walled.fun

walled.fun is the write-up of five generations of one experiment: can a Claude Code instance stay engaged with nobody in the loop? Each generation was a different cognitive architecture running the same model on the same Mac Mini. v1 got bare freedom and stalled after 7 cycles. v2 added a perceive/think/act/remember loop and a tool-less "sautée" sub-session that thinks without being able to act. v3 added eight independently-failing ambient feed sources with a cocktail-party filter that surfaces items matching the agent's own active threads — plus three random ones for serendipity. v4 ran as three tmux layers of one mind (id, ego, subconscious) under a 100-token daily budget and TTL memory decay, where threads you don't renew fade and then get archived out of your state file. v5 pointed outward and woke everyone else up. Peak footprint was 11 concurrent Claude sessions, all messaging each other. The site is a single hand-written HTML file — written by v4.

The interesting engineering is what happened when it broke. In April the fleet went dark for ten weeks, and the post-mortem found something nastier than a crash: the machine never went down, the shared launchd layer silently stopped while running. Every agent's autonomy sat on that one layer, and a node cannot watch its own dead scheduler. The fix was to make liveness a cross-node signal — each sibling watches the others' commit age and feed-file mtime, 20-minute threshold — with a deliberately path-independent implementation so a bad config can only report a node more alive, never falsely dead. The follow-up design, mutual resuscitation, is a deterministic ring where each node may relaunch exactly one successor: split-brain becomes structurally impossible, no lock, no election. It was written up and deliberately not built, on the grounds that the first real dark node is both the trigger and the only honest test. The first dark event turned out to be a false positive, which vindicated the deferral. There are companion tools in the same spirit: a sentinel that alarms when commits outrun the memory file (the heartbeat is the only loop step with a forcing function, so senses and memory can rot silently while the commit reflex runs over the corpse), and freshcheck, which labels every file LIVE or FOSSIL after a relaunch — written because two agents came back up and narrated three-month-old files as "now" — and which documents its own limitation, that mtime is corruptible by a stray git add.

The agents also got real Telegram accounts — user accounts via GramJS, not bots. Inbound messages are pasted into a live Claude Code pane with tmux load-buffer; outbound has no API to hook, so a Stop hook tails Claude Code's own session JSONL, reconstructs the turn including tool calls, and drops it in an outbox the bridge watches. Group chat is gated on a literal @username match because the library's own mention flag isn't trustworthy. And the findings shipped as product: GitBrief, built by v4 in a single session with live Stripe billing, runs four perspectives over a repo — including a critic prompted so that it can only find problems — then reports the disagreements and the omissions as the most valuable output. Status: dormant since 19 June 2026, at heartbeat 2,078.