Skip to main content

subc_daemon/
bootstrap.rs

1use std::{
2    collections::BTreeMap,
3    env,
4    error::Error,
5    ffi::OsString,
6    fmt, fs, io,
7    net::{IpAddr, Ipv4Addr, Ipv6Addr, SocketAddr},
8    path::{Path, PathBuf},
9    process,
10    time::Duration,
11};
12
13use fs4::{FileExt, TryLockError};
14use subc_protocol::PROTOCOL_VERSION;
15pub use subc_transport::user_connection_token;
16use subc_transport::{
17    authenticate_client, connection_file, generate_daemon_id, generate_key,
18    write_atomic_with_local_endpoint, AuthError, ConnectionFileError, ConnectionInfo, Endpoint,
19    LocalEndpoint, SCHEMA_VERSION,
20};
21use tokio::{
22    net::{TcpListener, TcpStream},
23    task::{JoinError, JoinHandle},
24    time::{sleep, timeout},
25};
26use tracing::{error, info, warn};
27
28use crate::{
29    daemon_config::{self, ConfiguredModule, DaemonConfigError},
30    local_listener::{self, BoundLocalListener, LocalListenerConfig},
31    server::{serve_listeners, serve_local_listener, ServerAuth, ServerError},
32    supervise::HealthConfig,
33    ConnectedClients, ControlHandler, DaemonSelfWatchdog, DaemonSelfWatchdogConfig,
34    ForwardingTable, Registry, RestartPolicy, Router, Supervisor, SupervisorHandle,
35    SupervisorProcessLiveness,
36};
37use std::sync::Arc;
38
39#[cfg(windows)]
40mod windows_connection_copy;
41
42pub const DEFAULT_SUBC_PORT: u16 = 8757;
43pub const SUBC_PORT_ENV: &str = "SUBC_PORT";
44use subc_transport::CONNECTION_FILE_NAME;
45const DAEMON_VERSION: &str = env!("CARGO_PKG_VERSION");
46const CONNECT_TIMEOUT: Duration = Duration::from_secs(2);
47const PROBE_AUTH_DEADLINE: Duration = Duration::from_secs(2);
48const START_LOCK_RETRIES: usize = 40;
49const START_LOCK_RETRY_DELAY: Duration = Duration::from_millis(25);
50
51#[derive(Debug, Clone, Copy, PartialEq, Eq)]
52pub enum ConnectionFileSource {
53    XdgRuntimeDir,
54    TempDirFallback,
55    Explicit,
56}
57
58impl ConnectionFileSource {
59    fn reason(self) -> &'static str {
60        match self {
61            Self::XdgRuntimeDir => "XDG_RUNTIME_DIR set and non-empty",
62            Self::TempDirFallback => "XDG_RUNTIME_DIR unset or empty",
63            Self::Explicit => "configured path",
64        }
65    }
66}
67
68impl fmt::Display for ConnectionFileSource {
69    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
70        formatter.write_str(match self {
71            Self::XdgRuntimeDir => "xdg_runtime_dir",
72            Self::TempDirFallback => "temp_dir_fallback",
73            Self::Explicit => "explicit",
74        })
75    }
76}
77
78/// Runtime bootstrap configuration. Production uses the default fixed port and
79/// optional daemon-config override; tests pass port 0 to let the OS assign a free
80/// loopback port and discover it from the connection file.
81#[derive(Debug, Clone, Default)]
82struct AdmissionFactsConfig {
83    carrier_module_id: Option<String>,
84    targets: Option<Vec<String>>,
85}
86
87/// Controls where module cgroups are prepared.
88///
89/// In-process daemons default to [`Self::Disabled`] so they never derive a
90/// production location from the host process. The shipped daemon explicitly uses
91/// [`Self::Current`]. The `ck-subc` binary accepts
92/// `SUBC_CGROUP_PLACEMENT=disabled` for isolated test processes; an unset
93/// variable retains `Current`, and every other value is rejected at startup.
94/// Tests that exercise placement can own a [`Self::Root`].
95#[derive(Debug, Clone, Default, PartialEq, Eq)]
96pub enum CgroupPlacementConfig {
97    #[default]
98    Disabled,
99    Current,
100    Root(PathBuf),
101}
102
103#[derive(Debug, Clone)]
104pub struct BootstrapConfig {
105    #[cfg(feature = "test-support")]
106    shutdown_intent_window: Duration,
107    pub operator_queue_wait: Duration,
108    pub operator_timeout: Duration,
109    pub operator_stuck_grace: Duration,
110    #[cfg(feature = "test-support")]
111    pub operator_script: Option<PathBuf>,
112    #[cfg(feature = "test-support")]
113    pub operator_events: Option<PathBuf>,
114    pub connection_file_path: PathBuf,
115    pub port: u16,
116    pub daemon_ver: String,
117    configured_modules: Vec<ConfiguredModule>,
118    storage_config: Option<daemon_config::StorageConfig>,
119    admission_facts: AdmissionFactsConfig,
120    /// Module ids whose scopes may carry the attributes that grant authority.
121    scope_authority_owners: Vec<String>,
122    federated_scope_owner: String,
123    privileged_operator_prompt: daemon_config::PrivilegedOperatorPrompt,
124    activity_config: daemon_config::ActivityConfig,
125    daemon_config_path: Option<PathBuf>,
126    /// Only a service-definition argument may name this authority store. An
127    /// embedded daemon with no explicit root never opens an installation store.
128    config_store_root: Option<PathBuf>,
129    defer_file_load: bool,
130    // Test-only stand-in for an additional field supplied by the file loader.
131    // Authority selection must carry it without knowing its name.
132    #[cfg(test)]
133    file_derived_test_probe: Option<Vec<String>>,
134    #[cfg(all(test, feature = "test-support"))]
135    test_store_factory:
136        Option<fn(crate::config_store::StoreRoot) -> crate::config_store::ConfigStore>,
137    #[cfg(all(test, feature = "test-support"))]
138    test_control: Option<fn(ControlHandler) -> ControlHandler>,
139    #[cfg(all(test, feature = "test-support"))]
140    test_listener_bind: Option<fn(&mut BootstrapConfig)>,
141    configured_port: Option<u16>,
142    /// Daemon-wide route.bind relay budget in milliseconds (the fallback for
143    /// any module without a per-module override). `None` = built-in default
144    /// (12s — see `control::DEFAULT_ROUTE_BIND_RELAY_TIMEOUT`).
145    route_bind_relay_default_ms: Option<u64>,
146    reserved_capabilities: BTreeMap<String, String>,
147    watchdog_config: DaemonSelfWatchdogConfig,
148    connection_file_source: ConnectionFileSource,
149    #[cfg(all(test, windows))]
150    profile_override: Option<PathBuf>,
151    /// Where module cgroups are prepared. Disabled unless a caller explicitly
152    /// opts in, because `Current` derives a host location from `/proc/self/cgroup`.
153    cgroup_placement: CgroupPlacementConfig,
154    /// Directory the supervisor writes per-module stdout/stderr capture files
155    /// into. `None` disables capture; the shipped binary supplies its real run
156    /// directory explicitly.
157    capture_logs_dir: Option<PathBuf>,
158    /// File the supervisor appends every module's terminal exits to, so exit
159    /// history survives a daemon restart. `None` keeps terminal history in
160    /// memory only (each module's ring). Absent by default for the same reason
161    /// as `capture_logs_dir`: an in-process daemon booted by a test must not
162    /// append its exits to the operator's real `terminals.jsonl`, where they
163    /// would show up in `ck module terminals`. The shipped binary supplies
164    /// `<run dir>/terminals.jsonl` explicitly.
165    terminal_journal_path: Option<PathBuf>,
166    /// Where the machine id is read from, or minted into when absent. `None`
167    /// serves no machine id. Absent by default for the same reason as
168    /// `capture_logs_dir`: an in-process daemon booted by a test must never
169    /// derive the operator's real data home and mint into it. The shipped binary
170    /// supplies `<data home>/cortexkit/machine-id` explicitly.
171    machine_id_path: Option<PathBuf>,
172
173    /// Where this daemon's same-user local listener lives (a Unix-domain socket
174    /// on macOS and Linux; nothing on Windows, where the pipe name is minted per
175    /// start). The default serves no local endpoint, so an in-process daemon
176    /// never touches the operator's run directory. The shipped binary names its
177    /// run directory; [`BootstrapConfig::with_local_listener_socket_dir`] names
178    /// any other, and an in-process daemon that wants the endpoint names one.
179    local_listener: LocalListenerConfig,
180    /// The live-children record: every supervised process this daemon has
181    /// running, kept so the next daemon can end the ones a crash left behind.
182    /// At startup, before any module is spawned, the previous daemon's record
183    /// here is swept. `None` keeps no record and sweeps nothing, for the same
184    /// reason as `capture_logs_dir`: an in-process daemon booted by a test
185    /// must never signal processes listed in the operator's real run
186    /// directory. The shipped binary supplies `<run dir>/live-children.json`.
187    live_children_path: Option<PathBuf>,
188    privacy_trampoline: Option<PathBuf>,
189}
190
191impl BootstrapConfig {
192    pub fn with_config_store_root(mut self, root: PathBuf) -> Result<Self, BootstrapError> {
193        crate::config_store::StoreRoot::from_service_definition(root.clone())
194            .map_err(|e| BootstrapError::ConfigStore(e.to_string()))?;
195        let parent = root.parent().ok_or_else(|| {
196            BootstrapError::ConfigStore("config store root needs a parent directory".into())
197        })?;
198        std::fs::read_dir(parent).map_err(|e| {
199            BootstrapError::ConfigStore(format!("config store parent is unreadable: {e}"))
200        })?;
201        self.config_store_root = Some(root);
202        Ok(self)
203    }
204    /// Shorten restart intent validity in daemon integration fixtures.
205    #[cfg(feature = "test-support")]
206    pub fn with_shutdown_intent_window(mut self, window: Duration) -> Self {
207        self.shutdown_intent_window = window;
208        self
209    }
210
211    pub fn new(connection_file_path: impl Into<PathBuf>, port: u16) -> Self {
212        Self {
213            #[cfg(feature = "test-support")]
214            shutdown_intent_window: Duration::from_secs(120),
215            connection_file_path: connection_file_path.into(),
216            operator_queue_wait: Duration::from_secs(150),
217            operator_timeout: Duration::from_secs(120),
218            operator_stuck_grace: Duration::from_secs(10),
219            #[cfg(feature = "test-support")]
220            operator_script: None,
221            #[cfg(feature = "test-support")]
222            operator_events: None,
223            port,
224            daemon_ver: DAEMON_VERSION.to_owned(),
225            configured_modules: Vec::new(),
226            storage_config: None,
227            admission_facts: AdmissionFactsConfig::default(),
228            scope_authority_owners: daemon_config::default_scope_authority_owners(),
229            federated_scope_owner: daemon_config::default_federated_scope_owner(),
230            privileged_operator_prompt: daemon_config::PrivilegedOperatorPrompt::default(),
231            activity_config: daemon_config::ActivityConfig::default(),
232            daemon_config_path: None,
233            config_store_root: None,
234            defer_file_load: false,
235            #[cfg(test)]
236            file_derived_test_probe: None,
237            #[cfg(all(test, feature = "test-support"))]
238            test_store_factory: None,
239            #[cfg(all(test, feature = "test-support"))]
240            test_control: None,
241            #[cfg(all(test, feature = "test-support"))]
242            test_listener_bind: None,
243            configured_port: None,
244            route_bind_relay_default_ms: None,
245            reserved_capabilities: BTreeMap::new(),
246            watchdog_config: DaemonSelfWatchdogConfig::default(),
247            connection_file_source: ConnectionFileSource::Explicit,
248            #[cfg(all(test, windows))]
249            profile_override: None,
250            cgroup_placement: CgroupPlacementConfig::default(),
251            capture_logs_dir: None,
252            terminal_journal_path: None,
253            machine_id_path: None,
254            local_listener: LocalListenerConfig { socket_dir: None },
255            live_children_path: None,
256            privacy_trampoline: None,
257        }
258    }
259
260    /// Keep the local listener's socket in `dir`. Embedding daemons and tests
261    /// pass a path inside their own fixture tree.
262    pub fn with_local_listener_socket_dir(mut self, dir: impl Into<PathBuf>) -> Self {
263        self.local_listener = LocalListenerConfig::in_socket_dir(dir);
264        self
265    }
266
267    /// Set where the local listener lives, as [`LocalListenerConfig`] describes.
268    pub fn with_local_listener(mut self, config: LocalListenerConfig) -> Self {
269        self.local_listener = config;
270        self
271    }
272
273    /// Keep the live-children record at `path`, and at startup end the
274    /// processes a previous daemon recorded there that are still running.
275    /// Embedding daemons and tests pass a path inside their own fixture tree.
276    pub fn with_live_children_record(mut self, path: impl Into<PathBuf>) -> Self {
277        self.live_children_path = Some(path.into());
278        self
279    }
280
281    /// macOS module launches require an explicit executable implementing the
282    /// subc-os trampoline entry point. Other systems ignore this setting.
283    pub fn with_privacy_trampoline(mut self, path: impl Into<PathBuf>) -> Self {
284        self.privacy_trampoline = Some(path.into());
285        self
286    }
287
288    /// Serve the machine id stored at `path`, minting it there at startup when
289    /// the file is absent. Embedding daemons and tests pass a path inside their
290    /// own fixture tree.
291    pub fn with_machine_id_path(mut self, path: impl Into<PathBuf>) -> Self {
292        self.machine_id_path = Some(path.into());
293        self
294    }
295
296    /// Selects module cgroup placement. The default is disabled.
297    pub fn with_cgroup_placement(mut self, placement: CgroupPlacementConfig) -> Self {
298        self.cgroup_placement = placement;
299        self
300    }
301
302    /// Redirects per-module stdout/stderr capture files out of the real run
303    /// directory. Tests that start an in-process daemon must call this with a
304    /// path inside their fixture tree.
305    pub fn with_capture_logs_dir(mut self, dir: impl Into<PathBuf>) -> Self {
306        self.capture_logs_dir = Some(dir.into());
307        self
308    }
309
310    /// Redirects the daemon-private journal, allowing embedded daemons and tests
311    /// to keep their observations out of the operator's live run directory.
312    pub fn with_terminal_journal_path(mut self, path: impl Into<PathBuf>) -> Self {
313        self.terminal_journal_path = Some(path.into());
314        self
315    }
316
317    pub fn from_env() -> Result<Self, BootstrapError> {
318        Self::from_env_with_daemon_config_path(daemon_config::default_config_path())
319    }
320
321    /// The SHIPPED BINARY's config, which is the only caller that should capture
322    /// child output into the operator's real run directory.
323    ///
324    /// Kept separate from `from_env` deliberately: see `capture_logs_dir` and
325    /// `terminal_journal_path` on this struct for why an absent value must mean
326    /// NO CAPTURE and NO JOURNAL rather than the operator's real run directory.
327    /// A run directory that cannot be resolved (a relative data home) refuses
328    /// startup instead of landing under the working directory.
329    pub fn from_env_for_daemon_binary() -> Result<Self, BootstrapError> {
330        let run_dir = daemon_config::daemon_run_dir().map_err(BootstrapError::RunDir)?;
331        let machine_id_path =
332            crate::machine_id::default_machine_id_path().map_err(BootstrapError::MachineId)?;
333        // The service supplies its store root after constructing this value.
334        // Do not parse live files until that store has selected recovery mode.
335        let (connection_file, source) =
336            connection_file_path_with_source(non_empty_os_var("XDG_RUNTIME_DIR"));
337        let mut config = Self::new(connection_file, DEFAULT_SUBC_PORT)
338            .with_connection_file_source(source)
339            .with_daemon_config_source(daemon_config::default_config_path(), None)
340            .with_local_listener(LocalListenerConfig::in_socket_dir(&run_dir))
341            .with_capture_logs_dir(run_dir.join("logs"))
342            .with_terminal_journal_path(run_dir.join("terminals.jsonl"))
343            .with_live_children_record(crate::live_children::record_path(&run_dir))
344            .with_machine_id_path(machine_id_path);
345        config.defer_file_load = true;
346        #[cfg(feature = "test-support")]
347        let config = {
348            let mut config = config;
349            config.operator_script =
350                std::env::var_os("SUBC_TEST_OPERATOR_SCRIPT").map(PathBuf::from);
351            config.operator_events =
352                std::env::var_os("SUBC_TEST_OPERATOR_EVENTS").map(PathBuf::from);
353            for (name, value) in [
354                (
355                    "SUBC_TEST_OPERATOR_QUEUE_WAIT_MS",
356                    &mut config.operator_queue_wait,
357                ),
358                (
359                    "SUBC_TEST_OPERATOR_TIMEOUT_MS",
360                    &mut config.operator_timeout,
361                ),
362                (
363                    "SUBC_TEST_OPERATOR_STUCK_MS",
364                    &mut config.operator_stuck_grace,
365                ),
366            ] {
367                if let Ok(ms) = std::env::var(name) {
368                    if let Ok(ms) = ms.parse::<u64>() {
369                        *value = Duration::from_millis(ms);
370                    }
371                }
372            }
373            config
374        };
375        Ok(config)
376    }
377
378    /// Load live-file values for an unstamped embedder. If a store root is
379    /// attached later, these values are provisional and replaced at run time.
380    /// An invalid live file can still fail this eager constructor; the binary
381    /// defers loading until store authority has been selected instead.
382    pub fn from_env_with_daemon_config_path(
383        daemon_config_path: impl AsRef<Path>,
384    ) -> Result<Self, BootstrapError> {
385        let daemon_config_path = daemon_config_path.as_ref().to_path_buf();
386        let (daemon_config, activity_config) =
387            daemon_config::load_with_activity(&daemon_config_path)
388                .map_err(BootstrapError::DaemonConfig)?
389                .map(|(config, activity)| (Some(config), activity))
390                .unwrap_or_default();
391        let config_port = daemon_config.as_ref().and_then(|config| config.port);
392        let storage_config = daemon_config
393            .as_ref()
394            .and_then(|config| config.storage.clone());
395        let admission_facts_carrier_module_id = daemon_config
396            .as_ref()
397            .and_then(|config| config.admission_facts_carrier_module_id.clone());
398        let admission_facts_targets = daemon_config
399            .as_ref()
400            .and_then(|config| config.admission_facts_targets.clone());
401        let scope_authority_owners = daemon_config
402            .as_ref()
403            .map(|config| config.scope_authority_owners.clone())
404            .unwrap_or_else(daemon_config::default_scope_authority_owners);
405        let federated_scope_owner = daemon_config
406            .as_ref()
407            .map(|config| config.federated_scope_owner.clone())
408            .unwrap_or_else(daemon_config::default_federated_scope_owner);
409        let privileged_operator_prompt = daemon_config
410            .as_ref()
411            .map(|config| config.privileged_operator_prompt)
412            .unwrap_or_default();
413        let route_bind_relay_default_ms = daemon_config
414            .as_ref()
415            .and_then(|config| config.route_bind_relay_timeout_ms);
416        let reserved_capabilities = daemon_config
417            .as_ref()
418            .map(|config| config.reserved_capabilities.clone())
419            .unwrap_or_default();
420        let configured_modules = daemon_config
421            .map(|config| config.modules)
422            .unwrap_or_default();
423
424        let port = match env::var(SUBC_PORT_ENV) {
425            Ok(raw) if !raw.trim().is_empty() => {
426                let port = raw
427                    .parse::<u16>()
428                    .map_err(|source| BootstrapError::InvalidPort { raw, source })?;
429                if let Some(config_port) = config_port {
430                    info!(
431                        env = SUBC_PORT_ENV,
432                        env_port = port,
433                        config_port,
434                        "SUBC_PORT overrides daemon config port"
435                    );
436                }
437                port
438            }
439            Ok(_) | Err(_) => config_port.unwrap_or(DEFAULT_SUBC_PORT),
440        };
441
442        let (connection_file_path, connection_file_source) =
443            connection_file_path_with_source(non_empty_os_var("XDG_RUNTIME_DIR"));
444        let config = Self::new(connection_file_path, port)
445            .with_configured_modules(configured_modules)
446            .with_storage_config(storage_config)
447            .with_admission_facts_config(admission_facts_carrier_module_id, admission_facts_targets)
448            .with_scope_authority_owners(scope_authority_owners)
449            .with_federated_scope_owner(federated_scope_owner)
450            .with_privileged_operator_prompt(privileged_operator_prompt)
451            .with_activity_config(activity_config)
452            .with_route_bind_relay_default_ms(route_bind_relay_default_ms)
453            .with_reserved_capabilities(reserved_capabilities)
454            .with_daemon_config_source(daemon_config_path, config_port)
455            .with_connection_file_source(connection_file_source);
456        #[cfg(test)]
457        let config = {
458            let mut config = config;
459            config.file_derived_test_probe = Some(config.scope_authority_owners.clone());
460            config
461        };
462        Ok(config)
463    }
464
465    /// Load live-file values unless a store root is already attached. Attaching
466    /// a root later makes eager values provisional: run time replaces them with
467    /// approved authority. An invalid live file can fail the eager constructor;
468    /// the binary's deferred path does not parse it before selecting authority.
469    pub fn with_daemon_config_path(
470        mut self,
471        daemon_config_path: impl AsRef<Path>,
472    ) -> Result<Self, BootstrapError> {
473        let daemon_config_path = daemon_config_path.as_ref().to_path_buf();
474        if self.config_store_root.is_some() {
475            self.defer_file_load = true;
476            return Ok(self.with_daemon_config_source(daemon_config_path, None));
477        }
478        let (daemon_config, activity_config) =
479            daemon_config::load_with_activity(&daemon_config_path)
480                .map_err(BootstrapError::DaemonConfig)?
481                .map(|(config, activity)| (Some(config), activity))
482                .unwrap_or_default();
483        let configured_port = daemon_config.as_ref().and_then(|config| config.port);
484        Ok(self
485            .with_parsed_daemon_config(daemon_config, activity_config)
486            .with_daemon_config_source(daemon_config_path, configured_port))
487    }
488
489    fn with_parsed_daemon_config(
490        self,
491        daemon_config: Option<daemon_config::DaemonConfig>,
492        activity_config: daemon_config::ActivityConfig,
493    ) -> Self {
494        let storage_config = daemon_config
495            .as_ref()
496            .and_then(|config| config.storage.clone());
497        let admission_facts_carrier_module_id = daemon_config
498            .as_ref()
499            .and_then(|config| config.admission_facts_carrier_module_id.clone());
500        let admission_facts_targets = daemon_config
501            .as_ref()
502            .and_then(|config| config.admission_facts_targets.clone());
503        let scope_authority_owners = daemon_config
504            .as_ref()
505            .map(|config| config.scope_authority_owners.clone())
506            .unwrap_or_else(daemon_config::default_scope_authority_owners);
507        let federated_scope_owner = daemon_config
508            .as_ref()
509            .map(|config| config.federated_scope_owner.clone())
510            .unwrap_or_else(daemon_config::default_federated_scope_owner);
511        let privileged_operator_prompt = daemon_config
512            .as_ref()
513            .map(|config| config.privileged_operator_prompt)
514            .unwrap_or_default();
515        let route_bind_relay_default_ms = daemon_config
516            .as_ref()
517            .and_then(|config| config.route_bind_relay_timeout_ms);
518        let reserved_capabilities = daemon_config
519            .as_ref()
520            .map(|config| config.reserved_capabilities.clone())
521            .unwrap_or_default();
522        let configured_modules = daemon_config
523            .map(|config| config.modules)
524            .unwrap_or_default();
525        self.with_configured_modules(configured_modules)
526            .with_storage_config(storage_config)
527            .with_admission_facts_config(admission_facts_carrier_module_id, admission_facts_targets)
528            .with_scope_authority_owners(scope_authority_owners)
529            .with_federated_scope_owner(federated_scope_owner)
530            .with_privileged_operator_prompt(privileged_operator_prompt)
531            .with_activity_config(activity_config)
532            .with_route_bind_relay_default_ms(route_bind_relay_default_ms)
533            .with_reserved_capabilities(reserved_capabilities)
534    }
535
536    pub fn with_configured_modules(
537        mut self,
538        modules: impl IntoIterator<Item = ConfiguredModule>,
539    ) -> Self {
540        self.configured_modules = modules.into_iter().collect();
541        self.configured_modules
542            .sort_by(|left, right| left.module_id.cmp(&right.module_id));
543        self
544    }
545
546    pub fn with_storage_config(
547        mut self,
548        storage_config: Option<daemon_config::StorageConfig>,
549    ) -> Self {
550        self.storage_config = storage_config;
551        self
552    }
553
554    pub fn with_admission_facts_config(
555        mut self,
556        carrier_module_id: Option<String>,
557        targets: Option<Vec<String>>,
558    ) -> Self {
559        self.admission_facts = AdmissionFactsConfig {
560            carrier_module_id,
561            targets,
562        };
563        self
564    }
565
566    /// Set the module ids whose scopes may carry `agent_id` and `delegates`.
567    /// Absent from a config, this is `daemon_config::default_scope_authority_owners`.
568    pub fn with_scope_authority_owners(mut self, owners: Vec<String>) -> Self {
569        self.scope_authority_owners = owners;
570        self
571    }
572
573    pub fn with_federated_scope_owner(mut self, owner: String) -> Self {
574        self.federated_scope_owner = owner;
575        self
576    }
577
578    /// Choose the privileged-operation policy for the daemon's entire lifetime.
579    pub fn with_privileged_operator_prompt(
580        mut self,
581        mode: daemon_config::PrivilegedOperatorPrompt,
582    ) -> Self {
583        self.privileged_operator_prompt = mode;
584        self
585    }
586
587    pub fn with_activity_config(mut self, config: daemon_config::ActivityConfig) -> Self {
588        self.activity_config = config;
589        self
590    }
591
592    /// Set the daemon-wide route.bind relay default (the fallback for any
593    /// module without a per-module override). `None` preserves the built-in
594    /// default (12s). `serve_bound_daemon` reads this at startup and threads
595    /// it into the control handler's daemon-wide field.
596    pub fn with_route_bind_relay_default_ms(mut self, ms: Option<u64>) -> Self {
597        self.route_bind_relay_default_ms = ms;
598        self
599    }
600
601    pub fn with_reserved_capabilities(
602        mut self,
603        reserved_capabilities: BTreeMap<String, String>,
604    ) -> Self {
605        self.reserved_capabilities = reserved_capabilities;
606        self
607    }
608
609    fn with_daemon_config_source(
610        mut self,
611        daemon_config_path: PathBuf,
612        configured_port: Option<u16>,
613    ) -> Self {
614        self.daemon_config_path = Some(daemon_config_path);
615        self.configured_port = configured_port;
616        self
617    }
618
619    fn with_connection_file_source(mut self, source: ConnectionFileSource) -> Self {
620        self.connection_file_source = source;
621        self
622    }
623
624    pub fn with_watchdog_config(mut self, watchdog_config: DaemonSelfWatchdogConfig) -> Self {
625        self.watchdog_config = watchdog_config;
626        self
627    }
628}
629
630/// Result of singleton discovery.
631// Built once per daemon start and immediately matched, so the size gap between
632// the variants costs nothing worth a box on the public shape.
633#[allow(clippy::large_enum_variant)]
634#[derive(Debug)]
635pub enum Outcome {
636    /// A live daemon authenticated from the connection file; this invocation should exit 0.
637    AlreadyRunning,
638    /// This process won the singleton race, owns bound loopback listener(s), and
639    /// has published a fresh connection file.
640    Bound(BoundDaemon),
641}
642
643#[derive(Debug)]
644pub struct BoundDaemon {
645    pub listeners: Vec<TcpListener>,
646    /// The same-user local listener, bound and not yet served. `None` when
647    /// binding it failed, in which case `local_endpoint` is `None` and
648    /// `local_endpoint_unavailable` says why.
649    pub local_listener: Option<BoundLocalListener>,
650    /// Why the daemon serves no local endpoint, when binding it failed. Reported
651    /// by `server.describe` so an operator can see it.
652    pub local_endpoint_unavailable: Option<String>,
653    /// What the connection file publishes under `local_endpoint`, so a client can
654    /// find the local listener without knowing how it was bound.
655    pub local_endpoint: Option<LocalEndpoint>,
656    pub connection_info: ConnectionInfo,
657    pub connection_file_path: PathBuf,
658    pub connection_file_source: ConnectionFileSource,
659    #[cfg(windows)]
660    connection_file_copy: Option<PathBuf>,
661    /// The machine id this daemon serves, established before any connection is
662    /// accepted. `None` when the config named no machine id path.
663    pub machine_id: Option<crate::machine_id::MachineId>,
664    /// Ownership of the run directory, held until the daemon exits so no other
665    /// daemon sweeps or writes this one's run state. `None` when the config
666    /// keeps no live-children record, and so has no run directory to own.
667    run_dir_lock: Option<crate::run_dir_lock::RunDirLock>,
668    /// The singleton start lock, kept when binding deferred publishing the
669    /// connection file. It stays held until the daemon has verified it is serving
670    /// and written the connection file, so a second daemon starting meanwhile
671    /// cannot claim the singleton too. `None` when the connection
672    /// file was already written at bind time.
673    publication_lock: Option<StartLock>,
674}
675
676/// Resolve subc's per-user TCP connection-file path.
677///
678/// `$XDG_RUNTIME_DIR/subc-connection.json` is preferred because the runtime
679/// directory is already per-user on Unix desktops. Without it, subc falls back
680/// to the system temp dir with a per-user token in the filename so different OS
681/// users do not collide on shared temp directories.
682pub fn connection_file_path() -> PathBuf {
683    connection_file_path_with_source(non_empty_os_var("XDG_RUNTIME_DIR")).0
684}
685
686fn connection_file_path_with_source(
687    runtime_dir: Option<OsString>,
688) -> (PathBuf, ConnectionFileSource) {
689    if let Some(runtime_dir) = runtime_dir.filter(|value| !value.is_empty()) {
690        return (
691            PathBuf::from(runtime_dir).join(CONNECTION_FILE_NAME),
692            ConnectionFileSource::XdgRuntimeDir,
693        );
694    }
695
696    (
697        env::temp_dir().join(format!("subc-{}.connection.json", user_connection_token())),
698        ConnectionFileSource::TempDirFallback,
699    )
700}
701
702/// Resolve, claim, and serve the per-user daemon singleton.
703///
704/// A second invocation is successful: if a live daemon authenticates from the
705/// existing connection file, this returns `Ok(())` after logging and the caller
706/// exits with status 0.
707pub async fn run() -> Result<(), BootstrapError> {
708    // `run` is a binary entry point, so it opts into the production cgroup and
709    // child-log locations. In-process callers keep both features disabled by default.
710    run_with_config(
711        BootstrapConfig::from_env_for_daemon_binary()?
712            .with_cgroup_placement(CgroupPlacementConfig::Current),
713    )
714    .await
715}
716
717#[cfg(feature = "test-support")]
718mod config_test_seam {
719    #[cfg(not(test))]
720    use super::BootstrapError;
721
722    pub(super) fn enabled() -> bool {
723        std::env::var_os("SUBC_TEST_CONFIG_GATE").as_deref() == Some(std::ffi::OsStr::new("1"))
724            && std::env::current_exe().ok().is_some_and(|path| {
725                path.file_name()
726                    .is_some_and(|name| name.to_string_lossy().starts_with("ckdev-"))
727            })
728    }
729
730    #[cfg(not(test))]
731    pub(super) fn open(
732        root: crate::config_store::StoreRoot,
733    ) -> Result<crate::config_store::ConfigStore, BootstrapError> {
734        if !enabled() {
735            return Ok(crate::config_store::ConfigStore::open(root));
736        }
737        // The service's isolated store root contains every seam artifact, so a
738        // test cannot accidentally open the operator's keychain or anchor.
739        crate::config_store::testing::TestSupport::new(
740            root.path().to_path_buf(),
741            Some(root.path().join("test-key")),
742        )
743        .and_then(|hooks| {
744            hooks
745                .with_anchor_failure_file(root.path().join("fail-anchor"))
746                .open()
747        })
748        .map_err(|error| BootstrapError::ConfigStore(error.to_string()))
749    }
750}
751
752/// Serve a daemon from an explicit config. This is the entry point the twelve
753/// sibling repos use to boot an in-process daemon in their integration tests.
754///
755/// # THIS INSTALLS NO TRACING SUBSCRIBER, SO THE DAEMON IS SILENT BY DEFAULT
756///
757/// The daemon's own diagnostics go through `tracing`, and `tracing` DISCARDS
758/// every event when no subscriber is installed. The shipped binary installs one
759/// in `main` (`init_tracing`); this function deliberately does not, because a
760/// library that installs a global subscriber fights with whatever the host
761/// process already set up.
762///
763/// The consequence for a test harness is not "less verbose": it is that the
764/// daemon has NOTHING TO SAY about any failure, and a missing instrument reads
765/// exactly like a clean one. PLEX found this on 2026-09-18 while trying to
766/// capture daemon logs beside an intermittent bind failure, and discovered the
767/// daemon had been silent in every conformance run that repo had ever done --
768/// so the one client-side error string was all the evidence that could exist,
769/// and they had spent a real investigation on a failure whose second source was
770/// never being recorded.
771///
772/// Install one in the harness before calling this, and assert it did something
773/// (they measured 236 daemon lines with the subscriber installed, 0 with the
774/// call commented out) -- otherwise the fix is itself unverified.
775pub async fn run_with_config(config: BootstrapConfig) -> Result<(), BootstrapError> {
776    let Some(start_lock) = singleton_start_lock(&config.connection_file_path).await? else {
777        info!("subc daemon already running");
778        return Ok(());
779    };
780    let (config, opened_store, start_lock) = tokio::task::spawn_blocking(move || {
781        let mut opened_store = config
782            .config_store_root
783            .as_ref()
784            .map(|root| {
785                let (root, lock) = crate::config_store::StoreWriterLock::acquire(root)
786                    .map_err(|e| BootstrapError::ConfigStore(e.to_string()))?;
787                #[cfg(all(test, feature = "test-support"))]
788                let store = if let Some(factory) = config.test_store_factory {
789                    crate::config_store::testing::TestSupport::new(root.path().to_path_buf(), None)
790                        .map_err(|e| BootstrapError::ConfigStore(e.to_string()))?;
791                    factory(root)
792                } else {
793                    crate::config_store::ConfigStore::open(root)
794                };
795                #[cfg(all(feature = "test-support", not(test)))]
796                let store = config_test_seam::open(root)?;
797                #[cfg(not(feature = "test-support"))]
798                let store = crate::config_store::ConfigStore::open(root);
799                Ok::<_, BootstrapError>((store, lock))
800            })
801            .transpose()?;
802        let config =
803            load_bootstrap_authority(config, opened_store.as_mut().map(|(store, _)| store))?;
804        #[cfg(all(test, feature = "test-support"))]
805        let config = {
806            let mut config = config;
807            if let Some(bind) = config.test_listener_bind {
808                // Tests observe authority's port selection but change only the
809                // bind, so they never take a running daemon's production port.
810                bind(&mut config);
811            }
812            config
813        };
814        Ok::<_, BootstrapError>((config, opened_store, start_lock))
815    })
816    .await
817    .map_err(|e| BootstrapError::ConfigStore(e.to_string()))??;
818    #[cfg(feature = "test-support")]
819    let shutdown_intent_window = config.shutdown_intent_window;
820    let forwarding = Arc::new(ForwardingTable::default());
821    forwarding.operator_confirms().configure_bootstrap(&config);
822    let configured_modules = config.configured_modules.clone();
823    let storage_config = config.storage_config.clone();
824    let admission_facts = config.admission_facts.clone();
825    let scope_authority_owners = config.scope_authority_owners.clone();
826    let federated_scope_owner = config.federated_scope_owner.clone();
827    let privileged_operator_prompt = config.privileged_operator_prompt;
828    let activity_config = config.activity_config;
829    let daemon_config_path = config.daemon_config_path.clone();
830    let configured_port = config.configured_port;
831    let route_bind_relay_default_ms = config.route_bind_relay_default_ms;
832    let reserved_capabilities = config.reserved_capabilities.clone();
833    let watchdog_config = config.watchdog_config.clone();
834    let cgroup_placement_config = config.cgroup_placement.clone();
835    let capture_logs_dir = config.capture_logs_dir.clone();
836    let terminal_journal_path = config.terminal_journal_path.clone();
837    let live_children_path = config.live_children_path.clone();
838    let privacy_trampoline = config.privacy_trampoline.clone();
839    #[cfg(all(test, feature = "test-support"))]
840    let test_control = config.test_control;
841    match ensure_singleton_with_start_lock(config, false, start_lock).await? {
842        Outcome::AlreadyRunning => {
843            info!("subc daemon already running");
844            Ok(())
845        }
846        Outcome::Bound(bound) => {
847            #[cfg(target_os = "linux")]
848            let cgroup_placement = prepare_cgroup_placement(&cgroup_placement_config);
849            #[cfg(not(target_os = "linux"))]
850            let _ = cgroup_placement_config;
851            serve_bound_daemon(
852                bound,
853                forwarding,
854                configured_modules,
855                storage_config,
856                admission_facts,
857                scope_authority_owners,
858                federated_scope_owner,
859                privileged_operator_prompt,
860                activity_config,
861                daemon_config_path,
862                opened_store,
863                configured_port,
864                route_bind_relay_default_ms,
865                reserved_capabilities,
866                watchdog_config,
867                capture_logs_dir,
868                terminal_journal_path,
869                live_children_path,
870                privacy_trampoline,
871                #[cfg(all(test, feature = "test-support"))]
872                test_control,
873                #[cfg(feature = "test-support")]
874                shutdown_intent_window,
875                #[cfg(target_os = "linux")]
876                cgroup_placement,
877            )
878            .await
879        }
880    }
881}
882
883// Keep authority selection before binding listeners or constructing control
884// authorization. Recovery must remain reachable even with invalid live files.
885fn load_bootstrap_authority(
886    config: BootstrapConfig,
887    mut store: Option<&mut crate::config_store::ConfigStore>,
888) -> Result<BootstrapConfig, BootstrapError> {
889    let store_state = store.as_mut().map(|store| store.status().state);
890    let recovery = matches!(
891        store_state,
892        Some(
893            crate::config_store::StoreState::StoreUnverified
894                | crate::config_store::StoreState::StoreMissing
895        )
896    );
897    let authority = if recovery {
898        BootstrapConfig::new(&config.connection_file_path, DEFAULT_SUBC_PORT)
899    } else if store_state == Some(crate::config_store::StoreState::Stamped) {
900        let store = store.expect("stamped store was just checked");
901        match stamped_authority(&config.connection_file_path, store) {
902            Ok(authority) => authority,
903            Err(reason) => {
904                // A stamped daemon never falls back to files, and exiting would
905                // leave the operator no way to reach recovery.
906                tracing::error!(%reason, "approved configuration unusable; starting in recovery");
907                store.enter_recovery(&reason);
908                BootstrapConfig::new(&config.connection_file_path, DEFAULT_SUBC_PORT)
909            }
910        }
911    } else if config.defer_file_load {
912        BootstrapConfig::from_env_with_daemon_config_path(
913            config
914                .daemon_config_path
915                .as_ref()
916                .expect("binary config path"),
917        )?
918    } else {
919        return Ok(config);
920    };
921    // Only installation paths and explicit binary arguments override authority.
922    // New file-derived fields are inherited whole from the selected authority.
923    Ok(BootstrapConfig {
924        connection_file_path: config.connection_file_path,
925        connection_file_source: config.connection_file_source,
926        daemon_ver: config.daemon_ver,
927        daemon_config_path: config.daemon_config_path,
928        config_store_root: config.config_store_root,
929        defer_file_load: config.defer_file_load,
930        operator_queue_wait: config.operator_queue_wait,
931        operator_timeout: config.operator_timeout,
932        operator_stuck_grace: config.operator_stuck_grace,
933        watchdog_config: config.watchdog_config,
934        cgroup_placement: config.cgroup_placement,
935        capture_logs_dir: config.capture_logs_dir,
936        terminal_journal_path: config.terminal_journal_path,
937        machine_id_path: config.machine_id_path,
938        local_listener: config.local_listener,
939        live_children_path: config.live_children_path,
940        privacy_trampoline: config.privacy_trampoline,
941        #[cfg(feature = "test-support")]
942        shutdown_intent_window: config.shutdown_intent_window,
943        #[cfg(feature = "test-support")]
944        operator_script: config.operator_script,
945        #[cfg(feature = "test-support")]
946        operator_events: config.operator_events,
947        #[cfg(all(test, windows))]
948        profile_override: config.profile_override,
949        #[cfg(all(test, feature = "test-support"))]
950        test_store_factory: config.test_store_factory,
951        #[cfg(all(test, feature = "test-support"))]
952        test_control: config.test_control,
953        #[cfg(all(test, feature = "test-support"))]
954        test_listener_bind: config.test_listener_bind,
955        ..authority
956    })
957}
958
959/// Authority from the approved daemon generation the store selects for this
960/// start. Every failure is a reason to enter recovery, never to exit.
961fn stamped_authority(
962    connection_file_path: &Path,
963    store: &mut crate::config_store::ConfigStore,
964) -> Result<BootstrapConfig, String> {
965    let generation = store
966        .select_boot_generation()
967        .map_err(|error| error.to_string())?;
968    let body = store.body().map_err(|error| error.to_string())?;
969    let layer = body
970        .generations
971        .get(&generation)
972        .and_then(|entry| entry.delivery.as_ref())
973        .and_then(|delivery| delivery.daemon.as_ref())
974        .ok_or("approved daemon snapshot is absent")?;
975    let bytes = layer
976        .bytes
977        .as_deref()
978        .ok_or("approved daemon bytes are absent")?;
979    let path = layer
980        .path
981        .as_deref()
982        .ok_or("approved daemon path is absent")?;
983    let doc = std::str::from_utf8(bytes).map_err(|error| error.to_string())?;
984    let (daemon, activity) = daemon_config::parse_with_activity(doc, Path::new(path))
985        .map_err(|error| format!("approved daemon configuration is invalid: {error}"))?;
986    let port = daemon.port;
987    Ok(
988        BootstrapConfig::new(connection_file_path, port.unwrap_or(DEFAULT_SUBC_PORT))
989            .with_parsed_daemon_config(Some(daemon), activity)
990            .with_daemon_config_source(PathBuf::from(path), port),
991    )
992}
993
994#[cfg(target_os = "linux")]
995fn prepare_cgroup_placement(config: &CgroupPlacementConfig) -> Option<subc_cgroup::Placement> {
996    let result = match config {
997        CgroupPlacementConfig::Disabled => return None,
998        CgroupPlacementConfig::Current => subc_cgroup::prepare_current(),
999        CgroupPlacementConfig::Root(root) => subc_cgroup::prepare_at(root),
1000    };
1001
1002    match result {
1003        Ok(Some(placement)) => Some(placement),
1004        Ok(None) => {
1005            warn!(
1006                placement = ?config,
1007                "module cgroup placement is disabled: configured cgroup root is not delegated"
1008            );
1009            None
1010        }
1011        Err(error) => {
1012            warn!(
1013                placement = ?config,
1014                error = %error,
1015                "module cgroup placement is disabled by an unexpected cgroup probe error"
1016            );
1017            None
1018        }
1019    }
1020}
1021
1022/// Target soft limit for open file descriptors, applied to the daemon before any
1023/// module is spawned so children inherit it. Multi-root modules (one process
1024/// aggregating every project root's sqlite stores, index caches, watchers, and
1025/// LSP pipes) trivially exceed the macOS default soft limit of 256; a launchd
1026/// user agent does not pass login-shell ulimits through, so the raise must
1027/// happen in-process.
1028#[cfg(unix)]
1029const NOFILE_TARGET: u64 = 65536;
1030
1031/// Raise RLIMIT_NOFILE to `NOFILE_TARGET` (clamped to the hard limit).
1032/// Best-effort: failure is logged and never fatal, since the daemon can run
1033/// under the inherited limit — modules with few roots just have less headroom.
1034#[cfg(unix)]
1035fn raise_nofile_limit() {
1036    #[cfg(target_os = "macos")]
1037    let ceiling = subc_os::child_spawn::output(
1038        std::process::Command::new("/usr/sbin/sysctl").args(["-n", "kern.maxfilesperproc"]),
1039    )
1040    .ok()
1041    .filter(|output| output.status.success())
1042    .and_then(|output| String::from_utf8(output.stdout).ok())
1043    .and_then(|value| value.trim().parse::<u64>().ok())
1044    .filter(|value| *value > 0);
1045    #[cfg(not(target_os = "macos"))]
1046    let ceiling = None;
1047    raise_nofile_limit_with_ceiling(ceiling);
1048}
1049
1050#[cfg(unix)]
1051fn raise_nofile_limit_with_ceiling(kernel_ceiling: Option<u64>) {
1052    match rlimit::Resource::NOFILE.get() {
1053        Ok((soft, hard)) => {
1054            // On XNU an infinite hard limit does not remove the separate
1055            // per-process kernel ceiling. Exceeding it makes setrlimit fail
1056            // without raising the inherited (often 256) soft limit at all.
1057            let target = NOFILE_TARGET
1058                .min(hard)
1059                .min(kernel_ceiling.unwrap_or(u64::MAX));
1060            if soft >= target {
1061                return;
1062            }
1063            match rlimit::Resource::NOFILE.set(target, hard) {
1064                Ok(()) => info!(
1065                    previous_soft = soft,
1066                    new_soft = target,
1067                    hard,
1068                    "raised open-file soft limit for daemon and module children"
1069                ),
1070                Err(err) => warn!(
1071                    soft,
1072                    hard,
1073                    error = %err,
1074                    "could not raise open-file soft limit; multi-root modules may exhaust descriptors"
1075                ),
1076            }
1077        }
1078        Err(err) => warn!(error = %err, "could not read open-file limit"),
1079    }
1080}
1081
1082/// CRT stdio-stream target on Windows (the `_setmaxstdio` maximum). Win32
1083/// HANDLEs — what Rust `File`, tokio sockets, and SQLite's Win32 VFS actually
1084/// consume — have a per-process quota in the millions and need no raise; the
1085/// C-runtime stream table (default 512) is the only low ceiling, and it is
1086/// per-process rather than inherited, so supervised modules linking the CRT
1087/// must raise their own. Raising it here covers the daemon itself.
1088#[cfg(windows)]
1089fn raise_nofile_limit() {
1090    const MAXSTDIO_TARGET: u32 = 8192;
1091    let current = rlimit::getmaxstdio();
1092    if current >= MAXSTDIO_TARGET {
1093        return;
1094    }
1095    match rlimit::setmaxstdio(MAXSTDIO_TARGET) {
1096        Ok(new_max) => info!(
1097            previous = current,
1098            new_max, "raised CRT stdio-stream limit for daemon"
1099        ),
1100        Err(err) => warn!(
1101            current,
1102            error = %err,
1103            "could not raise CRT stdio-stream limit"
1104        ),
1105    }
1106}
1107
1108#[cfg(not(any(unix, windows)))]
1109fn raise_nofile_limit() {}
1110
1111pub async fn run_with_daemon_config_path(
1112    config: BootstrapConfig,
1113    daemon_config_path: impl AsRef<Path>,
1114) -> Result<(), BootstrapError> {
1115    run_with_config(config.with_daemon_config_path(daemon_config_path)?).await
1116}
1117
1118#[allow(clippy::too_many_arguments)]
1119async fn serve_bound_daemon(
1120    bound: BoundDaemon,
1121    forwarding: Arc<ForwardingTable>,
1122    configured_modules: Vec<ConfiguredModule>,
1123    storage_config: Option<daemon_config::StorageConfig>,
1124    admission_facts: AdmissionFactsConfig,
1125    scope_authority_owners: Vec<String>,
1126    federated_scope_owner: String,
1127    privileged_operator_prompt: daemon_config::PrivilegedOperatorPrompt,
1128    activity_config: daemon_config::ActivityConfig,
1129    daemon_config_path: Option<PathBuf>,
1130    opened_store: Option<(
1131        crate::config_store::ConfigStore,
1132        crate::config_store::StoreWriterLock,
1133    )>,
1134    configured_port: Option<u16>,
1135    route_bind_relay_default_ms: Option<u64>,
1136    reserved_capabilities: BTreeMap<String, String>,
1137    watchdog_config: DaemonSelfWatchdogConfig,
1138    capture_logs_dir: Option<PathBuf>,
1139    terminal_journal_path: Option<PathBuf>,
1140    live_children_path: Option<PathBuf>,
1141    privacy_trampoline: Option<PathBuf>,
1142    #[cfg(all(test, feature = "test-support"))] test_control: Option<
1143        fn(ControlHandler) -> ControlHandler,
1144    >,
1145    #[cfg(feature = "test-support")] shutdown_intent_window: Duration,
1146    #[cfg(target_os = "linux")] cgroup_placement: Option<subc_cgroup::Placement>,
1147) -> Result<(), BootstrapError> {
1148    #[cfg(unix)]
1149    let mut terminate = tokio::signal::unix::signal(tokio::signal::unix::SignalKind::terminate())
1150        .map_err(BootstrapError::Signal)?;
1151    // Windows has no SIGTERM. Ctrl-C is a different event, not an equivalent
1152    // service-stop contract, so this shutdown handler is intentionally Unix-only.
1153    raise_nofile_limit();
1154    #[cfg(windows)]
1155    if let Err(error) = subc_jobobject::ensure_root_job() {
1156        // Keep the control plane available. The root-job API caches setup
1157        // failures, so later module spawns refuse instead of running unprotected.
1158        warn!(%error, "module launches disabled: Windows root job setup failed");
1159    }
1160
1161    info!(
1162        connection_file = %bound.connection_file_path.display(),
1163        connection_file_source = %bound.connection_file_source,
1164        connection_file_source_reason = bound.connection_file_source.reason(),
1165        endpoints = ?bound.connection_info.endpoints,
1166        configured_modules = configured_modules.len(),
1167        machine_id = bound.machine_id.as_ref().map(|id| id.as_str()).unwrap_or("none"),
1168        "subc daemon starting"
1169    );
1170
1171    // Before anything can spawn a module (the configured modules below, or a
1172    // client's start request once the listeners are served): a previous daemon
1173    // that died without its shutdown stop may have left children running, and
1174    // a fresh copy beside one would fight it for its port and stores. The
1175    // record is not a running daemon's because this daemon holds the lock on
1176    // the run directory the record lives in, and a running daemon holds that
1177    // lock for its whole life. (Claiming the singleton is not enough: it is
1178    // keyed on the connection file, which can live in a different runtime
1179    // directory from the run directory.) The lock stays held until this
1180    // function returns, which is when the daemon exits.
1181    let run_dir_lock = bound.run_dir_lock;
1182    let publication_lock = bound.publication_lock;
1183    if let Some(owner) = &run_dir_lock {
1184        #[cfg(target_os = "linux")]
1185        crate::live_children::sweep_orphans_with_placement(
1186            owner,
1187            &crate::live_children::AdoptedPids::none(),
1188            crate::live_children::SweepBounds::default(),
1189            cgroup_placement.as_ref(),
1190        )
1191        .await;
1192        #[cfg(not(target_os = "linux"))]
1193        crate::live_children::sweep_orphans(
1194            owner,
1195            &crate::live_children::AdoptedPids::none(),
1196            crate::live_children::SweepBounds::default(),
1197        )
1198        .await;
1199    }
1200
1201    let registry = Arc::new(Registry::default());
1202    let process_liveness = Arc::new(SupervisorProcessLiveness::new());
1203    let supervisor_handle = SupervisorHandle::new();
1204    let _host_health_task = AbortOnDrop::new(supervisor_handle.spawn_host_health_monitor());
1205    #[cfg(feature = "test-support")]
1206    supervisor_handle.set_shutdown_intent_window(shutdown_intent_window);
1207    let connected_clients = ConnectedClients::new();
1208    let daemon_incarnation = format!(
1209        "{:032x}",
1210        u128::from_be_bytes(bound.connection_info.daemon_id)
1211    );
1212    let config_runtime = if let Some((store, writer_lock)) = opened_store {
1213        let path = daemon_config_path.clone().ok_or_else(|| {
1214            BootstrapError::ConfigStore(
1215                "an explicit daemon config path is required with a config store".into(),
1216            )
1217        })?;
1218        let config_home = daemon_config::default_config_home();
1219        let incarnation = daemon_incarnation.clone();
1220        Some(Arc::new(
1221            tokio::task::spawn_blocking(move || {
1222                crate::config_gate::runtime::RuntimeOwner::new(
1223                    incarnation,
1224                    store,
1225                    crate::config_gate::capture::ConfigCapturer {
1226                        daemon_path: path,
1227                        config_home,
1228                        projects: Arc::new(crate::config_gate::capture::UnavailableProjectRegistry),
1229                    },
1230                )
1231                .retain_writer_lock(writer_lock)
1232                .map_err(io::Error::other)
1233            })
1234            .await
1235            .map_err(|e| BootstrapError::ConfigStore(e.to_string()))?
1236            .map_err(|e| BootstrapError::ConfigStore(e.to_string()))?,
1237        ))
1238    } else {
1239        None
1240    };
1241    let _config_plan_sweep = config_runtime.as_ref().map(|runtime| {
1242        let runtime = Arc::downgrade(runtime);
1243        AbortOnDrop::new(tokio::spawn(async move {
1244            let mut interval =
1245                tokio::time::interval(crate::config_gate::runtime::PLAN_SWEEP_INTERVAL);
1246            loop {
1247                interval.tick().await;
1248                let Some(runtime) = runtime.upgrade() else {
1249                    break;
1250                };
1251                match tokio::task::spawn_blocking(move || runtime.sweep_expired()).await {
1252                    Ok(Ok(())) => {}
1253                    result => tracing::warn!(?result, "configuration plan expiry sweep failed"),
1254                }
1255            }
1256        }))
1257    });
1258    let supervisor = Supervisor::new(Arc::clone(&registry), RestartPolicy::default())
1259        .with_process_liveness(process_liveness.clone())
1260        .with_forwarding(Arc::clone(&forwarding))
1261        .with_handle(supervisor_handle.clone())
1262        .with_connection_file_path(bound.connection_file_path.clone())
1263        .with_daemon_incarnation(daemon_incarnation.clone());
1264    let supervisor = match &config_runtime {
1265        Some(owner) => supervisor.with_config_runtime(owner.clone()),
1266        None => supervisor,
1267    };
1268    let supervisor = match privacy_trampoline {
1269        Some(path) => supervisor.with_privacy_trampoline(path),
1270        None => supervisor,
1271    };
1272    // ABSENT MEANS NO CAPTURE AND NO JOURNAL, NOT "THE REAL RUN DIRECTORY", and
1273    // the difference is a production-corruption hazard rather than a preference.
1274    // Both fields below follow the same rule: `None` for `capture_logs_dir`
1275    // means supervised output is not captured, and `None` for
1276    // `terminal_journal_path` means terminal history lives only in each
1277    // module's in-memory ring (`supervisor.terminals` still answers, with the
1278    // journal counters at zero).
1279    //
1280    // The terminal journal used to fall back to
1281    // `daemon_run_dir().join("terminals.jsonl")`, so an in-process test daemon
1282    // appended its fixture exits to the operator's journal, where they then
1283    // appeared in `ck module terminals`.
1284    //
1285    // The capture line used to be `unwrap_or_else(|| daemon_run_dir().join("logs"))`,
1286    // so ANY caller that did not set the field captured supervised children into
1287    // the operator's live `~/.local/share/cortexkit/run/logs/`. That is twelve
1288    // sibling repos whose integration tests boot an in-process daemon through
1289    // `run_with_config` -- none of which asked for it, and none of which can see
1290    // it from their side.
1291    //
1292    // Harmless while fixture module ids are fixture-shaped: this host carries 20
1293    // zero-byte files from subc's own tests (good-aft, missing-aft,
1294    // preview-consumer...). THE HAZARD IS A COLLISION. A fixture named "broca"
1295    // or "aft" appends to a PRODUCTION capture file that operators read
1296    // forensically and that placement gates count lines in -- with no residue to
1297    // notice, because the file legitimately exists and legitimately grows.
1298    //
1299    // Found by BROCA (2026-09-19) from the other side: their rigs spawn the
1300    // SHIPPED ck-subc and set XDG_CONFIG_HOME + XDG_RUNTIME_DIR but not
1301    // XDG_DATA_HOME, so every local rig run supervised a module named "broca"
1302    // and captured it into production's broca.stderr.log -- the same file I
1303    // count seal lines in before and after placing their binaries.
1304    //
1305    // The supervisor already treats `None` as no-capture and no-journal, so
1306    // this only removes invented defaults. The binary keeps capturing and
1307    // journaling via `BootstrapConfig::from_env_for_daemon_binary`.
1308
1309    // Keep both ledgers in the explicitly selected run directory. An embedded
1310    // daemon without a journal path must not write to the operator's live ledger.
1311    let activity_run_dir = terminal_journal_path
1312        .as_ref()
1313        .and_then(|path| path.parent())
1314        .map(Path::to_path_buf);
1315    let supervisor = match terminal_journal_path {
1316        Some(path) => supervisor.with_terminal_journal(path, daemon_incarnation.clone()),
1317        None => supervisor,
1318    };
1319    let supervisor = match capture_logs_dir {
1320        Some(dir) => supervisor.with_capture_logs_dir(dir),
1321        None => supervisor,
1322    };
1323    let supervisor = match live_children_path {
1324        Some(path) => supervisor.with_live_children_record(path),
1325        None => supervisor,
1326    };
1327    #[cfg(target_os = "linux")]
1328    let supervisor = supervisor.with_cgroup_placement(cgroup_placement);
1329    // Collect per-module route.bind relay overrides BEFORE handing the
1330    // `configured_modules` vector to the supervisor (which only needs each
1331    // module's `drain_timeout_ms`). Each entry was filled in by parse-time
1332    // resolution (per-module > daemon-wide > absent), so modules with no
1333    // override are absent from this map and the daemon-wide default applies.
1334    let route_bind_relay_timeouts = configured_modules
1335        .iter()
1336        .filter_map(|module| {
1337            module
1338                .route_bind_relay_timeout_ms
1339                .map(|ms| (module.module_id.clone(), Duration::from_millis(ms)))
1340        })
1341        .collect::<std::collections::BTreeMap<_, _>>();
1342    let control_start_clock = crate::clock::StartClock::capture();
1343    let mut control = ControlHandler::with_forwarding(Arc::clone(&registry), forwarding)
1344        .with_daemon_version(bound.connection_info.daemon_ver.clone())
1345        .with_process_liveness(process_liveness)
1346        .with_supervisor(supervisor_handle)
1347        // The same flag `begin_daemon_shutdown` sets below, so requirement
1348        // changes while modules exit during the shutdown log at INFO.
1349        .with_daemon_shutdown_flag(supervisor.daemon_shutdown_flag())
1350        .with_connected_clients(connected_clients.clone())
1351        .with_local_endpoint_unavailable(bound.local_endpoint_unavailable.clone())
1352        .with_storage_config(storage_config)
1353        .with_machine_id(bound.machine_id.clone())
1354        .with_admission_facts_config(admission_facts.carrier_module_id, admission_facts.targets)
1355        .with_scope_authority_owners(scope_authority_owners)
1356        .with_federated_scope_owner(federated_scope_owner)
1357        .with_privileged_operator_prompt(privileged_operator_prompt)
1358        .with_route_bind_relay_timeouts(route_bind_relay_timeouts)
1359        .with_daemon_provenance(
1360            bound.connection_info.pid,
1361            control_start_clock.started_at_ms(),
1362            std::env::current_exe().ok(),
1363            normalized_build_provenance(env!("SUBC_BUILD_GIT_SHA")),
1364            normalized_build_provenance(env!("SUBC_BUILD_LOCK_DIGEST")),
1365        )
1366        .with_daemon_start_clock(control_start_clock)
1367        .with_capability_config(
1368            configured_modules
1369                .iter()
1370                .map(|module| (module.module_id.clone(), module.enabled)),
1371            reserved_capabilities,
1372        );
1373    if let Some(run_dir) = activity_run_dir {
1374        control =
1375            control.with_activity_ledger_config(&run_dir, daemon_incarnation, activity_config);
1376    }
1377    let daemon_ready_runtime = config_runtime.clone();
1378    if let Some(runtime) = config_runtime {
1379        control = control.with_explicit_config_store(runtime);
1380    }
1381    #[cfg(feature = "test-support")]
1382    if config_test_seam::enabled() {
1383        control = control.with_config_os_prompt_policy_for_tests();
1384    }
1385    if let Some(ms) = route_bind_relay_default_ms {
1386        // A daemon-wide config value overrides the built-in default; a
1387        // `None` here leaves the ControlHandler's 12s default in place.
1388        control = control.with_route_bind_relay_timeout(Duration::from_millis(ms));
1389    }
1390    if let Some(config_path) = daemon_config_path {
1391        control = control.with_supervisor_rescan(supervisor.clone(), config_path, configured_port);
1392    }
1393    #[cfg(all(test, feature = "test-support"))]
1394    let control = if let Some(adapt) = test_control {
1395        // A test hook is not a substitute for the executable-name guard.
1396        let executable =
1397            std::env::current_exe().map_err(|e| BootstrapError::ConfigStore(e.to_string()))?;
1398        if !executable
1399            .file_name()
1400            .is_some_and(|name| name.to_string_lossy().starts_with("ckdev-"))
1401        {
1402            return Err(BootstrapError::ConfigStore(
1403                "test control requires a ckdev executable".into(),
1404            ));
1405        }
1406        adapt(control)
1407    } else {
1408        control
1409    };
1410    let control = Arc::new(control);
1411    let router = Arc::new(Router::with_control_handler(Arc::clone(&control)));
1412    let auth = ServerAuth::new(
1413        bound.connection_info.key.clone(),
1414        bound.connection_info.daemon_id,
1415        bound.connection_info.daemon_ver.clone(),
1416    )
1417    .with_connected_clients(connected_clients);
1418
1419    // Discovery must expose every accepted configured module, including disabled
1420    // ones. Registration creates status handles but neither starts monitors
1421    // nor launches processes: children need the published file for `--subc`.
1422    let mut registered_modules = Vec::with_capacity(configured_modules.len());
1423    for configured in configured_modules {
1424        let enabled = configured.enabled;
1425        let health = configured.health.clone();
1426        let module_id = configured.module_id.clone();
1427        let registered = match supervisor.register_configured_with_health(
1428            configured.module_spec(),
1429            enabled,
1430            health.clone(),
1431            configured.drain_timeout_ms,
1432            configured.restart,
1433            configured.register_timeout,
1434        ) {
1435            Ok(registered) => registered,
1436            Err(err) => {
1437                error!(module_id = %module_id, error = %err, "failed to supervise configured module; continuing daemon startup");
1438                continue;
1439            }
1440        };
1441        registered_modules.push(registered);
1442        // A relaxed threshold can silently outlive the temporary allowance
1443        // that required it, so report the extra tolerance on every boot.
1444        let default_threshold = HealthConfig::default().failure_threshold;
1445        if enabled && health.failure_threshold > default_threshold {
1446            warn!(
1447                module_id = %module_id,
1448                failure_threshold = health.failure_threshold,
1449                default_threshold,
1450                tolerance_secs = health.cadence.as_secs() * u64::from(health.failure_threshold),
1451                "health failure threshold is relaxed above the default; a wedged module stays unflagged for longer"
1452            );
1453        }
1454        info!(module_id = %module_id, enabled, "configured module registered");
1455    }
1456    control.refresh_capability_requirements();
1457
1458    // The local endpoint is served beside TCP and answers the same handshake.
1459    // It serves clients only, which the server enforces per connection.
1460    let local_serve_task = bound.local_listener.map(|local_listener| {
1461        AbortOnDrop::new(tokio::spawn(serve_local_listener(
1462            local_listener,
1463            Arc::clone(&router),
1464            auth.clone(),
1465        )))
1466    });
1467    // On Unix the task is dropped below to stop accepting at shutdown. Windows
1468    // leaves it to the process exit; naming it keeps the binding uniform.
1469    #[cfg(windows)]
1470    let _local_serve_task = local_serve_task;
1471    let mut serve_task =
1472        AbortOnDrop::new(tokio::spawn(serve_listeners(bound.listeners, router, auth)));
1473    verify_serving(&bound.connection_info).await?;
1474    // A published endpoint promises that authentication can complete now.
1475    // Orphan cleanup and control-plane construction can take longer than an
1476    // existing-daemon probe's deadline, so finish both before discovery.
1477    // The write fsyncs, so it runs on tokio's blocking pool (see
1478    // `publish_connection_file`). The singleton lock travels with it and comes
1479    // back only once the file is in place.
1480    let (published, publication_lock) = publish_connection_file(
1481        &bound.connection_file_path,
1482        #[cfg(windows)]
1483        bound.connection_file_copy.as_deref(),
1484        &bound.connection_info,
1485        bound.local_endpoint.as_ref(),
1486        publication_lock,
1487    )
1488    .await;
1489    published?;
1490    // Serving and discoverable: a daemon booted from a restart intent records
1491    // readiness, which makes that generation the daemon's active one. Without
1492    // this record the next start treats the warm-up as failed.
1493    if let Some(owner) = daemon_ready_runtime {
1494        if let Err(error) = owner.on_blocking(|owner| owner.record_daemon_ready()).await {
1495            warn!(%error, "daemon readiness was not recorded in the approved store");
1496        }
1497    }
1498    #[cfg(test)]
1499    startup_publication_hook::run(&bound.connection_file_path).await;
1500    drop(publication_lock);
1501    let _clock_step_task = AbortOnDrop::new(crate::watchdog::spawn_clock_step_monitor());
1502    // Off the startup path: it runs `systemctl`, and only ever logs.
1503    #[cfg(target_os = "linux")]
1504    let _kill_mode_check = AbortOnDrop::new(tokio::spawn(
1505        crate::systemd_kill_mode::warn_if_kill_mode_defeats_ordered_shutdown(),
1506    ));
1507    let _watchdog_task = AbortOnDrop::new(
1508        DaemonSelfWatchdog::new(
1509            bound.connection_info.clone(),
1510            bound.connection_file_path.clone(),
1511        )
1512        .with_config(watchdog_config)
1513        .spawn(),
1514    );
1515
1516    for registered in registered_modules {
1517        match supervisor
1518            .start_registered_configured_async(registered, None, true)
1519            .await
1520        {
1521            Ok(module) => {
1522                info!(module_id = %module.module_id(), "configured module supervised");
1523            }
1524            Err(err) => {
1525                error!(error = %err, "failed to start configured module; continuing daemon startup");
1526            }
1527        }
1528    }
1529
1530    control.refresh_capability_requirements();
1531    Arc::clone(&control).spawn_capability_deadline_loop();
1532    Arc::clone(&control).spawn_scope_expiry_loop();
1533    let _config_delivery_task = control.spawn_config_delivery_worker().map(AbortOnDrop::new);
1534    control.spawn_resource_sampler();
1535
1536    #[cfg(unix)]
1537    {
1538        tokio::select! {
1539            result = serve_task.join() => {
1540                return result.map_err(BootstrapError::ServeJoin)?.map_err(BootstrapError::Serve);
1541            }
1542            _ = terminate.recv() => {}
1543        }
1544        // First, before the notice, the drain, or any connection close: from
1545        // here on a module exit is recorded as `daemon_shutdown` and never
1546        // respawned. Also before allowing a second signal to cut the bounded
1547        // wait short, so the journal marker is always written.
1548        supervisor.begin_daemon_shutdown();
1549        // Stop the self-watchdog before closing the listener. Its next tick would
1550        // connect to that listener, fail, and log an ERROR indistinguishable from
1551        // a wedged daemon, once for every interval a planned stop lasts.
1552        drop(_watchdog_task);
1553        // Dropping the listener stops new accepts, not established connections:
1554        // their detached tasks must remain live throughout notice and drain. The
1555        // local endpoint stops accepting the same way.
1556        drop(serve_task);
1557        drop(local_serve_task);
1558        let second_sigterm = async {
1559            terminate.recv().await;
1560            std::time::Instant::now()
1561        };
1562        tokio::pin!(second_sigterm);
1563        let mut escalation_started = tokio::select! {
1564            biased;
1565            at = second_sigterm.as_mut() => {
1566                info!("second SIGTERM: abandoning daemon shutdown wait");
1567                Some(at)
1568            }
1569            result = supervisor.drain_for_daemon_shutdown() => {
1570                if let Err(error) = result {
1571                    warn!(%error, "daemon shutdown drain failed; exiting anyway");
1572                }
1573                None
1574            }
1575        };
1576        // Supervised modules lead their own process groups, so a service
1577        // manager's kill of this process's group does not reach them. (A
1578        // systemd unit with KillMode=control-group kills by cgroup instead and
1579        // does reach them; see `systemd_kill_mode`.) The daemon ends them
1580        // itself: EOF (or SIGTERM for a protocol none child) first, then
1581        // signals at each child's own drain deadline.
1582        let already_escalated = escalation_started.is_some();
1583        supervisor
1584            .end_children_for_daemon_shutdown(already_escalated, async {
1585                if already_escalated {
1586                    std::future::pending::<()>().await;
1587                } else {
1588                    escalation_started = Some(second_sigterm.await);
1589                }
1590            })
1591            .await;
1592        if let Some(at) = escalation_started {
1593            info!(
1594                second_sigterm_elapsed_ms = at.elapsed().as_millis() as u64,
1595                "daemon shutdown wait ended after second SIGTERM"
1596            );
1597        }
1598        info!("daemon exiting: SIGTERM shutdown complete");
1599        Ok(())
1600    }
1601    #[cfg(not(unix))]
1602    serve_task
1603        .join()
1604        .await
1605        .map_err(BootstrapError::ServeJoin)?
1606        .map_err(BootstrapError::Serve)
1607}
1608
1609fn normalized_build_provenance(value: &str) -> Option<String> {
1610    match value.trim() {
1611        "" | "unavailable" => None,
1612        value => Some(value.to_string()),
1613    }
1614}
1615
1616/// Find an existing daemon or atomically bind loopback TCP for this daemon.
1617///
1618/// The algorithm is intentionally connect-first: an endpoint from the connection
1619/// file is treated as live only after the TCP+key server-proof authenticates for
1620/// that file's key and daemon_id. Stale or foreign connection files are reclaimed
1621/// only while holding the per-user start lock; the TCP port is never the
1622/// singleton primitive.
1623pub async fn ensure_singleton(
1624    connection_file_path: impl AsRef<Path>,
1625    port: u16,
1626) -> Result<Outcome, BootstrapError> {
1627    ensure_singleton_with_config(BootstrapConfig::new(connection_file_path.as_ref(), port)).await
1628}
1629
1630pub async fn ensure_singleton_with_config(
1631    config: BootstrapConfig,
1632) -> Result<Outcome, BootstrapError> {
1633    ensure_singleton_inner(config, true).await
1634}
1635
1636async fn ensure_singleton_inner(
1637    config: BootstrapConfig,
1638    publish: bool,
1639) -> Result<Outcome, BootstrapError> {
1640    let Some(lock) = singleton_start_lock(&config.connection_file_path).await? else {
1641        return Ok(Outcome::AlreadyRunning);
1642    };
1643    ensure_singleton_with_start_lock(config, publish, lock).await
1644}
1645
1646async fn singleton_start_lock(path: &Path) -> Result<Option<StartLock>, BootstrapError> {
1647    if stable_peer(path).await? {
1648        return Ok(None);
1649    }
1650
1651    let lock = StartLock::acquire(path).await?;
1652
1653    // Re-probe after acquiring the start lock so a peer that won the race between
1654    // our first failed probe and the lock acquisition is observed instead of
1655    // overwritten.
1656    if stable_peer(path).await? {
1657        return Ok(None);
1658    }
1659    Ok(Some(lock))
1660}
1661
1662async fn ensure_singleton_with_start_lock(
1663    config: BootstrapConfig,
1664    publish: bool,
1665    lock: StartLock,
1666) -> Result<Outcome, BootstrapError> {
1667    #[cfg(windows)]
1668    let copy_path = windows_connection_copy::resolve(&config);
1669    let path = config.connection_file_path;
1670    // The held start lock and its peer check only rule out a daemon using this
1671    // connection file. A daemon started with another runtime directory but
1672    // the same data home shares the run directory while passing both, so the
1673    // run directory needs its own owner: refuse at once if a live daemon holds
1674    // it, before anything below reads or writes run or data-home state.
1675    let run_dir_lock = config
1676        .live_children_path
1677        .as_deref()
1678        .map(crate::run_dir_lock::RunDirLock::acquire)
1679        .transpose()?;
1680
1681    remove_stale_connection_file_if_present(&path)?;
1682    #[cfg(windows)]
1683    let copy_path = windows_connection_copy::reclaim(&path, copy_path).await;
1684
1685    // Established under the start lock, and under the run-directory lock when
1686    // there is one, and before binding: a corrupt file stops boot before
1687    // anything is published, and two daemons racing to start cannot both mint.
1688    //
1689    // Minting a new machine id fsyncs its file, so loading or minting runs on
1690    // tokio's blocking pool. Both locks travel with that work, so neither is
1691    // released while a mint is in flight.
1692    let (machine_id, lock, run_dir_lock) = match config.machine_id_path {
1693        None => (None, lock, run_dir_lock),
1694        Some(machine_id_path) => {
1695            let task_path = machine_id_path.clone();
1696            match run_blocking_holding((lock, run_dir_lock), move || {
1697                crate::machine_id::load_or_mint(&task_path)
1698            })
1699            .await
1700            {
1701                Ok((Ok(id), (lock, run_dir_lock))) => (Some(id), lock, run_dir_lock),
1702                Ok((Err(error), _locks)) => return Err(BootstrapError::MachineId(error)),
1703                Err(join_error) => {
1704                    return Err(BootstrapError::MachineId(
1705                        crate::machine_id::MachineIdFileError::Write {
1706                            path: machine_id_path,
1707                            source: io::Error::other(describe_blocking_task_failure(
1708                                "loading or minting the machine id",
1709                                &join_error,
1710                            )),
1711                        },
1712                    ))
1713                }
1714            }
1715        }
1716    };
1717
1718    let (listeners, endpoints) = bind_loopback(config.port).await?;
1719    // The local endpoint is an addition to the daemon, not a condition for it:
1720    // a bind failure leaves clients on TCP only, with the connection file naming
1721    // no local endpoint, and the reason reported by `server.describe`. A daemon
1722    // that refused to start would take every module down for a feature nothing
1723    // has to use yet.
1724    let (local_listener, local_endpoint, local_endpoint_unavailable) =
1725        if config.local_listener.configured_dir().is_some() {
1726            match local_listener::bind(&config.local_listener, &path) {
1727                Ok(local_listener) => {
1728                    let endpoint = local_listener.endpoint();
1729                    (Some(local_listener), Some(endpoint), None)
1730                }
1731                Err(error) => {
1732                    let reason = error.to_string();
1733                    warn!(
1734                        error = %reason,
1735                        "the local endpoint is unavailable; serving loopback TCP only"
1736                    );
1737                    (None, None, Some(reason))
1738                }
1739            }
1740        } else {
1741            (None, None, None)
1742        };
1743    let connection_info = ConnectionInfo {
1744        schema: SCHEMA_VERSION,
1745        wire_version: Some(PROTOCOL_VERSION),
1746        endpoints,
1747        key: generate_key().map_err(BootstrapError::GenerateConnectionFile)?,
1748        daemon_id: generate_daemon_id().map_err(BootstrapError::GenerateConnectionFile)?,
1749        pid: process::id(),
1750        daemon_ver: config.daemon_ver,
1751    };
1752
1753    // The low-level binding API historically publishes for callers that serve
1754    // these listeners themselves. The daemon entry point defers publication
1755    // until its own authentication server is running.
1756    let publication_lock = if publish {
1757        let (published, lock) = publish_connection_file(
1758            &path,
1759            #[cfg(windows)]
1760            copy_path.as_deref(),
1761            &connection_info,
1762            local_endpoint.as_ref(),
1763            lock,
1764        )
1765        .await;
1766        if let Err(error) = published {
1767            if let Some(local_listener) = local_listener.as_ref() {
1768                local_listener::release(local_listener);
1769            }
1770            drop(listeners);
1771            return Err(error);
1772        }
1773        drop(lock);
1774        None
1775    } else {
1776        Some(lock)
1777    };
1778
1779    Ok(Outcome::Bound(BoundDaemon {
1780        listeners,
1781        local_listener,
1782        local_endpoint,
1783        local_endpoint_unavailable,
1784        connection_info,
1785        connection_file_path: path,
1786        connection_file_source: config.connection_file_source,
1787        #[cfg(windows)]
1788        connection_file_copy: copy_path,
1789        machine_id,
1790        run_dir_lock,
1791        publication_lock,
1792    }))
1793}
1794
1795#[derive(Debug, Clone, Copy, PartialEq, Eq)]
1796enum Probe {
1797    Live,
1798    StaleOrAbsent,
1799}
1800
1801async fn stable_peer(path: &Path) -> Result<bool, BootstrapError> {
1802    if !matches!(probe_existing(path).await?, Probe::Live) {
1803        return Ok(false);
1804    }
1805    // A terminating peer can still authenticate while its listeners drain.
1806    // Require another proof after a short settling interval before reporting
1807    // a successful duplicate start. If the peer left, proceed with startup;
1808    // any remaining bind/ownership conflict is an error, not exit 0 (which
1809    // would suppress launchd's restart-on-failure policy).
1810    tokio::time::sleep(Duration::from_millis(1500)).await;
1811    Ok(matches!(probe_existing(path).await?, Probe::Live))
1812}
1813
1814async fn probe_existing(path: &Path) -> Result<Probe, BootstrapError> {
1815    let info = match connection_file::read(path) {
1816        Ok(info) => info,
1817        Err(source) if is_absent_or_stale_connection_file(&source) => {
1818            return Ok(Probe::StaleOrAbsent)
1819        }
1820        Err(source) => {
1821            return Err(BootstrapError::ConnectionFileRead {
1822                path: path.to_path_buf(),
1823                source,
1824            })
1825        }
1826    };
1827
1828    for endpoint in &info.endpoints {
1829        if matches!(probe_endpoint(&info, endpoint).await, Probe::Live) {
1830            return Ok(Probe::Live);
1831        }
1832    }
1833
1834    Ok(Probe::StaleOrAbsent)
1835}
1836
1837async fn probe_endpoint(info: &ConnectionInfo, endpoint: &Endpoint) -> Probe {
1838    let Ok(ip) = endpoint.host.parse::<IpAddr>() else {
1839        return Probe::StaleOrAbsent;
1840    };
1841    if !ip.is_loopback() {
1842        return Probe::StaleOrAbsent;
1843    }
1844    let addr = SocketAddr::new(ip, endpoint.port);
1845
1846    let mut stream = match timeout(CONNECT_TIMEOUT, TcpStream::connect(addr)).await {
1847        Ok(Ok(stream)) => stream,
1848        Ok(Err(_)) | Err(_) => return Probe::StaleOrAbsent,
1849    };
1850
1851    match authenticate_client(&mut stream, info, PROBE_AUTH_DEADLINE).await {
1852        Ok(()) => Probe::Live,
1853        Err(AuthError::DaemonIdMismatch)
1854        | Err(AuthError::InvalidServerProof)
1855        | Err(AuthError::UnexpectedEof { .. })
1856        | Err(AuthError::Timeout { .. })
1857        | Err(AuthError::JsonEncode { .. })
1858        | Err(AuthError::JsonDecode { .. })
1859        | Err(AuthError::Io { .. })
1860        | Err(AuthError::MessageTooLarge { .. })
1861        | Err(AuthError::KeyTooShort { .. })
1862        | Err(AuthError::Random(_))
1863        | Err(AuthError::InvalidClientAuth) => Probe::StaleOrAbsent,
1864    }
1865}
1866
1867async fn verify_serving(info: &ConnectionInfo) -> Result<(), BootstrapError> {
1868    // Scheduling a server task is not a readiness barrier. Authenticate once
1869    // against the mandatory IPv4 listener before exposing discovery to peers.
1870    // This uses the normal probe deadline, with no retries or widened budget.
1871    if let Some(endpoint) = info.endpoints.first() {
1872        if matches!(probe_endpoint(info, endpoint).await, Probe::Live) {
1873            return Ok(());
1874        }
1875    }
1876    Err(BootstrapError::StartupNotServing)
1877}
1878
1879/// Run blocking filesystem work (an fsync, a rename) on tokio's blocking-thread
1880/// pool rather than on the async worker thread that awaits it. On a slow or
1881/// busy disk an fsync can take seconds, and a runtime has few worker threads
1882/// (an in-process test daemon may share two with everything else), so doing it
1883/// inline stalls unrelated tasks for that long.
1884///
1885/// `guards` (lock handles) move into the blocking task and are handed back
1886/// with the work's result. If the awaiting future is dropped first, the work
1887/// still runs to completion on its thread and the guards are released only
1888/// after it ends, so a lock is never given up while the work it protects is
1889/// still writing.
1890///
1891/// `Err` means the task panicked, or was cancelled before it started because
1892/// the runtime is shutting down; the guards have then been released already.
1893async fn run_blocking_holding<G, T>(
1894    guards: G,
1895    work: impl FnOnce() -> T + Send + 'static,
1896) -> Result<(T, G), JoinError>
1897where
1898    G: Send + 'static,
1899    T: Send + 'static,
1900{
1901    tokio::task::spawn_blocking(move || {
1902        let output = work();
1903        (output, guards)
1904    })
1905    .await
1906}
1907
1908/// Explain, for a startup error, why blocking work produced no result.
1909fn describe_blocking_task_failure(work: &str, join_error: &JoinError) -> String {
1910    if join_error.is_panic() {
1911        format!("the background thread {work} panicked before finishing: {join_error}")
1912    } else {
1913        format!(
1914            "the background thread {work} was cancelled before finishing, because the async runtime is shutting down: {join_error}"
1915        )
1916    }
1917}
1918
1919/// Publish the connection file: owner-only permissions, written to a temporary
1920/// file, fsynced and renamed over `path` (all done by `write_atomic`). The
1921/// write runs on tokio's blocking pool through [`run_blocking_holding`], so the
1922/// fsync never holds an async worker thread.
1923///
1924/// `guard` is the singleton start lock (or whatever the caller must keep held
1925/// across publication); it is returned with the result. It comes back `None`
1926/// only when the write task panicked or was cancelled, in which case the
1927/// result is an error and the file must be treated as not published.
1928///
1929/// On Windows, `copy_path` asks for the profile copy as well. Its start lock is
1930/// taken before the primary write. Once `path` is written, an error returned
1931/// while preparing or writing the copy only logs a warning, and the result is
1932/// still success. The two writes share one blocking task but aren't one atomic
1933/// step, and a panic or cancellation of that task is still a publication
1934/// error.
1935async fn publish_connection_file<G: Send + 'static>(
1936    path: &Path,
1937    #[cfg(windows)] copy_path: Option<&Path>,
1938    info: &ConnectionInfo,
1939    local_endpoint: Option<&LocalEndpoint>,
1940    guard: G,
1941) -> (Result<(), BootstrapError>, Option<G>) {
1942    let task_path = path.to_path_buf();
1943    let task_info = info.clone();
1944    let task_endpoint = local_endpoint.cloned();
1945    #[cfg(windows)]
1946    let copy = windows_connection_copy::claim_for_publication(path, copy_path, info).await;
1947    let outcome = run_blocking_holding(guard, move || {
1948        write_connection_file(&task_path, &task_info, task_endpoint.as_ref())?;
1949        #[cfg(windows)]
1950        if let Some((copy_path, _lock)) = copy {
1951            if let Err(error) =
1952                write_connection_file(&copy_path, &task_info, task_endpoint.as_ref())
1953            {
1954                windows_connection_copy::warn_unavailable(&task_path, Some(&copy_path), &error);
1955            }
1956        }
1957        Ok(())
1958    })
1959    .await;
1960    match outcome {
1961        Ok((Ok(()), guard)) => (Ok(()), Some(guard)),
1962        Ok((Err(source), guard)) => (
1963            Err(BootstrapError::ConnectionFileWrite {
1964                path: path.to_path_buf(),
1965                source,
1966            }),
1967            Some(guard),
1968        ),
1969        Err(join_error) => (
1970            Err(BootstrapError::ConnectionFileWrite {
1971                path: path.to_path_buf(),
1972                source: ConnectionFileError::Io {
1973                    op: "write",
1974                    path: path.to_path_buf(),
1975                    source: io::Error::other(describe_blocking_task_failure(
1976                        "writing the connection file",
1977                        &join_error,
1978                    )),
1979                },
1980            }),
1981            None,
1982        ),
1983    }
1984}
1985
1986/// The synchronous connection-file write, including its fsync. Call it only
1987/// through [`publish_connection_file`], which keeps it off the async workers.
1988///
1989/// `local_endpoint` is written through subc-transport's endpoint-aware writer,
1990/// so the file carries `local_endpoint` when there is a local listener and is
1991/// byte-identical to before when there is none.
1992fn write_connection_file(
1993    path: &Path,
1994    info: &ConnectionInfo,
1995    local_endpoint: Option<&LocalEndpoint>,
1996) -> Result<(), ConnectionFileError> {
1997    #[cfg(test)]
1998    connection_file_write_hook::run(path);
1999    write_atomic_with_local_endpoint(path, info, local_endpoint)
2000}
2001
2002/// Lets a test run code inside the connection-file write, on whichever thread
2003/// performs it, for example to block there and check what else can still run.
2004#[cfg(test)]
2005mod connection_file_write_hook {
2006    use std::{
2007        path::{Path, PathBuf},
2008        sync::{Mutex, PoisonError},
2009    };
2010
2011    type Hook = Box<dyn FnOnce() + Send>;
2012
2013    /// Keyed by connection-file path, so a hook fires only for the test that
2014    /// installed it while other tests publish their own files in parallel.
2015    static HOOKS: Mutex<Vec<(PathBuf, Hook)>> = Mutex::new(Vec::new());
2016
2017    /// Run `hook` once, at the start of the next write of `path`.
2018    pub(super) fn install(path: &Path, hook: impl FnOnce() + Send + 'static) {
2019        HOOKS
2020            .lock()
2021            .unwrap_or_else(PoisonError::into_inner)
2022            .push((path.to_path_buf(), Box::new(hook)));
2023    }
2024
2025    pub(super) fn run(path: &Path) {
2026        let hook = {
2027            let mut hooks = HOOKS.lock().unwrap_or_else(PoisonError::into_inner);
2028            hooks
2029                .iter()
2030                .position(|(hooked, _)| hooked == path)
2031                .map(|index| hooks.swap_remove(index).1)
2032        };
2033        if let Some(hook) = hook {
2034            hook();
2035        }
2036    }
2037}
2038
2039/// A path-scoped barrier after publication and before any configured activation.
2040/// The async wait leaves the listeners free to answer discovery-time requests.
2041#[cfg(test)]
2042mod startup_publication_hook {
2043    use std::{
2044        path::{Path, PathBuf},
2045        sync::{Mutex, PoisonError},
2046    };
2047    use tokio::sync::oneshot;
2048
2049    type Barrier = (oneshot::Sender<()>, oneshot::Receiver<()>);
2050    static BARRIERS: Mutex<Vec<(PathBuf, Barrier)>> = Mutex::new(Vec::new());
2051
2052    pub(super) fn install(path: &Path) -> (oneshot::Receiver<()>, oneshot::Sender<()>) {
2053        let (entered_tx, entered_rx) = oneshot::channel();
2054        let (release_tx, release_rx) = oneshot::channel();
2055        BARRIERS
2056            .lock()
2057            .unwrap_or_else(PoisonError::into_inner)
2058            .push((path.to_path_buf(), (entered_tx, release_rx)));
2059        (entered_rx, release_tx)
2060    }
2061
2062    pub(super) async fn run(path: &Path) {
2063        let barrier = {
2064            let mut barriers = BARRIERS.lock().unwrap_or_else(PoisonError::into_inner);
2065            barriers
2066                .iter()
2067                .position(|(hooked, _)| hooked == path)
2068                .map(|index| barriers.swap_remove(index).1)
2069        };
2070        if let Some((entered, release)) = barrier {
2071            let _ = entered.send(());
2072            let _ = release.await;
2073        }
2074    }
2075}
2076
2077fn is_absent_or_stale_connection_file(err: &ConnectionFileError) -> bool {
2078    match err {
2079        ConnectionFileError::Io { source, .. } if source.kind() == io::ErrorKind::NotFound => true,
2080        ConnectionFileError::JsonRead { .. }
2081        | ConnectionFileError::UnsupportedSchema { .. }
2082        | ConnectionFileError::Invalid { .. }
2083        | ConnectionFileError::KeyTooShort { .. }
2084        // A live daemon always publishes the file owner-only (0600), so a file
2085        // with insecure permissions is never a daemon we should defer to: treat it
2086        // as stale and take over (which republishes a correct 0600 file).
2087        | ConnectionFileError::InsecurePermissions { .. } => true,
2088        ConnectionFileError::MissingParent { .. }
2089        | ConnectionFileError::MissingFileName { .. }
2090        // A writable ancestor is an operator misconfiguration, never evidence
2091        // about whether a daemon is live. Reclaiming the file would republish key
2092        // material into the same directory the refusal is about.
2093        | ConnectionFileError::InsecureParentDirectory { .. }
2094        | ConnectionFileError::Io { .. }
2095        | ConnectionFileError::JsonWrite { .. }
2096        | ConnectionFileError::Random(_)
2097        // A wire mismatch may identify a newer live daemon, so never reclaim its
2098        // connection file merely because this binary cannot speak its envelope.
2099        | ConnectionFileError::WireVersionMismatch { .. } => false,
2100    }
2101}
2102
2103fn remove_stale_connection_file_if_present(path: &Path) -> Result<(), BootstrapError> {
2104    match fs::remove_file(path) {
2105        Ok(()) => Ok(()),
2106        Err(err) if err.kind() == io::ErrorKind::NotFound => Ok(()),
2107        Err(source) => Err(BootstrapError::RemoveStale {
2108            path: path.to_path_buf(),
2109            source,
2110        }),
2111    }
2112}
2113
2114async fn bind_loopback(port: u16) -> Result<(Vec<TcpListener>, Vec<Endpoint>), BootstrapError> {
2115    bind_loopback_with(port, |port| TcpListener::bind((Ipv6Addr::LOCALHOST, port))).await
2116}
2117
2118async fn bind_loopback_with<F, Fut>(
2119    port: u16,
2120    mut bind_v6: F,
2121) -> Result<(Vec<TcpListener>, Vec<Endpoint>), BootstrapError>
2122where
2123    F: FnMut(u16) -> Fut,
2124    Fut: std::future::Future<Output = io::Result<TcpListener>>,
2125{
2126    let v4_host = Ipv4Addr::LOCALHOST;
2127    let v4 = TcpListener::bind((v4_host, port))
2128        .await
2129        .map_err(|source| BootstrapError::Bind {
2130            host: v4_host.to_string(),
2131            port,
2132            source,
2133        })?;
2134    let actual_port = v4
2135        .local_addr()
2136        .map_err(|source| BootstrapError::LocalAddr {
2137            host: v4_host.to_string(),
2138            source,
2139        })?
2140        .port();
2141
2142    let mut listeners = vec![v4];
2143    let mut endpoints = vec![Endpoint {
2144        host: v4_host.to_string(),
2145        port: actual_port,
2146    }];
2147
2148    let v6_host = Ipv6Addr::LOCALHOST;
2149    match bind_v6(actual_port).await {
2150        Ok(v6) => {
2151            listeners.push(v6);
2152            endpoints.push(Endpoint {
2153                host: v6_host.to_string(),
2154                port: actual_port,
2155            });
2156        }
2157        Err(err)
2158            if ipv6_loopback_unavailable(&err)
2159                || (port == 0 && err.kind() == io::ErrorKind::AddrInUse) =>
2160        {
2161            warn!(
2162                port = actual_port,
2163                error = %err,
2164                "IPv6 loopback unavailable; serving only IPv4 loopback"
2165            );
2166        }
2167        Err(source) => {
2168            drop(listeners);
2169            return Err(BootstrapError::Bind {
2170                host: v6_host.to_string(),
2171                port: actual_port,
2172                source,
2173            });
2174        }
2175    }
2176
2177    Ok((listeners, endpoints))
2178}
2179
2180fn ipv6_loopback_unavailable(err: &io::Error) -> bool {
2181    matches!(
2182        err.kind(),
2183        io::ErrorKind::AddrNotAvailable | io::ErrorKind::Unsupported
2184    ) || matches!(err.raw_os_error(), Some(47) | Some(49) | Some(97))
2185}
2186
2187struct AbortOnDrop<T> {
2188    handle: JoinHandle<T>,
2189}
2190
2191impl<T> AbortOnDrop<T> {
2192    fn new(handle: JoinHandle<T>) -> Self {
2193        Self { handle }
2194    }
2195
2196    async fn join(&mut self) -> Result<T, JoinError> {
2197        (&mut self.handle).await
2198    }
2199}
2200
2201impl<T> Drop for AbortOnDrop<T> {
2202    fn drop(&mut self) {
2203        if !self.handle.is_finished() {
2204            self.handle.abort();
2205        }
2206    }
2207}
2208
2209#[derive(Debug)]
2210struct StartLock {
2211    // Keep the locked file handle alive for the duration of bootstrap.
2212    _file: fs::File,
2213}
2214
2215impl Drop for StartLock {
2216    fn drop(&mut self) {
2217        // Closing alone leaves the lock held if a concurrent fork inherited
2218        // the open file description and has not exec'd yet. Unlock explicitly
2219        // so a restart can reclaim the singleton as soon as bootstrap ends.
2220        if let Err(error) = FileExt::unlock(&self._file) {
2221            tracing::warn!(%error, "failed to release the singleton start lock");
2222        }
2223    }
2224}
2225
2226impl StartLock {
2227    async fn acquire(connection_file_path: &Path) -> Result<Self, BootstrapError> {
2228        let path = start_lock_path(connection_file_path);
2229        for _ in 0..START_LOCK_RETRIES {
2230            let file = match open_owner_only_lock(&path) {
2231                Ok(file) => file,
2232                Err(source) => return Err(BootstrapError::StartLockCreate { path, source }),
2233            };
2234            match FileExt::try_lock(&file) {
2235                Ok(()) => return Ok(Self { _file: file }),
2236                Err(TryLockError::WouldBlock) => sleep(START_LOCK_RETRY_DELAY).await,
2237                Err(TryLockError::Error(source)) => {
2238                    return Err(BootstrapError::StartLockCreate { path, source });
2239                }
2240            }
2241        }
2242
2243        Err(BootstrapError::StartLockBusy {
2244            path,
2245            attempts: START_LOCK_RETRIES,
2246        })
2247    }
2248}
2249
2250pub(crate) fn open_owner_only_lock(path: &Path) -> io::Result<fs::File> {
2251    let mut options = fs::OpenOptions::new();
2252    options.read(true).write(true).create(true);
2253    #[cfg(unix)]
2254    {
2255        use std::os::unix::fs::OpenOptionsExt;
2256        options.mode(0o600);
2257    }
2258    options.open(path)
2259}
2260
2261fn start_lock_path(connection_file_path: &Path) -> PathBuf {
2262    let file_name = connection_file_path
2263        .file_name()
2264        .map(|name| name.to_string_lossy())
2265        .unwrap_or_else(|| CONNECTION_FILE_NAME.into());
2266    let lock_name = format!("{file_name}.start-lock");
2267    connection_file_path
2268        .parent()
2269        .filter(|parent| !parent.as_os_str().is_empty())
2270        .unwrap_or_else(|| Path::new("."))
2271        .join(lock_name)
2272}
2273
2274fn non_empty_os_var(key: &str) -> Option<OsString> {
2275    let value = env::var_os(key)?;
2276    if value.is_empty() {
2277        None
2278    } else {
2279        Some(value)
2280    }
2281}
2282
2283/// Bootstrap-layer errors are deliberately typed so startup never panics for
2284/// ordinary daemon-discovery races or stale filesystem state.
2285#[derive(Debug)]
2286#[non_exhaustive]
2287pub enum BootstrapError {
2288    ConfigStore(String),
2289    /// The bound server could not authenticate its own readiness probe.
2290    StartupNotServing,
2291    #[cfg(unix)]
2292    Signal(io::Error),
2293    InvalidPort {
2294        raw: String,
2295        source: std::num::ParseIntError,
2296    },
2297    ConnectionFileRead {
2298        path: PathBuf,
2299        source: ConnectionFileError,
2300    },
2301    ConnectionFileWrite {
2302        path: PathBuf,
2303        source: ConnectionFileError,
2304    },
2305    GenerateConnectionFile(ConnectionFileError),
2306    StartLockCreate {
2307        path: PathBuf,
2308        source: io::Error,
2309    },
2310    StartLockBusy {
2311        path: PathBuf,
2312        attempts: usize,
2313    },
2314    /// The run-directory lock file could not be created, opened or locked.
2315    RunDirLockCreate {
2316        path: PathBuf,
2317        source: io::Error,
2318    },
2319    /// Another live process, almost certainly a daemon started with a
2320    /// different runtime directory over the same data home, owns the run
2321    /// directory. The daemon does not start. `holder_pid` is read from the
2322    /// lock file when possible and is informational only.
2323    RunDirBusy {
2324        path: PathBuf,
2325        holder_pid: Option<u32>,
2326    },
2327    RemoveStale {
2328        path: PathBuf,
2329        source: io::Error,
2330    },
2331    Bind {
2332        host: String,
2333        port: u16,
2334        source: io::Error,
2335    },
2336    LocalAddr {
2337        host: String,
2338        source: io::Error,
2339    },
2340    DaemonConfig(DaemonConfigError),
2341    /// The machine id could not be established: its file is corrupt, unreadable
2342    /// or unwritable, or the data home is relative. The daemon does not start.
2343    MachineId(crate::machine_id::MachineIdFileError),
2344    /// The daemon run directory could not be resolved because the data home is
2345    /// relative. The daemon does not start.
2346    RunDir(daemon_config::DaemonRunDirError),
2347    Serve(ServerError),
2348    ServeJoin(tokio::task::JoinError),
2349}
2350
2351impl From<ServerError> for BootstrapError {
2352    fn from(error: ServerError) -> Self {
2353        Self::Serve(error)
2354    }
2355}
2356
2357impl fmt::Display for BootstrapError {
2358    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
2359        match self {
2360            Self::ConfigStore(error) => write!(f, "configuration store startup refused: {error}"),
2361            Self::StartupNotServing => write!(f, "bound daemon did not authenticate its readiness probe; connection file was not published"),
2362            #[cfg(unix)]
2363            Self::Signal(error) => write!(f, "failed to register SIGTERM handler: {error}"),
2364            Self::InvalidPort { raw, source } => {
2365                write!(f, "invalid {SUBC_PORT_ENV} value '{raw}': {source}")
2366            }
2367            Self::ConnectionFileRead { path, source } => write!(
2368                f,
2369                "failed to read connection file {}: {source}",
2370                path.display()
2371            ),
2372            Self::ConnectionFileWrite { path, source } => write!(
2373                f,
2374                "failed to publish connection file {}: {source}",
2375                path.display()
2376            ),
2377            Self::GenerateConnectionFile(err) => {
2378                write!(f, "failed to generate connection-file auth material: {err}")
2379            }
2380            Self::StartLockCreate { path, source } => {
2381                write!(
2382                    f,
2383                    "failed to create start lock {}: {source}",
2384                    path.display()
2385                )
2386            }
2387            Self::StartLockBusy { path, attempts } => write!(
2388                f,
2389                "start lock {} remained busy after {attempts} attempts",
2390                path.display()
2391            ),
2392            Self::RunDirLockCreate { path, source } => write!(
2393                f,
2394                "refusing to start: failed to lock run directory via {}: {source}",
2395                path.display()
2396            ),
2397            Self::RunDirBusy { path, holder_pid } => {
2398                write!(
2399                    f,
2400                    "refusing to start: run directory lock {} is held by another daemon",
2401                    path.display()
2402                )?;
2403                if let Some(pid) = holder_pid {
2404                    write!(f, " (pid {pid})")?;
2405                }
2406                write!(
2407                    f,
2408                    "; a live daemon owns this data home's run state (typically one started with a different XDG_RUNTIME_DIR)"
2409                )
2410            }
2411            Self::RemoveStale { path, source } => write!(
2412                f,
2413                "failed to remove stale connection file {}: {source}",
2414                path.display()
2415            ),
2416            Self::Bind { host, port, source } if source.kind() == io::ErrorKind::AddrInUse => {
2417                write!(
2418                    f,
2419                    "port {port} in use on loopback {host}: {source}; set the port in config"
2420                )
2421            }
2422            Self::Bind { host, port, source } => {
2423                write!(f, "failed to bind loopback TCP {host}:{port}: {source}")
2424            }
2425            Self::LocalAddr { host, source } => {
2426                write!(f, "failed to read local address for {host}: {source}")
2427            }
2428            Self::DaemonConfig(err) => write!(f, "failed to load daemon config: {err}"),
2429            Self::MachineId(err) => write!(f, "refusing to start: {err}"),
2430            Self::RunDir(err) => write!(f, "refusing to start: {err}"),
2431            Self::Serve(err) => write!(f, "daemon server failed: {err}"),
2432            Self::ServeJoin(err) => write!(f, "daemon server task failed: {err}"),
2433        }
2434    }
2435}
2436
2437impl Error for BootstrapError {
2438    fn source(&self) -> Option<&(dyn Error + 'static)> {
2439        match self {
2440            Self::ConfigStore(_) => None,
2441            Self::StartupNotServing => None,
2442            #[cfg(unix)]
2443            Self::Signal(source) => Some(source),
2444            Self::InvalidPort { source, .. } => Some(source),
2445            Self::ConnectionFileRead { source, .. }
2446            | Self::ConnectionFileWrite { source, .. }
2447            | Self::GenerateConnectionFile(source) => Some(source),
2448            Self::StartLockCreate { source, .. }
2449            | Self::RunDirLockCreate { source, .. }
2450            | Self::RemoveStale { source, .. }
2451            | Self::Bind { source, .. }
2452            | Self::LocalAddr { source, .. } => Some(source),
2453            Self::DaemonConfig(err) => Some(err),
2454            Self::MachineId(err) => Some(err),
2455            Self::RunDir(err) => Some(err),
2456            Self::Serve(err) => Some(err),
2457            Self::ServeJoin(err) => Some(err),
2458            Self::StartLockBusy { .. } | Self::RunDirBusy { .. } => None,
2459        }
2460    }
2461}
2462
2463#[cfg(all(test, feature = "test-support", unix))]
2464#[path = "bootstrap_recovery_tests.rs"]
2465mod recovery_tests;
2466
2467#[cfg(test)]
2468mod tests {
2469    mod activity {
2470        include!("activity_bootstrap_tests.rs");
2471    }
2472    mod startup_order {
2473        include!("startup_order_tests.rs");
2474    }
2475    use super::*;
2476    use crate::server::ServerAuth;
2477    use cortexkit_test_support::ScratchDir;
2478    #[cfg(target_os = "linux")]
2479    use std::collections::BTreeSet;
2480    use std::sync::Mutex;
2481    #[cfg(target_os = "linux")]
2482    use subc_control::ModuleProtocol;
2483    use subc_transport::{write_atomic, MIN_KEY_LEN};
2484    use tokio::io::AsyncReadExt;
2485    use tokio::task::JoinHandle;
2486
2487    #[cfg(unix)]
2488    use std::os::unix::fs::PermissionsExt;
2489
2490    static ENV_LOCK: Mutex<()> = Mutex::new(());
2491
2492    #[test]
2493    fn normalized_build_provenance_preserves_real_values() {
2494        assert_eq!(normalized_build_provenance("abc"), Some("abc".to_string()));
2495    }
2496
2497    #[test]
2498    fn config_recovery_bootstrap_preserves_all_loaded_fields_and_only_argument_overrides() {
2499        let _env_lock = ENV_LOCK.lock().unwrap();
2500        let scratch = unique_temp_dir("bootstrap-loaded-fields");
2501        let file = scratch.join("subc.jsonc");
2502        std::fs::write(
2503            &file,
2504            br#"{"version":1,"modules":{},"scope_authority_owners":["file-owner"]}"#,
2505        )
2506        .unwrap();
2507        let mut config = BootstrapConfig::new(scratch.join("connection.json"), 0)
2508            .with_daemon_config_source(file.clone(), None);
2509        config.defer_file_load = true;
2510        config.operator_timeout = Duration::from_secs(7);
2511        let selected = load_bootstrap_authority(config, None).unwrap();
2512        assert_eq!(
2513            selected.file_derived_test_probe,
2514            Some(vec!["file-owner".into()])
2515        );
2516        assert_eq!(selected.operator_timeout, Duration::from_secs(7));
2517        assert_eq!(selected.daemon_config_path, Some(file));
2518        assert_eq!(
2519            selected.connection_file_path,
2520            scratch.join("connection.json")
2521        );
2522    }
2523
2524    #[test]
2525    fn config_recovery_bootstrap_ignores_invalid_live_file_and_poisoned_settings() {
2526        let scratch = unique_temp_dir("recovery-bootstrap");
2527        let root = scratch.join("store");
2528        std::fs::create_dir_all(&root).unwrap();
2529        std::fs::write(
2530            root.join(crate::config_store::STORE_FILE),
2531            br#"{"format_version":999}"#,
2532        )
2533        .unwrap();
2534        let mut store = crate::config_store::ConfigStore::open_digest_only_for_tests(
2535            crate::config_store::StoreRoot::from_service_definition(root).unwrap(),
2536        );
2537        assert_eq!(
2538            store.status().state,
2539            crate::config_store::StoreState::StoreUnverified
2540        );
2541        let file = scratch.join("subc.jsonc");
2542        std::fs::write(&file, b"invalid live config").unwrap();
2543        let mut config = BootstrapConfig::new(scratch.join("connection.json"), 1)
2544            .with_daemon_config_source(file, Some(1))
2545            .with_scope_authority_owners(vec!["untrusted".into()])
2546            .with_reserved_capabilities(BTreeMap::from([("untrusted".into(), "untrusted".into())]));
2547        config.defer_file_load = true;
2548        config.file_derived_test_probe = Some(vec!["untrusted".into()]);
2549        let config = load_bootstrap_authority(config, Some(&mut store)).unwrap();
2550        assert_eq!(config.file_derived_test_probe, None);
2551        assert_eq!(config.port, DEFAULT_SUBC_PORT);
2552        assert!(config.configured_modules.is_empty());
2553        assert_eq!(
2554            config.scope_authority_owners,
2555            daemon_config::default_scope_authority_owners()
2556        );
2557        assert!(config.reserved_capabilities.is_empty());
2558        assert_eq!(
2559            config.privileged_operator_prompt,
2560            daemon_config::PrivilegedOperatorPrompt::default()
2561        );
2562    }
2563
2564    /// An approved generation whose daemon bytes listen on `port`. Changing
2565    /// the port is a setting only bootstrap reads, so it waits for a restart.
2566    fn approved_port_entry(
2567        kind: crate::config_store::CommitKind,
2568        port: u16,
2569    ) -> crate::config_store::HistoryEntry {
2570        use subc_protocol::config::{ConfigLayer, ConfigLayerSnapshot};
2571        crate::config_store::HistoryEntry {
2572            generation: subc_config::Generation::new(
2573                0,
2574                subc_config::ApprovalChannel::OsPrompt,
2575                subc_config::raw_digest(b"approved-plan"),
2576                subc_config::InputVector::default(),
2577                subc_config::Changeset::new(),
2578            ),
2579            kind,
2580            approval_id: "approval".into(),
2581            outcome: "approved".into(),
2582            tuples: Default::default(),
2583            configuration: Default::default(),
2584            delivery: Some(crate::config_store::DeliverySnapshot {
2585                daemon: Some(ConfigLayerSnapshot::file(
2586                    "subc",
2587                    ConfigLayer::User,
2588                    "/approved/subc.jsonc",
2589                    Some(format!(r#"{{"version":1,"port":{port},"modules":{{}}}}"#).into_bytes()),
2590                )),
2591                modules: Default::default(),
2592                declarations: Default::default(),
2593            }),
2594        }
2595    }
2596
2597    fn approve_port(
2598        store: &mut crate::config_store::ConfigStore,
2599        kind: crate::config_store::CommitKind,
2600        port: u16,
2601    ) {
2602        store
2603            .commit_generation(
2604                approved_port_entry(kind, port),
2605                Default::default(),
2606                Default::default(),
2607                1,
2608            )
2609            .unwrap();
2610    }
2611
2612    fn stamped_boot(
2613        scratch: &Path,
2614        store: &mut crate::config_store::ConfigStore,
2615    ) -> BootstrapConfig {
2616        let mut config = BootstrapConfig::new(scratch.join("connection.json"), 1)
2617            .with_daemon_config_source(scratch.join("subc.jsonc"), Some(1));
2618        config.defer_file_load = true;
2619        load_bootstrap_authority(config, Some(store)).unwrap()
2620    }
2621
2622    #[test]
2623    fn config_stamped_boot_uses_last_active_generation_until_restart_intent() {
2624        use crate::config_store::{CommitKind, ConfigStore, StoreRoot};
2625        let scratch = unique_temp_dir("stamped-boot-selection");
2626        let root = scratch.join("store");
2627        std::fs::create_dir_all(&root).unwrap();
2628        let open = || {
2629            ConfigStore::open_digest_only_for_tests(
2630                StoreRoot::from_service_definition(root.clone()).unwrap(),
2631            )
2632        };
2633        let mut store = open();
2634        approve_port(&mut store, CommitKind::Stamp, 41001);
2635        approve_port(&mut store, CommitKind::Approved, 41002);
2636        drop(store);
2637        let mut store = open();
2638        assert_eq!(
2639            stamped_boot(&scratch, &mut store).port,
2640            41001,
2641            "an ordinary start must not boot an approval still waiting for a daemon restart"
2642        );
2643        assert_eq!(store.assign_daemon_restart().unwrap(), Some(2));
2644        drop(store);
2645        let mut store = open();
2646        assert_eq!(stamped_boot(&scratch, &mut store).port, 41002);
2647    }
2648
2649    #[test]
2650    fn config_stamped_boot_enters_recovery_when_anchor_retry_fails_at_start() {
2651        use crate::config_store::{CommitKind, ConfigStore, StoreRoot, StoreState, TestStoreKey};
2652        let scratch = unique_temp_dir("stamped-boot-anchor-pending");
2653        let root = scratch.join("store");
2654        std::fs::create_dir_all(&root).unwrap();
2655        let key = TestStoreKey::default();
2656        let open = || {
2657            ConfigStore::open_sealed_for_tests(
2658                StoreRoot::from_service_definition(root.clone()).unwrap(),
2659                key.clone(),
2660            )
2661        };
2662        let mut store = open();
2663        approve_port(&mut store, CommitKind::Stamp, 41001);
2664        // Crash between publication and anchor: the bytes are renamed into
2665        // place but the anchor never catches up, and keeps failing at start.
2666        key.fail_anchor(true);
2667        assert!(store
2668            .commit_generation(
2669                approved_port_entry(CommitKind::Approved, 41001),
2670                Default::default(),
2671                Default::default(),
2672                2,
2673            )
2674            .is_err());
2675        drop(store);
2676        let mut store = open();
2677        let status = store.status();
2678        assert_eq!(status.state, StoreState::Stamped);
2679        assert!(status.anchor_pending);
2680        let config = stamped_boot(&scratch, &mut store);
2681        assert_eq!(config.port, DEFAULT_SUBC_PORT);
2682        assert!(config.configured_modules.is_empty());
2683        assert_eq!(store.status().state, StoreState::StoreUnverified);
2684    }
2685
2686    #[test]
2687    fn normalized_build_provenance_omits_unavailable_and_empty_values() {
2688        assert_eq!(normalized_build_provenance("unavailable"), None);
2689        assert_eq!(normalized_build_provenance(""), None);
2690    }
2691
2692    #[cfg(target_os = "linux")]
2693    fn current_cgroup_path_for_test() -> io::Result<PathBuf> {
2694        let cgroups = fs::read_to_string("/proc/self/cgroup")?;
2695        let relative = cgroups
2696            .lines()
2697            .find_map(|line| line.strip_prefix("0::"))
2698            .ok_or_else(|| {
2699                io::Error::new(io::ErrorKind::Unsupported, "cgroup v2 is unavailable")
2700            })?;
2701        Ok(Path::new("/sys/fs/cgroup").join(relative.trim_start_matches('/')))
2702    }
2703
2704    #[cfg(target_os = "linux")]
2705    fn module_cgroup_directories() -> io::Result<BTreeSet<OsString>> {
2706        let modules = current_cgroup_path_for_test()?.join("subc-modules");
2707        let entries = match fs::read_dir(modules) {
2708            Ok(entries) => entries,
2709            Err(error) if error.kind() == io::ErrorKind::NotFound => return Ok(BTreeSet::new()),
2710            Err(error) => return Err(error),
2711        };
2712        let mut directories = BTreeSet::new();
2713        for entry in entries {
2714            let entry = entry?;
2715            if entry.file_type()?.is_dir() {
2716                directories.insert(entry.file_name());
2717            }
2718        }
2719        Ok(directories)
2720    }
2721
2722    #[cfg(target_os = "linux")]
2723    async fn wait_for_path(path: &Path, task: &JoinHandle<Result<(), BootstrapError>>) {
2724        let deadline = tokio::time::Instant::now() + Duration::from_secs(5);
2725        while !path.exists() && tokio::time::Instant::now() < deadline {
2726            assert!(
2727                !task.is_finished(),
2728                "daemon exited before creating {}",
2729                path.display()
2730            );
2731            sleep(Duration::from_millis(10)).await;
2732        }
2733        assert!(path.exists(), "daemon did not create {}", path.display());
2734    }
2735
2736    /// An in-process daemon with the default (disabled) cgroup placement must
2737    /// leave the host's live module cgroup tree exactly as it found it.
2738    /// Red-checking this test (making it fail) creates a
2739    /// `cgroup-isolation-probe-*` directory in the live cgroup tree, which must be removed by hand.
2740    #[cfg(target_os = "linux")]
2741    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
2742    async fn run_with_config_does_not_reconcile_the_ambient_cgroup_by_default() {
2743        let temp = unique_temp_dir("bootstrap-cgroup-default-disabled");
2744        let module_id = format!("cgroup-isolation-probe-{}", process::id());
2745        let before = module_cgroup_directories().expect("read ambient module cgroups before boot");
2746        assert!(
2747            !before.contains(&OsString::from(&module_id)),
2748            "isolation probe cgroup already exists before this daemon starts"
2749        );
2750        let capture = temp.join("logs").join(format!("{module_id}.stderr.log"));
2751        let module = ConfiguredModule {
2752            carry_containment: false,
2753            module_id,
2754            program: PathBuf::from("sh"),
2755            args: vec!["-c".to_string(), "sleep 30".to_string()],
2756            env: Vec::new(),
2757            log: None,
2758            enabled: true,
2759            registration_channel: false,
2760            reserved: false,
2761            reserved_prefixes: Vec::new(),
2762            protocol: ModuleProtocol::None,
2763            overlap: Default::default(),
2764            health: HealthConfig::default(),
2765            drain_timeout_ms: None,
2766            register_timeout: crate::supervise::DEFAULT_REGISTER_TIMEOUT,
2767            route_bind_relay_timeout_ms: None,
2768            restart: RestartPolicy::default(),
2769        };
2770        let config = config_with_a_short_socket(BootstrapConfig::new(temp.join("c"), 0))
2771            .with_configured_modules([module])
2772            .with_capture_logs_dir(temp.join("logs"))
2773            .with_terminal_journal_path(temp.join("terminals.jsonl"));
2774        let task = tokio::spawn(run_with_config(config));
2775
2776        wait_for_path(&capture, &task).await;
2777        let after = module_cgroup_directories().expect("read ambient module cgroups after boot");
2778
2779        task.abort();
2780        assert!(task
2781            .await
2782            .expect_err("aborted daemon task must cancel")
2783            .is_cancelled());
2784        assert_eq!(
2785            before, after,
2786            "an in-process daemon must not create or reconcile ambient module cgroups"
2787        );
2788    }
2789
2790    #[cfg(target_os = "linux")]
2791    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
2792    async fn explicit_cgroup_root_is_prepared_inside_the_fixture_tree() {
2793        let temp = unique_temp_dir("bootstrap-cgroup-explicit-root");
2794        let cgroup_root = temp.join("cgroup");
2795        fs::create_dir(&cgroup_root).expect("create scratch cgroup root");
2796        fs::write(cgroup_root.join("cgroup.procs"), b"").expect("write scratch cgroup marker");
2797        let modules = cgroup_root.join("subc-modules");
2798        let config = config_with_a_short_socket(BootstrapConfig::new(temp.join("c"), 0))
2799            .with_cgroup_placement(CgroupPlacementConfig::Root(cgroup_root))
2800            .with_terminal_journal_path(temp.join("terminals.jsonl"));
2801        let task = tokio::spawn(run_with_config(config));
2802
2803        wait_for_path(&modules, &task).await;
2804
2805        task.abort();
2806        assert!(task
2807            .await
2808            .expect_err("aborted daemon task must cancel")
2809            .is_cancelled());
2810        assert!(
2811            fs::read_dir(&modules)
2812                .expect("read prepared modules directory")
2813                .next()
2814                .is_none(),
2815            "the delegation probe must clean up after itself"
2816        );
2817    }
2818
2819    struct EnvGuard {
2820        key: &'static str,
2821        previous: Option<OsString>,
2822    }
2823
2824    impl EnvGuard {
2825        fn set(key: &'static str, value: &Path) -> Self {
2826            let previous = env::var_os(key);
2827            env::set_var(key, value);
2828            Self { key, previous }
2829        }
2830
2831        fn set_str(key: &'static str, value: &str) -> Self {
2832            let previous = env::var_os(key);
2833            env::set_var(key, value);
2834            Self { key, previous }
2835        }
2836
2837        fn unset(key: &'static str) -> Self {
2838            let previous = env::var_os(key);
2839            env::remove_var(key);
2840            Self { key, previous }
2841        }
2842    }
2843
2844    impl Drop for EnvGuard {
2845        fn drop(&mut self) {
2846            match &self.previous {
2847                Some(value) => env::set_var(self.key, value),
2848                None => env::remove_var(self.key),
2849            }
2850        }
2851    }
2852
2853    fn unique_temp_dir(name: &str) -> ScratchDir {
2854        ScratchDir::new(name)
2855    }
2856
2857    fn temp_connection_file_path(name: &str) -> (ScratchDir, PathBuf) {
2858        let dir = unique_temp_dir(name);
2859        let path = dir.join("conn.json");
2860        (dir, path)
2861    }
2862
2863    fn auth_for(info: &ConnectionInfo) -> ServerAuth {
2864        ServerAuth::new(info.key.clone(), info.daemon_id, info.daemon_ver.clone())
2865    }
2866
2867    fn start_server(bound: BoundDaemon) -> JoinHandle<Result<(), ServerError>> {
2868        let auth = auth_for(&bound.connection_info);
2869        tokio::spawn(serve_listeners(
2870            bound.listeners,
2871            Arc::new(Router::with_default_self_handler()),
2872            auth,
2873        ))
2874    }
2875
2876    fn expect_bound(outcome: Outcome) -> BoundDaemon {
2877        match outcome {
2878            Outcome::Bound(bound) => bound,
2879            Outcome::AlreadyRunning => panic!("fresh connection file unexpectedly had a daemon"),
2880        }
2881    }
2882
2883    async fn connect_from_info(conn: &ConnectionInfo) -> io::Result<TcpStream> {
2884        let endpoint = conn
2885            .endpoints
2886            .first()
2887            .expect("test connection file should have an endpoint");
2888        let ip: IpAddr = endpoint.host.parse().unwrap();
2889        TcpStream::connect(SocketAddr::new(ip, endpoint.port)).await
2890    }
2891
2892    fn make_connection_info(port: u16) -> ConnectionInfo {
2893        ConnectionInfo {
2894            schema: SCHEMA_VERSION,
2895            wire_version: Some(PROTOCOL_VERSION),
2896            endpoints: vec![Endpoint {
2897                host: "127.0.0.1".to_owned(),
2898                port,
2899            }],
2900            key: generate_key().unwrap(),
2901            daemon_id: generate_daemon_id().unwrap(),
2902            pid: process::id(),
2903            daemon_ver: "test-subc".to_owned(),
2904        }
2905    }
2906
2907    fn write_raw_owner_only_connection_file(path: &Path, contents: &[u8]) {
2908        fs::write(path, contents).unwrap();
2909        #[cfg(unix)]
2910        fs::set_permissions(path, fs::Permissions::from_mode(0o600)).unwrap();
2911    }
2912
2913    fn assert_owner_only_connection_file(path: &Path) {
2914        // `path` is only inspected on Unix (mode bits); on Windows the owner-only
2915        // guarantee comes from the inherited %TEMP% ACL, nothing to assert here.
2916        #[cfg(unix)]
2917        {
2918            let mode = fs::metadata(path).unwrap().permissions().mode() & 0o777;
2919            assert_eq!(mode, 0o600);
2920        }
2921        #[cfg(not(unix))]
2922        let _ = path;
2923    }
2924
2925    #[test]
2926    fn connection_file_path_uses_xdg_runtime_dir_when_set() {
2927        let _env_lock = ENV_LOCK.lock().unwrap();
2928        let runtime_dir = unique_temp_dir("xdg-runtime");
2929        let _xdg = EnvGuard::set("XDG_RUNTIME_DIR", runtime_dir.path());
2930
2931        assert_eq!(
2932            connection_file_path(),
2933            runtime_dir.join(CONNECTION_FILE_NAME)
2934        );
2935    }
2936
2937    #[test]
2938    fn connection_file_path_source_is_xdg_runtime_dir_when_set() {
2939        let runtime_dir = OsString::from("/run/user/1000");
2940
2941        let (path, source) = connection_file_path_with_source(Some(runtime_dir));
2942
2943        assert_eq!(
2944            path,
2945            PathBuf::from("/run/user/1000").join(CONNECTION_FILE_NAME)
2946        );
2947        assert_eq!(source, ConnectionFileSource::XdgRuntimeDir);
2948    }
2949
2950    #[test]
2951    fn connection_file_path_falls_back_to_temp_dir_with_user_token_when_xdg_unset() {
2952        let _env_lock = ENV_LOCK.lock().unwrap();
2953        let _xdg = EnvGuard::unset("XDG_RUNTIME_DIR");
2954
2955        assert_eq!(
2956            connection_file_path(),
2957            env::temp_dir().join(format!("subc-{}.connection.json", user_connection_token()))
2958        );
2959    }
2960
2961    #[test]
2962    fn connection_file_path_source_is_temp_dir_when_xdg_unset() {
2963        let (path, source) = connection_file_path_with_source(None);
2964
2965        assert_eq!(
2966            path,
2967            env::temp_dir().join(format!("subc-{}.connection.json", user_connection_token()))
2968        );
2969        assert_eq!(source, ConnectionFileSource::TempDirFallback);
2970    }
2971
2972    /// Concurrent callers must derive one token. The former temp-file uid probe
2973    /// could fail transiently (same-tick name collision, fd exhaustion) and send
2974    /// the loser down the env-derived fallback with a different identity for the
2975    /// same user; this fence keeps identity independent of filesystem luck.
2976    #[test]
2977    fn user_connection_token_is_stable_under_concurrent_callers() {
2978        let expected = user_connection_token();
2979        let workers: Vec<_> = (0..32)
2980            .map(|_| {
2981                std::thread::spawn(|| (0..40).map(|_| user_connection_token()).collect::<Vec<_>>())
2982            })
2983            .collect();
2984        for worker in workers {
2985            for token in worker.join().expect("probe thread") {
2986                assert_eq!(token, expected, "token diverged under concurrent probes");
2987            }
2988        }
2989    }
2990
2991    #[test]
2992    fn connection_file_path_source_is_temp_dir_when_xdg_empty() {
2993        let (path, source) = connection_file_path_with_source(Some(OsString::new()));
2994
2995        assert_eq!(
2996            path,
2997            env::temp_dir().join(format!("subc-{}.connection.json", user_connection_token()))
2998        );
2999        assert_eq!(source, ConnectionFileSource::TempDirFallback);
3000    }
3001
3002    /// The shipped binary is the only caller that journals terminal exits into
3003    /// the real run directory, so its constructor must supply that path itself;
3004    /// the in-process default leaves it unset.
3005    #[test]
3006    fn daemon_binary_config_journals_and_captures_into_the_run_dir() {
3007        let _env_lock = ENV_LOCK.lock().unwrap();
3008        let root = unique_temp_dir("daemon-binary-config");
3009        let data_home = root.join("data");
3010        let _data = EnvGuard::set("XDG_DATA_HOME", &data_home);
3011        let _config = EnvGuard::set("XDG_CONFIG_HOME", &root.join("config"));
3012        let _port = EnvGuard::unset(SUBC_PORT_ENV);
3013
3014        let config = BootstrapConfig::from_env_for_daemon_binary().unwrap();
3015
3016        let run_dir = data_home.join("cortexkit").join("run");
3017        assert_eq!(
3018            config.terminal_journal_path,
3019            Some(run_dir.join("terminals.jsonl"))
3020        );
3021        assert_eq!(config.capture_logs_dir, Some(run_dir.join("logs")));
3022        assert_eq!(
3023            BootstrapConfig::new(root.join("connection.json"), 0).terminal_journal_path,
3024            None,
3025            "an in-process config must not journal anywhere unless asked to"
3026        );
3027    }
3028
3029    #[test]
3030    fn daemon_binary_config_refuses_a_relative_data_home() {
3031        let _env_lock = ENV_LOCK.lock().unwrap();
3032        let root = unique_temp_dir("daemon-binary-relative-data");
3033        let _data = EnvGuard::set_str("XDG_DATA_HOME", "relative-data-home");
3034        let _config = EnvGuard::set("XDG_CONFIG_HOME", &root.join("config"));
3035        let _port = EnvGuard::unset(SUBC_PORT_ENV);
3036
3037        let error = BootstrapConfig::from_env_for_daemon_binary()
3038            .expect_err("a relative data home must refuse the daemon binary's config");
3039        assert!(
3040            matches!(error, BootstrapError::RunDir(_)),
3041            "expected a run-directory refusal, got {error}"
3042        );
3043        assert!(error.to_string().contains("XDG_DATA_HOME"), "{error}");
3044    }
3045
3046    #[test]
3047    fn privileged_operator_prompt_is_loaded_by_both_bootstrap_config_paths() {
3048        let _env_lock = ENV_LOCK.lock().unwrap();
3049        let _port = EnvGuard::unset(SUBC_PORT_ENV);
3050        let (_dir, conn_path) = temp_connection_file_path("privileged-operator-prompt");
3051        let config_path = conn_path.with_file_name("subc.jsonc");
3052        for (doc, expected) in [
3053            (None, daemon_config::PrivilegedOperatorPrompt::Log),
3054            (
3055                Some(r#"{"version":1}"#),
3056                daemon_config::PrivilegedOperatorPrompt::Log,
3057            ),
3058            (
3059                Some(r#"{"version":1,"privileged_operator_prompt":"log"}"#),
3060                daemon_config::PrivilegedOperatorPrompt::Log,
3061            ),
3062            (
3063                Some(r#"{"version":1,"privileged_operator_prompt":"enforce"}"#),
3064                daemon_config::PrivilegedOperatorPrompt::Enforce,
3065            ),
3066        ] {
3067            if let Some(doc) = doc {
3068                fs::write(&config_path, doc).unwrap();
3069            }
3070            assert_eq!(
3071                BootstrapConfig::from_env_with_daemon_config_path(&config_path)
3072                    .unwrap()
3073                    .privileged_operator_prompt,
3074                expected
3075            );
3076            assert_eq!(
3077                BootstrapConfig::new(&conn_path, 0)
3078                    .with_daemon_config_path(&config_path)
3079                    .unwrap()
3080                    .privileged_operator_prompt,
3081                expected
3082            );
3083        }
3084    }
3085
3086    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
3087    async fn privileged_operator_prompt_is_snapshotted_by_running_daemon() {
3088        let dir = unique_temp_dir("privileged-operator-prompt-bootstrap");
3089        let connection_path = dir.join("connection.json");
3090        let config_path = dir.join("subc.jsonc");
3091        fs::write(
3092            &config_path,
3093            r#"{"version":1,"privileged_operator_prompt":"enforce"}"#,
3094        )
3095        .unwrap();
3096        let config = config_with_a_short_socket(BootstrapConfig::new(&connection_path, 0))
3097            .with_daemon_config_path(&config_path)
3098            .unwrap();
3099        let connection_path = config.connection_file_path.clone();
3100        let daemon = AbortOnDrop::new(tokio::spawn(run_with_config(config)));
3101        tokio::time::timeout(Duration::from_secs(5), async {
3102            while !connection_path.exists() {
3103                assert!(
3104                    !daemon.handle.is_finished(),
3105                    "daemon exited before publishing"
3106                );
3107                tokio::task::yield_now().await;
3108            }
3109        })
3110        .await
3111        .unwrap();
3112        let info = connection_file::read(&connection_path).unwrap();
3113        let mut client = connect_from_info(&info).await.unwrap();
3114        authenticate_client(&mut client, &info, Duration::from_secs(5))
3115            .await
3116            .unwrap();
3117        for (correlation, mode, expected) in [
3118            (1, "enforce", Vec::new()),
3119            (2, "log", vec!["privileged_operator_prompt"]),
3120            (3, "log", vec!["privileged_operator_prompt"]),
3121        ] {
3122            fs::write(
3123                &config_path,
3124                serde_json::json!({"version":1,"privileged_operator_prompt":mode}).to_string(),
3125            )
3126            .unwrap();
3127            let frame = crate::Frame::build(
3128                subc_protocol::FrameType::Request,
3129                subc_protocol::Flags::new(false, subc_protocol::Priority::Interactive, false),
3130                0,
3131                0,
3132                correlation,
3133                serde_json::to_vec(&subc_control::ClientControlRequest::SupervisorRescan {
3134                    preview: true,
3135                })
3136                .unwrap(),
3137            )
3138            .unwrap();
3139            crate::write_frame(&mut client, &frame).await.unwrap();
3140            let reply =
3141                tokio::time::timeout(Duration::from_secs(5), crate::read_frame(&mut client))
3142                    .await
3143                    .unwrap()
3144                    .unwrap()
3145                    .unwrap();
3146            assert_eq!(reply.header.ty, subc_protocol::FrameType::Response);
3147            assert_eq!(reply.header.corr, correlation);
3148            let subc_control::ClientControlResponse::SupervisorRescan { result } =
3149                serde_json::from_slice(&reply.body).unwrap()
3150            else {
3151                panic!("expected a rescan response");
3152            };
3153            assert_eq!(result.restart_required, expected);
3154        }
3155    }
3156
3157    #[test]
3158    fn configured_port_uses_default_config_and_env_override() {
3159        let _env_lock = ENV_LOCK.lock().unwrap();
3160        let (_dir, conn_path) = temp_connection_file_path("daemon-config-port");
3161        let config_path = conn_path.with_file_name("subc.jsonc");
3162
3163        let _port = EnvGuard::unset(SUBC_PORT_ENV);
3164        assert_eq!(
3165            BootstrapConfig::from_env_with_daemon_config_path(&config_path)
3166                .unwrap()
3167                .port,
3168            DEFAULT_SUBC_PORT
3169        );
3170
3171        fs::write(&config_path, r#"{ "version": 1, "port": 8123 }"#).unwrap();
3172        assert_eq!(
3173            BootstrapConfig::from_env_with_daemon_config_path(&config_path)
3174                .unwrap()
3175                .port,
3176            8123
3177        );
3178
3179        let _port = EnvGuard::set_str(SUBC_PORT_ENV, "9012");
3180        assert_eq!(
3181            BootstrapConfig::from_env_with_daemon_config_path(&config_path)
3182                .unwrap()
3183                .port,
3184            9012
3185        );
3186    }
3187
3188    #[tokio::test]
3189    async fn singleton_does_not_accept_a_peer_that_answers_once_then_exits() {
3190        let (_dir, path) = temp_connection_file_path("departing-peer");
3191        let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
3192        let info = make_connection_info(listener.local_addr().unwrap().port());
3193        write_connection_file(&path, &info, None).unwrap();
3194        let peer = tokio::spawn(async move {
3195            let (mut stream, _) = listener.accept().await.unwrap();
3196            subc_transport::authenticate_server(
3197                &mut stream,
3198                &info.key,
3199                &info.daemon_id,
3200                &info.daemon_ver,
3201                PROBE_AUTH_DEADLINE,
3202            )
3203            .await
3204            .unwrap();
3205        });
3206        let outcome = ensure_singleton(&path, 0).await.unwrap();
3207        peer.await.unwrap();
3208        assert!(
3209            matches!(outcome, Outcome::Bound(_)),
3210            "a departing peer is not AlreadyRunning"
3211        );
3212    }
3213
3214    #[tokio::test]
3215    async fn second_singleton_probe_against_served_tcp_daemon_reports_already_running() {
3216        let (_dir, path) = temp_connection_file_path("already-running");
3217
3218        let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
3219        let server = start_server(bound);
3220
3221        let second = ensure_singleton(&path, 0).await.unwrap();
3222        assert!(matches!(second, Outcome::AlreadyRunning));
3223
3224        server.abort();
3225        let _ = server.await;
3226    }
3227
3228    #[tokio::test]
3229    async fn daemon_connection_file_publishes_protocol_wire_version() {
3230        let (_dir, path) = temp_connection_file_path("wire-version");
3231        let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
3232        assert_eq!(bound.connection_info.wire_version, Some(PROTOCOL_VERSION));
3233        assert_eq!(
3234            connection_file::read(&path).unwrap().wire_version,
3235            Some(PROTOCOL_VERSION)
3236        );
3237
3238        drop(bound.listeners);
3239    }
3240
3241    #[tokio::test]
3242    async fn stale_unbound_connection_file_is_reclaimed() {
3243        let (_dir, path) = temp_connection_file_path("stale-reclaim");
3244        let stale = TcpListener::bind((Ipv4Addr::LOCALHOST, 0)).await.unwrap();
3245        let stale_port = stale.local_addr().unwrap().port();
3246        drop(stale);
3247        let stale_info = make_connection_info(stale_port);
3248        write_atomic(&path, &stale_info).unwrap();
3249
3250        let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
3251        assert_ne!(bound.connection_info.key, stale_info.key);
3252        drop(bound.listeners);
3253    }
3254
3255    #[cfg(unix)]
3256    #[tokio::test]
3257    async fn ensure_singleton_reclaims_insecure_connection_file() {
3258        let (_dir, path) = temp_connection_file_path("insecure-reclaim");
3259        let stale = TcpListener::bind((Ipv4Addr::LOCALHOST, 0)).await.unwrap();
3260        let stale_port = stale.local_addr().unwrap().port();
3261        drop(stale);
3262        let stale_info = make_connection_info(stale_port);
3263        write_atomic(&path, &stale_info).unwrap();
3264        fs::set_permissions(&path, fs::Permissions::from_mode(0o644)).unwrap();
3265
3266        let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
3267        assert_ne!(bound.connection_info.key, stale_info.key);
3268        assert_ne!(bound.connection_info.daemon_id, stale_info.daemon_id);
3269        assert_owner_only_connection_file(&path);
3270
3271        drop(bound.listeners);
3272    }
3273
3274    #[tokio::test]
3275    async fn ensure_singleton_reclaims_non_loopback_connection_file() {
3276        let (_dir, path) = temp_connection_file_path("non-loopback-reclaim");
3277        let mut stale_info = make_connection_info(8757);
3278        stale_info.endpoints = vec![Endpoint {
3279            host: "192.0.2.10".to_owned(),
3280            port: 8757,
3281        }];
3282        write_atomic(&path, &stale_info).unwrap();
3283
3284        let config = config_with_a_short_socket(BootstrapConfig::new(&path, 0));
3285        let bound = expect_bound(ensure_singleton_with_config(config).await.unwrap());
3286        assert_ne!(bound.connection_info.key, stale_info.key);
3287        assert_ne!(bound.connection_info.daemon_id, stale_info.daemon_id);
3288        assert!(bound
3289            .connection_info
3290            .endpoints
3291            .iter()
3292            .all(|endpoint| endpoint.host.parse::<IpAddr>().unwrap().is_loopback()));
3293        assert_owner_only_connection_file(&path);
3294
3295        drop(bound.listeners);
3296    }
3297
3298    #[tokio::test]
3299    async fn ensure_singleton_reclaims_invalid_connection_file_shapes() {
3300        let mut unsupported_schema = make_connection_info(8757);
3301        unsupported_schema.schema = SCHEMA_VERSION + 1;
3302
3303        let mut empty_endpoints = make_connection_info(8757);
3304        empty_endpoints.endpoints.clear();
3305
3306        let mut short_key = make_connection_info(8757);
3307        short_key.key = vec![0x5A; MIN_KEY_LEN - 1];
3308
3309        let cases = vec![
3310            (
3311                "unsupported-schema",
3312                serde_json::to_vec(&unsupported_schema).unwrap(),
3313                Some(unsupported_schema),
3314            ),
3315            (
3316                "empty-endpoints",
3317                serde_json::to_vec(&empty_endpoints).unwrap(),
3318                Some(empty_endpoints),
3319            ),
3320            (
3321                "short-key",
3322                serde_json::to_vec(&short_key).unwrap(),
3323                Some(short_key),
3324            ),
3325            ("invalid-json", b"{not valid connection json".to_vec(), None),
3326        ];
3327
3328        for (label, contents, old_info) in cases {
3329            let (_dir, path) = temp_connection_file_path(label);
3330            write_raw_owner_only_connection_file(&path, &contents);
3331
3332            let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
3333            if let Some(old_info) = old_info {
3334                assert_ne!(bound.connection_info.key, old_info.key, "{label}");
3335                assert_ne!(
3336                    bound.connection_info.daemon_id, old_info.daemon_id,
3337                    "{label}"
3338                );
3339            }
3340            assert!(bound.connection_info.key.len() >= MIN_KEY_LEN, "{label}");
3341            assert_ne!(bound.connection_info.daemon_id, [0u8; 16], "{label}");
3342            assert_owner_only_connection_file(&path);
3343
3344            drop(bound.listeners);
3345        }
3346    }
3347
3348    #[tokio::test]
3349    async fn foreign_reused_port_connection_file_is_reclaimed_after_auth_probe_fails() {
3350        let (_dir, path) = temp_connection_file_path("foreign-reclaim");
3351        let foreign = TcpListener::bind((Ipv4Addr::LOCALHOST, 0)).await.unwrap();
3352        let foreign_port = foreign.local_addr().unwrap().port();
3353        write_atomic(&path, &make_connection_info(foreign_port)).unwrap();
3354        let foreign_task = tokio::spawn(async move {
3355            if let Ok((mut stream, _)) = foreign.accept().await {
3356                let mut buf = [0u8; 64];
3357                let _ = stream.read(&mut buf).await;
3358            }
3359        });
3360
3361        let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
3362        assert!(bound
3363            .connection_info
3364            .endpoints
3365            .iter()
3366            .all(|endpoint| endpoint.port != foreign_port));
3367
3368        drop(bound.listeners);
3369        let _ = foreign_task.await;
3370    }
3371
3372    #[tokio::test]
3373    async fn stale_start_lock_file_is_reclaimable() {
3374        let (_dir, path) = temp_connection_file_path("start-lock-stale-file");
3375        let lock_path = start_lock_path(&path);
3376        drop(open_owner_only_lock(&lock_path).unwrap());
3377        assert!(lock_path.is_file());
3378
3379        let lock = StartLock::acquire(&path).await.unwrap();
3380        assert!(lock_path.is_file());
3381
3382        drop(lock);
3383        assert!(lock_path.is_file());
3384    }
3385
3386    #[tokio::test]
3387    async fn held_start_lock_blocks_second_acquire_until_release() {
3388        let (_dir, path) = temp_connection_file_path("start-lock-held");
3389        let lock_path = start_lock_path(&path);
3390        let first = StartLock::acquire(&path).await.unwrap();
3391
3392        let err = match StartLock::acquire(&path).await {
3393            Ok(_) => panic!("second acquire while held must stay busy"),
3394            Err(err) => err,
3395        };
3396        assert!(matches!(
3397            err,
3398            BootstrapError::StartLockBusy {
3399                ref path,
3400                attempts: START_LOCK_RETRIES,
3401            } if path == &lock_path
3402        ));
3403
3404        drop(first);
3405
3406        let second = StartLock::acquire(&path)
3407            .await
3408            .expect("released advisory lock should be reclaimable");
3409        drop(second);
3410    }
3411
3412    #[tokio::test]
3413    async fn bind_conflict_on_fixed_port_fails_loud_without_reselecting() {
3414        let (_dir, path) = temp_connection_file_path("bind-conflict");
3415        let occupied = TcpListener::bind((Ipv4Addr::LOCALHOST, 0)).await.unwrap();
3416        let occupied_port = occupied.local_addr().unwrap().port();
3417
3418        let err = ensure_singleton(&path, occupied_port).await.unwrap_err();
3419        assert!(matches!(
3420            err,
3421            BootstrapError::Bind { ref source, .. } if source.kind() == io::ErrorKind::AddrInUse
3422        ));
3423        assert!(err.to_string().contains("set the port in config"));
3424
3425        drop(occupied);
3426    }
3427
3428    #[tokio::test]
3429    async fn ephemeral_ipv6_collision_still_serves_ipv4() {
3430        let (listeners, endpoints) = bind_loopback_with(0, |port| async move {
3431            let occupied = TcpListener::bind((Ipv6Addr::LOCALHOST, port)).await?;
3432            let result = TcpListener::bind((Ipv6Addr::LOCALHOST, port)).await;
3433            drop(occupied);
3434            result
3435        })
3436        .await
3437        .expect("an incidental IPv6 collision must not prevent an ephemeral daemon from starting");
3438        assert_eq!(endpoints.len(), 1);
3439        assert_eq!(endpoints[0].host, "127.0.0.1");
3440        TcpStream::connect((Ipv4Addr::LOCALHOST, endpoints[0].port))
3441            .await
3442            .unwrap();
3443        drop(listeners);
3444    }
3445
3446    #[tokio::test]
3447    async fn unpublished_boot_holds_singleton_lock_until_publication() {
3448        let dir = ScratchDir::new("ul");
3449        let path = dir.join("c");
3450        let config = config_with_a_short_socket(BootstrapConfig::new(&path, 0));
3451        let bound = expect_bound(ensure_singleton_inner(config, false).await.unwrap());
3452        assert!(!path.exists());
3453        let contender = open_owner_only_lock(&start_lock_path(&path)).unwrap();
3454        assert!(
3455            matches!(FileExt::try_lock(&contender), Err(TryLockError::WouldBlock)),
3456            "an unpublished boot must keep its singleton claim"
3457        );
3458        drop(bound);
3459        FileExt::try_lock(&contender).unwrap();
3460    }
3461
3462    #[tokio::test]
3463    async fn start_lock_drop_releases_claim_with_a_shared_file_description() {
3464        let dir = ScratchDir::new("sl");
3465        let path = dir.join("c");
3466        let config = config_with_a_short_socket(BootstrapConfig::new(&path, 0));
3467        let bound = expect_bound(ensure_singleton_inner(config, false).await.unwrap());
3468        // try_clone shares the open file description just as fork does,
3469        // without relying on the timing of another test's child spawn.
3470        let inherited = bound
3471            .publication_lock
3472            .as_ref()
3473            .unwrap()
3474            ._file
3475            .try_clone()
3476            .unwrap();
3477        let contender = open_owner_only_lock(&start_lock_path(&path)).unwrap();
3478        assert!(matches!(
3479            FileExt::try_lock(&contender),
3480            Err(TryLockError::WouldBlock)
3481        ));
3482        drop(bound);
3483        FileExt::try_lock(&contender).expect("dropping bootstrap releases even an inherited lock");
3484        drop(inherited);
3485    }
3486
3487    #[tokio::test]
3488    async fn readiness_probe_requires_an_authenticating_server() {
3489        let dir = ScratchDir::new("unserved-readiness");
3490        let bound = expect_bound(
3491            ensure_singleton_inner(BootstrapConfig::new(dir.join("connection.json"), 0), false)
3492                .await
3493                .unwrap(),
3494        );
3495        assert!(matches!(
3496            verify_serving(&bound.connection_info).await,
3497            Err(BootstrapError::StartupNotServing)
3498        ));
3499    }
3500
3501    #[cfg(unix)]
3502    #[tokio::test]
3503    async fn orphan_sweep_finishes_before_connection_file_publication() {
3504        use crate::live_children::{ExecutableIdentity, LiveChild};
3505        let dir = ScratchDir::new("pas");
3506        // A one-letter connection file name, and no long socket path: the sweep
3507        // this test covers runs before the local endpoint is bound, so a longer
3508        // name would only add a way for the test to fail on something else.
3509        let path = dir.join("c");
3510        let record = dir.join("live-children.json");
3511        let ready = dir.join("ready");
3512        let mut child = std::process::Command::new("/bin/sh")
3513            .args([
3514                "-c",
3515                "trap '' TERM; : > \"$0\"; read -r _",
3516                ready.to_str().unwrap(),
3517            ])
3518            .stdin(std::process::Stdio::piped())
3519            .env("XDG_DATA_HOME", dir.path())
3520            .env("XDG_RUNTIME_DIR", dir.path())
3521            .env("XDG_CONFIG_HOME", dir.path())
3522            .spawn()
3523            .unwrap();
3524        timeout(Duration::from_secs(5), async {
3525            while !ready.exists() {
3526                sleep(Duration::from_millis(5)).await;
3527            }
3528        })
3529        .await
3530        .unwrap();
3531        let observed = subc_os::Process::open(child.id())
3532            .unwrap()
3533            .unwrap()
3534            .observe()
3535            .unwrap();
3536        crate::live_children::write_record(
3537            &record,
3538            &[LiveChild {
3539                container: None,
3540                module_id: "old-sleep".into(),
3541                pid: child.id(),
3542                protocol: subc_control::ModuleProtocol::None,
3543                start_time: Some(observed.start_time),
3544                executable: observed.executable.map(ExecutableIdentity::from),
3545                cgroup_name: None,
3546            }],
3547        )
3548        .unwrap();
3549        let config = config_with_a_short_socket(BootstrapConfig::new(&path, 0))
3550            .with_live_children_record(&record);
3551        let daemon = tokio::spawn(run_with_config(config));
3552        timeout(Duration::from_secs(5), async {
3553            while !path.exists() {
3554                sleep(Duration::from_millis(5)).await;
3555            }
3556        })
3557        .await
3558        .unwrap();
3559        assert!(path.exists());
3560        let ended = child.try_wait().unwrap().is_some();
3561        if !ended {
3562            child.kill().unwrap();
3563        }
3564        child.wait().unwrap();
3565        if ended {
3566            assert!(matches!(probe_existing(&path).await.unwrap(), Probe::Live));
3567        }
3568        daemon.abort();
3569        let _ = daemon.await;
3570        assert!(
3571            ended,
3572            "the published daemon must not still owe its orphan sweep"
3573        );
3574    }
3575
3576    /// Publishing the connection file fsyncs, which can take seconds on a busy
3577    /// disk, so the daemon must do it off the async worker thread. The hook
3578    /// blocks inside the write until another task on the same single-threaded
3579    /// runtime has run. Writing on the worker thread would starve that task: the
3580    /// hook then gives up after its deadline and the test fails, as it does if
3581    /// startup publishes without going through the hooked write at all.
3582    #[tokio::test(flavor = "current_thread")]
3583    async fn connection_file_publication_leaves_the_async_runtime_free() {
3584        use std::sync::atomic::{AtomicBool, Ordering};
3585
3586        let (_dir, path) = temp_connection_file_path("publish-off-runtime");
3587        let write_started = Arc::new(AtomicBool::new(false));
3588        let (progress_tx, progress_rx) = std::sync::mpsc::channel::<()>();
3589        let (verdict_tx, verdict_rx) = std::sync::mpsc::channel::<bool>();
3590        let hook_started = Arc::clone(&write_started);
3591        connection_file_write_hook::install(&path, move || {
3592            hook_started.store(true, Ordering::SeqCst);
3593            let progressed = progress_rx.recv_timeout(Duration::from_secs(5)).is_ok();
3594            let _ = verdict_tx.send(progressed);
3595        });
3596        // Signals only after the write has started, so a message can only come
3597        // from this task running while the write is still blocked.
3598        let ticker = tokio::spawn(async move {
3599            loop {
3600                if write_started.load(Ordering::SeqCst) {
3601                    let _ = progress_tx.send(());
3602                    return;
3603                }
3604                sleep(Duration::from_millis(5)).await;
3605            }
3606        });
3607
3608        let daemon = tokio::spawn(run_with_config(BootstrapConfig::new(&path, 0)));
3609        let published = timeout(Duration::from_secs(15), async {
3610            while !path.exists() && !daemon.is_finished() {
3611                sleep(Duration::from_millis(5)).await;
3612            }
3613        })
3614        .await;
3615        let daemon_exited = daemon.is_finished();
3616        daemon.abort();
3617        let daemon_result = daemon.await;
3618        ticker.abort();
3619
3620        assert!(
3621            published.is_ok() && !daemon_exited,
3622            "the daemon must start and publish its connection file: {daemon_result:?}"
3623        );
3624        match verdict_rx.try_recv() {
3625            Ok(true) => {}
3626            Ok(false) => panic!(
3627                "no other task on the runtime ran while the connection file was being written: the write is blocking the async worker thread"
3628            ),
3629            Err(_) => panic!("startup published the connection file without the hooked write"),
3630        }
3631    }
3632
3633    /// A write that never reports back must stop startup with an error naming
3634    /// the connection file, never be taken as published.
3635    #[tokio::test]
3636    async fn a_panicking_connection_file_write_fails_startup() {
3637        let (_dir, path) = temp_connection_file_path("publish-panics");
3638        connection_file_write_hook::install(&path, || {
3639            panic!("deliberate panic inside the connection-file write")
3640        });
3641
3642        let error = timeout(
3643            Duration::from_secs(15),
3644            run_with_config(BootstrapConfig::new(&path, 0)),
3645        )
3646        .await
3647        .expect("startup must end once the write task fails")
3648        .expect_err("startup must fail when the connection-file write panics");
3649
3650        assert!(
3651            matches!(&error, BootstrapError::ConnectionFileWrite { path: failed, .. } if failed == &path),
3652            "expected a connection-file write error, got {error:?}"
3653        );
3654        let message = error.to_string();
3655        assert!(message.contains(&path.display().to_string()), "{message}");
3656        assert!(message.contains("panicked"), "{message}");
3657        assert!(!path.exists(), "no connection file may be published");
3658    }
3659
3660    #[cfg(target_os = "macos")]
3661    #[test]
3662    fn nofile_raise_worker() {
3663        let Some(path) = std::env::var_os("SUBC_NOFILE_TEST_FILE") else {
3664            return;
3665        };
3666        let (_, hard) = rlimit::Resource::NOFILE.get().unwrap();
3667        rlimit::Resource::NOFILE.set(256, hard).unwrap();
3668        raise_nofile_limit_with_ceiling(Some(10240));
3669        std::fs::write(path, rlimit::Resource::NOFILE.get().unwrap().0.to_string()).unwrap();
3670    }
3671
3672    #[cfg(target_os = "macos")]
3673    #[test]
3674    fn macos_nofile_raise_clamps_to_kernel_ceiling() {
3675        let dir = ScratchDir::new("nofile-ceiling");
3676        let result = dir.join("soft");
3677        let status = std::process::Command::new(std::env::current_exe().unwrap())
3678            .args(["--exact", "bootstrap::tests::nofile_raise_worker"])
3679            .env("SUBC_NOFILE_TEST_FILE", &result)
3680            .env("XDG_DATA_HOME", dir.path())
3681            .env("XDG_RUNTIME_DIR", dir.path())
3682            .env("XDG_CONFIG_HOME", dir.path())
3683            .status()
3684            .unwrap();
3685        assert!(status.success());
3686        let soft: u64 = std::fs::read_to_string(result).unwrap().parse().unwrap();
3687        assert!(
3688            soft > 256 && soft <= 10240,
3689            "the inherited limit must be raised within the kernel ceiling; got {soft}"
3690        );
3691    }
3692
3693    #[tokio::test]
3694    async fn key_rotation_republishes_new_material_and_old_file_fails_auth() {
3695        let (_dir, path) = temp_connection_file_path("key-rotation");
3696        let first = expect_bound(ensure_singleton(&path, 0).await.unwrap());
3697        let old_info = first.connection_info.clone();
3698        let fixed_port = old_info.endpoints[0].port;
3699        drop(first.listeners);
3700
3701        let second = expect_bound(ensure_singleton(&path, fixed_port).await.unwrap());
3702        let new_info = second.connection_info.clone();
3703        assert_ne!(old_info.key, new_info.key);
3704        assert_ne!(old_info.daemon_id, new_info.daemon_id);
3705        let server = start_server(second);
3706
3707        let mut old_stream = connect_from_info(&old_info).await.unwrap();
3708        let old_auth = authenticate_client(&mut old_stream, &old_info, PROBE_AUTH_DEADLINE).await;
3709        assert!(
3710            old_auth.is_err(),
3711            "old key must not authenticate after restart"
3712        );
3713
3714        let reread = connection_file::read(&path).unwrap();
3715        let mut new_stream = connect_from_info(&reread).await.unwrap();
3716        authenticate_client(&mut new_stream, &reread, PROBE_AUTH_DEADLINE)
3717            .await
3718            .unwrap();
3719
3720        server.abort();
3721        let _ = server.await;
3722    }
3723
3724    #[cfg(unix)]
3725    #[tokio::test]
3726    async fn published_connection_file_permissions_are_owner_only() {
3727        let (_dir, path) = temp_connection_file_path("permissions");
3728        let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
3729
3730        let mode = fs::metadata(&path).unwrap().permissions().mode() & 0o777;
3731        assert_eq!(mode, 0o600);
3732
3733        drop(bound.listeners);
3734    }
3735
3736    /// A config with the socket in `dir`'s own short subdirectory.
3737    ///
3738    /// A Unix socket path is limited to just over a hundred bytes, and this
3739    /// harness's temporary directories use nearly all of that, so a test that
3740    /// boots a real daemon passes a short directory and a one-letter connection
3741    /// file name; anything longer makes the bind refuse. On Windows there is no
3742    /// socket path to keep short, so the config is returned unchanged.
3743    fn config_with_a_short_socket(config: BootstrapConfig) -> BootstrapConfig {
3744        let dir = config
3745            .connection_file_path
3746            .parent()
3747            .map(Path::to_path_buf)
3748            .unwrap_or_else(|| PathBuf::from("."));
3749        #[cfg(unix)]
3750        let config = BootstrapConfig {
3751            connection_file_path: dir.join("c"),
3752            local_listener: LocalListenerConfig::in_socket_dir(dir.join("s")),
3753            ..config
3754        };
3755        #[cfg(not(unix))]
3756        let _ = dir;
3757        config
3758    }
3759
3760    /// The shipped binary puts the socket in its own run directory. The file is
3761    /// named from the connection file's path, so a restart finds it again.
3762    #[cfg(unix)]
3763    #[test]
3764    fn the_daemon_binary_puts_its_socket_in_the_run_directory() {
3765        let _env_lock = ENV_LOCK.lock().unwrap();
3766        let root = ScratchDir::new("binary-local-socket");
3767        let data_home = root.join("data");
3768        let _data = EnvGuard::set("XDG_DATA_HOME", &data_home);
3769        let _config = EnvGuard::set("XDG_CONFIG_HOME", &root.join("config"));
3770        let _port = EnvGuard::unset(SUBC_PORT_ENV);
3771
3772        let config = BootstrapConfig::from_env_for_daemon_binary().unwrap();
3773
3774        let run_dir = data_home.join("cortexkit").join("run");
3775        assert_eq!(
3776            config.local_listener.configured_dir(),
3777            Some(run_dir.as_path()),
3778            "the binary serves from its run directory"
3779        );
3780        let socket = local_listener::socket_path_for(Some(&run_dir), &config.connection_file_path)
3781            .expect("the binary config names a directory");
3782        assert_eq!(socket.parent().unwrap(), run_dir);
3783        assert_eq!(
3784            socket.file_name().unwrap().to_string_lossy(),
3785            "subc.sock",
3786            "a fixed name: the run directory belongs to one data home"
3787        );
3788    }
3789
3790    /// An in-process daemon that names no directory serves TCP only: it publishes
3791    /// no local endpoint, creates nothing under the data home, and says nothing is
3792    /// unavailable, because nothing was asked for.
3793    /// The environment lock covers only the setup: the daemon under test reads
3794    /// its data home from the process environment while it boots, and holding
3795    /// the lock across that await would block every other test that sets one.
3796    #[cfg(unix)]
3797    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
3798    async fn an_in_process_daemon_without_a_configured_directory_serves_no_local_endpoint() {
3799        let root = ScratchDir::new("in-process-no-local-socket");
3800        let data_home = root.join("data");
3801        fs::create_dir_all(&data_home).unwrap();
3802        let connection_path = root.join("c");
3803        let _data;
3804        let _config;
3805        let _port;
3806        {
3807            let _env_lock = ENV_LOCK.lock().unwrap();
3808            _data = EnvGuard::set("XDG_DATA_HOME", &data_home);
3809            _config = EnvGuard::set("XDG_CONFIG_HOME", &root.join("config"));
3810            _port = EnvGuard::unset(SUBC_PORT_ENV);
3811        }
3812
3813        let bound = expect_bound(ensure_singleton(&connection_path, 0).await.unwrap());
3814
3815        assert!(bound.local_endpoint.is_none());
3816        assert!(bound.local_endpoint_unavailable.is_none());
3817        let published = connection_file::read_with_local_endpoint(&connection_path).unwrap();
3818        assert!(published.local_endpoint.is_none());
3819        assert!(
3820            local_listener::socket_path_for(None, &connection_path).is_none(),
3821            "no configured directory, no socket"
3822        );
3823        let mut leftovers = Vec::new();
3824        let mut pending = vec![data_home.clone()];
3825        while let Some(dir) = pending.pop() {
3826            for entry in fs::read_dir(&dir).unwrap() {
3827                let entry = entry.unwrap();
3828                let path = entry.path();
3829                if entry.file_type().unwrap().is_dir() {
3830                    pending.push(path.clone());
3831                }
3832                if entry.file_name().to_string_lossy().ends_with(".sock") {
3833                    leftovers.push(path);
3834                }
3835            }
3836        }
3837        assert!(
3838            leftovers.is_empty(),
3839            "an in-process daemon must not create a socket in the data home: {leftovers:?}"
3840        );
3841    }
3842
3843    /// A local endpoint that cannot be bound does not stop the daemon: it keeps
3844    /// serving TCP, publishes no local endpoint, and reports the reason where an
3845    /// operator reads daemon facts. Every module stays up for a feature nothing
3846    /// has to use.
3847    #[cfg(unix)]
3848    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
3849    async fn a_daemon_whose_socket_path_is_too_long_still_serves_tcp() {
3850        let dir = ScratchDir::new("tcp-only");
3851        let canonical = std::fs::canonicalize(&dir).unwrap();
3852        let connection_path = canonical.join("c");
3853        // The socket directory is deep enough that the socket's path is over the
3854        // Unix socket limit, so binding it must fail.
3855        let deep = canonical.join("x".repeat(80));
3856        fs::create_dir(&deep).unwrap();
3857        fs::set_permissions(&deep, fs::Permissions::from_mode(0o700)).unwrap();
3858        let config =
3859            BootstrapConfig::new(&connection_path, 0).with_local_listener_socket_dir(&deep);
3860        let bound = expect_bound(ensure_singleton_with_config(config).await.unwrap());
3861        assert!(
3862            bound.local_endpoint.is_none(),
3863            "a socket that cannot be bound is not published"
3864        );
3865        assert!(
3866            bound
3867                .local_endpoint_unavailable
3868                .as_deref()
3869                .is_some_and(|reason| reason.contains("too long")),
3870            "{:?}",
3871            bound.local_endpoint_unavailable
3872        );
3873        let published = connection_file::read_with_local_endpoint(&connection_path).unwrap();
3874        assert!(published.local_endpoint.is_none());
3875        assert!(
3876            !published.info.endpoints.is_empty(),
3877            "TCP is still published"
3878        );
3879
3880        // The daemon serves TCP as usual.
3881        let server = start_server(bound);
3882        let info = connection_file::read(&connection_path).unwrap();
3883        let mut client = connect_from_info(&info).await.unwrap();
3884        authenticate_client(&mut client, &info, PROBE_AUTH_DEADLINE)
3885            .await
3886            .expect("a TCP client still completes the handshake");
3887        server.abort();
3888        let _ = server.await;
3889    }
3890
3891    /// The published file names the local endpoint the daemon actually serves,
3892    /// and a client that trusts the file completes the handshake over it.
3893    #[cfg(unix)]
3894    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
3895    async fn published_connection_file_names_the_served_local_endpoint() {
3896        let (_dir, path) = temp_connection_file_path("local-endpoint");
3897        // The socket goes in the system temporary directory rather than inside
3898        // the harness's own directory, whose long lockdown-sandbox path leaves
3899        // no room inside the Unix socket path limit.
3900        let short_socket_dir = std::env::temp_dir().join(format!("subc-bt-{}", std::process::id()));
3901        let config =
3902            BootstrapConfig::new(&path, 0).with_local_listener_socket_dir(&short_socket_dir);
3903        let bound = expect_bound(ensure_singleton_with_config(config).await.unwrap());
3904        let endpoint = connection_file::read_with_local_endpoint(&path)
3905            .unwrap()
3906            .local_endpoint
3907            .expect("the daemon serves a local endpoint, so the file names it");
3908        assert_eq!(Some(&endpoint), bound.local_endpoint.as_ref());
3909        let LocalEndpoint::Unix { path: socket_path } = &endpoint else {
3910            panic!("a Unix daemon publishes a socket path");
3911        };
3912        assert_eq!(socket_path.parent().unwrap(), short_socket_dir);
3913        assert_eq!(
3914            fs::metadata(socket_path).unwrap().permissions().mode() & 0o777,
3915            0o600
3916        );
3917
3918        let router = Arc::new(Router::with_default_self_handler());
3919        let auth = auth_for(&bound.connection_info);
3920        let server = tokio::spawn(serve_local_listener(
3921            bound.local_listener.expect("the local listener is bound"),
3922            router,
3923            auth,
3924        ));
3925        let mut client = tokio::net::UnixStream::connect(socket_path).await.unwrap();
3926        authenticate_client(&mut client, &bound.connection_info, Duration::from_secs(5))
3927            .await
3928            .expect("a socket client completes the same handshake as a TCP client");
3929
3930        server.abort();
3931        drop(bound.listeners);
3932    }
3933
3934    /// A configured directory that is not a directory refuses at bind rather
3935    /// than publishing an endpoint nothing serves.
3936    #[cfg(unix)]
3937    #[test]
3938    fn a_socket_directory_that_is_a_file_refuses() {
3939        let dir = ScratchDir::new("local-listener-file-as-dir");
3940        let file = dir.join("not-a-directory");
3941        fs::write(&file, b"x").unwrap();
3942        let error = local_listener::bind(
3943            &LocalListenerConfig::in_socket_dir(file),
3944            &dir.join("connection.json"),
3945        )
3946        .expect_err("a path that is not a directory cannot hold the socket");
3947        assert!(matches!(error, ServerError::LocalAccept { .. }), "{error}");
3948    }
3949
3950    /// The socket sits directly in the configured directory under one fixed
3951    /// name, and a directory that would put the path over the platform limit is
3952    /// refused rather than shortened around.
3953    #[cfg(unix)]
3954    #[test]
3955    fn the_socket_path_stays_within_the_unix_limit() {
3956        let run_dir = PathBuf::from("/tmp/short-run");
3957        assert!(
3958            local_listener::socket_path_for(None, &run_dir.join("one.json")).is_none(),
3959            "no configured directory means no socket"
3960        );
3961        let socket = local_listener::socket_path_for(Some(&run_dir), &run_dir.join("one.json"))
3962            .expect("a configured directory gets a socket");
3963        assert_eq!(socket, run_dir.join("subc.sock"));
3964        assert_eq!(
3965            socket,
3966            local_listener::socket_path_for(Some(&run_dir), &run_dir.join("two.json")).unwrap(),
3967            "the name does not come from the connection file"
3968        );
3969
3970        let deep = PathBuf::from(format!("/tmp/{}", "x".repeat(60)));
3971        let long = local_listener::socket_path_for(Some(&deep), &deep.join("conn.json"))
3972            .expect("a configured directory gets a socket");
3973        assert!(
3974            long.as_os_str().len() < 104,
3975            "{} is {} bytes",
3976            long.display(),
3977            long.as_os_str().len()
3978        );
3979    }
3980}