Local chat and coding

Local Chat sends prompts to a selected installed model through the Desktop bridge and streams response events back to the interface. Local Chat can support code questions and selected file context. It does not establish a local coding agent with terminal access, repository editing, patch application, sandbox execution, or automatic source changes. Use the page as model-assisted chat, not as a substitute for a verified coding runtime.

Local chat

Open /chat after Ethen Desktop detects a runtime and at least one installed model. The page reads runtime status and inventory, can select the first installed model, manages sessions and messages in the interface, and submits chat through the bridge.

Response events include request-scoped deltas, completion, and error. The UI can stop a request, but cancellation is best effort when the bridge or runtime cannot terminate work immediately.

Browser fallback cannot generate. It returns an explanation that Ethen Desktop and a local runtime are required.

Local Chat streams model output through the Desktop bridge after runtime and model selection. Local Chat initializes runtime status and installed-model inventory and can auto-select the first installed model. Assistant output streams through request-scoped delta, completion, and error events. A stop action is available, although cancellation remains best effort when the bridge or runtime cannot stop work.

Local Chat initializes runtime status and installed-model inventory and can auto-select the first installed model; confirm the runtime and selected model before sending a prompt. File context can be inserted into the model request after a trust disclosure; inspect every file selected for context and remove secrets or unrelated material. A model response is advisory text and must not be described as an executed patch or verified repository change.

Browser fallback cannot generate and explains that Ethen Desktop and a runtime are required; use Local Chat for code explanation, review questions, planning, or discussion of deliberately selected file context. Assistant output streams through request-scoped delta, completion, and error events; stop generation when the response is no longer useful, while recognizing that cancellation is best effort. Treat a model response as advisory text, not as an executed patch or verified repository change.

Coding workflows

Use Local Chat for code explanation, design discussion, debugging ideas, test planning, and review of deliberately selected file context. Treat the model’s response as advisory text.

The inspected path does not provide terminal commands, repository writes, patches, branches, commits, sandbox execution, or automatic validation. Do not tell readers that a suggested edit has been applied.

A safe coding workflow is:

1. Select only the files needed for the question.
2. Ask for analysis or a proposed change.
3. Review the response for correctness and security.
4. Apply changes through the normal development environment.
5. Run tests and repository review independently.

Code discussion and selected file context are supported uses; repository execution is not. File context can be inserted into the model request after a trust disclosure. Chat sessions and messages are managed in the page, but durable persistence was not established. Browser fallback cannot generate and explains that Ethen Desktop and a runtime are required.

A stop action is available, although cancellation remains best effort when the bridge or runtime cannot stop work; use Local Chat for code explanation, review questions, planning, or discussion of deliberately selected file context. Browser fallback cannot generate and explains that Ethen Desktop and a runtime are required; stop generation when the response is no longer useful, while recognizing that cancellation is best effort. Confirm how the deployed environment stores sessions and files, where the runtime endpoint is located, and whether the runtime emits telemetry.

Local Chat initializes runtime status and installed-model inventory and can auto-select the first installed model; inspect every file selected for context and remove secrets or unrelated material. File context can be inserted into the model request after a trust disclosure; apply any suggested code change through the normal reviewed development workflow. Environment-specific review is required for session persistence, file retention, endpoint locality, and runtime telemetry environment-specific review.

Model selection

Select from the installed inventory reported by the connected runtime. An entry in Model Library is not enough, and a model listed as installed may still fail if the runtime cannot load it or the endpoint changes. Start with a short prompt, confirm that the selected model produced the response, and change one variable at a time when comparing models.

Only installed models reported by the detected runtime can be selected for Local Chat. Model Library is not a local inventory.

Choose a model based on the current runtime, available resources, context needs, and task. No universal model recommendation or hardware matrix is documented. A model that appears installed can still fail to load when memory or runtime compatibility is insufficient.

Record the exact selected identifier when comparing behavior. Similar names may represent different versions or quantizations.

Only installed models exposed by the detected runtime can be selected for local chat. No inspected file proves repository writes, terminal commands, patches, sandboxes, or coding-agent parity. Inspect every file selected for context and remove secrets or unrelated material. Stop generation when the response is no longer useful, while recognizing that cancellation is best effort.

Chat sessions and messages are managed in the page, but durable persistence was not established; inspect every file selected for context and remove secrets or unrelated material. Local Chat initializes runtime status and installed-model inventory and can auto-select the first installed model; apply any suggested code change through the normal reviewed development workflow.

A stop action is available, although cancellation remains best effort when the bridge or runtime cannot stop work; stop generation when the response is no longer useful, while recognizing that cancellation is best effort. Browser fallback cannot generate and explains that Ethen Desktop and a runtime are required; confirm the runtime and selected model before sending a prompt.

Context

File context is inserted into the model request after a trust disclosure. Select only the files needed for the question, review them for secrets or unrelated personal data, and verify where the configured runtime is hosted before sending file context. No fixed storage or deletion period is documented for selected files or local chat sessions.

For code-related work, use file context to ask for explanation, review suggestions, debugging ideas, or a proposed change in text. The response remains advisory. Apply any change through a normal repository workflow with version control, tests, and human review; Local Chat does not provide verified terminal access, sandbox execution, repository writes, or patch application.

Selected files become model-request context only after the trust disclosure appears. Review every file before selection. Remove secrets, credentials, personal data, unrelated proprietary content, and large generated files.

The endpoint determines where that context is processed. A configured endpoint may be remote. Do not rely on the Local Chat label without checking endpoint trust.

Fixed file retention, session persistence, and deletion behavior remain unverified for Local Chat. Keep sensitive context minimal and use organizational controls appropriate to the code.

File context enters the model request and should be reviewed as sensitive input. Apply any suggested code change through the normal reviewed development workflow. Confirm the runtime and selected model before sending a prompt.

No inspected file proves repository writes, terminal commands, patches, sandboxes, or coding-agent parity; stop generation when the response is no longer useful, while recognizing that cancellation is best effort. A stop action is available, although cancellation remains best effort when the bridge or runtime cannot stop work; confirm the runtime and selected model before sending a prompt.

Performance

Measure responsiveness with the selected model and actual context rather than relying on a catalog label. Longer prompts and attached files increase context work, while larger models can increase memory use and latency. If output stalls, reduce context, try a smaller installed model, and inspect the runtime before restarting Desktop.

Generation speed depends on model size, quantization, context length, CPU or GPU acceleration, memory bandwidth, runtime implementation, and concurrent work. A long file context can increase memory use and time to first output.

When performance is poor, reduce context, choose a smaller installed model, close competing workloads, and confirm that the runtime uses the expected acceleration path. Do not publish fixed performance claims from one device.

Stop generation when the response is no longer useful, while recognizing that cancellation can be delayed.

Latency and throughput depend on the model, context, runtime, hardware, and concurrent work. Use Local Chat for code explanation, review questions, planning, or discussion of deliberately selected file context.

Assistant output streams through request-scoped delta, completion, and error events; apply any suggested code change through the normal reviewed development workflow. Chat sessions and messages are managed in the page, but durable persistence was not established; use Local Chat for code explanation, review questions, planning, or discussion of deliberately selected file context.

Limitations

Autonomous code execution and durable repository context are outside the verified Local Chat contract. Keep those tasks in their dedicated reviewed tools.

Local Chat does not prove an offline or air-gapped workflow. The endpoint, runtime logs, model telemetry, file context, session storage, Desktop updates, and crash reporting can affect the data path.

No verified local coding-agent path provides terminal, sandbox, repository editing, patch application, or automatic tests. Session durability is also unverified.

Use Local Chat for controlled model-assisted reasoning and move execution into a reviewed development workflow. Keep the page as a draft until coding-runtime, persistence, file-handling, and endpoint-security contracts are established.

Terminal access, sandbox execution, patch application, and repository writes are not available as verified Local Chat capabilities.

File context can be inserted into the model request after a trust disclosure; confirm the runtime and selected model before sending a prompt. No inspected file proves repository writes, terminal commands, patches, sandboxes, or coding-agent parity; inspect every file selected for context and remove secrets or unrelated material.