Gas City

Since April 2026, more than 200 of my pull requests have merged into Gas City, a Go runtime for building orchestrations of coding agents. A supervisor reconciles desired against actual state every tick; agents run as tmux-hosted sessions inside per-repo rigs; work lives in Dolt-backed issue beads; formulas compile multi-step workflows that get slung to workers; and a mail plus external-messaging layer connects agents to each other and to channels like Slack. The maintenance work spans most of that surface: the reconciler, dispatch, session lifecycle, the beads layer, the HTTP API, formulas, and external messaging.

Maintaining an agent orchestrator is mostly defending one property: the city keeps running unattended. The early arc was crash-and-wedge work: recovering reconciler panics so one bad rig can’t cascade a whole city down, reaping session beads whose tmux session had died, preserving managed sessions across supervisor restarts, tearing down sessions that failed mid-create, and fixing the wedge where gc session close could strand its own work. The later arc was performance under real load, on cities whose stores had grown past two hundred thousand beads: hydration-free parallel status counts made the city-status endpoint roughly 100x faster, bead reads and formula feeds became bounded and deterministic, and the workflow-snapshot scan stopped re-listing metadata it already held.

A second thread is the fail-loud posture. The bd CLI silently falling back to on-disk mode became a loud error; an import revert that reported success became an error; formula filters that silently returned empty on unexpected output got validation; a malformed agent-settings override now says exactly what is wrong. And because agents both produce and consume the system’s text, some hardening is security-shaped: stripping system-reminder breakout sequences from user-controlled fields, rejecting worktree spawns under stale ancestor .git pointers, and symlink-escape checks on dispatch paths.

Maintainership is also the review side, and I’ve reviewed more than 200 contributor pull requests: triaging incoming work, mapping blast radius across the supervisor’s shared code paths before accepting a change, and keeping CI honest by quarantining flaky integration tests with tracking anchors, running the full suite when cross-cutting core paths change, and replacing sleeps with drains in racy tests. The adjacent projects here grew out of the same work: the dashboard that watches a running city, the packs that connect it to Slack and GitHub, and the mem benchmark that mines a city’s work exhaust for agentic memory.