chmonitor scheduled desk jobs

chmonitor's unattended maintenance is the herdr-desk Herdr plugin, config only in .herdr-desk.json plus repo-owned playbooks in docs/herdr-desk/.

Job	Cron	Owns
desk:github-issues	0,30 * * * *	triage, research, dispatch children
local:babysit	10,40 * * * *	red required CI, review, auto-merge, worktrees
local:prod	20,50 * * * *	live deploy verification, agent probe, usage, revert
local:improve	17 2 * * *	desk health, doc/skill drift, dead code, slowdowns

Why: one job doing all four duties made every duty compete for the same five child slots, so a busy issue queue silently starved PR maintenance. Separate jobs give each duty its own budget and its own manager.

How to apply:

• Stagger the minutes so two jobs never contend for one slot, and give each job its own agentName — the agent name is the session identity, so a shared name means two prompts racing for one pane.
• name in the desk config must equal the repo folder name. A stale "name": "chmonitor" in the herdr-desk repo registered a second chmonitor job on a different cron; the loser recorded a prompt failure.
• Required CI is unit-tests + dashboard. e2e-test, e2e-test-tsr, component-test, codecov/patch, promptfoo and Claude Issues are informational. Release-please PRs are human-only, never auto-merge.
• Production deploys itself on push to main; local:prod is the second pair of eyes and may only verify or restore service, never ship.
• Proof a branch landed before deleting its worktree: git diff <branch> origin/main -- <touched files>, not git branch --merged — squash merges change the sha, so a fully merged branch still reads "ahead N".

Lessons: tech-silent-scheduler-failure, feedback-issue-research-pr.