List all configured models with their current status (running/stopped). Returns {models: [...], count: N} where each entry is nested as {identification: {name, model_path}, status: {state, port, pid?}}. The value to pass as model_name to start_server / swap_server is identification.name exactly as returned — do NOT concatenate it with status.port or any other field.

Get the full configuration for a specific model by name

Start a model on an empty port. Fails with action='rejected_occupied' if a different model is already running on that port — use swap_server for that case. Both 'model_name' and 'port' are required; the port is always specified by the caller (ADR-010). The model_name must exactly match a model from list_models.

Stop whatever is running on this port. Idempotent: returns success with action='already_empty' if nothing was there. Returns action='stopped' on a successful termination.

Replace the model on this port with a different one. Primary use: an agent replacing its own brain on the harness's expected port. Performs the 5-phase swap (pre-flight, marker, stop, start, readiness) with rollback to the previous model on failure. Calling with the model already running on the port is a successful no-op (action='already_running'). Fails with action='rejected_empty' if the port is empty — use start_server for that case.

Cancel an in-flight start or swap on this port (ADR-014). Sets a cancel flag on the in-flight marker; the running op picks it up at the next phase boundary (typically within ~1 s during readiness poll). Returns success with marker_existed=False if there is no in-flight op — 'nothing to cancel' is a successful no-op, not an error. A cancel that arrives after the new process has been spawned and the lockfile written is ignored; the op completes and reports cancel_ignored_post_commit=True.

Get the status of all running llama-servers

List unmanaged llama-server processes on this node (ADR-015). An orphan is a live llama-server that llauncher did not launch — its (port, pid) does not match any live lockfile. Returns each orphan's pid, port (when discoverable from argv), and a cmdline_unreadable flag for processes whose argv could not be read. Adopt is intentionally not exposed in this revision — see ADR-015 §Deferred Work.

Fetch recent logs for a running server by port

Live in-server inference telemetry for a running server (ADR-LLNCH-019): phase (idle/prompt/generating), gen_tok_s, prompt_tok_s, slot counts, started_at. Safe tier — no prompt text. Local-node only. Returns a degraded envelope ({'available': false, 'reason': 'loading'|'no-metrics-flag'|'unreachable'}) rather than erroring when the target server can't be read.

Per-slot detail for a running server, including prompt text (ADR-LLNCH-019). Sensitive tier — grant separately from server_metrics. Local-node only. Returns {'available': false, 'reason': 'slots_disabled'} when the server was not started with --slots.

Update an existing model's configuration

Validate a model configuration without applying it

Add a new model configuration

Delete a model configuration. Refuses with action='rejected_in_use' (and the holding port) if the model is currently running anywhere; stop or swap first. Idempotent on a missing name (action='not_found').