Comma Personal Agent

The README describes separation of device identity from cloud work, an opt-in computer operations setting, Needs Review decisions, and no automatic rerun of interrupted side effects. The Admin test excerpts also show confirmation fields and checks for sensitive-field projections. Least privilege scores 1 because the product is described as always on and able to read files, run commands, and operate applications, while the supplied material does not establish default permissions or per-action authorization scope. Data-flow transparency and sensitive-data handling score 2: the docs explain model-provider charges, device/cloud responsibilities, and that the config volume holds an encryption key, but do not provide a complete lifecycle or retention account. Dependency security scores 1: package versions are pinned, but no vulnerability audit or mitigation evidence is supplied. External effects and rollback score 2: review gates, an operations toggle, and backup/restore guidance are documented, but recovery paths for every mutating operation are not shown. Source attribution scores 1: AFK Inc. and security/contribution contacts are named, but the prompt says publisher identity is unverified and the supplied material does not establish a verifiable maintainer identity.

Scores 2: the README's descriptions of task state, process restarts, persistence of accepted messages, and interrupted side effects are mutually consistent, and it points to Lean/TLA+ verification material; test excerpts provide partial support for Admin API contracts. Dependency availability scores 2: self-hosting docs list Docker Compose, public toolchains, and an external model account, but the supplied material does not include a complete dependency inventory or alternatives. Failure messages score 2: docs describe some failure semantics, and tests assert specific errors for missing build configuration and direct API requests; user-facing error coverage is not shown.

Scores 2: the product targets continuous personal tasks across computers, phones, cloud environments, and cooperating agents, but the supplied setup chiefly covers single-node Docker and Mac devices. Capability boundaries score 2: docs say meetings and Agent VMM require runtimes outside the repository and group cloud computers need Cloudflare configuration; they also distinguish drafts from autonomous work. Per-integration permission boundaries are not fully described. Trigger precision scores 2: Loops wake the agent only when a watched target needs attention, and Routines can brief at a chosen time; trigger configuration and false-wakeup controls are not shown. Environment fit scores 2: local self-hosting, development, Electron, and public deployment are documented, though deployment involves multiple services and an external model provider.

Scores 3 for clear repository layout, documentation index, architecture entry point, and install guidance covering requirements, configuration, startup, access, public deployment, upgrades, and backups. Naming stability scores 2: Tasks, Loops, and Needs Review are explained, but the supplied evidence is insufficient to assess consistency across code and docs. Examples and FAQ score 2: commands, configuration variables, and a development sign-in example are provided, but no FAQ is shown. Known limitations score 2: external runtime and cloud configuration requirements are stated, though coverage is limited. License scores 3: README and LICENSE identify AGPL-3.0-only and note that third-party components retain their own licenses. Versioning/changelog scores 1 because no release policy or changelog is included in the supplied material. Maintenance responsibility scores 2: AFK Inc., a security contact, and a CLA are named, but publisher identity and ongoing maintenance are not verified.

Scores 2: the task board, persistent execution, review decisions, scheduled briefings, and device operations describe usable work paths, but the supplied material contains no independent run results establishing task quality. Marginal value scores 2: cross-session task state, device collaboration, and agent orchestration offer a distinct product proposition, while some capabilities depend on external accounts or runtimes. Cost-benefit scores 2: docs disclose model-provider charges and say self-hosting needs no Comma credits or Stripe, but do not quantify resource costs or ongoing operational burden.

Scores 2: the README points to specific repository paths for formal verification, deployment, development, and architecture, and test excerpts show selected API and boundary assertions; the linked verification docs and full implementation are not part of the supplied material. Cross-source corroboration scores 2: README claims about confirmation, secret handling, and API boundaries have partial counterparts in test excerpts, but those excerpts are limited and cannot establish product-wide behavior. Fact/inference separation scores 2: the README explicitly says formal verification does not prove model judgment or task success, appropriately limiting that claim; other performance and product claims remain primarily project assertions.