failproofai 1.0.0-beta.2 → 1.0.0-beta.3

package/.next/standalone/crates/failproofaid/src/main.rs

@@ -1,3 +1,5 @@
+mod cloud_client;
+pub mod cloud_policies;
 mod lock;
 mod paths;
 mod server;
@@ -5,6 +7,7 @@ mod worker;
 
 use std::sync::Arc;
 use std::sync::atomic::{AtomicBool, Ordering};
+use std::time::Duration;
 
 fn main() {
     let args: Vec<String> = std::env::args().collect();
@@ -19,7 +22,27 @@ fn main() {
     }
 }
 
+/// Install a log subscriber, or the collector's diagnostics go nowhere.
+///
+/// `tracing` drops every event when no subscriber is registered, silently. The
+/// uploader reports "the server accepted the request but stored NONE of its
+/// events" through it — the single most important signal that a transform is
+/// systematically malformed — so without this that failure is invisible.
+///
+/// Writes to stderr, which is where the daemon's existing `eprintln!` output
+/// already goes and what systemd/launchd capture. `RUST_LOG` overrides the
+/// default level.
+fn init_logging() {
+    use tracing_subscriber::EnvFilter;
+    let filter = EnvFilter::try_from_default_env().unwrap_or_else(|_| EnvFilter::new("info"));
+    let _ = tracing_subscriber::fmt()
+        .with_env_filter(filter)
+        .with_writer(std::io::stderr)
+        .try_init();
+}
+
 fn run() -> Result<(), Box<dyn std::error::Error>> {
+    init_logging();
     let lock_path = paths::lock_path()?;
     paths::ensure_run_dir()?;
     let _singleton = lock::acquire(&lock_path)?;
@@ -40,16 +63,442 @@ fn run() -> Result<(), Box<dyn std::error::Error>> {
         let warm_worker = worker.clone();
         std::thread::spawn(move || warm_worker.warm());
     }
+    // The shutdown flag and its signal handler are installed BEFORE anything
+    // long-running starts, so the collector below can observe the same flag
+    // the socket server does. One SIGTERM then stops both; a second signal
+    // path would be one more thing to get wrong during shutdown.
+    let shutdown = Arc::new(AtomicBool::new(false));
+    install_signal_handler(shutdown.clone());
+
+    // Cloud policy integrity is a maintenance-lane responsibility, never a
+    // hook-path operation. The monitor is useful before cloud transport lands:
+    // it keeps the active generation and content-addressed artifact cache in
+    // agreement and reports when both verified copies have been lost.
+    let cloud_policy_store = cloud_policies::PolicyStore::new(paths::cloud_managed_policy_dir()?);
+    // One lane, resolving enrolment per tick rather than once at startup.
+    // `failproofai config --connect` writes its credential file without root,
+    // and this is a SYSTEM unit — so noticing it only on restart would put
+    // `sudo systemctl restart` back into a flow built to avoid it. The lane
+    // also does integrity repair whether or not the machine is enrolled.
+    let cloud_monitor = cloud_client::spawn_maintenance(
+        cloud_policy_store,
+        shutdown.clone(),
+        cloud_client::poll_interval_from_env(),
+        cloud_policy_reconcile_interval(),
+    );
+
+    // Log/hook collection. Runs on its own thread with its own Tokio runtime
+    // so it can never share fate with the accept loop — this daemon fails
+    // closed, so a collector fault would otherwise deny every tool call on the
+    // machine (see fpai_collect::supervisor). It observes the same `shutdown`
+    // flag the server and the cloud monitor do, so one SIGTERM stops all three.
+    //
+    // INERT until ingest is configured: `collector_tasks()` returns an empty
+    // list, `spawn_supervised` then declines to start a thread or a runtime at
+    // all, and the daemon behaves byte-for-byte as it did before this existed.
+    let collector = fpai_collect::spawn_supervised(collector_tasks(), shutdown.clone());
+
     let srv = server::Server::bind(&socket_path, worker)?;
     eprintln!("[failproofaid] listening on {}", socket_path.display());
 
-    let shutdown = Arc::new(AtomicBool::new(false));
-    install_signal_handler(shutdown.clone());
+    let run_result = srv.run_until(shutdown);
+
+    // Drain both background workers before returning. The collector gets a
+    // bounded budget rather than an unbounded join: process exit must not wait
+    // on a task that has wedged. Done before `?` so a server error still gets
+    // the collector a chance to flush instead of dropping buffered events.
+    if let Some(collector) = collector {
+        collector.join_with_flush(fpai_collect::DEFAULT_FLUSH_BUDGET);
+    }
+    let _ = cloud_monitor.join();
 
-    srv.run_until(shutdown)?;
+    run_result?;
     Ok(())
 }
 
+/// The collector tasks to supervise for this process.
+///
+/// Returns an empty list — and therefore starts no thread and no runtime —
+/// unless an ingest credential is configured AND at least one stream is
+/// enabled. That is the normal state, so an install that has not opted in
+/// pays nothing for this code path existing.
+///
+/// A configuration error is logged and treated as "off" rather than being
+/// propagated: collection failing to start must never stop the daemon from
+/// serving the socket, because the CLI fails closed and a daemon that refused
+/// to boot over a malformed `ingest.json` would deny every tool call on the
+/// machine. The error is loud so it is fixable, not silent.
+fn collector_tasks() -> Vec<fpai_collect::TaskSpec> {
+    let home = match paths::failproofai_home() {
+        Ok(home) => home,
+        Err(err) => {
+            eprintln!("[failproofaid] collector disabled: {err}");
+            return Vec::new();
+        }
+    };
+
+    let cfg = match fpai_collect::config::load(&home) {
+        Ok(cfg) => cfg,
+        Err(err) => {
+            eprintln!("[failproofaid] collector disabled: {err}");
+            return Vec::new();
+        }
+    };
+
+    if !cfg.is_enabled() {
+        return Vec::new();
+    }
+
+    // `is_enabled()` already established there is one.
+    let Some(ingest) = cfg.ingest.clone() else {
+        return Vec::new();
+    };
+
+    let uploader = match fpai_collect::Uploader::new(
+        ingest.url.clone(),
+        ingest.key.clone(),
+        cfg.failed_dir.clone(),
+    ) {
+        Ok(u) => std::sync::Arc::new(u),
+        Err(err) => {
+            eprintln!("[failproofaid] collector disabled: {err}");
+            return Vec::new();
+        }
+    };
+
+    eprintln!(
+        "[failproofaid] collector enabled: sessions={} hooks={} ({:?}) -> {}",
+        cfg.settings.sessions, cfg.settings.hooks, cfg.settings.hooks_verbosity, ingest.url,
+    );
+
+    // One `Delivery` shared by both tasks, so they share an upload semaphore
+    // and an in-flight set. Separate ones would let the watcher and a
+    // concurrent sweep POST the same batch twice.
+    let delivery = std::sync::Arc::new(fpai_collect::Delivery::new(uploader));
+
+    let watch_delivery = delivery.clone();
+    let watch_dirs = cfg.spool_dirs.clone();
+    let sweep_delivery = delivery;
+    let sweep_dirs = cfg.spool_dirs.clone();
+    let failed_dir = cfg.failed_dir.clone();
+
+    let mut tasks = Vec::new();
+
+    // Install the process health registry BEFORE any source starts, so no poll
+    // reports into a registry that does not exist yet. The writer publishes it
+    // on an interval and deletes it on clean shutdown — absence means "no
+    // daemon", where a stale file makes a stopped daemon look like a running
+    // one whose sources all went quiet.
+    let health = std::sync::Arc::new(fpai_collect::Health::new());
+    fpai_collect::health::install(health.clone());
+    let health_file = fpai_collect::health_path(&home);
+    tasks.push(fpai_collect::TaskSpec::new("health", move |sd| {
+        fpai_collect::health::writer_task(
+            health.clone(),
+            health_file.clone(),
+            fpai_collect::health::WRITE_INTERVAL,
+            sd,
+        )
+    }));
+
+    if cfg.settings.hooks {
+        // Hook activity: one source covering every CLI failproofai is
+        // installed in, because the store is CLI-agnostic — each row names its
+        // own integration. Reads the same store the dashboard's activity tab
+        // does, and never writes to it.
+        let store_dir = home.join("cache").join("hook-activity");
+        let state_dir = home.join("cursors").join("hooks");
+        let spool_dir = cfg.own_spool_dir.clone();
+        let verbosity = cfg.settings.hooks_verbosity;
+        let environment = cfg.settings.environment.clone();
+        tasks.push(fpai_collect::TaskSpec::new("hook-activity", move |sd| {
+            fpai_collect::sources::hooks::run(
+                store_dir.clone(),
+                state_dir.clone(),
+                spool_dir.clone(),
+                verbosity,
+                environment.clone(),
+                sd,
+            )
+        }));
+    }
+
+    if cfg.settings.sessions {
+        // Session transcripts, gated on the `sessions` opt-in because — unlike
+        // hook activity — these carry prompts, file contents and whatever was
+        // pasted into a terminal.
+        //
+        // Every source is registered the same way regardless of which engine it
+        // uses, so adding one is a row here plus its own module.
+        let spool = cfg.own_spool_dir.clone();
+        let env = cfg.settings.environment.clone();
+        let cursors = home.join("cursors");
+
+        use fpai_collect::sources::{claude, codex, copilot, openclaw, pi};
+        file_source(
+            &mut tasks,
+            "claude",
+            claude::FORMAT,
+            vec![claude_projects_root()],
+            claude::DEFAULT_AGENT_ID,
+            &spool,
+            &cursors,
+            &env,
+        );
+        // Subagent transcripts live under the SAME root, claimed by a second
+        // format. A separate source, not a second predicate: `is_source_file`
+        // is a bare fn, and each source needs its own cursor store — one store
+        // writes its whole map atomically, so two sharing a file would clobber
+        // each other and the loser would re-ship from zero after every restart.
+        file_source(
+            &mut tasks,
+            "claude-subagent",
+            claude::SUBAGENT_FORMAT,
+            vec![claude_projects_root()],
+            claude::SUBAGENT_DEFAULT_AGENT_ID,
+            &spool,
+            &cursors,
+            &env,
+        );
+        file_source(
+            &mut tasks,
+            "codex",
+            codex::FORMAT,
+            vec![codex_sessions_root()],
+            codex::DEFAULT_AGENT_ID,
+            &spool,
+            &cursors,
+            &env,
+        );
+        file_source(
+            &mut tasks,
+            "copilot",
+            copilot::FORMAT,
+            vec![copilot::session_state_root()],
+            copilot::DEFAULT_AGENT_ID,
+            &spool,
+            &cursors,
+            &env,
+        );
+        file_source(
+            &mut tasks,
+            "openclaw",
+            openclaw::FORMAT,
+            openclaw::default_roots(),
+            openclaw::DEFAULT_AGENT_ID,
+            &spool,
+            &cursors,
+            &env,
+        );
+        file_source(
+            &mut tasks,
+            "pi",
+            pi::FORMAT,
+            vec![pi::sessions_root()],
+            pi::DEFAULT_AGENT_ID,
+            &spool,
+            &cursors,
+            &env,
+        );
+
+        use fpai_collect::sources::{goose, hermes, opencode};
+        sqlite_source(
+            &mut tasks,
+            "goose",
+            goose::FORMAT,
+            goose::db_path(),
+            goose::DEFAULT_AGENT_ID,
+            &spool,
+            cursors.join("goose"),
+            &env,
+        );
+        sqlite_source(
+            &mut tasks,
+            "opencode",
+            opencode::FORMAT,
+            opencode::default_db_path(),
+            opencode::DEFAULT_AGENT_ID,
+            &spool,
+            cursors.join("opencode"),
+            &env,
+        );
+
+        // Hermes profiles are SEPARATE databases, and the SQLite poller keys its
+        // cursor on a fixed synthetic id — so two profiles sharing one state
+        // directory would clobber each other's watermark and each would re-read
+        // from zero after every restart. Each database gets its own.
+        for (i, db) in hermes::default_db_paths().into_iter().enumerate() {
+            let state = cursors.join("hermes").join(profile_dir_name(&db, i));
+            sqlite_source(
+                &mut tasks,
+                "hermes",
+                hermes::FORMAT,
+                db,
+                hermes::DEFAULT_AGENT_ID,
+                &spool,
+                state,
+                &env,
+            );
+        }
+    }
+
+    tasks.extend([
+        // Latency: delivers a batch within milliseconds of it being published.
+        fpai_collect::TaskSpec::new("spool-watcher", move |sd| {
+            fpai_collect::delivery::watch(watch_delivery.clone(), watch_dirs.clone(), sd)
+        }),
+        // Guarantee: delivers anything the watcher never saw — published while
+        // the daemon was stopped, or on a filesystem with no event support —
+        // and retries parked batches on a much slower cadence.
+        fpai_collect::TaskSpec::new("spool-sweeper", move |sd| {
+            fpai_collect::delivery::sweep(
+                sweep_delivery.clone(),
+                sweep_dirs.clone(),
+                failed_dir.clone(),
+                sd,
+            )
+        }),
+    ]);
+
+    tasks
+}
+
+/// Where Claude Code keeps its transcripts.
+///
+/// `CLAUDE_PROJECTS_PATH` overrides it, matching the env var the TypeScript
+/// side already honours, so a machine that has moved the directory is captured
+/// by both halves rather than one.
+fn claude_projects_root() -> std::path::PathBuf {
+    if let Some(p) = std::env::var_os("CLAUDE_PROJECTS_PATH") {
+        return std::path::PathBuf::from(p);
+    }
+    let home = std::env::var_os("HOME")
+        .map(std::path::PathBuf::from)
+        .unwrap_or_default();
+    home.join(".claude").join("projects")
+}
+
+/// Register one file-tailing source.
+#[allow(clippy::too_many_arguments)]
+fn file_source(
+    tasks: &mut Vec<fpai_collect::TaskSpec>,
+    name: &'static str,
+    format: fpai_collect::filetail::Format,
+    roots: Vec<std::path::PathBuf>,
+    default_agent_id: &'static str,
+    spool_dir: &std::path::Path,
+    cursor_root: &std::path::Path,
+    environment: &str,
+) {
+    let spool_dir = spool_dir.to_path_buf();
+    // One cursor store per source, never shared: the store writes its whole map
+    // atomically, so two sources sharing a file would clobber each other and the
+    // loser would re-read from zero after every restart.
+    let state_dir = cursor_root.join(name);
+    let environment = environment.to_string();
+    tasks.push(fpai_collect::TaskSpec::new(name, move |sd| {
+        fpai_collect::filetail::run(
+            fpai_collect::filetail::Spec {
+                format,
+                roots: roots.clone(),
+                spool_dir: spool_dir.clone(),
+                state_dir: state_dir.clone(),
+                poll_interval: std::time::Duration::from_secs(2),
+                params: fpai_collect::filetail::Params {
+                    agent_id: default_agent_id.to_string(),
+                    environment: environment.clone(),
+                    end_idle_mins: 10,
+                    max_read_bytes: 32 * 1024 * 1024,
+                    max_batch_bytes: fpai_collect::spool::DEFAULT_MAX_BATCH_BYTES,
+                    // Never the whole history by default. A normal machine holds
+                    // hundreds of megabytes of transcripts, and shipping all of
+                    // it on first start is not a reasonable default.
+                    since_days: Some(7),
+                },
+            },
+            sd,
+        )
+    }));
+}
+
+/// Register one SQLite-polling source.
+#[allow(clippy::too_many_arguments)]
+fn sqlite_source(
+    tasks: &mut Vec<fpai_collect::TaskSpec>,
+    name: &'static str,
+    format: fpai_collect::sqlitepoll::SqliteFormat,
+    db_path: std::path::PathBuf,
+    default_agent_id: &'static str,
+    spool_dir: &std::path::Path,
+    state_dir: std::path::PathBuf,
+    environment: &str,
+) {
+    let spool_dir = spool_dir.to_path_buf();
+    let environment = environment.to_string();
+    tasks.push(fpai_collect::TaskSpec::new(name, move |sd| {
+        fpai_collect::sqlitepoll::run(
+            fpai_collect::sqlitepoll::Spec {
+                format,
+                db_path: db_path.clone(),
+                spool_dir: spool_dir.clone(),
+                state_dir: state_dir.clone(),
+                poll_interval: std::time::Duration::from_secs(5),
+                params: fpai_collect::sqlitepoll::Params {
+                    agent_id: default_agent_id.to_string(),
+                    environment: environment.clone(),
+                    max_rows_per_poll: 2000,
+                    max_batch_bytes: fpai_collect::spool::DEFAULT_MAX_BATCH_BYTES,
+                    max_drain_passes: 20,
+                },
+            },
+            sd,
+        )
+    }));
+}
+
+/// A stable, filesystem-safe directory name for one Hermes profile database.
+///
+/// Derived from the profile directory rather than an index alone, so adding a
+/// profile cannot renumber an existing one and silently reset its watermark.
+fn profile_dir_name(db: &std::path::Path, index: usize) -> String {
+    let name = db
+        .parent()
+        .and_then(|p| p.file_name())
+        .and_then(|n| n.to_str())
+        .unwrap_or("default");
+    let safe: String = name
+        .chars()
+        .map(|c| if c.is_ascii_alphanumeric() { c } else { '-' })
+        .take(32)
+        .collect();
+    if safe.trim_matches('-').is_empty() {
+        format!("profile-{index}")
+    } else {
+        safe
+    }
+}
+
+/// Where OpenAI Codex keeps its rollout logs.
+fn codex_sessions_root() -> std::path::PathBuf {
+    if let Some(p) = std::env::var_os("CODEX_HOME") {
+        return std::path::PathBuf::from(p).join("sessions");
+    }
+    let home = std::env::var_os("HOME")
+        .map(std::path::PathBuf::from)
+        .unwrap_or_default();
+    home.join(".codex").join("sessions")
+}
+
+fn cloud_policy_reconcile_interval() -> Duration {
+    const DEFAULT_MS: u64 = 30_000;
+    const MINIMUM_MS: u64 = 100;
+    let configured = std::env::var("FAILPROOFAI_CLOUD_POLICY_RECONCILE_MS")
+        .ok()
+        .and_then(|value| value.parse::<u64>().ok())
+        .unwrap_or(DEFAULT_MS);
+    Duration::from_millis(configured.max(MINIMUM_MS))
+}
+
 /// Requests a clean shutdown (socket file removal, lock release via Drop)
 /// on SIGTERM/SIGINT instead of dying mid-accept-loop — this is how a
 /// systemd `stop`/launchd unload is expected to end the process.

package/.next/standalone/crates/failproofaid/src/paths.rs

@@ -49,6 +49,35 @@ pub fn worker_socket_path() -> io::Result<PathBuf> {
     Ok(run_dir()?.join("worker.sock"))
 }
 
+/// Local cache for cloud-managed policy generations. The override keeps
+/// tests and development runs away from a user's real policy directory.
+pub fn cloud_managed_policy_dir() -> io::Result<PathBuf> {
+    if let Some(path) = std::env::var_os("FAILPROOFAI_CLOUD_POLICY_DIR") {
+        return Ok(PathBuf::from(path));
+    }
+    let home = std::env::var_os("HOME").ok_or_else(|| {
+        io::Error::other("HOME is not set; cannot resolve cloud policy directory")
+    })?;
+    Ok(PathBuf::from(home)
+        .join(".failproofai")
+        .join("policies")
+        .join("cloud-managed"))
+}
+
+/// `~/.failproofai` — the root the collector reads its configuration from.
+///
+/// `FAILPROOFAI_HOME` overrides it so tests and development runs never touch a
+/// real user's config, and so a containerised daemon can be pointed at a
+/// mounted volume.
+pub fn failproofai_home() -> io::Result<PathBuf> {
+    if let Some(path) = std::env::var_os("FAILPROOFAI_HOME") {
+        return Ok(PathBuf::from(path));
+    }
+    let home = std::env::var_os("HOME")
+        .ok_or_else(|| io::Error::other("HOME is not set; cannot resolve the failproofai home"))?;
+    Ok(PathBuf::from(home).join(".failproofai"))
+}
+
 /// Creates the run directory (`0700`) if it doesn't exist yet. This
 /// directory holds a socket that evaluates security-relevant decisions, so
 /// a freshly created one is always locked to owner-only.

package/.next/standalone/crates/fpai-collect/Cargo.toml

@@ -0,0 +1,43 @@
+[package]
+name = "fpai-collect"
+version.workspace = true
+edition.workspace = true
+license-file.workspace = true
+repository.workspace = true
+description = "Session-log and hook-activity collection for failproofaid: fault-isolated supervisor, spool, and uploader. Carries no policy logic and never shares fate with the enforcement socket."
+publish = false
+
+[dependencies]
+serde = { version = "1", features = ["derive"] }
+serde_json = "1"
+# The collector is the only async part of the daemon. `failproofaid`'s socket
+# server stays blocking std threads on purpose (see supervisor.rs), so this
+# runtime is created here and never leaks into the enforcement path.
+tokio = { version = "1", features = ["rt-multi-thread", "macros", "time", "fs", "sync", "io-util"] }
+tracing = "0.1"
+# `rustls-tls` rather than the default native-tls: the daemon cross-compiles to
+# four targets and pulling OpenSSL onto each of them is the kind of build
+# dependency that only breaks on the leg nobody tests. `failproofaid` already
+# depends on reqwest with the same TLS backend, so this adds no new C code.
+reqwest = { version = "0.12", default-features = false, features = ["rustls-tls", "json"] }
+# inotify on Linux, FSEvents on macOS. Only ever a latency optimisation — the
+# sweeper is what guarantees delivery, so a watcher that fails to register
+# degrades to interval-only rather than losing anything.
+notify = "8"
+# Epoch-millis to RFC3339 with microseconds. Hand-rolling civil date maths is
+# a silent-failure risk — a leap-year slip would put every event on the wrong
+# day with nothing to notice it — and this pulls no C code onto any of the
+# four cross-compiled targets.
+# `bundled` vendors libsqlite3 rather than linking the host's, so the four
+# cross-compiled release binaries do not depend on whatever sqlite the target
+# happens to ship. It compiles C, which `ring` already does on every leg.
+#
+# A real libsqlite3 is not optional here: goose's sessions.db main file contains
+# ZERO tables, with all 1.25 MB of schema and rows in an uncheckpointed -wal.
+# Anything that parses the main file directly reads an empty database.
+rusqlite = { version = "0.40", features = ["bundled"] }
+time = { version = "0.3", default-features = false, features = ["std", "formatting", "macros"] }
+
+[dev-dependencies]
+tokio = { version = "1", features = ["rt-multi-thread", "macros", "time", "sync", "test-util"] }
+wiremock = "0.6"