llm-wiki

name	llm-wiki
description	Primary internal source for ordinary knowledge questions, including concepts, project or product lists, comparisons, recommendations, summaries, and document generation; also compiles immutable raw snapshots into the Schema-bound LLM Wiki and lints it. Use before public Web search unless the user explicitly asks for current/latest public information, a specific webpage, or a URL source.
toolsets	["llm_wiki","gbrain_query"]

LLM Wiki

For Query or Lint, call llm_wiki_context first with the matching operation and treat the returned AGENTS.md and Schema Bundle as authoritative. User-facing Ingest is queued through the intake tools below; its dedicated background Agent loads the Ingest context itself.

For a user-facing Ingest request, the main Agent is only an orchestrator:

1. Call llm_wiki_create_raw with source=current_message, conversation, attachments, or knowledge_file. For source=conversation, first call llm_wiki_conversation_documents with no ids to inspect the catalog. Use the full visible context to choose exactly the document ids that match the user's requested scope, including the current instruction when it adds a correction, classification, or boundary. Read those selected ids with the same tool, then synthesize one clean, self-contained Markdown document in raw_markdown. It must reflect the user's latest corrections; do not dump or concatenate chat transcripts. Pass both conversation_document_ids and raw_markdown to llm_wiki_create_raw. The server validates provenance and snapshots the Agent-authored Markdown exactly. It never guesses from a “recent N messages” window. Do not create an intermediate file with generic filesystem tools. The optional Raw title is only a provenance label, not a compiler instruction or Schema constraint.
2. Pass the returned complete raw_paths and opaque intake_id unchanged to llm_wiki_start_ingest. This queues (or reuses) the existing durable task and returns immediately. Tell the user the task id and that progress is available in the task center.
   • If the user only says compile, organize, or turn the material into Wiki, omit import_gbrain or set it to false. The task stops after Wiki publish and Lint.
   • Set import_gbrain=true only when the user explicitly asks to enter, import, or sync gbrain. The same task then continues with the validated gbrain PostgreSQL import.
   • The queue response proves only the selected Raw, task id, and requested stages. Do not predict or promise final page types, slugs, titles, page count, or relationships before the background Compiler finishes.

An explicit request to compile, organize, add, or turn selected material into Wiki is sufficient authority for this two-tool intake. Do not call request_user_input to reconfirm the scope or whether execution should start. Ask only when a genuinely missing source choice would change which user-owned material is ingested; Skill activation or tool availability is never such a choice.

Do not call llm_wiki_context, llm_wiki_publish, or llm_wiki_lint from the chat Session for Ingest. Those low-level protocol tools belong to the dedicated background Compiler Agent, which reads only the selected Raw snapshots, applies the active Schema, publishes complete pages, and runs Lint. It imports gbrain only when the queued task explicitly requests that second stage. Do not write raw/, wiki/, index.md, or log.md through generic filesystem tools.

For an explicitly identified duplicate or obsolete published page, call llm_wiki_retire_pages directly with the exact obsolete slug and an existing replacement slug. This is a deterministic maintenance operation: it does not queue the Compiler Agent or consume a compiler-model call. Do not infer a retirement merely from similar titles; the user must explicitly authorize the obsolete-to-replacement mapping. Do not read raw source files, list /knowledge, or inspect the Wiki before this call: the tool atomically verifies that both the obsolete page and replacement page satisfy the requested mapping. Keep sync_gbrain=false unless the user also asks to remove the obsolete page from gbrain. When enabled, gbrain performs a recoverable soft delete. Treat the tool result as authoritative. ok=true with either retired=true or already_retired=true completes the request; report the returned Lint and do not call any generic filesystem tool to verify it. If the retirement tool fails, report its exact error and stop. Do not use generic filesystem tools to inspect /knowledge or the physical knowledge root afterward: those tools intentionally have a different sandbox boundary and their denial is not evidence that the dedicated Wiki service is unreadable.

For Query, treat the published Markdown LLM Wiki as the complete source of truth and gbrain only as its structured acceleration/index layer. First call llm_wiki_context(operation="query"), then always call llm_wiki_query with the user's retrieval intent before deciding whether another knowledge path is needed. Do not query gbrain by default. Use allowlisted gbrain MCP tools only when entity relations, graph traversal, or structured filtering can materially improve the answer. A gbrain hit never replaces or skips the Markdown Wiki query; merge and deduplicate it as supplementary structure. Cite Wiki slugs and their sources; report a knowledge gap instead of reading raw files.

For Lint, call llm_wiki_lint. Use llm_wiki_compile when the user asks to verify gbrain compatibility. Neither operation repairs files.