Aura IDE

The tests show strong least-privilege and authority boundaries: restricted root turns refuse writable delegation instead of silently downgrading it, Agent permissions are frozen at submission, delegation tools disappear when unavailable, and traversal or link-like definition paths are rejected. Applying or discarding retained work requires approval and the primary workspace is revalidated afterward; isolated worktrees, checkpoints, journaled states, and cleanup-pending recovery provide thorough rollback evidence. The README identifies major external channels—hosted models, local servers, MCP, terminal, Git, and the mobile relay—but does not fully state what each receives, retention behavior, or processing responsibility, so data-flow transparency and external-effects handling are not complete. API keys are only described as settings entries, with no supplied evidence about encryption, storage, log redaction, or rotation, limiting sensitive-data handling to a thin score. Dependencies use broad minimum versions without a lockfile, supply-chain scanning, or vulnerability-response evidence, so dependency security is weak. The MIT notice and project naming provide attribution, but the copyright holder is only “Aura” and publisher identity remains unknown, preventing full source-attribution credit; unknown identity is not treated as suspicious.

README claims about approvals, isolated Agent work, visible failures, and retained changes align with the authority, boundary-hardening, and transaction tests. Tests cover persistence errors, locking, dirty worktrees, missing refs, Git failures, approval races, and cleanup failures while asserting structured failure classes and truthful state, supporting full self-consistency and failure-message scores. Python requirements, the executable entry point, and dependencies are declared, and Windows and source-install paths are described. Dependency availability is not full because versions are not locked or upper-bounded, some tests can skip when PySide6 is absent, and model services, Git, and local servers remain external availability requirements.

The material addresses Windows desktop users, macOS/Linux source users, hosted and local models, individual Agents and reusable Teams, and read-only, plan-review, and writable workflows, giving strong audience and scenario coverage. Permission enums, root-turn mutation restrictions, context-sensitive tool exposure, exclusion of private child instructions from the model-visible roster, and bounded diff inspection establish clear capability boundaries. Trigger behavior has implementation-level tests, but the supplied evidence does not include the complete system prompt, every tool-routing path, or false-trigger evaluations, so trigger precision is not full. Environment support spans Python 3.10+, multiple providers, and OpenAI-style local servers, but only Windows has the described packaged install; macOS/Linux require source execution, and local coding depends on tool-calling support and adequate context.

The README has clear sections for onboarding, Agents and Teams, models, safety controls, mobile access, and support, earning strong information-architecture credit. Windows setup is concrete and other systems are directed to an installation guide, but that guide's commands and troubleshooting content are not supplied. Aura, Agent, Team, and change-set terminology is generally consistent, although the generic package name “aura” and unusual “2.3.04” version format leave limited evidence of long-term naming stability. Practical prompts and a workflow example are present, but no actual FAQ is included. Hosted-provider costs, local-server requirements, and the desktop-uptime requirement are disclosed, though there is no comprehensive known-issues section. The complete MIT text justifies full license credit. A version and Releases path exist, but no changelog or compatibility policy is included. Issues, Discord, and sponsorship routes are named, yet no maintainer identity, support commitment, or security contact is established; with publisher identity unverified, maintenance responsibility remains thin.

The output is designed for practical review: users can inspect proposed diffs, tool activity, terminal output, checks, and failures. Agent changes remain as inspectable change sets, while binary and oversized content is explicitly represented with omission notices, size, hashes, and truncation state, giving strong output-usability evidence. Reusable Teams, graph-based workflows, model mixing, and isolated retained changes plausibly add value over a basic coding chat, but no comparative study or user outcome data is supplied, so marginal value is not full. The product is MIT-licensed and supports local models while clearly warning that hosted APIs cost extra; however, token usage, latency, resource consumption, and orchestration overhead are not quantified, limiting the cost-benefit score.

Important safety and recovery claims trace to focused assertions covering frozen authority, path boundaries, approval, transaction states, post-approval revalidation, and failure classification. Metadata independently exposes version, dependencies, entry point, and licensing, supporting strong claim traceability. Multiple test files corroborate the README's central control claims, while the license and project metadata add different source types. The deduction concerns promotional statements such as Aura being developed with Aura and the cited example-run outcome, which are supported only by the project's own narrative or referenced screenshots; the material also does not consistently label demonstrations, design goals, and verified behavior as separate categories.