Replace the footer's placeholder usage with a real observation ledger

The footer showed invented percentages. It now shows what two harnesses
actually report, or says it does not know.

lumbridge-core gains an append-only per-profile UsageLedger and a projection
that labels every derived value estimated, withholds a burn rate from a single
sample, withholds a window fraction with no reported ceiling, withholds an
exhaustion estimate that lands after the reset, and reports an expired window
as rolled over rather than freezing its last percentage. A missing fact renders
as missing, never as zero. (0012)

lumbridge-harness is the impure side: processes, clocks, and untrusted wire
text in, observations out. Three adapters:

- Codex's account/rateLimits/read over the app-server's JSON-RPC stdio. The
  client cannot express a request outside a two-variant enum and answers every
  server-to-client request with -32601, so a harness asking Lumbridge for a
  credential is refused by construction. (0013)
- Claude Code's session transcripts, as a byte-offset tail follower that
  reports nothing until the backlog is read to EOF — a partially-read backlog
  is indistinguishable from a burst of spend, and the first run against 20 MB
  reported forty-six billion tokens an hour. The parser models four counters,
  so the conversations in those files are not representable. (0014)
- Claude Code's five-hour and seven-day subscription windows, via a bridge
  installed as its statusLine command. 0014 had claimed no such surface
  existed; it does, and the record is corrected in place rather than quietly
  edited. Lumbridge does not read the OAuth credential to call the account
  usage endpoint, which is what comparable tools do — AGENTS.md forbids it,
  and 0015 says so rather than leaving the gap unexplained.

Also in here: a capability-check ordering fix in the workspace reducer, where
the applied-request replay table was consulted before the capability check and
so answered questions the caller had no right to ask; the GPUI spike wired to
the live probes with per-harness gauges and provenance chips; and a launcher
that matches its own window by PID, because GPUI sets WM_NAME but not
_NET_WM_NAME and a title match never succeeded.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Cargo.lock

@@ -867,6 +867,17 @@ dependencies = [
  "thiserror 2.0.20",
 ]
 
+[[package]]
+name = "lumbridge-harness"
+version = "0.0.1"
+dependencies = [
+ "lumbridge-core",
+ "nix",
+ "serde",
+ "serde_json",
+ "thiserror 2.0.20",
+]
+
 [[package]]
 name = "lumbridge-pty"
 version = "0.0.1"

Cargo.toml

@@ -3,6 +3,7 @@ members = [
     "apps/lumbridge",
     "crates/lumbridge-buzz",
     "crates/lumbridge-core",
+    "crates/lumbridge-harness",
     "crates/lumbridge-pty",
     "crates/lumbridge-runtime",
     "crates/lumbridge-storage",

README.md

@@ -25,9 +25,12 @@ responsive one-, three-, or five-panel workspace—five on a 3440 px ultrawide
 where every panel owns its own context, work surface, and decision shelf. One
 panel contains a styled actor-owned VT session and five comparison surfaces
 remain deterministic.
-Retained-history navigation is wired. We are still validating terminal text
+Retained-history navigation is wired. The usage footer is a live strip over an
-selection and mouse input, standalone runtime IPC/durability, ACP integration,
+append-only ledger with provenance, fed by two real adapters: Codex's
-packaging, and usage-data contracts before a large implementation.
+documented quota surface and Claude Code's session transcripts. A harness with
+no adapter renders an explicit gap rather than a zero. We are still validating
+terminal text selection and mouse input, standalone runtime IPC/durability, ACP
+integration, and packaging before a large implementation.
 
 ## Product shape
 

crates/lumbridge-core/src/lib.rs

@@ -2,8 +2,13 @@
 
 use std::fmt;
 
+mod usage;
 mod workspace;
 
+pub use usage::{
+    AccountProfile, AccountProfileId, FooterUsage, UsageConfidence, UsageError, UsageLedger,
+    UsageObservation, UsageProjection, UsageProvenance, UsageUnit, UsageWindow, format_duration_ms,
+};
 pub use workspace::{
     CommandOrigin, CommandRequestId, LayoutNode, PaneCloseDisposition, PaneDefinition, PaneId,
     PaneLaunchIntent, PanePlacement, PaneSurface, SplitAxis, SplitRatio, WorkspaceApplyOutcome,
@@ -96,15 +101,6 @@ impl fmt::Display for Platform {
     }
 }
 
-#[derive(Clone, Copy, Debug, Eq, PartialEq)]
-pub enum UsageProvenance {
-    ProviderReported,
-    HarnessReported,
-    LocallyMeasured,
-    Estimated,
-    Unavailable,
-}
-
 #[derive(Clone, Debug, Eq, PartialEq)]
 pub struct ProductStatus {
     pub platform: Platform,

crates/lumbridge-core/src/workspace.rs

@@ -464,6 +464,11 @@ impl WorkspaceState {
 
     /// Applies one idempotent command after checking its required capability.
     ///
+    /// Authority is checked before the replay table is consulted. A caller that
+    /// cannot perform a command also cannot learn whether its request ID was
+    /// already applied, so the retry path is not an oracle over the workspace's
+    /// history.
+    ///
     /// # Errors
     ///
     /// Rejects insufficient authority and invalid state transitions without
@@ -473,6 +478,11 @@ impl WorkspaceState {
         request: WorkspaceRequest,
         granted: WorkspaceCapability,
     ) -> Result<WorkspaceApplyOutcome, WorkspaceError> {
+        request.origin.validate()?;
+        let required = request.command.required_capability();
+        if granted < required {
+            return Err(WorkspaceError::CapabilityDenied { required, granted });
+        }
         if let Some(applied) = self.applied_requests.get(&request.request_id) {
             if applied == &request {
                 return Ok(WorkspaceApplyOutcome {
@@ -483,11 +493,6 @@ impl WorkspaceState {
             }
             return Err(WorkspaceError::RequestConflict(request.request_id));
         }
-        request.origin.validate()?;
-        let required = request.command.required_capability();
-        if granted < required {
-            return Err(WorkspaceError::CapabilityDenied { required, granted });
-        }
 
         let event = self.apply_command(request.command.clone())?;
         self.applied_requests
@@ -749,6 +754,33 @@ mod tests {
         assert_eq!(state.revision(), 1);
     }
 
+    #[test]
+    fn replaying_an_applied_request_still_requires_its_capability() {
+        let mut state = workspace();
+        let command = request(
+            "launch-agent",
+            WorkspaceCommand::SplitPane {
+                target: id(PaneId::new, "root"),
+                pane: pane("agent", true),
+                axis: SplitAxis::Horizontal,
+                ratio: SplitRatio::default(),
+                placement: PanePlacement::After,
+            },
+        );
+        state
+            .apply(command.clone(), WorkspaceCapability::Execute)
+            .expect("execute may launch");
+        // The replay table must not become an oracle: a caller holding less
+        // authority learns nothing about what has already been applied.
+        assert!(matches!(
+            state.apply(command, WorkspaceCapability::Observe),
+            Err(WorkspaceError::CapabilityDenied {
+                required: WorkspaceCapability::Execute,
+                granted: WorkspaceCapability::Observe,
+            })
+        ));
+    }
+
     #[test]
     fn reused_request_id_must_match_origin_and_command() {
         let mut state = workspace();

crates/lumbridge-harness/Cargo.toml

@@ -0,0 +1,20 @@
+[package]
+name = "lumbridge-harness"
+description = "Harness launch profiles and documented status/usage probes for Lumbridge"
+version.workspace = true
+edition.workspace = true
+rust-version.workspace = true
+license.workspace = true
+repository.workspace = true
+
+[dependencies]
+lumbridge-core = { path = "../lumbridge-core" }
+serde = { version = "1.0", features = ["derive"] }
+serde_json = "1.0"
+thiserror = "2.0"
+
+[target.'cfg(unix)'.dependencies]
+nix = { version = "0.28", features = ["process", "signal"] }
+
+[lints]
+workspace = true

crates/lumbridge-harness/src/child.rs

@@ -0,0 +1,324 @@
+//! A supervised, byte-capped stdio child.
+//!
+//! The child is launched with a cleared environment and a literal allowlist.
+//! It speaks a machine protocol, so it gets pipes rather than a PTY: a TTY
+//! would invite terminal control sequences in both directions for no benefit.
+//! Its stderr is discarded rather than captured, because a harness's diagnostic
+//! output is exactly the kind of text that ends up in a log carrying a path or
+//! a token fragment.
+//!
+//! Reading a pipe blocks. The child therefore splits into three owners: a
+//! reader that may block on its own thread, a writer, and a killer the
+//! supervising thread keeps. Shutdown kills first and joins second, so a child
+//! that has gone quiet — or hostile — can never hold the caller hostage.
+//!
+//! The child is put in its own process group and the group is what gets
+//! killed. A pipe reports end of file only when *every* write end closes, so
+//! killing the direct child alone is not enough: a launcher that execs the
+//! real program as a grandchild with inherited stdio — which is how the npm
+//! distribution of Codex works — leaves that grandchild holding our stdout,
+//! and the reader would block forever on a pipe nothing will ever write to.
+
+use std::io::{BufRead, BufReader, Read, Write};
+use std::process::{Child, ChildStdin, ChildStdout, Command, Stdio};
+use std::sync::{Arc, Mutex};
+
+use crate::HarnessError;
+
+/// Environment a probe child may inherit.
+///
+/// Literal names only: never a pattern and never a user-supplied name.
+///
+/// This deliberately mirrors but does not reuse the terminal crate's private
+/// allowlist. AGENTS.md keeps the harness and terminal boundaries separate,
+/// and the two lists genuinely differ: a probe needs `CODEX_HOME` so it reads
+/// the same account the user's panes use, and has no use for `SHELL`.
+/// `CODEX_HOME` is a filesystem path, never a secret.
+pub const PROBE_ENV_ALLOWLIST: &[&str] = &[
+    "CODEX_HOME",
+    "HOME",
+    "LANG",
+    "LC_ALL",
+    "LC_CTYPE",
+    "LOGNAME",
+    "PATH",
+    "TMPDIR",
+    "USER",
+    "XDG_CONFIG_HOME",
+    "XDG_DATA_HOME",
+    "XDG_RUNTIME_DIR",
+];
+
+/// A single frame longer than this is treated as a fault rather than buffered.
+/// A hostile or broken relay must not be able to grow this process's memory.
+pub(crate) const MAX_LINE_BYTES: usize = 256 * 1024;
+
+/// A freshly launched child, before its pipes are handed to their owners.
+pub(crate) struct ProbeChild {
+    child: Arc<Mutex<Child>>,
+    group: u32,
+    stdin: ChildStdin,
+    stdout: ChildStdout,
+}
+
+impl ProbeChild {
+    /// Launches a probe program with a minimal environment.
+    ///
+    /// This runs on the caller's thread, following the same pattern as the
+    /// runtime actor: start synchronously so a missing program is an immediate
+    /// error, then transfer the pipes to their threads.
+    ///
+    /// # Errors
+    ///
+    /// Returns [`HarnessError::EmptyProgram`] for a blank program name and
+    /// [`HarnessError::SpawnFailed`] if the process cannot start. The
+    /// underlying [`std::io::Error`] is deliberately reduced to its
+    /// [`std::io::ErrorKind`] so a filesystem path cannot travel in an error.
+    pub(crate) fn spawn(program: &str, arguments: &[&str]) -> Result<Self, HarnessError> {
+        if program.trim().is_empty() {
+            return Err(HarnessError::EmptyProgram);
+        }
+        let mut command = Command::new(program);
+        command
+            .args(arguments)
+            .env_clear()
+            .stdin(Stdio::piped())
+            .stdout(Stdio::piped())
+            .stderr(Stdio::null());
+        #[cfg(unix)]
+        {
+            use std::os::unix::process::CommandExt;
+            // The child leads its own group, so its group ID is its process ID
+            // and every descendant it does not deliberately detach joins it.
+            command.process_group(0);
+        }
+        for name in PROBE_ENV_ALLOWLIST {
+            if let Some(value) = std::env::var_os(name) {
+                command.env(name, value);
+            }
+        }
+        let mut child = command
+            .spawn()
+            .map_err(|error| HarnessError::SpawnFailed(error.kind()))?;
+        let broken = HarnessError::SpawnFailed(std::io::ErrorKind::BrokenPipe);
+        let group = child.id();
+        let stdin = child.stdin.take().ok_or(broken)?;
+        let stdout = child.stdout.take().ok_or(broken)?;
+        Ok(Self {
+            child: Arc::new(Mutex::new(child)),
+            group,
+            stdin,
+            stdout,
+        })
+    }
+
+    /// Splits the child into its three independent owners.
+    pub(crate) fn into_parts(self) -> (ProbeWriter, ProbeReader, ProbeKiller) {
+        (
+            ProbeWriter { stdin: self.stdin },
+            ProbeReader {
+                reader: BufReader::new(self.stdout),
+            },
+            ProbeKiller {
+                child: self.child,
+                group: self.group,
+            },
+        )
+    }
+}
+
+/// The writing half.
+pub(crate) struct ProbeWriter {
+    stdin: ChildStdin,
+}
+
+impl ProbeWriter {
+    /// Writes one newline-terminated frame.
+    ///
+    /// # Errors
+    ///
+    /// Returns [`HarnessError::Eof`] once the child has closed its input.
+    pub(crate) fn write_line(&mut self, line: &str) -> Result<(), HarnessError> {
+        self.stdin
+            .write_all(line.as_bytes())
+            .and_then(|()| self.stdin.write_all(b"\n"))
+            .and_then(|()| self.stdin.flush())
+            .map_err(|_| HarnessError::Eof)
+    }
+}
+
+/// The reading half. Its methods block, so it belongs on its own thread.
+pub(crate) struct ProbeReader {
+    reader: BufReader<ChildStdout>,
+}
+
+impl ProbeReader {
+    /// Reads one frame, refusing an oversized line instead of buffering it.
+    ///
+    /// # Errors
+    ///
+    /// Returns [`HarnessError::Eof`] at end of stream and
+    /// [`HarnessError::OversizedLine`] past [`MAX_LINE_BYTES`].
+    pub(crate) fn read_line(&mut self) -> Result<String, HarnessError> {
+        let mut buffer = Vec::new();
+        let read = self
+            .reader
+            .by_ref()
+            .take(MAX_LINE_BYTES as u64 + 1)
+            .read_until(b'\n', &mut buffer)
+            .map_err(|_| HarnessError::Eof)?;
+        if read == 0 {
+            return Err(HarnessError::Eof);
+        }
+        if read > MAX_LINE_BYTES {
+            return Err(HarnessError::OversizedLine(read));
+        }
+        String::from_utf8(buffer).map_err(|_| HarnessError::Malformed)
+    }
+}
+
+/// The half a supervisor keeps so it can always stop the child.
+#[derive(Clone)]
+pub(crate) struct ProbeKiller {
+    child: Arc<Mutex<Child>>,
+    group: u32,
+}
+
+impl ProbeKiller {
+    /// Ends the child and everything in its process group, then reaps it.
+    ///
+    /// Killing the whole group is what closes every write end of our stdout
+    /// pipe, which is what unblocks the reader. Idempotent: killing an
+    /// already-dead process is not an error worth reporting, and a poisoned
+    /// lock means another thread died mid-kill, which the operating system has
+    /// already resolved.
+    pub(crate) fn kill(&self) {
+        self.kill_group();
+        if let Ok(mut child) = self.child.lock() {
+            let _ = child.kill();
+            let _ = child.wait();
+        }
+    }
+
+    #[cfg(unix)]
+    fn kill_group(&self) {
+        use nix::sys::signal::{Signal, killpg};
+        use nix::unistd::Pid;
+
+        if let Ok(group) = i32::try_from(self.group) {
+            let _ = killpg(Pid::from_raw(group), Signal::SIGKILL);
+        }
+    }
+
+    #[cfg(not(unix))]
+    fn kill_group(&self) {}
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{MAX_LINE_BYTES, PROBE_ENV_ALLOWLIST, ProbeChild};
+    use crate::HarnessError;
+
+    #[test]
+    fn an_empty_program_is_refused_before_spawning() {
+        assert!(matches!(
+            ProbeChild::spawn("   ", &[]).err(),
+            Some(HarnessError::EmptyProgram)
+        ));
+    }
+
+    #[test]
+    fn a_missing_program_reports_a_kind_and_never_a_path() {
+        let Err(error) = ProbeChild::spawn("lumbridge-no-such-probe-binary", &[]) else {
+            panic!("a missing program cannot spawn");
+        };
+        assert!(matches!(error, HarnessError::SpawnFailed(_)));
+        assert!(
+            !error.to_string().contains("lumbridge-no-such-probe-binary"),
+            "an error must not carry a filesystem path"
+        );
+    }
+
+    #[test]
+    fn the_allowlist_is_literal_sorted_and_free_of_secret_bearing_names() {
+        let mut sorted = PROBE_ENV_ALLOWLIST.to_vec();
+        sorted.sort_unstable();
+        assert_eq!(sorted, PROBE_ENV_ALLOWLIST);
+        for name in PROBE_ENV_ALLOWLIST {
+            let upper = name.to_uppercase();
+            assert!(
+                !upper.contains("TOKEN")
+                    && !upper.contains("KEY")
+                    && !upper.contains("SECRET")
+                    && !upper.contains("PASSWORD"),
+                "{name} could carry a credential into a child"
+            );
+        }
+    }
+
+    #[test]
+    fn a_child_round_trips_one_frame() {
+        let Ok(child) = ProbeChild::spawn("cat", &[]) else {
+            return; // cat is not guaranteed on every supported platform.
+        };
+        let (mut writer, mut reader, killer) = child.into_parts();
+        writer
+            .write_line(r#"{"ok":true}"#)
+            .expect("cat accepts input");
+        assert_eq!(
+            reader.read_line().expect("cat echoes the line"),
+            "{\"ok\":true}\n"
+        );
+        killer.kill();
+    }
+
+    #[test]
+    fn an_oversized_frame_is_refused_rather_than_buffered() {
+        // Generated by the child so the test never deadlocks writing more than
+        // a pipe buffer into a process that is not reading.
+        let script = format!("head -c {} /dev/zero | tr '\\0' 'x'", MAX_LINE_BYTES + 16);
+        let Ok(child) = ProbeChild::spawn("sh", &["-c", &script]) else {
+            return;
+        };
+        let (_writer, mut reader, killer) = child.into_parts();
+        assert!(matches!(
+            reader.read_line(),
+            Err(HarnessError::OversizedLine(_))
+        ));
+        killer.kill();
+    }
+
+    #[test]
+    fn killing_a_child_unblocks_a_reader_blocked_on_a_grandchild() {
+        // A launcher that backgrounds a helper and waits is the shape of the
+        // npm Codex distribution: the helper inherits our stdout, so killing
+        // only the direct child would leave the reader blocked forever.
+        let Ok(child) = ProbeChild::spawn("sh", &["-c", "sleep 30 & wait"]) else {
+            return;
+        };
+        let (_writer, mut reader, killer) = child.into_parts();
+        let reading = std::thread::spawn(move || reader.read_line());
+        killer.kill();
+        let outcome = reading.join().expect("the reader thread must not panic");
+        assert!(
+            matches!(outcome, Err(HarnessError::Eof)),
+            "killing the group must end the read rather than hang it"
+        );
+    }
+
+    #[test]
+    fn killing_a_child_unblocks_its_reader() {
+        // `sleep` never writes, so the reader is blocked until the kill lands.
+        let Ok(child) = ProbeChild::spawn("sleep", &["30"]) else {
+            return;
+        };
+        let (_writer, mut reader, killer) = child.into_parts();
+        let reading = std::thread::spawn(move || reader.read_line());
+        killer.kill();
+        let outcome = reading.join().expect("the reader thread must not panic");
+        assert!(
+            matches!(outcome, Err(HarnessError::Eof)),
+            "a killed child must end the read rather than hang it"
+        );
+    }
+}

crates/lumbridge-harness/src/claude/mod.rs

@@ -0,0 +1,836 @@
+//! The Claude Code usage adapter, over two different surfaces.
+//!
+//! **Subscription windows** ([`statusline`]) are the real quota. Claude Code
+//! 2.1.80 and later pipe a `rate_limits` object carrying `five_hour` and
+//! `seven_day` — each with `used_percentage` and `resets_at` — to the
+//! configured `statusLine` command on every turn. Those are rate-limit headers
+//! the CLI already received on its own API responses, so reading them costs
+//! nothing and they are `ProviderReported`. Lumbridge does not call the
+//! account's usage endpoint for them: that would mean reading Claude Code's
+//! OAuth credential, which AGENTS.md forbids. The status line is pushed to a
+//! command the user installs, so no credential is ever touched.
+//!
+//! **Token consumption** ([`transcript`]) comes from the session transcripts
+//! Claude Code writes under `~/.claude/projects`. Those record the API's
+//! `usage` object per assistant turn. A running total is the harness's own
+//! record rather than a provider's statement of account, so it is
+//! `HarnessReported`, and it carries no ceiling — spend is not a quota.
+//!
+//! Both are tail followers. Each remembers a byte offset and reads only what
+//! was appended since, which is what makes the token total monotonic — the
+//! ledger rejects an observation that moves a profile backwards.
+
+mod statusline;
+mod transcript;
+
+use std::collections::BTreeMap;
+use std::fs::File;
+use std::io::{Read, Seek, SeekFrom};
+use std::path::{Path, PathBuf};
+use std::sync::Arc;
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::mpsc::{self, Receiver, SyncSender, TryRecvError};
+use std::thread::{self, JoinHandle};
+use std::time::Duration;
+
+use lumbridge_core::{AccountProfile, UsageObservation, UsageProvenance, UsageUnit};
+
+use crate::HarnessError;
+use crate::claude::statusline::parse_feed_line;
+pub use crate::claude::statusline::{ClaudeWindowKind, WindowReading};
+use crate::claude::transcript::MAX_RECORD_BYTES;
+use crate::clock::MonotonicWallClock;
+use crate::probe::{ProbeHealth, ProbeOutcome, UsageProbe};
+
+pub use transcript::TokenTally;
+
+/// Transcripts are appended to constantly; polling faster than this buys
+/// nothing but disk churn.
+const MIN_POLL_INTERVAL: Duration = Duration::from_secs(1);
+/// How long the worker sleeps between filesystem scans.
+const WORKER_TICK: Duration = Duration::from_millis(250);
+/// A bound on how much of the tree one scan will walk.
+const MAX_DEPTH: usize = 6;
+const MAX_FILES_PER_SCAN: usize = 4_096;
+/// A bound on how much any one poll will read, so a large backlog is absorbed
+/// across several polls instead of stalling one.
+const MAX_BYTES_PER_POLL: usize = 8 * 1024 * 1024;
+const EVENT_QUEUE: usize = 64;
+
+/// Where and how often to read transcripts.
+#[derive(Clone, Debug)]
+pub struct ClaudeCodeProbeOptions {
+    /// The Claude Code projects root. Defaults to `$CLAUDE_CONFIG_DIR/projects`
+    /// or `~/.claude/projects`.
+    pub projects_root: PathBuf,
+    /// Where the Lumbridge status-line bridge appends rate-limit readings.
+    /// Absent until the user installs the bridge, which is why the windows read
+    /// as unavailable rather than as zero before then.
+    pub rate_limit_feed: PathBuf,
+    pub poll_interval: Duration,
+}
+
+impl Default for ClaudeCodeProbeOptions {
+    fn default() -> Self {
+        Self {
+            projects_root: default_projects_root(),
+            rate_limit_feed: default_rate_limit_feed(),
+            poll_interval: Duration::from_secs(10),
+        }
+    }
+}
+
+/// Where the status-line bridge writes, under Lumbridge's own data root rather
+/// than inside the harness's configuration.
+///
+/// `LUMBRIDGE_CLAUDE_FEED` overrides it, and must be read here as well as in the
+/// bridge: the two halves have to resolve the same path or the bridge writes
+/// somewhere the probe never looks and the windows read as unavailable forever.
+#[must_use]
+pub fn default_rate_limit_feed() -> PathBuf {
+    if let Some(configured) = std::env::var_os("LUMBRIDGE_CLAUDE_FEED") {
+        return PathBuf::from(configured);
+    }
+    let root = std::env::var_os("XDG_DATA_HOME").map_or_else(
+        || {
+            std::env::var_os("HOME")
+                .map(PathBuf::from)
+                .unwrap_or_default()
+                .join(".local")
+                .join("share")
+        },
+        PathBuf::from,
+    );
+    root.join("lumbridge").join("claude-rate-limits.jsonl")
+}
+
+/// The documented default location of Claude Code's session transcripts.
+#[must_use]
+pub fn default_projects_root() -> PathBuf {
+    if let Some(configured) = std::env::var_os("CLAUDE_CONFIG_DIR") {
+        return PathBuf::from(configured).join("projects");
+    }
+    std::env::var_os("HOME")
+        .map(PathBuf::from)
+        .unwrap_or_default()
+        .join(".claude")
+        .join("projects")
+}
+
+#[derive(Debug)]
+enum WorkerEvent {
+    Observation(UsageObservation),
+    Health(ProbeHealth),
+}
+
+/// A running Claude Code usage probe.
+pub struct ClaudeCodeProbe {
+    profiles: Vec<AccountProfile>,
+    events: Option<Receiver<WorkerEvent>>,
+    stop: Arc<AtomicBool>,
+    worker: Option<JoinHandle<()>>,
+    health: ProbeHealth,
+    tokens: TokenTally,
+}
+
+impl ClaudeCodeProbe {
+    /// The static, account-free profiles this probe reports under: one per
+    /// documented subscription window, plus the transcript token total.
+    ///
+    /// The windows come first because they are the quota — what a user means
+    /// by "how much do I have left" — and the token total is the supporting
+    /// fact about spend.
+    fn profiles_for() -> Result<Vec<AccountProfile>, HarnessError> {
+        let mut profiles = Vec::new();
+        for kind in ClaudeWindowKind::ALL {
+            profiles.push(
+                AccountProfile::new(
+                    kind.profile_id(),
+                    "Claude Code",
+                    "Anthropic",
+                    kind.scope(),
+                    "subscription",
+                )
+                .map_err(|_| HarnessError::EmptyProgram)?,
+            );
+        }
+        profiles.push(
+            AccountProfile::new(
+                "claude-code-transcripts",
+                "Claude Code",
+                "Anthropic",
+                "session transcripts",
+                "subscription",
+            )
+            .map_err(|_| HarnessError::EmptyProgram)?,
+        );
+        Ok(profiles)
+    }
+
+    /// Starts the probe and its worker.
+    ///
+    /// A missing projects directory is not an error: the probe reports zero
+    /// and keeps looking, because Claude Code creates it on first use.
+    ///
+    /// # Errors
+    ///
+    /// Returns [`HarnessError::ProbeIntervalTooShort`] below one second.
+    pub fn start(options: ClaudeCodeProbeOptions) -> Result<Self, HarnessError> {
+        if options.poll_interval < MIN_POLL_INTERVAL {
+            return Err(HarnessError::ProbeIntervalTooShort);
+        }
+        let profiles = Self::profiles_for()?;
+        let worker_profiles = profiles.clone();
+        let (event_sender, events) = mpsc::sync_channel(EVENT_QUEUE);
+        let stop = Arc::new(AtomicBool::new(false));
+        let worker_stop = Arc::clone(&stop);
+        let worker = thread::Builder::new()
+            .name("lumbridge-claude-probe".to_owned())
+            .spawn(move || {
+                run_worker(&options, &worker_profiles, &event_sender, &worker_stop);
+            })
+            .map_err(|error| HarnessError::SpawnFailed(error.kind()))?;
+
+        Ok(Self {
+            profiles,
+            events: Some(events),
+            stop,
+            worker: Some(worker),
+            health: ProbeHealth::Starting,
+            tokens: TokenTally::default(),
+        })
+    }
+
+    /// The token split behind the latest observation.
+    ///
+    /// The ledger holds only the total; this keeps the breakdown available for
+    /// a detail view without a second pass over the transcripts.
+    #[must_use]
+    pub const fn tokens(&self) -> TokenTally {
+        self.tokens
+    }
+}
+
+impl UsageProbe for ClaudeCodeProbe {
+    fn poll(&mut self) -> ProbeOutcome {
+        let Some(events) = self.events.as_ref() else {
+            return ProbeOutcome::idle(self.health);
+        };
+        let mut observations = Vec::new();
+        loop {
+            match events.try_recv() {
+                Ok(WorkerEvent::Observation(observation)) => observations.push(observation),
+                Ok(WorkerEvent::Health(health)) => self.health = health,
+                Err(TryRecvError::Empty) => break,
+                Err(TryRecvError::Disconnected) => {
+                    if !self.health.is_faulted() {
+                        self.health = ProbeHealth::Stopped;
+                    }
+                    self.events = None;
+                    break;
+                }
+            }
+        }
+        ProbeOutcome::new(observations, self.health)
+    }
+
+    fn profiles(&self) -> &[AccountProfile] {
+        &self.profiles
+    }
+
+    fn shutdown(&mut self) {
+        self.stop.store(true, Ordering::Relaxed);
+        self.events = None;
+        if let Some(worker) = self.worker.take() {
+            let _ = worker.join();
+        }
+        if !self.health.is_faulted() {
+            self.health = ProbeHealth::Stopped;
+        }
+    }
+}
+
+impl Drop for ClaudeCodeProbe {
+    fn drop(&mut self) {
+        self.shutdown();
+    }
+}
+
+/// Per-transcript read position. The offset is what makes the total monotonic.
+#[derive(Default)]
+struct FollowState {
+    offsets: BTreeMap<PathBuf, u64>,
+    tokens: TokenTally,
+    /// Read position in the status-line feed, and the newest reading seen.
+    feed_offset: u64,
+    windows: BTreeMap<&'static str, WindowReading>,
+}
+
+/// Reads whatever the status-line bridge appended and keeps the newest line.
+///
+/// The feed is a state snapshot per line, not a stream of deltas, so only the
+/// last complete line matters — an older one is superseded, never summed.
+fn follow_feed(path: &Path, state: &mut FollowState, observed_at_ms: u64) {
+    let Ok(metadata) = std::fs::metadata(path) else {
+        return;
+    };
+    let length = metadata.len();
+    if length < state.feed_offset {
+        state.feed_offset = 0;
+    }
+    if length == state.feed_offset {
+        return;
+    }
+    let take = usize::try_from(length - state.feed_offset)
+        .unwrap_or(usize::MAX)
+        .min(MAX_BYTES_PER_POLL);
+    let Ok(mut file) = File::open(path) else {
+        return;
+    };
+    if file.seek(SeekFrom::Start(state.feed_offset)).is_err() {
+        return;
+    }
+    let mut buffer = vec![0_u8; take];
+    let Ok(read) = file.read(&mut buffer) else {
+        return;
+    };
+    buffer.truncate(read);
+    let chunk = String::from_utf8_lossy(&buffer);
+
+    let mut consumed = 0;
+    let mut newest = None;
+    for line in chunk.split_inclusive('\n') {
+        if !line.ends_with('\n') {
+            break;
+        }
+        consumed += line.len();
+        if line.len() <= MAX_RECORD_BYTES
+            && let Some(readings) = parse_feed_line(line.trim_end(), observed_at_ms)
+        {
+            newest = Some(readings);
+        }
+    }
+    state.feed_offset += consumed as u64;
+    if let Some(readings) = newest {
+        for (kind, reading) in readings {
+            state.windows.insert(kind.profile_id(), reading);
+        }
+    }
+}
+
+fn run_worker(
+    options: &ClaudeCodeProbeOptions,
+    profiles: &[AccountProfile],
+    events: &SyncSender<WorkerEvent>,
+    stop: &AtomicBool,
+) {
+    let token_profile = profiles
+        .iter()
+        .find(|profile| profile.id().as_str() == "claude-code-transcripts")
+        .cloned();
+    let mut clock = MonotonicWallClock::start();
+    let mut state = FollowState::default();
+    let poll_interval_ms = u64::try_from(options.poll_interval.as_millis()).unwrap_or(u64::MAX);
+    let mut last_poll_ms = 0;
+    let mut last_emitted = TokenTally::default();
+    let mut last_windows: BTreeMap<&'static str, WindowReading> = BTreeMap::new();
+    let mut primed = false;
+
+    loop {
+        if stop.load(Ordering::Relaxed) {
+            break;
+        }
+        let now_ms = clock.now_ms();
+        // While priming, scan every tick to work through the backlog quickly.
+        // After that, respect the poll interval.
+        if primed && now_ms.saturating_sub(last_poll_ms) < poll_interval_ms {
+            thread::sleep(WORKER_TICK);
+            continue;
+        }
+        last_poll_ms = now_ms;
+
+        let caught_up = scan(&options.projects_root, &mut state);
+        // The quota windows are independent of the transcript backlog: a
+        // status-line reading is a complete snapshot, so it can be reported
+        // immediately rather than waiting for priming.
+        follow_feed(&options.rate_limit_feed, &mut state, now_ms);
+        for kind in ClaudeWindowKind::ALL {
+            let Some(reading) = state.windows.get(kind.profile_id()).copied() else {
+                continue;
+            };
+            if last_windows.get(kind.profile_id()) == Some(&reading) {
+                continue;
+            }
+            last_windows.insert(kind.profile_id(), reading);
+            let Some(profile) = profiles
+                .iter()
+                .find(|profile| profile.id().as_str() == kind.profile_id())
+            else {
+                continue;
+            };
+            let observation = window_observation(profile, reading, now_ms);
+            if matches!(
+                events.try_send(WorkerEvent::Observation(observation)),
+                Err(mpsc::TrySendError::Disconnected(_))
+            ) {
+                return;
+            }
+        }
+
+        // Nothing is reported until the follower has read every existing
+        // transcript to its end. A partially-read backlog looks exactly like
+        // an enormous burst of spend, and the ledger would derive a burn rate
+        // from it — one early run reported forty-six billion tokens an hour,
+        // which was catch-up, not usage. The first reading must be a true
+        // baseline before any rate can mean anything.
+        if !caught_up {
+            thread::sleep(WORKER_TICK);
+            continue;
+        }
+        if !primed {
+            primed = true;
+            let _ = events.try_send(WorkerEvent::Health(ProbeHealth::Ready));
+        }
+
+        // Only speak when the number changed. An unchanged total re-recorded
+        // every poll would flush the ledger's bounded history of real readings.
+        if state.tokens == last_emitted {
+            continue;
+        }
+        last_emitted = state.tokens;
+        let Some(profile) = token_profile.as_ref() else {
+            continue;
+        };
+        let Ok(observation) = UsageObservation::counted(
+            profile.id().clone(),
+            UsageUnit::Tokens,
+            state.tokens.total(),
+            UsageProvenance::HarnessReported,
+            now_ms,
+        ) else {
+            continue;
+        };
+        // No `with_limit` and no `with_window`: the transcript states what was
+        // spent and says nothing about a ceiling or a reset. Attaching either
+        // would be an invention.
+        if matches!(
+            events.try_send(WorkerEvent::Observation(observation)),
+            Err(mpsc::TrySendError::Disconnected(_))
+        ) {
+            break;
+        }
+    }
+}
+
+/// Turns a window reading into the observation the ledger should hold.
+///
+/// A usable reading is `ProviderReported`: the CLI is relaying rate-limit
+/// headers from its own API responses, not computing a number. An absent
+/// window is an explicit gap — an account with no plan limits, or a window the
+/// API has stopped reporting — never a zero.
+fn window_observation(
+    profile: &AccountProfile,
+    reading: WindowReading,
+    observed_at_ms: u64,
+) -> UsageObservation {
+    let WindowReading::Usable { permille, window } = reading else {
+        return UsageObservation::unavailable(profile.id().clone(), observed_at_ms);
+    };
+    let Ok(observation) = UsageObservation::counted(
+        profile.id().clone(),
+        UsageUnit::WindowPermille,
+        permille,
+        UsageProvenance::ProviderReported,
+        observed_at_ms,
+    ) else {
+        return UsageObservation::unavailable(profile.id().clone(), observed_at_ms);
+    };
+    let observation = observation
+        .with_limit(1_000)
+        .unwrap_or_else(|_| UsageObservation::unavailable(profile.id().clone(), observed_at_ms));
+    let Some(window) = window else {
+        return observation;
+    };
+    observation
+        .clone()
+        .with_window(window)
+        .unwrap_or(observation)
+}
+
+/// Walks the projects tree and reads whatever was appended since last time.
+///
+/// Returns whether every transcript was read to its end. A `false` means the
+/// per-poll byte budget ran out with work remaining, and the running total is
+/// therefore mid-catch-up rather than current.
+fn scan(root: &Path, state: &mut FollowState) -> bool {
+    let mut transcripts = Vec::new();
+    collect_transcripts(root, 0, &mut transcripts);
+    let mut budget = MAX_BYTES_PER_POLL;
+    let mut caught_up = true;
+    for path in transcripts {
+        if budget == 0 {
+            return false;
+        }
+        if !follow(&path, state, &mut budget) {
+            caught_up = false;
+        }
+    }
+    caught_up
+}
+
+fn collect_transcripts(directory: &Path, depth: usize, found: &mut Vec<PathBuf>) {
+    if depth > MAX_DEPTH || found.len() >= MAX_FILES_PER_SCAN {
+        return;
+    }
+    let Ok(entries) = std::fs::read_dir(directory) else {
+        return;
+    };
+    for entry in entries.flatten() {
+        if found.len() >= MAX_FILES_PER_SCAN {
+            return;
+        }
+        let path = entry.path();
+        let Ok(file_type) = entry.file_type() else {
+            continue;
+        };
+        // Deliberately not following symlinks: a link inside the projects tree
+        // could otherwise point this reader at an arbitrary file.
+        if file_type.is_dir() {
+            collect_transcripts(&path, depth + 1, found);
+        } else if file_type.is_file() && path.extension().is_some_and(|ext| ext == "jsonl") {
+            found.push(path);
+        }
+    }
+}
+
+/// Reads one transcript from its remembered offset.
+///
+/// Returns whether this file is now read to its end.
+fn follow(path: &Path, state: &mut FollowState, budget: &mut usize) -> bool {
+    let offset = state.offsets.get(path).copied().unwrap_or(0);
+    let Ok(metadata) = std::fs::metadata(path) else {
+        return true;
+    };
+    let length = metadata.len();
+    if length < offset {
+        // The file shrank, so it is not the file we were reading. Start over
+        // rather than trusting an offset into different content.
+        state.offsets.insert(path.to_path_buf(), 0);
+        return false;
+    }
+    let pending = usize::try_from(length - offset).unwrap_or(usize::MAX);
+    if pending == 0 {
+        return true;
+    }
+    let take = pending.min(*budget);
+
+    let Ok(mut file) = File::open(path) else {
+        return true;
+    };
+    if file.seek(SeekFrom::Start(offset)).is_err() {
+        return true;
+    }
+    let mut buffer = vec![0_u8; take];
+    let Ok(read) = file.read(&mut buffer) else {
+        return true;
+    };
+    buffer.truncate(read);
+    // A transcript is UTF-8 JSON. A chunk boundary can split a multi-byte
+    // character, so decode lossily and stop at the last complete line — the
+    // offset only advances past whole records either way.
+    let chunk = String::from_utf8_lossy(&buffer);
+    let (tally, consumed) = transcript::consume_chunk(&chunk);
+    state.tokens.add(tally.tokens);
+    state
+        .offsets
+        .insert(path.to_path_buf(), offset + consumed as u64);
+    *budget = budget.saturating_sub(consumed);
+    // Caught up only if we took the whole remainder and parsed all of it. A
+    // trailing partial line counts as caught up: the record is still being
+    // written, and waiting for it is correct.
+    take == pending && pending.saturating_sub(consumed) < MAX_RECORD_BYTES
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{
+        ClaudeCodeProbe, ClaudeCodeProbeOptions, FollowState, TokenTally, default_projects_root,
+        scan,
+    };
+    use crate::HarnessError;
+    use crate::probe::{ProbeHealth, UsageProbe};
+    use lumbridge_core::{UsageProvenance, UsageUnit};
+    use std::io::Write;
+    use std::path::PathBuf;
+    use std::time::Duration;
+
+    const RECORD: &str = r#"{"type":"assistant","message":{"model":"claude-opus-5","content":[{"type":"text","text":"private"}],"usage":{"input_tokens":1,"output_tokens":10,"cache_creation_input_tokens":100,"cache_read_input_tokens":1000}}}"#;
+
+    fn scratch(tag: &str) -> PathBuf {
+        let root = std::env::temp_dir().join(format!(
+            "lumbridge-claude-probe-{}-{tag}",
+            std::process::id()
+        ));
+        let _ = std::fs::remove_dir_all(&root);
+        std::fs::create_dir_all(root.join("project-a")).expect("scratch dir");
+        root
+    }
+
+    fn append(path: &PathBuf, times: usize) {
+        let mut file = std::fs::OpenOptions::new()
+            .create(true)
+            .append(true)
+            .open(path)
+            .expect("transcript is writable");
+        for _ in 0..times {
+            writeln!(file, "{RECORD}").expect("append");
+        }
+    }
+
+    #[test]
+    fn a_missing_projects_root_reports_zero_rather_than_failing() {
+        let mut state = FollowState::default();
+        scan(
+            &PathBuf::from("/nonexistent/lumbridge/claude/root"),
+            &mut state,
+        );
+        assert_eq!(state.tokens, TokenTally::default());
+    }
+
+    #[test]
+    fn appended_records_accumulate_and_never_double_count() {
+        let root = scratch("accumulate");
+        let transcript = root.join("project-a").join("session.jsonl");
+        append(&transcript, 2);
+
+        let mut state = FollowState::default();
+        assert!(scan(&root, &mut state));
+        assert_eq!(state.tokens.total(), 2 * 1_111);
+
+        // A second scan with no new bytes must add nothing.
+        assert!(scan(&root, &mut state));
+        assert_eq!(state.tokens.total(), 2 * 1_111);
+
+        // Only the newly appended record is counted.
+        append(&transcript, 1);
+        assert!(scan(&root, &mut state));
+        assert_eq!(state.tokens.total(), 3 * 1_111);
+        let _ = std::fs::remove_dir_all(&root);
+    }
+
+    #[test]
+    fn nested_subagent_transcripts_are_counted_too() {
+        let root = scratch("nested");
+        let nested = root.join("project-a").join("session").join("subagents");
+        std::fs::create_dir_all(&nested).expect("nested dirs");
+        append(&root.join("project-a").join("session.jsonl"), 1);
+        append(&nested.join("agent-1.jsonl"), 1);
+
+        let mut state = FollowState::default();
+        assert!(scan(&root, &mut state));
+        assert_eq!(
+            state.tokens.total(),
+            2 * 1_111,
+            "a subagent's tokens are the user's tokens"
+        );
+        let _ = std::fs::remove_dir_all(&root);
+    }
+
+    #[test]
+    fn a_truncated_transcript_restarts_rather_than_reading_a_stale_offset() {
+        let root = scratch("truncated");
+        let transcript = root.join("project-a").join("session.jsonl");
+        append(&transcript, 3);
+        let mut state = FollowState::default();
+        assert!(scan(&root, &mut state));
+        let before = state.tokens.total();
+        assert_eq!(before, 3 * 1_111);
+
+        std::fs::write(&transcript, "").expect("truncate");
+        // The shrink is reported as not-caught-up on purpose: the offset was
+        // just reset, so the probe must re-scan before it trusts the total
+        // again rather than reporting mid-reset.
+        assert!(
+            !scan(&root, &mut state),
+            "a shrunk file asks to be read again"
+        );
+        assert!(scan(&root, &mut state), "and is caught up on the next pass");
+        assert_eq!(
+            state.tokens.total(),
+            before,
+            "a shrunk file resets its offset without inventing or losing tokens"
+        );
+        let _ = std::fs::remove_dir_all(&root);
+    }
+
+    #[test]
+    fn non_transcript_files_are_ignored() {
+        let root = scratch("ignored");
+        std::fs::write(root.join("project-a").join("notes.md"), RECORD).expect("write");
+        std::fs::write(root.join("project-a").join("history.json"), RECORD).expect("write");
+        let mut state = FollowState::default();
+        assert!(scan(&root, &mut state));
+        assert_eq!(state.tokens, TokenTally::default());
+        let _ = std::fs::remove_dir_all(&root);
+    }
+
+    #[test]
+    fn a_too_frequent_poll_interval_is_refused() {
+        let options = ClaudeCodeProbeOptions {
+            poll_interval: Duration::from_millis(10),
+            ..ClaudeCodeProbeOptions::default()
+        };
+        assert_eq!(
+            ClaudeCodeProbe::start(options).err(),
+            Some(HarnessError::ProbeIntervalTooShort)
+        );
+    }
+
+    #[test]
+    fn the_probe_reports_harness_reported_tokens_with_no_ceiling() {
+        let root = scratch("probe");
+        append(&root.join("project-a").join("session.jsonl"), 4);
+        let mut probe = ClaudeCodeProbe::start(ClaudeCodeProbeOptions {
+            projects_root: root.clone(),
+            rate_limit_feed: root.join("feed.jsonl"),
+            poll_interval: Duration::from_secs(1),
+        })
+        .expect("the probe starts");
+
+        let mut observations = Vec::new();
+        for _ in 0..200 {
+            observations.extend(probe.poll().into_observations());
+            if !observations.is_empty() {
+                break;
+            }
+            std::thread::sleep(Duration::from_millis(20));
+        }
+        probe.shutdown();
+        let _ = std::fs::remove_dir_all(&root);
+
+        let observation = observations.first().expect("a reading arrives");
+        assert_eq!(observation.provenance(), UsageProvenance::HarnessReported);
+        assert_eq!(observation.unit(), Some(UsageUnit::Tokens));
+        assert_eq!(observation.consumed(), 4 * 1_111);
+        assert_eq!(
+            observation.limit(),
+            None,
+            "a transcript states spend, never a ceiling"
+        );
+        assert_eq!(
+            observation.window(),
+            None,
+            "a transcript states spend, never a reset"
+        );
+    }
+
+    #[test]
+    fn a_status_line_feed_becomes_provider_reported_windows() {
+        let root = scratch("windows");
+        let feed = root.join("feed.jsonl");
+        // The shape the bridge writes, with a reset far enough ahead that the
+        // window is still open when the probe reads it.
+        let resets_at = (std::time::SystemTime::now()
+            .duration_since(std::time::UNIX_EPOCH)
+            .expect("a clock after the epoch")
+            .as_secs())
+            + 3_600;
+        std::fs::write(
+            &feed,
+            format!(
+                "{{\"rate_limits_available\":true,\"rate_limits\":{{\"five_hour\":{{\"used_percentage\":42,\"resets_at\":{resets_at}}},\"seven_day\":{{\"used_percentage\":8.5,\"resets_at\":{resets_at}}}}}}}\n"
+            ),
+        )
+        .expect("feed is writable");
+
+        let mut probe = ClaudeCodeProbe::start(ClaudeCodeProbeOptions {
+            projects_root: root.clone(),
+            rate_limit_feed: feed,
+            poll_interval: Duration::from_secs(1),
+        })
+        .expect("the probe starts");
+
+        let mut observations = Vec::new();
+        for _ in 0..200 {
+            observations.extend(probe.poll().into_observations());
+            if observations.len() >= 2 {
+                break;
+            }
+            std::thread::sleep(Duration::from_millis(20));
+        }
+        probe.shutdown();
+        let _ = std::fs::remove_dir_all(&root);
+
+        let five = observations
+            .iter()
+            .find(|observation| observation.profile().as_str() == "claude-code-five-hour")
+            .expect("the five-hour window is reported");
+        assert_eq!(
+            five.provenance(),
+            UsageProvenance::ProviderReported,
+            "the CLI relays the provider's own rate-limit headers"
+        );
+        assert_eq!(five.unit(), Some(UsageUnit::WindowPermille));
+        assert_eq!(five.consumed(), 420);
+        assert_eq!(
+            five.limit(),
+            Some(1_000),
+            "a percentage is a share of one whole"
+        );
+        assert!(five.window().is_some(), "a documented window has a reset");
+    }
+
+    #[test]
+    fn an_account_without_plan_limits_reports_a_gap_not_a_zero() {
+        let root = scratch("nolimits");
+        let feed = root.join("feed.jsonl");
+        std::fs::write(&feed, "{\"rate_limits_available\":false}\n").expect("writable");
+        let mut probe = ClaudeCodeProbe::start(ClaudeCodeProbeOptions {
+            projects_root: root.clone(),
+            rate_limit_feed: feed,
+            poll_interval: Duration::from_secs(1),
+        })
+        .expect("starts");
+        let mut observations = Vec::new();
+        for _ in 0..200 {
+            observations.extend(probe.poll().into_observations());
+            if !observations.is_empty() {
+                break;
+            }
+            std::thread::sleep(Duration::from_millis(20));
+        }
+        probe.shutdown();
+        let _ = std::fs::remove_dir_all(&root);
+        let window = observations
+            .iter()
+            .find(|observation| observation.profile().as_str() == "claude-code-five-hour")
+            .expect("the window is still named");
+        assert_eq!(window.provenance(), UsageProvenance::Unavailable);
+    }
+
+    #[test]
+    fn shutdown_is_prompt_and_idempotent() {
+        let root = scratch("shutdown");
+        let mut probe = ClaudeCodeProbe::start(ClaudeCodeProbeOptions {
+            projects_root: root.clone(),
+            rate_limit_feed: root.join("feed.jsonl"),
+            poll_interval: Duration::from_secs(60),
+        })
+        .expect("starts");
+        let started = std::time::Instant::now();
+        probe.shutdown();
+        probe.shutdown();
+        assert!(started.elapsed() < Duration::from_secs(5));
+        assert!(matches!(
+            probe.poll().health(),
+            ProbeHealth::Stopped | ProbeHealth::Starting | ProbeHealth::Ready
+        ));
+        let _ = std::fs::remove_dir_all(&root);
+    }
+
+    #[test]
+    fn the_default_root_follows_the_documented_location() {
+        let root = default_projects_root();
+        assert!(root.ends_with("projects"));
+    }
+}

crates/lumbridge-harness/src/claude/statusline.rs

@@ -0,0 +1,337 @@
+//! Claude Code's status-line rate-limit payload: the subscription window.
+//!
+//! Claude Code 2.1.80 and later pipe a `rate_limits` object to the configured
+//! `statusLine` command on every turn. The shape is documented inside the CLI
+//! itself, which describes `five_hour` and `seven_day` as
+//!
+//! ```text
+//! "five_hour": {   // present only while the API reports it and its resets_at has not passed
+//!   "used_percentage": number,  // Percentage of limit used (0-100)
+//!   "resets_at": number         // Unix epoch seconds when this window resets
+//! }
+//! ```
+//!
+//! These are the real quota windows, and they cost nothing to read: the CLI is
+//! relaying rate-limit headers that already came back on its own API responses.
+//! That makes them `ProviderReported` under decision 0013's test — the harness
+//! forwards the provider's number rather than computing one.
+//!
+//! `rate_limits_available` is the honest-gap signal, and the CLI documents it:
+//! "False when plan rate limits do not apply (API key, Bedrock, Vertex, or
+//! missing profile scope) — `rate_limits` will be null." An API-key user has no
+//! subscription window, and this field says so rather than leaving us guessing.
+//!
+//! Nothing else from the status-line payload is modelled. It also carries the
+//! session's cost, its transcript path, the current working directory, and the
+//! model — none of which this parser can represent.
+
+use lumbridge_core::UsageWindow;
+use serde::Deserialize;
+
+/// Percent is 0-100 on the wire; a spend limit may exceed 100 once breached.
+const MAX_PERCENT: f64 = 100.0;
+const PERMILLE_PER_PERCENT: f64 = 10.0;
+const MILLIS_PER_SECOND: i64 = 1_000;
+/// Documented window lengths, used to give each reading a window start so the
+/// ledger can scope a burn rate to one quota period.
+const FIVE_HOUR_MS: u64 = 5 * 60 * 60 * 1_000;
+const SEVEN_DAY_MS: u64 = 7 * 24 * 60 * 60 * 1_000;
+
+/// One line of the feed a Lumbridge status-line bridge appends.
+#[derive(Debug, Deserialize)]
+pub(crate) struct FeedLineWire {
+    #[serde(default)]
+    pub(crate) rate_limits_available: Option<bool>,
+    #[serde(default)]
+    pub(crate) rate_limits: Option<RateLimitsWire>,
+}
+
+#[derive(Debug, Default, Deserialize)]
+pub(crate) struct RateLimitsWire {
+    #[serde(default)]
+    pub(crate) five_hour: Option<WindowWire>,
+    #[serde(default)]
+    pub(crate) seven_day: Option<WindowWire>,
+}
+
+#[derive(Debug, Deserialize)]
+pub(crate) struct WindowWire {
+    /// The documented field. `utilization` is accepted as a sibling because the
+    /// OAuth usage endpoint spells the same quantity that way, and a rename
+    /// should degrade rather than silently go dark.
+    #[serde(default)]
+    pub(crate) used_percentage: Option<f64>,
+    #[serde(default)]
+    pub(crate) utilization: Option<f64>,
+    #[serde(default)]
+    pub(crate) resets_at: Option<i64>,
+}
+
+/// Which documented subscription window a reading describes.
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum ClaudeWindowKind {
+    FiveHour,
+    SevenDay,
+}
+
+impl ClaudeWindowKind {
+    pub(crate) const ALL: [Self; 2] = [Self::FiveHour, Self::SevenDay];
+
+    /// Static, account-free profile identifiers.
+    pub(crate) const fn profile_id(self) -> &'static str {
+        match self {
+            Self::FiveHour => "claude-code-five-hour",
+            Self::SevenDay => "claude-code-seven-day",
+        }
+    }
+
+    /// Named for the window, not for a model: these limits are account-scoped.
+    pub(crate) const fn scope(self) -> &'static str {
+        match self {
+            Self::FiveHour => "five-hour window",
+            Self::SevenDay => "seven-day window",
+        }
+    }
+
+    const fn length_ms(self) -> u64 {
+        match self {
+            Self::FiveHour => FIVE_HOUR_MS,
+            Self::SevenDay => SEVEN_DAY_MS,
+        }
+    }
+}
+
+/// What one window in the feed means at a point in time.
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum WindowReading {
+    /// A usable share of the window.
+    Usable {
+        permille: u64,
+        window: Option<UsageWindow>,
+    },
+    /// The window is absent, or the account has no plan limits at all.
+    Absent,
+}
+
+/// Reads one window against the moment it was observed.
+///
+/// Total: every input produces a decision and none of them panics.
+pub(crate) fn parse_window(
+    wire: &WindowWire,
+    kind: ClaudeWindowKind,
+    observed_at_ms: u64,
+) -> WindowReading {
+    let Some(percent) = wire.used_percentage.or(wire.utilization) else {
+        return WindowReading::Absent;
+    };
+    if !percent.is_finite() || percent < 0.0 {
+        return WindowReading::Absent;
+    }
+    // A spend limit can report over 100 once breached; a quota share cannot
+    // exceed the whole, so clamp rather than letting the ledger hold >100%.
+    //
+    // The value is clamped to 0..=100 and scaled by ten before rounding, so it
+    // is a whole number in 0..=1000 — well inside u64 and inside the range f64
+    // represents exactly. The conversion cannot truncate or lose a sign here.
+    #[expect(
+        clippy::cast_possible_truncation,
+        clippy::cast_sign_loss,
+        reason = "a rounded whole number clamped to 0..=1000"
+    )]
+    let permille = (percent.clamp(0.0, MAX_PERCENT) * PERMILLE_PER_PERCENT).round() as u64;
+
+    let window = wire
+        .resets_at
+        .and_then(unix_seconds_to_millis)
+        // The CLI documents that a window is present only while its resets_at
+        // has not passed. One that has is stale, so its start-and-reset pair is
+        // dropped and only the percentage survives.
+        .filter(|resets_at_ms| *resets_at_ms > observed_at_ms)
+        .and_then(|resets_at_ms| {
+            let started_at_ms = resets_at_ms.saturating_sub(kind.length_ms());
+            if started_at_ms <= observed_at_ms {
+                UsageWindow::new(started_at_ms, resets_at_ms).ok()
+            } else {
+                // The local clock disagrees with the provider about the start.
+                // Keep the reset, which is the half that bounds a forecast.
+                Some(UsageWindow::until(resets_at_ms))
+            }
+        });
+
+    WindowReading::Usable { permille, window }
+}
+
+fn unix_seconds_to_millis(seconds: i64) -> Option<u64> {
+    u64::try_from(seconds.checked_mul(MILLIS_PER_SECOND)?).ok()
+}
+
+/// Reads one feed line into a reading per documented window.
+///
+/// `rate_limits_available: false` is a positive statement that this account has
+/// no plan limits, so every window reads as absent rather than unknown.
+pub(crate) fn parse_feed_line(
+    line: &str,
+    observed_at_ms: u64,
+) -> Option<[(ClaudeWindowKind, WindowReading); 2]> {
+    let feed: FeedLineWire = serde_json::from_str(line).ok()?;
+    if feed.rate_limits_available == Some(false) {
+        return Some([
+            (ClaudeWindowKind::FiveHour, WindowReading::Absent),
+            (ClaudeWindowKind::SevenDay, WindowReading::Absent),
+        ]);
+    }
+    let limits = feed.rate_limits?;
+    let five = limits
+        .five_hour
+        .as_ref()
+        .map_or(WindowReading::Absent, |w| {
+            parse_window(w, ClaudeWindowKind::FiveHour, observed_at_ms)
+        });
+    let seven = limits
+        .seven_day
+        .as_ref()
+        .map_or(WindowReading::Absent, |w| {
+            parse_window(w, ClaudeWindowKind::SevenDay, observed_at_ms)
+        });
+    Some([
+        (ClaudeWindowKind::FiveHour, five),
+        (ClaudeWindowKind::SevenDay, seven),
+    ])
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{ClaudeWindowKind, WindowReading, parse_feed_line};
+
+    const HOUR_MS: u64 = 3_600_000;
+    const RESETS_AT_SECONDS: i64 = 1_800_000_000;
+    const RESETS_AT_MS: u64 = 1_800_000_000_000;
+
+    fn line(five: &str, seven: &str) -> String {
+        format!(
+            r#"{{"rate_limits_available":true,"rate_limits":{{"five_hour":{five},"seven_day":{seven}}}}}"#
+        )
+    }
+
+    #[test]
+    fn the_documented_shape_yields_both_windows() {
+        let body = line(
+            &format!(r#"{{"used_percentage":42,"resets_at":{RESETS_AT_SECONDS}}}"#),
+            &format!(r#"{{"used_percentage":8.5,"resets_at":{RESETS_AT_SECONDS}}}"#),
+        );
+        let readings = parse_feed_line(&body, RESETS_AT_MS - HOUR_MS).expect("a valid feed line");
+        let WindowReading::Usable { permille, window } = readings[0].1 else {
+            panic!("the five-hour window is usable");
+        };
+        assert_eq!(readings[0].0, ClaudeWindowKind::FiveHour);
+        assert_eq!(permille, 420);
+        let window = window.expect("a documented window length builds a window");
+        assert_eq!(window.resets_at_ms(), RESETS_AT_MS);
+        assert_eq!(
+            window.started_at_ms(),
+            Some(RESETS_AT_MS - 5 * HOUR_MS),
+            "the five-hour window starts five hours before it resets"
+        );
+
+        let WindowReading::Usable { permille, .. } = readings[1].1 else {
+            panic!("the seven-day window is usable");
+        };
+        assert_eq!(permille, 85, "a fractional percent keeps its tenth");
+    }
+
+    #[test]
+    fn an_api_key_account_reports_absent_rather_than_zero() {
+        // The CLI documents rate_limits_available as false for API key,
+        // Bedrock, Vertex, or a missing profile scope.
+        let readings = parse_feed_line(
+            r#"{"rate_limits_available":false,"rate_limits":null}"#,
+            RESETS_AT_MS,
+        )
+        .expect("a valid feed line");
+        assert!(readings.iter().all(|(_, r)| *r == WindowReading::Absent));
+    }
+
+    #[test]
+    fn a_missing_window_is_absent_without_affecting_the_other() {
+        let body = line(
+            &format!(r#"{{"used_percentage":10,"resets_at":{RESETS_AT_SECONDS}}}"#),
+            "null",
+        );
+        let readings = parse_feed_line(&body, RESETS_AT_MS - HOUR_MS).expect("valid");
+        assert!(matches!(readings[0].1, WindowReading::Usable { .. }));
+        assert_eq!(readings[1].1, WindowReading::Absent);
+    }
+
+    #[test]
+    fn a_reset_that_has_passed_keeps_the_percentage_and_drops_the_window() {
+        let body = line(
+            &format!(r#"{{"used_percentage":99,"resets_at":{RESETS_AT_SECONDS}}}"#),
+            "null",
+        );
+        let readings = parse_feed_line(&body, RESETS_AT_MS + 1).expect("valid");
+        assert_eq!(
+            readings[0].1,
+            WindowReading::Usable {
+                permille: 990,
+                window: None
+            },
+            "a window whose reset has passed is stale, not current"
+        );
+    }
+
+    #[test]
+    fn the_oauth_spelling_of_the_same_field_is_accepted() {
+        let body = line(
+            &format!(r#"{{"utilization":25,"resets_at":{RESETS_AT_SECONDS}}}"#),
+            "null",
+        );
+        let readings = parse_feed_line(&body, RESETS_AT_MS - HOUR_MS).expect("valid");
+        assert!(matches!(
+            readings[0].1,
+            WindowReading::Usable { permille: 250, .. }
+        ));
+    }
+
+    #[test]
+    fn an_over_range_or_nonsense_percentage_is_clamped_or_refused() {
+        let over = line(r#"{"used_percentage":140}"#, "null");
+        assert!(matches!(
+            parse_feed_line(&over, 0).expect("valid")[0].1,
+            WindowReading::Usable {
+                permille: 1_000,
+                window: None
+            }
+        ));
+        for bad in ["-1", "null"] {
+            let body = line(&format!(r#"{{"used_percentage":{bad}}}"#), "null");
+            assert_eq!(
+                parse_feed_line(&body, 0).expect("valid")[0].1,
+                WindowReading::Absent
+            );
+        }
+    }
+
+    #[test]
+    fn malformed_or_unrelated_lines_are_refused() {
+        for body in ["", "not json", "{}", r#"{"rate_limits":{}}"#] {
+            let parsed = parse_feed_line(body, 0);
+            assert!(
+                parsed.is_none_or(|readings| readings
+                    .iter()
+                    .all(|(_, r)| *r == WindowReading::Absent)),
+                "{body:?} must not produce a number"
+            );
+        }
+    }
+
+    #[test]
+    fn window_kinds_have_distinct_account_free_identifiers() {
+        assert_ne!(
+            ClaudeWindowKind::FiveHour.profile_id(),
+            ClaudeWindowKind::SevenDay.profile_id()
+        );
+        for kind in ClaudeWindowKind::ALL {
+            assert!(!kind.profile_id().contains('@'));
+        }
+    }
+}

crates/lumbridge-harness/src/claude/transcript.rs

@@ -0,0 +1,286 @@
+//! The Claude Code transcript reader: token counts in, nothing else out.
+//!
+//! Claude Code writes one JSONL record per event to
+//! `~/.claude/projects/<encoded-cwd>/<session>.jsonl`. Assistant records carry
+//! the API's `usage` object verbatim, which is the only documented local
+//! surface that reports what the harness actually spent.
+//!
+//! **These files contain the user's conversations.** This module is built so
+//! that reading one cannot surface their content: the wire types below model
+//! the record type, the model name, and the four token counters, and nothing
+//! else. Serde discards every unmodelled field while parsing, so message text,
+//! tool inputs, and file contents are never deserialized into a Lumbridge value
+//! at all. That is a structural guarantee, not a convention — there is no
+//! field here that could hold them.
+
+use serde::Deserialize;
+
+/// A transcript line is a single JSON object. Anything longer than this is not
+/// a record we can use, and buffering it would let one file grow this process.
+pub(crate) const MAX_RECORD_BYTES: usize = 1024 * 1024;
+
+/// The only fields this crate reads from a transcript record.
+#[derive(Debug, Deserialize)]
+pub(crate) struct RecordWire {
+    #[serde(rename = "type")]
+    pub(crate) record_type: Option<String>,
+    #[serde(default)]
+    pub(crate) message: Option<MessageWire>,
+}
+
+#[derive(Debug, Deserialize)]
+pub(crate) struct MessageWire {
+    #[serde(default)]
+    pub(crate) model: Option<String>,
+    #[serde(default)]
+    pub(crate) usage: Option<UsageWire>,
+}
+
+/// The API usage object as Claude Code records it.
+///
+/// `cache_creation`, `server_tool_use`, `iterations`, `service_tier`, `speed`,
+/// and `inference_geo` are all present on the wire and all deliberately
+/// unmodelled: a probe should not be able to read account or routing metadata
+/// it has no use for.
+///
+/// The wire names are pinned with explicit renames so the Rust field names can
+/// read naturally without the contract drifting from what Claude Code writes.
+#[derive(Debug, Default, Deserialize)]
+pub(crate) struct UsageWire {
+    #[serde(default, rename = "input_tokens")]
+    pub(crate) input: u64,
+    #[serde(default, rename = "output_tokens")]
+    pub(crate) output: u64,
+    #[serde(default, rename = "cache_creation_input_tokens")]
+    pub(crate) cache_creation: u64,
+    #[serde(default, rename = "cache_read_input_tokens")]
+    pub(crate) cache_read: u64,
+}
+
+/// Tokens the harness recorded, kept split by kind.
+///
+/// The four counters are priced very differently and mean different things, so
+/// they are carried separately rather than collapsed at the point of parsing.
+/// The ledger holds one number per profile; [`TokenTally::total`] is what it
+/// gets, and the split stays available for a detail view.
+#[derive(Clone, Copy, Debug, Default, Eq, PartialEq)]
+pub struct TokenTally {
+    pub input: u64,
+    pub output: u64,
+    pub cache_creation: u64,
+    pub cache_read: u64,
+}
+
+impl TokenTally {
+    /// Every token the API processed.
+    ///
+    /// Cache reads are included because they are billed and do count against a
+    /// subscription window, even though they are re-reads of context already
+    /// counted once. Excluding them would understate consumption; weighting
+    /// them by price would require a price list this adapter does not have and
+    /// would turn a counted fact into an estimate.
+    #[must_use]
+    pub const fn total(self) -> u64 {
+        self.input
+            .saturating_add(self.output)
+            .saturating_add(self.cache_creation)
+            .saturating_add(self.cache_read)
+    }
+
+    #[must_use]
+    pub const fn is_empty(self) -> bool {
+        self.total() == 0
+    }
+
+    pub(crate) fn add(&mut self, other: Self) {
+        self.input = self.input.saturating_add(other.input);
+        self.output = self.output.saturating_add(other.output);
+        self.cache_creation = self.cache_creation.saturating_add(other.cache_creation);
+        self.cache_read = self.cache_read.saturating_add(other.cache_read);
+    }
+}
+
+/// What one transcript line contributed.
+#[derive(Clone, Debug, Default, PartialEq, Eq)]
+pub(crate) struct RecordTally {
+    pub(crate) tokens: TokenTally,
+    /// The model that produced the record, when it named one.
+    pub(crate) model: Option<String>,
+}
+
+/// Reads one transcript line.
+///
+/// Total: a line that is not JSON, not an assistant record, or carries no
+/// usage contributes nothing. A malformed line is skipped rather than faulting
+/// the probe, because a transcript is append-only and the last line of a file
+/// being written is routinely incomplete.
+pub(crate) fn parse_record(line: &str) -> Option<RecordTally> {
+    let record: RecordWire = serde_json::from_str(line).ok()?;
+    if record.record_type.as_deref() != Some("assistant") {
+        return None;
+    }
+    let message = record.message?;
+    let usage = message.usage?;
+    let tokens = TokenTally {
+        input: usage.input,
+        output: usage.output,
+        cache_creation: usage.cache_creation,
+        cache_read: usage.cache_read,
+    };
+    if tokens.is_empty() {
+        return None;
+    }
+    Some(RecordTally {
+        tokens,
+        model: message.model,
+    })
+}
+
+/// Sums the complete lines in a chunk, reporting how many bytes were consumed.
+///
+/// A trailing partial line is left unconsumed so the next read picks it up
+/// whole. This is what makes the running total monotonic: every byte is
+/// counted exactly once, and a file that is still being appended to never
+/// double-counts or loses its last record.
+pub(crate) fn consume_chunk(chunk: &str) -> (RecordTally, usize) {
+    let mut tally = RecordTally::default();
+    let mut consumed = 0;
+    for line in chunk.split_inclusive('\n') {
+        if !line.ends_with('\n') {
+            break; // A partial record. Leave it for the next read.
+        }
+        consumed += line.len();
+        if line.len() > MAX_RECORD_BYTES {
+            continue;
+        }
+        if let Some(record) = parse_record(line.trim_end()) {
+            tally.tokens.add(record.tokens);
+            if record.model.is_some() {
+                tally.model = record.model;
+            }
+        }
+    }
+    (tally, consumed)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{TokenTally, consume_chunk, parse_record};
+
+    /// Shaped exactly like a real assistant record, with the content fields a
+    /// real one carries so the test proves they are ignored rather than absent.
+    ///
+    /// Deliberately one line: a transcript is one JSON object per line, and a
+    /// fixture that wrapped would not exercise the chunk boundary logic.
+    const ASSISTANT: &str = concat!(
+        r#"{"type":"assistant","uuid":"u1","sessionId":"s1","cwd":"/home/x","#,
+        r#""message":{"id":"msg_1","role":"assistant","model":"claude-opus-5","#,
+        r#""content":[{"type":"text","text":"SECRET CONVERSATION CONTENT"}],"#,
+        r#""usage":{"input_tokens":10,"output_tokens":200,"#,
+        r#""cache_creation_input_tokens":1000,"cache_read_input_tokens":50000,"#,
+        r#""service_tier":"standard","cache_creation":{"ephemeral_5m_input_tokens":7}}}}"#,
+    );
+
+    #[test]
+    fn an_assistant_record_yields_its_four_counters() {
+        let record = parse_record(ASSISTANT).expect("an assistant record with usage counts");
+        assert_eq!(
+            record.tokens,
+            TokenTally {
+                input: 10,
+                output: 200,
+                cache_creation: 1_000,
+                cache_read: 50_000,
+            }
+        );
+        assert_eq!(record.model.as_deref(), Some("claude-opus-5"));
+        assert_eq!(record.tokens.total(), 51_210);
+    }
+
+    #[test]
+    fn conversation_content_is_not_representable() {
+        // The record above carries message text. The parser's own output type
+        // has nowhere to put it, so this is a property of the types rather
+        // than of the parsing.
+        let record = parse_record(ASSISTANT).expect("parsed");
+        let rendered = format!("{record:?}");
+        assert!(
+            !rendered.contains("SECRET"),
+            "no debug rendering of a parsed record may include message content"
+        );
+    }
+
+    #[test]
+    fn non_assistant_records_contribute_nothing() {
+        for line in [
+            r#"{"type":"user","message":{"role":"user","content":"hello"}}"#,
+            r#"{"type":"attachment","attachment":{"content":"a file"}}"#,
+            r#"{"type":"system","subtype":"hook"}"#,
+        ] {
+            assert!(parse_record(line).is_none(), "{line} must not be counted");
+        }
+    }
+
+    #[test]
+    fn an_assistant_record_without_usage_contributes_nothing() {
+        assert!(
+            parse_record(r#"{"type":"assistant","message":{"model":"claude-opus-5"}}"#).is_none()
+        );
+    }
+
+    #[test]
+    fn malformed_lines_are_skipped_rather_than_faulting() {
+        for line in ["", "not json", "{", "[]", "\"a string\""] {
+            assert!(parse_record(line).is_none(), "{line:?} must not panic");
+        }
+    }
+
+    #[test]
+    fn a_partial_trailing_line_is_left_for_the_next_read() {
+        let chunk = format!("{ASSISTANT}\n{{\"type\":\"assis");
+        let (tally, consumed) = consume_chunk(&chunk);
+        assert_eq!(tally.tokens.total(), 51_210);
+        assert_eq!(
+            consumed,
+            ASSISTANT.len() + 1,
+            "only the complete record may be consumed"
+        );
+    }
+
+    #[test]
+    fn resuming_from_the_offset_counts_every_byte_exactly_once() {
+        let whole = format!("{ASSISTANT}\n{ASSISTANT}\n");
+        let (first, consumed) = consume_chunk(&whole[..ASSISTANT.len() + 4]);
+        let (second, _) = consume_chunk(&whole[consumed..]);
+        let mut total = first.tokens;
+        total.add(second.tokens);
+        assert_eq!(
+            total.total(),
+            2 * 51_210,
+            "a split read must not double-count or drop a record"
+        );
+    }
+
+    #[test]
+    fn an_oversized_line_is_consumed_but_not_parsed() {
+        let line = format!("{}\n", "x".repeat(super::MAX_RECORD_BYTES + 1));
+        let (tally, consumed) = consume_chunk(&line);
+        assert_eq!(tally.tokens.total(), 0);
+        assert_eq!(
+            consumed,
+            line.len(),
+            "the offset must still advance past it"
+        );
+    }
+
+    #[test]
+    fn totals_saturate_rather_than_overflowing() {
+        let tally = TokenTally {
+            input: u64::MAX,
+            output: u64::MAX,
+            cache_creation: 0,
+            cache_read: 0,
+        };
+        assert_eq!(tally.total(), u64::MAX);
+    }
+}

crates/lumbridge-harness/src/clock.rs

@@ -0,0 +1,92 @@
+//! The only place a wall clock enters the usage path.
+//!
+//! [`lumbridge_core::UsageLedger::record`] rejects an observation older than
+//! the newest one already recorded for that profile, because an append-only
+//! stream cannot move backwards. A raw `SystemTime::now()` can move backwards:
+//! an NTP step, a suspend/resume, or a manual clock change would poison a
+//! profile's stream permanently. This clock anchors wall time once and then
+//! advances it with a monotonic instant, so the timestamps it emits are
+//! non-decreasing by construction while still being comparable to the Unix
+//! reset times a provider reports.
+
+use std::time::{Instant, SystemTime, UNIX_EPOCH};
+
+/// A non-decreasing Unix-millisecond clock.
+#[derive(Debug)]
+pub struct MonotonicWallClock {
+    anchor_unix_ms: u64,
+    anchor: Instant,
+    last_emitted_ms: u64,
+}
+
+impl Default for MonotonicWallClock {
+    fn default() -> Self {
+        Self::start()
+    }
+}
+
+impl MonotonicWallClock {
+    /// Anchors to the system clock once, now.
+    #[must_use]
+    pub fn start() -> Self {
+        let anchor_unix_ms = SystemTime::now()
+            .duration_since(UNIX_EPOCH)
+            .ok()
+            .and_then(|elapsed| u64::try_from(elapsed.as_millis()).ok())
+            .unwrap_or(0);
+        Self::start_at(anchor_unix_ms, Instant::now())
+    }
+
+    /// Anchors to a caller-supplied wall time. Exists so tests can drive the
+    /// clock without waiting for real time to pass.
+    #[must_use]
+    pub(crate) const fn start_at(anchor_unix_ms: u64, anchor: Instant) -> Self {
+        Self {
+            anchor_unix_ms,
+            anchor,
+            last_emitted_ms: anchor_unix_ms,
+        }
+    }
+
+    /// The current time in Unix milliseconds, never earlier than the last value
+    /// this clock returned.
+    pub fn now_ms(&mut self) -> u64 {
+        let elapsed_ms = u64::try_from(self.anchor.elapsed().as_millis()).unwrap_or(u64::MAX);
+        let candidate = self.anchor_unix_ms.saturating_add(elapsed_ms);
+        self.last_emitted_ms = self.last_emitted_ms.max(candidate);
+        self.last_emitted_ms
+    }
+
+    #[must_use]
+    pub const fn last_emitted_ms(&self) -> u64 {
+        self.last_emitted_ms
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::MonotonicWallClock;
+    use std::time::Instant;
+
+    #[test]
+    fn the_clock_starts_at_its_anchor() {
+        let mut clock = MonotonicWallClock::start_at(1_700_000_000_000, Instant::now());
+        assert!(clock.now_ms() >= 1_700_000_000_000);
+    }
+
+    #[test]
+    fn emitted_times_never_move_backwards() {
+        let mut clock = MonotonicWallClock::start();
+        let first = clock.now_ms();
+        let second = clock.now_ms();
+        assert!(second >= first);
+        assert_eq!(clock.last_emitted_ms(), second);
+    }
+
+    #[test]
+    fn a_zero_anchor_still_produces_a_usable_stream() {
+        let mut clock = MonotonicWallClock::start_at(0, Instant::now());
+        let first = clock.now_ms();
+        assert!(clock.now_ms() >= first);
+    }
+}

crates/lumbridge-harness/src/codex/mod.rs

@@ -0,0 +1,500 @@
+//! The first real Lumbridge usage adapter: Codex `account/rateLimits/read`.
+//!
+//! Codex is the first harness because its quota surface is prose-documented
+//! upstream with a worked example, pinned by a checked-in JSON Schema, and
+//! genuinely provider-originated: the app-server forwards the backend's
+//! `x-codex-*-used-percent` response headers rather than computing a number.
+//! Lumbridge obtains it by launching `codex app-server` and letting the harness
+//! resolve its own credentials from `CODEX_HOME`. Lumbridge never reads an auth
+//! file, and [`crate::jsonrpc`] makes asking for a token unrepresentable.
+//!
+//! Three threads, each with one job. The caller's thread starts the child, so a
+//! missing harness is an immediate error rather than a silent fault. A reader
+//! thread does the blocking pipe reads. A worker thread owns the protocol clock
+//! and the writer.
+//!
+//! Shutdown is bounded and never waits on the far end. It kills the child's
+//! whole process group, sets a stop flag, and joins only the worker, which
+//! checks that flag every tick. The reader is deliberately never joined: it is
+//! blocked in a pipe read, and while killing the group is what normally closes
+//! that pipe, a descendant that escaped the group by calling `setsid` would
+//! otherwise hold shutdown open forever. A leaked thread parked on a dead file
+//! descriptor is a far better failure than a hung UI, since `Drop` calls
+//! shutdown.
+
+mod parse;
+mod session;
+
+use std::sync::Arc;
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::mpsc::{self, Receiver, RecvTimeoutError, SyncSender, TryRecvError, TrySendError};
+use std::thread::{self, JoinHandle};
+use std::time::Duration;
+
+use lumbridge_core::{AccountProfile, UsageObservation};
+
+use crate::HarnessError;
+use crate::child::{ProbeChild, ProbeKiller, ProbeReader, ProbeWriter};
+use crate::clock::MonotonicWallClock;
+use crate::jsonrpc::ServerRequestClass;
+use crate::probe::{ProbeHealth, ProbeOutcome, UsageProbe};
+
+use session::{CodexSession, SessionStep};
+
+/// The shortest interval a probe may poll an account quota at. A quota window
+/// is measured in minutes; polling faster only burns the user's process table.
+const MIN_POLL_INTERVAL: Duration = Duration::from_secs(1);
+/// How long the worker waits for a line before re-checking its poll timer.
+const WORKER_TICK: Duration = Duration::from_millis(100);
+const EVENT_QUEUE: usize = 64;
+const LINE_QUEUE: usize = 64;
+
+/// How the probe is launched.
+#[derive(Clone, Debug)]
+pub struct CodexProbeOptions {
+    /// The program to run. Defaults to `codex` on `PATH`.
+    pub program: String,
+    /// How often to request a fresh reading.
+    pub poll_interval: Duration,
+    /// Reported to the harness during the handshake. Not a credential.
+    pub client_name: String,
+    pub client_version: String,
+}
+
+impl Default for CodexProbeOptions {
+    fn default() -> Self {
+        Self {
+            program: "codex".to_owned(),
+            poll_interval: Duration::from_secs(60),
+            client_name: "lumbridge".to_owned(),
+            client_version: env!("CARGO_PKG_VERSION").to_owned(),
+        }
+    }
+}
+
+/// Replays a synthetic app-server transcript through the real session logic.
+///
+/// This is the same state machine the live probe runs, with the process, the
+/// socket, and the clock removed: each frame carries the moment it is to be
+/// treated as observed. It exists so an adapter can be verified end to end
+/// against a fixture, and so a captured (synthetic) transcript can be checked
+/// offline when upstream changes its wire shape.
+///
+/// Frames the session would answer are answered internally and dropped; only
+/// the observations it decided to record come back.
+#[must_use]
+pub fn replay_transcript(frames: &[(&str, u64)]) -> Vec<UsageObservation> {
+    let Ok(mut session) = CodexSession::new() else {
+        return Vec::new();
+    };
+    let _ = session.initialize("lumbridge-replay", "0");
+    let mut observations = Vec::new();
+    for (line, observed_at_ms) in frames {
+        // A replay polls whenever the session will let it, so a transcript
+        // does not have to encode the worker's timer.
+        let _ = session.poll_frame();
+        match session.on_line(line, *observed_at_ms) {
+            SessionStep::Observations(mut batch) => observations.append(&mut batch),
+            SessionStep::Idle
+            | SessionStep::Send(_)
+            | SessionStep::Health(_)
+            | SessionStep::Refused { .. } => {}
+        }
+    }
+    observations
+}
+
+/// What the worker needs to run the handshake and its poll timer.
+struct WorkerSetup {
+    client_name: String,
+    client_version: String,
+    poll_interval_ms: u64,
+}
+
+#[derive(Debug)]
+enum WorkerEvent {
+    Observations(Vec<UsageObservation>),
+    Health(ProbeHealth),
+    RefusedServerRequest(ServerRequestClass),
+}
+
+/// A running Codex quota probe.
+pub struct CodexProbe {
+    profiles: Vec<AccountProfile>,
+    events: Option<Receiver<WorkerEvent>>,
+    killer: ProbeKiller,
+    stop: Arc<AtomicBool>,
+    worker: Option<JoinHandle<()>>,
+    health: ProbeHealth,
+    refused_credential_requests: u64,
+}
+
+impl CodexProbe {
+    /// Starts the probe, its reader, and its worker.
+    ///
+    /// # Errors
+    ///
+    /// Returns [`HarnessError::ProbeIntervalTooShort`] below one second,
+    /// [`HarnessError::EmptyProgram`] for a blank program name, and
+    /// [`HarnessError::SpawnFailed`] when the harness cannot be launched.
+    pub fn start(options: CodexProbeOptions) -> Result<Self, HarnessError> {
+        if options.poll_interval < MIN_POLL_INTERVAL {
+            return Err(HarnessError::ProbeIntervalTooShort);
+        }
+        let CodexProbeOptions {
+            program,
+            poll_interval,
+            client_name,
+            client_version,
+        } = options;
+
+        let session = CodexSession::new()?;
+        let profiles = session.profiles().to_vec();
+        let (writer, reader, killer) = ProbeChild::spawn(&program, &["app-server"])?.into_parts();
+
+        let (line_sender, lines) = mpsc::sync_channel(LINE_QUEUE);
+        let (event_sender, events) = mpsc::sync_channel(EVENT_QUEUE);
+        let stop = Arc::new(AtomicBool::new(false));
+
+        // Deliberately not retained: the reader is never joined. See the module
+        // documentation for why waiting on a blocking pipe read is unsafe here.
+        thread::Builder::new()
+            .name("lumbridge-codex-probe-reader".to_owned())
+            .spawn(move || run_reader(reader, &line_sender))
+            .map_err(|error| HarnessError::SpawnFailed(error.kind()))?;
+
+        let setup = WorkerSetup {
+            client_name,
+            client_version,
+            poll_interval_ms: u64::try_from(poll_interval.as_millis()).unwrap_or(u64::MAX),
+        };
+        let worker_stop = Arc::clone(&stop);
+        let worker = thread::Builder::new()
+            .name("lumbridge-codex-probe".to_owned())
+            .spawn(move || {
+                run_worker(session, &setup, writer, &lines, &event_sender, &worker_stop);
+            })
+            .map_err(|error| HarnessError::SpawnFailed(error.kind()))?;
+
+        Ok(Self {
+            profiles,
+            events: Some(events),
+            killer,
+            stop,
+            worker: Some(worker),
+            health: ProbeHealth::Starting,
+            refused_credential_requests: 0,
+        })
+    }
+
+    /// How many times the harness asked this probe for a credential and was
+    /// refused. A non-zero count is not an error; it is evidence that the
+    /// refusal path ran.
+    #[must_use]
+    pub const fn refused_credential_requests(&self) -> u64 {
+        self.refused_credential_requests
+    }
+}
+
+impl UsageProbe for CodexProbe {
+    fn poll(&mut self) -> ProbeOutcome {
+        let Some(events) = self.events.as_ref() else {
+            return ProbeOutcome::idle(self.health);
+        };
+        let mut observations = Vec::new();
+        loop {
+            match events.try_recv() {
+                Ok(WorkerEvent::Observations(mut batch)) => observations.append(&mut batch),
+                Ok(WorkerEvent::Health(health)) => self.health = health,
+                Ok(WorkerEvent::RefusedServerRequest(class)) => {
+                    if class == ServerRequestClass::CredentialRefresh {
+                        self.refused_credential_requests =
+                            self.refused_credential_requests.saturating_add(1);
+                    }
+                }
+                Err(TryRecvError::Empty) => break,
+                Err(TryRecvError::Disconnected) => {
+                    if !self.health.is_faulted() {
+                        self.health = ProbeHealth::Stopped;
+                    }
+                    self.events = None;
+                    break;
+                }
+            }
+        }
+        ProbeOutcome::new(observations, self.health)
+    }
+
+    fn profiles(&self) -> &[AccountProfile] {
+        &self.profiles
+    }
+
+    fn shutdown(&mut self) {
+        // Order matters. Signal first so the worker cannot start another poll,
+        // then kill the whole group, then join only the worker: it wakes at
+        // most one tick later regardless of what the child is doing.
+        self.stop.store(true, Ordering::Relaxed);
+        self.killer.kill();
+        self.events = None;
+        if let Some(worker) = self.worker.take() {
+            let _ = worker.join();
+        }
+        if !self.health.is_faulted() {
+            self.health = ProbeHealth::Stopped;
+        }
+    }
+}
+
+impl Drop for CodexProbe {
+    fn drop(&mut self) {
+        self.shutdown();
+    }
+}
+
+/// Blocking reads, and nothing else.
+fn run_reader(mut reader: ProbeReader, lines: &SyncSender<Result<String, HarnessError>>) {
+    loop {
+        let outcome = reader.read_line();
+        let terminal = outcome.is_err();
+        if lines.send(outcome).is_err() || terminal {
+            break;
+        }
+    }
+}
+
+/// Owns the protocol clock and the writer. Holds no policy of its own.
+fn run_worker(
+    mut session: CodexSession,
+    setup: &WorkerSetup,
+    mut writer: ProbeWriter,
+    lines: &Receiver<Result<String, HarnessError>>,
+    events: &SyncSender<WorkerEvent>,
+    stop: &AtomicBool,
+) {
+    let mut clock = MonotonicWallClock::start();
+    let handshake = session.initialize(&setup.client_name, &setup.client_version);
+    if let Err(error) = writer.write_line(&handshake) {
+        let _ = events.try_send(WorkerEvent::Health(ProbeHealth::Faulted(error)));
+        return;
+    }
+
+    // Zero means "poll as soon as the handshake lets us", because
+    // `poll_frame` returns nothing until the session is ready.
+    let mut last_poll_ms = 0;
+    loop {
+        if stop.load(Ordering::Relaxed) {
+            break;
+        }
+        let now_ms = clock.now_ms();
+        if now_ms.saturating_sub(last_poll_ms) >= setup.poll_interval_ms
+            && let Some(frame) = session.poll_frame()
+        {
+            last_poll_ms = now_ms;
+            if writer.write_line(&frame).is_err() {
+                let _ =
+                    events.try_send(WorkerEvent::Health(ProbeHealth::Faulted(HarnessError::Eof)));
+                break;
+            }
+        }
+
+        match lines.recv_timeout(WORKER_TICK) {
+            Ok(Ok(line)) => {
+                if !dispatch(&mut session, &mut writer, events, &line, clock.now_ms()) {
+                    break;
+                }
+            }
+            Ok(Err(error)) => {
+                let _ = events.try_send(WorkerEvent::Health(ProbeHealth::Faulted(error)));
+                break;
+            }
+            Err(RecvTimeoutError::Timeout) => {}
+            Err(RecvTimeoutError::Disconnected) => break,
+        }
+    }
+}
+
+/// Applies one line. Returns whether the worker should keep running.
+fn dispatch(
+    session: &mut CodexSession,
+    writer: &mut ProbeWriter,
+    events: &SyncSender<WorkerEvent>,
+    line: &str,
+    now_ms: u64,
+) -> bool {
+    match session.on_line(line.trim_end(), now_ms) {
+        SessionStep::Idle => true,
+        SessionStep::Send(frame) => {
+            if writer.write_line(&frame).is_err() {
+                return false;
+            }
+            if session.is_ready() {
+                let _ = events.try_send(WorkerEvent::Health(ProbeHealth::Ready));
+            }
+            true
+        }
+        SessionStep::Refused { line, class } => {
+            let _ = events.try_send(WorkerEvent::RefusedServerRequest(class));
+            writer.write_line(&line).is_ok()
+        }
+        SessionStep::Observations(observations) => !matches!(
+            events.try_send(WorkerEvent::Observations(observations)),
+            Err(TrySendError::Disconnected(_))
+        ),
+        SessionStep::Health(health) => {
+            let keep_running = !health.is_faulted();
+            let _ = events.try_send(WorkerEvent::Health(health));
+            keep_running
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{CodexProbe, CodexProbeOptions};
+    use crate::HarnessError;
+    use crate::probe::{ProbeHealth, UsageProbe};
+    use std::time::Duration;
+
+    /// A synthetic launcher that ignores its arguments, backgrounds a helper
+    /// that inherits stdio, and waits. This is the shape of a wrapper script
+    /// that execs the real program as a grandchild, which is how the npm
+    /// distribution of Codex behaves.
+    fn launcher_script(tag: &str) -> Option<std::path::PathBuf> {
+        use std::io::Write;
+
+        let path = std::env::temp_dir().join(format!(
+            "lumbridge-probe-launcher-{}-{tag}.sh",
+            std::process::id()
+        ));
+        let mut file = std::fs::File::create(&path).ok()?;
+        file.write_all(b"#!/bin/sh\nsleep 600 &\nwait\n").ok()?;
+        drop(file);
+        #[cfg(unix)]
+        {
+            use std::os::unix::fs::PermissionsExt;
+            std::fs::set_permissions(&path, std::fs::Permissions::from_mode(0o755)).ok()?;
+        }
+        Some(path)
+    }
+
+    fn options(program: &str) -> CodexProbeOptions {
+        CodexProbeOptions {
+            program: program.to_owned(),
+            poll_interval: Duration::from_secs(1),
+            ..CodexProbeOptions::default()
+        }
+    }
+
+    #[test]
+    fn a_too_frequent_poll_interval_is_refused() {
+        let options = CodexProbeOptions {
+            poll_interval: Duration::from_millis(10),
+            ..CodexProbeOptions::default()
+        };
+        assert_eq!(
+            CodexProbe::start(options).err(),
+            Some(HarnessError::ProbeIntervalTooShort)
+        );
+    }
+
+    #[test]
+    fn an_empty_program_is_refused() {
+        assert_eq!(
+            CodexProbe::start(options("  ")).err(),
+            Some(HarnessError::EmptyProgram)
+        );
+    }
+
+    #[test]
+    fn a_missing_harness_fails_at_start_rather_than_silently() {
+        assert!(matches!(
+            CodexProbe::start(options("lumbridge-no-such-harness")).err(),
+            Some(HarnessError::SpawnFailed(_))
+        ));
+    }
+
+    #[test]
+    fn a_silent_harness_can_still_be_shut_down() {
+        let Some(script) = launcher_script("silent") else {
+            return;
+        };
+        let Some(program) = script.to_str() else {
+            return;
+        };
+        let Ok(mut probe) = CodexProbe::start(options(program)) else {
+            let _ = std::fs::remove_file(&script);
+            return;
+        };
+        assert_eq!(probe.profiles().len(), 2);
+        assert_eq!(probe.poll().health(), ProbeHealth::Starting);
+        probe.shutdown();
+        assert!(matches!(
+            probe.poll().health(),
+            ProbeHealth::Stopped | ProbeHealth::Faulted(_)
+        ));
+        let _ = std::fs::remove_file(&script);
+    }
+
+    #[test]
+    fn shutdown_is_idempotent() {
+        let Some(script) = launcher_script("idempotent") else {
+            return;
+        };
+        let Some(program) = script.to_str() else {
+            return;
+        };
+        let Ok(mut probe) = CodexProbe::start(options(program)) else {
+            let _ = std::fs::remove_file(&script);
+            return;
+        };
+        probe.shutdown();
+        probe.shutdown();
+        let _ = std::fs::remove_file(&script);
+    }
+
+    #[test]
+    fn shutdown_returns_promptly_behind_a_launcher_that_never_speaks() {
+        // The grandchild inherits our stdout and never writes, so the reader is
+        // blocked in a pipe read that will never complete on its own. Before
+        // the process-group kill and the bounded join, shutdown blocked here
+        // until the orphan exited — ten minutes, in this fixture.
+        let Some(script) = launcher_script("prompt") else {
+            return;
+        };
+        let Some(program) = script.to_str() else {
+            return;
+        };
+        let Ok(mut probe) = CodexProbe::start(options(program)) else {
+            let _ = std::fs::remove_file(&script);
+            return;
+        };
+        let started = std::time::Instant::now();
+        probe.shutdown();
+        let elapsed = started.elapsed();
+        let _ = std::fs::remove_file(&script);
+        assert!(
+            elapsed < Duration::from_secs(5),
+            "shutdown must not wait on the far end, took {elapsed:?}"
+        );
+    }
+
+    #[test]
+    fn a_harness_that_exits_immediately_faults_the_probe() {
+        let Ok(mut probe) = CodexProbe::start(options("true")) else {
+            return;
+        };
+        let mut health = probe.poll().health();
+        for _ in 0..200 {
+            if health != ProbeHealth::Starting {
+                break;
+            }
+            std::thread::sleep(Duration::from_millis(10));
+            health = probe.poll().health();
+        }
+        assert!(
+            matches!(health, ProbeHealth::Faulted(_)),
+            "a harness that closes its output must surface as a fault, got {health:?}"
+        );
+        probe.shutdown();
+    }
+}

crates/lumbridge-harness/src/codex/parse.rs

@@ -0,0 +1,397 @@
+//! The Codex rate-limit parser: hostile input in, a closed decision out.
+//!
+//! Every skew, roll, overflow, and out-of-range case is decided here, by pure
+//! functions over deserialized wire values with no process, no clock, and no
+//! network. That is what makes the honesty rules testable.
+//!
+//! Wire shape adapted from the openai/codex app-server protocol (Apache-2.0),
+//! commit `17e8101699c5062117d0d37f504313e8af53b043`:
+//! `codex-rs/app-server/README.md` section "7) Rate limits (`ChatGPT`)" and the
+//! checked-in schema `codex_app_server_protocol.v2.schemas.json`, where
+//! `RateLimitWindow` requires only `usedPercent` and both
+//! `windowDurationMins` and `resetsAt` are nullable. No upstream code is
+//! copied; only the observable wire contract is mirrored.
+
+use lumbridge_core::{AccountProfileId, UsageObservation, UsageProvenance, UsageUnit, UsageWindow};
+use serde::Deserialize;
+
+/// Percent is an integer 0..=100 on the wire, so the honest resolution of a
+/// Codex reading is one percent even though the ledger stores permille.
+const MAX_USED_PERCENT: i64 = 100;
+const PERMILLE_PER_PERCENT: u64 = 10;
+const MILLIS_PER_SECOND: i64 = 1_000;
+const MILLIS_PER_MINUTE: i64 = 60_000;
+/// A permille ceiling makes the window fraction a bounded quantity, which is
+/// what lets the ledger derive an exhaustion estimate. It is not an invented
+/// quota: a percentage is by definition a share of one whole.
+const PERMILLE_CEILING: u64 = 1_000;
+
+/// One quota window exactly as Codex sends it.
+///
+/// Only the three documented fields are deserialized. Sibling fields such as
+/// `rateLimitResetCredits`, `individualLimit`, `accountId`, and
+/// `rateLimitUpsell` are deliberately not modelled: a probe should not be able
+/// to accidentally read account metadata it has no use for.
+#[derive(Debug, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub(crate) struct RateLimitWindowWire {
+    /// Widened to `i64` so an out-of-range wire value reaches our own check
+    /// rather than failing inside serde with a less specific error.
+    pub(crate) used_percent: i64,
+    #[serde(default)]
+    pub(crate) window_duration_mins: Option<i64>,
+    #[serde(default)]
+    pub(crate) resets_at: Option<i64>,
+}
+
+/// The snapshot. Every field is optional in the upstream schema, including
+/// `primary`, so absence must be a first-class case rather than a parse error.
+#[derive(Debug, Default, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub(crate) struct RateLimitSnapshotWire {
+    #[serde(default)]
+    pub(crate) primary: Option<RateLimitWindowWire>,
+    #[serde(default)]
+    pub(crate) secondary: Option<RateLimitWindowWire>,
+}
+
+/// The `account/rateLimits/read` result body, which is also the shape of the
+/// `account/rateLimits/updated` notification's params.
+#[derive(Debug, Default, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub(crate) struct RateLimitsResultWire {
+    #[serde(default)]
+    pub(crate) rate_limits: RateLimitSnapshotWire,
+}
+
+/// Why a syntactically valid window still cannot be recorded as a fact.
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum UnusableReason {
+    /// A percentage outside 0..=100. The relay is not telling the truth.
+    PercentOutOfRange,
+    /// The window had already ended when the reading was taken. Decision 0012
+    /// requires an expired window to read as rolled over rather than as a
+    /// frozen percentage from a window that is gone.
+    WindowEnded,
+    /// A reset timestamp that cannot be represented as Unix milliseconds.
+    ResetsAtOutOfRange,
+}
+
+/// What the parser decided about one window.
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum WindowReading {
+    /// A usable share of the window, with a window attached when the wire
+    /// carried enough to build one.
+    Usable {
+        permille: u64,
+        window: Option<UsageWindow>,
+    },
+    /// The reading must be recorded as an explicit gap.
+    Unusable(UnusableReason),
+}
+
+/// Decides what a single Codex window means at a point in time.
+///
+/// Total: every input produces a decision, and none of them panics.
+pub(crate) fn parse_window(wire: &RateLimitWindowWire, observed_at_ms: u64) -> WindowReading {
+    if wire.used_percent < 0 || wire.used_percent > MAX_USED_PERCENT {
+        return WindowReading::Unusable(UnusableReason::PercentOutOfRange);
+    }
+    // Lossless: used_percent is already proven to be 0..=100.
+    let permille = u64::try_from(wire.used_percent).unwrap_or(0) * PERMILLE_PER_PERCENT;
+
+    let Some(resets_at_seconds) = wire.resets_at else {
+        // A percentage with no reset time is still a true percentage. It just
+        // cannot answer "when does this refill?".
+        return WindowReading::Usable {
+            permille,
+            window: None,
+        };
+    };
+    let Some(resets_at_ms) = unix_seconds_to_millis(resets_at_seconds) else {
+        return WindowReading::Unusable(UnusableReason::ResetsAtOutOfRange);
+    };
+    if observed_at_ms >= resets_at_ms {
+        return WindowReading::Unusable(UnusableReason::WindowEnded);
+    }
+
+    // The reset is a fact on its own. Codex declares the duration and the reset
+    // independently, so a missing or unreconcilable start must cost the start
+    // and nothing else: the reset still bounds the exhaustion forecast and
+    // still identifies which quota period this reading belongs to.
+    let window = wire
+        .window_duration_mins
+        .and_then(|minutes| window_start_ms(resets_at_ms, minutes))
+        // A start after the observation means the local clock disagrees with
+        // the provider's. That disagreement is about the start, not the reset.
+        .filter(|started_at_ms| *started_at_ms <= observed_at_ms)
+        .and_then(|started_at_ms| UsageWindow::new(started_at_ms, resets_at_ms).ok())
+        .unwrap_or_else(|| UsageWindow::until(resets_at_ms));
+
+    WindowReading::Usable {
+        permille,
+        window: Some(window),
+    }
+}
+
+fn unix_seconds_to_millis(seconds: i64) -> Option<u64> {
+    let millis = seconds.checked_mul(MILLIS_PER_SECOND)?;
+    u64::try_from(millis).ok()
+}
+
+fn window_start_ms(resets_at_ms: u64, duration_minutes: i64) -> Option<u64> {
+    if duration_minutes <= 0 {
+        return None;
+    }
+    let duration_ms = u64::try_from(duration_minutes.checked_mul(MILLIS_PER_MINUTE)?).ok()?;
+    resets_at_ms.checked_sub(duration_ms)
+}
+
+/// Turns a decision into the observation the ledger should hold.
+///
+/// A usable reading is `ProviderReported`: Codex forwards the backend's
+/// `x-codex-*-used-percent` response headers rather than computing the value,
+/// so the provider is the one making the claim. Anything else is an explicit
+/// gap, never a zero.
+pub(crate) fn to_observation(
+    profile: &AccountProfileId,
+    reading: WindowReading,
+    observed_at_ms: u64,
+) -> UsageObservation {
+    let WindowReading::Usable { permille, window } = reading else {
+        return UsageObservation::unavailable(profile.clone(), observed_at_ms);
+    };
+    let Ok(observation) = UsageObservation::counted(
+        profile.clone(),
+        UsageUnit::WindowPermille,
+        permille,
+        UsageProvenance::ProviderReported,
+        observed_at_ms,
+    ) else {
+        return UsageObservation::unavailable(profile.clone(), observed_at_ms);
+    };
+    let observation = observation
+        .with_limit(PERMILLE_CEILING)
+        .unwrap_or_else(|_| UsageObservation::unavailable(profile.clone(), observed_at_ms));
+    let Some(window) = window else {
+        return observation;
+    };
+    // A window that no longer contains the observation is rejected by core.
+    // Keeping the windowless observation preserves the provider's percentage
+    // instead of discarding a true fact over a timing disagreement.
+    observation
+        .clone()
+        .with_window(window)
+        .unwrap_or(observation)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{
+        RateLimitSnapshotWire, RateLimitWindowWire, RateLimitsResultWire, UnusableReason,
+        WindowReading, parse_window, to_observation,
+    };
+    use lumbridge_core::{AccountProfileId, UsageProvenance, UsageUnit, UsageWindow};
+
+    const HOUR_MS: u64 = 3_600_000;
+    const RESETS_AT_SECONDS: i64 = 1_730_947_200;
+    const RESETS_AT_MS: u64 = 1_730_947_200_000;
+
+    fn profile() -> AccountProfileId {
+        AccountProfileId::new("codex-app-server-primary").expect("a valid fixture ID")
+    }
+
+    fn wire(
+        used_percent: i64,
+        minutes: Option<i64>,
+        resets_at: Option<i64>,
+    ) -> RateLimitWindowWire {
+        RateLimitWindowWire {
+            used_percent,
+            window_duration_mins: minutes,
+            resets_at,
+        }
+    }
+
+    #[test]
+    fn the_documented_example_parses_to_a_bounded_window() {
+        // The worked example from the upstream README, read one hour before it
+        // resets: { "usedPercent": 25, "windowDurationMins": 15, ... } widened
+        // to a duration long enough to contain the observation.
+        let reading = parse_window(
+            &wire(25, Some(300), Some(RESETS_AT_SECONDS)),
+            RESETS_AT_MS - HOUR_MS,
+        );
+        let WindowReading::Usable { permille, window } = reading else {
+            panic!("the documented example must be usable");
+        };
+        assert_eq!(permille, 250);
+        let window = window.expect("a duration and a reset build a window");
+        assert_eq!(window.resets_at_ms(), RESETS_AT_MS);
+        assert_eq!(window.started_at_ms(), Some(RESETS_AT_MS - 300 * 60_000));
+    }
+
+    #[test]
+    fn a_reading_taken_after_the_reset_is_a_rolled_window_not_a_frozen_percentage() {
+        let reading = parse_window(
+            &wire(93, Some(300), Some(RESETS_AT_SECONDS)),
+            RESETS_AT_MS + 1,
+        );
+        assert_eq!(
+            reading,
+            WindowReading::Unusable(UnusableReason::WindowEnded)
+        );
+        let observation = to_observation(&profile(), reading, RESETS_AT_MS + 1);
+        assert_eq!(observation.provenance(), UsageProvenance::Unavailable);
+        assert_eq!(observation.consumed(), 0);
+    }
+
+    #[test]
+    fn an_absent_duration_keeps_the_percentage_and_the_reset() {
+        // Codex declares the duration and the reset independently. Dropping
+        // the reset for want of a duration would discard a fact the provider
+        // stated, and would let a burn baseline be drawn across two quota
+        // periods while the exhaustion guard never fired.
+        let reading = parse_window(
+            &wire(40, None, Some(RESETS_AT_SECONDS)),
+            RESETS_AT_MS - HOUR_MS,
+        );
+        assert_eq!(
+            reading,
+            WindowReading::Usable {
+                permille: 400,
+                window: Some(UsageWindow::until(RESETS_AT_MS))
+            }
+        );
+    }
+
+    #[test]
+    fn an_absent_reset_keeps_the_percentage_and_drops_the_window() {
+        let reading = parse_window(&wire(40, Some(300), None), 5_000);
+        assert_eq!(
+            reading,
+            WindowReading::Usable {
+                permille: 400,
+                window: None
+            }
+        );
+    }
+
+    #[test]
+    fn a_start_after_the_observation_costs_the_start_and_not_the_reset() {
+        // A 15 minute window resetting far in the future starts after now:
+        // the local clock and the provider disagree about the start only.
+        let reading = parse_window(&wire(25, Some(15), Some(RESETS_AT_SECONDS)), 1_000);
+        assert_eq!(
+            reading,
+            WindowReading::Usable {
+                permille: 250,
+                window: Some(UsageWindow::until(RESETS_AT_MS))
+            },
+            "clock disagreement must not discard the provider's reset time"
+        );
+    }
+
+    #[test]
+    fn an_out_of_range_percentage_is_refused() {
+        for percent in [-1, 101, i64::MAX, i64::MIN] {
+            assert_eq!(
+                parse_window(&wire(percent, Some(300), Some(RESETS_AT_SECONDS)), 0),
+                WindowReading::Unusable(UnusableReason::PercentOutOfRange),
+                "percent {percent} must not be recorded"
+            );
+        }
+    }
+
+    #[test]
+    fn an_unrepresentable_reset_is_refused_without_overflowing() {
+        for seconds in [-1, i64::MIN, i64::MAX] {
+            assert!(matches!(
+                parse_window(&wire(25, Some(300), Some(seconds)), 0),
+                WindowReading::Unusable(_)
+            ));
+        }
+    }
+
+    #[test]
+    fn a_zero_or_negative_duration_does_not_underflow() {
+        for minutes in [0, -5, i64::MIN] {
+            let reading = parse_window(&wire(25, Some(minutes), Some(RESETS_AT_SECONDS)), 1_000);
+            assert_eq!(
+                reading,
+                WindowReading::Usable {
+                    permille: 250,
+                    window: Some(UsageWindow::until(RESETS_AT_MS))
+                },
+                "a nonsensical duration must not cost the reset"
+            );
+        }
+    }
+
+    #[test]
+    fn a_weekly_window_matches_the_shape_a_live_account_returns() {
+        // Observed from a real account: usedPercent 18, windowDurationMins
+        // 10080 (one week), resetsAt in Unix seconds, secondary null.
+        let week_minutes: i64 = 10_080;
+        let observed_at_ms = RESETS_AT_MS - 24 * HOUR_MS;
+        let reading = parse_window(
+            &wire(18, Some(week_minutes), Some(RESETS_AT_SECONDS)),
+            observed_at_ms,
+        );
+        let WindowReading::Usable { permille, window } = reading else {
+            panic!("a weekly window is usable");
+        };
+        assert_eq!(permille, 180);
+        let window = window.expect("a week-long window is still a window");
+        assert_eq!(
+            window.started_at_ms(),
+            Some(RESETS_AT_MS - u64::try_from(week_minutes).expect("a positive duration") * 60_000)
+        );
+    }
+
+    #[test]
+    fn a_usable_reading_becomes_a_provider_reported_permille_observation() {
+        let reading = parse_window(
+            &wire(25, Some(300), Some(RESETS_AT_SECONDS)),
+            RESETS_AT_MS - HOUR_MS,
+        );
+        let observation = to_observation(&profile(), reading, RESETS_AT_MS - HOUR_MS);
+        assert_eq!(observation.provenance(), UsageProvenance::ProviderReported);
+        assert_eq!(observation.unit(), Some(UsageUnit::WindowPermille));
+        assert_eq!(observation.consumed(), 250);
+        assert_eq!(observation.limit(), Some(1_000));
+        assert!(observation.window().is_some());
+    }
+
+    #[test]
+    fn an_absent_primary_window_deserializes_rather_than_failing() {
+        let body: RateLimitsResultWire = serde_json::from_str(
+            r#"{"rateLimits":{"primary":null,"secondary":null,"rateLimitReachedType":null}}"#,
+        )
+        .expect("every snapshot field is optional upstream");
+        assert!(body.rate_limits.primary.is_none());
+        assert!(body.rate_limits.secondary.is_none());
+    }
+
+    #[test]
+    fn unmodelled_sibling_fields_are_ignored_not_rejected() {
+        // The upstream result also carries rateLimitResetCredits, accountId,
+        // individualLimit and rateLimitUpsell. A probe must tolerate them and
+        // must not model them.
+        let body: RateLimitsResultWire = serde_json::from_str(
+            r#"{"rateLimits":{"primary":{"usedPercent":25,"windowDurationMins":15,
+                "resetsAt":1730947200},"individualLimit":null,"accountId":"acct_x",
+                "planType":"plus"},"rateLimitResetCredits":{"availableCount":2}}"#,
+        )
+        .expect("unknown fields are ignored");
+        let primary = body.rate_limits.primary.expect("primary is present");
+        assert_eq!(primary.used_percent, 25);
+    }
+
+    #[test]
+    fn a_snapshot_with_no_windows_defaults_rather_than_panicking() {
+        let snapshot = RateLimitSnapshotWire::default();
+        assert!(snapshot.primary.is_none());
+        assert!(snapshot.secondary.is_none());
+    }
+}

crates/lumbridge-harness/src/codex/session.rs

@@ -0,0 +1,584 @@
+//! The Codex probe's protocol logic, as a pure state machine.
+//!
+//! Every decision the probe makes about what to send, what to record, and what
+//! to refuse happens here, driven by `&str` lines and an explicit timestamp.
+//! No process, no socket, and no clock are involved, so the whole protocol is
+//! testable from string fixtures. The thread in [`super::worker`] does nothing
+//! but move bytes between a child process and this type.
+
+use lumbridge_core::{AccountProfile, AccountProfileId, UsageObservation, UsageWindow};
+use serde_json::{Value, json};
+
+use crate::HarnessError;
+use crate::jsonrpc::{
+    AcceptedNotification, INVALID_REQUEST, Inbound, OutboundRequest, ServerRequestClass, classify,
+    encode_notification, encode_refusal, encode_request,
+};
+use crate::probe::ProbeHealth;
+
+use super::parse::{
+    RateLimitSnapshotWire, RateLimitsResultWire, WindowReading, parse_window, to_observation,
+};
+
+/// Re-emit an unchanged reading at least this often so the ledger keeps a
+/// baseline for its burn rate. Without a floor, a quiet account would hold one
+/// lone observation and the footer could never derive a rate; without a
+/// ceiling on how often we re-record, the bounded retention would be flushed
+/// of real history by identical rows.
+const MIN_REEMIT_INTERVAL_MS: u64 = 300_000;
+
+/// The two windows Codex reports are two separate facts about two separate
+/// quota periods. Merging them would mean silently choosing one.
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub(crate) enum WindowSlot {
+    Primary,
+    Secondary,
+}
+
+impl WindowSlot {
+    const ALL: [Self; 2] = [Self::Primary, Self::Secondary];
+
+    /// Static, account-free identifiers. A profile ID is persisted and must
+    /// never carry an account identity.
+    const fn profile_id(self) -> &'static str {
+        match self {
+            Self::Primary => "codex-app-server-primary",
+            Self::Secondary => "codex-app-server-secondary",
+        }
+    }
+
+    /// Codex rate limits are account-scoped, not model-scoped. Naming the
+    /// scope is honest; inventing a model name would not be.
+    const fn scope(self) -> &'static str {
+        match self {
+            Self::Primary => "primary window",
+            Self::Secondary => "secondary window",
+        }
+    }
+}
+
+/// What the session wants the caller to do next.
+#[derive(Clone, Debug)]
+pub(crate) enum SessionStep {
+    /// Nothing to do.
+    Idle,
+    /// Write this line to the child's stdin.
+    Send(String),
+    /// Record these observations.
+    Observations(Vec<UsageObservation>),
+    /// The probe's liveness changed.
+    Health(ProbeHealth),
+    /// A server-to-client request was refused. Write the line, and note why.
+    Refused {
+        line: String,
+        class: ServerRequestClass,
+    },
+}
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+enum Stage {
+    AwaitingInitialize,
+    Ready,
+}
+
+/// The last thing recorded for one profile, used to suppress identical rows.
+#[derive(Clone, Copy, Debug)]
+struct LastRecorded {
+    permille: u64,
+    window: Option<UsageWindow>,
+    at_ms: u64,
+}
+
+pub(crate) struct CodexSession {
+    profiles: Vec<AccountProfile>,
+    stage: Stage,
+    next_id: u64,
+    initialize_id: u64,
+    pending_read_id: Option<u64>,
+    last: [Option<LastRecorded>; 2],
+    refused_credential_requests: u64,
+}
+
+impl CodexSession {
+    pub(crate) fn new() -> Result<Self, HarnessError> {
+        let profiles = WindowSlot::ALL
+            .into_iter()
+            .map(|slot| {
+                AccountProfile::new(
+                    slot.profile_id(),
+                    "Codex",
+                    "ChatGPT",
+                    slot.scope(),
+                    "subscription",
+                )
+                .map_err(|_| HarnessError::EmptyProgram)
+            })
+            .collect::<Result<Vec<_>, _>>()?;
+        Ok(Self {
+            profiles,
+            stage: Stage::AwaitingInitialize,
+            next_id: 1,
+            initialize_id: 0,
+            pending_read_id: None,
+            last: [None, None],
+            refused_credential_requests: 0,
+        })
+    }
+
+    pub(crate) fn profiles(&self) -> &[AccountProfile] {
+        &self.profiles
+    }
+
+    /// Whether the handshake has completed and readings may be requested.
+    pub(crate) const fn is_ready(&self) -> bool {
+        matches!(self.stage, Stage::Ready)
+    }
+
+    #[cfg(test)]
+    pub(crate) const fn refused_credential_requests(&self) -> u64 {
+        self.refused_credential_requests
+    }
+
+    fn take_id(&mut self) -> u64 {
+        let id = self.next_id;
+        self.next_id = self.next_id.saturating_add(1);
+        id
+    }
+
+    /// The handshake frame. Codex requires `initialize` before any other call.
+    pub(crate) fn initialize(&mut self, client_name: &str, client_version: &str) -> String {
+        let id = self.take_id();
+        self.initialize_id = id;
+        encode_request(
+            id,
+            OutboundRequest::Initialize,
+            &json!({ "clientInfo": { "name": client_name, "version": client_version } }),
+        )
+    }
+
+    /// Requests a fresh reading, unless one is already outstanding.
+    pub(crate) fn poll_frame(&mut self) -> Option<String> {
+        if self.stage != Stage::Ready || self.pending_read_id.is_some() {
+            return None;
+        }
+        let id = self.take_id();
+        self.pending_read_id = Some(id);
+        Some(encode_request(
+            id,
+            OutboundRequest::ReadAccountRateLimits,
+            &json!({}),
+        ))
+    }
+
+    /// Consumes one line from the child.
+    pub(crate) fn on_line(&mut self, line: &str, observed_at_ms: u64) -> SessionStep {
+        let Some(inbound) = classify(line) else {
+            return SessionStep::Health(ProbeHealth::Faulted(HarnessError::Malformed));
+        };
+        match inbound {
+            Inbound::ServerRequest { id, class } => {
+                if class == ServerRequestClass::CredentialRefresh {
+                    self.refused_credential_requests =
+                        self.refused_credential_requests.saturating_add(1);
+                }
+                SessionStep::Refused {
+                    line: encode_refusal(&id),
+                    class,
+                }
+            }
+            Inbound::Result { id, result } => self.on_result(id, &result, observed_at_ms),
+            Inbound::Failure { id, code } => self.on_failure(id, code, observed_at_ms),
+            // Matching the kind exhaustively means adding a second accepted
+            // notification cannot silently reuse the rate-limit path.
+            Inbound::Notification {
+                kind: AcceptedNotification::AccountRateLimitsUpdated,
+                params,
+            } => self.on_snapshot_body(&params, observed_at_ms, true),
+            Inbound::Ignored => SessionStep::Idle,
+        }
+    }
+
+    fn on_result(&mut self, id: u64, result: &Value, observed_at_ms: u64) -> SessionStep {
+        if self.stage == Stage::AwaitingInitialize && id == self.initialize_id {
+            self.stage = Stage::Ready;
+            return SessionStep::Send(encode_notification("initialized", &json!({})));
+        }
+        if self.pending_read_id == Some(id) {
+            self.pending_read_id = None;
+            return self.on_snapshot_body(result, observed_at_ms, false);
+        }
+        SessionStep::Idle
+    }
+
+    fn on_failure(&mut self, id: u64, code: i64, observed_at_ms: u64) -> SessionStep {
+        if self.pending_read_id == Some(id) {
+            self.pending_read_id = None;
+            // Codex answers -32600 when the account cannot report a quota,
+            // which is a true statement about the account rather than a fault.
+            if code == INVALID_REQUEST {
+                let observations = self.unavailable_for_all(observed_at_ms);
+                self.forget_all();
+                return SessionStep::Observations(observations);
+            }
+            return SessionStep::Health(ProbeHealth::Faulted(HarnessError::Rejected(code)));
+        }
+        if id == self.initialize_id {
+            return SessionStep::Health(ProbeHealth::Faulted(HarnessError::Rejected(code)));
+        }
+        SessionStep::Idle
+    }
+
+    /// Turns a `rateLimits` value into observations.
+    ///
+    /// `sparse` marks a push notification. Upstream documents that a rolling
+    /// update may omit values and that omission "does not clear a previously
+    /// observed value", so an absent window on a push records nothing at all.
+    /// It must not record an explicit gap, and it must not be merged into the
+    /// last full read either: a merged composite would look provider-reported
+    /// while being partly a memory.
+    fn on_snapshot_body(&mut self, body: &Value, observed_at_ms: u64, sparse: bool) -> SessionStep {
+        let Ok(body) = serde_json::from_value::<RateLimitsResultWire>(body.clone()) else {
+            return SessionStep::Health(ProbeHealth::Faulted(HarnessError::Malformed));
+        };
+        let observations = self.observations_from(&body.rate_limits, observed_at_ms, sparse);
+        if observations.is_empty() {
+            SessionStep::Idle
+        } else {
+            SessionStep::Observations(observations)
+        }
+    }
+
+    fn observations_from(
+        &mut self,
+        snapshot: &RateLimitSnapshotWire,
+        observed_at_ms: u64,
+        sparse: bool,
+    ) -> Vec<UsageObservation> {
+        let mut observations = Vec::new();
+        for slot in WindowSlot::ALL {
+            let index = slot as usize;
+            let wire = match slot {
+                WindowSlot::Primary => snapshot.primary.as_ref(),
+                WindowSlot::Secondary => snapshot.secondary.as_ref(),
+            };
+            let Some(wire) = wire else {
+                if !sparse {
+                    // A full read that omits a window means the window is gone.
+                    if self.last[index].take().is_some() {
+                        observations.push(UsageObservation::unavailable(
+                            self.profile_id(index),
+                            observed_at_ms,
+                        ));
+                    }
+                }
+                continue;
+            };
+            let reading = parse_window(wire, observed_at_ms);
+            if !self.should_record(index, reading, observed_at_ms) {
+                continue;
+            }
+            observations.push(to_observation(
+                &self.profile_id(index),
+                reading,
+                observed_at_ms,
+            ));
+        }
+        observations
+    }
+
+    fn profile_id(&self, index: usize) -> AccountProfileId {
+        self.profiles[index].id().clone()
+    }
+
+    /// Suppresses an identical reading unless the re-emit floor has passed.
+    fn should_record(&mut self, index: usize, reading: WindowReading, observed_at_ms: u64) -> bool {
+        let WindowReading::Usable { permille, window } = reading else {
+            // A gap is only worth recording when it changes the story.
+            return self.last[index].take().is_some();
+        };
+        let unchanged = self.last[index].is_some_and(|last| {
+            last.permille == permille
+                && last.window == window
+                && observed_at_ms.saturating_sub(last.at_ms) < MIN_REEMIT_INTERVAL_MS
+        });
+        if unchanged {
+            return false;
+        }
+        self.last[index] = Some(LastRecorded {
+            permille,
+            window,
+            at_ms: observed_at_ms,
+        });
+        true
+    }
+
+    /// An explicit gap for every profile, stamped with when the gap was
+    /// observed.
+    ///
+    /// Stamping these from the last successful reading — or from the epoch when
+    /// there has never been one — would put them behind the ledger's newest
+    /// entry, and an append-only stream rejects a backwards timestamp. The gap
+    /// would then be silently dropped and a stale percentage would keep
+    /// rendering as current, which is the exact failure decision 0012 exists to
+    /// prevent.
+    fn unavailable_for_all(&self, observed_at_ms: u64) -> Vec<UsageObservation> {
+        (0..self.profiles.len())
+            .map(|index| UsageObservation::unavailable(self.profile_id(index), observed_at_ms))
+            .collect()
+    }
+
+    fn forget_all(&mut self) {
+        self.last = [None, None];
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{CodexSession, SessionStep, WindowSlot};
+    use crate::jsonrpc::ServerRequestClass;
+    use crate::probe::ProbeHealth;
+    use lumbridge_core::{UsageProvenance, UsageUnit};
+
+    const HOUR_MS: u64 = 3_600_000;
+    const RESETS_AT_SECONDS: i64 = 1_730_947_200;
+    const RESETS_AT_MS: u64 = 1_730_947_200_000;
+
+    fn session() -> CodexSession {
+        CodexSession::new().expect("the session profiles are valid")
+    }
+
+    fn ready() -> CodexSession {
+        let mut session = session();
+        let _ = session.initialize("lumbridge", "0.0.1");
+        let step = session.on_line(r#"{"jsonrpc":"2.0","id":1,"result":{}}"#, 0);
+        assert!(matches!(step, SessionStep::Send(_)), "handshake completes");
+        session
+    }
+
+    fn read_response(id: u64, percent: i64, minutes: i64) -> String {
+        format!(
+            r#"{{"jsonrpc":"2.0","id":{id},"result":{{"rateLimits":{{"primary":{{"usedPercent":{percent},"windowDurationMins":{minutes},"resetsAt":{RESETS_AT_SECONDS}}},"secondary":null}}}}}}"#
+        )
+    }
+
+    #[test]
+    fn the_two_windows_are_two_separate_account_free_profiles() {
+        let session = session();
+        let ids = session
+            .profiles()
+            .iter()
+            .map(|profile| profile.id().as_str().to_owned())
+            .collect::<Vec<_>>();
+        assert_eq!(
+            ids,
+            ["codex-app-server-primary", "codex-app-server-secondary"]
+        );
+        assert!(
+            !ids.iter().any(|id| id.contains('@')),
+            "a persisted profile ID must not carry an account identity"
+        );
+    }
+
+    #[test]
+    fn no_reading_is_requested_before_the_handshake_completes() {
+        let mut session = session();
+        let _ = session.initialize("lumbridge", "0.0.1");
+        assert!(
+            session.poll_frame().is_none(),
+            "the app-server requires initialize first"
+        );
+    }
+
+    #[test]
+    fn one_read_is_outstanding_at_a_time() {
+        let mut session = ready();
+        assert!(session.poll_frame().is_some());
+        assert!(
+            session.poll_frame().is_none(),
+            "a second read must wait for the first to answer"
+        );
+    }
+
+    #[test]
+    fn a_full_read_becomes_a_provider_reported_observation() {
+        let mut session = ready();
+        let frame = session.poll_frame().expect("ready sessions poll");
+        assert!(frame.contains("account/rateLimits/read"));
+        let step = session.on_line(&read_response(2, 25, 300), RESETS_AT_MS - HOUR_MS);
+        let SessionStep::Observations(observations) = step else {
+            panic!("a full read produces observations");
+        };
+        assert_eq!(observations.len(), 1, "only primary was present");
+        let observation = &observations[0];
+        assert_eq!(observation.provenance(), UsageProvenance::ProviderReported);
+        assert_eq!(observation.unit(), Some(UsageUnit::WindowPermille));
+        assert_eq!(observation.consumed(), 250);
+    }
+
+    #[test]
+    fn an_unchanged_reading_is_not_re_recorded_every_poll() {
+        let mut session = ready();
+        let _ = session.poll_frame();
+        let at = RESETS_AT_MS - HOUR_MS;
+        assert!(matches!(
+            session.on_line(&read_response(2, 25, 300), at),
+            SessionStep::Observations(_)
+        ));
+        let _ = session.poll_frame();
+        assert!(
+            matches!(
+                session.on_line(&read_response(3, 25, 300), at + 60_000),
+                SessionStep::Idle
+            ),
+            "an identical reading must not flush the bounded retention"
+        );
+        let _ = session.poll_frame();
+        assert!(
+            matches!(
+                session.on_line(&read_response(4, 26, 300), at + 120_000),
+                SessionStep::Observations(_)
+            ),
+            "a changed reading is always recorded"
+        );
+    }
+
+    #[test]
+    fn an_unchanged_reading_is_re_emitted_after_the_floor_so_a_rate_stays_derivable() {
+        let mut session = ready();
+        let _ = session.poll_frame();
+        let at = RESETS_AT_MS - 4 * HOUR_MS;
+        let _ = session.on_line(&read_response(2, 25, 300), at);
+        let _ = session.poll_frame();
+        assert!(matches!(
+            session.on_line(&read_response(3, 25, 300), at + 300_001),
+            SessionStep::Observations(_)
+        ));
+    }
+
+    #[test]
+    fn a_sparse_push_without_a_window_records_nothing() {
+        let mut session = ready();
+        let _ = session.poll_frame();
+        let at = RESETS_AT_MS - HOUR_MS;
+        let _ = session.on_line(&read_response(2, 25, 300), at);
+        // Upstream: nullable metadata absent from a rolling update "does not
+        // clear a previously observed value".
+        let push =
+            r#"{"jsonrpc":"2.0","method":"account/rateLimits/updated","params":{"rateLimits":{}}}"#;
+        assert!(
+            matches!(session.on_line(push, at + 1_000), SessionStep::Idle),
+            "a sparse push must neither clear nor merge"
+        );
+    }
+
+    #[test]
+    fn a_full_read_that_drops_a_window_records_an_explicit_gap() {
+        let mut session = ready();
+        let _ = session.poll_frame();
+        let at = RESETS_AT_MS - HOUR_MS;
+        let _ = session.on_line(&read_response(2, 25, 300), at);
+        let _ = session.poll_frame();
+        let empty =
+            r#"{"jsonrpc":"2.0","id":3,"result":{"rateLimits":{"primary":null,"secondary":null}}}"#;
+        let SessionStep::Observations(observations) = session.on_line(empty, at + 1_000) else {
+            panic!("a full read that loses a window is a change worth recording");
+        };
+        assert_eq!(observations.len(), 1);
+        assert_eq!(observations[0].provenance(), UsageProvenance::Unavailable);
+    }
+
+    #[test]
+    fn a_denial_stamps_its_gap_with_now_so_the_ledger_accepts_it() {
+        use lumbridge_core::UsageLedger;
+
+        let mut session = ready();
+        let _ = session.poll_frame();
+        let at = RESETS_AT_MS - HOUR_MS;
+        let SessionStep::Observations(good) = session.on_line(&read_response(2, 25, 300), at)
+        else {
+            panic!("the first read succeeds");
+        };
+        let _ = session.poll_frame();
+        let denied = r#"{"jsonrpc":"2.0","id":3,"error":{"code":-32600,"message":"chatgpt authentication required to read rate limits"}}"#;
+        let SessionStep::Observations(gaps) = session.on_line(denied, at + 60_000) else {
+            panic!("a denial is a fact about the account");
+        };
+
+        // The ledger is append-only, so a gap stamped behind the reading it
+        // supersedes would be refused and the stale reading would survive.
+        let mut ledger = UsageLedger::new();
+        for observation in good.into_iter().chain(gaps) {
+            ledger
+                .record(observation)
+                .expect("every observation must be accepted in order");
+        }
+        let projection = ledger.project(session.profiles()[0].id(), at + 60_000);
+        assert!(
+            !projection.is_available(),
+            "the denial must supersede the earlier reading"
+        );
+    }
+
+    #[test]
+    fn an_api_key_account_reports_unavailable_rather_than_a_fault() {
+        let mut session = ready();
+        let _ = session.poll_frame();
+        let denied = r#"{"jsonrpc":"2.0","id":2,"error":{"code":-32600,"message":"chatgpt authentication required to read rate limits"}}"#;
+        let SessionStep::Observations(observations) = session.on_line(denied, 5_000) else {
+            panic!("a quota-less account is a fact about the account");
+        };
+        assert_eq!(observations.len(), 2);
+        assert!(
+            observations
+                .iter()
+                .all(|observation| observation.provenance() == UsageProvenance::Unavailable)
+        );
+    }
+
+    #[test]
+    fn any_other_rejection_faults_the_probe() {
+        let mut session = ready();
+        let _ = session.poll_frame();
+        let denied = r#"{"jsonrpc":"2.0","id":2,"error":{"code":-32603,"message":"boom"}}"#;
+        assert!(matches!(
+            session.on_line(denied, 5_000),
+            SessionStep::Health(ProbeHealth::Faulted(_))
+        ));
+    }
+
+    #[test]
+    fn a_credential_refresh_request_is_refused_and_counted() {
+        let mut session = ready();
+        let step = session.on_line(
+            r#"{"jsonrpc":"2.0","id":99,"method":"account/chatgptAuthTokens/refresh"}"#,
+            0,
+        );
+        let SessionStep::Refused { line, class } = step else {
+            panic!("a credential request must be refused");
+        };
+        assert_eq!(class, ServerRequestClass::CredentialRefresh);
+        assert!(line.contains("-32601"));
+        assert!(
+            !line.contains("access"),
+            "a refusal must not echo the request"
+        );
+        assert_eq!(session.refused_credential_requests(), 1);
+    }
+
+    #[test]
+    fn a_malformed_line_faults_rather_than_being_guessed_at() {
+        let mut session = ready();
+        assert!(matches!(
+            session.on_line("this is not json", 0),
+            SessionStep::Health(ProbeHealth::Faulted(_))
+        ));
+    }
+
+    #[test]
+    fn window_slots_have_distinct_static_identifiers() {
+        assert_ne!(
+            WindowSlot::Primary.profile_id(),
+            WindowSlot::Secondary.profile_id()
+        );
+    }
+}

crates/lumbridge-harness/src/jsonrpc.rs

@@ -0,0 +1,278 @@
+//! A deliberately deaf JSON-RPC client.
+//!
+//! The Codex app-server can send requests *to* its client, including
+//! `account/chatgptAuthTokens/refresh`, which returns a bare access token.
+//! AGENTS.md forbids Lumbridge from scraping a harness's private credentials.
+//! Choosing not to call that method would be a convention; this module makes
+//! it unrepresentable instead.
+//!
+//! There is no free-form method string anywhere in this crate. Outbound
+//! requests come from a two-variant enum, and every inbound server request is
+//! answered with `-32601 method not found` regardless of what it asks for.
+//! Classification exists only so a refusal can be counted and named, which
+//! makes "we were asked for a credential and refused" an auditable event
+//! rather than an absence.
+
+use serde_json::{Value, json};
+
+/// JSON-RPC's "method not found". The only reply this client ever sends.
+const METHOD_NOT_FOUND: i64 = -32601;
+/// Codex returns this for a request that needs an account it does not have.
+pub(crate) const INVALID_REQUEST: i64 = -32600;
+
+/// Every request this crate is able to send.
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub(crate) enum OutboundRequest {
+    Initialize,
+    ReadAccountRateLimits,
+}
+
+impl OutboundRequest {
+    pub(crate) const fn method(self) -> &'static str {
+        match self {
+            Self::Initialize => "initialize",
+            Self::ReadAccountRateLimits => "account/rateLimits/read",
+        }
+    }
+}
+
+/// The notifications this client accepts. Anything else is ignored.
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub(crate) enum AcceptedNotification {
+    AccountRateLimitsUpdated,
+}
+
+impl AcceptedNotification {
+    pub(crate) fn from_method(method: &str) -> Option<Self> {
+        match method {
+            "account/rateLimits/updated" => Some(Self::AccountRateLimitsUpdated),
+            _ => None,
+        }
+    }
+}
+
+/// What a server-to-client request was asking for.
+///
+/// Purely descriptive. Every class is refused identically; the distinction
+/// exists so a probe can report that it declined a credential request.
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum ServerRequestClass {
+    /// A request that would hand Lumbridge a credential. Always refused.
+    CredentialRefresh,
+    /// A request to run something on the harness's behalf.
+    Execution,
+    /// A request for a human approval decision.
+    Approval,
+    /// A request for ambient information such as the current time.
+    Ambient,
+    /// Anything this client does not recognise.
+    Unrecognised,
+}
+
+impl ServerRequestClass {
+    pub(crate) fn classify(method: &str) -> Self {
+        match method {
+            "account/chatgptAuthTokens/refresh" => Self::CredentialRefresh,
+            "item/tool/call" | "attestation/generate" => Self::Execution,
+            "item/permissions/requestApproval" => Self::Approval,
+            "currentTime/read" => Self::Ambient,
+            _ => Self::Unrecognised,
+        }
+    }
+
+    #[must_use]
+    pub const fn label(self) -> &'static str {
+        match self {
+            Self::CredentialRefresh => "credential refresh",
+            Self::Execution => "execution",
+            Self::Approval => "approval",
+            Self::Ambient => "ambient information",
+            Self::Unrecognised => "unrecognised request",
+        }
+    }
+}
+
+/// A classified inbound line.
+#[derive(Clone, Debug)]
+pub(crate) enum Inbound {
+    /// A successful response to one of our two requests.
+    Result { id: u64, result: Value },
+    /// A failed response to one of our two requests.
+    Failure { id: u64, code: i64 },
+    /// An accepted push notification.
+    Notification {
+        kind: AcceptedNotification,
+        params: Value,
+    },
+    /// The server asked us for something. It will be refused.
+    ServerRequest {
+        id: Value,
+        class: ServerRequestClass,
+    },
+    /// A well-formed frame with nothing for us in it.
+    Ignored,
+}
+
+/// Encodes one of the two permitted requests.
+pub(crate) fn encode_request(id: u64, request: OutboundRequest, params: &Value) -> String {
+    json!({
+        "jsonrpc": "2.0",
+        "id": id,
+        "method": request.method(),
+        "params": params,
+    })
+    .to_string()
+}
+
+pub(crate) fn encode_notification(method: &'static str, params: &Value) -> String {
+    json!({ "jsonrpc": "2.0", "method": method, "params": params }).to_string()
+}
+
+/// The refusal sent for every server-to-client request.
+pub(crate) fn encode_refusal(id: &Value) -> String {
+    json!({
+        "jsonrpc": "2.0",
+        "id": id,
+        "error": { "code": METHOD_NOT_FOUND, "message": "method not found" },
+    })
+    .to_string()
+}
+
+/// Classifies one line of JSON-RPC without interpreting its payload.
+///
+/// Returns `None` when the line is not a JSON object at all, which the caller
+/// reports as [`crate::HarnessError::Malformed`].
+pub(crate) fn classify(line: &str) -> Option<Inbound> {
+    let frame: Value = serde_json::from_str(line).ok()?;
+    let object = frame.as_object()?;
+    let method = object.get("method").and_then(Value::as_str);
+    let id = object.get("id");
+
+    match (method, id) {
+        // A method with an id is the server asking us for something.
+        (Some(method), Some(id)) => Some(Inbound::ServerRequest {
+            id: id.clone(),
+            class: ServerRequestClass::classify(method),
+        }),
+        // A method without an id is a notification.
+        (Some(method), None) => Some(AcceptedNotification::from_method(method).map_or(
+            Inbound::Ignored,
+            |kind| Inbound::Notification {
+                kind,
+                params: object.get("params").cloned().unwrap_or(Value::Null),
+            },
+        )),
+        // An id without a method is a response to one of ours.
+        (None, Some(id)) => {
+            let id = id.as_u64()?;
+            if let Some(error) = object.get("error") {
+                let code = error.get("code").and_then(Value::as_i64).unwrap_or(0);
+                return Some(Inbound::Failure { id, code });
+            }
+            Some(Inbound::Result {
+                id,
+                result: object.get("result").cloned().unwrap_or(Value::Null),
+            })
+        }
+        (None, None) => Some(Inbound::Ignored),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{
+        AcceptedNotification, Inbound, OutboundRequest, ServerRequestClass, classify,
+        encode_refusal, encode_request,
+    };
+    use serde_json::{Value, json};
+
+    #[test]
+    fn only_two_methods_can_be_sent() {
+        assert_eq!(OutboundRequest::Initialize.method(), "initialize");
+        assert_eq!(
+            OutboundRequest::ReadAccountRateLimits.method(),
+            "account/rateLimits/read"
+        );
+    }
+
+    #[test]
+    fn a_credential_refresh_request_is_classified_and_refused() {
+        let line = r#"{"jsonrpc":"2.0","id":9,"method":"account/chatgptAuthTokens/refresh"}"#;
+        let Some(Inbound::ServerRequest { id, class }) = classify(line) else {
+            panic!("a method with an id is a server request");
+        };
+        assert_eq!(class, ServerRequestClass::CredentialRefresh);
+        let refusal: Value =
+            serde_json::from_str(&encode_refusal(&id)).expect("the refusal is valid JSON");
+        assert_eq!(refusal["error"]["code"], json!(-32601));
+        assert!(
+            refusal.get("result").is_none(),
+            "a refusal must never carry a result"
+        );
+    }
+
+    #[test]
+    fn every_server_request_class_is_refused_the_same_way() {
+        for method in [
+            "account/chatgptAuthTokens/refresh",
+            "item/tool/call",
+            "item/permissions/requestApproval",
+            "currentTime/read",
+            "something/entirely/new",
+        ] {
+            let line = format!(r#"{{"jsonrpc":"2.0","id":1,"method":"{method}"}}"#);
+            let Some(Inbound::ServerRequest { id, .. }) = classify(&line) else {
+                panic!("{method} must classify as a server request");
+            };
+            let refusal: Value = serde_json::from_str(&encode_refusal(&id)).expect("valid JSON");
+            assert_eq!(refusal["error"]["code"], json!(-32601));
+        }
+    }
+
+    #[test]
+    fn only_the_rate_limit_notification_is_accepted() {
+        let accepted = classify(
+            r#"{"jsonrpc":"2.0","method":"account/rateLimits/updated","params":{"rateLimits":{}}}"#,
+        );
+        assert!(matches!(
+            accepted,
+            Some(Inbound::Notification {
+                kind: AcceptedNotification::AccountRateLimitsUpdated,
+                ..
+            })
+        ));
+        let ignored =
+            classify(r#"{"jsonrpc":"2.0","method":"thread/tokenUsage/updated","params":{}}"#);
+        assert!(matches!(ignored, Some(Inbound::Ignored)));
+    }
+
+    #[test]
+    fn responses_and_failures_are_separated() {
+        assert!(matches!(
+            classify(r#"{"jsonrpc":"2.0","id":7,"result":{"rateLimits":{}}}"#),
+            Some(Inbound::Result { id: 7, .. })
+        ));
+        assert!(matches!(
+            classify(r#"{"jsonrpc":"2.0","id":7,"error":{"code":-32600,"message":"nope"}}"#),
+            Some(Inbound::Failure {
+                id: 7,
+                code: -32600
+            })
+        ));
+    }
+
+    #[test]
+    fn malformed_and_non_object_lines_are_rejected() {
+        for line in ["", "not json", "[1,2,3]", "\"a string\"", "{"] {
+            assert!(classify(line).is_none(), "{line:?} must not classify");
+        }
+    }
+
+    #[test]
+    fn a_request_encodes_without_a_free_form_method() {
+        let encoded = encode_request(1, OutboundRequest::ReadAccountRateLimits, &json!({}));
+        let frame: Value = serde_json::from_str(&encoded).expect("valid JSON");
+        assert_eq!(frame["method"], json!("account/rateLimits/read"));
+        assert_eq!(frame["jsonrpc"], json!("2.0"));
+    }
+}

crates/lumbridge-harness/src/lib.rs

@@ -0,0 +1,68 @@
+//! Documented status and usage probes for supervised coding harnesses.
+//!
+//! `lumbridge-core` owns the usage ledger and stays free of IO, clocks, and
+//! async. This crate is the other side of that boundary: it runs processes,
+//! reads a clock, parses untrusted wire text, and hands back
+//! [`lumbridge_core::UsageObservation`] values. Nothing here interprets a
+//! usage number; it only decides which observation is honest to record.
+//!
+//! The rule every probe follows: Lumbridge observes documented surfaces of a
+//! harness it launched, and never reads that harness's credentials. A probe
+//! that cannot obtain a reading records
+//! [`lumbridge_core::UsageObservation::unavailable`] rather than a zero.
+
+#![forbid(unsafe_code)]
+
+mod child;
+pub mod claude;
+mod clock;
+pub mod codex;
+mod jsonrpc;
+mod probe;
+
+pub use child::PROBE_ENV_ALLOWLIST;
+pub use clock::MonotonicWallClock;
+pub use jsonrpc::ServerRequestClass;
+pub use probe::{ProbeHealth, ProbeOutcome, UsageProbe};
+
+use thiserror::Error;
+
+/// Failures a probe can report.
+///
+/// Every variant is [`Copy`] and carries only a static discriminator, an
+/// [`std::io::ErrorKind`], or a number. No variant can capture a string from a
+/// child process or the filesystem, so an error can never smuggle a path,
+/// a credential, or attacker-chosen text into a log or a UI surface.
+#[derive(Clone, Copy, Debug, Eq, Error, PartialEq)]
+pub enum HarnessError {
+    #[error("a probe program name must not be empty")]
+    EmptyProgram,
+    #[error("the environment variable {0} is not on the probe allowlist")]
+    EnvNotAllowlisted(&'static str),
+    #[error("a probe poll interval must be at least one second")]
+    ProbeIntervalTooShort,
+    #[error("the probe program could not be started ({0:?})")]
+    SpawnFailed(std::io::ErrorKind),
+    #[error("the probe program closed its output")]
+    Eof,
+    #[error("the probe program sent a {0} byte line, over the accepted limit")]
+    OversizedLine(usize),
+    #[error("the probe program sent a line that is not a JSON-RPC frame")]
+    Malformed,
+    #[error("the harness rejected the request with JSON-RPC code {0}")]
+    Rejected(i64),
+    #[error("the harness has no subscription account to report a quota for")]
+    NotAuthenticated,
+}
+
+impl HarnessError {
+    /// Whether this failure means "there is no number", as opposed to
+    /// "the probe broke".
+    ///
+    /// Both record an unavailable observation, but only the second is a fault
+    /// worth surfacing as a broken probe.
+    #[must_use]
+    pub const fn is_expected_gap(self) -> bool {
+        matches!(self, Self::NotAuthenticated)
+    }
+}

crates/lumbridge-harness/src/probe.rs

@@ -0,0 +1,106 @@
+//! The contract between a harness probe and the usage ledger.
+//!
+//! This trait lives here rather than in `lumbridge-core` because implementing
+//! it implies a fallible external call, and core stays IO-free, clock-free,
+//! and async-free. [`lumbridge_core::UsageObservation`] is already the shared
+//! type; nothing further needs to cross the boundary.
+
+use lumbridge_core::{AccountProfile, UsageObservation};
+
+use crate::HarnessError;
+
+/// What a probe knows about its own liveness.
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum ProbeHealth {
+    /// Launched, handshake not finished.
+    Starting,
+    /// Handshake complete and readings are expected.
+    Ready,
+    /// Running, but this account has no subscription quota to report. This is
+    /// a correct terminal state, not a fault: API-key users have no window.
+    NoQuotaAccount,
+    /// Broken. The observations it produced before this are still facts.
+    Faulted(HarnessError),
+    /// Deliberately shut down.
+    Stopped,
+}
+
+impl ProbeHealth {
+    #[must_use]
+    pub const fn is_faulted(self) -> bool {
+        matches!(self, Self::Faulted(_))
+    }
+
+    /// A short phrase a UI may render beside the probe's profiles.
+    #[must_use]
+    pub const fn label(self) -> &'static str {
+        match self {
+            Self::Starting => "probe starting",
+            Self::Ready => "probe ready",
+            Self::NoQuotaAccount => "no subscription quota on this account",
+            Self::Faulted(_) => "probe faulted",
+            Self::Stopped => "probe stopped",
+        }
+    }
+}
+
+/// One round of probe output.
+///
+/// It carries observations and liveness and nothing else. There is no variant
+/// able to carry a command, a plan, or a capability request, so a probe cannot
+/// become an actor on the workspace no matter what the harness sends it.
+#[derive(Clone, Debug)]
+pub struct ProbeOutcome {
+    observations: Vec<UsageObservation>,
+    health: ProbeHealth,
+}
+
+impl ProbeOutcome {
+    #[must_use]
+    pub const fn new(observations: Vec<UsageObservation>, health: ProbeHealth) -> Self {
+        Self {
+            observations,
+            health,
+        }
+    }
+
+    #[must_use]
+    pub const fn idle(health: ProbeHealth) -> Self {
+        Self {
+            observations: Vec::new(),
+            health,
+        }
+    }
+
+    #[must_use]
+    pub fn observations(&self) -> &[UsageObservation] {
+        &self.observations
+    }
+
+    /// Takes the observations out for recording, leaving the outcome empty.
+    #[must_use]
+    pub fn into_observations(self) -> Vec<UsageObservation> {
+        self.observations
+    }
+
+    #[must_use]
+    pub const fn health(&self) -> ProbeHealth {
+        self.health
+    }
+}
+
+/// A source of usage observations for one or more account profiles.
+pub trait UsageProbe {
+    /// Drains whatever the probe has ready.
+    ///
+    /// Implementations must return promptly: this is called from the UI's
+    /// frame loop and must never wait on a child process, a socket, or a lock
+    /// held across IO.
+    fn poll(&mut self) -> ProbeOutcome;
+
+    /// The profiles this probe can produce observations for.
+    fn profiles(&self) -> &[AccountProfile];
+
+    /// Stops the probe and releases its child process.
+    fn shutdown(&mut self);
+}

crates/lumbridge-harness/tests/codex_ledger_path.rs

@@ -0,0 +1,175 @@
+//! End-to-end: a Codex transcript through the real ledger and footer.
+//!
+//! The unit tests prove the adapter builds the right observations. These prove
+//! the observations still obey every decision 0012 rule once they reach
+//! `lumbridge-core`, which is the property that actually matters to a user
+//! reading the footer.
+//!
+//! Every transcript below is hand-written from the documented wire shape in
+//! openai/codex (Apache-2.0), commit
+//! `17e8101699c5062117d0d37f504313e8af53b043`, `codex-rs/app-server/README.md`
+//! section "7) Rate limits (`ChatGPT`)". No real account, transcript, or
+//! credential is used, and all timestamps are small synthetic integers.
+
+use lumbridge_core::{
+    AccountProfileId, FooterUsage, UsageConfidence, UsageLedger, UsageProvenance,
+};
+use lumbridge_harness::codex::replay_transcript;
+
+const HOUR_MS: u64 = 3_600_000;
+/// A synthetic reset five hours after the epoch, in seconds.
+const RESETS_AT_SECONDS: u64 = 5 * 3_600;
+const RESETS_AT_MS: u64 = 5 * HOUR_MS;
+
+fn initialize_response() -> &'static str {
+    r#"{"jsonrpc":"2.0","id":1,"result":{"userAgent":"codex"}}"#
+}
+
+fn read_response(id: u64, percent: u64) -> String {
+    format!(
+        r#"{{"jsonrpc":"2.0","id":{id},"result":{{"rateLimits":{{"primary":{{"usedPercent":{percent},"windowDurationMins":300,"resetsAt":{RESETS_AT_SECONDS}}},"secondary":null,"rateLimitReachedType":null}}}}}}"#
+    )
+}
+
+fn primary() -> AccountProfileId {
+    AccountProfileId::new("codex-app-server-primary").expect("the probe's static ID is valid")
+}
+
+fn ledger_from(frames: &[(&str, u64)]) -> UsageLedger {
+    let mut ledger = UsageLedger::new();
+    for observation in replay_transcript(frames) {
+        ledger
+            .record(observation)
+            .expect("the adapter emits observations in time order");
+    }
+    ledger
+}
+
+#[test]
+fn a_two_reading_transcript_produces_a_provider_reported_footer() {
+    let first = read_response(2, 20);
+    let second = read_response(3, 35);
+    let ledger = ledger_from(&[
+        (initialize_response(), 0),
+        (first.as_str(), HOUR_MS),
+        (second.as_str(), 2 * HOUR_MS),
+    ]);
+
+    let projection = ledger.project(&primary(), 2 * HOUR_MS);
+    assert_eq!(projection.provenance(), UsageProvenance::ProviderReported);
+    assert_eq!(projection.confidence(), UsageConfidence::Exact);
+    assert_eq!(projection.consumed_permille(), Some(350));
+    assert_eq!(projection.resets_in_ms(), Some(3 * HOUR_MS));
+    // 200 permille burned over one hour, derived rather than reported.
+    assert_eq!(projection.burn_per_hour(), Some(150));
+    assert_eq!(projection.burn_provenance(), UsageProvenance::Estimated);
+}
+
+#[test]
+fn a_single_reading_yields_a_percentage_but_no_rate() {
+    let only = read_response(2, 20);
+    let ledger = ledger_from(&[(initialize_response(), 0), (only.as_str(), HOUR_MS)]);
+
+    let projection = ledger.project(&primary(), HOUR_MS);
+    assert_eq!(projection.consumed_permille(), Some(200));
+    assert_eq!(
+        projection.burn_per_hour(),
+        None,
+        "one provider reading is not a rate"
+    );
+    assert_eq!(projection.exhaustion_in_ms(), None);
+}
+
+#[test]
+fn an_api_key_account_renders_the_unavailable_path_rather_than_a_zero() {
+    // Codex answers -32600 with "chatgpt authentication required to read rate
+    // limits" for an account that has no subscription quota.
+    let denied = r#"{"jsonrpc":"2.0","id":2,"error":{"code":-32600,"message":"chatgpt authentication required to read rate limits"}}"#;
+    let ledger = ledger_from(&[(initialize_response(), 0), (denied, HOUR_MS)]);
+
+    let projection = ledger.project(&primary(), HOUR_MS);
+    assert!(!projection.is_available());
+    assert_eq!(projection.consumed_permille(), None);
+
+    let profile = lumbridge_core::AccountProfile::new(
+        "codex-app-server-primary",
+        "Codex",
+        "ChatGPT",
+        "primary window",
+        "subscription",
+    )
+    .expect("a valid profile");
+    let footer = FooterUsage::new(&profile, &projection);
+    assert_eq!(footer.window(), "usage unavailable");
+    assert_eq!(footer.burn(), "burn rate unavailable");
+    assert_eq!(footer.trust(), "no usage source");
+    assert!(
+        !footer.to_string().contains('%'),
+        "an account with no quota must not render a percentage"
+    );
+}
+
+#[test]
+fn a_reading_taken_after_the_reset_never_renders_a_frozen_percentage() {
+    let stale = read_response(2, 90);
+    let ledger = ledger_from(&[
+        (initialize_response(), 0),
+        // Observed one millisecond after the window it describes ended.
+        (stale.as_str(), RESETS_AT_MS + 1),
+    ]);
+
+    let projection = ledger.project(&primary(), RESETS_AT_MS + 1);
+    assert!(!projection.is_available());
+    assert_eq!(
+        projection.consumed_permille(),
+        None,
+        "an ended window must not keep rendering its last percentage"
+    );
+}
+
+#[test]
+fn a_credential_refresh_request_produces_no_observation_at_all() {
+    let refresh = r#"{"jsonrpc":"2.0","id":77,"method":"account/chatgptAuthTokens/refresh"}"#;
+    let reading = read_response(2, 20);
+    let observations = replay_transcript(&[
+        (initialize_response(), 0),
+        (refresh, 1_000),
+        (reading.as_str(), HOUR_MS),
+    ]);
+    assert_eq!(
+        observations.len(),
+        1,
+        "the refused credential request must not enter the usage stream"
+    );
+}
+
+#[test]
+fn a_repeated_identical_reading_does_not_flood_the_ledger() {
+    let a = read_response(2, 20);
+    let b = read_response(3, 20);
+    let c = read_response(4, 20);
+    let ledger = ledger_from(&[
+        (initialize_response(), 0),
+        (a.as_str(), HOUR_MS),
+        (b.as_str(), HOUR_MS + 60_000),
+        (c.as_str(), HOUR_MS + 120_000),
+    ]);
+    assert_eq!(
+        ledger.observation_count(&primary()),
+        1,
+        "identical readings inside the re-emit floor must not evict history"
+    );
+}
+
+#[test]
+fn a_malformed_transcript_records_nothing_rather_than_guessing() {
+    let observations = replay_transcript(&[
+        (initialize_response(), 0),
+        ("}{ not json", 1_000),
+        (
+            r#"{"jsonrpc":"2.0","id":2,"result":{"rateLimits":"a string"}}"#,
+            2_000,
+        ),
+    ]);
+    assert!(observations.is_empty());
+}

docs/ARCHITECTURE.md

@@ -28,7 +28,8 @@ messages should be real from the beginning.
 - `lumbridge-pty`: portable PTY and process-tree adapters.
 - `lumbridge-terminal`: VT parsing, scrollback, selection, search, render model.
 - `lumbridge-acp`: ACP client, capability negotiation, transcript normalization.
-- `lumbridge-harness`: manifests, launch profiles, hooks, PTY fallback adapters.
+- `lumbridge-harness`: manifests, launch profiles, hooks, PTY fallback adapters,
+  and documented status/usage probes.
 - `lumbridge-provider`: BYOK providers and provider-neutral usage records.
 - `lumbridge-storage`: SQLite migrations, event log, snapshots, retention.
 - `lumbridge-remote`: OpenSSH/Tailscale command construction, framed stdio,
@@ -41,10 +42,10 @@ messages should be real from the beginning.
 - `lumbridge`: installable application entry point.
 
 The scaffold currently contains `lumbridge-core`, `lumbridge-storage`,
-`lumbridge-buzz`, `lumbridge-pty`, `lumbridge-runtime`, `lumbridge-terminal`, and
+`lumbridge-buzz`, `lumbridge-pty`, `lumbridge-runtime`, `lumbridge-terminal`,
-the entry point. `lumbridge-core` also contains the first typed workspace
+`lumbridge-harness`, and the entry point. `lumbridge-core` also contains the
-command reducer. Larger runtime crates are added after their architecture spikes
+first typed workspace command reducer and the usage ledger. Larger runtime
-pass.
+crates are added after their architecture spikes pass.
 
 ## Terminal path
 
@@ -200,6 +201,52 @@ Observations include account profile, provider, harness, model, units, time
 window, reset time, provenance, confidence, and source timestamp. Projections are
 derived views that can be recomputed as forecasting improves.
 
+`lumbridge-core` now implements that stream. A bounded per-profile `UsageLedger`
+accepts observations in time order, and `project` derives the footer view:
+window fraction, reset, burn rate, and exhaustion. Facts keep their reported
+provenance; every derived value is labelled estimated. The projection withholds
+a burn rate from a single sample, withholds a window fraction without a reported
+ceiling, withholds an exhaustion estimate that falls after the reset, treats an
+unavailable observation as invalidating older facts, and reports an expired
+window as rolled over rather than freezing its last percentage. `FooterUsage`
+renders the five product-spec questions and returns explicit unavailable phrases
+instead of placeholder numbers. See decision 0012.
+
+`lumbridge-harness` is the other side of that boundary: it runs processes,
+reads a clock, and parses untrusted wire text, then hands back observations.
+Its first adapter probes Codex's documented `account/rateLimits/read` surface
+over the app-server's JSON-RPC stdio protocol. The client cannot express a
+request outside a two-variant enum and answers every server-to-client request
+with `-32601`, so a harness asking Lumbridge for a credential is refused by
+construction. All protocol decisions live in a pure state machine driven by
+`&str` lines and an explicit timestamp, with a `replay_transcript` entry point
+that verifies the whole path against a fixture without a process, a clock, or a
+network. See decision 0013.
+
+Its second adapter reads Claude Code's session transcripts, the only local
+surface that reports what that harness spent. It is a tail follower with a
+per-file byte offset, so its total is monotonic, and it reports nothing until
+it has read the existing backlog to the end — a partially-read backlog is
+indistinguishable from a burst of spend, and the ledger would derive a rate
+from it. The transcript carries no window or ceiling, so that profile reports
+consumption against none. The parser models four token counters and nothing
+else, so the conversations in those files are not representable in a Lumbridge
+value. See decision 0014.
+
+Claude Code's subscription windows come from a different surface: the CLI pipes
+a `rate_limits` object carrying the five-hour and seven-day windows to whatever
+`statusLine` command the user has configured, on every turn. A small installed
+bridge writes those fields — and only those — to a local feed the probe tails,
+so the windows are `ProviderReported` and no credential is ever read. Reading
+Claude Code's OAuth token would be more capable and is what comparable tools
+do; Lumbridge does not, because AGENTS.md forbids it. See decision 0015.
+
+The GPUI shell now runs both probes; a probe contributes its own profiles on
+top of the declared ones, so the Claude windows appear once they report. Its
+footer strip
+shows every declared harness at once, and a harness with no adapter renders
+the honest gap.
+
 Subscription balance is provider-specific and sometimes unavailable. BYOK calls
 usually expose token counts but cost still depends on cached tokens, reasoning,
 tool calls, and current pricing. Adapters normalize facts without erasing their

docs/UI_SPIKE_SCORECARD.md

@@ -97,7 +97,13 @@ Terminal, Browser, Markdown, and Review panels beside the selection or
 reattaches an existing session. Every Terminal panel receives an independent
 actor-owned VT session; the three seed non-terminal surfaces keep deterministic
 updates running. Counters
-separate external PTY batches/lines from total model updates. The GPUI footer
+separate external PTY batches/lines from total model updates. The GPUI footer's
+usage zone is no longer a constant string: it renders a `lumbridge-core`
+projection for the selected panel's account profile, with a provenance chip and
+explicit unavailable phrases, and its sidebar attention count and runtime rows
+are derived from live panel and ledger state. The Floem shell still renders the
+static footer fixtures, so the two candidates are no longer comparable on that
+strip. The GPUI footer also
 reports dispatch-to-element-build p50/p95 over a bounded 256-sample window. It
 is deliberately not called key-to-present or frame-present latency: neither
 candidate exposes a reliable public cross-platform post-present callback. True

docs/decisions/0012-usage-ledger-and-footer-provenance.md

@@ -0,0 +1,67 @@
+# 0012: The footer renders a derived projection, never a stored percentage
+
+Status: accepted. Supplemented by decisions 0013 and 0014, which supply the
+first two real adapters. Every rule below still holds.
+
+Decision 0002 established that usage observations carry provenance. This
+decision fixes how those observations become the five answers the footer owes
+the user, and what the footer must do when it cannot answer.
+
+## The contract
+
+`lumbridge-core` owns an append-only `UsageLedger` of `UsageObservation`s per
+`AccountProfile`. An observation is a fact: consumed units, an optional ceiling,
+an optional provider window, a provenance, and the time it was observed. The
+ledger never edits an observation and refuses one that predates the newest entry
+for that profile, because an append-only stream cannot move backwards.
+
+`UsageLedger::project` derives the footer view. Reported facts keep their own
+provenance. Anything the projection computes is labelled `Estimated` even when
+every input was provider-reported, because a derived rate is not a reported
+fact.
+
+## Rules the projection enforces
+
+- A burn rate needs at least two observations inside the same window, the same
+  unit, and a minimum elapsed sample. One reading yields no rate at all rather
+  than a rate of zero.
+- A window fraction requires a reported ceiling. A self-hosted endpoint has no
+  quota, so it reports consumption and explicitly says no ceiling exists.
+- An exhaustion estimate is withheld when the window resets first. Telling a
+  user they will run out after the quota has already refilled is worse than
+  telling them nothing.
+- An `Unavailable` observation invalidates older facts for that profile instead
+  of letting a stale reading keep rendering as current.
+- An expired window is reported as rolled over, not as a frozen percentage from
+  a window that ended.
+- Staleness is surfaced, not hidden: past the staleness budget the value still
+  renders but its confidence drops to unknown and the footer says so.
+
+## What the UI may do
+
+A compact surface may substitute the short provenance label. It may not drop the
+label, round an estimate into a reported fact, or substitute a placeholder
+number for a missing one. `FooterUsage` therefore returns explicit phrases —
+`usage unavailable`, `reset time unavailable`, `burn rate unavailable`,
+`no usage source` — and the GPUI shell renders unavailable values in the quiet
+tone with the provenance chip coloured by source rather than by value.
+
+A pane with no harness attached shows the detached roll-up. It never inherits
+another pane's account, because attributing one agent's spend to another is the
+same class of untruth as inventing the number.
+
+## Status of the numbers
+
+The GPUI shell's observations are still a declared spike fixture. The path is
+real: the fixture records through the same ledger and reads back through the
+same projection, so the honest-gap behaviour above is exercised on every frame
+rather than asserted only in tests.
+
+The first real adapter now exists in `lumbridge-harness` (decision 0013) and its
+integration tests drive a synthetic Codex transcript through this ledger and
+this footer, proving every rule above survives the adapter path. Pointing the
+running shell at a live account spawns a process against the user's own
+subscription, so it stays a separate, user-consented step.
+
+All arithmetic is integer arithmetic and fractions are carried as permille, so a
+rendered percentage cannot drift from the stored fact.

docs/decisions/0013-first-usage-adapter-codex-app-server.md

@@ -0,0 +1,116 @@
+# 0013: The first usage adapter is the Codex app-server rate-limit probe
+
+Status: accepted; wire contract verified against one live account.
+
+Decision 0012 built a usage ledger and a footer that refuse to invent numbers,
+but every observation reaching them was a declared fixture. This decision picks
+the first real source and records the rules that choice establishes for the
+adapters after it.
+
+## Why Codex first
+
+Of the seven captured harnesses, Codex is the only one whose quota surface is
+all four of: prose-documented upstream with a worked example
+(`codex-rs/app-server/README.md`, "7) Rate limits (ChatGPT)"), pinned by a
+checked-in JSON Schema, genuinely provider-originated, and readable without
+Lumbridge touching a credential. `account/rateLimits/read` returns
+`usedPercent`, `windowDurationMins`, and `resetsAt`; the app-server parses
+those out of the backend's `x-codex-*-used-percent` response headers.
+
+Lumbridge launches `codex app-server` and lets the harness resolve its own
+authentication from `CODEX_HOME`. That satisfies the product spec's
+"observes only documented status, usage, and protocol surfaces" without any of
+the credential handling AGENTS.md forbids. Upstream is Apache-2.0; only the
+observable wire contract is mirrored, and its source, commit, and license are
+recorded in the parser and the tests.
+
+## The provenance test for every future adapter
+
+Ask whether the harness **computes** the value or **forwards** it.
+
+- Forwards a provider's own number: `ProviderReported`.
+- Computes, counts, or estimates it: `HarnessReported`, or `LocallyMeasured`
+  when Lumbridge did the counting.
+
+Codex forwards, so its windows are provider-reported. The ACP `usage_update`
+notification, by contrast, is a *context window* gauge, and at least one
+captured harness fills it with an explicit local estimate. It must never be fed
+through the same path: a context window is not a quota, and rendering one as
+"62% used" would imply a subscription limit that does not exist.
+
+## Rules this adapter establishes
+
+- **Refusal is structural, not procedural.** The app-server can send its client
+  `account/chatgptAuthTokens/refresh`, which returns a bare access token. The
+  client has no free-form method string: outbound requests come from a
+  two-variant enum, and every server-to-client request is answered `-32601`
+  regardless of what it asks for. Asking for a credential is unrepresentable
+  rather than merely declined, and refusals are counted so the path is
+  auditable.
+- **An account with no quota is a fact, not a fault.** Codex answers `-32600`
+  for an API-key user. That records an unavailable observation for both
+  windows. Any other rejection faults the probe.
+- **Two windows are two profiles.** `primary` and `secondary` get separate,
+  account-free profile identifiers. Merging them would mean silently choosing
+  one. A profile ID is persisted, so it never carries an account identity.
+- **A sparse push neither clears nor merges.** Upstream documents that nullable
+  values absent from a rolling update "do not clear a previously observed
+  value". An absent window on a notification therefore records nothing. It is
+  also not merged into the last full read: a merged composite would present
+  itself as provider-reported while being partly a memory.
+- **Clock disagreement costs the window, not the number.** A window whose
+  computed start is after the observation is dropped while the provider's
+  percentage is kept. A reading taken at or after `resetsAt` is unusable
+  entirely, because 0012 requires an ended window to read as rolled over.
+- **Identical readings are suppressed, with a floor.** The ledger retains a
+  bounded history per profile; re-recording an unchanged percentage every poll
+  would evict real history and destroy the burn baseline. An unchanged reading
+  is re-emitted only after five minutes, which keeps a rate derivable without
+  flooding.
+- **Errors cannot carry text.** `HarnessError` is `Copy` and holds only static
+  discriminators, an `io::ErrorKind`, and numbers. A spawn failure deliberately
+  discards the underlying error so a filesystem path cannot travel into a log,
+  and a rejection keeps the JSON-RPC code while discarding the backend message.
+- **The child is contained.** Cleared environment with a literal allowlist,
+  pipes rather than a PTY, stderr discarded, and a per-line byte cap. The child
+  leads its own process group and the group is what gets killed, because a pipe
+  reports end of file only when every write end closes: a launcher that execs
+  the real program as a grandchild with inherited stdio — which is how the npm
+  distribution of Codex works — would otherwise leave a reader blocked on a pipe
+  nothing will ever write to. Shutdown signals, kills the group, and joins only
+  the worker. The reader is never joined, so a descendant that escapes the group
+  can leak a parked thread but cannot hang the caller, which matters because
+  `Drop` calls shutdown.
+
+## Verified against a live account
+
+One reading was taken with the user's consent on 2026-08-31, against
+`codex-cli 0.145.0` installed as a standalone native binary. It confirmed the
+whole contract: `usedPercent` an integer, `resetsAt` Unix seconds, `secondary`
+null, and no unmodelled fields inside the window object. The result body also
+carried `rateLimitResetCredits`, `rateLimitsByLimitId`, `planType`, `limitId`,
+`limitName`, `individualLimit`, `credits`, and `spendControlReached`, all
+correctly ignored by the parser. Two unsolicited notifications arrived during
+the exchange — `configWarning` and `remoteControl/status/changed` — which
+exercised the ignore path with real traffic. No server-to-client request
+arrived, so the refusal path remains proven only by test.
+
+The live account's `windowDurationMins` was **10080**: a weekly window, not the
+fifteen-minute example in the upstream README. That exposed a real defect no
+fixture built from the documented example would have caught —
+`format_duration_ms` rendered it as "151h 30m" — and days were added to the
+formatter as a result. Fixtures agree with documentation; only a real account
+disagrees with your assumptions.
+
+## What is still not true
+
+The whole test suite runs without the `codex` binary, a `ChatGPT` account, or
+network access, and that property must be preserved: CI must never depend on an
+installed harness. A single live reading verifies the shape of one account on
+one plan; it does not exercise the `secondary` window, the `-32600` no-quota
+path, or a window that rolls mid-read. Those remain covered by synthetic
+transcripts only.
+
+Wiring the probe into the running shell means spawning a process against the
+user's own subscription account on every launch. Until that is wired, the shell
+keeps rendering its declared fixture.

docs/decisions/0014-claude-code-transcript-usage.md

@@ -0,0 +1,103 @@
+# 0014: Claude Code usage comes from its transcripts, and reports spend without a ceiling
+
+Status: accepted for token spend. **Corrected by decision 0015**: this record
+originally claimed Claude Code exposes no quota window. That was wrong — it
+does, through the status line — and the claim is struck below rather than
+quietly edited out, because the reasoning that followed from it shaped the
+adapter.
+
+Decision 0013 established the provenance test for a usage adapter: does the
+harness forward a provider's number, or record its own? Claude Code is the case
+that makes the second half of that test matter.
+
+## Correction
+
+The original text asserted that "Claude Code publishes no quota API and no
+remaining-balance command" and that a user asking how much they have left
+"gets the honest answer — Lumbridge can show what was spent, and cannot show
+what remains, because Claude Code does not say."
+
+**Both statements are false.** Claude Code 2.1.80 and later pipe a
+`rate_limits` object with `five_hour` and `seven_day` windows to the configured
+`statusLine` command on every turn, and the CLI documents that shape itself.
+There is also an account usage endpoint. The correct claim is narrower and is
+the only one this record still makes: *the transcript* carries spend and not
+quota. Decision 0015 covers the window surface.
+
+## The surface
+
+What Claude Code writes to disk per session is the API's `usage` object for
+every assistant turn,
+appended to a per-session JSONL transcript under
+`$CLAUDE_CONFIG_DIR/projects` (default `~/.claude/projects`). Each assistant
+record carries `input_tokens`, `output_tokens`,
+`cache_creation_input_tokens`, and `cache_read_input_tokens`, alongside the
+model that produced it.
+
+Summing those is how this adapter reports spend. **A subscription's window,
+ceiling, and reset are not in the transcript**, so this profile reports
+consumption against no ceiling rather than implying a quota — but that is a
+fact about this file, not about Claude Code, and the quota is read separately
+under decision 0015.
+
+The `statusLine` hook does require mutating the user's `settings.json`, which
+is why the transcript came first: reading a file the harness already writes
+needs no change to the user's configuration. That was a good reason to build
+this adapter, and a bad reason to conclude the other surface did not exist.
+
+## Why harness-reported
+
+Claude Code forwards the provider's per-response counts faithfully, so a single
+record's numbers originate with the provider. But a running total is the
+*harness's* record — an append-only log Lumbridge is reading, not a provider's
+statement of account. Under 0013's rule the value is therefore
+`HarnessReported`, and the footer says `harness · approximate`.
+
+## Content blindness is structural, not procedural
+
+**These files contain the user's conversations.** The parser models exactly
+four integers, a record type, and a model name. Every other field — message
+text, tool inputs, file contents, `service_tier`, `inference_geo`,
+`server_tool_use` — is discarded by serde during parsing, because there is no
+field in the wire types that could hold it. A test asserts that no debug
+rendering of a parsed record can contain message content.
+
+That is the same discipline as the Codex probe's refusal path: the safe
+behaviour is made unrepresentable rather than merely avoided.
+
+## What "tokens" counts
+
+The reported number is `input + output + cache_creation + cache_read`.
+
+Cache reads are included because they are billed and do count against a
+subscription window, even though they are re-reads of context already counted
+once. Excluding them would understate consumption. Weighting them by price
+would require a price list this adapter does not have, and would turn a counted
+fact into an estimate. In practice cache reads dominate the total — on the
+transcripts this was verified against they were roughly 98% of it — so the
+derived burn rate is largely a measure of context re-read per turn rather than
+of new work. The split is preserved in `TokenTally` for a detail view that can
+say so.
+
+## Priming before reporting
+
+The probe is a tail follower: it remembers a byte offset per transcript and
+reads only what was appended since, so its running total is monotonic by
+construction. The ledger rejects an observation that moves a profile backwards,
+so this is a correctness requirement, not an optimisation.
+
+It also **reports nothing until it has read every existing transcript to its
+end**. A partially-read backlog is indistinguishable from an enormous burst of
+spend, and the ledger will derive a burn rate from it: the first run against a
+20 MB backlog with an 8 MB per-poll budget reported **forty-six billion tokens
+an hour**, which was catch-up, not usage. The first reading must be a true
+baseline before any rate derived from it can mean anything. A shrinking file
+resets its own offset rather than trusting a position into different content.
+
+## Bounds
+
+A scan walks at most six directory levels and 4096 files, never follows
+symlinks out of the tree, reads at most 8 MB per poll, and refuses a single
+record over 1 MB while still advancing past it. A transcript's last line is
+routinely a partial record still being written; it is left unconsumed for the
+next read rather than parsed or skipped.

docs/decisions/0015-claude-code-subscription-windows.md

@@ -0,0 +1,97 @@
+# 0015: Claude Code's subscription windows come from the status line, not the credential
+
+Status: accepted; bridge installed and verified end to end.
+
+Decision 0014 concluded that Claude Code cannot report a remaining balance.
+That was wrong, and the error was one of not looking rather than of reasoning:
+the transcript genuinely carries no quota, and that fact was generalised into a
+claim about the harness. Claude Code reports its quota through three other
+channels.
+
+## The four surfaces
+
+| Surface | Carries | Cost | Needs a credential |
+|---|---|---|---|
+| **Status line `rate_limits`** | `five_hour`, `seven_day` — `used_percentage` + `resets_at` | none; relayed from headers the CLI already received | no |
+| Account usage endpoint | the same windows, plus per-model scoped limits | tight budget; 429s under polling | **yes** |
+| Stream `rate_limit_info` | `status`, `rateLimitType`, `unifiedWindows` | none | no, but requires driving the harness |
+| `/usage` in a PTY | rendered text | a turn | no, but screen-scraped |
+
+The status-line shape is documented inside the CLI itself:
+
+```text
+"five_hour": {   // present only while the API reports it and its resets_at has not passed
+  "used_percentage": number,  // Percentage of limit used (0-100)
+  "resets_at": number         // Unix epoch seconds when this window resets
+}
+```
+
+`rate_limits_available` is the honest-gap signal, also documented: "False when
+plan rate limits do not apply (API key, Bedrock, Vertex, or missing profile
+scope)". An account with no subscription window says so, rather than leaving
+Lumbridge to infer it from an absence.
+
+## Why the status line and not the endpoint
+
+The account usage endpoint is more authoritative — it answers without an active
+session and exposes per-model scoped limits. It also requires reading Claude
+Code's OAuth credential, and AGENTS.md says Lumbridge "must not scrape their
+private credentials." Other tools in this space do read it; that is their call
+to make, and it is a real capability difference.
+
+Lumbridge keeps the rule. The status line is pushed *to* a command the user
+installs, so no credential is ever touched, and the data is free because the
+CLI is relaying rate-limit headers it already received on its own API calls.
+The cost is that a window only appears once a session has run, and that
+installation edits `settings.json`.
+
+The endpoint is recorded here as a rejected-for-now alternative rather than an
+unconsidered one. Reopening it means amending AGENTS.md first.
+
+## Provenance
+
+`ProviderReported`. Under decision 0013's test the CLI forwards the provider's
+number rather than computing one — the same standing as Codex's forwarded
+`x-codex-*` headers. The footer shows `provider · exact`.
+
+## The bridge
+
+`scripts/claude-statusline-bridge.sh` is installed as Claude Code's
+`statusLine` command. It appends the rate-limit fields — and only those — to
+`$XDG_DATA_HOME/lumbridge/claude-rate-limits.jsonl`, and prints a status line.
+
+Three properties it must hold:
+
+- **It cannot fail.** A status-line command that errors or hangs degrades the
+  user's session, so every step is guarded and it always exits 0 having
+  printed something.
+- **It copies nothing else.** The payload also carries the session's cost,
+  transcript path, working directory, and model. The bridge builds a fresh
+  record from three numeric fields per window rather than filtering the
+  original, so there is no path by which the rest travels.
+- **It stays bounded.** The feed rotates past 256 KB. Lines are appended rather
+  than overwritten so concurrent sessions do not clobber each other; the probe
+  reads the newest line, because each line is a complete snapshot and an older
+  one is superseded, never summed.
+
+Installation is a separate, explicit, reversible script — not something a probe
+does on the user's behalf. It backs up `settings.json`, writes through a
+temporary file, refuses to replace a `statusLine` someone else configured, and
+supports `--dry-run` and `--uninstall`.
+
+## Window handling
+
+Each window becomes its own account-free profile — `claude-code-five-hour` and
+`claude-code-seven-day` — because two windows are two facts about two quota
+periods, and merging them means silently picking one. The pane maps to the
+five-hour window: it is the limit that actually stops work.
+
+A reading's window start is derived from `resets_at` minus the documented
+window length, which is what lets the ledger scope a burn rate to one quota
+period. Two rules follow decision 0012 unchanged: a window whose `resets_at`
+has passed is stale, so its reset is dropped and only the percentage survives;
+and a start the local clock places in the future costs the start, not the
+reset, which is the half that bounds a forecast.
+
+A percentage above 100 — possible on a gateway spend limit once breached — is
+clamped to the whole rather than stored as more than a full window.

scripts/claude-statusline-bridge.py

@@ -0,0 +1,108 @@
+"""Extract Claude Code's rate-limit windows from a status-line payload.
+
+Invoked by `claude-statusline-bridge.sh` with the feed path as argv[1] and the
+status-line JSON on stdin. Appends only the rate-limit fields to the feed and
+prints a status line.
+
+Must never fail: a status-line command that errors degrades the user's Claude
+Code session, so every step is guarded and this always exits 0 having printed.
+"""
+
+import json
+import os
+import sys
+
+# Keep the feed bounded. The probe only needs the newest line; older ones are
+# superseded snapshots, not history worth keeping.
+MAX_FEED_BYTES = 256 * 1024
+WINDOW_NAMES = ("five_hour", "seven_day")
+PERCENT_KEYS = ("used_percentage", "utilization", "resets_at")
+
+
+def extract(payload):
+    """Return the record to append: rate-limit fields and nothing else.
+
+    The payload also carries the session's cost, transcript path, working
+    directory, and model. None of that is Lumbridge's business, and none of it
+    is copied — this builds a fresh dict rather than filtering the original.
+    """
+    record = {"rate_limits_available": payload.get("rate_limits_available")}
+    limits = payload.get("rate_limits")
+    if not isinstance(limits, dict):
+        return record
+    windows = {}
+    for name in WINDOW_NAMES:
+        window = limits.get(name)
+        if not isinstance(window, dict):
+            continue
+        kept = {
+            key: window[key]
+            for key in PERCENT_KEYS
+            if isinstance(window.get(key), (int, float))
+            and not isinstance(window.get(key), bool)
+        }
+        if kept:
+            windows[name] = kept
+    if windows:
+        record["rate_limits"] = windows
+    return record
+
+
+def status_line(record):
+    if record.get("rate_limits_available") is False:
+        return "lumbridge · no plan limits on this account"
+    windows = record.get("rate_limits")
+    if not windows:
+        return "lumbridge · no window reported yet"
+    parts = []
+    for name, label in (("five_hour", "5h"), ("seven_day", "7d")):
+        window = windows.get(name)
+        if not window:
+            continue
+        percent = window.get("used_percentage", window.get("utilization"))
+        if percent is None:
+            continue
+        parts.append(f"{label} {max(0.0, 100.0 - float(percent)):.0f}% left")
+    if not parts:
+        return "lumbridge · no window reported yet"
+    return "lumbridge · " + " · ".join(parts)
+
+
+def append(feed, record):
+    if not feed:
+        return
+    try:
+        # Rotate rather than grow without bound. The probe resets its read
+        # offset when the file shrinks, so a rotation costs one reading.
+        if os.path.exists(feed) and os.path.getsize(feed) > MAX_FEED_BYTES:
+            open(feed, "w").close()
+        # Append so concurrent Claude Code sessions do not clobber each other;
+        # a short line written in append mode lands intact.
+        with open(feed, "a") as handle:
+            handle.write(json.dumps(record, separators=(",", ":")) + "\n")
+    except Exception:
+        pass  # A feed we cannot write is not worth breaking a session over.
+
+
+def main():
+    feed = sys.argv[1] if len(sys.argv) > 1 else ""
+    try:
+        payload = json.load(sys.stdin)
+    except Exception:
+        print("lumbridge · unreadable status payload")
+        return 0
+    if not isinstance(payload, dict):
+        print("lumbridge · unreadable status payload")
+        return 0
+    record = extract(payload)
+    append(feed, record)
+    print(status_line(record))
+    return 0
+
+
+try:
+    sys.exit(main())
+except Exception:
+    # Last resort: never let a status line take a session down.
+    print("lumbridge")
+    sys.exit(0)

scripts/claude-statusline-bridge.sh

@@ -0,0 +1,31 @@
+#!/bin/sh
+# Lumbridge status-line bridge for Claude Code.
+#
+# Claude Code 2.1.80+ pipes a JSON payload to the configured `statusLine`
+# command on every turn. That payload carries `rate_limits.five_hour` and
+# `rate_limits.seven_day` — the real subscription windows, relayed from
+# rate-limit headers the CLI already received. Reading them here costs nothing
+# and needs no credential, which is why Lumbridge takes this route rather than
+# calling the account usage endpoint.
+#
+# This wrapper exists so a missing python3 degrades to a printed message
+# instead of a broken status line. It `exec`s the parser so stdin stays the
+# payload.
+#
+# Install with scripts/install-claude-statusline.sh; uninstall by removing the
+# statusLine block it adds to settings.json.
+
+set -u
+
+FEED="${LUMBRIDGE_CLAUDE_FEED:-${XDG_DATA_HOME:-$HOME/.local/share}/lumbridge/claude-rate-limits.jsonl}"
+HERE=$(dirname "$0")
+
+if ! command -v python3 >/dev/null 2>&1; then
+  # No parser available. Say so rather than printing a number we cannot read.
+  echo "lumbridge · python3 not found"
+  exit 0
+fi
+
+mkdir -p "$(dirname "$FEED")" 2>/dev/null || true
+
+exec python3 "$HERE/claude-statusline-bridge.py" "$FEED"

scripts/install-claude-statusline.sh

@@ -0,0 +1,113 @@
+#!/bin/sh
+# Installs the Lumbridge status-line bridge into Claude Code's settings.
+#
+# This is the one thing Lumbridge cannot do without touching the user's
+# configuration, so it is a separate, explicit, reversible step rather than
+# something a probe does on your behalf.
+#
+# It sets `statusLine` in $CLAUDE_CONFIG_DIR/settings.json (default
+# ~/.claude/settings.json) to run scripts/claude-statusline-bridge.sh. Claude
+# Code then pipes its rate-limit payload there on every turn, which is how
+# Lumbridge learns the five-hour and seven-day subscription windows without
+# ever reading a credential.
+#
+#   install:    scripts/install-claude-statusline.sh
+#   uninstall:  scripts/install-claude-statusline.sh --uninstall
+#   preview:    scripts/install-claude-statusline.sh --dry-run
+#
+# An existing statusLine is backed up and reported, never silently replaced.
+
+set -eu
+
+MODE=install
+case "${1:-}" in
+  --uninstall) MODE=uninstall ;;
+  --dry-run) MODE=dry-run ;;
+  "") ;;
+  *) echo "usage: $0 [--uninstall|--dry-run]" >&2; exit 2 ;;
+esac
+
+CONFIG_DIR="${CLAUDE_CONFIG_DIR:-$HOME/.claude}"
+SETTINGS="$CONFIG_DIR/settings.json"
+HERE=$(cd "$(dirname "$0")" && pwd)
+BRIDGE="$HERE/claude-statusline-bridge.sh"
+
+if [ ! -x "$BRIDGE" ]; then
+  echo "bridge not executable: $BRIDGE" >&2
+  exit 1
+fi
+if ! command -v python3 >/dev/null 2>&1; then
+  echo "python3 is required to edit settings.json safely" >&2
+  exit 1
+fi
+
+python3 - "$SETTINGS" "$BRIDGE" "$MODE" <<'PYTHON'
+import json
+import os
+import shutil
+import sys
+import time
+
+settings_path, bridge, mode = sys.argv[1], sys.argv[2], sys.argv[3]
+desired = {"type": "command", "command": bridge}
+
+settings = {}
+if os.path.exists(settings_path):
+    try:
+        with open(settings_path) as handle:
+            settings = json.load(handle)
+    except Exception as error:
+        print(f"refusing to edit unreadable settings: {error}", file=sys.stderr)
+        raise SystemExit(1)
+    if not isinstance(settings, dict):
+        print("refusing to edit: settings.json is not an object", file=sys.stderr)
+        raise SystemExit(1)
+
+current = settings.get("statusLine")
+
+if mode == "uninstall":
+    if isinstance(current, dict) and current.get("command") == bridge:
+        settings.pop("statusLine", None)
+        action = "removed the Lumbridge status line"
+    else:
+        print("no Lumbridge status line installed; nothing to do")
+        raise SystemExit(0)
+elif isinstance(current, dict) and current.get("command") == bridge:
+    print("Lumbridge status line already installed; nothing to do")
+    raise SystemExit(0)
+else:
+    if current is not None:
+        # Someone else's status line is here. Report it and stop rather than
+        # replacing a thing the user set up.
+        print("a different statusLine is already configured:", file=sys.stderr)
+        print(f"  {json.dumps(current)}", file=sys.stderr)
+        print(
+            "remove it first, or merge the bridge into it by hand:\n"
+            f"  {bridge}",
+            file=sys.stderr,
+        )
+        raise SystemExit(1)
+    settings["statusLine"] = desired
+    action = "installed the Lumbridge status line"
+
+if mode == "dry-run":
+    print("would write:")
+    print(json.dumps({"statusLine": settings.get("statusLine")}, indent=2))
+    raise SystemExit(0)
+
+os.makedirs(os.path.dirname(settings_path), exist_ok=True)
+if os.path.exists(settings_path):
+    backup = f"{settings_path}.lumbridge-backup-{int(time.time())}"
+    shutil.copy2(settings_path, backup)
+    print(f"backed up existing settings to {backup}")
+
+# Write via a temporary file so an interrupted run cannot leave settings.json
+# truncated — this file configures the user's editor session.
+temporary = f"{settings_path}.lumbridge-tmp"
+with open(temporary, "w") as handle:
+    json.dump(settings, handle, indent=2)
+    handle.write("\n")
+os.replace(temporary, settings_path)
+print(f"{action} in {settings_path}")
+print("restart or start a Claude Code session for it to take effect")
+PYTHON

scripts/open-metal-lumbridge.sh

@@ -44,11 +44,31 @@ if [[ ! -x "$binary_path" ]]; then
   cargo build --locked --manifest-path "$manifest_path"
 fi
 
-existing_id=$(wmctrl -l | awk -v title="$window_title" 'index($0, title) { print $1; exit }')
+# GPUI's X11 client sets WM_NAME but not _NET_WM_NAME, so wmctrl's window list
+# reports its title as N/A and a title match never succeeds. xdotool reads
+# WM_NAME, and matching the launched PID avoids picking up another client that
+# merely mentions the window title.
+window_for_pid() {
+  local pid=$1 candidate
+  for candidate in $(xdotool search --name 'GPUI workspace' 2>/dev/null); do
+    if [[ "$(xdotool getwindowpid "$candidate" 2>/dev/null || true)" == "$pid" ]]; then
+      printf '%s\n' "$candidate"
+      return 0
+    fi
+  done
+  return 1
+}
+
+existing_id=''
+app_pid=$(pgrep -f "^$binary_path$" | head -n 1 || true)
+if [[ -n "$app_pid" ]]; then
+  existing_id=$(window_for_pid "$app_pid" || true)
+fi
 if [[ -z "$existing_id" ]]; then
   "$binary_path" >"$XDG_RUNTIME_DIR/lumbridge-gpui.log" 2>&1 &
+  app_pid=$!
   for _attempt in $(seq 1 80); do
-    existing_id=$(wmctrl -l | awk -v title="$window_title" 'index($0, title) { print $1; exit }')
+    existing_id=$(window_for_pid "$app_pid" || true)
     [[ -n "$existing_id" ]] && break
     sleep 0.1
   done
@@ -59,23 +79,43 @@ if [[ -z "$existing_id" ]]; then
   exit 1
 fi
 
-# Gigabyte G34WQC is 3440x1440 at X=0. Lumbridge owns its left half; Bacon
+# Gigabyte G34WQC is 3440x1440 at X=0. Lumbridge owns the whole panel: the
-# remains visible beside it on the right half of the same monitor.
+# workspace is a wall of agents and every pixel of width buys another column.
-bacon_id=$(wmctrl -l | awk -v title="$bacon_title" 'index($0, title) { print $1; exit }')
+# Set LUMBRIDGE_HALF=1 to fall back to the left half beside Bacon.
-if [[ -n "$bacon_id" ]]; then
+if [[ "${LUMBRIDGE_HALF:-0}" == "1" ]]; then
-  wmctrl -i -r "$bacon_id" -b remove,maximized_vert,maximized_horz
+  bacon_id=$(wmctrl -l | awk -v title="$bacon_title" 'index($0, title) { print $1; exit }')
-  # GNOME's X11 move coordinate is scaled on this display even though window
+  if [[ -n "$bacon_id" ]]; then
-  # sizes are not; 860 places the frame at physical X=1720.
+    wmctrl -i -r "$bacon_id" -b remove,maximized_vert,maximized_horz
-  wmctrl -i -r "$bacon_id" -e 0,860,0,1720,1400
+    # GNOME's X11 move coordinate is scaled on this display even though window
+    # sizes are not; 860 places the frame at physical X=1720.
+    wmctrl -i -r "$bacon_id" -e 0,860,0,1720,1400
+  fi
+  wmctrl -i -r "$existing_id" -b remove,fullscreen
+  wmctrl -i -r "$existing_id" -b remove,maximized_vert,maximized_horz
+  wmctrl -i -r "$existing_id" -e 0,0,0,1720,1400
+  wmctrl -i -a "$existing_id"
+  # GPUI's undecorated X11 client ignores direct position requests under GNOME.
+  # Use the window manager's move interaction to settle it into the left half.
+  xdotool key alt+F7
+  sleep 0.15
+  xdotool mousemove 860 660 click 1
+  echo "Lumbridge is on the left half of the Gigabyte display (window $existing_id)."
+  exit 0
 fi
 
+# Fullscreen drops the title bar and the shell owns all 3440x1440. Move it onto
+# the Gigabyte first: a fullscreen request applies to whichever output the
+# window is currently on, so placing it after would fullscreen the wrong panel.
+wmctrl -i -r "$existing_id" -b remove,fullscreen
 wmctrl -i -r "$existing_id" -b remove,maximized_vert,maximized_horz
 wmctrl -i -r "$existing_id" -e 0,0,0,1720,1400
 wmctrl -i -a "$existing_id"
-# GPUI's undecorated X11 client ignores direct position requests under GNOME.
-# Use the window manager's move interaction to settle it into the left half.
 xdotool key alt+F7
 sleep 0.15
 xdotool mousemove 860 660 click 1
+sleep 0.2
+wmctrl -i -r "$existing_id" -b add,fullscreen
+sleep 0.3
 
-echo "Lumbridge is visible in window $existing_id on the left half of the Gigabyte display."
+geometry=$(xdotool getwindowgeometry "$existing_id" | awk '/Geometry/ { print $2 }')
+echo "Lumbridge is fullscreen on the Gigabyte display at ${geometry} (window $existing_id)."

spikes/gpui-shell/Cargo.lock

@@ -3043,6 +3043,17 @@ dependencies = [
  "thiserror 2.0.20",
 ]
 
+[[package]]
+name = "lumbridge-harness"
+version = "0.0.1"
+dependencies = [
+ "lumbridge-core",
+ "nix 0.28.0",
+ "serde",
+ "serde_json",
+ "thiserror 2.0.20",
+]
+
 [[package]]
 name = "lumbridge-pty"
 version = "0.0.1"
@@ -3065,6 +3076,8 @@ name = "lumbridge-spike-gpui"
 version = "0.0.1"
 dependencies = [
  "gpui",
+ "lumbridge-core",
+ "lumbridge-harness",
  "lumbridge-runtime",
  "lumbridge-spike-model",
  "lumbridge-storage",

spikes/gpui-shell/Cargo.toml

@@ -9,6 +9,8 @@ publish = false
 
 [dependencies]
 gpui = "0.2.2"
+lumbridge-core = { path = "../../crates/lumbridge-core" }
+lumbridge-harness = { path = "../../crates/lumbridge-harness" }
 lumbridge-runtime = { path = "../../crates/lumbridge-runtime" }
 lumbridge-spike-model = { path = "../ui-shell-model" }
 lumbridge-storage = { path = "../../crates/lumbridge-storage" }

spikes/gpui-shell/src/main.rs

@@ -1,4 +1,5 @@
 mod panel_registry;
+mod usage_feed;
 
 use std::collections::{BTreeMap, VecDeque};
 use std::path::PathBuf;
@@ -8,13 +9,14 @@ use gpui::{
     App, Application, Bounds, Context, FocusHandle, FontWeight, KeyBinding, KeyDownEvent, Pixels,
     Size, Window, WindowBounds, WindowOptions, actions, div, prelude::*, px, relative, rgb, size,
 };
+use lumbridge_core::UsageProvenance;
 use lumbridge_runtime::{
     CommandConfig, PtyOptions, RuntimeActorError, RuntimeActorOptions, RuntimeCommand,
     RuntimeEvent, RuntimeRegistry, RuntimeRegistryError, TerminalSize,
 };
 use lumbridge_spike_model::{
-    ActionOutcome, FOOTER_RIGHT, OutputSource, PaneId as FixturePaneId, PaneStatus, ShellAction,
+    ActionOutcome, OutputSource, PaneId as FixturePaneId, PaneStatus, ShellAction, ShellModel,
-    ShellModel, SurfaceKind, WORKSPACES,
+    SurfaceKind, WORKSPACES,
 };
 use lumbridge_storage::Store;
 use lumbridge_terminal::{
@@ -24,6 +26,7 @@ use lumbridge_terminal::{
 };
 
 use panel_registry::{PanelId, PanelKind, PanelRegistry, SeedPane};
+use usage_feed::{UsageFeed, UsageSegment};
 
 const BG: u32 = 0x090c12;
 const PANEL: u32 = 0x101620;
@@ -43,7 +46,7 @@ const WORKSPACE_SNAPSHOT_ID: &str = "lumbridge-code-gpui-spike";
 const SIDEBAR_WIDTH: f32 = 248.0;
 const APP_HEADER_HEIGHT: f32 = 48.0;
 const TAB_BAR_HEIGHT: f32 = 38.0;
-const APP_FOOTER_HEIGHT: f32 = 32.0;
+const APP_FOOTER_HEIGHT: f32 = 46.0;
 const TERMINAL_CONTENT_VERTICAL_INSET: f32 = 24.0;
 const TERMINAL_HORIZONTAL_INSET: f32 = 32.0;
 const TERMINAL_CELL_WIDTH: f32 = 8.4;
@@ -84,10 +87,90 @@ struct LumbridgeShell {
     live_terminals: BTreeMap<PanelId, LiveTerminalState>,
     store: Option<Store>,
     persistence_status: String,
+    usage: UsageFeed,
+    /// Which surface tab each panel is showing. A panel is absent until the
+    /// user picks something other than the surface it provides natively.
+    surfaces: BTreeMap<PanelId, SurfaceTab>,
+    /// Which decision-shelf choice each panel has selected. Selecting one is
+    /// inert by design: it prepares nothing and runs nothing.
+    shelf_choice: BTreeMap<PanelId, usize>,
+    active_worktree: usize,
     add_panel_chooser_open: bool,
     root_focus: FocusHandle,
 }
 
+/// The surfaces a pane can be viewed through.
+///
+/// A pane is the durable unit of work and its surface is a view over that work.
+/// Switching one never launches a process, moves the pane, or changes which
+/// agent owns the session — it only changes what is drawn.
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+enum SurfaceTab {
+    Terminal,
+    Browser,
+    Tools,
+    Context,
+    Goal,
+    Review,
+}
+
+impl SurfaceTab {
+    const ALL: [Self; 6] = [
+        Self::Terminal,
+        Self::Browser,
+        Self::Tools,
+        Self::Context,
+        Self::Goal,
+        Self::Review,
+    ];
+
+    const fn label(self) -> &'static str {
+        match self {
+            Self::Terminal => "TERMINAL",
+            Self::Browser => "BROWSER",
+            Self::Tools => "TOOLS",
+            Self::Context => "CONTEXT",
+            Self::Goal => "GOAL",
+            Self::Review => "REVIEW",
+        }
+    }
+
+    const fn ordinal(self) -> u64 {
+        match self {
+            Self::Terminal => 0,
+            Self::Browser => 1,
+            Self::Tools => 2,
+            Self::Context => 3,
+            Self::Goal => 4,
+            Self::Review => 5,
+        }
+    }
+
+    /// The one surface a pane of this kind actually provides.
+    const fn native_for(kind: SurfaceKind) -> Self {
+        match kind {
+            SurfaceKind::Terminal => Self::Terminal,
+            SurfaceKind::Browser => Self::Browser,
+            SurfaceKind::Markdown => Self::Context,
+            SurfaceKind::Review => Self::Review,
+        }
+    }
+
+    /// Why this pane cannot show this surface. Stated plainly, because the
+    /// product boundary says an unsupported surface is shown as unavailable
+    /// rather than simulated.
+    const fn unavailable_reason(self) -> &'static str {
+        match self {
+            Self::Terminal => "Only a terminal pane owns a live shell.",
+            Self::Browser => "No isolated web engine is wired yet.",
+            Self::Tools => "Tool calls arrive with the ACP client.",
+            Self::Context => "Context projection is not built yet.",
+            Self::Goal => "Goal tracking is not built yet.",
+            Self::Review => "Review is not wired to Git yet.",
+        }
+    }
+}
+
 #[derive(Clone)]
 struct PanelView {
     id: PanelId,
@@ -488,6 +571,8 @@ impl LumbridgeShell {
                 if this
                     .update(cx, |shell, cx| {
                         shell.dispatch(ShellAction::SyntheticStreamTick);
+                        shell.usage.advance();
+                        shell.usage.tick();
                         cx.notify();
                     })
                     .is_err()
@@ -554,6 +639,10 @@ impl LumbridgeShell {
             live_terminals,
             store,
             persistence_status,
+            usage: UsageFeed::start(),
+            surfaces: BTreeMap::new(),
+            shelf_choice: BTreeMap::new(),
+            active_worktree: 0,
             add_panel_chooser_open: false,
             root_focus,
         }
@@ -778,6 +867,54 @@ impl LumbridgeShell {
         cx.notify();
     }
 
+    /// Changes which surface a pane is viewed through.
+    ///
+    /// This is a view change only. It never touches the PTY, the agent, or the
+    /// pane's execution target, so a pane keeps running whatever it was running.
+    fn select_surface(
+        &mut self,
+        pane: PanelId,
+        tab: SurfaceTab,
+        window: &mut Window,
+        cx: &mut Context<Self>,
+    ) {
+        self.timing.mark_dispatch();
+        self.surfaces.insert(pane, tab);
+        if self.panels.select(pane) {
+            self.persist_panels();
+        }
+        window.focus(&self.root_focus);
+        cx.notify();
+    }
+
+    /// Records which suggestion the user picked. Deliberately does nothing else.
+    fn select_shelf_choice(
+        &mut self,
+        pane: PanelId,
+        index: usize,
+        window: &mut Window,
+        cx: &mut Context<Self>,
+    ) {
+        self.timing.mark_dispatch();
+        if self.shelf_choice.get(&pane) == Some(&index) {
+            self.shelf_choice.remove(&pane);
+        } else {
+            self.shelf_choice.insert(pane, index);
+        }
+        if self.panels.select(pane) {
+            self.persist_panels();
+        }
+        window.focus(&self.root_focus);
+        cx.notify();
+    }
+
+    fn select_worktree(&mut self, index: usize, window: &mut Window, cx: &mut Context<Self>) {
+        self.timing.mark_dispatch();
+        self.active_worktree = index.min(WORKSPACES.len().saturating_sub(1));
+        window.focus(&self.root_focus);
+        cx.notify();
+    }
+
     fn select_pane_at(&mut self, index: usize, window: &mut Window, cx: &mut Context<Self>) {
         self.timing.mark_dispatch();
         if self.panels.select_attached_at(index) {
@@ -1215,26 +1352,38 @@ impl LumbridgeShell {
                 }
             },
         );
-        let surface = match pane.kind {
+        let native = SurfaceTab::native_for(pane.kind);
-            SurfaceKind::Terminal => "TERMINAL",
+        let shown = self.surface_for(pane.id, pane.kind);
-            SurfaceKind::Markdown => "CONTEXT",
+        let tabs = SurfaceTab::ALL
-            SurfaceKind::Browser => "BROWSER",
-            SurfaceKind::Review => "REVIEW",
-        };
-        let tabs = ["TERMINAL", "BROWSER", "TOOLS", "CONTEXT", "GOAL", "REVIEW"]
             .into_iter()
-            .map(|label| {
+            .map(|tab| {
+                let active = tab == shown;
+                let provided = tab == native;
                 div()
+                    .id(("surface-tab", pane_id.get() * 16 + tab.ordinal()))
+                    .cursor_pointer()
                     .h_full()
                     .flex()
                     .items_center()
                     .px_3()
                     .text_xs()
-                    .text_color(rgb(if label == surface { ACCENT } else { MUTED }))
+                    // Provided but unselected reads as available; unprovided
-                    .when(label == surface, |view| {
+                    // stays dim, so the tab strip shows what this pane can do
-                        view.border_b_2().border_color(rgb(ACCENT))
+                    // before you click rather than after.
-                    })
+                    .text_color(rgb(if active {
-                    .child(label)
+                        ACCENT
+                    } else if provided {
+                        TEXT
+                    } else {
+                        MUTED
+                    }))
+                    .when(active, |view| view.border_b_2().border_color(rgb(ACCENT)))
+                    .hover(|view| view.bg(rgb(PANEL_ACTIVE)).text_color(rgb(TEXT)))
+                    .child(tab.label())
+                    .on_click(cx.listener(move |shell, _, window, cx| {
+                        cx.stop_propagation();
+                        shell.select_surface(pane_id, tab, window, cx);
+                    }))
             })
             .collect::<Vec<_>>();
 
@@ -1304,6 +1453,12 @@ impl LumbridgeShell {
                                     .border_color(rgb(BORDER))
                                     .text_xs()
                                     .text_color(rgb(if can_detach { MUTED } else { BORDER }))
+                                    .when(can_detach, |view| {
+                                        view.hover(|view| {
+                                            view.border_color(rgb(ATTENTION))
+                                                .text_color(rgb(ATTENTION))
+                                        })
+                                    })
                                     .child(if can_detach {
                                         "− DETACH"
                                     } else {
@@ -1347,7 +1502,67 @@ impl LumbridgeShell {
             .into_any_element()
     }
 
+    /// The surface a panel is currently being viewed through.
+    fn surface_for(&self, id: PanelId, kind: SurfaceKind) -> SurfaceTab {
+        self.surfaces
+            .get(&id)
+            .copied()
+            .unwrap_or_else(|| SurfaceTab::native_for(kind))
+    }
+
+    /// What a pane shows when asked for a surface it does not provide.
+    ///
+    /// It says so, and it says what is still running underneath, because the
+    /// pane's process identity does not change when the view does.
+    fn unavailable_surface(&self, pane: &PanelView, shown: SurfaceTab) -> gpui::AnyElement {
+        let native = SurfaceTab::native_for(pane.kind);
+        div()
+            .flex()
+            .flex_col()
+            .size_full()
+            .min_w_0()
+            .min_h_0()
+            .overflow_hidden()
+            .bg(rgb(BG))
+            .border_y_1()
+            .border_color(rgb(BORDER))
+            .child(
+                div()
+                    .flex()
+                    .flex_col()
+                    .gap_2()
+                    .p_4()
+                    .child(
+                        div()
+                            .text_sm()
+                            .text_color(rgb(TEXT))
+                            .child(format!("{} unavailable", shown.label())),
+                    )
+                    .child(
+                        div()
+                            .text_xs()
+                            .text_color(rgb(MUTED))
+                            .child(shown.unavailable_reason()),
+                    )
+                    .child(
+                        div()
+                            .mt_2()
+                            .text_xs()
+                            .text_color(rgb(SUCCESS))
+                            .child(format!(
+                                "{} is still this pane's live surface. Nothing was stopped.",
+                                native.label()
+                            )),
+                    ),
+            )
+            .into_any_element()
+    }
+
     fn pane_work_surface(&self, pane: &PanelView) -> gpui::AnyElement {
+        let shown = self.surface_for(pane.id, pane.kind);
+        if shown != SurfaceTab::native_for(pane.kind) {
+            return self.unavailable_surface(pane, shown);
+        }
         let external = pane.output_source == OutputSource::External;
         let content = if external {
             self.terminal_view(pane.id)
@@ -1387,35 +1602,22 @@ impl LumbridgeShell {
             .into_any_element()
     }
 
-    fn pane_decision_region(&self, pane: &PanelView) -> gpui::AnyElement {
+    fn pane_decision_region(&self, pane: &PanelView, cx: &mut Context<Self>) -> gpui::AnyElement {
-        let choice = |label: &'static str, detail: &'static str, attention: bool| {
+        let pane_id = pane.id;
-            div()
+        let picked = self.shelf_choice.get(&pane_id).copied();
-                .flex()
+        // Every choice names the capability it would need. None of them holds
-                .items_center()
+        // that capability: choosing prepares, the command plane executes.
-                .justify_between()
+        let choices: [(&'static str, &'static str, &'static str, bool); 3] = if pane.needs_input() {
-                .min_w_0()
-                .px_2()
-                .py_1()
-                .rounded(px(4.0))
-                .border_1()
-                .border_color(rgb(if attention { ATTENTION } else { BORDER }))
-                .bg(rgb(PANEL_ALT))
-                .text_xs()
-                .text_color(rgb(TEXT))
-                .child(label)
-                .child(div().truncate().text_color(rgb(MUTED)).child(detail))
-        };
-        let choices = if pane.needs_input() {
             [
-                ("Review request", "inspect scope", true),
+                ("Review request", "inspect scope", "needs Observe", true),
-                ("Steer…", "edit response", false),
+                ("Steer…", "edit response", "needs Execute", false),
-                ("Dismiss", "leave inert", false),
+                ("Dismiss", "leave inert", "inert", false),
             ]
         } else {
             [
-                ("Continue", "keep moving", false),
+                ("Continue", "keep moving", "needs Execute", false),
-                ("Review plan", "inspect commands", false),
+                ("Review plan", "inspect commands", "needs Observe", false),
-                ("Ask…", "refine prompt", false),
+                ("Ask…", "refine prompt", "inert", false),
             ]
         };
         div()
@@ -1447,8 +1649,55 @@ impl LumbridgeShell {
                     ),
             )
             .child(div().flex().flex_col().gap_1().mt_2().children(
-                choices.map(|(label, detail, attention)| choice(label, detail, attention)),
+                choices.into_iter().enumerate().map(
+                    |(index, (label, detail, _capability, attention))| {
+                        let chosen = picked == Some(index);
+                        div()
+                            .id(("shelf-choice", pane_id.get() * 8 + index as u64))
+                            .cursor_pointer()
+                            .flex()
+                            .items_center()
+                            .justify_between()
+                            .min_w_0()
+                            .px_2()
+                            .py_1()
+                            .rounded(px(4.0))
+                            .border_1()
+                            .border_color(rgb(if chosen {
+                                ACCENT
+                            } else if attention {
+                                ATTENTION
+                            } else {
+                                BORDER
+                            }))
+                            .bg(rgb(if chosen { PANEL_ACTIVE } else { PANEL_ALT }))
+                            .hover(|view| view.border_color(rgb(ACCENT)))
+                            .text_xs()
+                            .text_color(rgb(TEXT))
+                            .child(label)
+                            .child(div().truncate().text_color(rgb(MUTED)).child(detail))
+                            .on_click(cx.listener(move |shell, _, window, cx| {
+                                cx.stop_propagation();
+                                shell.select_shelf_choice(pane_id, index, window, cx);
+                            }))
+                    },
+                ),
             ))
+            .child(
+                div()
+                    .mt_2()
+                    .text_xs()
+                    .text_color(rgb(if picked.is_some() { ACCENT } else { MUTED }))
+                    .child(picked.map_or_else(
+                        || "Pick one to see what it would need".to_owned(),
+                        |index| {
+                            format!(
+                                "{} · {} · nothing has run",
+                                choices[index].0, choices[index].2
+                            )
+                        },
+                    )),
+            )
             .into_any_element()
     }
 
@@ -1461,6 +1710,7 @@ impl LumbridgeShell {
         let pane_id = pane.id;
         div()
             .id(("workspace-panel", pane_id.get()))
+            .cursor_pointer()
             .on_click(cx.listener(move |shell, _, window, cx| {
                 shell.select_pane(pane_id, window, cx);
             }))
@@ -1494,7 +1744,7 @@ impl LumbridgeShell {
                     .flex_none()
                     .min_h_0()
                     .overflow_hidden()
-                    .child(self.pane_decision_region(pane)),
+                    .child(self.pane_decision_region(pane, cx)),
             )
             .into_any_element()
     }
@@ -1795,6 +2045,209 @@ impl LumbridgeShell {
                     ),
             )
     }
+
+    /// Attached panels whose agent is blocked on a human answer.
+    ///
+    /// The count is derived rather than declared: a sidebar that claims one
+    /// pane needs input while no pane is blocked is the same class of untruth
+    /// as an invented usage number.
+    fn attention_panels(&self) -> Vec<PanelView> {
+        self.panels
+            .attached_ids()
+            .into_iter()
+            .map(|id| self.panel_view(id))
+            .filter(PanelView::needs_input)
+            .collect()
+    }
+
+    /// Where sessions are owned, and what each host is actually doing.
+    fn runtime_rows(&self) -> Vec<RuntimeRow> {
+        let running = self
+            .live_terminals
+            .values()
+            .filter(|terminal| matches!(terminal.status, LiveRuntimeStatus::Running { .. }))
+            .count();
+        let total = self.live_terminals.len();
+        let local = RuntimeRow {
+            label: "metal",
+            detail: format!("{running}/{total} live PTYs"),
+            tone: if running == 0 { MUTED } else { SUCCESS },
+        };
+        // Remote runtimes are a Phase 1 deliverable. Until the SSH stdio
+        // transport exists, the only honest status for a saved host is that
+        // nothing is connected to it.
+        let remote = RuntimeRow {
+            label: "amd-server",
+            detail: "saved host · no runtime connected".to_owned(),
+            tone: MUTED,
+        };
+        // The usage adapter is neither a host nor an endpoint, but its health
+        // belongs beside them: it is the reason the footer can or cannot answer.
+        let adapter = RuntimeRow {
+            label: "usage adapter",
+            detail: format!(
+                "{} · {}/{} profiles reporting",
+                self.usage.probe_status(),
+                self.usage.reporting_profile_count(),
+                self.usage.declared_profile_count()
+            ),
+            tone: if self.usage.reporting_profile_count() == 0 {
+                MUTED
+            } else {
+                SUCCESS
+            },
+        };
+        vec![local, remote, adapter]
+    }
+
+    /// The standing of every harness, always, regardless of what is selected.
+    ///
+    /// The old footer answered only for the selected pane, so the moment you
+    /// focused a plain shell your quota vanished. This strip keeps every
+    /// harness on screen and leads with what is left, because that is the
+    /// question. The selected harness expands with its rate and its trust.
+    fn footer_usage_zone(&self) -> impl IntoElement {
+        let active = self
+            .panels
+            .panel(self.panels.selected())
+            .and_then(|panel| self.usage.profile_for_seed(panel.seed))
+            .cloned();
+        let segments = self.usage.strip();
+        let orphan = active.is_none();
+        div()
+            .flex()
+            .items_center()
+            .gap_4()
+            .when(orphan, |view| {
+                view.child(
+                    div()
+                        .flex_none()
+                        .text_color(rgb(MUTED))
+                        .child("no harness on this pane"),
+                )
+            })
+            .children(segments.into_iter().map(|segment| {
+                let selected = active.as_ref() == Some(&segment.id);
+                usage_segment(segment, selected)
+            }))
+    }
+}
+
+/// The provenance of a value is carried by the colour of its meter, so the
+/// strip reads at a glance without a legend on every segment.
+const fn provenance_color(provenance: UsageProvenance) -> u32 {
+    match provenance {
+        UsageProvenance::ProviderReported => SUCCESS,
+        UsageProvenance::HarnessReported => ACCENT,
+        UsageProvenance::LocallyMeasured => TEXT,
+        UsageProvenance::Estimated => ATTENTION,
+        UsageProvenance::Unavailable => MUTED,
+    }
+}
+
+/// A quota meter. The filled part is spent; the empty part is what is left,
+/// which is the way round an engineer reads it.
+fn usage_meter(consumed_permille: Option<u64>, color: u32) -> impl IntoElement {
+    let Some(permille) = consumed_permille else {
+        // No capsule. A full-length empty gauge reads as "plenty left" from
+        // across the room, which is the opposite of the truth. A hairline is
+        // visibly not a measurement.
+        return div()
+            .w(px(80.0))
+            .h(px(2.0))
+            .flex_none()
+            .bg(rgb(BORDER))
+            .into_any_element();
+    };
+    let clamped = u16::try_from(permille.min(1_000)).unwrap_or(1_000);
+    div()
+        .w(px(80.0))
+        .h(px(8.0))
+        .flex_none()
+        .rounded(px(4.0))
+        .bg(rgb(BORDER_QUIET))
+        .overflow_hidden()
+        .child(
+            div()
+                .h_full()
+                .w(relative(f32::from(clamped) / 1_000.0))
+                .bg(rgb(color)),
+        )
+        .into_any_element()
+}
+
+fn usage_segment(segment: UsageSegment, selected: bool) -> impl IntoElement {
+    let color = provenance_color(segment.provenance);
+    let name_color = if selected { TEXT } else { MUTED };
+    div()
+        .flex()
+        .items_center()
+        .gap_2()
+        .when(selected, |view| {
+            view.px_2().py_1().rounded(px(4.0)).bg(rgb(PANEL_ACTIVE))
+        })
+        .child(
+            div()
+                .flex_none()
+                .text_color(rgb(name_color))
+                .child(segment.label),
+        )
+        .child(usage_meter(segment.consumed_permille, color))
+        .child(
+            div()
+                .flex_none()
+                .text_color(rgb(if segment.headline.is_some() {
+                    TEXT
+                } else {
+                    MUTED
+                }))
+                .child(segment.headline.unwrap_or_else(|| "no reading".to_owned())),
+        )
+        .when_some(segment.reset, |view, reset| {
+            view.child(div().flex_none().text_color(rgb(MUTED)).child(reset))
+        })
+        // Only the harness you are looking at spends footer width on its rate
+        // and its trust. The rest stay one glance wide.
+        .when(selected, |view| {
+            view.child(footer_separator())
+                .child(
+                    div()
+                        .flex_none()
+                        .text_color(rgb(provenance_color(segment.burn_provenance)))
+                        .child(segment.burn),
+                )
+                .child(provenance_chip(&segment.trust, segment.provenance))
+        })
+}
+
+/// One host or endpoint row in the sidebar's runtime list.
+struct RuntimeRow {
+    label: &'static str,
+    detail: String,
+    tone: u32,
+}
+
+fn footer_separator() -> impl IntoElement {
+    div().text_color(rgb(BORDER)).child("│")
+}
+
+/// The trust marker. Its colour is the provenance, never the value.
+fn provenance_chip(text: &str, provenance: UsageProvenance) -> impl IntoElement {
+    let color = match provenance {
+        UsageProvenance::ProviderReported => SUCCESS,
+        UsageProvenance::HarnessReported => ACCENT,
+        UsageProvenance::LocallyMeasured => TEXT,
+        UsageProvenance::Estimated => ATTENTION,
+        UsageProvenance::Unavailable => MUTED,
+    };
+    div()
+        .px_2()
+        .py(px(1.0))
+        .rounded(px(3.0))
+        .border_1()
+        .border_color(rgb(color))
+        .text_color(rgb(color))
+        .child(text.to_owned())
 }
 
 impl Render for LumbridgeShell {
@@ -1827,6 +2280,7 @@ impl Render for LumbridgeShell {
             "{running_runtime_count}/{} LIVE PTYS",
             self.live_terminals.len()
         );
+        let attention_panels = self.attention_panels();
         let sidebar = div()
             .flex()
             .flex_col()
@@ -1841,11 +2295,30 @@ impl Render for LumbridgeShell {
                     .pt_4()
                     .pb_2()
                     .text_xs()
-                    .text_color(rgb(ATTENTION))
+                    .text_color(rgb(if attention_panels.is_empty() {
-                    .child("ATTENTION · 1"),
+                        MUTED
+                    } else {
+                        ATTENTION
+                    }))
+                    .child(format!("ATTENTION · {}", attention_panels.len())),
             )
-            .child(
+            .when(attention_panels.is_empty(), |view| {
+                view.child(
+                    div()
+                        .mx_2()
+                        .mb_3()
+                        .px_3()
+                        .py_2()
+                        .text_xs()
+                        .text_color(rgb(MUTED))
+                        .child("No pane is waiting on you"),
+                )
+            })
+            .children(attention_panels.into_iter().map(|pane| {
+                let id = pane.id;
                 div()
+                    .id(("attention-panel", id.get()))
+                    .cursor_pointer()
                     .mx_2()
                     .mb_3()
                     .px_3()
@@ -1854,20 +2327,19 @@ impl Render for LumbridgeShell {
                     .bg(rgb(PANEL_ACTIVE))
                     .border_1()
                     .border_color(rgb(ATTENTION))
-                    .child(
+                    .hover(|view| view.bg(rgb(PANEL_ALT)))
-                        div()
+                    .child(div().text_sm().text_color(rgb(TEXT)).child(pane.title))
-                            .text_sm()
-                            .text_color(rgb(TEXT))
-                            .child("Claude Code · UI"),
-                    )
                     .child(
                         div()
                             .mt_1()
                             .text_xs()
                             .text_color(rgb(MUTED))
-                            .child("Waiting for a split decision"),
+                            .child(pane.target),
-                    ),
+                    )
-            )
+                    .on_click(cx.listener(move |shell, _, window, cx| {
+                        shell.select_pane(id, window, cx);
+                    }))
+            }))
             .child(
                 div()
                     .px_4()
@@ -1877,18 +2349,20 @@ impl Render for LumbridgeShell {
                     .child("WORKTREES"),
             )
             .children(WORKSPACES.into_iter().enumerate().map(|(index, name)| {
+                let active = index == self.active_worktree;
                 div()
+                    .id(("worktree", index as u64))
+                    .cursor_pointer()
                     .mx_2()
                     .mb_1()
                     .px_3()
                     .py_2()
                     .rounded(px(5.0))
-                    .when(index == 0, |view| {
+                    .when(active, |view| view.bg(rgb(PANEL_ALT)).text_color(rgb(TEXT)))
-                        view.bg(rgb(PANEL_ALT)).text_color(rgb(TEXT))
+                    .when(!active, |view| view.text_color(rgb(MUTED)))
-                    })
+                    .hover(|view| view.bg(rgb(PANEL_ACTIVE)).text_color(rgb(TEXT)))
-                    .when(index != 0, |view| view.text_color(rgb(MUTED)))
                     .child(name)
-                    .when(index == 0, |view| {
+                    .when(active, |view| {
                         view.child(
                             div()
                                 .mt_1()
@@ -1897,6 +2371,9 @@ impl Render for LumbridgeShell {
                                 .child("main · metal"),
                         )
                     })
+                    .on_click(cx.listener(move |shell, _, window, cx| {
+                        shell.select_worktree(index, window, cx);
+                    }))
             }))
             .child(
                 div()
@@ -1907,30 +2384,13 @@ impl Render for LumbridgeShell {
                     .text_color(rgb(MUTED))
                     .child("RUNTIMES"),
             )
-            .child(
+            .children(self.runtime_rows().into_iter().map(|row| {
                 div()
                     .px_4()
                     .py_1()
-                    .text_sm()
+                    .child(div().text_sm().text_color(rgb(row.tone)).child(row.label))
-                    .text_color(rgb(SUCCESS))
+                    .child(div().text_xs().text_color(rgb(MUTED)).child(row.detail))
-                    .child(format!("metal · {runtime_summary}")),
+            }))
-            )
-            .child(
-                div()
-                    .px_4()
-                    .py_1()
-                    .text_sm()
-                    .text_color(rgb(MUTED))
-                    .child("amd-server · connected"),
-            )
-            .child(
-                div()
-                    .px_4()
-                    .py_1()
-                    .text_sm()
-                    .text_color(rgb(MUTED))
-                    .child("spark-1 · sleeping"),
-            )
             .child(
                 div()
                     .px_4()
@@ -1959,6 +2419,7 @@ impl Render for LumbridgeShell {
                     .rounded(px(4.0))
                     .border_1()
                     .border_color(rgb(BORDER_QUIET))
+                    .hover(|view| view.border_color(rgb(ACCENT)).text_color(rgb(TEXT)))
                     .text_xs()
                     .text_color(rgb(MUTED))
                     .child(format!("↪ {title}"))
@@ -1994,6 +2455,8 @@ impl Render for LumbridgeShell {
             .bg(rgb(PANEL))
             .border_b_1()
             .border_color(rgb(BORDER_QUIET))
+            // One tab, because there is one workspace. Two more used to sit
+            // here looking like navigation and doing nothing.
             .child(
                 div()
                     .h_full()
@@ -2003,21 +2466,7 @@ impl Render for LumbridgeShell {
                     .border_b_2()
                     .border_color(rgb(ACCENT))
                     .text_sm()
-                    .child("Agent workspace"),
+                    .child(WORKSPACES[self.active_worktree]),
-            )
-            .child(
-                div()
-                    .px_3()
-                    .text_sm()
-                    .text_color(rgb(MUTED))
-                    .child("Architecture.md"),
-            )
-            .child(
-                div()
-                    .px_3()
-                    .text_sm()
-                    .text_color(rgb(MUTED))
-                    .child("Review"),
             )
             .child(div().flex_1())
             .child(div().px_3().text_xs().text_color(rgb(MUTED)).child(format!(
@@ -2145,7 +2594,7 @@ impl Render for LumbridgeShell {
                             .ml_3()
                             .text_sm()
                             .text_color(rgb(MUTED))
-                            .child("Lumbridge Code / main"),
+                            .child(format!("{} / main", WORKSPACES[self.active_worktree])),
                     )
                     .child(div().flex_1())
                     .child(
@@ -2180,17 +2629,26 @@ impl Render for LumbridgeShell {
                     .flex()
                     .items_center()
                     .justify_between()
-                    .h(px(32.0))
+                    .h(px(APP_FOOTER_HEIGHT))
                     .flex_none()
                     .px_3()
+                    .gap_4()
                     .bg(rgb(PANEL_ALT))
                     .border_t_1()
                     .border_color(rgb(BORDER_QUIET))
                     .text_xs()
                     .text_color(rgb(MUTED))
-                    .child(footer_left)
+                    .child(
-                    .child(timing)
+                        div()
-                    .child(FOOTER_RIGHT),
+                            .flex()
+                            .flex_col()
+                            .flex_none()
+                            .min_w_0()
+                            .child(div().truncate().child(footer_left))
+                            .child(div().truncate().child(timing)),
+                    )
+                    .child(div().flex_1())
+                    .child(self.footer_usage_zone()),
             )
             .when(self.model.command_palette().is_open(), |view| {
                 view.child(self.command_palette())

spikes/gpui-shell/src/usage_feed.rs

@@ -0,0 +1,555 @@
+//! Footer usage for the spike, fed by real probes through the real ledger.
+//!
+//! Profiles are *declared*; numbers are not. A profile exists as soon as the
+//! shell knows a pane is running a harness, so the footer can name the account
+//! it cannot read. Observations only ever come from a [`UsageProbe`]. A harness
+//! with no adapter yet therefore renders the honest-gap path under its own
+//! name — "Claude Code · Anthropic · …  usage unavailable" — rather than
+//! borrowing another pane's number or showing a zero.
+//!
+//! The only adapter that exists today is the Codex quota probe. It is started
+//! at launch unless `LUMBRIDGE_CODEX_PROBE=0` is set, and if the `codex` binary
+//! is absent the probe simply fails to start and its profiles keep rendering as
+//! unavailable.
+
+use std::collections::BTreeMap;
+
+use lumbridge_core::{
+    AccountProfile, AccountProfileId, FooterUsage, UsageLedger, UsageProjection, UsageProvenance,
+    format_duration_ms,
+};
+use lumbridge_harness::claude::{ClaudeCodeProbe, ClaudeCodeProbeOptions};
+use lumbridge_harness::codex::{CodexProbe, CodexProbeOptions};
+use lumbridge_harness::{MonotonicWallClock, ProbeHealth, UsageProbe};
+
+use crate::panel_registry::SeedPane;
+
+/// Harnesses the seeded panes run, whether or not an adapter exists for them.
+struct DeclaredProfile {
+    seed: SeedPane,
+    id: &'static str,
+    harness: &'static str,
+    provider: &'static str,
+    model: &'static str,
+    account: &'static str,
+}
+
+/// The Codex pane maps onto the probe's own primary-window profile, so a live
+/// reading lands under the pane that produced it. The rest are declarations
+/// with no adapter behind them yet.
+const DECLARED: &[DeclaredProfile] = &[
+    DeclaredProfile {
+        seed: SeedPane::CodexRuntime,
+        id: "codex-app-server-primary",
+        harness: "Codex",
+        provider: "ChatGPT",
+        model: "primary window",
+        account: "subscription",
+    },
+    // The Claude pane maps to the five-hour window, not the token total: it is
+    // the limit that actually stops work, and it is what "how much do I have
+    // left" means. The transcript total is still declared by the probe and
+    // still shows in the strip once it reports.
+    DeclaredProfile {
+        seed: SeedPane::ClaudeUi,
+        id: "claude-code-five-hour",
+        harness: "Claude Code",
+        provider: "Anthropic",
+        model: "five-hour window",
+        account: "subscription",
+    },
+    DeclaredProfile {
+        seed: SeedPane::PiDocs,
+        id: "spark-1-local",
+        harness: "Pi",
+        provider: "spark-1",
+        model: "laguna-s-2.1",
+        account: "self-hosted",
+    },
+];
+
+pub(crate) struct UsageFeed {
+    ledger: UsageLedger,
+    clock: MonotonicWallClock,
+    profiles: BTreeMap<AccountProfileId, AccountProfile>,
+    seeds: Vec<(SeedPane, AccountProfileId)>,
+    probes: Vec<Box<dyn UsageProbe>>,
+    health: BTreeMap<AccountProfileId, ProbeHealth>,
+    probe_status: String,
+}
+
+impl UsageFeed {
+    pub(crate) fn start() -> Self {
+        let mut profiles = BTreeMap::new();
+        let mut seeds = Vec::new();
+        for declared in DECLARED {
+            let Ok(profile) = AccountProfile::new(
+                declared.id,
+                declared.harness,
+                declared.provider,
+                declared.model,
+                declared.account,
+            ) else {
+                continue;
+            };
+            seeds.push((declared.seed, profile.id().clone()));
+            profiles.insert(profile.id().clone(), profile);
+        }
+
+        let mut feed = Self {
+            ledger: UsageLedger::new(),
+            clock: MonotonicWallClock::start(),
+            profiles,
+            seeds,
+            probes: Vec::new(),
+            health: BTreeMap::new(),
+            probe_status: "no usage adapter running".to_owned(),
+        };
+        feed.start_codex_probe();
+        feed.start_claude_probe();
+        feed
+    }
+
+    /// Starts the Codex quota probe unless the user opted out.
+    ///
+    /// A failure here is not an error state for the shell: the profiles stay
+    /// declared and render as unavailable, which is exactly what they should do
+    /// when no adapter can run.
+    fn start_codex_probe(&mut self) {
+        if std::env::var_os("LUMBRIDGE_CODEX_PROBE").is_some_and(|value| value == "0") {
+            self.probe_status = "codex probe disabled".to_owned();
+            return;
+        }
+        match CodexProbe::start(CodexProbeOptions::default()) {
+            Ok(probe) => {
+                for profile in probe.profiles() {
+                    self.profiles
+                        .entry(profile.id().clone())
+                        .or_insert_with(|| profile.clone());
+                    self.health
+                        .insert(profile.id().clone(), ProbeHealth::Starting);
+                }
+                self.probes.push(Box::new(probe));
+                self.probe_status = "codex probe starting".to_owned();
+            }
+            Err(error) => {
+                self.probe_status = format!("codex probe unavailable · {error}");
+            }
+        }
+    }
+
+    /// Starts the Claude Code transcript probe unless the user opted out.
+    ///
+    /// This one reads files rather than launching anything, so it has no
+    /// harness to fail to find — a missing projects directory simply reports
+    /// nothing until Claude Code creates it.
+    fn start_claude_probe(&mut self) {
+        if std::env::var_os("LUMBRIDGE_CLAUDE_PROBE").is_some_and(|value| value == "0") {
+            return;
+        }
+        match ClaudeCodeProbe::start(ClaudeCodeProbeOptions::default()) {
+            Ok(probe) => {
+                for profile in probe.profiles() {
+                    self.profiles
+                        .entry(profile.id().clone())
+                        .or_insert_with(|| profile.clone());
+                    self.health
+                        .insert(profile.id().clone(), ProbeHealth::Starting);
+                }
+                self.probes.push(Box::new(probe));
+            }
+            Err(error) => {
+                self.probe_status = format!("claude probe unavailable · {error}");
+            }
+        }
+    }
+
+    /// Which declared profile a seeded panel belongs to.
+    ///
+    /// Panels created at runtime are plain shells with no harness, so they map
+    /// to nothing rather than borrowing another pane's account.
+    pub(crate) fn profile_for_seed(&self, seed: Option<SeedPane>) -> Option<&AccountProfileId> {
+        let seed = seed?;
+        self.seeds
+            .iter()
+            .find(|(candidate, _)| *candidate == seed)
+            .map(|(_, id)| id)
+    }
+
+    /// Drains every probe into the ledger. Returns whether anything changed.
+    pub(crate) fn tick(&mut self) -> bool {
+        let mut changed = false;
+        for probe in &mut self.probes {
+            let outcome = probe.poll();
+            let health = outcome.health();
+            for profile in probe.profiles() {
+                if self.health.insert(profile.id().clone(), health) != Some(health) {
+                    changed = true;
+                }
+            }
+            for observation in outcome.into_observations() {
+                // A rejected observation is a real signal, not a crash: the
+                // ledger refuses anything that would move a profile's stream
+                // backwards. Dropping it preserves the append-only invariant.
+                if self.ledger.record(observation).is_ok() {
+                    changed = true;
+                }
+            }
+        }
+        if changed {
+            self.probe_status = self.summarize_health();
+        }
+        changed
+    }
+
+    fn summarize_health(&self) -> String {
+        if self.probes.is_empty() {
+            return "no usage adapter running".to_owned();
+        }
+        let ready = self
+            .health
+            .values()
+            .filter(|health| matches!(health, ProbeHealth::Ready))
+            .count();
+        let faulted = self
+            .health
+            .values()
+            .filter(|health| health.is_faulted())
+            .count();
+        if faulted > 0 {
+            return format!("{} probes · {faulted} faulted", self.probes.len());
+        }
+        format!("{} probes · {ready} ready", self.probes.len())
+    }
+
+    pub(crate) fn probe_status(&self) -> &str {
+        &self.probe_status
+    }
+
+    fn projection(&self, id: &AccountProfileId) -> UsageProjection {
+        self.ledger.project(id, self.clock.last_emitted_ms())
+    }
+
+    /// Advances the feed's read clock. Kept separate from [`Self::tick`] so
+    /// rendering never mutates the clock mid-frame.
+    pub(crate) fn advance(&mut self) {
+        let _ = self.clock.now_ms();
+    }
+
+    /// Every harness the footer should account for, in a stable order.
+    ///
+    /// The seeded panes come first so the strip does not reorder itself as
+    /// readings arrive, then any probe profile that has actually reported. A
+    /// probe profile with nothing to say — Codex's secondary window on an
+    /// account that has none — is left out rather than shown as an empty rail,
+    /// because the strip is for harnesses the user is running, not for every
+    /// row a probe could theoretically fill.
+    pub(crate) fn strip(&self) -> Vec<UsageSegment> {
+        let mut ordered: Vec<&AccountProfileId> = self.seeds.iter().map(|(_, id)| id).collect();
+        for id in self.profiles.keys() {
+            if !ordered.contains(&id) && self.projection(id).is_available() {
+                ordered.push(id);
+            }
+        }
+        let mut segments: Vec<UsageSegment> = ordered
+            .into_iter()
+            .filter_map(|id| self.segment(id))
+            .collect();
+        // Two windows on one account would otherwise both read "CODEX". A
+        // strip that names two different quotas the same thing is worse than
+        // a longer label.
+        let duplicated: Vec<String> = segments
+            .iter()
+            .filter(|segment| {
+                segments
+                    .iter()
+                    .filter(|other| other.label == segment.label)
+                    .count()
+                    > 1
+            })
+            .map(|segment| segment.label.clone())
+            .collect();
+        for segment in &mut segments {
+            if duplicated.contains(&segment.label) {
+                segment.label = format!("{} {}", segment.label, segment.model.to_uppercase());
+            }
+        }
+        segments
+    }
+
+    fn segment(&self, id: &AccountProfileId) -> Option<UsageSegment> {
+        let profile = self.profiles.get(id)?;
+        let projection = self.projection(id);
+        let footer = FooterUsage::new(profile, &projection);
+        let consumed_permille = projection.consumed_permille();
+        Some(UsageSegment {
+            id: id.clone(),
+            label: profile.harness().to_uppercase(),
+            model: profile.model().to_owned(),
+            consumed_permille,
+            // "How much do I have left" is the question an engineer actually
+            // asks. Consumption stays available in the expanded detail.
+            // With a ceiling, lead with what is left. Without one, the honest
+            // headline is what was spent — a profile that reports real
+            // consumption but no quota must not read as "no reading".
+            //
+            // Show a decimal only when the value actually has one. Codex
+            // reports whole percents, so "81.0% left" would claim a tenth of a
+            // percent of resolution that no one measured.
+            headline: consumed_permille
+                .map(|permille| {
+                    let left = 1_000_u64.saturating_sub(permille);
+                    if left % 10 == 0 {
+                        format!("{}% left", left / 10)
+                    } else {
+                        format!("{}.{}% left", left / 10, left % 10)
+                    }
+                })
+                .or_else(|| {
+                    let consumed = projection.consumed()?;
+                    let unit = projection.unit()?;
+                    Some(format!("{} used", unit.format_amount(consumed)))
+                }),
+            // A profile that reports spend but no ceiling and no reset should
+            // say so where the reset would go, rather than leaving a silent
+            // gap that reads as "we just haven't shown it yet".
+            reset: projection.resets_in_ms().map_or_else(
+                || {
+                    (projection.is_available() && projection.limit().is_none())
+                        .then(|| "no quota reported".to_owned())
+                },
+                |remaining| Some(format!("resets {}", format_duration_ms(remaining))),
+            ),
+            burn: footer.burn(),
+            burn_provenance: projection.burn_provenance(),
+            trust: footer.trust(),
+            provenance: projection.provenance(),
+        })
+    }
+
+    /// Profiles that have produced at least one usable reading.
+    pub(crate) fn reporting_profile_count(&self) -> usize {
+        self.profiles
+            .keys()
+            .filter(|id| self.projection(id).is_available())
+            .count()
+    }
+
+    pub(crate) fn declared_profile_count(&self) -> usize {
+        self.profiles.len()
+    }
+
+    pub(crate) fn shutdown(&mut self) {
+        for probe in &mut self.probes {
+            probe.shutdown();
+        }
+        self.probes.clear();
+    }
+}
+
+impl Drop for UsageFeed {
+    fn drop(&mut self) {
+        self.shutdown();
+    }
+}
+
+/// One harness's standing in the footer strip.
+///
+/// `remaining` and `reset` are `None` when there is nothing to report. That is
+/// deliberately not an empty string: the renderer has to decide what a gap
+/// looks like rather than printing a blank where a number belongs.
+pub(crate) struct UsageSegment {
+    pub(crate) id: AccountProfileId,
+    pub(crate) label: String,
+    pub(crate) model: String,
+    pub(crate) consumed_permille: Option<u64>,
+    /// What to lead with: how much is left when a ceiling is known, how much
+    /// was spent when it is not, and nothing at all when there is no reading.
+    pub(crate) headline: Option<String>,
+    pub(crate) reset: Option<String>,
+    pub(crate) burn: String,
+    pub(crate) burn_provenance: UsageProvenance,
+    pub(crate) trust: String,
+    pub(crate) provenance: UsageProvenance,
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{DECLARED, UsageFeed};
+    use crate::panel_registry::SeedPane;
+    use lumbridge_core::UsageProvenance;
+
+    /// Never starts a probe, so the test cannot depend on an installed harness,
+    /// on the user's transcripts, or on a status-line bridge being present.
+    fn declared_only_feed() -> UsageFeed {
+        // SAFETY-FREE: this only sets environment variables for this process.
+        unsafe {
+            std::env::set_var("LUMBRIDGE_CODEX_PROBE", "0");
+            std::env::set_var("LUMBRIDGE_CLAUDE_PROBE", "0");
+        }
+        UsageFeed::start()
+    }
+
+    #[test]
+    fn a_declared_harness_with_no_adapter_shows_a_gap_under_its_own_name() {
+        let feed = declared_only_feed();
+        let segment = feed
+            .strip()
+            .into_iter()
+            .find(|segment| segment.label.starts_with("CLAUDE CODE"))
+            .expect("Claude Code is declared");
+        assert!(
+            segment.headline.is_none(),
+            "a harness with no adapter must not show a number"
+        );
+        assert_eq!(segment.trust, "no usage source");
+    }
+
+    #[test]
+    fn panels_without_a_seed_have_no_profile() {
+        let feed = declared_only_feed();
+        assert!(feed.profile_for_seed(None).is_none());
+        assert!(
+            feed.profile_for_seed(Some(SeedPane::Architecture))
+                .is_none()
+        );
+    }
+
+    #[test]
+    fn nothing_reports_until_a_probe_produces_a_reading() {
+        let feed = declared_only_feed();
+        assert_eq!(feed.declared_profile_count(), DECLARED.len());
+        assert_eq!(
+            feed.reporting_profile_count(),
+            0,
+            "declaring a profile must not imply a reading"
+        );
+    }
+
+    #[test]
+    fn the_codex_pane_maps_onto_the_probe_profile_id() {
+        let feed = declared_only_feed();
+        let id = feed
+            .profile_for_seed(Some(SeedPane::CodexRuntime))
+            .expect("the Codex pane declares a profile");
+        assert_eq!(
+            id.as_str(),
+            "codex-app-server-primary",
+            "a live reading must land under the pane that produced it"
+        );
+    }
+
+    #[test]
+    fn the_strip_lists_every_declared_harness_even_with_no_readings() {
+        let feed = declared_only_feed();
+        let strip = feed.strip();
+        assert_eq!(strip.len(), DECLARED.len());
+        let labels: Vec<&str> = strip.iter().map(|segment| segment.label.as_str()).collect();
+        assert_eq!(labels, ["CODEX", "CLAUDE CODE", "PI"]);
+        for segment in &strip {
+            assert!(
+                segment.headline.is_none() && segment.reset.is_none(),
+                "a harness with no adapter must report no figure at all"
+            );
+            assert!(!segment.provenance.is_available());
+            assert_eq!(segment.provenance, UsageProvenance::Unavailable);
+        }
+    }
+
+    #[test]
+    fn a_fractional_reading_keeps_its_decimal() {
+        use lumbridge_core::{UsageObservation, UsageUnit, UsageWindow};
+
+        let mut feed = declared_only_feed();
+        let id = feed
+            .profile_for_seed(Some(SeedPane::CodexRuntime))
+            .expect("declared")
+            .clone();
+        feed.advance();
+        let now_ms = feed.clock.last_emitted_ms();
+        feed.ledger
+            .record(
+                UsageObservation::counted(
+                    id.clone(),
+                    UsageUnit::WindowPermille,
+                    185,
+                    UsageProvenance::ProviderReported,
+                    now_ms.saturating_sub(1_000),
+                )
+                .expect("available")
+                .with_limit(1_000)
+                .expect("a positive limit")
+                .with_window(UsageWindow::until(now_ms + 3_600_000))
+                .expect("inside the window"),
+            )
+            .expect("ordered");
+        let segment = feed
+            .strip()
+            .into_iter()
+            .find(|segment| segment.id == id)
+            .expect("present");
+        assert_eq!(segment.headline.as_deref(), Some("81.5% left"));
+    }
+
+    #[test]
+    fn a_segment_reports_what_is_left_not_what_was_used() {
+        use lumbridge_core::{UsageObservation, UsageUnit, UsageWindow};
+
+        let mut feed = declared_only_feed();
+        let id = feed
+            .profile_for_seed(Some(SeedPane::CodexRuntime))
+            .expect("the Codex pane declares a profile")
+            .clone();
+        // The feed reads a real wall clock, so the fixture has to sit inside a
+        // window that is still open now rather than at an arbitrary epoch.
+        feed.advance();
+        let now_ms = feed.clock.last_emitted_ms();
+        feed.ledger
+            .record(
+                UsageObservation::counted(
+                    id.clone(),
+                    UsageUnit::WindowPermille,
+                    180,
+                    UsageProvenance::ProviderReported,
+                    now_ms.saturating_sub(1_000),
+                )
+                .expect("available")
+                .with_limit(1_000)
+                .expect("a positive limit")
+                .with_window(UsageWindow::until(now_ms + 3_600_000))
+                .expect("inside the window"),
+            )
+            .expect("ordered");
+
+        let segment = feed
+            .strip()
+            .into_iter()
+            .find(|segment| segment.id == id)
+            .expect("the Codex segment is present");
+        assert_eq!(
+            segment.headline.as_deref(),
+            Some("82% left"),
+            "a whole-percent source must not render a tenth it never measured"
+        );
+        assert_eq!(segment.consumed_permille, Some(180));
+        assert_eq!(
+            segment.reset.as_deref(),
+            Some("resets 1h 0m"),
+            "a windowed reading reports its reset"
+        );
+    }
+
+    #[test]
+    fn a_segment_with_no_reading_offers_no_figure_to_render() {
+        let feed = declared_only_feed();
+        for segment in feed.strip() {
+            assert!(segment.headline.is_none());
+            assert!(segment.reset.is_none());
+            assert!(
+                !segment.burn.contains("/hr"),
+                "a rate needs readings this segment does not have"
+            );
+            assert_eq!(segment.consumed_permille, None);
+        }
+    }
+}

spikes/ui-shell-model/src/lib.rs

@@ -205,8 +205,17 @@ pub const WORKSPACES: [&str; 5] = [
     "ACP adapters",
     "Usage telemetry",
 ];
+/// Static footer strings retained only by the Floem comparison shell.
+///
+/// These are layout fixtures, not usage. They are not a usage source and must
+/// not be rendered as one: the GPUI shell now derives its footer from
+/// `lumbridge_core::UsageLedger`, where every value carries a provenance and a
+/// missing fact renders as missing. Delete these once Floem consumes the ledger
+/// or the candidate is retired. See decision 0012.
 pub const FOOTER_LEFT: &str = "6 panes · 3 remote · 1 needs input";
+/// See [`FOOTER_LEFT`]: a layout fixture, not a usage reading.
 pub const FOOTER_CENTER: &str = "Codex · ChatGPT subscription · 62% window remaining";
+/// See [`FOOTER_LEFT`]: a layout fixture, not a usage reading.
 pub const FOOTER_RIGHT: &str = "burn 8.4%/hr · resets in 2h 14m";
 
 #[derive(Clone, Debug, Eq, PartialEq)]