Surface cache_creation_input_tokens in usage (token-cost diagnostics)

Decode cache_creation_input_tokens from the result event and surface it
alongside cache_read in the spike console and the app transcript row, so a
session's prompt-cache write/read split is visible. Cache writes (billed at a
premium) were previously invisible, making cache-thrash impossible to diagnose
from the UI.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

Sources/NucleicApp/TranscriptRow.swift

@@ -189,6 +189,8 @@ struct TranscriptRow: View {
     private func usageText(_ usage: Usage) -> String {
         var parts: [String] = []
         if let i = usage.inputTokens { parts.append("in \(i)") }
+        if let r = usage.cachedInputTokens, r > 0 { parts.append("cache-read \(r)") }
+        if let w = usage.cacheCreationInputTokens, w > 0 { parts.append("cache-write \(w)") }
         if let o = usage.outputTokens { parts.append("out \(o)") }
         if let cost = usage.costUSD { parts.append(String(format: "$%.4f", cost)) }
         return parts.joined(separator: " · ")

Sources/NucleicCore/Claude/ClaudeStreamDecoder.swift

@@ -308,6 +308,7 @@ public final class ClaudeStreamDecoder {
             usage = Usage(
                 inputTokens: usageObject["input_tokens"]?.intValue,
                 cachedInputTokens: usageObject["cache_read_input_tokens"]?.intValue,
+                cacheCreationInputTokens: usageObject["cache_creation_input_tokens"]?.intValue,
                 outputTokens: usageObject["output_tokens"]?.intValue,
                 reasoningTokens: nil,
                 costUSD: root["total_cost_usd"]?.numberValue)

Sources/NucleicProtocol/AgentEvent.swift

@@ -182,16 +182,23 @@ public struct FileChange: Sendable, Codable, Equatable {
 public struct Usage: Sendable, Codable, Equatable {
     public let inputTokens: Int?
     public let cachedInputTokens: Int?
+    /// Tokens written to the prompt cache this turn (`cache_creation_input_tokens`).
+    /// Billed at a premium; a healthy multi-turn session writes the stable prefix
+    /// once and then reads it. A large value on *every* turn is the signature of a
+    /// prefix that keeps changing (cache thrash) — see token-cost diagnostics.
+    public let cacheCreationInputTokens: Int?
     public let outputTokens: Int?
     public let reasoningTokens: Int?
     public let costUSD: Double?
 
     public init(
-        inputTokens: Int? = nil, cachedInputTokens: Int? = nil, outputTokens: Int? = nil,
+        inputTokens: Int? = nil, cachedInputTokens: Int? = nil,
+        cacheCreationInputTokens: Int? = nil, outputTokens: Int? = nil,
         reasoningTokens: Int? = nil, costUSD: Double? = nil
     ) {
         self.inputTokens = inputTokens
         self.cachedInputTokens = cachedInputTokens
+        self.cacheCreationInputTokens = cacheCreationInputTokens
         self.outputTokens = outputTokens
         self.reasoningTokens = reasoningTokens
         self.costUSD = costUSD

Sources/nucleic-spike/Spike.swift

@@ -189,7 +189,7 @@ struct Spike {
             print("\(prefix) ▶ approval resolved by \(resolved.decidedBy)")
         case .usage(let usage):
             let cost = usage.costUSD.map { String(format: "$%.4f", $0) } ?? "n/a"
-            print("\(prefix) 📊 in=\(usage.inputTokens ?? 0) cached=\(usage.cachedInputTokens ?? 0) out=\(usage.outputTokens ?? 0) cost=\(cost)")
+            print("\(prefix) 📊 in=\(usage.inputTokens ?? 0) cacheRead=\(usage.cachedInputTokens ?? 0) cacheWrite=\(usage.cacheCreationInputTokens ?? 0) out=\(usage.outputTokens ?? 0) cost=\(cost)")
         case .rateLimit(let rl):
             let resets = rl.resetsAt.map { " resets=\($0)" } ?? ""
             print("\(prefix) ⏳ rate-limit \(rl.rateLimitType ?? "?") status=\(rl.status ?? "?")\(resets)")