Gas Town - multi-agent workspace manager
| Rating | Summary | |
|---|---|---|
| Quality | solid (17/24) | Actively maintained with a v1.1.0 release and rich documentation, held back by a missing dependency manifest and no visible test infrastructure. |
| PAI Relevance | integrate (0.50) | Addresses a genuine gap in PAI's multi-agent coordination at scale; the Go CLI is subprocess-wrappable, and the Seance and git-backed persistence patterns are directly applicable to PAI's Delegation and Agents subsystems. |
17/24 — maintained / adequately-documented / solid
Failed:
Passed:
Failed:
Passed:
Failed:
Passed:
| Dimension | Score | Assessment |
|---|---|---|
| Harvest Value | 1 | The Seance pattern (discovering predecessor sessions via .events.jsonl logs to recover prior decisions and context) is directly applicable to PAI's ContextSearch skill and cross-session memory architecture; the git-backed Hooks persistence model is worth studying as an alternative to PAI's current file-based WORK/ substrate. |
| Integration Readiness | 1 | Gas Town is Go-based rather than TypeScript, but exposes a clean gt CLI that PAI can call via subprocess from skills or hooks with moderate adapter code — the same pattern PAI uses to wrap other external binaries. |
| Overlap Risk | 1 | Partial overlap with PAI's Delegation skill (parallel work routing) and Agents skill (agent composition and voices); Gas Town's persistent workspace scoping, Mayor coordinator pattern, and Refinery merge queue are not covered by existing PAI capabilities. |
| Gap Fill | 1 | PAI has limited coverage in coordinating 10+ concurrent long-running agent sessions with isolated persistent state and automatic recovery; Gas Town's Mayor/Rig/Polecat model and three-tier watchdog (Witness/Deacon/Dogs) address a real coordination gap at scale that PAI's current Delegation skill does not handle. |
Composite: 0.50
Capture-to-Knowledge Pipeline's clean-room Haiku validation step: Clone Gas Town, initialize a workspace for the pipeline, and define a named agent identity for the validation role with the pipeline's working directory as its root. Subsequent Haiku validation sessions will bootstrap via the Seance pattern from .events.jsonl — no manual context-setting at session start, and every validation decision is automatically git-committed to an auditable, replayable log.
Running Claude Code agents concurrently across Conservancy, Aphoria, and fab: Register each active repo as a Gas Town managed workspace and enable the three-tier Witness/Deacon/Dogs watchdog. Each project gets an isolated git worktree; a crashed or silent agent is restarted with its last persisted state rather than silently dropped — making 5–8 simultaneous sessions across those repos reliable without per-session babysitting.
Conservancy editorial pipeline handoffs between recurring LLM passes: Define named Gas Town mailboxes for each pipeline stage (intake, editorial, publish-gate) and wire the pipeline so each agent deposits a structured handoff message on exit. The next agent reads it on startup rather than requiring a re-briefing prompt — a session crash or timeout no longer breaks the pipeline's continuity, and the full handoff history is git-backed for rollback.
Category: AI Agent Frameworks
In this category: VoltAgent--voltagent (appraised 2026-05-27, excellent 21/24, integrate)
Standing: Gas Town is purpose-built for persistent multi-agent coding workflows with deep git integration and fleet watchdog infrastructure; it complements VoltAgent's broader TypeScript agent platform by operating at a different layer — Gas Town manages workspace lifecycle, session persistence, and merge queuing while VoltAgent handles agent composition and routing abstraction.
Density: 8/10 — Available: README (8KB, substantive with architecture diagrams, concept glossary, and multi-path install instructions), repository metadata (stars, forks, dates, license, release history), landscape cluster data, prior appraisal record. Missing: go.mod dependency manifest, test file listings, CI workflow files, full docs/ directory contents, CHANGELOG.
This repo carried a prior landscape score of poor (4.1/24) — that score is inconsistent with the current README depth, v1.1.0 release, and exceptional adoption signal (15K+ stars in ~5 months). This appraisal supersedes the prior record with a full probe-based evaluation. The Wasteland federated coordination network (DoltHub-backed, multi-dimensional reputation stamps) is architecturally ambitious and appears partially speculative; it was not penalized in quality scoring but warrants monitoring for delivery. The macOS code-signing friction (unsigned Go binaries SIGKILL'd by Gatekeeper) is a genuine usability barrier for the primary target audience and should be tracked as a recurring caveat.