fix: truncate billing replay at compaction boundary, clamp exact overshoot

src/billing.go

@@ -71,8 +71,15 @@ func computeBillingSnapshot(fullEntries, turnEntries []TranscriptEntry) map[stri
 	acc := map[billingKey]*billingTier{}
 	totals := billingTier{}
 	for _, call := range calls {
+		// After a compaction the request no longer contains the pre-compact
+		// conversation — only the summary entry and what follows. Replaying
+		// from the start would lay out content that isn't in the request,
+		// and the (usage-derived, never-rescaled) assistant pieces would
+		// overflow the positional cut into the fresh-input tier.
+		history := fullEntries[:call.entryIdx]
+		history = history[compactReplayStart(history):]
 		pieces := append(append([]billingPiece{}, staticPieces...),
-			conversationPieces(fullEntries[:call.entryIdx], skillBodyNames, toolNames)...)
+			conversationPieces(history, skillBodyNames, toolNames)...)
 		pieces = reconcileToUsage(pieces, float64(call.read+call.write+call.fresh))
 		cutByPosition(pieces, float64(call.read), float64(call.write), acc)
 
@@ -87,6 +94,22 @@ func computeBillingSnapshot(fullEntries, turnEntries []TranscriptEntry) map[stri
 	return renderBillingSnapshot(len(calls), totals, acc, counts)
 }
 
+// compactReplayStart returns the index in entries where the live request
+// content begins: the entry AFTER the last compact boundary (the summary
+// user entry itself is in the request, so it is included). 0 when the
+// session has never compacted.
+func compactReplayStart(entries []TranscriptEntry) int {
+	start := 0
+	for i, e := range entries {
+		if e.Type == "system" && e.Subtype == "compact_boundary" {
+			start = i + 1
+		} else if e.IsCompactSummary {
+			start = i
+		}
+	}
+	return start
+}
+
 type billingCall struct {
 	entryIdx                   int // index in fullEntries of the call's FIRST entry: its request is the prefix before it
 	entryEnd                   int // one past the call's LAST entry
@@ -226,6 +249,17 @@ func conversationPieces(entries []TranscriptEntry, skillBodyNames map[string]str
 			if e.Message == nil {
 				continue
 			}
+			if e.IsCompactSummary {
+				// The summary stands in for the compacted conversation — it
+				// is session-length cost, not something the user typed.
+				for _, c := range e.Message.Content {
+					if c.Type == "text" {
+						add("prior_assistant", "compact_summary", kindUsage,
+							float64(measuredOrEstimate(c.Text, "prose")), false)
+					}
+				}
+				continue
+			}
 			for _, c := range e.Message.Content {
 				switch c.Type {
 				case "text":
@@ -359,9 +393,13 @@ func toolLane(name string) (string, string) {
 	}
 }
 
-// reconcileToUsage makes Σ pieces == total exactly. Overshoot shrinks only
-// the estimated pieces (usage-derived ones are already exact); undershoot
-// appends the explicit `unattributed` tail piece.
+// reconcileToUsage makes Σ pieces == total exactly. Overshoot shrinks the
+// estimated pieces first (usage-derived ones are normally already exact);
+// if the usage-derived pieces alone still exceed the measured total — the
+// request dropped content we can't see (compaction we failed to detect,
+// context editing) — they are scaled down too: the per-call exactness
+// contract outranks per-piece exactness. Undershoot appends the explicit
+// `unattributed` tail piece.
 func reconcileToUsage(pieces []billingPiece, total float64) []billingPiece {
 	sum, estSum := 0.0, 0.0
 	for _, p := range pieces {
@@ -370,16 +408,27 @@ func reconcileToUsage(pieces []billingPiece, total float64) []billingPiece {
 			estSum += p.tokens
 		}
 	}
+	exactSum := sum - estSum
 	switch {
-	case sum > total && estSum > 0:
-		target := total - (sum - estSum)
-		if target < 0 {
-			target = 0
+	case sum > total:
+		if estSum > 0 {
+			target := total - exactSum
+			if target < 0 {
+				target = 0
+			}
+			scale := target / estSum
+			for i := range pieces {
+				if !pieces[i].exact {
+					pieces[i].tokens *= scale
+				}
+			}
 		}
-		scale := target / estSum
-		for i := range pieces {
-			if !pieces[i].exact {
-				pieces[i].tokens *= scale
+		if exactSum > total {
+			scale := total / exactSum
+			for i := range pieces {
+				if pieces[i].exact {
+					pieces[i].tokens *= scale
+				}
 			}
 		}
 	case sum < total:
@@ -447,7 +496,7 @@ func countNewEvents(turnEntries []TranscriptEntry, skillBodyNames map[string]str
 	for _, e := range turnEntries {
 		switch e.Type {
 		case "user":
-			if e.Message == nil {
+			if e.Message == nil || e.IsCompactSummary {
 				continue
 			}
 			for _, c := range e.Message.Content {

src/billing_test.go

@@ -260,3 +260,112 @@ func TestDeferredCatalogSplitsBuiltinFromMcp(t *testing.T) {
 		t.Errorf("expected catalog_deltas under mcp_servers: %v", mcp["items"])
 	}
 }
+
+// After /compact the request contains only the summary + post-compact
+// entries. The replay must truncate at the boundary: pre-compact content
+// must not be laid out, and the summary lands in prior_assistant.
+func TestBillingCompactBoundaryTruncatesReplay(t *testing.T) {
+	preCompact := []TranscriptEntry{userPromptEntry("the original ask")}
+	// A huge pre-compact call: its blocks are usage-derived (exact) replay
+	// pieces. Before the fix these overflowed the next call's cut into the
+	// fresh-input tier.
+	preCompact = append(preCompact, assistantCall(t, "m1",
+		&Usage{InputTokens: 500, CacheCreationInputTokens: 10_000, OutputTokens: 80_000},
+		Content{Type: "thinking", Thinking: "redacted"},
+		Content{Type: "text", Text: strings.Repeat("big ", 50)},
+	)...)
+
+	boundary := TranscriptEntry{Type: "system", Subtype: "compact_boundary"}
+	summary := TranscriptEntry{Type: "user", IsCompactSummary: true,
+		Message: &Message{Content: ContentSlice{
+			{Type: "text", Text: strings.Repeat("summary of prior work ", 100)},
+		}}}
+
+	u2 := &Usage{InputTokens: 60, CacheReadInputTokens: 9_000,
+		CacheCreationInputTokens: 400, OutputTokens: 30}
+	post := assistantCall(t, "m2", u2, Content{Type: "text", Text: "continuing"})
+
+	entries := append(append(preCompact, boundary, summary), post...)
+	turn := entries[len(entries)-len(post):]
+
+	snap := computeBillingSnapshot(entries, turn)
+	if snap == nil {
+		t.Fatal("expected billing snapshot")
+	}
+
+	read, write, fresh, output, rows := billingColumnSums(snap)
+	closeEnough := func(got, want int) bool {
+		d := got - want
+		if d < 0 {
+			d = -d
+		}
+		return d <= rows
+	}
+	if !closeEnough(read, u2.CacheReadInputTokens) || !closeEnough(write, u2.CacheCreationInputTokens) ||
+		!closeEnough(fresh, u2.InputTokens) || !closeEnough(output, u2.OutputTokens) {
+		t.Errorf("Σ lanes = read %d / write %d / fresh %d / output %d, want %d/%d/%d/%d (±%d)",
+			read, write, fresh, output, u2.CacheReadInputTokens, u2.CacheCreationInputTokens,
+			u2.InputTokens, u2.OutputTokens, rows)
+	}
+
+	lanes := snap["lanes"].(map[string]interface{})
+	pa, ok := lanes["prior_assistant"].(map[string]interface{})
+	if !ok {
+		t.Fatal("expected prior_assistant lane (compact summary)")
+	}
+	foundSummary := false
+	for _, it := range pa["items"].([]map[string]interface{}) {
+		if it["name"] == "compact_summary" && it["total"].(int) > 0 {
+			foundSummary = true
+		}
+		if it["name"] == "thinking" && it["total"].(int) > 0 {
+			t.Errorf("pre-compact thinking leaked into the replay: %v", it)
+		}
+	}
+	if !foundSummary {
+		t.Errorf("expected compact_summary item in prior_assistant: %v", pa["items"])
+	}
+	if up, ok := lanes["user_prompts"].(map[string]interface{}); ok {
+		if up["total"].(int) > 0 {
+			t.Errorf("pre-compact user prompt leaked into the replay: %v", up)
+		}
+	}
+}
+
+// Safety net: when usage-derived replay pieces alone exceed the call's
+// measured prompt (undetected truncation), they must be scaled down so the
+// per-call exactness contract still holds — the overshoot must never land
+// in the fresh-input tier.
+func TestBillingExactOvershootIsClamped(t *testing.T) {
+	entries := []TranscriptEntry{userPromptEntry("hi")}
+	entries = append(entries, assistantCall(t, "m1",
+		&Usage{InputTokens: 20, OutputTokens: 50_000},
+		Content{Type: "thinking", Thinking: "redacted"},
+		Content{Type: "text", Text: "done"},
+	)...)
+	u2 := &Usage{InputTokens: 40, CacheReadInputTokens: 1_000, OutputTokens: 10}
+	entries = append(entries, assistantCall(t, "m2", u2, Content{Type: "text", Text: "ok"})...)
+
+	snap := computeBillingSnapshot(entries, entries)
+	if snap == nil {
+		t.Fatal("expected billing snapshot")
+	}
+
+	wantRead := u2.CacheReadInputTokens
+	wantFresh := 20 + u2.InputTokens
+	wantOut := 50_000 + u2.OutputTokens
+
+	read, write, fresh, output, rows := billingColumnSums(snap)
+	closeEnough := func(got, want int) bool {
+		d := got - want
+		if d < 0 {
+			d = -d
+		}
+		return d <= rows
+	}
+	if !closeEnough(read, wantRead) || !closeEnough(write, 0) ||
+		!closeEnough(fresh, wantFresh) || !closeEnough(output, wantOut) {
+		t.Errorf("Σ lanes = read %d / write %d / fresh %d / output %d, want %d/0/%d/%d (±%d)",
+			read, write, fresh, output, wantRead, wantFresh, wantOut, rows)
+	}
+}

src/transcript.go

@@ -9,15 +9,19 @@ import (
 )
 
 type TranscriptEntry struct {
-	Type          string         `json:"type"`
-	UUID          string         `json:"uuid"`
-	Timestamp     string         `json:"timestamp"`
-	Slug          string         `json:"slug,omitempty"`
-	AITitle       string         `json:"aiTitle,omitempty"` // populated on type:"ai-title" events
-	Version       string         `json:"version,omitempty"` // Claude Code CLI version stamped on every user/assistant entry
-	Message       *Message       `json:"message,omitempty"`
-	ToolUseResult *ToolUseResult `json:"toolUseResult,omitempty"`
-	Attachment    *Attachment    `json:"attachment,omitempty"`
+	Type      string `json:"type"`
+	Subtype   string `json:"subtype,omitempty"` // e.g. "compact_boundary" on type:"system"
+	UUID      string `json:"uuid"`
+	Timestamp string `json:"timestamp"`
+	Slug      string `json:"slug,omitempty"`
+	AITitle   string `json:"aiTitle,omitempty"` // populated on type:"ai-title" events
+	Version   string `json:"version,omitempty"` // Claude Code CLI version stamped on every user/assistant entry
+	// Set on the user entry that carries the compaction summary — the text
+	// that REPLACES the pre-compact conversation in subsequent requests.
+	IsCompactSummary bool           `json:"isCompactSummary,omitempty"`
+	Message          *Message       `json:"message,omitempty"`
+	ToolUseResult    *ToolUseResult `json:"toolUseResult,omitempty"`
+	Attachment       *Attachment    `json:"attachment,omitempty"`
 }
 
 // Attachment covers the subset of `type:"attachment"` records we extract