llm-wiki boundary

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

Boundary and evidence

llm-wiki is an in-scope memory/knowledge/context-engineering system for maintaining a personal or team wiki through Claude Code. Its whole-system boundary includes the shipped plugin and copied-command entry paths, wiki initialization, schema and page templates, and the agent instructions for ingest, import, retrieval, access logging, pruning, L1 rule verification, lint and status. These paths jointly specify how retained knowledge is produced, admitted, maintained and later consumed. The boundary is not limited to the initialization library or shell scripts (SRC-1, SKILL.md, setup.sh, scripts/init-wiki.sh, .claude-plugin/plugin.json).

The classification follows the shipped maintainer role rather than an assumption that this repository owns an agent runtime:

Wiki maintainer for a personal or team knowledge base. You process source material and distribute extracted knowledge across wiki pages, maintain cross-references, and ensure structural integrity. --- SKILL.md

The plugin's marketplace entry points at the repository root; its manifest supplies the required wiki-root directory configuration. The root skill consumes that setting and dispatches its subcommands. The alternate setup.sh calls the same initialization script and optionally copies SKILL.md into a project's .claude/commands/wiki.md, replacing plugin placeholders with filesystem paths. The root wiki.md is a migration pointer, not a second maintained command implementation (SRC-1, .claude-plugin/marketplace.json, .claude-plugin/plugin.json, SKILL.md, setup.sh, wiki.md).

The setup consumer is explicit:

One-time. Creates the wiki skeleton in the wiki root. Pages and the config file are created by a script; you collect the answers and run it. Do not create or edit wiki pages yourself. --- SKILL.md

Initialization reads the selected tool's Schema, Dashboard, Hub and Access-Log templates and writes missing pages plus llm-wiki.yml; subsequent skill commands read that configuration and the generated schema. Ingest maintains pages and routing lines; query reloads selected retained pages, synthesizes an answer and appends access records; prune consumes those records; lint supplies the shipped quality-check procedure. These are instructions for the host agent, not independently executable implementations of those maintenance algorithms (SRC-1, scripts/init-wiki.sh, templates/logseq/Schema.md, templates/obsidian/Schema.md, SKILL.md).

Phase 1 - Targeted Read (Stage 2, only the chosen pages): - Read ONLY the 3-5 pages chosen in Phase 0 (max 3 loaded simultaneously, JIT — batch if needed) --- SKILL.md

The immutable evidence is the acquired Git repository at 11538c1195666db24e25e23cc64bb1d70587aef9, with an applicability cutoff of 2026-10-17. HEAD matched that pin and the checkout had no reported changes at inspection. The overall tier is doc-grounded: the central knowledge loop is a model-consumed prompt contract. The executable initialization and installation paths supply a narrower implementation layer. No target execution, provider call or independent experiment was performed; source inspection cannot establish actual compliance, retrieval accuracy, safety or sustained operating quality.

Functional exclusions and dependencies

Claude Code supplies model execution, tool access, plugin loading and automatic L1 memory loading. Those external host implementations are not frozen here, so their loading guarantees, permission enforcement and actual model behavior cannot be established. llm-wiki's instructions for consulting and modifying a configured L1 directory remain included. In particular, the source disclaims an action-time freshness gate (SRC-1, SKILL.md):

Out of scope: warning at the moment a due rule is about to justify an action. Claude Code loads L1, not llm-wiki — there is no hook to enforce that gate, so do not claim one. --- SKILL.md

Logseq and Obsidian are external storage/UI consumers; their implementations and link resolution behavior are not inspected. Git, Bash and Python provide local execution and persistence facilities. User confirmation and any periodic scheduler are external inputs, not autonomous services shipped by the target. User wiki contents, actual L1 memory files, URL sources passed to ingest, external inspiration links and external release services are not captured. Accordingly, this boundary supports the shipped format and workflow contracts, not compatibility demonstrated in a real vault, actual memory loading, deployed scheduling, ingested fact correctness or the security of real credentials (SRC-1, README.md, SKILL.md, setup.sh, scripts/init-wiki.sh).

Top-level coverage

The tracked tree was enumerated at the frozen commit. “Included; not initially inspected” below denotes available supporting evidence, not an access gap or a missing operational path. All shipped operational entry paths, generated-state templates and maintenance/consumption instructions were inspected. The source register covers the repository at the pin; later jobs may inspect its remaining supporting documents without adding a source.

No inaccessible material shipped operational path was identified. Available supporting files not initially read must be inspected before making claims specific to them; their availability is not itself corroboration of the main skill's claims.

Source register