The mesh test expects the single line of a pair, dialed by the smaller id. Tests that send at a fixed rate pin adaptive to false, so the admission limit follows that rate.
MoShiro__MinAgentVersion makes the backend refuse agents below it at hello (fleet_version_required); fleet.min_server_version or MOSHIRO_MIN_SERVER_VERSION makes the agent drop a backend below it. Either side waits five seconds and shakes hands again, because the other one may be updating.
Heartbeats carry per-level counts, the most frequent log types (a stable hash of the message) and the newest id, no text. The node keeps its last 2048 lines; moshiro.logs.fetch lets the backend pull them. The log page pulls when it is open, and the backend pulls warnings and errors once a minute after a node reports an error.
The control layer keeps doubling after the first minute up to ten minutes, and the QUIC redial doubles after twelve short attempts up to five minutes, instead of retrying every minute or every five seconds forever.
udp_echo and tcp_echo run at the lowest rate that still proves the loss bound and keeps the RTT p99 confidence interval within tolerance, go to the maximum when the line loses packets or its RTT rises, and back off exponentially when it is unreachable. Every echo row carries a weight (the time it stands for, in grid intervals) and the analysis ops accumulate loss and RTT by it, so p50/p95/p99 and loss stay time-averaged when the rate changes. State changes are logged for audit.
The Backend now keeps the time since a link's control layer was last up, because retries move the control state in and out of backoff and its own timestamp keeps resetting. The overview exposes it as control_down_since and the card shows it next to the lost-contact notice.
A line shows when either end is selected; the selection is kept in ?agents=. Cards use container queries to tighten the columns and drop the hint text when narrow.
Rows are laid out as protocol, cells and value on one line per metric. Each direction gets its own axis with ticks on the hour; cells without data and the bucket still being filled are grey.
POST /api/admin/links/reset clears every manual link setting and turns automatic linking on. In the default rules a listening node in CN ranks below a listening node elsewhere, so the CN side dials; the links are pushed again when a node's country is first resolved or changes.
tcp.retrans.up counts retransmitted segments sent by the dialing side, tcp.retrans.down those sent by the listening side. The cards show RTT as a round trip (⇄) with forward (→) and return (←) rows for QUIC loss and TCP retransmission.
A pair of nodes now has a single line, dialed by the non-listening side or by the smaller id when both listen, so RTT is measured once. The board draws it undirected and colours it by RTT and loss; a manual setting applies to the pair.
buildah unshare fails on the CI host with memfd_create: Invalid argument, so extract-node-binaries.sh failed right after a successful native build. Exporting an OCI archive needs no mount (the toolchain archive already works that way); extract-oci-files.py reads the files from the layers, later layers win and whiteouts delete.
Identical restore images accumulate one tag per commit, and buildah images <name> prints a row per tag, so the single-id check in build-linux-agent.sh failed for every commit after the first. moshiro_image_id returns one id and a test covers the multi-tag case.
Every Linux step now runs under ci/guard-step.py. The first step that fails, is cancelled or times out marks the pipeline; running steps get SIGTERM (so traps release flock and Buildah state) then SIGKILL, steps that have not started exit at once, and any process a step leaves behind is killed. A pipeline deadline is the last resort. build-linux-agent.sh now prints the lookup error, store configuration and the restore images present when the prepared restore image cannot be found.
The restore step failed after restoring every dependency because the final cache cleanup (buildah unshare reporting memfd_create) was not guarded. Cleanup failures now only warn, and a test requires every cleanup call to be guarded.
justme