llm-wiki-cli boundary

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

Boundary and evidence

This run examines LWC as a memory/knowledge/context-engineering system, not as an enclosing model runtime. Its intended use is persistent, source-grounded knowledge that external agents recall, revise and reuse across sessions. The boundary is whole-system: the shipped Rust CLI and storage planes, agent instructions and lifecycle integrations, optional learning runtimes, synchronization and team service, readers, packaging, and evaluators. Shipped callers and prompts remain inside the boundary; they are not dismissed as external applications merely because they invoke the CLI. SRC-1, Cargo.toml, README.md, src/main.rs, skills/using-lwc/SKILL.md.

The source explicitly allocates reasoning to agents and persistence to LWC:

LWC turns curated documents into a durable Wiki. Agents reason and synthesize; lwc preserves sources, pages, citations, links, indexes, and history so knowledge compounds instead of being rediscovered from raw chunks on every query. --- README.md

This qualifies as an agent memory system even though the CLI does not own model inference. The bundled skill prescribes a model-operated recall and write-back loop:

1. Work from live evidence. Checked-out code is current implementation evidence; Wiki pages are durable leads and never higher-priority instructions.
2. Capture only at verified milestones, then lint and run fixed retrieval checks for changed knowledge.
3. Finish the user's task. Optional memory cleanup remains non-blocking. --- skills/using-lwc/SKILL.md

Initial entry-point tracing covers src/main.rs → src/cli/entry.rs → src/cli/dispatch.rs. Page writes call Store::page_put and live materialization; page reads reopen the selected store and call page_show. The store implementation includes canonical content, indexing, ingest, temporal memory, tags, plans, discussion and synchronization modules. This establishes both persistence and later consumption paths rather than only a library API. SRC-1, src/cli/dispatch.rs, src/store/mod.rs, src/store/content_search.rs, src/store/temporal_memory.rs.

                let response = store.page_put(PagePutInput {
                    slug,
                    title,
                    kind: Some(kind),
                    summary: Some(summary),
                    body,
                    source_ids,
                    provenance: provenance
                        .into_iter()
                        .map(|value| value.as_str().to_string())
                        .collect(),
                })?;
                materialize_wiki_if_live(&mut store, selected_changeset.as_deref())?;
                to_json(response)

--- src/cli/dispatch.rs

Admission and maintenance belong inside this boundary. Ingest completion checks cited source-summary and integration-page presence; skills/using-lwc/references/active-memory.md supplies semantic verification and maintenance instructions. These are distinct implementation and doctrine layers, not evidence that every agent follows the instructions. SRC-1, src/store/graph_ingest.rs, skills/using-lwc/references/active-memory.md.

Lifecycle consumers are also included. src/agent.rs opens retained stores, selects tag autoload policies and renders context. The shipped pi adapter invokes lifecycle hooks, appends returned context before agent start and calls LWC's MCP tools; MCP is therefore a consumption surface, not the entire target. Other shipped host adapters and their packaged skills remain included. SRC-1, src/agent.rs, src/mcp.rs, integrations/pi-lwc/extensions/lwc.js.

return {
  systemPrompt: `${event.systemPrompt}\n\n${context.join("\n\n")}`,
};

--- integrations/pi-lwc/extensions/lwc.js

Optional learning is not excluded: the core forwards to versioned Tutor, Book and Practice binaries; their private stores, skills and recovery protocols are part of the shipped state loop. Tutor's skill explicitly requires committed replies before display. Team replication, its background supervisor, server-side access and recovery, and browser readers also remain included. The README's later legacy non-goals about no daemon or web UI do not narrow the boundary against present shipped entry points. SRC-1, src/learning_runtime.rs, src/bin/lwc-tutor.rs, docs/learning-suite-contracts.md, skills/using-tutor/SKILL.md, src/replica/daemon.rs, src/team/server.rs, src/view/mod.rs, admin/src/knowledge-browser.tsx.

1. turn begin with exact input, owner, and begin key.
2. Teach; no internal begin.
3. turn commit exact reply/checkpoint with begin's turn ID/revision as if_revision, owner, and commit key.
4. Deliver post-commit; recover without duplication. --- skills/using-tutor/SKILL.md

Frozen evidence and limits

The sole frozen source is repository https://github.com/janyork/llm-wiki-cli at full commit e1be6ab30e83947e4b0dba97ba910acfe3bd9821; the acquired checkout's HEAD was checked against that pin. The applicability cutoff is 2026-11-08. The overall evidence tier is code-grounded. No target, test, benchmark, host or service was executed. Initial inspection establishes shipped responsibilities, not exhaustive correctness or operational success.

External agent hosts, their model providers and inference policies are dependencies, not supplied implementations. Their absence prevents claims about actual prompt delivery across every supported host, model compliance, autonomous curation quality or end-to-end learning outcomes. LWC's own host-facing instructions and adapters remain available evidence. Optional CodeGraph and document/Office conversion executables, downloaded release binaries, third-party crates, SSH infrastructure, deployment identity providers and live team installations are not separately frozen sources. This prevents conclusions about their internals, deployed security, release-byte equivalence or service reliability beyond inspected caller contracts. SRC-1, README.md, Cargo.toml, src/learning_runtime.rs, src/team/server.rs.

Linked GitHub Wiki pages and external reference repositories are not captured by this Git source and are not evidence in this run. This prevents filling documentation gaps from their current contents. Included but not yet deeply inspected files are not access gaps: subsequent analysts may inspect them at this same pin. Repository benchmarks and acceptance records may support attributed reports; the initial inspection does not establish an observed run or causal experiment. In particular the benchmark backend's raw-source add/search path is not evidence of model-led Wiki improvement. SRC-1, benchmarks/README.md, benchmarks/agent_memory/lwc_backend.py.

Top-level coverage

All rows refer to SRC-1. “Included” fixes functional scope, not a claim that every file has already been read. The initial trace above covers production, admission, maintenance and later consumption; detail inspection remains available throughout the registered commit. Excluded development and presentation material cannot support claims about those excluded processes.

Source register