fix(hass): stop Good Night/Sleep killing white noise

Manual on was stomped by scene force-off before the HR-gated
automation could keep the plug on. Leave ownership to manual +
white_noise_with_bedtime_audiobook; Good Morning still turns it off.

Author: edmundmiller

config/agents/rules/03-version-control.md

@@ -13,6 +13,7 @@ waiver_path: .agents/waivers/AGENT-03.md
 - **Diff policy:** prefer `sem diff` over `git diff` (entity-level changes, better for agent review). Use native `git diff` only when sem lacks needed flags/output.
 - **Selective staging:** use `git hunks list` / `git hunks add <hunk-id>` — non-interactive, deterministic hunk IDs.
 - **Dotfiles exception**: sibling layout (`../dotfiles.branch`) via `.envrc` override
+- **Commit scope:** before committing, compare staged paths with the current request. “Commit/push” authorizes only those paths; if they are already landed and only unrelated paths remain, report that state instead of touching them.
 
 ## Non-interactive defaults (agents)
 

config/agents/rules/15-agent-behavior.md

@@ -18,5 +18,6 @@ Before changing code:
 - For external state changes, a successful command/API response is not verification. Re-read the authoritative state or user-visible artifact that should change before claiming success.
 - Do not clean up unrelated code; mention it separately.
 - Define success criteria for non-trivial tasks, then verify them.
+- A skipped or no-op check (`no files to check`, zero tests collected, missing validator) is not verification. Run a check that exercises the changed artifact, and never report the no-op as passed.
 
 Every changed line should trace directly to the user request.

modules/services/hass/_domains/sleep/default.nix

@@ -335,7 +335,6 @@ in
           "input_boolean.goodnight" = "on";
           "select.master_suite_current_mode" = "sleep";
           "switch.adaptive_lighting_sleep_mode_living_space" = "on";
-          "switch.eve_energy_20ebu4101" = "off"; # Whitenoise waits for final Sleep phase
           "cover.smartwings_window_covering" = "closed";
           "light.essentials_a19_a60" = "off";
           "light.essentials_a19_a60_2" = "off";
@@ -356,7 +355,6 @@ in
           "input_boolean.goodnight" = "on";
           "select.master_suite_current_mode" = "sleep";
           "switch.adaptive_lighting_sleep_mode_living_space" = "on";
-          "switch.eve_energy_20ebu4101" = "off"; # Whitenoise waits for sleep_done + bedroom audiobook
           "switch.desk_monitor" = "off";
           "switch.desk_pop" = "off";
           "cover.smartwings_window_covering" = "closed";

modules/services/hass/_tests/eval-automations.nix

@@ -514,12 +514,12 @@ let
       msg = "Good Morning scene doesn't turn off whitenoise";
     }
     {
-      test = (goodNightScene.entities."switch.eve_energy_20ebu4101" or null) == "off";
-      msg = "Good Night scene must leave whitenoise off until audiobook speaker gate";
+      test = !(goodNightScene.entities ? "switch.eve_energy_20ebu4101");
+      msg = "Good Night scene must not force whitenoise off (manual/automation owns it)";
     }
     {
-      test = (sleepScene.entities."switch.eve_energy_20ebu4101" or null) == "off";
-      msg = "Sleep scene must leave whitenoise off until audiobook speaker gate";
+      test = !(sleepScene.entities ? "switch.eve_energy_20ebu4101");
+      msg = "Sleep scene must not force whitenoise off (manual/automation owns it)";
     }
     {
       test = (sleepScene.entities."switch.desk_monitor" or null) == "off";

skills/catalog/autonomous-agent-loop/SKILL.md

@@ -12,7 +12,7 @@ Use this skill when the task should finish without repeated user nudges.
 1. Restate the requested end state as an outcome, not a task list.
 2. Define concrete verification evidence before editing: commands, diffs, rendered output, smoke checks, logs, or artifact paths.
 3. If a durable goal tool exists and no active goal covers the work, create one. Do not invent a token budget.
-4. For common Edmund workflows, prefer the prompt templates in `~/.pi/agent/prompts/`: `goalize.md` to start durable work and `goal-continue-audit.md` to recover from early stops. Keep project-specific details in the goal text or repo docs instead of creating many templates.
+4. Prefer existing `goalize` and `goal-continue-audit` prompt templates for common start/recovery loops. Keep project-specific details in the goal text or repo docs instead of creating many templates.
 5. Inspect repo/session state before changing files; preserve unrelated user changes.
 
 ## Work loop
@@ -24,6 +24,7 @@ Repeat until done or blocked:
 3. Inspect fresh evidence.
 4. Update the plan based on what happened.
 5. Continue without waiting for the user unless a decision/access blocker is real.
+6. For background jobs, use a blocking or longest bounded wait when available. After an unchanged status, increase the interval; never poll again immediately unless the status changed or a real deadline is near.
 
 Do not stop at: