Auto-claude-code-research-in-sleep boundary

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

Boundary and evidence

ARIS is in scope as a model-dependent research workflow, not an enclosing agent runtime. Its intended use is to coordinate literature review, idea selection, experiment implementation, iterative review and optional paper production through host-executed skills. The source-native description is:

ARIS is a research harness: composable Markdown skills that orchestrate the ML research lifecycle through cross-model adversarial collaboration. --- AGENT_GUIDE.md

This classification rests on SRC-1, AGENT_GUIDE.md and skills/research-pipeline/SKILL.md. The latter is a shipped caller, not an external application: it chains /idea-discovery, /experiment-bridge, /auto-review-loop and optional /paper-writing, specifying artifact handoffs and per-stage state transitions. Its explicit orchestration contract is:

It orchestrates up to four major workflows in sequence. Workflow 3 (paper writing) is optional and controlled by AUTO_WRITE. --- skills/research-pipeline/SKILL.md

The functional boundary includes the repository's mainline and host-specific skill prompts, shared contracts, installer/update paths, reviewer bridges, persistence and validation helpers, monitoring, templates and tests. Shipped callers, retention maintenance and evaluators remain inside the target even when their implementation has not yet been inspected. The complete Git artifact is frozen; initial inspection establishes only a partial loop, not whole-system coverage.

The traced entry path is the Claude plugin manifest → setup command → registered Codex MCP bridge, alongside the research-pipeline skill → phase-state CLI → persisted JSON → resume selection. The bridge's complete implementation was inspected through request dispatch, subprocess launch, result conversion and saved thread options (SRC-1, .claude-plugin/plugin.json, commands/setup.md, mcp-servers/codex-exec/server.py). The phase helper's complete implementation was inspected through creation, status mutation, acceptance/provisional recording, persistence and resume consumption (SRC-1, tools/run_state.py). The caller explicitly requires re-auditing unfinished acceptance:

A stage left done (gate failed/ambiguous, or the run crashed before the gate) is re-validated on the next resume — the acceptance obligation is never skipped. --- skills/research-pipeline/SKILL.md

These are implementation and instruction observations, not evidence that a host followed the instructions or that recorded verdict identities are genuine. The model-family classifier imported by the phase helper has not yet been inspected; its classification accuracy is not established here (SRC-1, tools/run_state.py).

Initial coverage and remaining inspection

The following is an inspection map, not a source allowlist. Uninspected shipped paths remain eligible at the same registered commit. None is an access failure.

Inventory is based on SRC-1's tracked tree at the reviewed commit. This table explicitly leaves material loop paths uninspected; consequently the boundary is not whole-system. The optional knowledge subsystem is part of ARIS, not a separate frozen source. Its opening specification identifies persistent paper/idea/experiment/claim entities and helper-mediated operations, but the helper and downstream consumers need further inspection before claiming an implemented closed memory loop (SRC-1, skills/research-wiki/SKILL.md, AGENT_GUIDE.md).

Dependencies and time boundary

External dependencies include the enclosing agent hosts and their tool/skill dispatch, authenticated model services and CLIs, MCP transport clients, Python and the shell, project working directories, and task-specific compute and literature services. Setup explicitly requires an installed/authenticated Codex CLI and host MCP registration; the shipped bridge starts the external CLI rather than implementing its agent execution engine (SRC-1, commands/setup.md, mcp-servers/codex-exec/server.py). Host scheduling, context compaction, provider behavior and GPU execution are not attributed to ARIS itself. Their implementations are outside this source set, preventing conclusions about host isolation, model reliability and successful unattended execution.

The only evidence source is the acquired repository at full commit 26b95cfa0d8747078e9e43b42e20952709e561b8; checkout HEAD was confirmed against that pin. The applicability cutoff is the supplied run date, 2026-10-15, not a claim about when every upstream feature was deployed. No other repository, branch, live page or later revision is evidence. Overall evidence is code-grounded, with Markdown workflow instructions treated as doctrine/design rather than independently observed operation. No target commands, tests or model calls were executed, and no execution trace or causal experiment was supplied. This prevents claims of operational success, measured performance or causal effectiveness.

Source register