feat(cache): add cache token audit logging behind OPENCODE_EXPERIMENTAL_CACHE_AUDIT

packages/core/src/flag/flag.ts

@@ -86,6 +86,7 @@ export const Flag = {
   OPENCODE_EXPERIMENTAL_WORKSPACES: OPENCODE_EXPERIMENTAL || truthy("OPENCODE_EXPERIMENTAL_WORKSPACES"),
   OPENCODE_EXPERIMENTAL_EVENT_SYSTEM: OPENCODE_EXPERIMENTAL || truthy("OPENCODE_EXPERIMENTAL_EVENT_SYSTEM"),
   OPENCODE_EXPERIMENTAL_SESSION_SWITCHING: OPENCODE_EXPERIMENTAL || truthy("OPENCODE_EXPERIMENTAL_SESSION_SWITCHING"),
+  OPENCODE_EXPERIMENTAL_CACHE_AUDIT: truthy("OPENCODE_EXPERIMENTAL_CACHE_AUDIT"),
 
   // Evaluated at access time (not module load) because tests, the CLI, and
   // external tooling set these env vars at runtime.

packages/opencode/src/cli/cmd/tui/feature-plugins/sidebar/context.tsx

@@ -1,7 +1,8 @@
 import type { AssistantMessage } from "@opencode-ai/sdk/v2"
 import type { TuiPlugin, TuiPluginApi } from "@opencode-ai/plugin/tui"
+import { Flag } from "@opencode-ai/core/flag/flag"
 import type { InternalTuiPlugin } from "../../plugin/internal"
-import { createMemo } from "solid-js"
+import { Show, createMemo } from "solid-js"
 
 const id = "internal:sidebar-context"
 
@@ -22,27 +23,54 @@ function View(props: { api: TuiPluginApi; session_id: string }) {
       return {
         tokens: 0,
         percent: null,
+        cacheInput: 0,
+        cacheNew: 0,
+        cacheRead: 0,
+        cacheWrite: 0,
+        cacheHitPercent: null,
+        cacheOutput: 0,
       }
     }
 
     const tokens =
       last.tokens.input + last.tokens.output + last.tokens.reasoning + last.tokens.cache.read + last.tokens.cache.write
     const model = props.api.state.provider.find((item) => item.id === last.providerID)?.models[last.modelID]
+    const cacheInput = last.tokens.input + last.tokens.cache.read + last.tokens.cache.write
+    const cacheHitPercent = cacheInput > 0 ? ((last.tokens.cache.read / cacheInput) * 100).toFixed(1) : null
     return {
       tokens,
       percent: model?.limit.context ? Math.round((tokens / model.limit.context) * 100) : null,
+      cacheInput,
+      cacheNew: last.tokens.input,
+      cacheRead: last.tokens.cache.read,
+      cacheWrite: last.tokens.cache.write,
+      cacheHitPercent,
+      cacheOutput: last.tokens.output,
     }
   })
 
   return (
-    <box>
-      <text fg={theme().text}>
-        <b>Context</b>
-      </text>
-      <text fg={theme().textMuted}>{state().tokens.toLocaleString()} tokens</text>
-      <text fg={theme().textMuted}>{state().percent ?? 0}% used</text>
-      <text fg={theme().textMuted}>{money.format(cost())} spent</text>
-    </box>
+    <>
+      <box>
+        <text fg={theme().text}>
+          <b>Context</b>
+        </text>
+        <text fg={theme().textMuted}>{state().tokens.toLocaleString()} tokens</text>
+        <text fg={theme().textMuted}>{state().percent ?? 0}% used</text>
+        <text fg={theme().textMuted}>{money.format(cost())} spent</text>
+      </box>
+      <Show when={Flag.OPENCODE_EXPERIMENTAL_CACHE_AUDIT && state().cacheHitPercent != null}>
+        <box>
+          <text fg={theme().text}><b>Cache Audit</b></text>
+          <text fg={theme().textMuted}>{state().cacheInput.toLocaleString()} input tokens</text>
+          <text fg={theme().textMuted}>  {state().cacheNew.toLocaleString()} new</text>
+          <text fg={theme().textMuted}>  {state().cacheRead.toLocaleString()} cache read</text>
+          <text fg={theme().textMuted}>  {state().cacheWrite.toLocaleString()} cache write</text>
+          <text fg={theme().textMuted}>{state().cacheHitPercent}% hit rate</text>
+          <text fg={theme().textMuted}>{state().cacheOutput.toLocaleString()} output tokens</text>
+        </box>
+      </Show>
+    </>
   )
 }
 

packages/opencode/src/session/processor.ts

@@ -25,6 +25,7 @@ import { SyncEvent } from "@/sync"
 import { SessionEvent } from "@/v2/session-event"
 import { ModelV2 } from "@opencode-ai/core/model"
 import { ProviderV2 } from "@opencode-ai/core/provider"
+import { Flag } from "@opencode-ai/core/flag/flag"
 import * as DateTime from "effect/DateTime"
 import { RuntimeFlags } from "@/effect/runtime-flags"
 
@@ -510,6 +511,13 @@ export const layer: Layer.Layer<
               usage: value.usage,
               metadata: value.providerMetadata,
             })
+            if (Flag.OPENCODE_EXPERIMENTAL_CACHE_AUDIT) {
+              const totalInputTokens = usage.tokens.input + usage.tokens.cache.read + usage.tokens.cache.write
+              const cacheHitPercent = totalInputTokens > 0 ? ((usage.tokens.cache.read / totalInputTokens) * 100).toFixed(1) : "0.0"
+              slog.info(
+                `[CACHE] ${ctx.model.id}  input=${totalInputTokens} (cache_read=${usage.tokens.cache.read} cache_write=${usage.tokens.cache.write} new=${usage.tokens.input})  hit=${cacheHitPercent}%  output=${usage.tokens.output}  total=${usage.tokens.total ?? 0}`,
+              )
+            }
             if (!ctx.assistantMessage.summary) {
               // TODO(v2): Temporary dual-write while migrating session messages to v2 events.
               if (flags.experimentalEventSystem) {

packages/opencode/src/session/prompt.ts

@@ -1579,7 +1579,7 @@ NOTE: At any point in time through this workflow you should feel free to ask the
             text: nextPrompt.text.join("\n"),
             files: nextPrompt.files,
             agents: nextPrompt.agents,
-            references: nextPrompt.references,
+
           },
         })
       }

packages/opencode/src/session/session.ts

@@ -419,6 +419,7 @@ export const getUsage = (input: { model: Provider.Model; usage: LanguageModelUsa
   }
 
   const contextTokens = inputTokens
+  const rawInputTokens = inputTokens
   const costInfo =
     input.model.cost?.tiers
       ?.filter((item) => item.tier.type === "context" && contextTokens > item.tier.size)
@@ -439,6 +440,7 @@ export const getUsage = (input: { model: Provider.Model; usage: LanguageModelUsa
         .toNumber(),
     ),
     tokens,
+    rawInputTokens,
   }
 }