> ## Documentation Index > Fetch the complete documentation index at: https://docs.nika.sh/llms.txt > Use this file to discover all available pages before exploring further. # Resident server > Operate the stable nika serve firer on a container, VM or bare server. export const ProjectContract = () => <> Live project control plane. The released engine discovers nika.yaml from the working directory toward its ancestors. The file is optional when built-in defaults are enough. Its project-shape grammar is closed: nika · ceiling · arm · traces · registry. Current released reader limit. Direct workflow operations accept the traces and registry project-policy rungs. The cadence path behind nika arm and nika serve currently accepts only nika, optional ceiling and arm. Keep those profiles separate until the engine readers converge. ; `nika serve` is the released resident firer. It reads the `arm:` registry from `nika.yaml`, waits on the wall clock and ends each due beat at the same firing law as `nika arm fire`. `nika serve` is not an HTTP API. It accepts no SDK job request and exposes no SSE endpoint. Do not point the root `Nika` client at it. ## Rehearse each operating mode ```sh theme={"system"} nika serve --dry nika serve --once nika serve ``` | Mode | Behavior | | -------- | -------------------------------------------------- | | `--dry` | Preview calendar-due beats, run and record nothing | | `--once` | Perform one due sweep, then exit | | no mode | Remain resident and wait for the next slot | `--dry` is intentionally narrower than a real fire. It does not take the beat lock or apply the complete missed-run, overlap, cost and execution policy. ## Reload without losing the last good registry

HOT RELOAD · INVALID EDITS DO NOT REPLACE VALID POLICY

    {`nika.yaml changes
             │
             ├── parses + validates ── replace registry
             │
             └── refuses ───────────── report error
                                         └── keep last good registry`}

The server retries a refused edit on a later tick. This keeps a typo visible without silently replacing valid policy with an empty schedule. ## Read process and beat outcomes separately Each beat prints one decision line and appends project state. A completed `--once` sweep exits clean even when an individual workflow failed or paused, so supervision must read the beat lines and ledgers, not only the outer process exit. ```text theme={"system"} resident process ├── stdout line per beat ├── .nika/arm//history.ndjson └── .nika/traces/.ndjson ``` ## Delivery and shutdown Ctrl-C and SIGTERM stop the resident loop at its signal boundaries. The ledger uses durable claims and receipts, which yields at-least-once delivery: ```text theme={"system"} claim appended ── workflow starts ── receipt appended │ │ └── crash here ────┴── visible unmatched claim ``` Exactly-once is never promised. Make the workflow's external effects idempotent when replay would otherwise duplicate harm. ## Deployment checklist * Set a stable project `cwd` containing `nika.yaml`. * Install an explicit `nika` binary path. * Protect provider keys outside project source. * Persist `.nika/traces/` and `.nika/arm/` according to policy. * Capture stdout decision lines. * Send SIGTERM and allow the active firing boundary to settle. ## Continue The schedule and policy this process consumes. Let launchd or systemd own wakeups instead. Persist the evidence behind each decision line. Boot, supervise, probe and investigate the process. Choose the right process boundary.