LACP boundary

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

Boundary and evidence

This run analyses LACP as shipped in the registered repository, not only its memory library or Claude plugin. The intended use is to establish how its agent execution controls, coordination, retained state, memory maintenance and later consumption fit together. The target is in scope because shipped paths both wrap model-dependent coding agents and implement a model/tool loop themselves (SRC-1, bin/lacp, tui/repl.py, tui/providers.py).

Target class: agent operating layer. Here this means a shipped local system that configures and coordinates agent operation, governs selected execution paths, and supplies context and retained state. It does not imply that every path is an external-runtime client. The class includes LACP's native REPL runtime alongside its host integrations, orchestration and improvement plane.

The documentation declares a narrower host relationship:

LACP is a local execution wrapper and control plane. The coding-agent host still owns its model, tools, session, and interface. LACP owns the contract around a run: where it may execute, which risk tier applies, when approval is required, what evidence must be kept, and how a failed change is recovered. --- README.md

That is doctrine, not a sufficient description of every shipped entry point. The top-level dispatcher also launches a native REPL without an external coding-agent CLI (SRC-1):

if [[ $# -lt 1 ]]; then # No args: launch LACP native REPL exec python3 "${ROOT}/tui/repl.py" fi --- bin/lacp

The REPL requests provider output with conversation state and tool definitions, dispatches returned calls, and supplies their results to another turn. Its Anthropic provider calls the model API (SRC-1, tui/repl.py, tui/providers.py):

async for event in self.provider.stream( messages=self.messages, system=self.system_prompt, tools=tools if isinstance(self.provider, AnthropicProvider) else None, ): --- tui/repl.py

with client.messages.stream(**kwargs) as stream: --- tui/providers.py

if tool_name.startswith("mcp_") and self.mcp_manager: result = self.mcp_manager.call_tool(tool_name, tool_input) else: result = execute_tool(tool_name, tool_input) --- tui/repl.py

Add tool results to conversation

self.messages.append({"role": "user", "content": tool_results}) --- tui/repl.py

Boundary kind: whole-system. All shipped functional paths at this commit are included: native REPL and dashboards; external CLI adoption and streaming; policy routing and execution; orchestration, harnesses and evidence gates; installation and configuration; learning-event capture, retrieval, evaluation, promotion and rollback; session hooks, handoffs and self-memory; ingestion, vault curation, indexing and retrieval; and automated improvement/evaluation. A shipped caller is not excluded because it supplies state or calls a helper. Alternative distribution paths, including the plugin's separate hook scripts, remain included rather than assumed equivalent (SRC-1, bin/lacp, plugin/hooks/hooks.json, Formula/lacp.rb, CLAUDE.md).

The initial state-loop trace reaches writers, maintenance drivers, reloaders and consumers. The native REPL's memory tools write/read keyed JSON entries; its session module saves and reloads conversations, and the REPL restores loaded messages. The stop hook records SMS episodes and updates its self-model; the session-start hook obtains SMS context. Brain expansion invokes synchronization, consolidation, curation, archival, indexing and health checks. A separate memory-pipeline driver invokes extraction, index building, promotion suggestions and retrieval benchmarking. These paths establish functional coverage, not that every store is mutually integrated or every documented learning phase is automatically invoked (SRC-1, tui/tools.py, tui/sessions.py, tui/repl.py, hooks/stop_quality_gate.py, hooks/session_start.py, bin/lacp-brain-expand, automation/scripts/run_memory_pipeline.sh).

Minimum retained evidence for the host-memory producer/consumer pair:

write_episode(episode) _debug(f"SMS: episode recorded (significance={significance:.2f}, files={len(files_modified)})") --- hooks/stop_quality_gate.py

from self_memory_system import build_session_context sms_context = build_session_context() if sms_context: injections.append((7, "sms", f"Agent memory (SMS):\n{sms_context}")) --- hooks/session_start.py

The improvement plane is also shipped, not an external operator responsibility: the idle agent constructs an experiment prompt, invokes Claude/Codex/Hermes, compares a new health score to its baseline and keeps or discards the change. Evaluation code and stored score artifacts belong to the boundary; stored scores alone do not establish causal improvement (SRC-1, tui/idle_agent.py, autoresearch/evaluate.py, autoresearch/evaluate_memory.py).

Top-level coverage

Each tracked top-level entry is classified below. “Included” fixes functional scope; it does not assert full-file inspection of every descendant. The inspected scope in the source register and the table distinguish initial tracing from material still to inspect. No material shipped path was inaccessible during boundary declaration.

Frozen revision, dependencies and limits

SRC-1 is frozen at e31f134b689afd6a71d94b0420682d1f528d6fa1; checkout HEAD was checked against the acquired revision. The applicability cutoff is 2026-10-14, the supplied run date. Evidence is code-grounded through static source inspection. No target command, tests, examples, providers or services were executed. This boundary supports implementation and intent findings, not verified installation, performance, real task outcomes or causal effects.

External host implementations (Claude, Codex, Hermes and other supported CLIs), model services/SDKs and authentication systems, Obsidian and its CLI, external MCP servers/indexers such as QMD, GitNexus and smart-connections, terminal/session tools, and remote sandbox services are dependencies, not independently frozen sources. LACP's adapters, configuration and prompts for them are included; their internal behavior, availability, security and cross-version compatibility are not established. In particular, external-host tool/session ownership must not be attributed to LACP's wrapper; the native REPL's own responsibilities must not be attributed to an excluded host (SRC-1, README.md, CLAUDE.md, tui/providers.py, bin/lacp-stream, bin/lacp-brain-stack, bin/lacp-sandbox-run).

Private installed settings, vault contents, session histories, credentials, live indexes and deployed schedules were not supplied. Their absence prevents conclusions about actual retained knowledge, operational coverage, retrieval quality and scheduling reliability. It is not an access gap in the registered repository. Repository-authored status and evaluation artifacts may support attributed reported operation only within their own stated conditions.

Source register