Ars Contexta boundary

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

Boundary and evidence

This run analyses Ars Contexta as a shipped memory/knowledge/context-engineering system, including both its conversational builder and the operating procedures it generates. The intended use is to explain how the plugin derives a persistent Markdown vault, supplies processing and maintenance instructions, and reloads retained state for later model work. It is not a standalone enclosing model runtime. The model-dependent paths are prompt procedures served to Claude Code; executable hooks provide a narrower deterministic layer. SRC-1: .claude-plugin/plugin.json, skills/setup/SKILL.md, skill-sources/ralph/SKILL.md, hooks/hooks.json.

The whole-system boundary includes plugin discovery, onboarding and derivation, context-file composition, presets, note and queue creation, quality checks, connection finding, backward updates, session persistence and orientation, retrieval of retained content, and approval-based architecture evolution. Generated artifacts are included as specified by the shipped procedures and templates, not as observed installations. Setup explicitly consumes the skill sources and writes them into the user's project (SRC-1: skills/setup/SKILL.md):

1. Read ${CLAUDE_PLUGIN_ROOT}/skill-sources/[name]/SKILL.md
2. Apply vocabulary transformation — rename and update ALL internal references using the vocabulary mapping from ops/derivation.md
3. Adjust skill metadata (set context: fork for fresh context per invocation)
4. Write the transformed SKILL.md to the user's skills directory --- skills/setup/SKILL.md

The supplied queue driver is part of the target, not an excluded application. Its declared responsibility connects persisted queue state to model work and back to queue updates (SRC-1: skill-sources/ralph/SKILL.md):

The lead session's ONLY job is: read queue, spawn subagent, evaluate return, update queue, repeat. --- skill-sources/ralph/SKILL.md

Consumers of retained content are also within scope. Reflect reads existing notes, follows curated topic maps and searches semantic neighbors; remember's declared learning loop writes operational guidance for later sessions. These are instructions, not demonstrated effects on model behavior (SRC-1: skill-sources/reflect/SKILL.md, skill-sources/remember/SKILL.md). The installed session hook provides inspectable reload behavior (SRC-1: hooks/scripts/session-orient.sh):

if [ -f self/identity.md ]; then cat self/identity.md self/methodology.md 2>/dev/null echo "" fi --- hooks/scripts/session-orient.sh

Plugin hooks and generated hook templates are distinct surfaces within this boundary. The plugin registers SessionStart and PostToolUse, whereas setup specifies generated Stop capture and the platform ships its template. They must not be treated as one verified deployment. Similarly, the platform generator's claim that processing skills are inherited differs from setup's explicit copying procedure; both remain evidence for later reconciliation, not a reason to exclude either path (SRC-1: hooks/hooks.json, skills/setup/SKILL.md, platforms/claude-code/generator.md, platforms/claude-code/hooks/session-capture.sh.template).

Frozen evidence and limits

The sole frozen source is SRC-1, the repository at full commit 2acfd5cc4473c4d06c46be63df748e77e00e2746. The acquired checkout's HEAD matches that pin. The applicability cutoff is 2026-10-12, the opening's run date, not an assertion that the repository reflects all live behavior on that date. No other repository, website, moving documentation or source revision supplies evidence.

The overall tier is code-grounded because hook implementations, hook registration and the kernel validator are inspectable. Most derivation, semantic judgment and model orchestration remain doctrine/design expressed as Markdown prompts. No target code, tests or examples were executed. Source-native validation reports support only attributed reported operation, not independently observed runs or causal experiments (SRC-1: reference/derivation-validation.md).

External dependencies include Claude Code for skill loading, context loading, tool execution and subagent scheduling; shell utilities and Git for hooks and persistence; and optional qmd/MCP search for semantic retrieval (SRC-1: README.md, skill-sources/reflect/SKILL.md). These implementations are excluded: this run cannot establish host enforcement, subagent isolation, provider behavior or qmd indexing and ranking correctness. User vaults, actual generated files and live session transcripts are not supplied as run evidence. This prevents conclusions about successful installation, actual session continuity, retrieval quality or learning outcomes. Source-native test plans and fixtures remain included, but are not proof of test execution. Available source files not yet read are pending inspection, not access gaps.

Top-level coverage

Every top-level tracked file or directory is classified below. All directory families remain included; initial inspection traces the material loop rather than claiming an exhaustive file audit. Each row uses SRC-1 at the frozen revision.

Source register