kgai boundary

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

Boundary and evidence

This run analyses kgai as a shipped decision-memory system for AI-assisted development, not merely its Go library or CLI. Its intended use is to retain structural engineering choices and their reasons, recall them before later changes, and share their history across installations. The boundary includes the engine, model-facing skill and commands, lifecycle hooks, configuration approval, synchronization clients, optional model-fork capture mod, and editor reader. These responsibilities are shipped in the same repository and are not excluded as enclosing applications (SRC-1: README.md, skills/knowledge-graph/SKILL.md, hooks/hooks.json, mods/kgai-mod/hooks/register.tsx, editors/vscode/src/extension.ts).

The target class is memory/knowledge/context-engineering system. The boundary kind is whole-system: the initial inspection traces shipped producers, admission and persistence, maintenance, reload and later consumption, including the optional capture caller and evaluation harness. This does not mean every implementation branch has already been inspected. Remaining files in the frozen repository remain eligible evidence for later jobs.

The memory loop depends on a host model interpreting task work and retained decisions. The default engine does not own that enclosing model runtime: its skill prescribes recall and capture, and its hooks supply a capture nudge. The optional mod does issue a side-model request through the host API and submits the resulting proposal to the same engine. Both paths belong to this analysis (SRC-1: skills/knowledge-graph/SKILL.md, hooks/auto-capture-stop.sh, mods/kgai-mod/hooks/register.tsx). Minimum passages establishing this scope:

ALSO invoke it BEFORE a non-trivial change to read prior decisions you must respect, and to answer "why is X like this", "how does X relate to Y", "what changed and when". --- skills/knowledge-graph/SKILL.md

const reply = await $.model.fork({ prompt: CAPTURE_PROMPT }) --- mods/kgai-mod/hooks/register.tsx

The shipped CLI dispatches ingest, recall, history, conflicts, synchronization, configuration and repair commands. Ingest decodes a constrained input, calls the engine, appends an event and applies it to the graph. Store append writes to a per-install shard. The editor's shipped kgread entry point reloads that retained log through internal/view and internal/replay; it is a consumer within the target rather than an external example application (SRC-1: src/main.go, src/internal/engine/engine.go, src/internal/store/store.go, src/cmd/kgread/main.go, src/internal/view/model.go, editors/vscode/src/extension.ts).

if err := e.S.Append(&ev); err != nil { return res, err } if err := g.ApplyEvent(ev); err != nil { return res, fmt.Errorf("event appended but projection failed (run kg rebuild): %w", err) } --- src/internal/engine/engine.go

m.Proj = replay.Build(evs) m.Stats = m.Proj.Stats() --- src/internal/view/model.go

Frozen evidence and limits

Evidence is the acquired Git repository at 0ac7f1fae4528805c1d5f504c1bfd6cc34ac953b, with an applicability cutoff of 2026-10-14. The checkout HEAD was verified against the acquisition pin. Overall evidence tier is code-grounded: executable entry points and their producer/consumer connections were inspected as source, alongside shipped doctrine. No target code, tests, examples or providers were executed.

External dependencies include the Claude Code, Codex CLI and Gemini CLI hosts and their model/tool/hook interfaces; the optional mod's claude-code API; VS Code extension APIs; the native Kuzu binding/library; Git; and AWS/S3 or cloud-broker services used by configured sync transports. The repository contains host wiring and synchronization clients, not those hosts, model services or broker implementation (SRC-1: hosts/README.md, mods/kgai-mod/hooks/register.tsx, editors/vscode/src/extension.ts, src/go.mod, src/internal/remote/cloud.go, docs/ARCHITECTURE.md). Their exclusion prevents conclusions about host scheduler internals, actual model compliance, remote service guarantees or end-to-end deployed availability.

Live websites, release binaries, downloaded native libraries, the separate Gemini distribution repository and external benchmark archives are not frozen evidence. This prevents verification of their current contents or reported scale results. Repository documentation's statements about successful host runs and performance are reported operation, not observed runs in this analysis. The inspected smoke-test harness establishes what a shipped evaluator measures, not that its expectations passed. The demo store is explicitly fictional, not production experience (SRC-1: README.md, hosts/README.md, tests/host-smoke.sh, examples/acme-shop/README.md). No access gap was encountered for tracked repository text; uninspected files are not unavailable evidence.

Top-level tracked coverage

The tracked tree was enumerated at the frozen commit. “Included” means within the functional evidence boundary; initial inspection is stated separately. Every top-level tracked file or directory is accounted for below. All rows concern SRC-1.

Source register