context pressure carries usage and ceiling as named fields

BackendEvent::ContextPressure was (f32, usize) whose second element was the
context limit. A positional tuple crossing a module boundary made every
consumer guess: the TUI guessed limit and was right, the HTTP layer named it
tokens and was wrong. So every non-TUI surface rendered the ceiling as the
usage — a constant 250000 that looked like a measurement. Nothing failed and
nothing logged.

Name the fields, and carry tokens_used explicitly rather than leaving it to
be reconstructed as pressure x limit. bifrost_pressure already computed the
token count and discarded it.

src/api/models.rs

@@ -378,7 +378,11 @@ pub enum StreamEvent {
     #[serde(rename = "compaction_warning")]
     CompactionWarning { pressure: f32, tier: u8 },
     #[serde(rename = "context_pressure")]
-    ContextPressure { pressure: f32, tokens: usize },
+    ContextPressure {
+        pressure: f32,
+        tokens_used: usize,
+        context_limit: usize,
+    },
     #[serde(rename = "inference_strain")]
     InferenceStrain {
         attempt: u32,
@@ -482,7 +486,15 @@ impl From<crate::backend::BackendEvent> for StreamEvent {
                 pressure,
             },
             BE::CompactionWarning { pressure, tier } => Self::CompactionWarning { pressure, tier },
-            BE::ContextPressure(pressure, tokens) => Self::ContextPressure { pressure, tokens },
+            BE::ContextPressure {
+                pressure,
+                tokens_used,
+                context_limit,
+            } => Self::ContextPressure {
+                pressure,
+                tokens_used,
+                context_limit,
+            },
             BE::InferenceStrain {
                 attempt,
                 status,
@@ -578,9 +590,15 @@ impl From<StreamEvent> for crate::backend::BackendEvent {
             StreamEvent::CompactionWarning { pressure, tier } => {
                 BE::CompactionWarning { pressure, tier }
             }
-            StreamEvent::ContextPressure { pressure, tokens } => {
-                BE::ContextPressure(pressure, tokens)
-            }
+            StreamEvent::ContextPressure {
+                pressure,
+                tokens_used,
+                context_limit,
+            } => BE::ContextPressure {
+                pressure,
+                tokens_used,
+                context_limit,
+            },
             StreamEvent::InferenceStrain {
                 attempt,
                 status,
@@ -797,4 +815,43 @@ mod tests {
         assert!(!typed.content.has_images());
         assert!(carrying.content.has_images());
     }
+
+    #[test]
+    fn context_pressure_keeps_usage_and_ceiling_distinct_on_the_wire() {
+        // Regression, 2026-08-12. `BackendEvent::ContextPressure` was a
+        // positional `(f32, usize)` whose second element was the context
+        // *limit*. This layer named that element `tokens`, so every HTTP
+        // surface rendered the ceiling as the usage — a constant 250000
+        // that looked exactly like a measurement. Assert the two numbers
+        // are carried separately and cannot be confused again.
+        let ev = StreamEvent::ContextPressure {
+            pressure: 0.37,
+            tokens_used: 92_500,
+            context_limit: 250_000,
+        };
+        let wire: serde_json::Value = serde_json::to_value(&ev).unwrap();
+        assert_eq!(wire["tokens_used"], 92_500);
+        assert_eq!(wire["context_limit"], 250_000);
+        // The old, ambiguous name must not reappear.
+        assert!(
+            wire.get("tokens").is_none(),
+            "`tokens` was the ambiguous name that caused the bug; \
+             a surface reading it must fail loudly, not read a ceiling"
+        );
+
+        // And it must survive the trip back into the backend enum.
+        let back: crate::backend::BackendEvent = ev.into();
+        match back {
+            crate::backend::BackendEvent::ContextPressure {
+                tokens_used,
+                context_limit,
+                ..
+            } => {
+                assert_eq!(tokens_used, 92_500);
+                assert_eq!(context_limit, 250_000);
+                assert_ne!(tokens_used, context_limit);
+            }
+            other => panic!("expected ContextPressure, got {other:?}"),
+        }
+    }
 }

src/backend/mod.rs

@@ -82,7 +82,19 @@ pub enum BackendEvent {
     /// Continuous context pressure update (sub-threshold).
     /// Fires every round so the TUI ctx counter reflects live state
     /// rather than only updating when a warning crosses a threshold.
-    ContextPressure(f32, usize),
+    ///
+    /// Named fields, deliberately. This was `(f32, usize)` until 2026-08-12,
+    /// and a positional tuple crossing a module boundary made every consumer
+    /// guess what the second element meant. The TUI guessed `limit` (right);
+    /// the HTTP layer guessed `tokens` (wrong), so every non-TUI surface
+    /// rendered the context *ceiling* as if it were the usage — a constant
+    /// 250000 that looked like a measurement. Nothing failed and nothing
+    /// logged. Carry the names.
+    ContextPressure {
+        pressure: f32,
+        tokens_used: usize,
+        context_limit: usize,
+    },
     /// Inference strain — the voice is hoarse, providers are slow.
     /// Correlates to health over time.
     InferenceStrain {

src/server/turn.rs

@@ -28,13 +28,13 @@ fn bifrost_pressure(
     counter: &TokenCounter,
     messages: &[BifrostMessage],
     context_limit: usize,
-) -> f32 {
+) -> (usize, f32) {
     let tokens: usize = messages
         .iter()
         .map(|m| counter.count(&m.content.as_text()))
         .sum();
     let limit = context_limit.max(1);
-    (tokens as f32 / limit as f32).min(1.0)
+    (tokens, (tokens as f32 / limit as f32).min(1.0))
 }
 
 /// Helper: bump adaptive delay when we hit a 429. No decay — once bumped,
@@ -364,7 +364,7 @@ pub(crate) async fn run_turn(
             last_pulse = Instant::now();
         }
 
-        let pressure = bifrost_pressure(&counter, &messages, context_limit);
+        let (tokens_used, pressure) = bifrost_pressure(&counter, &messages, context_limit);
         tracing::info!(
             turn_round = tool_round,
             agent = %agent_id,
@@ -375,7 +375,11 @@ pub(crate) async fn run_turn(
         );
         let max_tokens = pressure_to_max_tokens(pressure, output_limit);
         let _ = tx
-            .send(Ok(BackendEvent::ContextPressure(pressure, context_limit)))
+            .send(Ok(BackendEvent::ContextPressure {
+                pressure,
+                tokens_used,
+                context_limit,
+            }))
             .await;
 
         if max_rounds == 0 {

src/ui/app/mod.rs

@@ -397,8 +397,12 @@ impl App {
                         BackendEvent::CompactionWarning { pressure, tier } => {
                             self.dispatch(TuiEvent::CompactionWarning { pressure, tier });
                         }
-                        BackendEvent::ContextPressure(p, limit) => {
-                            self.dispatch(TuiEvent::PressureChanged(p, limit));
+                        BackendEvent::ContextPressure {
+                            pressure,
+                            context_limit,
+                            ..
+                        } => {
+                            self.dispatch(TuiEvent::PressureChanged(pressure, context_limit));
                         }
                         BackendEvent::InferenceStrain {
                             attempt, status, ..

src/ui/chat/events.rs

@@ -296,11 +296,19 @@ impl ChatState {
                     self.pending_consciousness
                         .push(BackendEvent::CompactionWarning { pressure, tier });
                 }
-                BackendEvent::ContextPressure(p, limit) => {
-                    self.pressure = p;
-                    self.context_limit = Some(limit);
+                BackendEvent::ContextPressure {
+                    pressure,
+                    tokens_used,
+                    context_limit,
+                } => {
+                    self.pressure = pressure;
+                    self.context_limit = Some(context_limit);
                     self.pending_consciousness
-                        .push(BackendEvent::ContextPressure(p, limit));
+                        .push(BackendEvent::ContextPressure {
+                            pressure,
+                            tokens_used,
+                            context_limit,
+                        });
                 }
                 BackendEvent::InferenceStrain {
                     attempt,