Skip to content

memory.dashboard.launcher_core

memory.dashboard.launcher_core

Shared launcher core for the memory and zettelkasten dashboard stacks.

Both dashboards spawn the same shape of process tree — a FastAPI backend and (in a source checkout) a Vite dev server — and historically each had its own near-identical __main__.py. This module holds that logic once; each dashboard builds a :class:LauncherConfig and calls :func:run. The chat agent now runs in-process in the backend (cursor-sdk's hermetic bridge ships a bundled node binary), so there is no longer a separate Node sidecar to launch.

Beyond de-duplication it adds three pieces of robustness over the original copy-paste launchers:

  • Health-confirmed port skip. A busy port is no longer assumed to be our dashboard. The launcher hits /api/health and only skips when the live server identifies itself as our service; an unrelated process holding the port is reported as a conflict instead of silently masking a broken stack.
  • Wait-for-health. After spawning the backend the launcher polls its health endpoint and only prints the "open this URL" line once it actually answers, so users don't click through to a connection-refused page.
  • Restart-with-backoff. A component that dies is respawned with exponential backoff (resetting after a stable run) instead of just being declared dead, so a transient crash self-heals without a manual relaunch.

LauncherConfig dataclass

Everything that differs between the two dashboards.

backend_port / frontend_port are the base defaults. The actual ports are resolved per-workspace in :func:resolve_ports: an explicit env override (backend_port_env / frontend_port_env) always wins, else the base is shifted by a deterministic per-workspace offset so two checkouts of the same dashboard (e.g. angelo and a second clone) land on different ports and don't fight over one.

Source code in memory/dashboard/launcher_core.py 
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
@dataclass(frozen=True)
class LauncherConfig:
    """Everything that differs between the two dashboards.

    ``backend_port`` / ``frontend_port`` are the *base* defaults. The actual
    ports are resolved per-workspace in :func:`resolve_ports`: an explicit env
    override (``backend_port_env`` / ``frontend_port_env``) always wins, else the
    base is shifted by a deterministic per-workspace offset so two checkouts of
    the same dashboard (e.g. angelo and a second clone) land on different ports
    and don't fight over one.
    """

    dashboard_dir: Path
    backend_app: str          # uvicorn import path, e.g. "memory.dashboard.backend.main:app"
    service_name: str         # base value returned by /api/health, e.g. "memory-dashboard"
    backend_port: int
    frontend_port: int
    workspace: Path
    install_hint: str = 'pip install -e ".[zettelkasten,dashboard]"'
    backend_port_env: str = "DASHBOARD_BACKEND_PORT"
    frontend_port_env: str = "DASHBOARD_FRONTEND_PORT"

    @property
    def frontend_dir(self) -> Path:
        return self.dashboard_dir / "frontend"

workspace_instance_id 

workspace_instance_id(workspace: Path | str) -> str

Stable short id for a workspace, derived from its resolved path.

Deterministic (not random) on purpose: the launcher and the backend each compute it independently, and a relaunch must recognise its own already running dashboard rather than spawning a duplicate. Two different checkouts therefore get two different ids and never claim each other's port.

Source code in memory/dashboard/launcher_core.py 
82
83
84
85
86
87
88
89
90
91
def workspace_instance_id(workspace: Path | str) -> str:
    """Stable short id for a workspace, derived from its resolved path.

    Deterministic (not random) on purpose: the launcher and the backend each
    compute it independently, and a relaunch must recognise its *own* already
    running dashboard rather than spawning a duplicate. Two different checkouts
    therefore get two different ids and never claim each other's port.
    """
    resolved = str(Path(workspace).expanduser().resolve())
    return hashlib.sha1(resolved.encode("utf-8")).hexdigest()[:8]

current_instance_id 

current_instance_id() -> str

The instance id for the running process.

Prefers the value the launcher exported (:data:INSTANCE_ENV); falls back to deriving it from ANGELO_WORKSPACE (or cwd) so a backend started directly by uvicorn — outside the launcher — still reports a consistent identity.

Source code in memory/dashboard/launcher_core.py 
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
def current_instance_id() -> str:
    """The instance id for the running process.

    Prefers the value the launcher exported (:data:`INSTANCE_ENV`); falls back to
    deriving it from ``ANGELO_WORKSPACE`` (or cwd) so a backend started directly
    by uvicorn — outside the launcher — still reports a consistent identity.
    """
    pre = os.environ.get(INSTANCE_ENV)
    if pre:
        return pre
    ws = os.environ.get("ANGELO_WORKSPACE", os.getcwd())
    return workspace_instance_id(ws)

service_identity 

service_identity(base_service: str, instance_id: str | None = None) -> str

Workspace-scoped service name, e.g. zettelkasten-dashboard@a1b2c3d4.

This is what the health endpoint reports and what the launcher matches on, so a sibling workspace's dashboard reads as foreign instead of being adopted.

Source code in memory/dashboard/launcher_core.py 
108
109
110
111
112
113
114
def service_identity(base_service: str, instance_id: str | None = None) -> str:
    """Workspace-scoped service name, e.g. ``zettelkasten-dashboard@a1b2c3d4``.

    This is what the health endpoint reports and what the launcher matches on, so
    a sibling workspace's dashboard reads as *foreign* instead of being adopted.
    """
    return f"{base_service}@{instance_id or current_instance_id()}"

port_offset 

port_offset(instance_id: str, span: int = PORT_OFFSET_SPAN) -> int

Deterministic 0..span-1 port shift for a workspace instance id.

Source code in memory/dashboard/launcher_core.py 
117
118
119
def port_offset(instance_id: str, span: int = PORT_OFFSET_SPAN) -> int:
    """Deterministic 0..span-1 port shift for a workspace instance id."""
    return int(instance_id, 16) % max(1, span)

resolve_ports 

resolve_ports(config: LauncherConfig, instance_id: str) -> tuple[int, int]

Resolve (backend, frontend) ports: explicit env override else base+offset.

Source code in memory/dashboard/launcher_core.py 
132
133
134
135
136
137
138
139
140
141
def resolve_ports(config: LauncherConfig, instance_id: str) -> tuple[int, int]:
    """Resolve (backend, frontend) ports: explicit env override else base+offset."""
    offset = port_offset(instance_id)
    backend = _env_port(config.backend_port_env)
    if backend is None:
        backend = config.backend_port + offset
    frontend = _env_port(config.frontend_port_env)
    if frontend is None:
        frontend = config.frontend_port + offset
    return backend, frontend

resolved_ports 

resolved_ports(workspace: Path | str, base_backend: int, base_frontend: int, backend_port_env: str, frontend_port_env: str) -> tuple[int, int]

Resolve a workspace's (backend, frontend) ports without a LauncherConfig.

Same rule as :func:resolve_ports (explicit env override else a deterministic per-workspace offset), but callable from places that don't build a full :class:LauncherConfig — notably the launch_dashboard MCP tools and the angelo CLI. Sharing this is what keeps the python -m launcher, the MCP tool, and angelo kill agreeing on which port a given workspace's dashboard actually uses (a past divergence spawned duplicate backends).

Source code in memory/dashboard/launcher_core.py 
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
def resolved_ports(
    workspace: Path | str,
    base_backend: int,
    base_frontend: int,
    backend_port_env: str,
    frontend_port_env: str,
) -> tuple[int, int]:
    """Resolve a workspace's (backend, frontend) ports without a LauncherConfig.

    Same rule as :func:`resolve_ports` (explicit env override else a deterministic
    per-workspace offset), but callable from places that don't build a full
    :class:`LauncherConfig` — notably the ``launch_dashboard`` MCP tools and the
    ``angelo`` CLI. Sharing this is what keeps the ``python -m`` launcher, the MCP
    tool, and ``angelo kill`` agreeing on *which* port a given workspace's
    dashboard actually uses (a past divergence spawned duplicate backends).
    """
    instance_id = workspace_instance_id(workspace)
    offset = port_offset(instance_id)
    backend = _env_port(backend_port_env)
    if backend is None:
        backend = base_backend + offset
    frontend = _env_port(frontend_port_env)
    if frontend is None:
        frontend = base_frontend + offset
    return backend, frontend

package_version 

package_version() -> str

Installed angelo package version, or a sentinel when unresolvable.

Source code in memory/dashboard/launcher_core.py 
179
180
181
182
183
184
def package_version() -> str:
    """Installed ``angelo`` package version, or a sentinel when unresolvable."""
    try:
        return importlib.metadata.version("angelo")
    except Exception:
        return "0.0.0+unknown"

build_id 

build_id() -> str

Stable per-process identifier of the running code: <version>+<hash>.

Two processes from the same install report the same id; an upgrade (version bump) or any source edit (fingerprint change) yields a different id. Cached so the (cheap) fingerprint scan runs at most once per process.

Source code in memory/dashboard/launcher_core.py 
232
233
234
235
236
237
238
239
240
241
242
def build_id() -> str:
    """Stable per-process identifier of the running code: ``<version>+<hash>``.

    Two processes from the same install report the same id; an upgrade (version
    bump) or any source edit (fingerprint change) yields a different id. Cached so
    the (cheap) fingerprint scan runs at most once per process.
    """
    global _BUILD_ID_CACHE
    if _BUILD_ID_CACHE is None:
        _BUILD_ID_CACHE = f"{package_version()}+{_code_fingerprint()}"
    return _BUILD_ID_CACHE

ensure_node_deps 

ensure_node_deps(name: str, comp_dir: Path, npm: str | None) -> bool

Install a Node component's dependencies if missing. Returns readiness.

Source code in memory/dashboard/launcher_core.py 
262
263
264
265
266
267
268
269
270
271
272
273
274
def ensure_node_deps(name: str, comp_dir: Path, npm: str | None) -> bool:
    """Install a Node component's dependencies if missing. Returns readiness."""
    if (comp_dir / "node_modules").is_dir():
        return True
    if npm is None:
        print(f"[launcher] {name}: npm not found on PATH — skipping "
              "(install Node.js to enable it).", file=sys.stderr)
        return False
    print(f"[launcher] installing {name} dependencies (first run)...")
    if subprocess.run([npm, "install"], cwd=str(comp_dir)).returncode != 0:
        print(f"[launcher] {name}: `npm install` failed — skipping.", file=sys.stderr)
        return False
    return True

preflight 

preflight(config: LauncherConfig) -> dict[str, bool]

Decide which components can run for this install, prepping Node deps.

The backend always runs (and now hosts the chat agent in-process — no Node sidecar). Vite is a dev-only hot-reload layer that needs the frontend source plus a Node runtime. A plain pip install ships the built UI (frontend/dist) but not the frontend source — so installed users get a backend-served UI, while a source checkout gets the full live stack. Returns {"dev", "vite"} flags.

Source code in memory/dashboard/launcher_core.py 
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
def preflight(config: LauncherConfig) -> dict[str, bool]:
    """Decide which components can run for this install, prepping Node deps.

    The backend always runs (and now hosts the chat agent in-process — no Node
    sidecar). Vite is a dev-only hot-reload layer that needs the frontend
    *source* plus a Node runtime. A plain `pip install` ships the built UI
    (frontend/dist) but not the frontend source — so installed users get a
    backend-served UI, while a source checkout gets the full live stack.
    Returns {"dev", "vite"} flags.
    """
    frontend_dir = config.frontend_dir
    have_built_ui = (frontend_dir / "dist" / "index.html").is_file()
    have_frontend_src = (frontend_dir / "package.json").is_file()

    if not have_built_ui and not have_frontend_src:
        print(
            "[launcher] ERROR: no dashboard UI found — neither a built bundle at\n"
            f"    {frontend_dir / 'dist'}\n"
            f"  nor frontend source at {frontend_dir}.\n"
            "  Reinstall angelo; for a source checkout use:\n"
            f"    {config.install_hint}\n",
            file=sys.stderr,
        )
        sys.exit(1)

    npm = shutil.which("npm")

    run_vite = False
    if have_frontend_src:
        run_vite = ensure_node_deps("frontend", frontend_dir, npm)
        if not run_vite and not have_built_ui:
            print("[launcher] ERROR: frontend dependencies could not be installed "
                  "and no built bundle exists to serve.", file=sys.stderr)
            sys.exit(1)

    return {"dev": have_frontend_src, "vite": run_vite}

dashboard_url 

dashboard_url(port: int) -> str

Browser-facing dashboard URL. Uses 127.0.0.1 (not localhost) to match the address uvicorn actually binds — localhost may resolve to IPv6 ::1 first and incur a failed hop before falling back to the IPv4-only backend.

Source code in memory/dashboard/launcher_core.py 
335
336
337
338
339
def dashboard_url(port: int) -> str:
    """Browser-facing dashboard URL. Uses 127.0.0.1 (not ``localhost``) to match
    the address uvicorn actually binds — ``localhost`` may resolve to IPv6 ``::1``
    first and incur a failed hop before falling back to the IPv4-only backend."""
    return f"http://127.0.0.1:{port}"

http_ok 

http_ok(url: str, timeout: float = 1.0, expect_service: str | None = None) -> bool

GET url; True on HTTP 200 (and matching service, if requested).

Any connection/parse failure is False — a non-HTTP process squatting on the port, a different app, or a backend that hasn't finished booting all read as "not (our) healthy service".

Source code in memory/dashboard/launcher_core.py 
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
def http_ok(url: str, timeout: float = 1.0, expect_service: str | None = None) -> bool:
    """GET ``url``; True on HTTP 200 (and matching service, if requested).

    Any connection/parse failure is False — a non-HTTP process squatting on the
    port, a different app, or a backend that hasn't finished booting all read as
    "not (our) healthy service".
    """
    try:
        with urlopen(url, timeout=timeout) as resp:  # noqa: S310 (localhost only)
            if resp.status != 200:
                return False
            if expect_service is None:
                return True
            try:
                payload = json.loads(resp.read().decode("utf-8"))
            except (ValueError, UnicodeDecodeError):
                return False
            return isinstance(payload, dict) and payload.get("service") == expect_service
    except (URLError, OSError, ValueError):
        return False

fetch_health 

fetch_health(url: str, timeout: float = 1.0) -> dict | None

GET url and return the parsed health JSON, or None on any failure.

Unlike :func:http_ok (a bool), this surfaces the whole payload so callers can read the service/build fields and decide ours-vs-stale-vs-foreign. A non-200, non-JSON, or non-object response reads as None.

Source code in memory/dashboard/launcher_core.py 
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
def fetch_health(url: str, timeout: float = 1.0) -> dict | None:
    """GET ``url`` and return the parsed health JSON, or ``None`` on any failure.

    Unlike :func:`http_ok` (a bool), this surfaces the whole payload so callers
    can read the ``service``/``build`` fields and decide ours-vs-stale-vs-foreign.
    A non-200, non-JSON, or non-object response reads as ``None``.
    """
    try:
        with urlopen(url, timeout=timeout) as resp:  # noqa: S310 (localhost only)
            if resp.status != 200:
                return None
            payload = json.loads(resp.read().decode("utf-8"))
    except (URLError, OSError, ValueError, UnicodeDecodeError):
        return None
    return payload if isinstance(payload, dict) else None

classify_port 

classify_port(port: int, *, url: str | None = None, expect_service: str | None = None, expect_build: str | None = None, probe_in_use=None, probe_health=None, probe_payload=None) -> str

Return 'free', 'ours', 'stale', or 'foreign' for a component's port.

  • 'free' — nothing is listening.
  • 'ours' — a healthy backend whose identity (expect_service) matches and, when expect_build is given, whose reported build matches.
  • 'stale' — our backend by identity, but running a different build than expect_build (old code after an upgrade or source edit). Only returned when expect_build is provided.
  • 'foreign' — a busy port with no/failed health check, or a different service.

Components with no health endpoint (url is None) can only be 'free' or 'foreign'. When expect_build is omitted the legacy http_ok bool path is used, so existing callers/tests keep their two-way ours/foreign behaviour.

Probes default to the module-level helpers, resolved at call time so tests (and any future swap) can monkeypatch them.

Source code in memory/dashboard/launcher_core.py 
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
def classify_port(
    port: int,
    *,
    url: str | None = None,
    expect_service: str | None = None,
    expect_build: str | None = None,
    probe_in_use=None,
    probe_health=None,
    probe_payload=None,
) -> str:
    """Return 'free', 'ours', 'stale', or 'foreign' for a component's port.

    - 'free'    — nothing is listening.
    - 'ours'    — a healthy backend whose identity (``expect_service``) matches and,
                  when ``expect_build`` is given, whose reported ``build`` matches.
    - 'stale'   — our backend by identity, but running a *different* build than
                  ``expect_build`` (old code after an upgrade or source edit). Only
                  returned when ``expect_build`` is provided.
    - 'foreign' — a busy port with no/failed health check, or a different service.

    Components with no health endpoint (``url is None``) can only be 'free' or
    'foreign'. When ``expect_build`` is omitted the legacy ``http_ok`` bool path
    is used, so existing callers/tests keep their two-way ours/foreign behaviour.

    Probes default to the module-level helpers, resolved at call time so tests
    (and any future swap) can monkeypatch them.
    """
    probe_in_use = probe_in_use or port_in_use
    if not probe_in_use(port):
        return "free"
    if not url:
        return "foreign"

    if expect_build is not None:
        probe_payload = probe_payload or fetch_health
        payload = probe_payload(url)
        if not isinstance(payload, dict):
            return "foreign"
        if expect_service is not None and payload.get("service") != expect_service:
            return "foreign"
        if payload.get("build") != expect_build:
            return "stale"
        return "ours"

    probe_health = probe_health or http_ok
    if probe_health(url, expect_service=expect_service):
        return "ours"
    return "foreign"

kill_listeners 

kill_listeners(port: int) -> int

SIGTERM every process LISTENing on a TCP port. Returns the count signalled.

Cross-platform (netstat on Windows, lsof elsewhere) and best-effort: any enumeration failure or unkillable pid is swallowed so callers can treat a 0 return as "nothing recycled" rather than an error.

Source code in memory/dashboard/launcher_core.py 
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
def kill_listeners(port: int) -> int:
    """SIGTERM every process LISTENing on a TCP ``port``. Returns the count signalled.

    Cross-platform (netstat on Windows, lsof elsewhere) and best-effort: any
    enumeration failure or unkillable pid is swallowed so callers can treat a 0
    return as "nothing recycled" rather than an error.
    """
    pids: set[int] = set()
    try:
        if sys.platform == "win32":
            out = subprocess.run(
                ["netstat", "-ano", "-p", "TCP"],
                capture_output=True, text=True, timeout=5,
            ).stdout
            for line in out.splitlines():
                parts = line.split()
                if (len(parts) >= 5 and parts[3] == "LISTENING"
                        and parts[1].endswith(f":{port}")):
                    try:
                        pids.add(int(parts[4]))
                    except ValueError:
                        pass
        else:
            out = subprocess.run(
                ["lsof", "-ti", f"tcp:{port}", "-sTCP:LISTEN"],
                capture_output=True, text=True, timeout=5,
            ).stdout
            for tok in out.split():
                try:
                    pids.add(int(tok))
                except ValueError:
                    pass
    except (OSError, subprocess.SubprocessError):
        return 0

    killed = 0
    for pid in pids:
        try:
            os.kill(pid, signal.SIGTERM)
            killed += 1
        except OSError:
            pass
    return killed

wait_for_free 

wait_for_free(port: int, timeout: float = 5.0, interval: float = 0.25, *, probe=None, sleep=sleep, clock=monotonic) -> bool

Poll until port is no longer listening (e.g. after a recycle kill).

Source code in memory/dashboard/launcher_core.py 
476
477
478
479
480
481
482
483
484
485
486
def wait_for_free(port: int, timeout: float = 5.0, interval: float = 0.25,
                  *, probe=None, sleep=time.sleep, clock=time.monotonic) -> bool:
    """Poll until ``port`` is no longer listening (e.g. after a recycle kill)."""
    probe = probe or port_in_use
    deadline = clock() + timeout
    while True:
        if not probe(port):
            return True
        if clock() >= deadline:
            return False
        sleep(interval)

wait_for_health 

wait_for_health(url: str, timeout: float = HEALTH_WAIT_SECONDS, interval: float = 0.25, expect_service: str | None = None, *, probe=None, sleep=sleep, clock=monotonic) -> bool

Poll url until it answers healthy or timeout elapses.

Source code in memory/dashboard/launcher_core.py 
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
def wait_for_health(
    url: str,
    timeout: float = HEALTH_WAIT_SECONDS,
    interval: float = 0.25,
    expect_service: str | None = None,
    *,
    probe=None,
    sleep=time.sleep,
    clock=time.monotonic,
) -> bool:
    """Poll ``url`` until it answers healthy or ``timeout`` elapses."""
    probe = probe or http_ok
    deadline = clock() + timeout
    while True:
        if probe(url, expect_service=expect_service):
            return True
        if clock() >= deadline:
            return False
        sleep(interval)

backoff_delay 

backoff_delay(attempt: int, base: float = RESTART_BASE_SECONDS, cap: float = RESTART_CAP_SECONDS) -> float

Exponential backoff for restart attempt (1-based), capped.

Source code in memory/dashboard/launcher_core.py 
513
514
515
516
517
518
def backoff_delay(attempt: int, base: float = RESTART_BASE_SECONDS,
                  cap: float = RESTART_CAP_SECONDS) -> float:
    """Exponential backoff for restart ``attempt`` (1-based), capped."""
    if attempt < 1:
        attempt = 1
    return min(cap, base * (2 ** (attempt - 1)))

next_restart 

next_restart(attempts: int, uptime: float, *, max_restarts: int = MAX_RESTARTS, stable_reset: float = STABLE_RESET_SECONDS, base: float = RESTART_BASE_SECONDS, cap: float = RESTART_CAP_SECONDS) -> tuple[str, float, int]

Decide what to do when a component exits.

Returns (action, delay_seconds, new_attempts) where action is "restart" or "giveup". A component that had been up at least stable_reset seconds has its failure counter reset first, so an occasional late crash self-heals rather than counting toward the flap limit.

Source code in memory/dashboard/launcher_core.py 
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
def next_restart(
    attempts: int,
    uptime: float,
    *,
    max_restarts: int = MAX_RESTARTS,
    stable_reset: float = STABLE_RESET_SECONDS,
    base: float = RESTART_BASE_SECONDS,
    cap: float = RESTART_CAP_SECONDS,
) -> tuple[str, float, int]:
    """Decide what to do when a component exits.

    Returns ``(action, delay_seconds, new_attempts)`` where action is "restart"
    or "giveup". A component that had been up at least ``stable_reset`` seconds
    has its failure counter reset first, so an occasional late crash self-heals
    rather than counting toward the flap limit.
    """
    effective = 0 if uptime >= stable_reset else attempts
    if effective >= max_restarts:
        return ("giveup", 0.0, effective)
    return ("restart", backoff_delay(effective + 1, base, cap), effective + 1)

run 

run(config: LauncherConfig) -> None

Preflight, start everything, wait for health, then supervise.

Source code in memory/dashboard/launcher_core.py 
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
def run(config: LauncherConfig) -> None:
    """Preflight, start everything, wait for health, then supervise."""
    # Windows consoles default to cp1252, which can't render Vite's arrows etc.
    if hasattr(sys.stdout, "reconfigure"):
        sys.stdout.reconfigure(encoding="utf-8", errors="replace")

    # Hand the resolved project root to every child (notably the backend, whose
    # in-process chat agent reads the Cursor API key and graph relative to it;
    # otherwise a pip install would resolve them under site-packages).
    os.environ["ANGELO_WORKSPACE"] = str(config.workspace)

    # Per-workspace identity + ports. Exporting these into our own environment
    # propagates them to every child: the backend reports the workspace-scoped
    # service id from its health endpoint, and the Vite child reads the port envs
    # (see frontend/vite.config.ts) so its dev server and proxy match the backend
    # this launcher actually started.
    instance_id = workspace_instance_id(config.workspace)
    backend_port, frontend_port = resolve_ports(config, instance_id)
    identity = service_identity(config.service_name, instance_id)
    os.environ[INSTANCE_ENV] = instance_id
    os.environ[config.backend_port_env] = str(backend_port)
    os.environ[config.frontend_port_env] = str(frontend_port)
    print(f"[launcher] workspace {config.workspace} -> instance {instance_id} "
          f"(backend :{backend_port}, frontend :{frontend_port})")

    caps = preflight(config)
    components = _build_components(
        config, caps,
        backend_port=backend_port,
        frontend_port=frontend_port,
        expect_service=identity,
    )

    # In a source checkout the live Vite server is the UI; otherwise the backend
    # serves the bundled build, so that's where users should point their browser.
    ui_port = frontend_port if caps["vite"] else backend_port

    supervised = [comp for comp in components if _start_component(comp)]

    if not supervised:
        print("[launcher] everything already running, nothing to do")
        print(f"[launcher] dashboard: {dashboard_url(ui_port)}")
        return

    # Wait for the backend to actually answer before advertising the URL.
    backend = next((c for c in supervised if c.name == "backend"), None)
    if backend is not None and backend.url is not None:
        if wait_for_health(backend.url, expect_service=backend.expect_service):
            print("[launcher] backend healthy")
        else:
            print(f"[launcher] WARNING: backend did not answer {backend.url} within "
                  f"{HEALTH_WAIT_SECONDS:.0f}s — it may still be starting; check the "
                  "[backend] logs above.", file=sys.stderr)

    print(f"[launcher] dashboard: {dashboard_url(ui_port)}  (Ctrl+C to stop)")

    stop = threading.Event()
    try:
        _supervise(supervised, stop)
    except KeyboardInterrupt:
        stop.set()
        _shutdown(supervised)