Interactive chat entry points
On a local kind target, use /inference to choose Foundry (Azure login), Copilot CLI, or the Agent's native Orka Provider. Host Foundry and Copilot modes are refused on remote targets; configure a cluster Provider there instead. /inference-foundry opens saved configuration, Azure browse or manual endpoint setup using the host's authenticated Azure CLI. Foundry host mode uses Entra authentication and native model function calls with KMX's supported cluster HTTP tools. See local Foundry inference for scope and setup.
Quickstart's post-wizard chat and direct Orka chat now use the same entry point, backend, renderer and command loop:
kmx chat hello-world-agent
kmx chat --runtime orka --namespace orka-system hello-world-agentThe default runtime selection discovers Orka's API and checks the named Agent in orka-system (or --namespace). A matching Orka Agent uses the wizard chat shell: sticky agent/location header, tools, /agent, inference selection, verbose controls, response timings and reusable connections. Discovery/read errors are reported rather than silently connecting to a different runtime.
If no Orka Agent matches, the command says so. There is no fallback: the Kagent create-only adapter is not registered for chat or discovery. Its former resumable chat, A2A approval flow and in-chat preset governance remain removed, and --runtime kagent is refused rather than resolved to Orka. Those runtime-specific semantics are not implemented by the Orka Task API.
An optional message after the Agent name is sent once before the interactive prompt. Orka chat is always a session, so that message is not a one-shot; use kmx run for one Task and an answer-only stdout. The --interactive flag is removed; v0.4.1 and earlier require it. Direct Orka chat starts with the Agent's current Provider; use /inference-copilot on a local kind target to choose Copilot, or /inference-local to return to the Agent's Provider. The wizard retains its explicitly chosen inference source when entering this same shell.
Shell Agent-name completion reads Orka names in the selected namespace, and respects the namespace flags. No additional runtime CLI installation is needed for Orka chat.
Typing / opens a floating completion menu above the message editor. It filters as you type and lists only the active backend's commands. Up/Down selects; Tab or Enter inserts a completion, and Enter on the completed command runs it. The menu clears on submission and does not appear in approval/answer prompts.
Full-screen chat reflows when terminal dimensions change, retaining typed input and the completion selection. It redraws the sticky header and anchors a fresh editor below the existing transcript rather than erasing with obsolete cursor coordinates. Resize itself never submits a message or approval. The non-full-screen raw-input fallback retains its conservative resize-abort behavior.
Chat /lift and console Lift only print Lift moved: kmx lift <ref>@<digest> --to-context … --evidence …. They perform no release workflow or source-less copy. Host Foundry/Copilot controls remain independent; live source tools/inference editing is retired.
Message editing and scrollback
Full-screen Orka chat uses a retained transcript viewport with a pinned header and message editor. The same interface is used after quickstart and by Orka chat.
- Ctrl-W: delete the word before the cursor.
- Ctrl-A / Ctrl-E: move to the beginning/end of the message.
- Left / Right: move within the message; edits happen at the cursor.
- Up / Down: recall sent messages; Down past the newest restores the draft. When slash completions are open, Up/Down selects a completion instead.
- PageUp / PageDown, Shift-Up / Shift-Down, or the mouse wheel: scroll conversation history while leaving the header and editor fixed.
- Ctrl-Home / Ctrl-End: oldest history / latest messages.
- Shift-Enter / Alt-Enter: insert a newline. Enter sends when idle.
History can be browsed and the next message drafted while a response is running; Enter does not submit another request until the current one finishes. Ctrl-C cancels the active worker and restores the terminal. /agent resets conversation and recall history on a successful switch. Tool-call headings and payloads are indented underneath the agent heading, including after wrapping on narrow windows.