MindBase boundary

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

Boundary and evidence

This run analyses MindBase as a persistent, model-maintained knowledge system. Its intended use is to retain user sources, maintain synthesized Markdown knowledge, and supply that retained knowledge to later questions and agent sessions. It is not classified merely as its MCP tool interface or its core library. The repository ships the server that coordinates operations, instructions for an externally hosted builder, and a session-start consumer of persisted context (SRC-1, README.md, apps/server/src/ops/runner.ts, apps/plugin/agents/builder.md, apps/plugin/hooks/session-start).

The declared purpose is explicit:

An open-source implementation of Andrej Karpathy's LLM Wiki idea: an AI that builds and maintains a wiki from your sources. Not RAG-in-a-vector-DB. A real markdown wiki on your disk, that an LLM gardens for you between conversations. --- README.md

The implementation connects model output to persistent knowledge updates. The server's build path gathers retained inputs, calls its model-completion helper, then applies the returned actions (SRC-1, apps/server/src/ops/runner.ts):

const { system, user } = buildPrompt({ core, sources, today: new Date().toISOString().slice(0, 10) }); const out = await completeJson(ctx, { system, user, schema: buildSchema, maxTokens: 8192 });

emit({ kind: 'phase', phase: 'writing' }); const result = await applyActions(ctx.projectRoot, out.actions); --- apps/server/src/ops/runner.ts

Retained state is subsequently reloaded, rather than existing only in the completion that produced it (SRC-1, apps/server/src/ops/gather.ts):

const [context, indexYaml, readme] = await Promise.all([read('context.md'), read('index.yaml'), read('README.md')]); return { context, indexYaml, readme };

--- apps/server/src/ops/gather.ts

The functional boundary includes the entire shipped repository artifact: core storage, compilation, retrieval and model adapters; server routes and background workers; web and command-line callers; MCP tools, resources and prompts; plugin commands, agents, hooks and templates; capture clients; and shipped tests and evaluation instructions. These are potential evidence for later jobs, not uniformly inspected implementations. No shipped caller is declared external simply because it invokes an API.

The boundary kind is complete artifact, partial loop, not whole-system. Initial inspection traced the server's contribution planning/application, build, research and lint coordinator; action persistence; project-input gathering; the ask route's answer/file-back integration; MCP registration and contribution persistence; and plugin skill, builder and session-start instructions. It did not exhaustively trace every capture client, background worker, legacy compilation path, search implementation or evaluator. Those are available but not yet inspected, not access gaps or deliberate functional exclusions. This prevents an initial claim of exhaustive lifecycle coverage, uniform approval, uniform storage semantics or equivalent behavior across clients. In particular, the README's approval doctrine is not a blanket implementation guarantee: the inspected server runner has distinct contribute, build and research paths (SRC-1, README.md, apps/server/src/ops/runner.ts).

Initial coverage and remaining inspection

The tracked tree was enumerated at the frozen commit. The following map records inspection status, not a source-file allowlist.

External dependencies are not separate analysed systems: the Node/Express and MCP SDK runtime dependencies, model providers such as Ollama or optional cloud services, and enclosing clients such as Claude Code provide execution infrastructure. MindBase's shipped adapters, commands and hooks remain inside the boundary; the providers' model internals and the hosts' scheduling, permissions and prompt-obedience enforcement do not. This prevents attributing host-level enforcement or provider quality to MindBase itself. These dependency surfaces are identified by SRC-1, apps/server/src/index.ts, apps/mcp/src/index.ts, README.md and apps/plugin/agents/builder.md; their external implementations were not acquired.

The frozen evidence is the acquired Git checkout at c5273d2336c4d686df41c2a1495f14f085707f51, verified as HEAD with no working-tree changes reported. The applicability cutoff is 2026-11-08, the supplied run date, not a claim about the date of the commit. Later repository revisions, linked external idea sources, npm releases, live websites, user knowledge directories and provider services are excluded from evidence. Their absence prevents conclusions about current releases, actual user retention, live deployment or model-output quality. There is no unavailable repository file established by this inspection. No target program, tests, examples or model calls were executed. The overall tier is code-grounded; there is no registered observed run or causal experiment.

Source register