feat: cache statistics in terminal and WebUI

Add cache metrics (Anthropic cache_creation_input_tokens,
cache_read_input_tokens and OpenAI prompt_tokens_details.cached_tokens)
to the agent runtime with display in both terminal and WebUI.

- llm/client.go: Parse cache metrics from API response (both formats)
- loop/loop.go: Accumulate cache metrics across iterations
- odek.go: Expose TotalCacheCreationTokens/ReadTokens/CachedTokens
- render/render.go: Summary() method after final answer
- serve.go: Send cache metrics in done WebSocket message
- ui/index.html: Show cache stats in per-message and session stats

Author: jkyberneees

cmd/odek/serve.go

@@ -498,6 +498,9 @@ func handlePrompt(
 
 	contextTokens := agent.TotalInputTokens()
 	outputTokens := agent.TotalOutputTokens()
+	cacheCreate := agent.TotalCacheCreationTokens()
+	cacheRead := agent.TotalCacheReadTokens()
+	cached := agent.TotalCachedTokens()
 	*sessionInputTokens += contextTokens
 	*sessionOutputTokens += outputTokens
 
@@ -506,6 +509,9 @@ func handlePrompt(
 		"latency":             latency.Seconds(),
 		"contextTokens":       contextTokens,
 		"outputTokens":        outputTokens,
+		"cacheCreationTokens": cacheCreate,
+		"cacheReadTokens":     cacheRead,
+		"cachedTokens":        cached,
 		"sessionContextTokens": *sessionInputTokens,
 		"sessionOutputTokens": *sessionOutputTokens,
 	})

cmd/odek/ui/index.html

@@ -1399,14 +1399,33 @@ <h3>Delete session?</h3>
             parts.push('⚡ ' + (event.latency < 1 ? (event.latency * 1000).toFixed(0) + 'ms' : event.latency.toFixed(1) + 's'));
             if (event.contextTokens != null) parts.push('⌂ ' + formatNum(event.contextTokens));
             if (event.outputTokens != null) parts.push('⎇ ' + formatNum(event.outputTokens));
+            // Cache metrics — show only when non-zero
+            if (event.cacheCreationTokens > 0) parts.push('⊕ ' + formatNum(event.cacheCreationTokens));
+            if (event.cacheReadTokens > 0) parts.push('⊙ ' + formatNum(event.cacheReadTokens));
+            if (event.cachedTokens > 0) parts.push('⊡ ' + formatNum(event.cachedTokens));
             stats.textContent = parts.join('  ·  ');
             lastAssistant.appendChild(stats);
           }
         }
         // Update session-level token stats in top bar
         const sessionStatsEl = document.getElementById('session-stats');
         if (event.sessionContextTokens != null && event.sessionOutputTokens != null) {
-          sessionStatsEl.textContent = '∑ ⌂ ' + formatNum(event.sessionContextTokens) + '  ·  ⎇ ' + formatNum(event.sessionOutputTokens);
+          const sessParts = ['∑ ⌂ ' + formatNum(event.sessionContextTokens) + '  ·  ⎇ ' + formatNum(event.sessionOutputTokens)];
+          if (event.cacheReadTokens > 0 || event.cacheCreationTokens > 0 || event.cachedTokens > 0) {
+            // Count total session cache stats (accumulated across the session)
+            // We store cache totals on the session-stats element's dataset
+            const el = document.getElementById('session-stats');
+            const cc = (parseInt(el.dataset.cacheCreate || '0') + (event.cacheCreationTokens || 0));
+            const cr = (parseInt(el.dataset.cacheRead || '0') + (event.cacheReadTokens || 0));
+            const cd = (parseInt(el.dataset.cached || '0') + (event.cachedTokens || 0));
+            el.dataset.cacheCreate = cc;
+            el.dataset.cacheRead = cr;
+            el.dataset.cached = cd;
+            if (cr > 0) sessParts.push('⊙ ' + formatNum(cr));
+            if (cc > 0) sessParts.push('⊕ ' + formatNum(cc));
+            if (cd > 0) sessParts.push('⊡ ' + formatNum(cd));
+          }
+          sessionStatsEl.textContent = sessParts.join('  ·  ');
           sessionStatsEl.classList.add('visible');
         }
         if (sessionId) loadSessions();

internal/loop/loop.go

@@ -207,6 +207,9 @@ func (e *Engine) RunWithMessages(ctx context.Context, messages []llm.Message) (s
 	// Reset token accounting for this run
 	e.TotalInputTokens = 0
 	e.TotalOutputTokens = 0
+	e.TotalCacheCreationTokens = 0
+	e.TotalCacheReadTokens = 0
+	e.TotalCachedTokens = 0
 	return e.runLoop(ctx, messages)
 }
 
@@ -281,6 +284,7 @@ func (e *Engine) runLoop(ctx context.Context, messages []llm.Message) (string, [
 		e.TotalOutputTokens += result.OutputTokens
 
 		// Accumulate cache metrics
+		// Accumulate cache metrics across iterations
 		e.TotalCacheCreationTokens += result.CacheCreationTokens
 		e.TotalCacheReadTokens += result.CacheReadTokens
 		e.TotalCachedTokens += result.CachedTokens
@@ -289,6 +293,13 @@ func (e *Engine) runLoop(ctx context.Context, messages []llm.Message) (string, [
 		if len(result.ToolCalls) == 0 {
 			if e.renderer != nil {
 				e.renderer.FinalAnswer(result.Content)
+				e.renderer.Summary(
+					e.TotalInputTokens,
+					e.TotalOutputTokens,
+					e.TotalCacheCreationTokens,
+					e.TotalCacheReadTokens,
+					e.TotalCachedTokens,
+				)
 			}
 			// Append final assistant message so callers (e.g. WebUI) get
 			// the final text in the messages slice and can stream it.

internal/render/render.go

@@ -204,6 +204,33 @@ func (r *Renderer) FinalAnswer(text string) {
 	fmt.Fprintln(r.w)
 }
 
+// Summary prints a run summary line with total token and cache statistics.
+// Emitted after the final answer when at least one stat is non-zero.
+// Shows: total input/output tokens, cache creation/read/cached tokens.
+func (r *Renderer) Summary(inTokens, outTokens, cacheCreate, cacheRead, cached int) {
+	if r.disable() {
+		return
+	}
+	if inTokens == 0 && outTokens == 0 {
+		return
+	}
+	parts := []string{
+		fmt.Sprintf("⌂ %d in", inTokens),
+		fmt.Sprintf("⎇ %d out", outTokens),
+	}
+	if cacheCreate > 0 {
+		parts = append(parts, fmt.Sprintf("⊕ %d created", cacheCreate))
+	}
+	if cacheRead > 0 {
+		parts = append(parts, fmt.Sprintf("⊙ %d read", cacheRead))
+	}
+	if cached > 0 {
+		parts = append(parts, fmt.Sprintf("⊡ %d cached", cached))
+	}
+	fmt.Fprintln(r.w, r.style(gray, "── "+strings.Join(parts, " · ")))
+	fmt.Fprintln(r.w)
+}
+
 // Error prints a non-fatal loop error with a cross emoji.
 func (r *Renderer) Error(err error) {
 	if r.disable() || err == nil {