← the Atlas

Remote terminals — the future work

Analysis·seedling·proposed·

What to build on top of the multi-host switch — candidate features crossed from ecosystem research (17 source-grounded capabilities) with kolu's shipped primitives, each with its mechanism, demand proof, and honest gates.

What this note is. The switch (W4) makes kolu multi-host. This note is the researched answer to “what next?” — built from two ground-up passes, not memory: an inventory of every shipped primitive (the hosting-side final API) and already-named future item in this repo, and a source-grounded survey of what tmux/mosh/Eternal Terminal/Zellij/VS Code Remote/Coder/Gitpod/Tailscale/Warp/tmate/Blink/Wave actually prove users want. Each candidate below names its demand proof, the shipped mechanism it rides, and its gates. Nothing here is committed work — this is the menu for the next planning pass.

The unfair advantage — why kolu can do what none of them can

Every surveyed product solves one slice: tmux owns persistence, mosh owns roaming, tmate owns sharing, Coder owns fleet cost, Omnara/Warp-Remote own agent notification. Kolu’s three primitives — a daemon on every host that owns the PTYs, a warm multi-host pool, and typed per-terminal agent urgency — cover all the slices at once, and enable one thing structurally impossible elsewhere: products that know what an agent is waiting for. No surveyed tool can rank a fleet by “which agent needs a human,” because none of them know what an agent is. That is the axis to build along; everything below is ordered by how directly it exploits it. (Tools surveyed, for the record: tmux + tmux-resurrect/continuum · mosh · Eternal Terminal · Zellij (incl. its web client) · VS Code Remote-SSH + Tunnels · GitHub Codespaces · Coder · Gitpod · DevPod · Tailscale (SSH, session recording, Taildrop) · Warp (Remote Control, session sharing) · Termius · Blink Shell · Upterm · tmate · Wave Terminal — plus the agent-attention products: Omnara, Happy Coder, Claude Code’s own agent view.)

Tier 1 — the attention products (build on W5’s wire, exploit the advantage directly)

F1 · W5 cross-host attention (already planned — the ecosystem research now validates it hard). PWA badge summed across bindings, per-host switcher-chip counts, OS notification whose click = switch + focus. Whole products exist doing only this for single machines (Omnara, Happy Coder, Warp Remote Control, Claude Code’s own agent view) — proof the pain is real and monetizable; kolu gets the cross-host version from one urgency subscription per warm binding. Deliberately urgency-count-only on the wire (bounded cost, and immune to the clock-skew gap). Gate: W4 merged. Size: small — the wire member and the fold exist.

F2 · The cross-host dock (B-lite) (already named as the first escalation; the research renames its category: a fleet dashboard ranked by what matters). Foreign hosts’ agents listed in your dock, urgency-ranked, click = switch + focus. Coder’s workspace dashboard and Termius host groups prove fleet-view demand; none can rank by agent attention. The terminals contract has reserved the host axis since W1. Gate: demand — “if switching proves insufficient” (the plan’s own bar). Size: medium, pure aggregation layer, zero daemon changes.

F3 · Agent-aware idle suspension (new — the sharpest differentiated idea the research surfaced). Not sleeping terminals, one level up: kolu already has per-terminal sleep (records persist, wake resumes the agent) — but sleeping is manual today, and F3 is about the machine (stop paying for an idle cloud host). Two layers, same signal: an auto-sleep policy for terminals (sleep a terminal when its agent completes and nothing awaits) is the natural precursor and could ship first; Coder/Gitpod/DevPod all auto-stop idle environments to save money, and their users file the same gap: the tools can’t tell “human stepped away” from “agent mid-task” (a cited Coder discussion asks for exactly this). Kolu knows — padi distinguishes an awaiting agent from a working one per terminal. Feature: for suspendable hosts (cloud boxes, pu-style), suspend when no agent is mid-task AND nothing is awaiting input, wake on demand through the warm pool. Gates: W4; a host-lifecycle verb kolu deliberately doesn’t have today (the registry has no controls — this feature is the first honest consumer for one, per prove-then-extract); real demand from someone running cloud hosts. Size: medium-large.

Tier 2 — the daemon-ownership dividends (persistence beyond what tmux can express)

F4 · Agent-aware resurrection — CORRECTED: mostly shipped already. The survey pitched this as the tmux-resurrect killer (“restore the agent, not just the terminal”) — but a ground check shows padi already has it: wake reads the frozen agent identity back and re-launches via resumeCommand (endpoint.ts), with resumeIds as the per-terminal resume opt-in on session.restore (surface.ts). The honest residual is only framing and automation: making “your fleet rebooted; your agents came back mid-conversation” a demonstrated, documented cross-host story (and deciding whether post-reboot restore should offer resume by default). Size: small — polish on shipped machinery, not a feature.

F5 · Auto port-forward with preview URLs — promoted to its own note: port-preview, incl. the two-hop architecture (headless kolu + remote padi) with the visual explanation. Codespaces/VS Code prove the demand (detect localhost URLs in terminal output → clickable forwarded links); the agent-era version is sharper: agents start dev servers constantly and print the URL. padi watches PTY output (it already has activity sensors), announces detected ports on its surface, kolu-server proxies through the existing bound connection (the preview route already reads through the bound session — same shape, one more route). Click the chip → the remote dev server in a tab. Gate: W4 for the multi-host version (single-host works today). Size: medium. This is likely the highest delight-per-effort item on the list.

F6 · Session recording & replay (agent audit trails). Tailscale ships SSH session recording for compliance; kolu’s angle is stronger: transcript.exportHtml already exists for agent transcripts, and the daemon sees all PTY bytes — recording an agent’s terminal doubles as an agent-governance audit trail, a need the ecosystem is just waking to. Gate: demand-driven. Size: medium.

Tier 3 — already-named, unblocked mechanics (ship when convenient)

F7 · Honest clocks (the deferred W4 piece 5) — the named fast-follow: consume control.core.clockNow per bind (RTT-halved offset on the session), translate host-stamped times in the client’s formatters. Prerequisite hygiene for any cross-host feature that displays time; W5 was shaped to not need it. Size: small.

F8 · padi-tui –host (ledger L24) — the CLI’s remote leg, ready since W3.1. One correction from the ground-truth pass: L24’s text cites getHostSession, which S10 deleted — the implementation composes makeSession({ connectOnce: sshConnector({ binary: "padi", … }) }). A tui is a dial; it never converges (#1313). Size: small.

F9 · kolu-tui — the graduation proof already in Future work: consume padiSurface over socket/ssh without kolu-server; done only when that named path renders a live canvas. Unblocks L2. Gate: demand.

Parked with reasons (the research argued against, for now)

The constraint ledger (what any of this must respect)

The parity tail (remote-bind-parity) gates “kolu fully works over a remote bind” — #1701’s watcher fix, the one load-source dig, R3’s kill-fault primitive, and the R5 reconcile-poll series are the debts; the watcher genus especially bites anything that adds more remote watchers. The N-bindings standing cost stays bounded only if multi-host wires stay urgency-count-sized (the W5 rule — F2 must honor it too). The ssh-user caveat is the trust model until a sharing story exists. The drishti gate binds every @kolu/surface* touch. And the clock gap (F7) silently corrupts any cross-host timestamp display until paid.