LLM Wiki v2 boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

This run analyses the shipped LLM Wiki v2 repository as a knowledge system: how documents and client memories become retained Markdown notes, how later questions consume those notes, and how users and code maintain the retained state. The target includes both classic and classifier-routed CROW operation. It is not limited to the Python library API or to its MCP endpoint. The repository ships callers, persistence, retrieval, maintenance, prompts and evaluation drivers, so the boundary is whole-system, not an excluded-host library boundary.

The classification rests on the shipped loop, not only the product description. SRC-1 src/llmw2/wiki.py constructs the store, models, prompts, Librarian and Researcher. SRC-1 src/llmw2/agents/librarian.py routes incoming material, chooses merge or creation, saves the source, writes a note, links related notes and logs the change. SRC-1 src/llmw2/bundle/store.py writes notes and indexes, retains raw text, and reloads folder state. SRC-1 src/llmw2/agents/researcher.py reloads the tree for a subsequent question, selects notes and produces an answer from them. The latter explicitly constrains returned citation paths to the selected notes:

citations = [c for c in dict.fromkeys(_cited(text)) if c in read]

--- src/llmw2/agents/researcher.py

The shipped prompts state the intended division of responsibilities; these are design evidence, not proof of factual preservation:

You decide where knowledge goes and write the notes; code creates the files, the indexes and the links. --- src/llmw2/agents/prompts/librarian/system.md

The same retained state is exposed through the CLI, HTTP/web application and optional MCP memory. The console script is bound to llmw2.cli:main in SRC-1 pyproject.toml; SRC-1 src/llmw2/__main__.py calls that same entry point. The CLI's ingest, sync, rebuild and ask branches call the Wiki facade. SRC-1 src/llmw2/server.py routes HTTP ingest/upload/ask requests to its current Wiki and serves the web assets; SRC-1 src/llmw2/web/queue.js sends uploads or text ingests, and SRC-1 src/llmw2/web/app.js posts questions. These are included shipped consumers, not external applications.

SRC-1 src/llmw2/memory.py connects MCP tools to the same loop. In particular, remember calls wiki.ingest, recall calls wiki.ask, and read_note reads the store. Its client-model setting can relay model requests through a client while the wiki still runs its filing and retrieval code. This makes the MCP integration part of the system, but does not make Claude Code, Codex or another MCP client's enclosing agent loop part of the target. The retained-memory write enters the same filing implementation:

return self.run(lambda wiki: filed(wiki.ingest(text, title=title or None, author=by))) --- src/llmw2/memory.py

Maintenance includes source synchronization and origin tracking, replacement/archiving, deletion and purge, route audits, taxonomy health, user-managed areas, batch planning and folder splitting, plus rebuilding from retained raw copies. These responsibilities stay inside the boundary even when initiated by a user or driver. SRC-1 src/llmw2/sources/sync.py turns connector changes into wiki.ingest_many calls and stores per-connector cursor state. SRC-1 src/llmw2/wiki.py implements rebuild by clearing retained state while keeping folders, then filing again without a plan:

with self.store.lock(): self.store.clear(keep_folders=True) return self.ingest_many(items, concurrency=concurrency, on_result=on_result, plan=False) --- src/llmw2/wiki.py

The evidence is frozen at SRC-1 commit e0a205afe60bd1913d083a3ed4cdab81e38e4c26, with applicability cutoff 2026-11-08. The checkout HEAD was checked against that pin. The overall tier is code-grounded: executable source was traced from shipped entry points into retention and later consumption. No target code, tests, examples, benchmarks or providers were executed. Prompt goals, benchmark tables and author explanations remain declared or reported evidence, not independently observed success.

Top-level coverage

The tracked tree was enumerated. Included means eligible evidence within the functional boundary, not a claim that every file has already been read. The initial inspection below traces the material lifecycle; remaining available files may be inspected at the same pin. No material lifecycle path was unavailable.

External dependencies and evidence limits

Model providers and classifier servers, Python dependencies (including the optional MCP SDK), operating system/filesystem behavior, browser engines and external source connectors are dependencies, not separately frozen implementations. Their integration code in SRC-1 is included; their independent guarantees are not. Client-side model compliance and the enclosing MCP client's scheduling, tools and session memory are outside the target, preventing claims that wiki instructions enforce those clients' behavior. The separate Crow desktop repository is outside this source pin, preventing conclusions about its packaging, UI or agent behavior (SRC-1 README.md, src/llmw2/memory.py, pyproject.toml).

The external papers and noncommitted real benchmark corpora mentioned in SRC-1 README.md and benchmarks/README.md are not frozen sources here. Repository descriptions and retained tables can support attributed reports, but not verification of the underlying private corpus, external experiments or broad accuracy/scaling claims. No live browsing or additional repository is needed to establish this boundary. Missing execution evidence limits operational conclusions; available-but-unread files are not described as inaccessible.

Source register