kandev/docs/specs/go-dev-launcher/spec.md

status, created

status	created
draft	2026-08-07

Go dev launcher and minimal Node surface

Why

make dev is the last Kandev entrypoint that runs through the TypeScript launcher. make start (Makefile:277) and every make service-* target already exec the native Go binary at apps/backend/bin/kandev; only make dev still shells out to pnpm -C cli dev -- dev (Makefile:195), which runs tsx src/cli.ts.

That single call site keeps roughly 3,000 lines of TypeScript alive under apps/cli/src/** — dev.ts, plus run.ts, start.ts, service/, and supervisor/, which are already dead code because their Go equivalents (launcher.runInstalled, launcher.runStart, launcher.runService, launcher/supervisor.go) shipped and took over the Makefile targets. None of that TypeScript is published: apps/cli/package.json declares files: ["bin/cli.js", "bin/native-shim.js"], so the npm kandev package is already a four-line Node shim that execs the Go binary. The src/ tree exists solely to serve make dev.

The cost of keeping it is not just lines. The dev launch chain today is bash → make → pnpm → node → make → sh → kandev, and that chain is why apps/backend/cmd/winjob exists at all: MSYS bash, native Win32 processes, and Node disagree on signal propagation, so Ctrl-C leaks processes on Windows unless the backend is wrapped in a Job Object (apps/cli/src/dev.ts:169-199). The Go launcher already solves this generically with CREATE_NEW_PROCESS_GROUP plus taskkill /F /T (launcher/process_group_windows.go), so moving dev to Go removes the winjob dependency from the dev path rather than reimplementing it.

Behavior also drifts between the two implementations. PR #2394 ("harden launcher port ownership") had to land the same health-token and explicit-port-preflight fix twice — once in apps/cli/src/{args,health,ports,run,shared,dev}.ts and once in apps/backend/internal/launcher/{env,health,ports,run,start}.go. Every launcher change pays that double-implementation tax until dev moves.

What

kandev dev becomes a native launcher command

The Go launcher gains a third command alongside run and start. It accepts the same surface the TypeScript dev accepted:

Input	Behavior
dev, --dev	Select dev mode.
--port, --backend-port, KANDEV_BACKEND_PORT, KANDEV_PORT	Backend port. Existing preflight applies: an explicitly requested occupied port is a hard error naming the port and its source.
--web-internal-port, KANDEV_WEB_PORT	Internal Vite port. Rejected outside dev mode.
--web-port	Rejected with the existing "use --web-internal-port for dev mode" message.
--verbose, --debug, --headless, --no-browser	Unchanged meaning.

Omitted ports keep the current preferred-then-random selection (38429 backend, 37429 web, 39429 agentctl), and the web and agentctl ports must not collide with the backend port or with each other.

Dev state stays isolated from production

Dev mode roots Kandev at <repoRoot>/.kandev-dev so make dev never mutates the user's ~/.kandev state and so make clean-db (which removes .kandev-dev/) matches what make dev writes. The repo root is discovered by walking up from the working directory for a parent containing both apps/backend and apps/web; failing to find one is a usage error, not a silent fallback.

KANDEV_DATABASE_PATH resolution keeps its current three-way rule:

• Inside a Kandev task workspace — signalled by KANDEV_TASK_ID being set, or by the repo root living under ~/.kandev/tasks/ — any inherited KANDEV_DATABASE_PATH is assumed to be leaked from the parent backend and is cleared, so the backend derives its DB from KANDEV_HOME_DIR. The launcher prints that it detected a task workspace.
• Otherwise an explicit KANDEV_DATABASE_PATH is honored as an escape hatch.
• Otherwise the DB is <repoRoot>/.kandev-dev/data/kandev.db.

Dev mode always sets KANDEV_DEBUG_DEV_MODE=true, which is the profile selector; the concrete dev defaults stay in profiles.yaml. The launcher must not restate them.

Production databases are backed up before dev touches them

When the resolved DB path is not under a .kandev-dev segment, the launcher snapshots it to ~/.kandev/data/backups/dev-prod-db-<timestamp>.db before starting the backend, retaining the five newest such snapshots and leaving other backup families (kandev-*, manual-*) untouched. This is what makes make dev-prod-db safe.

A backup failure aborts startup with a non-zero exit. The guard exists precisely to protect production state before dev mode touches it; continuing past a failed backup would remove the guarantee that justified it.

The Vite dev server is a supervised child

The launcher starts the frontend with pnpm -C apps --filter @kandev/web dev, passing PORT, HOSTNAME=127.0.0.1, and KANDEV_API_BASE_URL, setting KANDEV_DEBUG=true and VITE_KANDEV_DEBUG=true, and explicitly removing any inherited VITE_KANDEV_API_PORT so a host .env, Docker env, or CI variable cannot reintroduce cross-origin browser API calls. On Windows, pnpm is invoked through cmd.exe /c because Node-style .cmd shims cannot be spawned directly.

Browser traffic still enters through the Go backend, which reverse-proxies to the Vite port via KANDEV_WEB_INTERNAL_URL. The launcher waits for the backend health endpoint (dev timeout: 600s, versus 45s for release) and then for the Vite URL to answer, then opens the backend URL. The Vite readiness probe accepts any HTTP response and does not require the launcher health token — Vite is not a Kandev process and does not emit one.

If either child exits, the launcher shuts the whole tree down and exits with the child's code, matching current behavior.

Startup output advertises network addresses

The startup banner for dev, start, and run prints the localhost backend URL and then one network: URL for each unique non-loopback, non-link-local host address on the machine. This includes addresses from local-network and Tailscale interfaces so that a user running Kandev on a remote VM can identify an address to open from another machine.

The banner lists IPv4 addresses before IPv6 addresses, omits link-local addresses, and wraps IPv6 hosts in brackets. Network-address discovery is best effort: if the host cannot enumerate its interfaces, the launcher still starts and prints the localhost URL. When KANDEV_SERVER_HOST restricts the backend to loopback or specific addresses, the banner omits unreachable interface addresses and only advertises matching binds.

Backend restart from the UI still rebuilds

The restart supervisor (ADR 0019) writes a launch manifest and listens on a control socket so Settings can restart the backend. In dev that restart must rebuild from source, which is what makes the affordance useful after editing Go code. The dev launcher therefore keeps make -C apps/backend dev as the backend child command rather than re-execing itself.

That creates a constraint the TypeScript launcher never had: the launcher and the backend are the same binary (bin/kandev, dispatched by the __backend argument), so a rebuild would overwrite the running launcher's own executable — which fails outright on Windows. make dev therefore execs a copy of the freshly built binary from a distinct path, and the copy supervises make -C apps/backend dev. The copy is a build artifact, not a second shipped binary.

make dev stops depending on Node and winjob

dev: doctor
	@$(MAKE) -C $(BACKEND_DIR) build-kandev
	@cp bin/kandev → bin/kandev-launcher
	@exec $(BACKEND_DIR)/bin/kandev-launcher dev $(DEV_FLAGS)

build-kandev is the only root prebuild: the copied supervisor binary does not need agentctl helpers before it starts the supervised child. The child's make -C apps/backend dev runs dev: build-dev, which builds the native agentctl and, on hosts other than Linux/amd64, one linux/amd64 helper, plus mock-agent, acpdbg, and winjob. Full four-platform helper builds remain under build-runtime for release and service bundles.

build-winjob leaves the dev path: the Go launcher's process-group handling replaces it, and the root dev target no longer invokes the winjob build step. cmd/winjob itself stays in the tree for any other consumer. The lean dev: build-dev target still builds winjob once because the backend's Windows process handling remains part of the local development binary set.

apps/cli shrinks to the published shim

After dev lands in Go, apps/cli/src/** has no consumer and is deleted in full, together with tsconfig.json and vitest.config.ts. What remains is exactly what npm publishes plus its test:

apps/cli/
├── package.json
└── bin/
    ├── cli.js              # 4-line exec of the native binary
    ├── native-shim.js      # platform-package resolution
    └── native-shim.test.mjs

native-shim.test.mjs moves from Vitest to the Node built-in test runner, so apps/cli needs no devDependencies at all and make test-cli becomes node --test. The package keeps its name, version, bin, files, and optionalDependencies untouched: the npm release flow (scripts/release/publish-npm.sh, .github/workflows/release.yml) reads and bumps apps/cli/package.json and must keep working unchanged.

Node remains required for exactly three things, and this spec does not attempt to remove them: building and serving the web app (Vite), the published npm shim, and repo tooling (prettier, commitlint, Playwright). That is the floor.

Non-goals

• Removing Node from the web build or from apps/web. Vite is the dev server.
• Changing what the npm kandev package ships or how it resolves @kdlbs/runtime-*.
• Removing apps/backend/cmd/winjob from the repository.
• Changing make start, make service-*, or the desktop/Tauri launch path, all of which already use the Go binary.
• Changing E2E backend spawning, which already execs KANDEV_BIN __backend directly (apps/web/e2e/fixtures/backend.ts:287).

Failure modes

Condition	Expected behavior
Repo root not found from CWD	Exit 2 with a message telling the user to run from the repo.
Explicit backend or web port occupied	Exit 1 naming the port and its source (flag vs. env). No browser is opened.
Production DB backup fails	Exit 1 before the backend starts; the message names the DB path.
Backend fails health within 600s	Print the captured backend output, shut down the tree, exit 1.
Vite never becomes ready	Shut down the tree, exit 1.
Either child exits	Shut down the tree; exit with that child's code, or 0 if it was signalled.
pnpm missing	Surface the spawn error rather than hanging on the readiness probe.
Ctrl-C	Whole tree terminates, including on Windows, without winjob.
Network interface enumeration fails	Continue startup and omit network: lines; keep the localhost URL.

Persistence guarantees

• .kandev-dev/ is the only state make dev writes, unless the user explicitly sets KANDEV_DATABASE_PATH outside a task workspace.
• dev-prod-db-*.db snapshots are retained five deep in ~/.kandev/data/backups/ and never prune other backup families.
• The supervisor manifest and control socket live under <KANDEV_HOME_DIR>/supervisor/ at mode 0600, unchanged from start.

Scenarios

• GIVEN a host with non-loopback LAN and Tailscale IPv4 addresses, WHEN the dev, start, or run launcher prints its startup banner, THEN it prints one network: URL for each address using the backend port.
• GIVEN a host with duplicate addresses, loopback addresses, or link-local addresses, WHEN the startup banner is printed, THEN duplicate, loopback, and link-local addresses are omitted.
• GIVEN a host with an IPv6 address, WHEN the startup banner is printed, THEN the address is rendered as http://[<address>]:<port>.
• GIVEN interface enumeration returns an error, WHEN the launcher starts, THEN it continues startup and prints the localhost URL without a network: line.
• GIVEN KANDEV_SERVER_HOST is a loopback or specific bind address, WHEN the startup banner is printed, THEN it omits unrelated network addresses.