Gas Town - multi-agent workspace manager
Repository is ~5 months old (created 2025-12-16) but has reached v1.1.0 as of 2026-05-07 — a genuine versioned release rather than a placeholder tag. The README describes a stable feature surface: Rigs, Polecats, Hooks, Convoys, Molecules, Refinery, Escalation, Scheduler, Seance, and Wasteland are all documented with design docs linked. 187 open issues is elevated but appropriate for an actively developing project at this adoption level. Not yet at 1.x long-term-stable confidence, but clearly past prototype stage — call it late beta / early stable.
README is among the more thorough encountered: architecture diagram (Mermaid), concept-by-concept breakdown with emoji anchors, prerequisites matrix, four distinct install paths (Homebrew, npm, go install, Docker Compose), and a quick-start sequence. Linked sub-docs exist for Molecules, Escalation, Scheduler, Wasteland, and a Glossary — evidence of real documentation investment beyond the README. Deducting one point because the dependency manifest is unavailable, API/programmatic docs are not referenced, and the README truncates mid-sentence in the Quick Start section, suggesting the scrape cut off content that exists upstream.
Go is a strong choice for a CLI workspace manager — good concurrency primitives, single-binary distribution, fast compilation. The architecture shows thoughtful separation (per-rig Witness, cross-rig Deacon, Refinery merge queue, capacity-governed Scheduler). However, no dependency manifest was available for hygiene inspection, no test coverage or CI badge is referenced in the README, and the project depends on several external binaries at specific versions (Dolt 1.82.4+, beads 0.55.4+, sqlite3, tmux, Claude Code CLI) rather than internalizing them — which creates a brittle dependency chain that is difficult to audit. Score is held at 3 pending visibility into test infrastructure.
Last commit is 2026-05-12 — the same day as this appraisal. v1.1.0 was released five days prior (2026-05-07). This is a project with someone actively at the keyboard. Commit cadence appears daily or near-daily based on the release recency gap. No signals of abandonment whatsoever.
15,110 stars and 1,387 forks accumulated in approximately five months is an exceptional trajectory — this is well beyond organic growth and suggests sustained community momentum, likely amplified by coverage in AI-developer circles. The forks-to-stars ratio (~9%) is healthy and suggests developers are actually working with the code, not just bookmarking it.
Overall: 4.1/5
Category: Multi-Agent Orchestration Known alternatives in vault: gordonbrander--busytown (2.5/5, the only prior vault entry in this category) Differentiation: Gas Town is categorically more complete than busytown. Where busytown appears conceptual/lightweight, Gas Town ships a full supervisor hierarchy (Mayor → Deacon → Witness → Dogs), persistent git-worktree state via the Beads ledger, a Bors-style merge queue (Refinery), multi-runtime support (Claude Code, Copilot, Codex, Gemini), an escalation routing system with severity tiers, a capacity governor (Scheduler), session-continuation via Seance, and a federated work network (Wasteland). There is no meaningful overlap — Gas Town is the production-grade answer to the same problem busytown sketches. Alternatives do not meaningfully outperform it in any documented dimension at this time. Partial overlap with NorthwoodsSentinel--brook (fleet monitoring) and VoltAgent--voltagent (agent engineering framework) but those serve adjacent, not identical, concerns. Gap or crowd: The Multi-Agent Orchestration category had a single thin-coverage entry at 2.5/5. This fills the gap decisively and upgrades the category's top pick substantially.
Score: 5/5
Harvestable: (1) Seance pattern — .events.jsonl log-based session discovery for querying predecessor agent context is directly extractable as a PAI skill hook. (2) Beads ledger — git-backed structured work-item tracking with prefixed IDs is a portable knowledge-graph primitive. (3) Three-tier watchdog topology (per-agent Witness → cross-rig Deacon → dispatched Dogs) is an architecture pattern worth adopting for any PAI system that runs background agents. (4) Escalation severity routing (P0/P1/P2 → Deacon → Mayor → Overseer) maps cleanly onto PAI interrupt handling. (5) Convoy/mountain autonomous skip logic for epic-scale tasks. (6) Refinery bisecting merge queue as a pattern for agent output verification before integration.
Integration path: Gas Town can serve directly as the orchestration substrate for a PAI system's agent fleet — the Mayor role maps to a PAI coordinator skill, Polecats map to ephemeral task-agents, and Rigs map to project-scoped contexts. The gt seance and gt escalate commands are immediately usable as PAI tool-calls. The Wasteland federated layer offers a longer-term integration point for cross-system knowledge exchange. Install via Homebrew, configure a Town workspace, and wire gt mayor attach into the PAI bootstrap sequence.
Overlap with existing: gordonbrander--busytown is the only vault entry in the same category — Gas Town supersedes it without meaningful redundancy. NorthwoodsSentinel--brook handles fleet monitoring (a subset of what Gas Town's Witness/Deacon subsystem does internally), creating partial functional duplication if both are in active use. VoltAgent--voltagent overlaps on agent-engineering primitives but operates at a different abstraction level.
Adoption cost: Significant. Requires Dolt, beads/bd CLI, sqlite3, tmux, and at least one agent runtime (Claude Code CLI) as external prerequisites at specific versions. The Gas Town conceptual vocabulary (Rigs, Crew, Polecats, Hooks, Beads, Convoys, Molecules, Wisps) has a real learning curve. Docker Compose path reduces environment friction but does not eliminate the mental model investment. Integration into an existing PAI system means either adopting Gas Town's workspace conventions wholesale or surgically extracting patterns — both paths are non-trivial.
Gas Town is the most feature-complete multi-agent orchestration system in the vault by a wide margin, and its adoption trajectory (15k stars in five months) signals genuine practitioner uptake rather than hype-only interest. The primary risks are dependency brittleness (five external binaries at pinned versions), lack of visible test infrastructure, and the early-stage nature of some subsystems (Wasteland in particular reads as aspirational). The Seance and Beads patterns alone justify vault inclusion for harvest purposes even if full integration is deferred. Recommend adding with a note to monitor the v1.2.x cycle for stabilization signals on the Wasteland and Molecules subsystems before committing to deep PAI integration. Supersedes gordonbrander--busytown as the category top pick.