Orbital Project Agent

The evidence shows workspace scoping, request_access for paths outside the workspace, selectable autonomy, structured approvals, credential-request modals, OS-keychain storage, key masking, and CI secret cleanup. These support ordinary-use least privilege, confirmation, and sensitive-data handling. Tool categories and external-effect surfaces are enumerated, scheduled runs are visible, and session logs are append-only. However, authenticated browser sessions, a shared browser profile, shell and file writes, triggers, and third-party connectors create a broad authority surface without operation-by-operation confirmation rules or complete retention policies. The local-only claim conflicts with CI and test comments documenting a production telemetry endpoint and local spooling; default payloads, destination, retention, and user controls are not explained, so data-flow transparency is thin. Dependency care is demonstrated by the justified MCP upper bound, but most Python packages have broad lower bounds and no supplied lockfile, vulnerability scan, or provenance verification. Pausing, archives, and append-only logs aid contextual recovery, but there is no general rollback for file changes, commands, or remote actions. GPL notices and attribution to Orbital Contributors provide limited provenance, while the publisher and third-party asset attribution remain unclear.

The README, dependency declaration, CI, and test fixtures jointly support desktop packaging, multiple providers, keychain use, telemetry controls, and cross-platform testing. Every queued task is required to end Completed or Blocked, with re-prompting and forced blocking described; test markers and installer smoke tests provide concrete skip, timeout, and failure messages, earning full credit for failure reporting. Dependency availability is reasonably addressed through Windows/macOS builds, optional extras, the MCP compatibility cap, and packaged installers, but external agent CLIs, model providers, accounts, browser components, and a relay absent from the checkout remain availability conditions. Self-consistency is reduced by Orbital versus agent-os naming, README v0.4.2 versus package version 0.12.3, and the tension between local-only positioning and documented production telemetry.

The material addresses persistent projects, queued work, one-off tasks, cross-agent handoffs, a mainland-China TokenDance path, and global provider choices, but gives little guidance for regulated enterprises, highly sensitive projects, or offline use. Capability boundaries are unusually explicit through the Is/Is Not table, beta labels, scheduler-driven autonomy limitation, platform statements, and transport distinctions. Schedules and file watchers are distinguished, and queue ordering, pausing, and blocking semantics are described, but trigger matching, debouncing, concurrency conflicts, and approval behavior for risky triggered work are not fully specified. Windows and macOS have build paths and an optional macOS integration degrades gracefully; Linux desktop support, Intel macOS coverage, and supported-version ranges for external workers are not established.

The workspace and global directory trees clearly assign responsibilities to sessions, memory, archives, queues, tool results, and output artifacts, earning full information-architecture credit. Quick Start covers installers, providers, account connection, and project creation, but omits source installation, upgrades, uninstallation, troubleshooting, and installer-integrity verification. Demonstrations, screenshots, examples, and feature deep dives are extensive, although there is no systematic FAQ. Beta status, incomplete full autonomy, unsigned builds, resource-dependent tests, and the external relay are disclosed, but limitations remain scattered. The full GPL-3.0 text, README badge, and SPDX headers align, fully supporting the license criterion. Releases and version notes are referenced, but package version 0.12.3, README feature version v0.4.2, and the agent-os/Orbital naming split weaken naming and version clarity; no complete changelog is supplied. Maintenance responsibility is limited to the generic Orbital Contributors label, without a named maintainer, support route, or security-update owner.

State, decisions, lessons, queues, logs, and deliverables are stored as ordinary files with indexing, archival, explicit task outcomes, and cross-agent handoff, making the outputs highly reusable. Project-level context and replaceable workers offer plausible marginal value over isolated agent sessions, but seamless continuation and competitor comparisons are mainly first-party README claims rather than behavior corroborated by the supplied production code. The source mentions sub-five-minute setup, bring-your-own API keys, budget controls, prefix caching, and paid live turns, yet does not quantify model expense, storage growth, background resource use, browser-session risk, or maintenance overhead. Cost-benefit handling is therefore adequate rather than thorough.

Many claims point to concrete filenames, directory layouts, environment variables, versions, test markers, CI jobs, and specification identifiers. CI and fixtures cross-corroborate telemetry suppression, keychain avoidance, platform constraints, and packaged-daemon smoke tests. Key agent behavior, approval governance, memory consolidation, and competitor comparisons, however, are chiefly documented in the README without corresponding production implementation files in the supplied evidence. Comments acknowledging stale and pending screenshots improve candor, but promotional conclusions, competitor assertions, approximately 95% cache-hit claims, and exact continuation claims are not clearly separated into demonstrated facts, inference, and planned validation. Fact-inference separation is therefore thin.