# A Local Workbench for Operators, Not Another Chat Window Evaluate erm as a supervised Linux workspace for typed commands, media and local operational knowledge, with observable services and explicit permissions. Canonical HTML: ## Put the operation where the operator already works An engineer investigating a service failure moves between a terminal, a runbook, audio controls and an assistant window. Adding another chat interface can simply create another place to maintain context. The more useful question is which small operations can become easier without making their effects harder to inspect. erm approaches that question as a local Erlang/OTP workspace. The documented voice, media and knowledge components are interesting because they can share service state while keeping the visible interface replaceable. That is a more specific proposition than “an AI operating system.” The first audience is a Linux developer or operations team willing to evaluate native dependencies and a bounded set of actions. It is not a claim that an unmodified consumer desktop can be replaced without integration work. ## Services own the work; windows present it The [Erlang/OTP supervision model](https://www.erlang.org/doc/system/sup_princ.html) provides a structure for restarting child processes according to a chosen strategy. It does not automatically persist application state, restore an external player or make a window survive a whole-machine restart. For an erm evaluation, identify which process owns each operation and what state must survive an interface restart. A GTKGS window should expose that state rather than become the only place where an operation exists. Test the ownership boundary instead of assuming it from the presence of a supervisor. Start with a media command and a visible service-health view. Interrupt the interface, reconnect it and check whether the underlying operation remains understood and controllable. Treat stronger session-survival guarantees as separate acceptance requirements. ## Prefer typed actions to a model's invented command The [local voice reference](erm_local_voice.md) describes a direct path for common media actions and optional model assistance for unfamiliar phrases. It also separates speech recognition, intent handling, media control and speech output. That makes a failed interaction easier to locate than a single opaque “assistant failed” message. A transcript is not the completed action. The pilot should observe the player state or other intended effect, then check that the spoken and visible responses match the result. Include negated phrases, incomplete utterances, duplicate transcripts and an unavailable speech-output dependency. Do not make an arbitrary shell command the default interpretation of an unrecognised sentence. Keep the action vocabulary explicit and require a separate permission decision before adding an operation that can change files, access secrets or affect a remote service. ## Local is a data-flow property, not a label Local Whisper or Piper integration and a local model endpoint are useful components, but the complete deployment also includes model retrieval, configuration, logs and any enabled network integrations. Inspect that path before promising that every interaction stays offline. Define which documents the knowledge interface may retrieve and which model may receive their contents. A model listening on loopback says where the immediate request goes; it does not establish every subsequent action of that model service. Voice recognition is also not an authorisation system. This evaluation does not assert biometric authentication, safe voice-triggered spending or permission to administer a machine merely because a wake phrase was heard. Keep sensitive operations out of the initial action set. ## Demonstrate one interruption and one recovery Choose a handful of everyday commands and one approved runbook lookup. Measure whether they complete, how long feedback takes and whether the operator can explain a failure using the service-health view. Compare this with the same task performed through existing controls. Disconnect a native dependency, stop a model service and restart the UI. Record which component reports the failure, whether an action is repeated unexpectedly and what the operator must do to recover. A polished success animation is less informative than a clearly bounded failure. For a team pilot, preserve configuration and model identities with the result. A successful test on one microphone, GPU or desktop session should not be advertised as coverage of every Linux workstation. ## Implementation and evaluation boundary The feature detail here comes from the dated voice/media guide and readable application declaration. No microphone, native GTKGS surface, model, player or workstation recovery test was exercised for this article. Other recent desktop applications are not enumerated as verified additions by this review. Begin with [the implementation guide](erm_local_voice.md). Evaluate erm as an inspectable local workbench for a specific operator, then expand only the actions whose behaviour and recovery have been demonstrated.