Skip to article
Qingping Champ
← All field notesPROJECT WIKI

Anera · From task to artifact

A desktop execution workbench with persistent context and delivery checks.

Reviewed 5 sections

What it does

Anera is a desktop workbench for taking a task through research, tool use, a persistent workspace, browser checks, and artifact delivery. A visitor can inspect how an agent got to its result instead of seeing only a final chat message.

The project independently reconstructs publicly observable Arena Agent Mode behavior using public pages, logs, and recordings. It is not an official Arena implementation or a claim of exact feature, visual, latency, or cost parity.

Section sources

Execution and workspace

A TypeScript service coordinates the model, tools, session state, and task workspace; a React/Vite interface presents the work. Session messages and events are stored in file-backed journals. Anera's session persistence should not be described as Napier's SQLite ledger.

Workspace versions use a restore journal with prepared, installed, and settled phases. This makes recovery of an interrupted restore an explicit state transition instead of treating a partly copied directory as a completed version.

IllustrationAnera's task-to-artifact workflow · illustrationRead the source ↗
Loading image…

Anera's task-to-artifact workflow · illustration

Section sources

Resume and context

Explicit resume accepts eligible cancelled, failed, timed-out, and interrupted runs. It reloads stored messages and events, reconciles tool calls that did not execute, collapses redundant tool tails, and stages a new turn with a run.resumed event.

Durable evidence is separate from the context shown to the model. Long consumed tool outputs may be projected as excerpts while their originals remain available through hash-bound read_context. A shorter prompt does not mean deleting the underlying execution record.

Section sources

Sandbox and controls

Local execution requires an operating-system sandbox by default: Seatbelt on macOS or Bubblewrap on Linux. Workspace writes, visibility of the host home directory, and network access are controlled by the execution environment, not solely by model instructions.

Usage is recorded. Optional per-turn model and token hard limits default to zero, meaning disabled; cancellation, timeout, and no-progress controls are separate. A hard budget should not be claimed to be active merely because usage is visible.

Section sources

Access and limits

The repository was public at review time and no software license file was found. Publicly readable code is not by itself a license grant. Running it requires configured model access and a supported local sandbox.

Mobile support, a complete GitHub Connector, and exact Arena parity remain outside the stated completed scope. Documentation and code explain mechanisms; they do not establish production reliability or performance on every task.

Section sources
← Explore another note