[igorseixashr]

Related Issue

No existing issue found for this (searched issues/PRs for Device#, IOS prompt truncation).

Description

For IOS/IOS-XE the persistent SSH session waits for a bare # or > until show version has been parsed and self.prompt becomes <hostname>#/<hostname>>.

On IOS 15.x (seen on CISCO1905 with 15.2(4)M3 and 15.7(3)M9) the license table in show version contains a line starting with Device#:

Device#   PID                   SN
-------------------------------------------------
*0        CISCO1905BR-SEC/K9    XXXXXXXXXXX

readuntil(('>', '#')) stops at that line, so the captured output starts at \t PID\t\t\tSN. Hostname and uptime are never parsed (Cannot parse uptime), the prompt stays bare, and every following command returns the tail of the previous command's output.

Symptom: the device, interfaces and bgp tables stay empty for these nodes while sqPoller reports OK for every service.

Type of change

• Bug fix (non-breaking change which fixes an issue)

New Behavior

Right after the interactive session is opened in IosXENode._ssh_connect(), the hostname is taken from the login prompt (last line, ending in # or >, without spaces) and self.prompt is anchored on it immediately, before privilege escalation and terminal length 0.

Contrast to Current Behavior

Today the prompt is only anchored after show version is parsed, which is exactly the command that gets truncated on these platforms. With the change the prompt is anchored before any command runs.

Discussion: Benefits and Drawbacks

• Fixes silent data loss (empty tables with OK status) on IOS 15.x routers whose show version contains Device#.
• Backwards-compatible: if the login prompt can't be read, self.prompt keeps the default ('>', '#') and the current flow is unchanged. The later anchoring from show version still runs.
• Only IosXENode / IOSNode are touched; other NOS use different connect paths.

Testing

Tested with sq-poller --run-once=process against real devices, on 0.25.0 (same hunk as develop at 75198b5):

Device	OS	Before	After
CISCO1905BR-SEC/K9	IOS 15.2(4)M3	device 0 rows, interfaces 0, bgp 0	device 1, interfaces 13, bgp 3
CISCO1905BR-SEC/K9	IOS 15.7(3)M9	device 0 rows	device OK
WS-C2960X	IOS 15.2(2)E6 / 15.2(7)E13	OK	OK (no regression)
C9300-24S	IOS-XE 17.15.3	OK	OK (no regression)

No unit test added: this path needs a live interactive SSH session and I couldn't find existing tests covering _ssh_connect(). Happy to add one if you point me to a preferred pattern.

Changes to the Documentation

None.

Proposed Release Note Entry

IOS/IOS-XE: anchor the CLI prompt on the device hostname at connect time, fixing truncated show version and empty device/interfaces/bgp data on IOS 15.x routers whose output contains Device#.

[igorseixashr]

Pushed a follow-up commit: some devices have ASCII-art login banners containing many # characters, so the first readuntil(('>', '#')) returned inside the banner and the real hostname# prompt was never seen. The connect path now drains the banner until the device goes quiet (2s idle, bounded) and then anchors the prompt on the last line.

Re-tested with --run-once=process: two CISCO1905 (15.2(4)M3) with such a banner went from empty device/interfaces to 1 device row and 15 interfaces each; no regression on 1905 without banner art, 2960X (15.2(2)E6, 15.2(7)E13) or C9300 (17.15.3).

[igorseixashr]

Added a third commit, related but broader than the IOS prompt fix: reconnect-and-retry in the base _ssh_gather when the persistent SSH connection was dropped by the device.

Seen on NX-OS 7.3(13)N1(1): the device closes an idle SSH connection ~40s after the last channel closes, and the first command of the next poll fails with SSH connection closed (the node shows as dead for that cycle). With the change, 4 such switches polled every 5 min logged 4 reconnects and device succeeded in 19/20 cycles (previously failing from the 2nd cycle on).

Happy to split this into a separate PR if you prefer to review it independently.