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        control = control.with_supervisor_rescan_entries(&configured_modules);
1393    }
1394    #[cfg(all(test, feature = "test-support"))]
1395    let control = if let Some(adapt) = test_control {
1396        // A test hook is not a substitute for the executable-name guard.
1397        let executable =
1398            std::env::current_exe().map_err(|e| BootstrapError::ConfigStore(e.to_string()))?;
1399        if !executable
1400            .file_name()
1401            .is_some_and(|name| name.to_string_lossy().starts_with("ckdev-"))
1402        {
1403            return Err(BootstrapError::ConfigStore(
1404                "test control requires a ckdev executable".into(),
1405            ));
1406        }
1407        adapt(control)
1408    } else {
1409        control
1410    };
1411    let control = Arc::new(control);
1412    let router = Arc::new(Router::with_control_handler(Arc::clone(&control)));
1413    let auth = ServerAuth::new(
1414        bound.connection_info.key.clone(),
1415        bound.connection_info.daemon_id,
1416        bound.connection_info.daemon_ver.clone(),
1417    )
1418    .with_connected_clients(connected_clients);
1419
1420    // Discovery must expose every accepted configured module, including disabled
1421    // ones. Registration creates status handles but neither starts monitors
1422    // nor launches processes: children need the published file for `--subc`.
1423    let mut registered_modules = Vec::with_capacity(configured_modules.len());
1424    for configured in configured_modules {
1425        let enabled = configured.enabled;
1426        let health = configured.health.clone();
1427        let module_id = configured.module_id.clone();
1428        let registered = match supervisor.register_configured_with_health(
1429            configured.module_spec(),
1430            enabled,
1431            health.clone(),
1432            configured.drain_timeout_ms,
1433            configured.restart,
1434            configured.register_timeout,
1435        ) {
1436            Ok(registered) => registered,
1437            Err(err) => {
1438                error!(module_id = %module_id, error = %err, "failed to supervise configured module; continuing daemon startup");
1439                continue;
1440            }
1441        };
1442        registered_modules.push(registered);
1443        // A relaxed threshold can silently outlive the temporary allowance
1444        // that required it, so report the extra tolerance on every boot.
1445        let default_threshold = HealthConfig::default().failure_threshold;
1446        if enabled && health.failure_threshold > default_threshold {
1447            warn!(
1448                module_id = %module_id,
1449                failure_threshold = health.failure_threshold,
1450                default_threshold,
1451                tolerance_secs = health.cadence.as_secs() * u64::from(health.failure_threshold),
1452                "health failure threshold is relaxed above the default; a wedged module stays unflagged for longer"
1453            );
1454        }
1455        info!(module_id = %module_id, enabled, "configured module registered");
1456    }
1457    control.refresh_capability_requirements();
1458
1459    // The local endpoint is served beside TCP and answers the same handshake.
1460    // It serves clients only, which the server enforces per connection.
1461    let local_serve_task = bound.local_listener.map(|local_listener| {
1462        AbortOnDrop::new(tokio::spawn(serve_local_listener(
1463            local_listener,
1464            Arc::clone(&router),
1465            auth.clone(),
1466        )))
1467    });
1468    // On Unix the task is dropped below to stop accepting at shutdown. Windows
1469    // leaves it to the process exit; naming it keeps the binding uniform.
1470    #[cfg(windows)]
1471    let _local_serve_task = local_serve_task;
1472    let mut serve_task =
1473        AbortOnDrop::new(tokio::spawn(serve_listeners(bound.listeners, router, auth)));
1474    verify_serving(&bound.connection_info).await?;
1475    // A published endpoint promises that authentication can complete now.
1476    // Orphan cleanup and control-plane construction can take longer than an
1477    // existing-daemon probe's deadline, so finish both before discovery.
1478    // The write fsyncs, so it runs on tokio's blocking pool (see
1479    // `publish_connection_file`). The singleton lock travels with it and comes
1480    // back only once the file is in place.
1481    let (published, publication_lock) = publish_connection_file(
1482        &bound.connection_file_path,
1483        #[cfg(windows)]
1484        bound.connection_file_copy.as_deref(),
1485        &bound.connection_info,
1486        bound.local_endpoint.as_ref(),
1487        publication_lock,
1488    )
1489    .await;
1490    published?;
1491    // Serving and discoverable: a daemon booted from a restart intent records
1492    // readiness, which makes that generation the daemon's active one. Without
1493    // this record the next start treats the warm-up as failed.
1494    if let Some(owner) = daemon_ready_runtime {
1495        if let Err(error) = owner.on_blocking(|owner| owner.record_daemon_ready()).await {
1496            warn!(%error, "daemon readiness was not recorded in the approved store");
1497        }
1498    }
1499    #[cfg(test)]
1500    startup_publication_hook::run(&bound.connection_file_path).await;
1501    drop(publication_lock);
1502    let _clock_step_task = AbortOnDrop::new(crate::watchdog::spawn_clock_step_monitor());
1503    // Off the startup path: it runs `systemctl`, and only ever logs.
1504    #[cfg(target_os = "linux")]
1505    let _kill_mode_check = AbortOnDrop::new(tokio::spawn(
1506        crate::systemd_kill_mode::warn_if_kill_mode_defeats_ordered_shutdown(),
1507    ));
1508    let _watchdog_task = AbortOnDrop::new(
1509        DaemonSelfWatchdog::new(
1510            bound.connection_info.clone(),
1511            bound.connection_file_path.clone(),
1512        )
1513        .with_config(watchdog_config)
1514        .spawn(),
1515    );
1516
1517    for registered in registered_modules {
1518        match supervisor
1519            .start_registered_configured_async(registered, None, true)
1520            .await
1521        {
1522            Ok(module) => {
1523                info!(module_id = %module.module_id(), "configured module supervised");
1524            }
1525            Err(err) => {
1526                error!(error = %err, "failed to start configured module; continuing daemon startup");
1527            }
1528        }
1529    }
1530
1531    control.refresh_capability_requirements();
1532    Arc::clone(&control).spawn_capability_deadline_loop();
1533    Arc::clone(&control).spawn_scope_expiry_loop();
1534    let _config_delivery_task = control.spawn_config_delivery_worker().map(AbortOnDrop::new);
1535    control.spawn_resource_sampler();
1536
1537    #[cfg(unix)]
1538    {
1539        tokio::select! {
1540            result = serve_task.join() => {
1541                return result.map_err(BootstrapError::ServeJoin)?.map_err(BootstrapError::Serve);
1542            }
1543            _ = terminate.recv() => {}
1544        }
1545        // First, before the notice, the drain, or any connection close: from
1546        // here on a module exit is recorded as `daemon_shutdown` and never
1547        // respawned. Also before allowing a second signal to cut the bounded
1548        // wait short, so the journal marker is always written.
1549        supervisor.begin_daemon_shutdown();
1550        // Stop the self-watchdog before closing the listener. Its next tick would
1551        // connect to that listener, fail, and log an ERROR indistinguishable from
1552        // a wedged daemon, once for every interval a planned stop lasts.
1553        drop(_watchdog_task);
1554        // Dropping the listener stops new accepts, not established connections:
1555        // their detached tasks must remain live throughout notice and drain. The
1556        // local endpoint stops accepting the same way.
1557        drop(serve_task);
1558        drop(local_serve_task);
1559        let second_sigterm = async {
1560            terminate.recv().await;
1561            std::time::Instant::now()
1562        };
1563        tokio::pin!(second_sigterm);
1564        let mut escalation_started = tokio::select! {
1565            biased;
1566            at = second_sigterm.as_mut() => {
1567                info!("second SIGTERM: abandoning daemon shutdown wait");
1568                Some(at)
1569            }
1570            result = supervisor.drain_for_daemon_shutdown() => {
1571                if let Err(error) = result {
1572                    warn!(%error, "daemon shutdown drain failed; exiting anyway");
1573                }
1574                None
1575            }
1576        };
1577        // Supervised modules lead their own process groups, so a service
1578        // manager's kill of this process's group does not reach them. (A
1579        // systemd unit with KillMode=control-group kills by cgroup instead and
1580        // does reach them; see `systemd_kill_mode`.) The daemon ends them
1581        // itself: EOF (or SIGTERM for a protocol none child) first, then
1582        // signals at each child's own drain deadline.
1583        let already_escalated = escalation_started.is_some();
1584        supervisor
1585            .end_children_for_daemon_shutdown(already_escalated, async {
1586                if already_escalated {
1587                    std::future::pending::<()>().await;
1588                } else {
1589                    escalation_started = Some(second_sigterm.await);
1590                }
1591            })
1592            .await;
1593        if let Some(at) = escalation_started {
1594            info!(
1595                second_sigterm_elapsed_ms = at.elapsed().as_millis() as u64,
1596                "daemon shutdown wait ended after second SIGTERM"
1597            );
1598        }
1599        info!("daemon exiting: SIGTERM shutdown complete");
1600        Ok(())
1601    }
1602    #[cfg(not(unix))]
1603    serve_task
1604        .join()
1605        .await
1606        .map_err(BootstrapError::ServeJoin)?
1607        .map_err(BootstrapError::Serve)
1608}
1609
1610fn normalized_build_provenance(value: &str) -> Option<String> {
1611    match value.trim() {
1612        "" | "unavailable" => None,
1613        value => Some(value.to_string()),
1614    }
1615}
1616
1617/// Find an existing daemon or atomically bind loopback TCP for this daemon.
1618///
1619/// The algorithm is intentionally connect-first: an endpoint from the connection
1620/// file is treated as live only after the TCP+key server-proof authenticates for
1621/// that file's key and daemon_id. Stale or foreign connection files are reclaimed
1622/// only while holding the per-user start lock; the TCP port is never the
1623/// singleton primitive.
1624pub async fn ensure_singleton(
1625    connection_file_path: impl AsRef<Path>,
1626    port: u16,
1627) -> Result<Outcome, BootstrapError> {
1628    ensure_singleton_with_config(BootstrapConfig::new(connection_file_path.as_ref(), port)).await
1629}
1630
1631pub async fn ensure_singleton_with_config(
1632    config: BootstrapConfig,
1633) -> Result<Outcome, BootstrapError> {
1634    ensure_singleton_inner(config, true).await
1635}
1636
1637async fn ensure_singleton_inner(
1638    config: BootstrapConfig,
1639    publish: bool,
1640) -> Result<Outcome, BootstrapError> {
1641    let Some(lock) = singleton_start_lock(&config.connection_file_path).await? else {
1642        return Ok(Outcome::AlreadyRunning);
1643    };
1644    ensure_singleton_with_start_lock(config, publish, lock).await
1645}
1646
1647async fn singleton_start_lock(path: &Path) -> Result<Option<StartLock>, BootstrapError> {
1648    if stable_peer(path).await? {
1649        return Ok(None);
1650    }
1651
1652    let lock = StartLock::acquire(path).await?;
1653
1654    // Re-probe after acquiring the start lock so a peer that won the race between
1655    // our first failed probe and the lock acquisition is observed instead of
1656    // overwritten.
1657    if stable_peer(path).await? {
1658        return Ok(None);
1659    }
1660    Ok(Some(lock))
1661}
1662
1663async fn ensure_singleton_with_start_lock(
1664    config: BootstrapConfig,
1665    publish: bool,
1666    lock: StartLock,
1667) -> Result<Outcome, BootstrapError> {
1668    #[cfg(windows)]
1669    let copy_path = windows_connection_copy::resolve(&config);
1670    let path = config.connection_file_path;
1671    // The held start lock and its peer check only rule out a daemon using this
1672    // connection file. A daemon started with another runtime directory but
1673    // the same data home shares the run directory while passing both, so the
1674    // run directory needs its own owner: refuse at once if a live daemon holds
1675    // it, before anything below reads or writes run or data-home state.
1676    let run_dir_lock = config
1677        .live_children_path
1678        .as_deref()
1679        .map(crate::run_dir_lock::RunDirLock::acquire)
1680        .transpose()?;
1681
1682    remove_stale_connection_file_if_present(&path)?;
1683    #[cfg(windows)]
1684    let copy_path = windows_connection_copy::reclaim(&path, copy_path).await;
1685
1686    // Established under the start lock, and under the run-directory lock when
1687    // there is one, and before binding: a corrupt file stops boot before
1688    // anything is published, and two daemons racing to start cannot both mint.
1689    //
1690    // Minting a new machine id fsyncs its file, so loading or minting runs on
1691    // tokio's blocking pool. Both locks travel with that work, so neither is
1692    // released while a mint is in flight.
1693    let (machine_id, lock, run_dir_lock) = match config.machine_id_path {
1694        None => (None, lock, run_dir_lock),
1695        Some(machine_id_path) => {
1696            let task_path = machine_id_path.clone();
1697            match run_blocking_holding((lock, run_dir_lock), move || {
1698                crate::machine_id::load_or_mint(&task_path)
1699            })
1700            .await
1701            {
1702                Ok((Ok(id), (lock, run_dir_lock))) => (Some(id), lock, run_dir_lock),
1703                Ok((Err(error), _locks)) => return Err(BootstrapError::MachineId(error)),
1704                Err(join_error) => {
1705                    return Err(BootstrapError::MachineId(
1706                        crate::machine_id::MachineIdFileError::Write {
1707                            path: machine_id_path,
1708                            source: io::Error::other(describe_blocking_task_failure(
1709                                "loading or minting the machine id",
1710                                &join_error,
1711                            )),
1712                        },
1713                    ))
1714                }
1715            }
1716        }
1717    };
1718
1719    let (listeners, endpoints) = bind_loopback(config.port).await?;
1720    // The local endpoint is an addition to the daemon, not a condition for it:
1721    // a bind failure leaves clients on TCP only, with the connection file naming
1722    // no local endpoint, and the reason reported by `server.describe`. A daemon
1723    // that refused to start would take every module down for a feature nothing
1724    // has to use yet.
1725    let (local_listener, local_endpoint, local_endpoint_unavailable) =
1726        if config.local_listener.configured_dir().is_some() {
1727            match local_listener::bind(&config.local_listener, &path) {
1728                Ok(local_listener) => {
1729                    let endpoint = local_listener.endpoint();
1730                    (Some(local_listener), Some(endpoint), None)
1731                }
1732                Err(error) => {
1733                    let reason = error.to_string();
1734                    warn!(
1735                        error = %reason,
1736                        "the local endpoint is unavailable; serving loopback TCP only"
1737                    );
1738                    (None, None, Some(reason))
1739                }
1740            }
1741        } else {
1742            (None, None, None)
1743        };
1744    let connection_info = ConnectionInfo {
1745        schema: SCHEMA_VERSION,
1746        wire_version: Some(PROTOCOL_VERSION),
1747        endpoints,
1748        key: generate_key().map_err(BootstrapError::GenerateConnectionFile)?,
1749        daemon_id: generate_daemon_id().map_err(BootstrapError::GenerateConnectionFile)?,
1750        pid: process::id(),
1751        daemon_ver: config.daemon_ver,
1752    };
1753
1754    // The low-level binding API historically publishes for callers that serve
1755    // these listeners themselves. The daemon entry point defers publication
1756    // until its own authentication server is running.
1757    let publication_lock = if publish {
1758        let (published, lock) = publish_connection_file(
1759            &path,
1760            #[cfg(windows)]
1761            copy_path.as_deref(),
1762            &connection_info,
1763            local_endpoint.as_ref(),
1764            lock,
1765        )
1766        .await;
1767        if let Err(error) = published {
1768            if let Some(local_listener) = local_listener.as_ref() {
1769                local_listener::release(local_listener);
1770            }
1771            drop(listeners);
1772            return Err(error);
1773        }
1774        drop(lock);
1775        None
1776    } else {
1777        Some(lock)
1778    };
1779
1780    Ok(Outcome::Bound(BoundDaemon {
1781        listeners,
1782        local_listener,
1783        local_endpoint,
1784        local_endpoint_unavailable,
1785        connection_info,
1786        connection_file_path: path,
1787        connection_file_source: config.connection_file_source,
1788        #[cfg(windows)]
1789        connection_file_copy: copy_path,
1790        machine_id,
1791        run_dir_lock,
1792        publication_lock,
1793    }))
1794}
1795
1796#[derive(Debug, Clone, Copy, PartialEq, Eq)]
1797enum Probe {
1798    Live,
1799    StaleOrAbsent,
1800}
1801
1802async fn stable_peer(path: &Path) -> Result<bool, BootstrapError> {
1803    if !matches!(probe_existing(path).await?, Probe::Live) {
1804        return Ok(false);
1805    }
1806    // A terminating peer can still authenticate while its listeners drain.
1807    // Require another proof after a short settling interval before reporting
1808    // a successful duplicate start. If the peer left, proceed with startup;
1809    // any remaining bind/ownership conflict is an error, not exit 0 (which
1810    // would suppress launchd's restart-on-failure policy).
1811    tokio::time::sleep(Duration::from_millis(1500)).await;
1812    Ok(matches!(probe_existing(path).await?, Probe::Live))
1813}
1814
1815async fn probe_existing(path: &Path) -> Result<Probe, BootstrapError> {
1816    let info = match connection_file::read(path) {
1817        Ok(info) => info,
1818        Err(source) if is_absent_or_stale_connection_file(&source) => {
1819            return Ok(Probe::StaleOrAbsent)
1820        }
1821        Err(source) => {
1822            return Err(BootstrapError::ConnectionFileRead {
1823                path: path.to_path_buf(),
1824                source,
1825            })
1826        }
1827    };
1828
1829    for endpoint in &info.endpoints {
1830        if matches!(probe_endpoint(&info, endpoint).await, Probe::Live) {
1831            return Ok(Probe::Live);
1832        }
1833    }
1834
1835    Ok(Probe::StaleOrAbsent)
1836}
1837
1838async fn probe_endpoint(info: &ConnectionInfo, endpoint: &Endpoint) -> Probe {
1839    let Ok(ip) = endpoint.host.parse::<IpAddr>() else {
1840        return Probe::StaleOrAbsent;
1841    };
1842    if !ip.is_loopback() {
1843        return Probe::StaleOrAbsent;
1844    }
1845    let addr = SocketAddr::new(ip, endpoint.port);
1846
1847    let mut stream = match timeout(CONNECT_TIMEOUT, TcpStream::connect(addr)).await {
1848        Ok(Ok(stream)) => stream,
1849        Ok(Err(_)) | Err(_) => return Probe::StaleOrAbsent,
1850    };
1851
1852    match authenticate_client(&mut stream, info, PROBE_AUTH_DEADLINE).await {
1853        Ok(()) => Probe::Live,
1854        Err(AuthError::DaemonIdMismatch)
1855        | Err(AuthError::InvalidServerProof)
1856        | Err(AuthError::UnexpectedEof { .. })
1857        | Err(AuthError::Timeout { .. })
1858        | Err(AuthError::JsonEncode { .. })
1859        | Err(AuthError::JsonDecode { .. })
1860        | Err(AuthError::Io { .. })
1861        | Err(AuthError::MessageTooLarge { .. })
1862        | Err(AuthError::KeyTooShort { .. })
1863        | Err(AuthError::Random(_))
1864        | Err(AuthError::InvalidClientAuth) => Probe::StaleOrAbsent,
1865    }
1866}
1867
1868async fn verify_serving(info: &ConnectionInfo) -> Result<(), BootstrapError> {
1869    // Scheduling a server task is not a readiness barrier. Authenticate once
1870    // against the mandatory IPv4 listener before exposing discovery to peers.
1871    // This uses the normal probe deadline, with no retries or widened budget.
1872    if let Some(endpoint) = info.endpoints.first() {
1873        if matches!(probe_endpoint(info, endpoint).await, Probe::Live) {
1874            return Ok(());
1875        }
1876    }
1877    Err(BootstrapError::StartupNotServing)
1878}
1879
1880/// Run blocking filesystem work (an fsync, a rename) on tokio's blocking-thread
1881/// pool rather than on the async worker thread that awaits it. On a slow or
1882/// busy disk an fsync can take seconds, and a runtime has few worker threads
1883/// (an in-process test daemon may share two with everything else), so doing it
1884/// inline stalls unrelated tasks for that long.
1885///
1886/// `guards` (lock handles) move into the blocking task and are handed back
1887/// with the work's result. If the awaiting future is dropped first, the work
1888/// still runs to completion on its thread and the guards are released only
1889/// after it ends, so a lock is never given up while the work it protects is
1890/// still writing.
1891///
1892/// `Err` means the task panicked, or was cancelled before it started because
1893/// the runtime is shutting down; the guards have then been released already.
1894async fn run_blocking_holding<G, T>(
1895    guards: G,
1896    work: impl FnOnce() -> T + Send + 'static,
1897) -> Result<(T, G), JoinError>
1898where
1899    G: Send + 'static,
1900    T: Send + 'static,
1901{
1902    tokio::task::spawn_blocking(move || {
1903        let output = work();
1904        (output, guards)
1905    })
1906    .await
1907}
1908
1909/// Explain, for a startup error, why blocking work produced no result.
1910fn describe_blocking_task_failure(work: &str, join_error: &JoinError) -> String {
1911    if join_error.is_panic() {
1912        format!("the background thread {work} panicked before finishing: {join_error}")
1913    } else {
1914        format!(
1915            "the background thread {work} was cancelled before finishing, because the async runtime is shutting down: {join_error}"
1916        )
1917    }
1918}
1919
1920/// Publish the connection file: owner-only permissions, written to a temporary
1921/// file, fsynced and renamed over `path` (all done by `write_atomic`). The
1922/// write runs on tokio's blocking pool through [`run_blocking_holding`], so the
1923/// fsync never holds an async worker thread.
1924///
1925/// `guard` is the singleton start lock (or whatever the caller must keep held
1926/// across publication); it is returned with the result. It comes back `None`
1927/// only when the write task panicked or was cancelled, in which case the
1928/// result is an error and the file must be treated as not published.
1929///
1930/// On Windows, `copy_path` asks for the profile copy as well. Its start lock is
1931/// taken before the primary write. Once `path` is written, an error returned
1932/// while preparing or writing the copy only logs a warning, and the result is
1933/// still success. The two writes share one blocking task but aren't one atomic
1934/// step, and a panic or cancellation of that task is still a publication
1935/// error.
1936async fn publish_connection_file<G: Send + 'static>(
1937    path: &Path,
1938    #[cfg(windows)] copy_path: Option<&Path>,
1939    info: &ConnectionInfo,
1940    local_endpoint: Option<&LocalEndpoint>,
1941    guard: G,
1942) -> (Result<(), BootstrapError>, Option<G>) {
1943    let task_path = path.to_path_buf();
1944    let task_info = info.clone();
1945    let task_endpoint = local_endpoint.cloned();
1946    #[cfg(windows)]
1947    let copy = windows_connection_copy::claim_for_publication(path, copy_path, info).await;
1948    let outcome = run_blocking_holding(guard, move || {
1949        write_connection_file(&task_path, &task_info, task_endpoint.as_ref())?;
1950        #[cfg(windows)]
1951        if let Some((copy_path, _lock)) = copy {
1952            if let Err(error) =
1953                write_connection_file(&copy_path, &task_info, task_endpoint.as_ref())
1954            {
1955                windows_connection_copy::warn_unavailable(&task_path, Some(&copy_path), &error);
1956            }
1957        }
1958        Ok(())
1959    })
1960    .await;
1961    match outcome {
1962        Ok((Ok(()), guard)) => (Ok(()), Some(guard)),
1963        Ok((Err(source), guard)) => (
1964            Err(BootstrapError::ConnectionFileWrite {
1965                path: path.to_path_buf(),
1966                source,
1967            }),
1968            Some(guard),
1969        ),
1970        Err(join_error) => (
1971            Err(BootstrapError::ConnectionFileWrite {
1972                path: path.to_path_buf(),
1973                source: ConnectionFileError::Io {
1974                    op: "write",
1975                    path: path.to_path_buf(),
1976                    source: io::Error::other(describe_blocking_task_failure(
1977                        "writing the connection file",
1978                        &join_error,
1979                    )),
1980                },
1981            }),
1982            None,
1983        ),
1984    }
1985}
1986
1987/// The synchronous connection-file write, including its fsync. Call it only
1988/// through [`publish_connection_file`], which keeps it off the async workers.
1989///
1990/// `local_endpoint` is written through subc-transport's endpoint-aware writer,
1991/// so the file carries `local_endpoint` when there is a local listener and is
1992/// byte-identical to before when there is none.
1993fn write_connection_file(
1994    path: &Path,
1995    info: &ConnectionInfo,
1996    local_endpoint: Option<&LocalEndpoint>,
1997) -> Result<(), ConnectionFileError> {
1998    #[cfg(test)]
1999    connection_file_write_hook::run(path);
2000    write_atomic_with_local_endpoint(path, info, local_endpoint)
2001}
2002
2003/// Lets a test run code inside the connection-file write, on whichever thread
2004/// performs it, for example to block there and check what else can still run.
2005#[cfg(test)]
2006mod connection_file_write_hook {
2007    use std::{
2008        path::{Path, PathBuf},
2009        sync::{Mutex, PoisonError},
2010    };
2011
2012    type Hook = Box<dyn FnOnce() + Send>;
2013
2014    /// Keyed by connection-file path, so a hook fires only for the test that
2015    /// installed it while other tests publish their own files in parallel.
2016    static HOOKS: Mutex<Vec<(PathBuf, Hook)>> = Mutex::new(Vec::new());
2017
2018    /// Run `hook` once, at the start of the next write of `path`.
2019    pub(super) fn install(path: &Path, hook: impl FnOnce() + Send + 'static) {
2020        HOOKS
2021            .lock()
2022            .unwrap_or_else(PoisonError::into_inner)
2023            .push((path.to_path_buf(), Box::new(hook)));
2024    }
2025
2026    pub(super) fn run(path: &Path) {
2027        let hook = {
2028            let mut hooks = HOOKS.lock().unwrap_or_else(PoisonError::into_inner);
2029            hooks
2030                .iter()
2031                .position(|(hooked, _)| hooked == path)
2032                .map(|index| hooks.swap_remove(index).1)
2033        };
2034        if let Some(hook) = hook {
2035            hook();
2036        }
2037    }
2038}
2039
2040/// A path-scoped barrier after publication and before any configured activation.
2041/// The async wait leaves the listeners free to answer discovery-time requests.
2042#[cfg(test)]
2043mod startup_publication_hook {
2044    use std::{
2045        path::{Path, PathBuf},
2046        sync::{Mutex, PoisonError},
2047    };
2048    use tokio::sync::oneshot;
2049
2050    type Barrier = (oneshot::Sender<()>, oneshot::Receiver<()>);
2051    static BARRIERS: Mutex<Vec<(PathBuf, Barrier)>> = Mutex::new(Vec::new());
2052
2053    pub(super) fn install(path: &Path) -> (oneshot::Receiver<()>, oneshot::Sender<()>) {
2054        let (entered_tx, entered_rx) = oneshot::channel();
2055        let (release_tx, release_rx) = oneshot::channel();
2056        BARRIERS
2057            .lock()
2058            .unwrap_or_else(PoisonError::into_inner)
2059            .push((path.to_path_buf(), (entered_tx, release_rx)));
2060        (entered_rx, release_tx)
2061    }
2062
2063    pub(super) async fn run(path: &Path) {
2064        let barrier = {
2065            let mut barriers = BARRIERS.lock().unwrap_or_else(PoisonError::into_inner);
2066            barriers
2067                .iter()
2068                .position(|(hooked, _)| hooked == path)
2069                .map(|index| barriers.swap_remove(index).1)
2070        };
2071        if let Some((entered, release)) = barrier {
2072            let _ = entered.send(());
2073            let _ = release.await;
2074        }
2075    }
2076}
2077
2078fn is_absent_or_stale_connection_file(err: &ConnectionFileError) -> bool {
2079    match err {
2080        ConnectionFileError::Io { source, .. } if source.kind() == io::ErrorKind::NotFound => true,
2081        ConnectionFileError::JsonRead { .. }
2082        | ConnectionFileError::UnsupportedSchema { .. }
2083        | ConnectionFileError::Invalid { .. }
2084        | ConnectionFileError::KeyTooShort { .. }
2085        // A live daemon always publishes the file owner-only (0600), so a file
2086        // with insecure permissions is never a daemon we should defer to: treat it
2087        // as stale and take over (which republishes a correct 0600 file).
2088        | ConnectionFileError::InsecurePermissions { .. } => true,
2089        ConnectionFileError::MissingParent { .. }
2090        | ConnectionFileError::MissingFileName { .. }
2091        // A writable ancestor is an operator misconfiguration, never evidence
2092        // about whether a daemon is live. Reclaiming the file would republish key
2093        // material into the same directory the refusal is about.
2094        | ConnectionFileError::InsecureParentDirectory { .. }
2095        | ConnectionFileError::Io { .. }
2096        | ConnectionFileError::JsonWrite { .. }
2097        | ConnectionFileError::Random(_)
2098        // A wire mismatch may identify a newer live daemon, so never reclaim its
2099        // connection file merely because this binary cannot speak its envelope.
2100        | ConnectionFileError::WireVersionMismatch { .. } => false,
2101    }
2102}
2103
2104fn remove_stale_connection_file_if_present(path: &Path) -> Result<(), BootstrapError> {
2105    match fs::remove_file(path) {
2106        Ok(()) => Ok(()),
2107        Err(err) if err.kind() == io::ErrorKind::NotFound => Ok(()),
2108        Err(source) => Err(BootstrapError::RemoveStale {
2109            path: path.to_path_buf(),
2110            source,
2111        }),
2112    }
2113}
2114
2115async fn bind_loopback(port: u16) -> Result<(Vec<TcpListener>, Vec<Endpoint>), BootstrapError> {
2116    bind_loopback_with(port, |port| TcpListener::bind((Ipv6Addr::LOCALHOST, port))).await
2117}
2118
2119async fn bind_loopback_with<F, Fut>(
2120    port: u16,
2121    mut bind_v6: F,
2122) -> Result<(Vec<TcpListener>, Vec<Endpoint>), BootstrapError>
2123where
2124    F: FnMut(u16) -> Fut,
2125    Fut: std::future::Future<Output = io::Result<TcpListener>>,
2126{
2127    let v4_host = Ipv4Addr::LOCALHOST;
2128    let v4 = TcpListener::bind((v4_host, port))
2129        .await
2130        .map_err(|source| BootstrapError::Bind {
2131            host: v4_host.to_string(),
2132            port,
2133            source,
2134        })?;
2135    let actual_port = v4
2136        .local_addr()
2137        .map_err(|source| BootstrapError::LocalAddr {
2138            host: v4_host.to_string(),
2139            source,
2140        })?
2141        .port();
2142
2143    let mut listeners = vec![v4];
2144    let mut endpoints = vec![Endpoint {
2145        host: v4_host.to_string(),
2146        port: actual_port,
2147    }];
2148
2149    let v6_host = Ipv6Addr::LOCALHOST;
2150    match bind_v6(actual_port).await {
2151        Ok(v6) => {
2152            listeners.push(v6);
2153            endpoints.push(Endpoint {
2154                host: v6_host.to_string(),
2155                port: actual_port,
2156            });
2157        }
2158        Err(err)
2159            if ipv6_loopback_unavailable(&err)
2160                || (port == 0 && err.kind() == io::ErrorKind::AddrInUse) =>
2161        {
2162            warn!(
2163                port = actual_port,
2164                error = %err,
2165                "IPv6 loopback unavailable; serving only IPv4 loopback"
2166            );
2167        }
2168        Err(source) => {
2169            drop(listeners);
2170            return Err(BootstrapError::Bind {
2171                host: v6_host.to_string(),
2172                port: actual_port,
2173                source,
2174            });
2175        }
2176    }
2177
2178    Ok((listeners, endpoints))
2179}
2180
2181fn ipv6_loopback_unavailable(err: &io::Error) -> bool {
2182    matches!(
2183        err.kind(),
2184        io::ErrorKind::AddrNotAvailable | io::ErrorKind::Unsupported
2185    ) || matches!(err.raw_os_error(), Some(47) | Some(49) | Some(97))
2186}
2187
2188struct AbortOnDrop<T> {
2189    handle: JoinHandle<T>,
2190}
2191
2192impl<T> AbortOnDrop<T> {
2193    fn new(handle: JoinHandle<T>) -> Self {
2194        Self { handle }
2195    }
2196
2197    async fn join(&mut self) -> Result<T, JoinError> {
2198        (&mut self.handle).await
2199    }
2200}
2201
2202impl<T> Drop for AbortOnDrop<T> {
2203    fn drop(&mut self) {
2204        if !self.handle.is_finished() {
2205            self.handle.abort();
2206        }
2207    }
2208}
2209
2210#[derive(Debug)]
2211struct StartLock {
2212    // Keep the locked file handle alive for the duration of bootstrap.
2213    _file: fs::File,
2214}
2215
2216impl Drop for StartLock {
2217    fn drop(&mut self) {
2218        // Closing alone leaves the lock held if a concurrent fork inherited
2219        // the open file description and has not exec'd yet. Unlock explicitly
2220        // so a restart can reclaim the singleton as soon as bootstrap ends.
2221        if let Err(error) = FileExt::unlock(&self._file) {
2222            tracing::warn!(%error, "failed to release the singleton start lock");
2223        }
2224    }
2225}
2226
2227impl StartLock {
2228    async fn acquire(connection_file_path: &Path) -> Result<Self, BootstrapError> {
2229        let path = start_lock_path(connection_file_path);
2230        for _ in 0..START_LOCK_RETRIES {
2231            let file = match open_owner_only_lock(&path) {
2232                Ok(file) => file,
2233                Err(source) => return Err(BootstrapError::StartLockCreate { path, source }),
2234            };
2235            match FileExt::try_lock(&file) {
2236                Ok(()) => return Ok(Self { _file: file }),
2237                Err(TryLockError::WouldBlock) => sleep(START_LOCK_RETRY_DELAY).await,
2238                Err(TryLockError::Error(source)) => {
2239                    return Err(BootstrapError::StartLockCreate { path, source });
2240                }
2241            }
2242        }
2243
2244        Err(BootstrapError::StartLockBusy {
2245            path,
2246            attempts: START_LOCK_RETRIES,
2247        })
2248    }
2249}
2250
2251pub(crate) fn open_owner_only_lock(path: &Path) -> io::Result<fs::File> {
2252    let mut options = fs::OpenOptions::new();
2253    options.read(true).write(true).create(true);
2254    #[cfg(unix)]
2255    {
2256        use std::os::unix::fs::OpenOptionsExt;
2257        options.mode(0o600);
2258    }
2259    options.open(path)
2260}
2261
2262fn start_lock_path(connection_file_path: &Path) -> PathBuf {
2263    let file_name = connection_file_path
2264        .file_name()
2265        .map(|name| name.to_string_lossy())
2266        .unwrap_or_else(|| CONNECTION_FILE_NAME.into());
2267    let lock_name = format!("{file_name}.start-lock");
2268    connection_file_path
2269        .parent()
2270        .filter(|parent| !parent.as_os_str().is_empty())
2271        .unwrap_or_else(|| Path::new("."))
2272        .join(lock_name)
2273}
2274
2275fn non_empty_os_var(key: &str) -> Option<OsString> {
2276    let value = env::var_os(key)?;
2277    if value.is_empty() {
2278        None
2279    } else {
2280        Some(value)
2281    }
2282}
2283
2284/// Bootstrap-layer errors are deliberately typed so startup never panics for
2285/// ordinary daemon-discovery races or stale filesystem state.
2286#[derive(Debug)]
2287#[non_exhaustive]
2288pub enum BootstrapError {
2289    ConfigStore(String),
2290    /// The bound server could not authenticate its own readiness probe.
2291    StartupNotServing,
2292    #[cfg(unix)]
2293    Signal(io::Error),
2294    InvalidPort {
2295        raw: String,
2296        source: std::num::ParseIntError,
2297    },
2298    ConnectionFileRead {
2299        path: PathBuf,
2300        source: ConnectionFileError,
2301    },
2302    ConnectionFileWrite {
2303        path: PathBuf,
2304        source: ConnectionFileError,
2305    },
2306    GenerateConnectionFile(ConnectionFileError),
2307    StartLockCreate {
2308        path: PathBuf,
2309        source: io::Error,
2310    },
2311    StartLockBusy {
2312        path: PathBuf,
2313        attempts: usize,
2314    },
2315    /// The run-directory lock file could not be created, opened or locked.
2316    RunDirLockCreate {
2317        path: PathBuf,
2318        source: io::Error,
2319    },
2320    /// Another live process, almost certainly a daemon started with a
2321    /// different runtime directory over the same data home, owns the run
2322    /// directory. The daemon does not start. `holder_pid` is read from the
2323    /// lock file when possible and is informational only.
2324    RunDirBusy {
2325        path: PathBuf,
2326        holder_pid: Option<u32>,
2327    },
2328    RemoveStale {
2329        path: PathBuf,
2330        source: io::Error,
2331    },
2332    Bind {
2333        host: String,
2334        port: u16,
2335        source: io::Error,
2336    },
2337    LocalAddr {
2338        host: String,
2339        source: io::Error,
2340    },
2341    DaemonConfig(DaemonConfigError),
2342    /// The machine id could not be established: its file is corrupt, unreadable
2343    /// or unwritable, or the data home is relative. The daemon does not start.
2344    MachineId(crate::machine_id::MachineIdFileError),
2345    /// The daemon run directory could not be resolved because the data home is
2346    /// relative. The daemon does not start.
2347    RunDir(daemon_config::DaemonRunDirError),
2348    Serve(ServerError),
2349    ServeJoin(tokio::task::JoinError),
2350}
2351
2352impl From<ServerError> for BootstrapError {
2353    fn from(error: ServerError) -> Self {
2354        Self::Serve(error)
2355    }
2356}
2357
2358impl fmt::Display for BootstrapError {
2359    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
2360        match self {
2361            Self::ConfigStore(error) => write!(f, "configuration store startup refused: {error}"),
2362            Self::StartupNotServing => write!(f, "bound daemon did not authenticate its readiness probe; connection file was not published"),
2363            #[cfg(unix)]
2364            Self::Signal(error) => write!(f, "failed to register SIGTERM handler: {error}"),
2365            Self::InvalidPort { raw, source } => {
2366                write!(f, "invalid {SUBC_PORT_ENV} value '{raw}': {source}")
2367            }
2368            Self::ConnectionFileRead { path, source } => write!(
2369                f,
2370                "failed to read connection file {}: {source}",
2371                path.display()
2372            ),
2373            Self::ConnectionFileWrite { path, source } => write!(
2374                f,
2375                "failed to publish connection file {}: {source}",
2376                path.display()
2377            ),
2378            Self::GenerateConnectionFile(err) => {
2379                write!(f, "failed to generate connection-file auth material: {err}")
2380            }
2381            Self::StartLockCreate { path, source } => {
2382                write!(
2383                    f,
2384                    "failed to create start lock {}: {source}",
2385                    path.display()
2386                )
2387            }
2388            Self::StartLockBusy { path, attempts } => write!(
2389                f,
2390                "start lock {} remained busy after {attempts} attempts",
2391                path.display()
2392            ),
2393            Self::RunDirLockCreate { path, source } => write!(
2394                f,
2395                "refusing to start: failed to lock run directory via {}: {source}",
2396                path.display()
2397            ),
2398            Self::RunDirBusy { path, holder_pid } => {
2399                write!(
2400                    f,
2401                    "refusing to start: run directory lock {} is held by another daemon",
2402                    path.display()
2403                )?;
2404                if let Some(pid) = holder_pid {
2405                    write!(f, " (pid {pid})")?;
2406                }
2407                write!(
2408                    f,
2409                    "; a live daemon owns this data home's run state (typically one started with a different XDG_RUNTIME_DIR)"
2410                )
2411            }
2412            Self::RemoveStale { path, source } => write!(
2413                f,
2414                "failed to remove stale connection file {}: {source}",
2415                path.display()
2416            ),
2417            Self::Bind { host, port, source } if source.kind() == io::ErrorKind::AddrInUse => {
2418                write!(
2419                    f,
2420                    "port {port} in use on loopback {host}: {source}; set the port in config"
2421                )
2422            }
2423            Self::Bind { host, port, source } => {
2424                write!(f, "failed to bind loopback TCP {host}:{port}: {source}")
2425            }
2426            Self::LocalAddr { host, source } => {
2427                write!(f, "failed to read local address for {host}: {source}")
2428            }
2429            Self::DaemonConfig(err) => write!(f, "failed to load daemon config: {err}"),
2430            Self::MachineId(err) => write!(f, "refusing to start: {err}"),
2431            Self::RunDir(err) => write!(f, "refusing to start: {err}"),
2432            Self::Serve(err) => write!(f, "daemon server failed: {err}"),
2433            Self::ServeJoin(err) => write!(f, "daemon server task failed: {err}"),
2434        }
2435    }
2436}
2437
2438impl Error for BootstrapError {
2439    fn source(&self) -> Option<&(dyn Error + 'static)> {
2440        match self {
2441            Self::ConfigStore(_) => None,
2442            Self::StartupNotServing => None,
2443            #[cfg(unix)]
2444            Self::Signal(source) => Some(source),
2445            Self::InvalidPort { source, .. } => Some(source),
2446            Self::ConnectionFileRead { source, .. }
2447            | Self::ConnectionFileWrite { source, .. }
2448            | Self::GenerateConnectionFile(source) => Some(source),
2449            Self::StartLockCreate { source, .. }
2450            | Self::RunDirLockCreate { source, .. }
2451            | Self::RemoveStale { source, .. }
2452            | Self::Bind { source, .. }
2453            | Self::LocalAddr { source, .. } => Some(source),
2454            Self::DaemonConfig(err) => Some(err),
2455            Self::MachineId(err) => Some(err),
2456            Self::RunDir(err) => Some(err),
2457            Self::Serve(err) => Some(err),
2458            Self::ServeJoin(err) => Some(err),
2459            Self::StartLockBusy { .. } | Self::RunDirBusy { .. } => None,
2460        }
2461    }
2462}
2463
2464#[cfg(all(test, feature = "test-support", unix))]
2465#[path = "bootstrap_recovery_tests.rs"]
2466mod recovery_tests;
2467
2468#[cfg(test)]
2469mod tests {
2470    mod activity {
2471        include!("activity_bootstrap_tests.rs");
2472    }
2473    mod startup_order {
2474        include!("startup_order_tests.rs");
2475    }
2476    use super::*;
2477    use crate::server::ServerAuth;
2478    use cortexkit_test_support::ScratchDir;
2479    #[cfg(target_os = "linux")]
2480    use std::collections::BTreeSet;
2481    use std::sync::Mutex;
2482    #[cfg(target_os = "linux")]
2483    use subc_control::ModuleProtocol;
2484    use subc_transport::{write_atomic, MIN_KEY_LEN};
2485    use tokio::io::AsyncReadExt;
2486    use tokio::task::JoinHandle;
2487
2488    #[cfg(unix)]
2489    use std::os::unix::fs::PermissionsExt;
2490
2491    static ENV_LOCK: Mutex<()> = Mutex::new(());
2492
2493    #[test]
2494    fn normalized_build_provenance_preserves_real_values() {
2495        assert_eq!(normalized_build_provenance("abc"), Some("abc".to_string()));
2496    }
2497
2498    #[test]
2499    fn config_recovery_bootstrap_preserves_all_loaded_fields_and_only_argument_overrides() {
2500        let _env_lock = ENV_LOCK.lock().unwrap();
2501        let scratch = unique_temp_dir("bootstrap-loaded-fields");
2502        let file = scratch.join("subc.jsonc");
2503        std::fs::write(
2504            &file,
2505            br#"{"version":1,"modules":{},"scope_authority_owners":["file-owner"]}"#,
2506        )
2507        .unwrap();
2508        let mut config = BootstrapConfig::new(scratch.join("connection.json"), 0)
2509            .with_daemon_config_source(file.clone(), None);
2510        config.defer_file_load = true;
2511        config.operator_timeout = Duration::from_secs(7);
2512        let selected = load_bootstrap_authority(config, None).unwrap();
2513        assert_eq!(
2514            selected.file_derived_test_probe,
2515            Some(vec!["file-owner".into()])
2516        );
2517        assert_eq!(selected.operator_timeout, Duration::from_secs(7));
2518        assert_eq!(selected.daemon_config_path, Some(file));
2519        assert_eq!(
2520            selected.connection_file_path,
2521            scratch.join("connection.json")
2522        );
2523    }
2524
2525    #[test]
2526    fn config_recovery_bootstrap_ignores_invalid_live_file_and_poisoned_settings() {
2527        let scratch = unique_temp_dir("recovery-bootstrap");
2528        let root = scratch.join("store");
2529        std::fs::create_dir_all(&root).unwrap();
2530        std::fs::write(
2531            root.join(crate::config_store::STORE_FILE),
2532            br#"{"format_version":999}"#,
2533        )
2534        .unwrap();
2535        let mut store = crate::config_store::ConfigStore::open_digest_only_for_tests(
2536            crate::config_store::StoreRoot::from_service_definition(root).unwrap(),
2537        );
2538        assert_eq!(
2539            store.status().state,
2540            crate::config_store::StoreState::StoreUnverified
2541        );
2542        let file = scratch.join("subc.jsonc");
2543        std::fs::write(&file, b"invalid live config").unwrap();
2544        let mut config = BootstrapConfig::new(scratch.join("connection.json"), 1)
2545            .with_daemon_config_source(file, Some(1))
2546            .with_scope_authority_owners(vec!["untrusted".into()])
2547            .with_reserved_capabilities(BTreeMap::from([("untrusted".into(), "untrusted".into())]));
2548        config.defer_file_load = true;
2549        config.file_derived_test_probe = Some(vec!["untrusted".into()]);
2550        let config = load_bootstrap_authority(config, Some(&mut store)).unwrap();
2551        assert_eq!(config.file_derived_test_probe, None);
2552        assert_eq!(config.port, DEFAULT_SUBC_PORT);
2553        assert!(config.configured_modules.is_empty());
2554        assert_eq!(
2555            config.scope_authority_owners,
2556            daemon_config::default_scope_authority_owners()
2557        );
2558        assert!(config.reserved_capabilities.is_empty());
2559        assert_eq!(
2560            config.privileged_operator_prompt,
2561            daemon_config::PrivilegedOperatorPrompt::default()
2562        );
2563    }
2564
2565    /// An approved generation whose daemon bytes listen on `port`. Changing
2566    /// the port is a setting only bootstrap reads, so it waits for a restart.
2567    fn approved_port_entry(
2568        kind: crate::config_store::CommitKind,
2569        port: u16,
2570    ) -> crate::config_store::HistoryEntry {
2571        use subc_protocol::config::{ConfigLayer, ConfigLayerSnapshot};
2572        crate::config_store::HistoryEntry {
2573            generation: subc_config::Generation::new(
2574                0,
2575                subc_config::ApprovalChannel::OsPrompt,
2576                subc_config::raw_digest(b"approved-plan"),
2577                subc_config::InputVector::default(),
2578                subc_config::Changeset::new(),
2579            ),
2580            kind,
2581            approval_id: "approval".into(),
2582            outcome: "approved".into(),
2583            tuples: Default::default(),
2584            configuration: Default::default(),
2585            delivery: Some(crate::config_store::DeliverySnapshot {
2586                daemon: Some(ConfigLayerSnapshot::file(
2587                    "subc",
2588                    ConfigLayer::User,
2589                    "/approved/subc.jsonc",
2590                    Some(format!(r#"{{"version":1,"port":{port},"modules":{{}}}}"#).into_bytes()),
2591                )),
2592                modules: Default::default(),
2593                declarations: Default::default(),
2594            }),
2595        }
2596    }
2597
2598    fn approve_port(
2599        store: &mut crate::config_store::ConfigStore,
2600        kind: crate::config_store::CommitKind,
2601        port: u16,
2602    ) {
2603        store
2604            .commit_generation(
2605                approved_port_entry(kind, port),
2606                Default::default(),
2607                Default::default(),
2608                1,
2609            )
2610            .unwrap();
2611    }
2612
2613    fn stamped_boot(
2614        scratch: &Path,
2615        store: &mut crate::config_store::ConfigStore,
2616    ) -> BootstrapConfig {
2617        let mut config = BootstrapConfig::new(scratch.join("connection.json"), 1)
2618            .with_daemon_config_source(scratch.join("subc.jsonc"), Some(1));
2619        config.defer_file_load = true;
2620        load_bootstrap_authority(config, Some(store)).unwrap()
2621    }
2622
2623    #[test]
2624    fn config_stamped_boot_uses_last_active_generation_until_restart_intent() {
2625        use crate::config_store::{CommitKind, ConfigStore, StoreRoot};
2626        let scratch = unique_temp_dir("stamped-boot-selection");
2627        let root = scratch.join("store");
2628        std::fs::create_dir_all(&root).unwrap();
2629        let open = || {
2630            ConfigStore::open_digest_only_for_tests(
2631                StoreRoot::from_service_definition(root.clone()).unwrap(),
2632            )
2633        };
2634        let mut store = open();
2635        approve_port(&mut store, CommitKind::Stamp, 41001);
2636        approve_port(&mut store, CommitKind::Approved, 41002);
2637        drop(store);
2638        let mut store = open();
2639        assert_eq!(
2640            stamped_boot(&scratch, &mut store).port,
2641            41001,
2642            "an ordinary start must not boot an approval still waiting for a daemon restart"
2643        );
2644        assert_eq!(store.assign_daemon_restart().unwrap(), Some(2));
2645        drop(store);
2646        let mut store = open();
2647        assert_eq!(stamped_boot(&scratch, &mut store).port, 41002);
2648    }
2649
2650    #[test]
2651    fn config_stamped_boot_enters_recovery_when_anchor_retry_fails_at_start() {
2652        use crate::config_store::{CommitKind, ConfigStore, StoreRoot, StoreState, TestStoreKey};
2653        let scratch = unique_temp_dir("stamped-boot-anchor-pending");
2654        let root = scratch.join("store");
2655        std::fs::create_dir_all(&root).unwrap();
2656        let key = TestStoreKey::default();
2657        let open = || {
2658            ConfigStore::open_sealed_for_tests(
2659                StoreRoot::from_service_definition(root.clone()).unwrap(),
2660                key.clone(),
2661            )
2662        };
2663        let mut store = open();
2664        approve_port(&mut store, CommitKind::Stamp, 41001);
2665        // Crash between publication and anchor: the bytes are renamed into
2666        // place but the anchor never catches up, and keeps failing at start.
2667        key.fail_anchor(true);
2668        assert!(store
2669            .commit_generation(
2670                approved_port_entry(CommitKind::Approved, 41001),
2671                Default::default(),
2672                Default::default(),
2673                2,
2674            )
2675            .is_err());
2676        drop(store);
2677        let mut store = open();
2678        let status = store.status();
2679        assert_eq!(status.state, StoreState::Stamped);
2680        assert!(status.anchor_pending);
2681        let config = stamped_boot(&scratch, &mut store);
2682        assert_eq!(config.port, DEFAULT_SUBC_PORT);
2683        assert!(config.configured_modules.is_empty());
2684        assert_eq!(store.status().state, StoreState::StoreUnverified);
2685    }
2686
2687    #[test]
2688    fn normalized_build_provenance_omits_unavailable_and_empty_values() {
2689        assert_eq!(normalized_build_provenance("unavailable"), None);
2690        assert_eq!(normalized_build_provenance(""), None);
2691    }
2692
2693    #[cfg(target_os = "linux")]
2694    fn current_cgroup_path_for_test() -> io::Result<PathBuf> {
2695        let cgroups = fs::read_to_string("/proc/self/cgroup")?;
2696        let relative = cgroups
2697            .lines()
2698            .find_map(|line| line.strip_prefix("0::"))
2699            .ok_or_else(|| {
2700                io::Error::new(io::ErrorKind::Unsupported, "cgroup v2 is unavailable")
2701            })?;
2702        Ok(Path::new("/sys/fs/cgroup").join(relative.trim_start_matches('/')))
2703    }
2704
2705    #[cfg(target_os = "linux")]
2706    fn module_cgroup_directories() -> io::Result<BTreeSet<OsString>> {
2707        let modules = current_cgroup_path_for_test()?.join("subc-modules");
2708        let entries = match fs::read_dir(modules) {
2709            Ok(entries) => entries,
2710            Err(error) if error.kind() == io::ErrorKind::NotFound => return Ok(BTreeSet::new()),
2711            Err(error) => return Err(error),
2712        };
2713        let mut directories = BTreeSet::new();
2714        for entry in entries {
2715            let entry = entry?;
2716            if entry.file_type()?.is_dir() {
2717                directories.insert(entry.file_name());
2718            }
2719        }
2720        Ok(directories)
2721    }
2722
2723    #[cfg(target_os = "linux")]
2724    async fn wait_for_path(path: &Path, task: &JoinHandle<Result<(), BootstrapError>>) {
2725        let deadline = tokio::time::Instant::now() + Duration::from_secs(5);
2726        while !path.exists() && tokio::time::Instant::now() < deadline {
2727            assert!(
2728                !task.is_finished(),
2729                "daemon exited before creating {}",
2730                path.display()
2731            );
2732            sleep(Duration::from_millis(10)).await;
2733        }
2734        assert!(path.exists(), "daemon did not create {}", path.display());
2735    }
2736
2737    /// An in-process daemon with the default (disabled) cgroup placement must
2738    /// leave the host's live module cgroup tree exactly as it found it.
2739    /// Red-checking this test (making it fail) creates a
2740    /// `cgroup-isolation-probe-*` directory in the live cgroup tree, which must be removed by hand.
2741    #[cfg(target_os = "linux")]
2742    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
2743    async fn run_with_config_does_not_reconcile_the_ambient_cgroup_by_default() {
2744        let temp = unique_temp_dir("bootstrap-cgroup-default-disabled");
2745        let module_id = format!("cgroup-isolation-probe-{}", process::id());
2746        let before = module_cgroup_directories().expect("read ambient module cgroups before boot");
2747        assert!(
2748            !before.contains(&OsString::from(&module_id)),
2749            "isolation probe cgroup already exists before this daemon starts"
2750        );
2751        let capture = temp.join("logs").join(format!("{module_id}.stderr.log"));
2752        let module = ConfiguredModule {
2753            script: None,
2754            carry_containment: false,
2755            module_id,
2756            program: PathBuf::from("sh"),
2757            args: vec!["-c".to_string(), "sleep 30".to_string()],
2758            env: Vec::new(),
2759            log: None,
2760            enabled: true,
2761            registration_channel: false,
2762            reserved: false,
2763            reserved_prefixes: Vec::new(),
2764            protocol: ModuleProtocol::None,
2765            overlap: Default::default(),
2766            health: HealthConfig::default(),
2767            drain_timeout_ms: None,
2768            register_timeout: crate::supervise::DEFAULT_REGISTER_TIMEOUT,
2769            route_bind_relay_timeout_ms: None,
2770            restart: RestartPolicy::default(),
2771        };
2772        let config = config_with_a_short_socket(BootstrapConfig::new(temp.join("c"), 0))
2773            .with_configured_modules([module])
2774            .with_capture_logs_dir(temp.join("logs"))
2775            .with_terminal_journal_path(temp.join("terminals.jsonl"));
2776        let task = tokio::spawn(run_with_config(config));
2777
2778        wait_for_path(&capture, &task).await;
2779        let after = module_cgroup_directories().expect("read ambient module cgroups after boot");
2780
2781        task.abort();
2782        assert!(task
2783            .await
2784            .expect_err("aborted daemon task must cancel")
2785            .is_cancelled());
2786        assert_eq!(
2787            before, after,
2788            "an in-process daemon must not create or reconcile ambient module cgroups"
2789        );
2790    }
2791
2792    #[cfg(target_os = "linux")]
2793    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
2794    async fn explicit_cgroup_root_is_prepared_inside_the_fixture_tree() {
2795        let temp = unique_temp_dir("bootstrap-cgroup-explicit-root");
2796        let cgroup_root = temp.join("cgroup");
2797        fs::create_dir(&cgroup_root).expect("create scratch cgroup root");
2798        fs::write(cgroup_root.join("cgroup.procs"), b"").expect("write scratch cgroup marker");
2799        let modules = cgroup_root.join("subc-modules");
2800        let config = config_with_a_short_socket(BootstrapConfig::new(temp.join("c"), 0))
2801            .with_cgroup_placement(CgroupPlacementConfig::Root(cgroup_root))
2802            .with_terminal_journal_path(temp.join("terminals.jsonl"));
2803        let task = tokio::spawn(run_with_config(config));
2804
2805        wait_for_path(&modules, &task).await;
2806
2807        task.abort();
2808        assert!(task
2809            .await
2810            .expect_err("aborted daemon task must cancel")
2811            .is_cancelled());
2812        assert!(
2813            fs::read_dir(&modules)
2814                .expect("read prepared modules directory")
2815                .next()
2816                .is_none(),
2817            "the delegation probe must clean up after itself"
2818        );
2819    }
2820
2821    struct EnvGuard {
2822        key: &'static str,
2823        previous: Option<OsString>,
2824    }
2825
2826    impl EnvGuard {
2827        fn set(key: &'static str, value: &Path) -> Self {
2828            let previous = env::var_os(key);
2829            env::set_var(key, value);
2830            Self { key, previous }
2831        }
2832
2833        fn set_str(key: &'static str, value: &str) -> Self {
2834            let previous = env::var_os(key);
2835            env::set_var(key, value);
2836            Self { key, previous }
2837        }
2838
2839        fn unset(key: &'static str) -> Self {
2840            let previous = env::var_os(key);
2841            env::remove_var(key);
2842            Self { key, previous }
2843        }
2844    }
2845
2846    impl Drop for EnvGuard {
2847        fn drop(&mut self) {
2848            match &self.previous {
2849                Some(value) => env::set_var(self.key, value),
2850                None => env::remove_var(self.key),
2851            }
2852        }
2853    }
2854
2855    fn unique_temp_dir(name: &str) -> ScratchDir {
2856        ScratchDir::new(name)
2857    }
2858
2859    fn temp_connection_file_path(name: &str) -> (ScratchDir, PathBuf) {
2860        let dir = unique_temp_dir(name);
2861        let path = dir.join("conn.json");
2862        (dir, path)
2863    }
2864
2865    fn auth_for(info: &ConnectionInfo) -> ServerAuth {
2866        ServerAuth::new(info.key.clone(), info.daemon_id, info.daemon_ver.clone())
2867    }
2868
2869    fn start_server(bound: BoundDaemon) -> JoinHandle<Result<(), ServerError>> {
2870        let auth = auth_for(&bound.connection_info);
2871        tokio::spawn(serve_listeners(
2872            bound.listeners,
2873            Arc::new(Router::with_default_self_handler()),
2874            auth,
2875        ))
2876    }
2877
2878    fn expect_bound(outcome: Outcome) -> BoundDaemon {
2879        match outcome {
2880            Outcome::Bound(bound) => bound,
2881            Outcome::AlreadyRunning => panic!("fresh connection file unexpectedly had a daemon"),
2882        }
2883    }
2884
2885    async fn connect_from_info(conn: &ConnectionInfo) -> io::Result<TcpStream> {
2886        let endpoint = conn
2887            .endpoints
2888            .first()
2889            .expect("test connection file should have an endpoint");
2890        let ip: IpAddr = endpoint.host.parse().unwrap();
2891        TcpStream::connect(SocketAddr::new(ip, endpoint.port)).await
2892    }
2893
2894    fn make_connection_info(port: u16) -> ConnectionInfo {
2895        ConnectionInfo {
2896            schema: SCHEMA_VERSION,
2897            wire_version: Some(PROTOCOL_VERSION),
2898            endpoints: vec![Endpoint {
2899                host: "127.0.0.1".to_owned(),
2900                port,
2901            }],
2902            key: generate_key().unwrap(),
2903            daemon_id: generate_daemon_id().unwrap(),
2904            pid: process::id(),
2905            daemon_ver: "test-subc".to_owned(),
2906        }
2907    }
2908
2909    fn write_raw_owner_only_connection_file(path: &Path, contents: &[u8]) {
2910        fs::write(path, contents).unwrap();
2911        #[cfg(unix)]
2912        fs::set_permissions(path, fs::Permissions::from_mode(0o600)).unwrap();
2913    }
2914
2915    fn assert_owner_only_connection_file(path: &Path) {
2916        // `path` is only inspected on Unix (mode bits); on Windows the owner-only
2917        // guarantee comes from the inherited %TEMP% ACL, nothing to assert here.
2918        #[cfg(unix)]
2919        {
2920            let mode = fs::metadata(path).unwrap().permissions().mode() & 0o777;
2921            assert_eq!(mode, 0o600);
2922        }
2923        #[cfg(not(unix))]
2924        let _ = path;
2925    }
2926
2927    #[test]
2928    fn connection_file_path_uses_xdg_runtime_dir_when_set() {
2929        let _env_lock = ENV_LOCK.lock().unwrap();
2930        let runtime_dir = unique_temp_dir("xdg-runtime");
2931        let _xdg = EnvGuard::set("XDG_RUNTIME_DIR", runtime_dir.path());
2932
2933        assert_eq!(
2934            connection_file_path(),
2935            runtime_dir.join(CONNECTION_FILE_NAME)
2936        );
2937    }
2938
2939    #[test]
2940    fn connection_file_path_source_is_xdg_runtime_dir_when_set() {
2941        let runtime_dir = OsString::from("/run/user/1000");
2942
2943        let (path, source) = connection_file_path_with_source(Some(runtime_dir));
2944
2945        assert_eq!(
2946            path,
2947            PathBuf::from("/run/user/1000").join(CONNECTION_FILE_NAME)
2948        );
2949        assert_eq!(source, ConnectionFileSource::XdgRuntimeDir);
2950    }
2951
2952    #[test]
2953    fn connection_file_path_falls_back_to_temp_dir_with_user_token_when_xdg_unset() {
2954        let _env_lock = ENV_LOCK.lock().unwrap();
2955        let _xdg = EnvGuard::unset("XDG_RUNTIME_DIR");
2956
2957        assert_eq!(
2958            connection_file_path(),
2959            env::temp_dir().join(format!("subc-{}.connection.json", user_connection_token()))
2960        );
2961    }
2962
2963    #[test]
2964    fn connection_file_path_source_is_temp_dir_when_xdg_unset() {
2965        let (path, source) = connection_file_path_with_source(None);
2966
2967        assert_eq!(
2968            path,
2969            env::temp_dir().join(format!("subc-{}.connection.json", user_connection_token()))
2970        );
2971        assert_eq!(source, ConnectionFileSource::TempDirFallback);
2972    }
2973
2974    /// Concurrent callers must derive one token. The former temp-file uid probe
2975    /// could fail transiently (same-tick name collision, fd exhaustion) and send
2976    /// the loser down the env-derived fallback with a different identity for the
2977    /// same user; this fence keeps identity independent of filesystem luck.
2978    #[test]
2979    fn user_connection_token_is_stable_under_concurrent_callers() {
2980        let expected = user_connection_token();
2981        let workers: Vec<_> = (0..32)
2982            .map(|_| {
2983                std::thread::spawn(|| (0..40).map(|_| user_connection_token()).collect::<Vec<_>>())
2984            })
2985            .collect();
2986        for worker in workers {
2987            for token in worker.join().expect("probe thread") {
2988                assert_eq!(token, expected, "token diverged under concurrent probes");
2989            }
2990        }
2991    }
2992
2993    #[test]
2994    fn connection_file_path_source_is_temp_dir_when_xdg_empty() {
2995        let (path, source) = connection_file_path_with_source(Some(OsString::new()));
2996
2997        assert_eq!(
2998            path,
2999            env::temp_dir().join(format!("subc-{}.connection.json", user_connection_token()))
3000        );
3001        assert_eq!(source, ConnectionFileSource::TempDirFallback);
3002    }
3003
3004    /// The shipped binary is the only caller that journals terminal exits into
3005    /// the real run directory, so its constructor must supply that path itself;
3006    /// the in-process default leaves it unset.
3007    #[test]
3008    fn daemon_binary_config_journals_and_captures_into_the_run_dir() {
3009        let _env_lock = ENV_LOCK.lock().unwrap();
3010        let root = unique_temp_dir("daemon-binary-config");
3011        let data_home = root.join("data");
3012        let _data = EnvGuard::set("XDG_DATA_HOME", &data_home);
3013        let _config = EnvGuard::set("XDG_CONFIG_HOME", &root.join("config"));
3014        let _port = EnvGuard::unset(SUBC_PORT_ENV);
3015
3016        let config = BootstrapConfig::from_env_for_daemon_binary().unwrap();
3017
3018        let run_dir = data_home.join("cortexkit").join("run");
3019        assert_eq!(
3020            config.terminal_journal_path,
3021            Some(run_dir.join("terminals.jsonl"))
3022        );
3023        assert_eq!(config.capture_logs_dir, Some(run_dir.join("logs")));
3024        assert_eq!(
3025            BootstrapConfig::new(root.join("connection.json"), 0).terminal_journal_path,
3026            None,
3027            "an in-process config must not journal anywhere unless asked to"
3028        );
3029    }
3030
3031    #[test]
3032    fn daemon_binary_config_refuses_a_relative_data_home() {
3033        let _env_lock = ENV_LOCK.lock().unwrap();
3034        let root = unique_temp_dir("daemon-binary-relative-data");
3035        let _data = EnvGuard::set_str("XDG_DATA_HOME", "relative-data-home");
3036        let _config = EnvGuard::set("XDG_CONFIG_HOME", &root.join("config"));
3037        let _port = EnvGuard::unset(SUBC_PORT_ENV);
3038
3039        let error = BootstrapConfig::from_env_for_daemon_binary()
3040            .expect_err("a relative data home must refuse the daemon binary's config");
3041        assert!(
3042            matches!(error, BootstrapError::RunDir(_)),
3043            "expected a run-directory refusal, got {error}"
3044        );
3045        assert!(error.to_string().contains("XDG_DATA_HOME"), "{error}");
3046    }
3047
3048    #[test]
3049    fn privileged_operator_prompt_is_loaded_by_both_bootstrap_config_paths() {
3050        let _env_lock = ENV_LOCK.lock().unwrap();
3051        let _port = EnvGuard::unset(SUBC_PORT_ENV);
3052        let (_dir, conn_path) = temp_connection_file_path("privileged-operator-prompt");
3053        let config_path = conn_path.with_file_name("subc.jsonc");
3054        for (doc, expected) in [
3055            (None, daemon_config::PrivilegedOperatorPrompt::Log),
3056            (
3057                Some(r#"{"version":1}"#),
3058                daemon_config::PrivilegedOperatorPrompt::Log,
3059            ),
3060            (
3061                Some(r#"{"version":1,"privileged_operator_prompt":"log"}"#),
3062                daemon_config::PrivilegedOperatorPrompt::Log,
3063            ),
3064            (
3065                Some(r#"{"version":1,"privileged_operator_prompt":"enforce"}"#),
3066                daemon_config::PrivilegedOperatorPrompt::Enforce,
3067            ),
3068        ] {
3069            if let Some(doc) = doc {
3070                fs::write(&config_path, doc).unwrap();
3071            }
3072            assert_eq!(
3073                BootstrapConfig::from_env_with_daemon_config_path(&config_path)
3074                    .unwrap()
3075                    .privileged_operator_prompt,
3076                expected
3077            );
3078            assert_eq!(
3079                BootstrapConfig::new(&conn_path, 0)
3080                    .with_daemon_config_path(&config_path)
3081                    .unwrap()
3082                    .privileged_operator_prompt,
3083                expected
3084            );
3085        }
3086    }
3087
3088    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
3089    async fn privileged_operator_prompt_is_snapshotted_by_running_daemon() {
3090        let dir = unique_temp_dir("privileged-operator-prompt-bootstrap");
3091        let connection_path = dir.join("connection.json");
3092        let config_path = dir.join("subc.jsonc");
3093        fs::write(
3094            &config_path,
3095            r#"{"version":1,"privileged_operator_prompt":"enforce"}"#,
3096        )
3097        .unwrap();
3098        let config = config_with_a_short_socket(BootstrapConfig::new(&connection_path, 0))
3099            .with_daemon_config_path(&config_path)
3100            .unwrap();
3101        let connection_path = config.connection_file_path.clone();
3102        let daemon = AbortOnDrop::new(tokio::spawn(run_with_config(config)));
3103        tokio::time::timeout(Duration::from_secs(5), async {
3104            while !connection_path.exists() {
3105                assert!(
3106                    !daemon.handle.is_finished(),
3107                    "daemon exited before publishing"
3108                );
3109                tokio::task::yield_now().await;
3110            }
3111        })
3112        .await
3113        .unwrap();
3114        let info = connection_file::read(&connection_path).unwrap();
3115        let mut client = connect_from_info(&info).await.unwrap();
3116        authenticate_client(&mut client, &info, Duration::from_secs(5))
3117            .await
3118            .unwrap();
3119        for (correlation, mode, expected) in [
3120            (1, "enforce", Vec::new()),
3121            (2, "log", vec!["privileged_operator_prompt"]),
3122            (3, "log", vec!["privileged_operator_prompt"]),
3123        ] {
3124            fs::write(
3125                &config_path,
3126                serde_json::json!({"version":1,"privileged_operator_prompt":mode}).to_string(),
3127            )
3128            .unwrap();
3129            let frame = crate::Frame::build(
3130                subc_protocol::FrameType::Request,
3131                subc_protocol::Flags::new(false, subc_protocol::Priority::Interactive, false),
3132                0,
3133                0,
3134                correlation,
3135                serde_json::to_vec(&subc_control::ClientControlRequest::SupervisorRescan {
3136                    preview: true,
3137                })
3138                .unwrap(),
3139            )
3140            .unwrap();
3141            crate::write_frame(&mut client, &frame).await.unwrap();
3142            let reply =
3143                tokio::time::timeout(Duration::from_secs(5), crate::read_frame(&mut client))
3144                    .await
3145                    .unwrap()
3146                    .unwrap()
3147                    .unwrap();
3148            assert_eq!(reply.header.ty, subc_protocol::FrameType::Response);
3149            assert_eq!(reply.header.corr, correlation);
3150            let subc_control::ClientControlResponse::SupervisorRescan { result } =
3151                serde_json::from_slice(&reply.body).unwrap()
3152            else {
3153                panic!("expected a rescan response");
3154            };
3155            assert_eq!(result.restart_required, expected);
3156        }
3157    }
3158
3159    #[test]
3160    fn configured_port_uses_default_config_and_env_override() {
3161        let _env_lock = ENV_LOCK.lock().unwrap();
3162        let (_dir, conn_path) = temp_connection_file_path("daemon-config-port");
3163        let config_path = conn_path.with_file_name("subc.jsonc");
3164
3165        let _port = EnvGuard::unset(SUBC_PORT_ENV);
3166        assert_eq!(
3167            BootstrapConfig::from_env_with_daemon_config_path(&config_path)
3168                .unwrap()
3169                .port,
3170            DEFAULT_SUBC_PORT
3171        );
3172
3173        fs::write(&config_path, r#"{ "version": 1, "port": 8123 }"#).unwrap();
3174        assert_eq!(
3175            BootstrapConfig::from_env_with_daemon_config_path(&config_path)
3176                .unwrap()
3177                .port,
3178            8123
3179        );
3180
3181        let _port = EnvGuard::set_str(SUBC_PORT_ENV, "9012");
3182        assert_eq!(
3183            BootstrapConfig::from_env_with_daemon_config_path(&config_path)
3184                .unwrap()
3185                .port,
3186            9012
3187        );
3188    }
3189
3190    #[tokio::test]
3191    async fn singleton_does_not_accept_a_peer_that_answers_once_then_exits() {
3192        let (_dir, path) = temp_connection_file_path("departing-peer");
3193        let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
3194        let info = make_connection_info(listener.local_addr().unwrap().port());
3195        write_connection_file(&path, &info, None).unwrap();
3196        let peer = tokio::spawn(async move {
3197            let (mut stream, _) = listener.accept().await.unwrap();
3198            subc_transport::authenticate_server(
3199                &mut stream,
3200                &info.key,
3201                &info.daemon_id,
3202                &info.daemon_ver,
3203                PROBE_AUTH_DEADLINE,
3204            )
3205            .await
3206            .unwrap();
3207        });
3208        let outcome = ensure_singleton(&path, 0).await.unwrap();
3209        peer.await.unwrap();
3210        assert!(
3211            matches!(outcome, Outcome::Bound(_)),
3212            "a departing peer is not AlreadyRunning"
3213        );
3214    }
3215
3216    #[tokio::test]
3217    async fn second_singleton_probe_against_served_tcp_daemon_reports_already_running() {
3218        let (_dir, path) = temp_connection_file_path("already-running");
3219
3220        let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
3221        let server = start_server(bound);
3222
3223        let second = ensure_singleton(&path, 0).await.unwrap();
3224        assert!(matches!(second, Outcome::AlreadyRunning));
3225
3226        server.abort();
3227        let _ = server.await;
3228    }
3229
3230    #[tokio::test]
3231    async fn daemon_connection_file_publishes_protocol_wire_version() {
3232        let (_dir, path) = temp_connection_file_path("wire-version");
3233        let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
3234        assert_eq!(bound.connection_info.wire_version, Some(PROTOCOL_VERSION));
3235        assert_eq!(
3236            connection_file::read(&path).unwrap().wire_version,
3237            Some(PROTOCOL_VERSION)
3238        );
3239
3240        drop(bound.listeners);
3241    }
3242
3243    #[tokio::test]
3244    async fn stale_unbound_connection_file_is_reclaimed() {
3245        let (_dir, path) = temp_connection_file_path("stale-reclaim");
3246        let stale = TcpListener::bind((Ipv4Addr::LOCALHOST, 0)).await.unwrap();
3247        let stale_port = stale.local_addr().unwrap().port();
3248        drop(stale);
3249        let stale_info = make_connection_info(stale_port);
3250        write_atomic(&path, &stale_info).unwrap();
3251
3252        let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
3253        assert_ne!(bound.connection_info.key, stale_info.key);
3254        drop(bound.listeners);
3255    }
3256
3257    #[cfg(unix)]
3258    #[tokio::test]
3259    async fn ensure_singleton_reclaims_insecure_connection_file() {
3260        let (_dir, path) = temp_connection_file_path("insecure-reclaim");
3261        let stale = TcpListener::bind((Ipv4Addr::LOCALHOST, 0)).await.unwrap();
3262        let stale_port = stale.local_addr().unwrap().port();
3263        drop(stale);
3264        let stale_info = make_connection_info(stale_port);
3265        write_atomic(&path, &stale_info).unwrap();
3266        fs::set_permissions(&path, fs::Permissions::from_mode(0o644)).unwrap();
3267
3268        let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
3269        assert_ne!(bound.connection_info.key, stale_info.key);
3270        assert_ne!(bound.connection_info.daemon_id, stale_info.daemon_id);
3271        assert_owner_only_connection_file(&path);
3272
3273        drop(bound.listeners);
3274    }
3275
3276    #[tokio::test]
3277    async fn ensure_singleton_reclaims_non_loopback_connection_file() {
3278        let (_dir, path) = temp_connection_file_path("non-loopback-reclaim");
3279        let mut stale_info = make_connection_info(8757);
3280        stale_info.endpoints = vec![Endpoint {
3281            host: "192.0.2.10".to_owned(),
3282            port: 8757,
3283        }];
3284        write_atomic(&path, &stale_info).unwrap();
3285
3286        let config = config_with_a_short_socket(BootstrapConfig::new(&path, 0));
3287        let bound = expect_bound(ensure_singleton_with_config(config).await.unwrap());
3288        assert_ne!(bound.connection_info.key, stale_info.key);
3289        assert_ne!(bound.connection_info.daemon_id, stale_info.daemon_id);
3290        assert!(bound
3291            .connection_info
3292            .endpoints
3293            .iter()
3294            .all(|endpoint| endpoint.host.parse::<IpAddr>().unwrap().is_loopback()));
3295        assert_owner_only_connection_file(&path);
3296
3297        drop(bound.listeners);
3298    }
3299
3300    #[tokio::test]
3301    async fn ensure_singleton_reclaims_invalid_connection_file_shapes() {
3302        let mut unsupported_schema = make_connection_info(8757);
3303        unsupported_schema.schema = SCHEMA_VERSION + 1;
3304
3305        let mut empty_endpoints = make_connection_info(8757);
3306        empty_endpoints.endpoints.clear();
3307
3308        let mut short_key = make_connection_info(8757);
3309        short_key.key = vec![0x5A; MIN_KEY_LEN - 1];
3310
3311        let cases = vec![
3312            (
3313                "unsupported-schema",
3314                serde_json::to_vec(&unsupported_schema).unwrap(),
3315                Some(unsupported_schema),
3316            ),
3317            (
3318                "empty-endpoints",
3319                serde_json::to_vec(&empty_endpoints).unwrap(),
3320                Some(empty_endpoints),
3321            ),
3322            (
3323                "short-key",
3324                serde_json::to_vec(&short_key).unwrap(),
3325                Some(short_key),
3326            ),
3327            ("invalid-json", b"{not valid connection json".to_vec(), None),
3328        ];
3329
3330        for (label, contents, old_info) in cases {
3331            let (_dir, path) = temp_connection_file_path(label);
3332            write_raw_owner_only_connection_file(&path, &contents);
3333
3334            let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
3335            if let Some(old_info) = old_info {
3336                assert_ne!(bound.connection_info.key, old_info.key, "{label}");
3337                assert_ne!(
3338                    bound.connection_info.daemon_id, old_info.daemon_id,
3339                    "{label}"
3340                );
3341            }
3342            assert!(bound.connection_info.key.len() >= MIN_KEY_LEN, "{label}");
3343            assert_ne!(bound.connection_info.daemon_id, [0u8; 16], "{label}");
3344            assert_owner_only_connection_file(&path);
3345
3346            drop(bound.listeners);
3347        }
3348    }
3349
3350    #[tokio::test]
3351    async fn foreign_reused_port_connection_file_is_reclaimed_after_auth_probe_fails() {
3352        let (_dir, path) = temp_connection_file_path("foreign-reclaim");
3353        let foreign = TcpListener::bind((Ipv4Addr::LOCALHOST, 0)).await.unwrap();
3354        let foreign_port = foreign.local_addr().unwrap().port();
3355        write_atomic(&path, &make_connection_info(foreign_port)).unwrap();
3356        let foreign_task = tokio::spawn(async move {
3357            if let Ok((mut stream, _)) = foreign.accept().await {
3358                let mut buf = [0u8; 64];
3359                let _ = stream.read(&mut buf).await;
3360            }
3361        });
3362
3363        let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
3364        assert!(bound
3365            .connection_info
3366            .endpoints
3367            .iter()
3368            .all(|endpoint| endpoint.port != foreign_port));
3369
3370        drop(bound.listeners);
3371        let _ = foreign_task.await;
3372    }
3373
3374    #[tokio::test]
3375    async fn stale_start_lock_file_is_reclaimable() {
3376        let (_dir, path) = temp_connection_file_path("start-lock-stale-file");
3377        let lock_path = start_lock_path(&path);
3378        drop(open_owner_only_lock(&lock_path).unwrap());
3379        assert!(lock_path.is_file());
3380
3381        let lock = StartLock::acquire(&path).await.unwrap();
3382        assert!(lock_path.is_file());
3383
3384        drop(lock);
3385        assert!(lock_path.is_file());
3386    }
3387
3388    #[tokio::test]
3389    async fn held_start_lock_blocks_second_acquire_until_release() {
3390        let (_dir, path) = temp_connection_file_path("start-lock-held");
3391        let lock_path = start_lock_path(&path);
3392        let first = StartLock::acquire(&path).await.unwrap();
3393
3394        let err = match StartLock::acquire(&path).await {
3395            Ok(_) => panic!("second acquire while held must stay busy"),
3396            Err(err) => err,
3397        };
3398        assert!(matches!(
3399            err,
3400            BootstrapError::StartLockBusy {
3401                ref path,
3402                attempts: START_LOCK_RETRIES,
3403            } if path == &lock_path
3404        ));
3405
3406        drop(first);
3407
3408        let second = StartLock::acquire(&path)
3409            .await
3410            .expect("released advisory lock should be reclaimable");
3411        drop(second);
3412    }
3413
3414    #[tokio::test]
3415    async fn bind_conflict_on_fixed_port_fails_loud_without_reselecting() {
3416        let (_dir, path) = temp_connection_file_path("bind-conflict");
3417        let occupied = TcpListener::bind((Ipv4Addr::LOCALHOST, 0)).await.unwrap();
3418        let occupied_port = occupied.local_addr().unwrap().port();
3419
3420        let err = ensure_singleton(&path, occupied_port).await.unwrap_err();
3421        assert!(matches!(
3422            err,
3423            BootstrapError::Bind { ref source, .. } if source.kind() == io::ErrorKind::AddrInUse
3424        ));
3425        assert!(err.to_string().contains("set the port in config"));
3426
3427        drop(occupied);
3428    }
3429
3430    #[tokio::test]
3431    async fn ephemeral_ipv6_collision_still_serves_ipv4() {
3432        let (listeners, endpoints) = bind_loopback_with(0, |port| async move {
3433            let occupied = TcpListener::bind((Ipv6Addr::LOCALHOST, port)).await?;
3434            let result = TcpListener::bind((Ipv6Addr::LOCALHOST, port)).await;
3435            drop(occupied);
3436            result
3437        })
3438        .await
3439        .expect("an incidental IPv6 collision must not prevent an ephemeral daemon from starting");
3440        assert_eq!(endpoints.len(), 1);
3441        assert_eq!(endpoints[0].host, "127.0.0.1");
3442        TcpStream::connect((Ipv4Addr::LOCALHOST, endpoints[0].port))
3443            .await
3444            .unwrap();
3445        drop(listeners);
3446    }
3447
3448    #[tokio::test]
3449    async fn unpublished_boot_holds_singleton_lock_until_publication() {
3450        let dir = ScratchDir::new("ul");
3451        let path = dir.join("c");
3452        let config = config_with_a_short_socket(BootstrapConfig::new(&path, 0));
3453        let bound = expect_bound(ensure_singleton_inner(config, false).await.unwrap());
3454        assert!(!path.exists());
3455        let contender = open_owner_only_lock(&start_lock_path(&path)).unwrap();
3456        assert!(
3457            matches!(FileExt::try_lock(&contender), Err(TryLockError::WouldBlock)),
3458            "an unpublished boot must keep its singleton claim"
3459        );
3460        drop(bound);
3461        FileExt::try_lock(&contender).unwrap();
3462    }
3463
3464    #[tokio::test]
3465    async fn start_lock_drop_releases_claim_with_a_shared_file_description() {
3466        let dir = ScratchDir::new("sl");
3467        let path = dir.join("c");
3468        let config = config_with_a_short_socket(BootstrapConfig::new(&path, 0));
3469        let bound = expect_bound(ensure_singleton_inner(config, false).await.unwrap());
3470        // try_clone shares the open file description just as fork does,
3471        // without relying on the timing of another test's child spawn.
3472        let inherited = bound
3473            .publication_lock
3474            .as_ref()
3475            .unwrap()
3476            ._file
3477            .try_clone()
3478            .unwrap();
3479        let contender = open_owner_only_lock(&start_lock_path(&path)).unwrap();
3480        assert!(matches!(
3481            FileExt::try_lock(&contender),
3482            Err(TryLockError::WouldBlock)
3483        ));
3484        drop(bound);
3485        FileExt::try_lock(&contender).expect("dropping bootstrap releases even an inherited lock");
3486        drop(inherited);
3487    }
3488
3489    #[tokio::test]
3490    async fn readiness_probe_requires_an_authenticating_server() {
3491        let dir = ScratchDir::new("unserved-readiness");
3492        let bound = expect_bound(
3493            ensure_singleton_inner(BootstrapConfig::new(dir.join("connection.json"), 0), false)
3494                .await
3495                .unwrap(),
3496        );
3497        assert!(matches!(
3498            verify_serving(&bound.connection_info).await,
3499            Err(BootstrapError::StartupNotServing)
3500        ));
3501    }
3502
3503    #[cfg(unix)]
3504    #[tokio::test]
3505    async fn orphan_sweep_finishes_before_connection_file_publication() {
3506        use crate::live_children::{ExecutableIdentity, LiveChild};
3507        let dir = ScratchDir::new("pas");
3508        // A one-letter connection file name, and no long socket path: the sweep
3509        // this test covers runs before the local endpoint is bound, so a longer
3510        // name would only add a way for the test to fail on something else.
3511        let path = dir.join("c");
3512        let record = dir.join("live-children.json");
3513        let ready = dir.join("ready");
3514        let mut child = std::process::Command::new("/bin/sh")
3515            .args([
3516                "-c",
3517                "trap '' TERM; : > \"$0\"; read -r _",
3518                ready.to_str().unwrap(),
3519            ])
3520            .stdin(std::process::Stdio::piped())
3521            .env("XDG_DATA_HOME", dir.path())
3522            .env("XDG_RUNTIME_DIR", dir.path())
3523            .env("XDG_CONFIG_HOME", dir.path())
3524            .spawn()
3525            .unwrap();
3526        timeout(Duration::from_secs(5), async {
3527            while !ready.exists() {
3528                sleep(Duration::from_millis(5)).await;
3529            }
3530        })
3531        .await
3532        .unwrap();
3533        let observed = subc_os::Process::open(child.id())
3534            .unwrap()
3535            .unwrap()
3536            .observe()
3537            .unwrap();
3538        crate::live_children::write_record(
3539            &record,
3540            &[LiveChild {
3541                container: None,
3542                module_id: "old-sleep".into(),
3543                pid: child.id(),
3544                protocol: subc_control::ModuleProtocol::None,
3545                start_time: Some(observed.start_time),
3546                executable: observed.executable.map(ExecutableIdentity::from),
3547                cgroup_name: None,
3548            }],
3549        )
3550        .unwrap();
3551        let config = config_with_a_short_socket(BootstrapConfig::new(&path, 0))
3552            .with_live_children_record(&record);
3553        let daemon = tokio::spawn(run_with_config(config));
3554        timeout(Duration::from_secs(5), async {
3555            while !path.exists() {
3556                sleep(Duration::from_millis(5)).await;
3557            }
3558        })
3559        .await
3560        .unwrap();
3561        assert!(path.exists());
3562        let ended = child.try_wait().unwrap().is_some();
3563        if !ended {
3564            child.kill().unwrap();
3565        }
3566        child.wait().unwrap();
3567        if ended {
3568            assert!(matches!(probe_existing(&path).await.unwrap(), Probe::Live));
3569        }
3570        daemon.abort();
3571        let _ = daemon.await;
3572        assert!(
3573            ended,
3574            "the published daemon must not still owe its orphan sweep"
3575        );
3576    }
3577
3578    /// Publishing the connection file fsyncs, which can take seconds on a busy
3579    /// disk, so the daemon must do it off the async worker thread. The hook
3580    /// blocks inside the write until another task on the same single-threaded
3581    /// runtime has run. Writing on the worker thread would starve that task: the
3582    /// hook then gives up after its deadline and the test fails, as it does if
3583    /// startup publishes without going through the hooked write at all.
3584    #[tokio::test(flavor = "current_thread")]
3585    async fn connection_file_publication_leaves_the_async_runtime_free() {
3586        use std::sync::atomic::{AtomicBool, Ordering};
3587
3588        let (_dir, path) = temp_connection_file_path("publish-off-runtime");
3589        let write_started = Arc::new(AtomicBool::new(false));
3590        let (progress_tx, progress_rx) = std::sync::mpsc::channel::<()>();
3591        let (verdict_tx, verdict_rx) = std::sync::mpsc::channel::<bool>();
3592        let hook_started = Arc::clone(&write_started);
3593        connection_file_write_hook::install(&path, move || {
3594            hook_started.store(true, Ordering::SeqCst);
3595            let progressed = progress_rx.recv_timeout(Duration::from_secs(5)).is_ok();
3596            let _ = verdict_tx.send(progressed);
3597        });
3598        // Signals only after the write has started, so a message can only come
3599        // from this task running while the write is still blocked.
3600        let ticker = tokio::spawn(async move {
3601            loop {
3602                if write_started.load(Ordering::SeqCst) {
3603                    let _ = progress_tx.send(());
3604                    return;
3605                }
3606                sleep(Duration::from_millis(5)).await;
3607            }
3608        });
3609
3610        let daemon = tokio::spawn(run_with_config(BootstrapConfig::new(&path, 0)));
3611        let published = timeout(Duration::from_secs(15), async {
3612            while !path.exists() && !daemon.is_finished() {
3613                sleep(Duration::from_millis(5)).await;
3614            }
3615        })
3616        .await;
3617        let daemon_exited = daemon.is_finished();
3618        daemon.abort();
3619        let daemon_result = daemon.await;
3620        ticker.abort();
3621
3622        assert!(
3623            published.is_ok() && !daemon_exited,
3624            "the daemon must start and publish its connection file: {daemon_result:?}"
3625        );
3626        match verdict_rx.try_recv() {
3627            Ok(true) => {}
3628            Ok(false) => panic!(
3629                "no other task on the runtime ran while the connection file was being written: the write is blocking the async worker thread"
3630            ),
3631            Err(_) => panic!("startup published the connection file without the hooked write"),
3632        }
3633    }
3634
3635    /// A write that never reports back must stop startup with an error naming
3636    /// the connection file, never be taken as published.
3637    #[tokio::test]
3638    async fn a_panicking_connection_file_write_fails_startup() {
3639        let (_dir, path) = temp_connection_file_path("publish-panics");
3640        connection_file_write_hook::install(&path, || {
3641            panic!("deliberate panic inside the connection-file write")
3642        });
3643
3644        let error = timeout(
3645            Duration::from_secs(15),
3646            run_with_config(BootstrapConfig::new(&path, 0)),
3647        )
3648        .await
3649        .expect("startup must end once the write task fails")
3650        .expect_err("startup must fail when the connection-file write panics");
3651
3652        assert!(
3653            matches!(&error, BootstrapError::ConnectionFileWrite { path: failed, .. } if failed == &path),
3654            "expected a connection-file write error, got {error:?}"
3655        );
3656        let message = error.to_string();
3657        assert!(message.contains(&path.display().to_string()), "{message}");
3658        assert!(message.contains("panicked"), "{message}");
3659        assert!(!path.exists(), "no connection file may be published");
3660    }
3661
3662    #[cfg(target_os = "macos")]
3663    #[test]
3664    fn nofile_raise_worker() {
3665        let Some(path) = std::env::var_os("SUBC_NOFILE_TEST_FILE") else {
3666            return;
3667        };
3668        let (_, hard) = rlimit::Resource::NOFILE.get().unwrap();
3669        rlimit::Resource::NOFILE.set(256, hard).unwrap();
3670        raise_nofile_limit_with_ceiling(Some(10240));
3671        std::fs::write(path, rlimit::Resource::NOFILE.get().unwrap().0.to_string()).unwrap();
3672    }
3673
3674    #[cfg(target_os = "macos")]
3675    #[test]
3676    fn macos_nofile_raise_clamps_to_kernel_ceiling() {
3677        let dir = ScratchDir::new("nofile-ceiling");
3678        let result = dir.join("soft");
3679        let status = std::process::Command::new(std::env::current_exe().unwrap())
3680            .args(["--exact", "bootstrap::tests::nofile_raise_worker"])
3681            .env("SUBC_NOFILE_TEST_FILE", &result)
3682            .env("XDG_DATA_HOME", dir.path())
3683            .env("XDG_RUNTIME_DIR", dir.path())
3684            .env("XDG_CONFIG_HOME", dir.path())
3685            .status()
3686            .unwrap();
3687        assert!(status.success());
3688        let soft: u64 = std::fs::read_to_string(result).unwrap().parse().unwrap();
3689        assert!(
3690            soft > 256 && soft <= 10240,
3691            "the inherited limit must be raised within the kernel ceiling; got {soft}"
3692        );
3693    }
3694
3695    #[tokio::test]
3696    async fn key_rotation_republishes_new_material_and_old_file_fails_auth() {
3697        let (_dir, path) = temp_connection_file_path("key-rotation");
3698        let first = expect_bound(ensure_singleton(&path, 0).await.unwrap());
3699        let old_info = first.connection_info.clone();
3700        let fixed_port = old_info.endpoints[0].port;
3701        drop(first.listeners);
3702
3703        let second = expect_bound(ensure_singleton(&path, fixed_port).await.unwrap());
3704        let new_info = second.connection_info.clone();
3705        assert_ne!(old_info.key, new_info.key);
3706        assert_ne!(old_info.daemon_id, new_info.daemon_id);
3707        let server = start_server(second);
3708
3709        let mut old_stream = connect_from_info(&old_info).await.unwrap();
3710        let old_auth = authenticate_client(&mut old_stream, &old_info, PROBE_AUTH_DEADLINE).await;
3711        assert!(
3712            old_auth.is_err(),
3713            "old key must not authenticate after restart"
3714        );
3715
3716        let reread = connection_file::read(&path).unwrap();
3717        let mut new_stream = connect_from_info(&reread).await.unwrap();
3718        authenticate_client(&mut new_stream, &reread, PROBE_AUTH_DEADLINE)
3719            .await
3720            .unwrap();
3721
3722        server.abort();
3723        let _ = server.await;
3724    }
3725
3726    #[cfg(unix)]
3727    #[tokio::test]
3728    async fn published_connection_file_permissions_are_owner_only() {
3729        let (_dir, path) = temp_connection_file_path("permissions");
3730        let bound = expect_bound(ensure_singleton(&path, 0).await.unwrap());
3731
3732        let mode = fs::metadata(&path).unwrap().permissions().mode() & 0o777;
3733        assert_eq!(mode, 0o600);
3734
3735        drop(bound.listeners);
3736    }
3737
3738    /// A config with the socket in `dir`'s own short subdirectory.
3739    ///
3740    /// A Unix socket path is limited to just over a hundred bytes, and this
3741    /// harness's temporary directories use nearly all of that, so a test that
3742    /// boots a real daemon passes a short directory and a one-letter connection
3743    /// file name; anything longer makes the bind refuse. On Windows there is no
3744    /// socket path to keep short, so the config is returned unchanged.
3745    fn config_with_a_short_socket(config: BootstrapConfig) -> BootstrapConfig {
3746        let dir = config
3747            .connection_file_path
3748            .parent()
3749            .map(Path::to_path_buf)
3750            .unwrap_or_else(|| PathBuf::from("."));
3751        #[cfg(unix)]
3752        let config = BootstrapConfig {
3753            connection_file_path: dir.join("c"),
3754            local_listener: LocalListenerConfig::in_socket_dir(dir.join("s")),
3755            ..config
3756        };
3757        #[cfg(not(unix))]
3758        let _ = dir;
3759        config
3760    }
3761
3762    /// The shipped binary puts the socket in its own run directory. The file is
3763    /// named from the connection file's path, so a restart finds it again.
3764    #[cfg(unix)]
3765    #[test]
3766    fn the_daemon_binary_puts_its_socket_in_the_run_directory() {
3767        let _env_lock = ENV_LOCK.lock().unwrap();
3768        let root = ScratchDir::new("binary-local-socket");
3769        let data_home = root.join("data");
3770        let _data = EnvGuard::set("XDG_DATA_HOME", &data_home);
3771        let _config = EnvGuard::set("XDG_CONFIG_HOME", &root.join("config"));
3772        let _port = EnvGuard::unset(SUBC_PORT_ENV);
3773
3774        let config = BootstrapConfig::from_env_for_daemon_binary().unwrap();
3775
3776        let run_dir = data_home.join("cortexkit").join("run");
3777        assert_eq!(
3778            config.local_listener.configured_dir(),
3779            Some(run_dir.as_path()),
3780            "the binary serves from its run directory"
3781        );
3782        let socket = local_listener::socket_path_for(Some(&run_dir), &config.connection_file_path)
3783            .expect("the binary config names a directory");
3784        assert_eq!(socket.parent().unwrap(), run_dir);
3785        assert_eq!(
3786            socket.file_name().unwrap().to_string_lossy(),
3787            "subc.sock",
3788            "a fixed name: the run directory belongs to one data home"
3789        );
3790    }
3791
3792    /// An in-process daemon that names no directory serves TCP only: it publishes
3793    /// no local endpoint, creates nothing under the data home, and says nothing is
3794    /// unavailable, because nothing was asked for.
3795    /// The environment lock covers only the setup: the daemon under test reads
3796    /// its data home from the process environment while it boots, and holding
3797    /// the lock across that await would block every other test that sets one.
3798    #[cfg(unix)]
3799    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
3800    async fn an_in_process_daemon_without_a_configured_directory_serves_no_local_endpoint() {
3801        let root = ScratchDir::new("in-process-no-local-socket");
3802        let data_home = root.join("data");
3803        fs::create_dir_all(&data_home).unwrap();
3804        let connection_path = root.join("c");
3805        let _data;
3806        let _config;
3807        let _port;
3808        {
3809            let _env_lock = ENV_LOCK.lock().unwrap();
3810            _data = EnvGuard::set("XDG_DATA_HOME", &data_home);
3811            _config = EnvGuard::set("XDG_CONFIG_HOME", &root.join("config"));
3812            _port = EnvGuard::unset(SUBC_PORT_ENV);
3813        }
3814
3815        let bound = expect_bound(ensure_singleton(&connection_path, 0).await.unwrap());
3816
3817        assert!(bound.local_endpoint.is_none());
3818        assert!(bound.local_endpoint_unavailable.is_none());
3819        let published = connection_file::read_with_local_endpoint(&connection_path).unwrap();
3820        assert!(published.local_endpoint.is_none());
3821        assert!(
3822            local_listener::socket_path_for(None, &connection_path).is_none(),
3823            "no configured directory, no socket"
3824        );
3825        let mut leftovers = Vec::new();
3826        let mut pending = vec![data_home.clone()];
3827        while let Some(dir) = pending.pop() {
3828            for entry in fs::read_dir(&dir).unwrap() {
3829                let entry = entry.unwrap();
3830                let path = entry.path();
3831                if entry.file_type().unwrap().is_dir() {
3832                    pending.push(path.clone());
3833                }
3834                if entry.file_name().to_string_lossy().ends_with(".sock") {
3835                    leftovers.push(path);
3836                }
3837            }
3838        }
3839        assert!(
3840            leftovers.is_empty(),
3841            "an in-process daemon must not create a socket in the data home: {leftovers:?}"
3842        );
3843    }
3844
3845    /// A local endpoint that cannot be bound does not stop the daemon: it keeps
3846    /// serving TCP, publishes no local endpoint, and reports the reason where an
3847    /// operator reads daemon facts. Every module stays up for a feature nothing
3848    /// has to use.
3849    #[cfg(unix)]
3850    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
3851    async fn a_daemon_whose_socket_path_is_too_long_still_serves_tcp() {
3852        let dir = ScratchDir::new("tcp-only");
3853        let canonical = std::fs::canonicalize(&dir).unwrap();
3854        let connection_path = canonical.join("c");
3855        // The socket directory is deep enough that the socket's path is over the
3856        // Unix socket limit, so binding it must fail.
3857        let deep = canonical.join("x".repeat(80));
3858        fs::create_dir(&deep).unwrap();
3859        fs::set_permissions(&deep, fs::Permissions::from_mode(0o700)).unwrap();
3860        let config =
3861            BootstrapConfig::new(&connection_path, 0).with_local_listener_socket_dir(&deep);
3862        let bound = expect_bound(ensure_singleton_with_config(config).await.unwrap());
3863        assert!(
3864            bound.local_endpoint.is_none(),
3865            "a socket that cannot be bound is not published"
3866        );
3867        assert!(
3868            bound
3869                .local_endpoint_unavailable
3870                .as_deref()
3871                .is_some_and(|reason| reason.contains("too long")),
3872            "{:?}",
3873            bound.local_endpoint_unavailable
3874        );
3875        let published = connection_file::read_with_local_endpoint(&connection_path).unwrap();
3876        assert!(published.local_endpoint.is_none());
3877        assert!(
3878            !published.info.endpoints.is_empty(),
3879            "TCP is still published"
3880        );
3881
3882        // The daemon serves TCP as usual.
3883        let server = start_server(bound);
3884        let info = connection_file::read(&connection_path).unwrap();
3885        let mut client = connect_from_info(&info).await.unwrap();
3886        authenticate_client(&mut client, &info, PROBE_AUTH_DEADLINE)
3887            .await
3888            .expect("a TCP client still completes the handshake");
3889        server.abort();
3890        let _ = server.await;
3891    }
3892
3893    /// The published file names the local endpoint the daemon actually serves,
3894    /// and a client that trusts the file completes the handshake over it.
3895    #[cfg(unix)]
3896    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
3897    async fn published_connection_file_names_the_served_local_endpoint() {
3898        let (_dir, path) = temp_connection_file_path("local-endpoint");
3899        // The socket goes in the system temporary directory rather than inside
3900        // the harness's own directory, whose long lockdown-sandbox path leaves
3901        // no room inside the Unix socket path limit.
3902        let short_socket_dir = std::env::temp_dir().join(format!("subc-bt-{}", std::process::id()));
3903        let config =
3904            BootstrapConfig::new(&path, 0).with_local_listener_socket_dir(&short_socket_dir);
3905        let bound = expect_bound(ensure_singleton_with_config(config).await.unwrap());
3906        let endpoint = connection_file::read_with_local_endpoint(&path)
3907            .unwrap()
3908            .local_endpoint
3909            .expect("the daemon serves a local endpoint, so the file names it");
3910        assert_eq!(Some(&endpoint), bound.local_endpoint.as_ref());
3911        let LocalEndpoint::Unix { path: socket_path } = &endpoint else {
3912            panic!("a Unix daemon publishes a socket path");
3913        };
3914        assert_eq!(socket_path.parent().unwrap(), short_socket_dir);
3915        assert_eq!(
3916            fs::metadata(socket_path).unwrap().permissions().mode() & 0o777,
3917            0o600
3918        );
3919
3920        let router = Arc::new(Router::with_default_self_handler());
3921        let auth = auth_for(&bound.connection_info);
3922        let server = tokio::spawn(serve_local_listener(
3923            bound.local_listener.expect("the local listener is bound"),
3924            router,
3925            auth,
3926        ));
3927        let mut client = tokio::net::UnixStream::connect(socket_path).await.unwrap();
3928        authenticate_client(&mut client, &bound.connection_info, Duration::from_secs(5))
3929            .await
3930            .expect("a socket client completes the same handshake as a TCP client");
3931
3932        server.abort();
3933        drop(bound.listeners);
3934    }
3935
3936    /// A configured directory that is not a directory refuses at bind rather
3937    /// than publishing an endpoint nothing serves.
3938    #[cfg(unix)]
3939    #[test]
3940    fn a_socket_directory_that_is_a_file_refuses() {
3941        let dir = ScratchDir::new("local-listener-file-as-dir");
3942        let file = dir.join("not-a-directory");
3943        fs::write(&file, b"x").unwrap();
3944        let error = local_listener::bind(
3945            &LocalListenerConfig::in_socket_dir(file),
3946            &dir.join("connection.json"),
3947        )
3948        .expect_err("a path that is not a directory cannot hold the socket");
3949        assert!(matches!(error, ServerError::LocalAccept { .. }), "{error}");
3950    }
3951
3952    /// The socket sits directly in the configured directory under one fixed
3953    /// name, and a directory that would put the path over the platform limit is
3954    /// refused rather than shortened around.
3955    #[cfg(unix)]
3956    #[test]
3957    fn the_socket_path_stays_within_the_unix_limit() {
3958        let run_dir = PathBuf::from("/tmp/short-run");
3959        assert!(
3960            local_listener::socket_path_for(None, &run_dir.join("one.json")).is_none(),
3961            "no configured directory means no socket"
3962        );
3963        let socket = local_listener::socket_path_for(Some(&run_dir), &run_dir.join("one.json"))
3964            .expect("a configured directory gets a socket");
3965        assert_eq!(socket, run_dir.join("subc.sock"));
3966        assert_eq!(
3967            socket,
3968            local_listener::socket_path_for(Some(&run_dir), &run_dir.join("two.json")).unwrap(),
3969            "the name does not come from the connection file"
3970        );
3971
3972        let deep = PathBuf::from(format!("/tmp/{}", "x".repeat(60)));
3973        let long = local_listener::socket_path_for(Some(&deep), &deep.join("conn.json"))
3974            .expect("a configured directory gets a socket");
3975        assert!(
3976            long.as_os_str().len() < 104,
3977            "{} is {} bytes",
3978            long.display(),
3979            long.as_os_str().len()
3980        );
3981    }
3982}