rdmsm4x-changelog-20260903-1526-spoke-preflight-gate-and-publish-job

[2026-09-03 15:26:16 EDT · rdmsm4x · claude@rdmsm4x session 158c9525 (notes-republish-pending-changelog)]

Four fleet-maintenance defects reported by codex sessions on jdmbair13m5 and rdmbair13m5 at 14:52-14:55 EDT are fixed at the source, distributed to all six Macs, and verified by re-running the exact reported commands on the spokes; the dev.dataroo.net publish job is now bounded and timed per stage, and the agent board it feeds regenerates again. Ticket: TASK-20260903-30 (resolved).

Scope

Hosts: all six (rdmsm4x, rdmbair13m5, rdmbair15m5, jdmbair13m5, rdmpw3265m, rdmpw3275m).
Repos: fleet/maintenance submodule (e74d218), lib/agentkit (5eb5bfa, 9417907), both pushed to GitHub backup and git.ecs0.net fleet.
Files outside git: ~/.agent-coordination/agent_msg.zsh (six hosts), ~/bin/ticket (five spokes), ~/dev/lib/agentkit/lib/ plain copies (four spokes), ~/scripts/agent_checkpoint.zsh (six hosts) plus the six untracked snapshots under ~/dev/fleet/tooling/host-scripts/, ~/.agent-coordination/AGENT_COORDINATION.md, ~/.claude/skills/fleet-ticketing/SKILL.md, ~/dev/scripts/publish_site.zsh (untracked in the macos-scripts repo).

What changed and why

Bus inbox 59 s -> <1 s. cmd_inbox ran one awk per front-matter field per message (about seven processes per message) over 3,396 messages. v1.4.0 reads all front matter in one awk pass and loads the seen-list once. Output proven byte-identical on the live mailbox (2,850 rows, diff empty; --all 3,395, --unread 1,451). Installed with checksum guards; the previous copy is kept beside it as .bak-v1.3.0-20260903. Spoke timings after install: 0-2 s.
Ticket CLI on pointer-only spokes. The retirement removed ~/dev/issues from spokes, so the mandatory preflight returned BROKEN on every spoke and the fleet-ticketing skill's "browsing stays local" was no longer true. ~/bin/ticket on each spoke is now a 30-line proxy that runs every subcommand on rdmsm4x with the spoke's identity, using the same env contract as the CLI's own proxy_to_authority. Verified per spoke: claims, show, and the conformance probe. rdmbair15m5 had a dangling symlink at that path (moved to ~/bin/ticket.dangling-symlink-20260903, not deleted).
agentkit preflight v1.1.1 accepts a ticket proxy found on PATH before declaring the CLI missing. v1.1 (5eb5bfa) was wrong: a perl substitution with | as delimiter turned the escaped || into an empty alternation and inserted the block at file top; all three suites still passed because both paths resolve on rdmsm4x. Caught only by re-running the codex repro on jdmbair13m5; repaired in 9417907 with exact-string edits. lib/ copied without .git to the four spokes that lacked it; the rdmbair13m5 clone fast-forwarded via its origin.
agent_checkpoint.zsh used Bash-only ${VAR@Q} under a zsh shebang (exit 1 "bad substitution" at the JSON heredoc; and even under Bash it would have emitted single-quoted, non-JSON strings). Replaced with a zsh json_str escaper and (qq) quoting for the ssh branch. Validated with hostile input through python's json parser; the rule-30 fallback on jdmbair13m5 now writes valid JSON with the right host and agent.
Policy text. AGENT_COORDINATION.md section 7 gained a pointer-only-spoke paragraph (backup AGENT_COORDINATION.md.bak-pre-spoke-preflight-20260903-151956); the ticketing skill notes the proxy.
publish_site.zsh v1.1. Every stage now runs under coreutils timeout and its rc/seconds are appended to ~/dev/data/publish.log as a stages line; gen_agent_board.py was below exit and had not run since 2026-08-29 (the board Rich reads was five days stale). First v1.1 run failed rc 127 at 15:21 because zsh does not word-split ${TMO:+$TMO 900}; fixed to an array wrapper. Verified run: rc 0 in 69 s (merge 0 s, projects 2 s, site 6 s, ecs0 61 s, board 0 s); ~/dataroo.net/wiki/agents.html regenerated 15:24. Backup: publish_site.zsh.bak-20260903-152118.

Host	Preflight	Time
jdmbair13m5	rc 1 (ACT: aging HIGH messages)	2 s
rdmbair13m5	rc 1	1 s
rdmbair15m5	rc 1	2 s
rdmsm4x	rc 1	~3 s

rc 1 is the intended signal (unacked HIGH messages aging), not BROKEN. The fleet-wide non-acking of HIGH messages (685 for claude@rdmsm4x, oldest 11.8 days) is a separate, pre-existing finding (nightly bus-unread-high).

Undo

Bus: cp ~/.agent-coordination/agent_msg.zsh.bak-v1.3.0-20260903 ~/.agent-coordination/agent_msg.zsh on any host; git -C ~/dev/fleet/maintenance revert e74d218.
agentkit: git -C ~/dev/lib/agentkit revert 9417907 5eb5bfa; re-copy lib/ to spokes.
Proxy: remove ~/bin/ticket on a spoke (on rdmbair15m5 restore the moved symlink only if ~/dev/issues returns).
Checkpoint: ~/dev/fleet/tooling/host-scripts/*/agent_checkpoint.zsh were updated in place; the original is in this session's scratchpad and in the spoke copies' history (sha 17be5e91).
publish_site.zsh: cp ~/dev/scripts/publish_site.zsh.bak-20260903-152118 ~/dev/scripts/publish_site.zsh.

Left open

Root cause of the 2-6 h gaps in devsite delivery is unproven; the new stages log line attributes the next one. Not sleep (pmset shows none); not stormguard (it does not stop LaunchAgents).
Fleet-wide HIGH-message acking discipline (nightly bus-unread-high, 612 unread).
~/scripts has no distribution mechanism; the six fleet/tooling/host-scripts snapshots are untracked.

Secrets