Mysti AI Coding Team

The default asks before edits and the product exposes read-only, ask-permission, and full-access levels; autonomous-mode safety classes, block patterns, audit logging, and active-child protection provide meaningful least-privilege, confirmation, and external-effect controls. Full-access and aggressive autonomy can nevertheless authorize broad actions. The README identifies file context, external CLIs/local endpoints, telemetry categories, and an opt-out, but does not provide a per-provider destination map, retention policy, or complete network inventory. Cursor and LocalAI keys can be placed directly in ordinary VS Code settings, while no SecretStorage, redaction, or log-leak controls are evidenced, so sensitive-data handling is thin. No lockfile evidence, vulnerability scanning, dependency policy, or supply-chain controls are supplied, yielding zero for dependency security. No file-change undo, snapshot, or recovery procedure is shown, so rollback is weak. Repository, author, company, contributors, and licensing are attributed, but publisher identity remains unverified and the attribution evidence is not exhaustive.

The provider matrix, configuration, README, and test helpers show design attention to provider status, missing or unauthenticated tools, timeouts, hanging streams, thrown errors, and cancellation. Material inconsistencies reduce self-consistency: the README announces v0.3.4 while package.json is 0.4.0; the documented default maximum discussion rounds is 3 while the manifest says 2; the README example uses mysti.claudePath while the manifest defines mysti.claudeCodePath; and provider counts alternate between 10 and 11. Installation prerequisites and CLI commands are extensive, but operation depends on external CLIs, accounts, authentication, daemons, or local services. Test doubles can simulate failures, yet no production error copy or concrete user recovery flow is provided, so failure messaging receives only thin credit.

Documentation thoroughly addresses single-agent work, two-agent brainstorming, security review, architecture decisions, local privacy, developer personas, and five collaboration strategies. Permission modes, the two-agent constraint, round limits, compaction, and provider prerequisites establish useful boundaries, but unsupported tasks and model-quality limits are not clearly stated. Mentions, commands, menus, and an explicit brainstorm toggle make routing reasonably precise; however, activation on startup and numerous workspace-file patterns are broad. VS Code compatibility, cross-platform shortcuts, executable paths, local endpoints, timeouts, and model controls support environment fit, though several setup paths are npm/Unix-oriented and compatibility with every external tool is not demonstrated by the supplied source.

The README has strong navigation, feature sections, configuration tables, and a dedicated-document index, while Marketplace installation, CLI setup, authentication, and quick-start instructions are detailed. Apache-2.0 metadata matches the complete license file, justifying full license credit. Naming and release stability lose substantial credit because of claudePath versus claudeCodePath, v0.3.4 versus 0.4.0, conflicting round defaults, and inconsistent provider counts. Examples are plentiful, but there is no actual FAQ or systematic troubleshooting section. External CLI, subscription, and two-CLI requirements are disclosed, while known safety limits and a compatibility matrix are sparse. Author, company, issue tracker, and contribution routes identify maintainers, but no support lifecycle, release owner, or security-reporting route is shown. The “what’s new” section is not a complete changelog aligned to the current package version.

The source describes a usable chat surface with Markdown/Mermaid rendering, task progress, plan selection, context management, agent routing, and synthesized answers; test factories also expose structured brainstorm and mention streams. This supports ordinary output usability and a plausible marginal benefit from contrasting agents. However, the supplied evidence is mostly product documentation and test scaffolding, with neither core implementation nor representative textual outputs proving that synthesis is actually better than one agent. Auto-convergence, round caps, local models, compaction, and token settings can control cost, but two-agent workflows inherently increase latency and model use, while “no additional cost” and “zero latency” claims omit API, subscription, and hardware qualifications. Cost-benefit is therefore adequate rather than thorough.

The manifest and test helpers make some concrete claims traceable, including provider enumeration, permission defaults, strategies, rounds, timeouts, and simulated failure states. Core claims about safety classification, audit logging, telemetry behavior, convergence, and superior results lack corresponding implementation evidence in the supplied files, limiting traceability. README and package.json corroborate several features but conflict over versions, setting names, round defaults, and provider counts, weakening cross-source support. Marketing conclusions such as “superior solutions,” “complete privacy,” “zero latency,” and never losing work are not clearly separated from verified facts or qualifications, so fact/inference separation is thin.