fix(tui): gauge shows live window fill, not cumulative session tokens

internal/client/client.go

@@ -38,7 +38,9 @@ type Event struct {
 	Model     string `json:"model"`
 	Sandbox   bool   `json:"sandbox"`
 
-	// done — token economics for the turn and the session
+	// done — token economics for the turn and the session. ContextTokens is the
+	// live window fill for the turn (drops again after history trims); the
+	// Session* pair are cumulative totals that only grow.
 	Latency              float64 `json:"latency"`
 	ContextTokens        int     `json:"contextTokens"`
 	OutputTokens         int     `json:"outputTokens"`

internal/tui/commands.go

@@ -221,13 +221,13 @@ func (m *Model) showStats() {
 		}
 		mean := sumLat / float64(len(m.turnStats))
 
-		ctxVal := th.statsValue.Render(human(m.sessCtxTok))
+		ctxVal := th.statsValue.Render(human(m.winCtxTok))
 		if m.maxContext > 0 {
-			ratio := float64(m.sessCtxTok) / float64(m.maxContext)
+			ratio := float64(m.winCtxTok) / float64(m.maxContext)
 			if ratio > 1 {
 				ratio = 1
 			}
-			ctxVal = th.statsValue.Render(human(m.sessCtxTok)+"/"+humanCtx(m.maxContext)) +
+			ctxVal = th.statsValue.Render(human(m.winCtxTok)+"/"+humanCtx(m.maxContext)) +
 				"  " + m.gaugeColor(ratio).Render(gaugeGlyph(ratio)) +
 				" " + th.statsDim.Render(fmt.Sprintf("%d%%", int(ratio*100+0.5)))
 		}

internal/tui/last_gaps_test.go

@@ -84,7 +84,7 @@ func TestShowStatsBranches(t *testing.T) {
 	m.thinkOn = true
 	m.sessionID = "sess-123"
 	m.maxContext = 100
-	m.sessCtxTok = 250 // 250% → clamped to 100%
+	m.winCtxTok = 250 // 250% → clamped to 100%
 	m.sessionStart = time.Now().Add(-time.Minute)
 	m.turnStats = []turnStats{
 		{latency: 1.0, thought: true},
@@ -304,12 +304,12 @@ func TestCtxGaugeClamps(t *testing.T) {
 	m := newTestModel()
 	m.maxContext = 100
 
-	m.sessCtxTok = -5 // corrupt/negative accounting → clamp to 0%
+	m.winCtxTok = -5 // corrupt/negative accounting → clamp to 0%
 	if out := plain(m.ctxGauge(true)); !strings.Contains(out, "0%") {
 		t.Errorf("negative ratio gauge = %q, want 0%%", out)
 	}
 
-	m.sessCtxTok = 250 // overrun → clamp to 100%
+	m.winCtxTok = 250 // overrun → clamp to 100%
 	if out := plain(m.ctxGauge(false)); !strings.Contains(out, "100%") {
 		t.Errorf("overrun ratio gauge = %q, want 100%%", out)
 	}

internal/tui/model.go

@@ -112,6 +112,7 @@ type Model struct {
 
 	sessCtxTok  int
 	sessOutTok  int
+	winCtxTok   int // live context-window fill: last turn's contextTokens
 	lastLatency float64
 
 	maxContext   int         // active model's context window (0 = unknown → gauge hidden)
@@ -494,6 +495,7 @@ func (m *Model) handleEvent(ev client.Event) (tea.Model, tea.Cmd) {
 		m.status = "ready"
 		m.sessCtxTok = ev.SessionContextTokens
 		m.sessOutTok = ev.SessionOutputTokens
+		m.winCtxTok = ev.ContextTokens
 		m.lastLatency = ev.Latency
 
 	case "error":
@@ -564,6 +566,7 @@ func (m *Model) clearConversation() {
 	m.sessionStart = time.Time{}
 	m.sessCtxTok = 0
 	m.sessOutTok = 0
+	m.winCtxTok = 0
 	m.lastLatency = 0
 	m.refresh()
 }

internal/tui/panels.go

@@ -238,6 +238,7 @@ func (m *Model) handleSessionDetail(msg sessionDetailMsg) {
 	m.sessionStart = time.Time{}
 	m.sessCtxTok = 0
 	m.sessOutTok = 0
+	m.winCtxTok = 0
 	m.lastLatency = 0
 	m.msgs = m.msgs[:0]
 	m.convCount = -1 // transcript swapped for the resumed one — drop the cache

internal/tui/stats_test.go

@@ -78,6 +78,40 @@ func TestNoStatLineWhileStreaming(t *testing.T) {
 	}
 }
 
+// TestGaugeFollowsTurnFill is a regression test: the header gauge must track
+// the last turn's contextTokens (the live window fill, which drops after odek
+// trims history), not sessionContextTokens, which is cumulative and only grows.
+func TestGaugeFollowsTurnFill(t *testing.T) {
+	m := driveTurn(t, client.Event{
+		Type: "done", Latency: 1,
+		ContextTokens: 900, OutputTokens: 100,
+		SessionContextTokens: 50000, SessionOutputTokens: 300,
+	})
+	m.model = "big"
+	m.models = []client.ModelInfo{{ID: "big", MaxContext: 1000}}
+	m.resolveMaxContext()
+	if out := plain(m.header()); !strings.Contains(out, "90%") {
+		t.Errorf("gauge should reflect the turn fill (90%%), got:\n%s", out)
+	}
+
+	// Next turn comes back smaller after a history trim — the gauge drops,
+	// even though the cumulative session total kept growing.
+	m.msgs = append(m.msgs, message{role: roleAsst, streaming: true})
+	m.curIdx = len(m.msgs) - 1
+	m.busy = true
+	m.handleEvent(client.Event{
+		Type: "done", Latency: 1,
+		ContextTokens: 300, OutputTokens: 100,
+		SessionContextTokens: 51200, SessionOutputTokens: 400,
+	})
+	if m.winCtxTok != 300 {
+		t.Fatalf("winCtxTok = %d, want 300", m.winCtxTok)
+	}
+	if out := plain(m.header()); !strings.Contains(out, "30%") {
+		t.Errorf("gauge should drop after a trim (30%%), got:\n%s", out)
+	}
+}
+
 func TestContextGauge(t *testing.T) {
 	m := newTestModel()
 	m.model = "big"
@@ -87,6 +121,7 @@ func TestContextGauge(t *testing.T) {
 		t.Fatalf("maxContext = %d, want 1000", m.maxContext)
 	}
 	m.sessCtxTok = 380
+	m.winCtxTok = 380
 
 	out := plain(m.header())
 	for _, want := range []string{"○", "38%", "380/1k"} {

internal/tui/view.go

@@ -87,13 +87,16 @@ func (m *Model) header() string {
 
 // ctxGauge renders the context-window usage indicator for the header right
 // cluster: a pressure-tinted fill glyph, a percentage, and (when not compact)
-// the used/max fraction. Returns "" when the model's budget is unknown so the
-// header silently keeps its prior shape rather than guessing.
+// the used/max fraction. Usage is the last turn's contextTokens — the live
+// window fill, which drops again after odek trims history — never the
+// cumulative session total, which only grows. Returns "" when the model's
+// budget is unknown so the header silently keeps its prior shape rather than
+// guessing.
 func (m *Model) ctxGauge(compact bool) string {
 	if m.maxContext <= 0 {
 		return ""
 	}
-	ratio := float64(m.sessCtxTok) / float64(m.maxContext)
+	ratio := float64(m.winCtxTok) / float64(m.maxContext)
 	if ratio < 0 {
 		ratio = 0
 	}
@@ -105,7 +108,7 @@ func (m *Model) ctxGauge(compact bool) string {
 	if !compact {
 		// used via human() so it matches the adjacent "∑ ⌂ …" summary; max via
 		// humanCtx() for a tidy whole-k budget.
-		g += " " + m.th.headerMeta.Render(human(m.sessCtxTok)+"/"+humanCtx(m.maxContext))
+		g += " " + m.th.headerMeta.Render(human(m.winCtxTok)+"/"+humanCtx(m.maxContext))
 	}
 	return g
 }