FAQ and troubleshooting
Start with getting started for installation and a first answer and suite workspace for author/eval/push/digest Lift.
Why is kmx orka missing?
Check kmx version and which binary is on your PATH. The historical v0.1.0 release predates Orka; v0.2.0 and newer provide these commands. Install the pinned release as described in getting started, then check kmx orka --help. @latest and the installer default select the newest published tag, not the tip of main.
Does installing Orka govern my application?
No. Installation creates platform resources and, by default, a local Provider. It does not adopt an existing application or redirect its traffic. Author an native workspace member with kmx create, then evaluate, publish and Lift with private evidence. Placement readiness is not an answer. Execution and enforcement belong to the selected runtime.
Why does status show a version different from the pin?
kmx orka status reads the running controller image. The pin is what kmx would install, not proof of what someone else installed. An unreadable cluster is not an empty cluster. See Orka installation for inspection, no-write planning, dry-run and upgrade limits.
What kagent support exists now?
Orka remains default. Explicit native kagent workspaces use exact v0.10.3 Go/linux-amd64: create, eval --bindings FILE, evaluated push, native receipt diff and create-only digest Lift. Every occupied target refuses, even with matching evaluated UIDs. Eval is unsecure-only, one fresh blocking A2A Task per case, never retried; trusted-proxy refuses before admission/send. No Python, RemoteMCP, standalone run/chat, suite status/retirement, installer/discovery, console/quickstart or AKS kagent payload is implied. See native workspace profiles and runtime boundaries.
Live v0.10.2 create and bundle lift/status/evaluation/replay/readers/receipts are removed outright. Old receipts are unreadable; no conversion/migration or deletion of saved files/resources occurs. Reference extraction/export remains independent, without native RecordedExecution or a manifest consumer. Author native cases. Partial placement may leave resources; no retry/adoption/rollback is implied.
Model and tool troubleshooting
Empty replies and input-required
This was the former Kagent chat path's human-in-the-loop state. The current packaged native evaluation accepts only completed answers and has no continuation operation for input-required; Orka turns likewise complete or fail. Historical note, for transcripts that still show it: the small-model path used to re-sample a question-only response up to twice, but not after a tool call or with an explicit session; there is no guarantee that a system instruction suppresses questions. Malformed ask_user arguments from smaller models can also fail invocation. The committed keyless model is qwen2.5:3b; test any replacement by invoking it.
The tool worked but the answer is wrong
A model may misquote valid tool results. Compare the task's actual tool call/result with the prose summary. CI asserting a tool path is not proof that a model reasons correctly or copies identifiers faithfully.
Hosted model authentication fails
For native Orka, provision the Provider's named Secret and key separately, without a trailing newline. Provider readiness is not proof of a successful model call. Inspect the Task and the model/deployment identifier as well as Provider and Agent readiness. See models.md for native configuration.
Host Copilot authentication is managed by the installed Copilot CLI, not a KMX cluster token exchange; see Copilot inference. Host Foundry uses its own endpoint and credential path; see Foundry inference. Never put token values in command arguments or committed YAML.
A local model disappeared after a restart
The local Ollama model cache is an emptyDir; a pod restart may require another model pull. Pod restarts are not cluster deletion. kmx local down removes the whole named kind cluster and its persistent data. Export needed resources, Secrets and volumes before replacing it; see Orka's data-retention limits.