Megastructure Arc 2.1.5

• API agents stream replies by default. New API agents stream, and upgrading moves existing agents and saved setups on built-in providers to streaming. The Response deadline, 300 seconds by default, now limits only the wait for new output, so a streamed reply that keeps producing output can run past five minutes, and every reply keeps its hard cap. Custom OpenAI-compatible endpoints saved with buffered replies keep them. A provider whose rejection names the streaming setting still gets a buffered request, and an endpoint that ignores the streaming request and answers in one piece with a JSON reply keeps that reply without a second request. Agent settings shows Reply delivery for API and self-hosted server agents: Streaming, Buffered replies, or Buffered replies (provider declined streaming) while a running agent's provider is declining it.
• API and self-hosted server agents recover from provider failures. When a provider is rate-limited, overloaded or briefly failing, Arc retries only the unfinished model step, after about 20 seconds and then about 3 minutes; tool work the agent already finished is never run again. A stream the provider ends with an unexplained error gets one retry, after about 3 minutes. Waits honor the provider's Retry-After, and when waiting would take recovery past ten minutes in total, Arc pauses and shows the earliest retry time. Arc never automatically re-sends a request the provider refused (a rejected key or exhausted quota), or a request to an API provider whose outcome is unknown (the connection dropped after it was sent); in those cases, and when retries run out, the agent holds the unfinished step until you act. A self-hosted server or model file that drops the connection still gets one immediate re-send once it answers again, as before. This replaces the hidden retries after 20 and 60 seconds.
• While Arc recovers, a card in the room shows the provider's error, how many replies are waiting, and copyable recovery details with the provider's request or generation ID when it sends one, plus the countdown and retries used while a retry is scheduled. Retry unfinished step now sends the step at once unless a provider cooldown is running, Pause retries holds a scheduled retry, and stopping the agent cancels recovery. Replies sent to the agent meanwhile wait and run in order once the step completes. The unfinished step, finished tool work, token counts and retry counts are saved, so recovery continues when the agent resumes after Arc restarts; a request cut off by a restart stays paused and is marked as having an unknown outcome. Arc will not resume a step after the agent's model, key, folder or room access has changed. Provider recovery is on by default; turn it off per agent in Agent settings → Advanced, and a failed step still waits for you to retry it or stop the agent. Agents running a model file show the same card in local-model wording.
• Failures link to the setting involved. When an API agent's reply times out, or a provider rejects a request and names its temperature, thinking level or output-token value, the failure offers Open relevant settings. It opens that agent's Agent settings at Response deadline (expanding Advanced), Output tokens per step, Temperature or Thinking, without saving anything or retrying the request. A reply ended by the per-reply hard cap points to Output tokens per step rather than the Response deadline.
• Agent settings say where each value comes from. Effective values are labeled Arc default, Agent setting, or Model default or limit. New API and self-hosted server agents, and a Temperature field you clear, use the model's or server's own sampling default instead of Arc's 0.2; agents created earlier keep their saved temperature until you clear it. Temperature is disabled, with a note, for models that ignore it, such as models on Anthropic's API and OpenAI's GPT-5 models. A context window a new agent takes from its provider's catalog is labeled as a model value; one you type stays an agent setting.
• Agent session journals are private to your account. On macOS, each agent's session journal, which can hold full tool results and saved recovery state, is made readable only by your user account whenever Arc opens it.

• API agents no longer ask for more output than their model allows. When a provider's catalog lists a model's output limit, an API agent's per-step output budget is held to it, and Agent settings shows Held to N by the model's output limit. A larger value you saved is kept and shown.
• Provider errors that mention “upstream” no longer switch an agent to buffered replies. Arc re-sends a request without streaming only when the provider's rejection names the stream or stream_options setting, so an unrelated rejection, such as one that mentions an upstream or downstream service, is reported once instead of being sent again and leaving the agent on buffered replies.

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