Megastructure Arc 2.1.6

• API agents added from a connection see pasted images when their model can. Add agent never certified these agents, so Arc could not tell whether their model saw images and told them the pictures were not shown. Now, when such an agent starts, Arc asks its model once with a small image and keeps the answer, so images on messages and from the File button reach a model that sees them; a model that cannot see keeps the text references and the note. Until that check has run, a provider catalog that lists the model's input types, such as OpenRouter's, decides for a newly added agent. A check that fails on the endpoint's side, such as a server error, a dropped connection or the Response deadline, is asked again at a start at least ten minutes later; a rejected key, an empty balance or a rate limit asks again at the next start. Agent settings → Limits gains Images for API and self-hosted server agents: Arc decides, Shown or Not shown, with what Arc decided and why, or why there is no answer yet. A change applies from the agent's next turn.
• Model files see images with their vision projector. When a model file's folder holds one vision projector (an mmproj .gguf, which vision models ship beside their weights), Arc loads it with the model and counts it when it sizes the context window to fit memory. When the agent starts, it checks once that its model sees an image. A model that sees gets pasted and attached images and, in a Full Access room, the computer tool; where Arc can drive the desktop, the model first takes the pointing test once, so its clicks use the coordinates it really answers in. A timeout in either check is not a result, and the next start tries again. Through the HTTP API, a model-file setup can name a specific projector or turn it off (projector: a path, auto or off); Agent settings has no control for it. A model file added before 2.1.6 has no projector on record: add it again, then add its agents from the new entry.
• Images from MCP services land in the room. When a tool from an MCP service you granted an agent returns an image, for example from an image-generation server that draws on this computer or calls a hosted image API, Arc saves it in the room as an image artifact, up to four per call, with alt text built from the service, the tool and the call's prompt. The agent gets a reference it can attach to a message instead of a cut-off block of image data. MCP services still need a Full Access room. For an image that is unreadable or larger than 8 MB, the result says why it was not saved.
• Reply delivery is a setting for Chat Completions API agents. In Agent settings → Limits, an API agent that uses the Chat Completions protocol, including a custom OpenAI-compatible endpoint, can choose Arc default: Streaming, Streaming or Buffered replies; a change applies from the agent's next turn and also asks a provider that declined streaming for a stream again. A custom endpoint saved with buffered replies keeps them until you choose Streaming or Arc default here; upgrading does not change it. API agents on Anthropic Messages or Responses, and self-hosted server agents, always stream, so their Reply delivery is read-only. While a running agent's provider is declining streaming, the setting shows Buffered replies (provider declined streaming). When an agent set to Buffered replies runs past its Response deadline, the error also suggests switching Reply delivery to Streaming.
• A built-in UI review recipe. Agents can fetch arc_get_skill with skill_id arc-ui-review, which arc_browser_open now names, for the steps of a screen review: screenshot the page, post the critique with the image, mock up the change as an HTML or SVG artifact, mark up a screenshot, and post before and after side by side. It says what a mock-up page may load: inline styles and earlier screenshots, with no scripts or outside files. It comes in English, Simplified Chinese, Traditional Chinese and Japanese, and you can edit or reset it with the other built-in instructions in Settings → Arc instructions.

• Timeout and dropped-connection errors name the recovery card's Retry. When a reply runs past its Response deadline or the hard limit for a streamed reply, or the provider or model server closes the connection, the error no longer says to send a new message, which would only wait in the agent's queue behind the held step. It says to choose Retry unfinished step now, after raising Response deadline (Agent settings → Advanced) or lowering Output tokens per step (Agent settings → Limits) if needed.
• A refused retry during a provider cooldown gives the time in your time zone. Retry unfinished step now, sent while a provider cooldown runs, was refused with the earliest retry time in UTC while the room card showed local time. The refusal now gives local time with its UTC offset, for example 2026-10-15T00:30:00-07:00.
• An agent whose start cleared a stale failed test takes limits changes again. When an agent's start cleared a failed connection test left by an older Arc build, the agent took no later limits change until it was restarted. It now takes them from its next turn.
• Starting a Full Access agent's turn no longer holds up the hub. Before such an agent starts work, Arc lists the programming tools on the computer (Python, Node.js, Java, .NET and others) for its instructions. It checked them one at a time while every other request to the hub waited, which on a busy Windows machine could pass 15 seconds and time the start out. Arc now checks them side by side and reuses the list for ten minutes, or until PATH changes; a tool that did not answer in time is checked again at the next start.
• A quick first edit to a traffic limit is kept. The Concurrent requests, Server slots and Requests per minute fields in Settings → Connections & models could reset themselves just after appearing, undoing a value typed in that moment: clearing Server slots and typing 3 could save 13. They now change only when the saved limits change.
• Arc's command line writes UTF-8 when its output goes to a file or pipe. On Windows, arc --help and arc-daemon.exe --help failed with UnicodeEncodeError when redirected, as did any command printing text outside the system code page, such as a Chinese or Japanese path, and error messages showed such text as backslash escapes. Redirected output and errors are now UTF-8; a console window is unchanged.

Windows 10/11 (64-bit), or macOS 11+ on Apple Silicon. Bundled Local AI requires macOS 26+. Computer use requires macOS 14+.