NanoClaw proposes persistent coding sessions with an optional remote terminal

NanoClaw opened a two-part code-mode stack on September 12 to move beyond its existing chat-loop runner. The core proposal gives an agent group a persistent Claude Code session inside its container, with a workspace that an operator can inspect and reattach to. A separate community-portal proposal supplies an opt-in remote terminal and an abstract chat-surface provider. Both remain open. The split is useful product information: a persistent local coding session, a remotely accessible terminal and a platform-specific conversation are different capabilities, with different prerequisites and exposure.

The facts

- Code mode is selected per group and defaults off; its second runner uses Claude Code under tmux in the existing agent image. - The proposed sandbox commands create, list, attach, inspect status and diff, or stop a session; the code conversation starts and resumes by its own ID. - Sandbox commands are host-only, and the permission posture is supplied through host-owned read-only settings rather than mutable agent instructions alone. - The core adds approval cards for boundary crossings and mailbox delivery into the persistent session. - The companion remote terminal uses a loopback SSH door with public-key authentication, a waiting room for unknown keys and no general shell, forwarding or subsystem access. - Concrete chat-platform adapters and development environments are outside the core; remote access must be enabled, and session surfaces depend on a managed platform registering its side.

Why it matters

This brings NanoClaw closer to a supervised coding workstation that can remain useful after the person disconnects. It also makes the small-core philosophy more demanding: optional modules must agree on lifecycle, permissions and session identity without silently taking ownership from the host. The current README supplies that modular baseline, while the proposed seams show how persistent execution and remote access might be added without turning every chat group into a terminal service.

Current

Inspected on 2026-09-13. The NanoClaw release baseline is v2.3.0 published 2026-08-24T11:45:32Z. The main source was open (not merged) when captured. Proposals, merged mainline changes and packaged releases are different availability states; the linked release baseline does not imply that open proposals have shipped.

Evidence

Primary evidence: nanocoai/nanoclaw PR #3783 (https://github.com/nanocoai/nanoclaw/pull/3783). Supporting context: nanocoai/nanoclaw PR #3784 (https://github.com/nanocoai/nanoclaw/pull/3784); NanoClaw — official product README (https://raw.githubusercontent.com/nanocoai/nanoclaw/main/README.md). Sources were read directly through public pages or GitHub API bodies. Contributor test results are attributed claims, not independently reproduced results.

Operator take

Review the stack in order. Prove a default chat group remains unchanged, then exercise create, detach, idle, reattach and stop for a code-mode group before enabling the remote door. Test key revocation while a cold sandbox is being attached, not merely whether a known key connects. The proposal reports versioned seam checks and real-client tests, but those do not amount to a verified deployment on a reader's host or a completed chat integration.

Caveat

Both pull requests are open and stacked. The core explicitly excludes remote terminals; the companion consumes its interfaces but excludes concrete chat platforms. The image needs tmux and a rebuild, and migration behavior depends on the host database mode. Contributor test results are attributed evidence, not a release or isolation certification.