NativeAgent 0.6.3

A turn never fails. The agent's reply ends when the agent finishes it, not
when a limit, a guard or a restart says so.

One way out

The tool loop has one exit: the agent answering with nothing left to do.
Stop, and a question that waits on you, are the only other ways out, and a
wait resumes.
Whatever happens mid-reply is the next step's input, in one line the agent
reads: the provider's length limit was reached, the stream was cut after so
many seconds with so much kept, a tool call was malformed, nothing came
back, the working context was trimmed. What was already written stays with
you; the reply continues in place.
Round caps, the wall clock, retry caps and the guard that read the agent's
output and called it repetition are gone. "I started repeating myself, so I
stopped here" and "I stopped before I could finish" no longer exist.
A transient provider failure retries with backoff until the model answers
or you stop; "Reconnecting to the model, try N…" shows the attempts.
If the agent keeps doing the same thing without getting further, it asks
you in the chat — keep going, or change course — instead of stopping.
Taking over the Mac mid-reply pauses the reply; handing control back
resumes it exactly there.

A reply in progress is checkpointed step by step. If the app quits or
crashes mid-reply, the next launch resumes the same reply where it was,
with the recorded results of the steps already done, and finishes it. No
"I lost my answer, say it again."
An answer that finished while its Telegram, Slack or iPhone delivery was
interrupted is delivered on the next launch without another model call.