wrote /home/user/Projects/overdeck/.worktrees/fleet-s1/fleet/inventory.yml: 3 buildbox(es) from registry — ['debian1', 'debian2', 'debian3']

PLAY [fleet — path bins] *******************************************************

PLAY [fleet — seat settings template] ******************************************

PLAY [fleet — systemd user units (laptop only this pass — see README)] *********

TASK [Gathering Facts] *********************************************************
ok: [laptop]

TASK [systemd_units : converge resource-control slice unit files] **************
[WARNING]: Deprecation warnings can be disabled by setting `deprecation_warnings=False` in ansible.cfg.
[DEPRECATION WARNING]: INJECT_FACTS_AS_VARS default to `True` is deprecated, top-level facts will not be auto injected after the change. This feature will be removed from ansible-core version 2.24.
Origin: /home/user/Projects/overdeck/.worktrees/fleet-s1/fleet/roles/systemd_units/tasks/main.yml:5:11

3   ansible.builtin.copy:
4     src: "{{ playbook_dir }}/../modules/monitor/systemd/user/{{ item }}"
5     dest: "{{ ansible_env.HOME }}/.config/systemd/user/{{ item }}"
            ^ column 11

Use `ansible_facts["fact_name"]` (no `ansible_` prefix) instead.

--- before: /home/user/.config/systemd/user/agent.slice
+++ after: /home/user/Projects/overdeck/.worktrees/fleet-s1/modules/monitor/systemd/user/agent.slice
@@ -2,8 +2,14 @@
 Description=Interactive AI agent runtimes (codex/claude/cursor-agent)
 
 [Slice]
-CPUQuota=150%
-CPUWeight=1
+# No CPUQuota: a hard quota starved every agent tool shell under contention
+# (nr_throttled=2.88M, ~300h stalled, 2026-08-15). CPUWeight is proportional
+# and cannot starve — desktop (weight 100) still outranks agents, agents
+# split what remains. This was a runtime drop-in override
+# (agent.slice.d/50-no-starvation.conf) before 2026-08-15; folded into the
+# base declaration here so the correct value cannot regress if the drop-in
+# is ever lost.
+CPUWeight=50
 TasksMax=3072
 # No MemoryHigh/MemoryMax/MemorySwapMax: agent work is legitimately heavy, and a fixed
 # number cannot tell a 35G build from a runaway. Memory pressure is handled by

changed: [laptop] => (item=agent.slice)
changed: [laptop] => (item=build.slice)
changed: [laptop] => (item=ci.slice)
changed: [laptop] => (item=guard.slice)
changed: [laptop] => (item=human.slice)
changed: [laptop] => (item=interactive.slice)
changed: [laptop] => (item=unsafe.slice)
changed: [laptop] => (item=app-org.gnome.Terminal.slice)

TASK [systemd_units : reload systemd user daemon if any slice changed (apply mode only)] ***
skipping: [laptop]

PLAY RECAP *********************************************************************
laptop                     : ok=2    changed=1    unreachable=0    failed=0    skipped=1    rescued=0    ignored=0