fix(cli): correct token context calculation for CLI providers

CLI providers (Claude CLI) report cache_read_input_tokens as the full
cached context for each turn, unlike the embedded/API flow. This change
adds CLI-aware token calculation that uses the correct formula:
cacheRead + cacheWrite + input = total context tokens.

- Add deriveCliContextTokens() for CLI-specific calculation
- Apply CLI detection via isCliProvider() before calculating
- Update session-usage.ts, status.ts, session-utils.fs.ts to use
  CLI-aware calculation
- Add verbose logging for token flow debugging

src/agents/cli-runner.ts

@@ -261,6 +261,13 @@ export async function runCliAgent(params: {
             );
           }
 
+          log.info(
+            `[cli-runner] streaming result usage: input=${streamResult.usage?.input} ` +
+              `output=${streamResult.usage?.output} cacheRead=${streamResult.usage?.cacheRead} ` +
+              `cacheWrite=${streamResult.usage?.cacheWrite} total=${streamResult.usage?.total} ` +
+              `sessionId=${streamResult.sessionId}`,
+          );
+
           return {
             text: streamResult.text,
             sessionId: streamResult.sessionId,
@@ -327,6 +334,12 @@ export async function runCliAgent(params: {
     const text = output.text?.trim();
     const payloads = text ? [{ text }] : undefined;
 
+    log.info(
+      `[cli-runner] FINAL agentMeta.usage: input=${output.usage?.input} ` +
+        `output=${output.usage?.output} cacheRead=${output.usage?.cacheRead} ` +
+        `cacheWrite=${output.usage?.cacheWrite} total=${output.usage?.total}`,
+    );
+
     return {
       payloads,
       meta: {

src/agents/usage.ts

@@ -93,3 +93,22 @@ export function derivePromptTokens(usage?: {
   const sum = input + cacheRead + cacheWrite;
   return sum > 0 ? sum : undefined;
 }
+
+/**
+ * CLI-specific context token calculation.
+ * For CLI, we use the latest turn's values directly (not accumulated across turns)
+ * since cache_read_input_tokens already represents the full cached context.
+ * Formula: cacheRead + cacheWrite + input = total context for this turn
+ */
+export function deriveCliContextTokens(usage?: {
+  input?: number;
+  cacheRead?: number;
+  cacheWrite?: number;
+}): number | undefined {
+  if (!usage) return undefined;
+  const input = usage.input ?? 0;
+  const cacheRead = usage.cacheRead ?? 0;
+  const cacheWrite = usage.cacheWrite ?? 0;
+  const sum = input + cacheRead + cacheWrite;
+  return sum > 0 ? sum : undefined;
+}

src/auto-reply/reply/agent-runner-execution.ts

@@ -211,6 +211,11 @@ export async function runAgentTurnWithFallback(params: {
                 // CLI backends bypass SessionManager, so persist response to transcript
                 // directly to ensure it's available on TUI reload.
                 // Record even empty responses to maintain message turn consistency.
+                const usageForTranscript = result.meta.agentMeta?.usage;
+                logVerbose(
+                  `[agent-runner-execution] CLI transcript persist: usage=${JSON.stringify(usageForTranscript)} ` +
+                    `provider=${provider} model=${model}`,
+                );
                 const assistantPersistResult = appendAssistantMessageToTranscript({
                   message: cliText || "",
                   sessionId: params.followupRun.run.sessionId,
@@ -219,7 +224,7 @@ export async function runAgentTurnWithFallback(params: {
                   createIfMissing: true,
                   provider,
                   model,
-                  usage: result.meta.agentMeta?.usage,
+                  usage: usageForTranscript,
                 });
                 if (!assistantPersistResult.ok) {
                   logVerbose(

src/auto-reply/reply/agent-runner.ts

@@ -41,6 +41,7 @@ import { incrementCompactionCount } from "./session-updates.js";
 import type { TypingController } from "./typing.js";
 import { createTypingSignaler } from "./typing-mode.js";
 import { emitDiagnosticEvent, isDiagnosticsEnabled } from "../../infra/diagnostic-events.js";
+import { logVerbose } from "../../globals.js";
 
 const BLOCK_REPLY_SEND_TIMEOUT_MS = 15_000;
 
@@ -377,6 +378,12 @@ export async function runReplyAgent(params: {
       activeSessionEntry?.contextTokens ??
       DEFAULT_CONTEXT_TOKENS;
 
+    logVerbose(
+      `[agent-runner] usage from runResult: input=${usage?.input} output=${usage?.output} ` +
+        `cacheRead=${usage?.cacheRead} cacheWrite=${usage?.cacheWrite} total=${usage?.total} ` +
+        `provider=${providerUsed} model=${modelUsed} cliSessionId=${cliSessionId}`,
+    );
+
     await persistSessionUsageUpdate({
       storePath,
       sessionKey,

src/auto-reply/reply/session-usage.ts

@@ -1,5 +1,10 @@
 import { setCliSessionId } from "../../agents/cli-session.js";
-import { hasNonzeroUsage, type NormalizedUsage } from "../../agents/usage.js";
+import { isCliProvider } from "../../agents/model-selection.js";
+import {
+  deriveCliContextTokens,
+  hasNonzeroUsage,
+  type NormalizedUsage,
+} from "../../agents/usage.js";
 import {
   type SessionSystemPromptReport,
   type SessionEntry,
@@ -19,10 +24,13 @@ export async function persistSessionUsageUpdate(params: {
   logLabel?: string;
 }): Promise<void> {
   const { storePath, sessionKey } = params;
-  if (!storePath || !sessionKey) return;
+  if (!storePath || !sessionKey) {
+    return;
+  }
 
   const label = params.logLabel ? `${params.logLabel} ` : "";
-  if (hasNonzeroUsage(params.usage)) {
+  const hasUsage = hasNonzeroUsage(params.usage);
+  if (hasUsage) {
     try {
       await updateSessionStoreEntry({
         storePath,
@@ -30,12 +38,18 @@ export async function persistSessionUsageUpdate(params: {
         update: async (entry) => {
           const input = params.usage?.input ?? 0;
           const output = params.usage?.output ?? 0;
-          const promptTokens =
+          const cacheRead = params.usage?.cacheRead ?? 0;
-            input + (params.usage?.cacheRead ?? 0) + (params.usage?.cacheWrite ?? 0);
+          const cacheWrite = params.usage?.cacheWrite ?? 0;
+          // For CLI providers, use the full context for this turn (cacheRead + cacheWrite + input)
+          const isCli = isCliProvider(params.providerUsed ?? "", undefined);
+          const promptTokens = isCli
+            ? (deriveCliContextTokens(params.usage) ?? input)
+            : input + cacheRead + cacheWrite;
+          const newTotalTokens = promptTokens > 0 ? promptTokens : (params.usage?.total ?? input);
           const patch: Partial<SessionEntry> = {
             inputTokens: input,
             outputTokens: output,
-            totalTokens: promptTokens > 0 ? promptTokens : (params.usage?.total ?? input),
+            totalTokens: newTotalTokens,
             modelProvider: params.providerUsed ?? entry.modelProvider,
             model: params.modelUsed ?? entry.model,
             contextTokens: params.contextTokensUsed ?? entry.contextTokens,

src/auto-reply/status.ts

@@ -3,9 +3,14 @@ import fs from "node:fs";
 import { lookupContextTokens } from "../agents/context.js";
 import { DEFAULT_CONTEXT_TOKENS, DEFAULT_MODEL, DEFAULT_PROVIDER } from "../agents/defaults.js";
 import { resolveModelAuthMode } from "../agents/model-auth.js";
-import { resolveConfiguredModelRef } from "../agents/model-selection.js";
+import { isCliProvider, resolveConfiguredModelRef } from "../agents/model-selection.js";
 import { resolveSandboxRuntimeStatus } from "../agents/sandbox.js";
-import { derivePromptTokens, normalizeUsage, type UsageLike } from "../agents/usage.js";
+import {
+  deriveCliContextTokens,
+  derivePromptTokens,
+  normalizeUsage,
+  type UsageLike,
+} from "../agents/usage.js";
 import type { ClawdbotConfig } from "../config/config.js";
 import {
   resolveMainSessionKey,
@@ -29,6 +34,7 @@ import {
   resolveModelCostConfig,
 } from "../utils/usage-format.js";
 import { VERSION } from "../version.js";
+import { logVerbose } from "../globals.js";
 import { listChatCommands, listChatCommandsForConfig } from "./commands-registry.js";
 import { listPluginCommands } from "../plugins/commands.js";
 import type { SkillCommandSpec } from "../agents/skills.js";
@@ -155,6 +161,7 @@ const formatQueueDetails = (queue?: QueueStatus) => {
 const readUsageFromSessionLog = (
   sessionId?: string,
   sessionEntry?: SessionEntry,
+  provider?: string,
 ):
   | {
       input: number;
@@ -165,17 +172,26 @@ const readUsageFromSessionLog = (
     }
   | undefined => {
   // Transcripts are stored at the session file path (fallback: ~/.clawdbot/sessions/<SessionId>.jsonl)
-  if (!sessionId) return undefined;
+  if (!sessionId) {
+    logVerbose(`[status/readUsageFromSessionLog] no sessionId, returning undefined`);
+    return undefined;
+  }
   const logPath = resolveSessionFilePath(sessionId, sessionEntry);
-  if (!fs.existsSync(logPath)) return undefined;
+  logVerbose(`[status/readUsageFromSessionLog] reading transcript: ${logPath}`);
+  if (!fs.existsSync(logPath)) {
+    logVerbose(`[status/readUsageFromSessionLog] transcript file not found`);
+    return undefined;
+  }
 
   try {
     const lines = fs.readFileSync(logPath, "utf-8").split(/\n+/);
+    logVerbose(`[status/readUsageFromSessionLog] transcript has ${lines.length} lines`);
     let input = 0;
     let output = 0;
     let promptTokens = 0;
     let model: string | undefined;
     let lastUsage: ReturnType<typeof normalizeUsage> | undefined;
+    let usageEntriesFound = 0;
 
     for (const line of lines) {
       if (!line.trim()) continue;
@@ -190,18 +206,37 @@ const readUsageFromSessionLog = (
         };
         const usageRaw = parsed.message?.usage ?? parsed.usage;
         const usage = normalizeUsage(usageRaw);
-        if (usage) lastUsage = usage;
+        if (usage) {
+          usageEntriesFound++;
+          logVerbose(
+            `[status/readUsageFromSessionLog] entry #${usageEntriesFound}: ` +
+              `input=${usage.input} output=${usage.output} cacheRead=${usage.cacheRead} ` +
+              `cacheWrite=${usage.cacheWrite} total=${usage.total}`,
+          );
+          lastUsage = usage;
+        }
         model = parsed.message?.model ?? parsed.model ?? model;
       } catch {
         // ignore bad lines
       }
     }
 
-    if (!lastUsage) return undefined;
+    logVerbose(`[status/readUsageFromSessionLog] found ${usageEntriesFound} usage entries`);
+    if (!lastUsage) {
+      logVerbose(`[status/readUsageFromSessionLog] no lastUsage found, returning undefined`);
+      return undefined;
+    }
     input = lastUsage.input ?? 0;
     output = lastUsage.output ?? 0;
-    promptTokens = derivePromptTokens(lastUsage) ?? lastUsage.total ?? input + output;
+    // For CLI providers, use full context for this turn (cacheRead + cacheWrite + input)
+    const isCli = isCliProvider(provider ?? "", undefined);
+    promptTokens = isCli
+      ? (deriveCliContextTokens(lastUsage) ?? lastUsage.total ?? input + output)
+      : (derivePromptTokens(lastUsage) ?? lastUsage.total ?? input + output);
     const total = lastUsage.total ?? promptTokens + output;
+    logVerbose(
+      `[status/readUsageFromSessionLog] lastUsage: input=${lastUsage.input} output=${lastUsage.output} cacheRead=${lastUsage.cacheRead} cacheWrite=${lastUsage.cacheWrite} total=${lastUsage.total} → derived promptTokens=${promptTokens} total=${total}`,
+    );
     if (promptTokens === 0 && total === 0) return undefined;
     return { input, output, promptTokens, total, model };
   } catch {
@@ -299,9 +334,12 @@ export function buildStatusMessage(args: StatusArgs): string {
   // Prefer prompt-size tokens from the session transcript when it looks larger
   // (cached prompt tokens are often missing from agent meta/store).
   if (args.includeTranscriptUsage) {
-    const logUsage = readUsageFromSessionLog(entry?.sessionId, entry);
+    const logUsage = readUsageFromSessionLog(entry?.sessionId, entry, provider);
     if (logUsage) {
       const candidate = logUsage.promptTokens || logUsage.total;
+      logVerbose(
+        `[status] transcript: promptTokens=${logUsage.promptTokens} total=${logUsage.total} candidate=${candidate} currentTotal=${totalTokens} willReplace=${!totalTokens || totalTokens === 0 || candidate > totalTokens}`,
+      );
       if (!totalTokens || totalTokens === 0 || candidate > totalTokens) {
         totalTokens = candidate;
       }

src/gateway/session-utils.fs.ts

@@ -4,7 +4,8 @@ import os from "node:os";
 import path from "node:path";
 
 import { CURRENT_SESSION_VERSION } from "@mariozechner/pi-coding-agent";
-import type { NormalizedUsage } from "../agents/usage.js";
+import { isCliProvider } from "../agents/model-selection.js";
+import { deriveCliContextTokens, type NormalizedUsage } from "../agents/usage.js";
 import { resolveSessionTranscriptPath } from "../config/sessions.js";
 import { stripEnvelope } from "./chat-sanitize.js";
 import type { SessionPreviewItem } from "./session-utils.types.js";
@@ -141,13 +142,21 @@ function appendMessageToTranscriptImpl(
   if (params.role === "assistant") {
     messageBody.stopReason = params.stopReason ?? "cli_backend";
     const u = params.usage;
-    messageBody.usage = {
+    // For CLI providers, use CLI-specific calculation (cacheRead + cacheWrite + input)
+    // representing the full context for this turn
+    const isCli = isCliProvider(params.provider ?? "", undefined);
+    const derivedTotal = isCli
+      ? (deriveCliContextTokens(u) ?? 0)
+      : (u?.input ?? 0) + (u?.cacheRead ?? 0) + (u?.cacheWrite ?? 0);
+    const usageToWrite = {
       input: u?.input ?? 0,
       output: u?.output ?? 0,
       cacheRead: u?.cacheRead,
       cacheWrite: u?.cacheWrite,
-      totalTokens: u?.total ?? (u?.input ?? 0) + (u?.output ?? 0),
+      totalTokens:
+        u?.total ?? (derivedTotal > 0 ? derivedTotal : (u?.input ?? 0) + (u?.output ?? 0)),
     };
+    messageBody.usage = usageToWrite;
     if (params.provider) messageBody.provider = params.provider;
     if (params.model) messageBody.model = params.model;
   }