com-google-android-apps-chromecast-app-control-hvac-orchestrated

name	com.google.android.apps.chromecast.app.control-hvac-orchestrated
clawperator-skill-type	orchestrated
description	Agent-driven Google Home HVAC controller for one named climate action per run.

Runtime program for an assisted orchestrated Google Home HVAC skill.

This first pass is intentionally one action per run. The runtime contract is:

• --action <temperature|mode|fan_speed|climate_state>
• --value <target>
• --unit-name "<label>"

This pass is assisted from nearby patterns, not from-scratch only:

• the retained recording is the main route evidence for controller entry, temperature adjustment, mode-sheet entry, and fan-speed-sheet entry
• the existing replay siblings informed nearby naming and the fresh-session verification pattern
• power verification specifically reuses the same fresh-session reread idea as the replay power skill because this recording did not include an on/off tap

Truthfulness boundary:

• this skill is somewhat generic across Google Home HVAC units that expose the same controller structure
• this pass does not claim every Google Home climate device has identical mode names, fan-speed options, or controller layout
• this pass supports power control as an assisted nearby-pattern branch even though the retained recording did not itself include a power-button action

The currently supported runtime agent is codex. This skill runs through Codex with danger-full-access sandbox posture so the runtime agent can reach the live adb target through the local Clawperator CLI.

Immediate execution rules:

• Do not summarize the plan. Start with a real Clawperator CLI command.
• Stay inside com.google.android.apps.chromecast.app.
• Do not open launcher search, the Google app, Assistant, Chrome, Settings, or any unrelated app.
• Use one live device command at a time.
• Do not use skills install, skills list, skills get, or any other local skills-store management command during a live run.
• Do not call clawperator skills run from inside this skill. That would recurse back into the wrapper instead of operating the device.
• Do not run clawperator --help, clawperator exec --help, or any other help or introspection command during a live run.
• Do not run clawperator exec ... --validate-only during a live run.
• Do not use exec best-effort.
• Do not use the flat wait command for this skill. Use bounded exec payloads with wait_for_node instead.
• Do not emit a final success result unless terminal verification came from a fresh-session reread of the Google Home controller state.
• If the visible UI differs materially from this route, fail or mark the run indeterminate truthfully instead of improvising a different product journey.

Inputs:

• raw argv passed through by clawperator skills run
• clawperator skills run provides the selected device as CLAWPERATOR_DEVICE_ID
• canonical invocation shape: clawperator skills run com.google.android.apps.chromecast.app.control-hvac-orchestrated --device <device_serial> -- --action temperature --value 24 --unit-name "Panasonic"
• valid actions:
  • temperature
  • mode
  • fan_speed
  • climate_state

Runtime contract:

• emit goal: { "kind": "control_hvac", "action": <action>, "value": <value>, "unit_name": <unit_name> }
• emit inputs: { "action": <action>, "value": <value>, "unit_name": <unit_name> }
• preserve this checkpoint subset in this exact order:
  • app_opened
  • controller_opened
  • current_state_read
  • action_applied
  • terminal_state_verified
• each checkpoint must include both id and status
• checkpoint status must be one of ok, failed, skipped
• terminalVerification status must be one of verified, failed, not_run
• do not emit source; Clawperator injects it from skill.json
• emit exactly one terminal [Clawperator-Skill-Result] frame on stdout

Recorded route and nearby-pattern route:

1. Open Google Home from a fresh app session.
2. Ensure the Home tab is active.
3. Scroll horizontally to the Climate chip if needed.
4. Scroll to the requested climate tile and long-press it to open the controller.
5. Verify the controller toolbar title exactly matches unit_name.
6. Read the current state for the requested action before changing anything.
7. Apply exactly one requested action.
8. Close and reopen Google Home, reopen the same controller, and reread the resulting state for terminal verification.

Action branches:

• temperature
  • read resourceId=com.google.android.apps.chromecast.app:id/low_value
  • if the visible setpoint already matches the requested integer, treat the action step as a no-op but still do the fresh-session reread for proof
  • otherwise use Increase temperature or Decrease temperature until the requested integer is reached
  • reread low_value after each adjustment attempt instead of assuming a tap landed
• mode
  • read the current mode from the Mode action tile
  • click the Mode ... tile
  • wait for the Select a mode sheet
  • choose the requested visible mode label
  • wait until the controller view is back before moving on
• fan_speed
  • read the current fan speed from the Fan speed action tile
  • click the Fan speed ... tile
  • wait for the Fan speed pop-up
  • in the bottom sheet, choose the requested visible option label exactly as rendered there; on the proving device this sheet uses lowercase labels such as auto, high, low, and med
  • wait until the controller view is back before moving on
• climate_state
  • read the current state before acting
  • interpret a numeric low_value as powered on
  • interpret literal Off as powered off
  • only click the power button when the current state differs from the requested state
  • emit normalized terminal verification text as lowercase on or off so it stays aligned with the declared value input
  • use the fresh-session reread as the only trustworthy terminal proof, just like the replay power sibling

Navigation policy:

• the only allowed Google Home path is the recorded Home -> Climate -> named tile -> long press controller route
• do not explore Favorites, Devices, Activity, Settings, or device settings as part of this run
• if the controller title does not match the requested unit label after open, fail truthfully instead of applying the action to a different unit

Verification policy:

• terminal verification must come from a fresh-session reread
• for temperature, verify the reopened controller low_value
• for mode, verify the reopened Mode action tile body text
• for fan_speed, verify the reopened Fan speed action tile body text
• for climate_state, verify the reopened normalized lowercase power state from low_value
• do not trust the immediate in-place controller after a change as final proof

Execution templates:

• prefer the exact exec --execution route shown in the harness prompt over inventing new exploratory commands
• for temperature, use a bounded controller-entry exec, then a bounded read-current-temperature exec, then one or more bounded adjustment execs, and finally a fresh-session reread exec
• for fan_speed, use a bounded controller-entry exec, then a bounded read-current-fan-speed exec, then a bounded open-sheet-and-click-option exec, and finally a fresh-session reread exec

Known caveats:

• this first pass assumes the named climate unit is reachable from the Google Home Home tab under Climate
• power support is informed by the replay sibling because the retained recording did not include a power-button tap
• mode and fan-speed option labels are device-specific and must match the visible option text
• this skill proves the state visible in Google Home, not the physical HVAC hardware response

Debugging support:

• the harness writes a per-run prompt file, agent stdout log, agent stderr log, and metadata file into a temporary run directory
• set CLAWPERATOR_SKILL_RETAIN_LOGS=1 to keep those artifacts after a successful run
• set CLAWPERATOR_SKILL_DEBUG=1 to also stream agent stdout to stderr while retaining the run directory
• set CLAWPERATOR_SKILL_LOG_DIR=<dir> to place retained run directories under a predictable parent path during local debugging

Emission rules:

• no extra prose after the final frame
• the final non-empty stdout suffix must be:
  1. [Clawperator-Skill-Result]
  2. one JSON object line containing the emitted SkillResult
• the emitted SkillResult must include result before status; use an evidence-shaped JSON result for a confirmed final state and result: null when no truthful final state is available

Recording note:

• recording-context.json and references/compare-baseline.export.json are authoring and compare evidence only
• they are not runtime inputs to this skill