[p-r-a-v-i-n]

Host.upload_file passes block_size=255 * 1024 to sftp.put. With the default block_size=-1, asyncssh sizes each write from the max_write_len the server advertises. OpenSSH's sftp-server drops the session on a packet over 256 KiB and advertises a max-write-length of 255 KiB, which leaves room for the packet header. The drukbox gateway in front of it advertises more, so every upload of a file of 256 KiB or more failed: chat thread files, workspace files, and browser state. An explicit block_size is used as given, so no advertised limit can raise it. The gateway issue is czpython/drukbox#85 and its fix is czpython/drukbox#86; this keeps Druks inside what sftp-server takes whatever a gateway advertises.

• Host.upload_file raises SandboxUploadError, beside SandboxDownloadError, when its connect or transfer fails with an asyncssh.Error or an OSError. The error names the remote path and the cause, for example failed to upload /…/files/F1/plan.pdf: SFTPConnectionLost: Connection lost. A refused or timed-out connect is an OSError, not an asyncssh.Error, and the upload connects again because the host that get_running_sandbox returns has its SSH connection closed.
• chat_read_file answers 502 with that message on a SandboxUploadError. Before, the error left the route as a bare 500, and the agent could not tell the person which file failed or why.
• SandboxUploadError is a SandboxError. An agent call uploads its prompt as the run's stdin file, so a failed prompt upload now becomes a HarnessSandboxError in Host._start_run, which retries, as other sandbox failures at the run's start do. Before, the call failed without a retry.
• The fake SFTP client in test_sandbox_host.py records the block_size of each put, and the upload test asserts 255 KiB. A new host test covers a lost SFTP session and a refused connect; the route test covers the 502.

[czpython]

The block-size change is right. Two fixes and a nit are inline.

The PR body has the same error as the comment in host.py. sftp-server drops a packet over 256 KiB, and 255 KiB is the max-write-length it advertises. The body also calls czpython/drukbox#85 the gateway fix. That is the issue. The fix is czpython/drukbox#86.

[czpython]

One more change, inline. Sorry, we missed it in the first review.

[czpython]

Instead of catching these errors on the route, catch them in Host.upload_file. Raise a SandboxUploadError there, beside SandboxDownloadError in sandbox/exceptions.py. This way the route does not know the details of the upload: SandboxUploadError is all it catches, and routes.py drops its asyncssh import. The two error cases then belong in a Host.upload_file test, and the route test needs one case.