})

For structured filtering/sorting, use schema-v2 search_options:

@omni("knowledge.search", {
  "query": "architecture",
  "mode": "link_graph",
  "max_results": 5,
  "search_options": {
    "schema": "omni.link_graph.search_options.v2",
    "match_strategy": "exact",
    "sort_terms": [{"field": "title", "order": "asc"}],
    "filters": {
      "link_to": {"seeds": ["design-doc"], "recursive": true, "max_distance": 2},
      "tags": {"any": ["architecture", "design"]}
    }
  }
})

mode="link_graph" responses include parsed_query (residual free-text after directive extraction) and normalized effective search_options from the Rust planner.

link_graph_toc

Get Table of Contents for LLM context (all notes overview).

Parameter	Type	Default	Description
limit	int	100	Maximum notes to return

Example:

@omni("knowledge.link_graph_toc", {"limit": 50})

link_graph_hybrid_search

Hybrid search combining LinkGraph reasoning + vector search fallback.

Parameter	Type	Default	Description
query	str	-	Search query (required)
max_results	int	10	Maximum results
use_hybrid	bool	true	Use vector fallback

Example:

@omni("knowledge.link_graph_hybrid_search", {"query": "architecture MCP", "use_hybrid": true})

link_graph_stats

Get knowledge base statistics.

Example:

@omni("knowledge.link_graph_stats")

link_graph_refresh_index

Trigger LinkGraph index refresh through the common backend API. Useful for operations/debugging with -v monitor output.

Parameter	Type	Default	Description
changed_paths	list[str]	[]	Changed paths for delta refresh planning
force_full	bool	false	Force full rebuild instead of delta path

Example:

@omni("knowledge.link_graph_refresh_index", {"changed_paths": ["docs/architecture/kernel.md"]})
@omni("knowledge.link_graph_refresh_index", {"force_full": true})

link_graph_links

Find notes linked to/from a specific note.

Parameter	Type	Default	Description
note_id	str	-	Note ID (required)
direction	str	"both"	"to", "from", or "both"

Example:

@omni("knowledge.link_graph_links", {"note_id": "architecture", "direction": "both"})

link_graph_find_related

Find notes related to a given note using link-graph traversal.

Parameter	Type	Default	Description
note_id	str	-	Starting note ID (required)
max_distance	int	2	Maximum link distance
limit	int	20	Maximum results

Example:

@omni("knowledge.link_graph_find_related", {"note_id": "agent-skills", "max_distance": 2})

Core Concepts

Topic	Description	Reference
Development Context	Project rules, scopes, guardrails	context.md
Writing Memory	Writing style guidelines	writing.md
Session Summarization	Trajectory capture pattern	session.md

Best Practices

• Search first: Before adding new knowledge, search for duplicates
• Use categories: Organize entries by category (patterns/solutions/errors/techniques)
• Add tags: Use consistent tags for better retrieval
• Include examples: Code examples improve AI understanding

ingest_document

Ingest a document (PDF, Markdown, etc.) with full RAG pipeline: parse, chunk, optional entity extraction and graph storage, then vector store.

Parameter	Type	Default	Description
file_path	str	-	Local path or PDF URL (e.g. https://arxiv.org/pdf/2601.03192) (required)
chunking_strategy	str	"semantic"	sentence, paragraph, sliding_window, semantic
extract_entities	bool	true	Extract entities and store in knowledge graph
store_in_graph	bool	true	Store extracted entities/relations in graph

When file_path is a URL, the file is downloaded to project data (.data/knowledge/downloads) then processed.

Example:

@omni("knowledge.ingest_document", {"file_path": "docs/guide.pdf"})
@omni("knowledge.ingest_document", {"file_path": "https://arxiv.org/pdf/2601.03192"})

Advanced

• Semantic vs Text Search: Use knowledge_search for semantic understanding, search_documentation for exact matches
• Batch Ingest: Call ingest_knowledge with clean=false for incremental updates
• Session Continuity: Use session_id to link related sessions
• PDF from URL: Use ingest_document with a PDF URL; file is saved under project data then ingested