muxel Multi-Agent Workspace

The evidence openly identifies high-impact capabilities including PTYs, agent processes, environment variables, Git worktrees, the gh CLI, SSH, roaming-layout synchronization, shared MEMORY.md, scheduled runs, and in-app updates. Release write permission is scoped to the release job, and the Windows installer is described as per-user and non-admin. Deductions apply because there is no evidence of agent sandboxing, per-operation confirmations, precise remote-sync data boundaries, or protection for sensitive environment variables and SSH credentials. Dependency controls include a stated Cargo.lock pin and versioned CI actions, but the root GPUI Git dependencies lack explicit rev fields and no audit, vulnerability-response, or supply-chain verification policy is shown. The worktree review/dispose flow offers limited recovery context, but no general undo or backup design is documented. ProjectHax LLC, the repository, and major upstreams are attributed, though the publisher is unverified and no complete third-party attribution inventory is present.

The README, Cargo workspace, and CI/release workflows are broadly consistent about platforms, builds, package formats, and release mechanics. CI includes formatting, Clippy, builds, and tests, while release jobs contain explicit failures or warnings for unresolved AppImage dependencies, missing artifacts, and absent changelog sections. Deductions apply because most product capabilities are asserted in the README without the corresponding implementation or tests in the supplied evidence; availability still depends on Git upstreams, system libraries, package managers, and network-fetched installers. Visible failure handling is concentrated in packaging rather than user-facing runtime failures involving SSH, PTYs, updates, scheduling, or persistence.

The audience and scenarios are exceptionally clear: concurrent coding agents and shells, cross-project monitoring, local and remote development, worktrees, scheduling, and cross-platform desktop operation. Editable presets cover programs, arguments, models, system prompts, environments, and status markers, while environment guidance distinguishes operating systems, architectures, graphics stacks, and native-build constraints. Deductions apply because capability boundaries are incomplete: Windows tmux limitations and signing fallback are stated, but supported agent protocols, remote security boundaries, and degradation behavior are not. Loop timing is described at a useful high level, yet concurrency, missed-run, timezone, and duplicate-prevention semantics are absent.

The README has clear feature, installation, source-build, CI, release, and licensing sections. Product and crate naming is consistent, workspace version 0.2.0 aligns with a vX.Y.Z tagging scheme, and installation covers three operating systems, package choices, source commands, and Linux dependencies. GPL-3.0-only is consistent between Cargo metadata and the full license, the dual-license arrangement is plainly explained, and releases derive notes from versioned CHANGELOG sections. Deductions apply because FEATURES.md, LICENSING.md, CONTRIBUTING.md, CHANGELOG.md, and the toolchain pin are referenced but not supplied; routine usage examples and an FAQ are absent; limitations are concrete but not comprehensive. ProjectHax LLC is identified as author and commercial licensor, but no maintainer contact, support expectation, response policy, or security-reporting path is shown.

The described output is a directly usable native workspace combining real terminals, status displays, notifications, editing, diff viewing, worktrees, and persistence, offering credible marginal value for managing multiple coding agents. GPL use is available without a license fee, prebuilt packages are broad, and the costly first GPUI build plus platform dependencies are disclosed. Deductions apply because there are no demonstrated end-to-end workflows, substantive examples, benchmarks, or user evidence establishing usability or status-detection quality. Most differentiated value remains a README claim, and ongoing CPU, memory, network, remote-host, notification, and multi-agent service costs are not quantified.

Version, license, authorship, workspace structure, platform matrix, package formats, signing conditions, and release rules can be traced across Cargo metadata and both workflows. The release workflow also documents a concrete AppImage compatibility failure and associated regression guards. Deductions apply because central claims such as status detection, shared memory, SSH roaming layouts, editor behavior, and worktree interactions appear only in the README, with no supplied implementation or focused tests for corroboration; statements such as “verified locally” lack accompanying static evidence. Facts and caveats are generally distinguished, but feature marketing does not systematically separate implemented behavior, design intent, and platform exceptions.