15. Things that could be added next

15.1. Small, groundwork done

• publish=<addr>:<hostport>:<guestport>: passt's -t takes an address; ports_to_passt in proxymous_passt/src/passt_spawn.c builds it.

• An instantaneous cpu rate: dump pid reports the average over the life of the process; a rate needs two samples and the half-second beat of the switch could hold the first.

15.2. From the windows test on a work machine

The syslog of a failed add win was readable only with the source open and the guest at hand. Each entry below removes one trip.

• why <guest>: the startup state qga_where prints, the ping status of doorways, the exit text of the agent launcher and the qemu stderr tail the death KERR carries, answered to the client as one message instead of scattered over a syslog the user has to grep on the server.

• A windows launcher that returns at once: QGA_WIN_SCRIPT_LAUNCH uses start /b, and cmd.exe came back 33s later on the test machine. The switch now waits for windows instead of failing, but the cause (Defender scanning the exes from the cd, or inherited pipe handles) is unmeasured. powershell Start-Process or wmic process call create detaches without inheriting the qemu-ga pipes; C:\cloonix-agent.log timestamps say which of the two it was.

• Second family of demo helpers: twelve demos keep an ssh() with a log= argument and a longer error text, thirteen a put() that traces, twenty-seven a place() built on cli(). Converging each pair on one form moves about 700 more lines into demos/lib/demolib.py; the error texts change, so it goes with a run of debug/demo_test.py.

15.3. From the toy runs of October 2026

• The green of a capturable interface: the topology shows eths as soon as sniff=on is accepted, before its snf is up; a webshark sync waits for the snf (timer_sync_wait of ovs_snf.c). The endp_type could turn eths at the ok of the snf instead.

• Sniff off with a webshark open: refused while the webshark runs (handle_snf of web_ctl_post.c); the webshark of that interface could be closed by the same click instead.

15.4. From the testitall run of October 18-19 2026

• The interface of a new bridge: once in a night of tst_max.sh, ovs_cmd_add_lan_endp found no interface _b_<lan> after the add-br that msg_lan_add_name sends without waiting for its answer, and syslog showed no failed add-br. bridge_intf_up of ovs_cmd.c now waits 1 s for it and tells which case it was: a late interface, or a bridge ovs no longer holds because its lan was deleted meanwhile.

• The loop of suid_power waiting for its children: one turn of 3.3 s with 68 ms of cpu while a container with 32 eth is made, its slowest callback the timers: the creation runs in a timer of suid_power, and the 3000 voluntary switches are the 1 ms naps of execute_cmd_out (utils.c) waiting for its commands, the ip -batch that makes and moves 32 veth pairs among them. The creation run in a child of suid_power, as the dtach consoles are, would keep the loop free.

• A turn of proxymous holding 14 s: 1.4 s of cpu in one turn, the host at 97 % softirq during the demos. The late-loop warning of channel.c now names the slowest callback of that turn and its time, so the next one says which channel it was.

• One event, two lines from two processes: the death of a container (crun_beat of suid_power, crun_sigdiag_resp of the switch), a refused byte pipe (doorways and the agent of the guest).

15.5. From the testitall run of October 23-24 2026

• The tcp of the nat: testbigflows.py finds 242 Mbit/s for four tcp streams from a kvm to the host through the nat, against 1377 Mbit/s of udp on the same path, and the flow control of passt_shuttle.c (flow_ctrl_update) stopped the tap read 52 times in that phase. Its backlog thresholds, measured against a tcp flow, would say whether they are the limit.

• A move of an eth asked by two clients at once: the client moves a cabled eth itself (cable_first of cloonix: it reads dump commands, unplugs, then plugs), so two clients moving the same eth at the same time act on a lan that is no longer the one of the eth, and the server refuses the second plug (eth0 is already cabled to lan ..., once in the night of tst_max.sh). A move done by the server in one request would close that window.

• The drop warnings of a flood: one line every 10 s per process, so a c2c saturated for an hour writes some 360 lines of the tap is not taking frames (rxtx.c of ovs_c2c, intf_flags.c): 263 and 262 in this run. A gap growing from 10 s to 60 s while the drops go on would keep the count and the first warning, with fewer lines.

15.6. From the asterisk toy, November 2026

• Asterisk 24: the LTS of October 2026 replaces 22 at its first point release, a version and a sha256 in get_asterisk.sh.

• A call heard in the Demo panel: the panel plays the narration clips; a recording of a call (MixMonitor, 8 kHz wav) played the same way would let the user hear the a2b of voice_quality.py clean, at 5% loss, then through the jitter buffer.

• A ladder of the SIP dialogs in webshark: the dissectors give the Call-ID, the CSeq and the tags of each message, what sngrep draws as a ladder in a terminal of the pbx.

• The reordering of an a2b: the jitter of an a2b keeps the order of the packets; a setting that lets a late packet be passed would show the reordering an RTP receiver repairs.