tokentracker-cli 1.5.1 → 1.5.2

package/src/lib/codex-rollout-parser.js

@@ -34,6 +34,7 @@ const { readCodexServiceTier, isPriorityServiceTier, isUltrafastServiceTier } =
 const {
   createUsageRecordState,
   extractTokenUsageRecord,
+  isCodexTotalOnlyResetSentinel,
   noteLineAfterUsageRecord,
   noteTokenCount,
   noteUsageRecord,
@@ -72,6 +73,11 @@ const CODEX_RESUME_INVALIDATED = "TOKENTRACKER_CODEX_RESUME_INVALIDATED";
 // input (including cache), not the accumulated session or day totals.
 const OPENAI_LONG_CONTEXT_INPUT_THRESHOLD = 272_000;
 
+// Session token summaries use the same fork-prefix rules as rollout.js; the
+// reset-sentinel predicate is shared verbatim from codex-usage-record.js.
+// Context/performance callers retain their existing event view.
+const LEDGER_FORK_REPLAY_GAP_MS = 500;
+
 // ---------------------------------------------------------------------------
 // Timezone helpers
 // ---------------------------------------------------------------------------
@@ -461,7 +467,7 @@ function formatToolDisplayName(name) {
 // Usage normalization
 // ---------------------------------------------------------------------------
 
-function normalizeUsage(u) {
+function normalizeUsage(u, { preserveReportedTotal = false } = {}) {
   const out = {};
   for (const k of [
     "input_tokens",
@@ -477,7 +483,7 @@ function normalizeUsage(u) {
   // Codex reports input inclusive of cached_input_tokens; keep our schema
   // convention: non-cached input and cached input tracked separately.
   out.input_tokens = Math.max(0, out.input_tokens - out.cached_input_tokens);
-  out.total_tokens =
+  if (!preserveReportedTotal) out.total_tokens =
     out.input_tokens +
     out.cached_input_tokens +
     out.cache_creation_input_tokens +
@@ -596,6 +602,7 @@ async function parseCodexRolloutFile(filePath, {
   sourceHandle = null,
   collectBreakdowns = true,
   collectModelUsage = false,
+  ledgerUsage = false,
   onObject = null,
   onUsage = null,
 } = {}) {
@@ -638,6 +645,12 @@ async function parseCodexRolloutFile(filePath, {
   let provider = isResuming ? resumeState.provider || null : null;
   let cliVersion = isResuming ? resumeState.cliVersion || null : null;
   let forkedFromId = isResuming ? resumeState.forkedFromId || null : null;
+  const ledgerRolloutDate = path.basename(primaryFilePath).match(/^rollout-(\d{4}-\d{2}-\d{2})T/)?.[1] || null;
+  let ledgerCurrentDate = null;
+  // Same gating as rollout.js: the replay prefix only ever sits at the head of
+  // a freshly-forked file, so only a scan from offset 0 can see it.
+  let ledgerReplayActive = initialOffset === 0;
+  let ledgerPreviousTokenMs = null;
   let parentThreadId = isResuming ? resumeState.parentThreadId || null : null;
   let threadSource = isResuming ? resumeState.threadSource || null : null;
   let agentNickname = isResuming ? resumeState.agentNickname || null : null;
@@ -1050,6 +1063,9 @@ async function parseCodexRolloutFile(filePath, {
     contentHasher,
     sourceHandle,
   })) {
+    if (ledgerUsage && obj?.type === "turn_context" && typeof obj?.payload?.current_date === "string") {
+      ledgerCurrentDate = obj.payload.current_date.trim().match(/^(\d{4}-\d{2}-\d{2})/)?.[1] || null;
+    }
     const modelEvent = applyCodexModelEvent(modelAttributionState, obj);
     const attributedModel = currentCodexModel(modelAttributionState);
     model = attributedModel || model;
@@ -1114,7 +1130,7 @@ async function parseCodexRolloutFile(filePath, {
     }
 
     if (usageRecord) {
-      const delta = normalizeUsage(usageRecord.usage);
+      const delta = normalizeUsage(usageRecord.usage, { preserveReportedTotal: ledgerUsage });
       noteUsageRecord(
         usageRecordState,
         // Without a response_id a compaction cannot be deduplicated.
@@ -1147,15 +1163,36 @@ async function parseCodexRolloutFile(filePath, {
         );
       }
       const rawDelta = consumeUsageDelta(usageDeltaState, lastUsage, totalUsage);
-      const delta = rawDelta ? normalizeUsage(rawDelta) : null;
+      const delta = rawDelta && !(ledgerUsage && isCodexTotalOnlyResetSentinel(lastUsage, totalUsage))
+        ? normalizeUsage(rawDelta, { preserveReportedTotal: ledgerUsage }) : null;
       // Timing consumers receive canonical, deduplicated usage below. A null
       // delta is a snapshot, not another model response.
       if (compaction) countCompactionRecord(compaction);
       const isNewResumeEvent = noteTokenEvent(ts, lastUsage, totalUsage);
       if (inRequestedRange) {
         const eventSessionId = sessionId || rolloutSessionIdFromPath(primaryFilePath) || primaryFilePath;
-        const eventKey = `${eventSessionId}:${ts}:${usageEventSignature(lastUsage, totalUsage)}`;
-        if (!isNewResumeEvent || (seenTokenEvents && seenTokenEvents.has(eventKey))) {
+        const eventKey = ledgerUsage ? `${eventSessionId}:${ts}`
+          : `${eventSessionId}:${ts}:${usageEventSignature(lastUsage, totalUsage)}`;
+        const alreadySeen = seenTokenEvents && seenTokenEvents.has(eventKey);
+        let replayed = false;
+        if (ledgerUsage && forkedFromId && delta?.total_tokens > 0) {
+          if (ledgerReplayActive) {
+            const tokenMs = Date.parse(ts);
+            if (!Number.isFinite(tokenMs) || (ledgerPreviousTokenMs !== null && tokenMs < ledgerPreviousTokenMs)) {
+              ledgerReplayActive = false;
+            } else {
+              if (ledgerPreviousTokenMs !== null && tokenMs - ledgerPreviousTokenMs >= LEDGER_FORK_REPLAY_GAP_MS) ledgerReplayActive = false;
+              replayed = ledgerReplayActive && ledgerPreviousTokenMs !== null
+                && tokenMs - ledgerPreviousTokenMs < LEDGER_FORK_REPLAY_GAP_MS;
+              ledgerPreviousTokenMs = tokenMs;
+            }
+          }
+          replayed ||= Boolean(ledgerRolloutDate && ledgerCurrentDate && ledgerCurrentDate < ledgerRolloutDate);
+        }
+        // Remember skipped inherited rows too: another readable copy of the
+        // same session must not reintroduce its prefix after the first live turn.
+        if (ledgerUsage && seenTokenEvents && delta?.total_tokens > 0) seenTokenEvents.add(eventKey);
+        if (!isNewResumeEvent || alreadySeen || replayed) {
           if (typeof onUsage === "function") onUsage({ timestamp: ts, delta: null, rawUsage: lastUsage, model: attributedModel });
           attributeTurn(null);
         } else {

package/src/lib/codex-usage-record.js

@@ -176,6 +176,26 @@ function formatRecordOnlyWarning(count) {
   return `${count} Codex session(s) report usage only via token_usage_record; not counted yet (see #652)`;
 }
 
+// A reset sentinel is a token_count whose total drops to exact zero while the
+// previous total was positive and every component of both snapshots is zero —
+// Codex restarting a session's cumulative counters, not a real response. Both
+// the hourly ledger (rollout.js) and the session ledger mode
+// (codex-rollout-parser.js) must reject it identically, so the predicate lives
+// here as the single copy of the rule.
+function isCodexTotalOnlyResetSentinel(lastUsage, totalUsage) {
+  const componentTotal = (usage) => [
+    usage?.input_tokens,
+    usage?.cached_input_tokens,
+    usage?.cache_creation_input_tokens ?? usage?.cache_write_input_tokens,
+    usage?.output_tokens,
+    usage?.reasoning_output_tokens,
+  ].reduce((sum, value) => sum + Math.max(0, Number(value) || 0), 0);
+  return Number(lastUsage?.total_tokens || 0) > 0
+    && Number(totalUsage?.total_tokens || 0) === 0
+    && componentTotal(lastUsage) === 0
+    && componentTotal(totalUsage) === 0;
+}
+
 module.exports = {
   MAX_COMPACTION_RESPONSE_IDS,
   createCompactionResponseIds,
@@ -184,6 +204,7 @@ module.exports = {
   awaitsCompactedLine,
   createUsageRecordState,
   extractTokenUsageRecord,
+  isCodexTotalOnlyResetSentinel,
   isCodexTurnEndEvent,
   isRecordOnly,
   noteLineAfterUsageRecord,

package/src/lib/pricing/curated-overrides.json

@@ -75,6 +75,27 @@
       "cache_write": 6.25,
       "note": "Opus 4.8 not yet in LiteLLM. Pin to the standard Opus tier (same as 4.6/4.7); remove once LiteLLM carries it."
     },
+    "claude-opus-4-5": {
+      "input": 5,
+      "output": 25,
+      "cache_read": 0.5,
+      "cache_write": 6.25,
+      "note": "Canonical undated Opus 4.5 alias (LiteLLM bare key, sourced from https://platform.claude.com/docs/en/about-claude/pricing, cached 2026-10-31): standard Opus tier $5/$25 per MTok, cache_read 0.1x, cache_write 1.25x — same rates as claude-opus-4-5-20250414. Previously LiteLLM-only; promoted so cloud account views price the bare id (issue #929)."
+    },
+    "claude-opus-4-6": {
+      "input": 5,
+      "output": 25,
+      "cache_read": 0.5,
+      "cache_write": 6.25,
+      "note": "Opus 4.6 standard tier (LiteLLM entry sourced from https://platform.claude.com/docs/en/about-claude/pricing, cached 2026-10-31): $5/$25 per MTok, cache_read 0.1x, cache_write 1.25x. The edge table pins the same rate; curated pins it so local billing and -thinking variants no longer depend on LiteLLM fuzzy matching (issue #929)."
+    },
+    "claude-sonnet-4-6": {
+      "input": 3,
+      "output": 15,
+      "cache_read": 0.3,
+      "cache_write": 3.75,
+      "note": "Sonnet 4.6 standard tier (LiteLLM entry sourced from https://platform.claude.com/docs/en/about-claude/pricing, cached 2026-10-31): $3/$15 per MTok, cache_read 0.1x, cache_write 1.25x. The edge table pins the same rate; curated pins it so local billing no longer depends on the LiteLLM cache (issue #929)."
+    },
     "claude-sonnet-5": {
       "input": 2,
       "output": 10,
@@ -137,6 +158,13 @@
       "priority_multiplier": 2,
       "note": "Official Standard USD/MTok rates verified 2026-09-24: https://developers.openai.com/api/docs/models/gpt-6-sol. Same rules as gpt-6-astra: requests above 272K input tokens use 2x input/cache and 1.5x output for the full request, applied only to observed long-context request subsets; Fast mode is published as 2x the applicable rates (priority_multiplier). Reasoning-effort suffixes use the same base rates."
     },
+    "gpt-6-luna": {
+      "input": 0.1,
+      "output": 0.5,
+      "cache_read": 0.01,
+      "cache_write": 0.125,
+      "note": "GPT-6 Luna lightweight tier, Standard USD/MTok (LiteLLM entry sourced from https://developers.openai.com/api/docs/pricing, cached 2026-10-31): $0.10/$0.50 per MTok, $0.01 cache read, $0.125 cache write. Distinct from gpt-5.6-luna. Previously LiteLLM-only, so cloud account views showed it unpriced; promoted so cloud prices it and curated/edge parity guards the key (issue #929)."
+    },
     "gpt-5.6-sol": {
       "input": 4,
       "output": 20,
@@ -158,6 +186,42 @@
       "cache_write": 0.25,
       "note": "GPT-5.6 lightweight tier. Official short-context rates: $0.20/$1.20 per MTok (previous curated values $1/$6 overcounted 5x). See gpt-5.6-sol."
     },
+    "gpt-5.5": {
+      "input": 5,
+      "output": 30,
+      "cache_read": 0.5,
+      "note": "Official Standard USD/MTok rates (LiteLLM entry sourced from https://developers.openai.com/api/docs/pricing, cached 2026-10-31): $5/$30 per MTok, $0.50 cache read, no published cache-write rate. The largest production LiteLLM-only model — codex's primary SKU ($1,118 of cloud account usage showed unpriced between 2026-01-01 and 2026-10-31); promoted so cloud account views price it and curated/edge parity guards the key (issue #929)."
+    },
+    "gpt-5.4": {
+      "input": 2.5,
+      "output": 15,
+      "cache_read": 0.25,
+      "note": "Official Standard USD/MTok rates (LiteLLM entry sourced from https://developers.openai.com/api/docs/pricing, cached 2026-10-31): $2.50/$15 per MTok, $0.25 cache read. The edge table has pinned this rate since before v1.5.1; curated pins the same value so local billing no longer depends on the LiteLLM cache (issue #929). -pro/-nano/-mini keep their own SKUs."
+    },
+    "gpt-5.4-pro": {
+      "input": 30,
+      "output": 180,
+      "cache_read": 3,
+      "note": "Official Pro USD/MTok rates, matching the edge table's gpt-5.4-pro entry ($30/$180, $3 cache read; https://developers.openai.com/api/docs/pricing). Companion pin for the gpt-5.4 promotion: without a curated Pro entry, the curated contained match captures provider-qualified dated snapshots (openai/gpt-5.4-pro-2026-03-05) at the base $2.50/$15 rate before the LiteLLM reverse lookup reaches the Pro SKU — 12x under local vs edge (issue #929)."
+    },
+    "gpt-5.4-mini": {
+      "input": 0.75,
+      "output": 4.5,
+      "cache_read": 0.075,
+      "note": "Official Standard USD/MTok rates (LiteLLM entry sourced from https://developers.openai.com/api/docs/pricing, cached 2026-10-31): $0.75/$4.50 per MTok, $0.075 cache read. Same promotion rationale as gpt-5.4 (issue #929)."
+    },
+    "gpt-5.2": {
+      "input": 1.75,
+      "output": 14,
+      "cache_read": 0.175,
+      "note": "Official Standard USD/MTok rates (LiteLLM entry sourced from https://developers.openai.com/api/docs/pricing, cached 2026-10-31): $1.75/$14 per MTok, $0.175 cache read. Same promotion rationale as gpt-5.4 (issue #929); -pro stays on its own 10x LiteLLM SKU."
+    },
+    "gpt-5.3-codex-spark": {
+      "input": 1.75,
+      "output": 14,
+      "cache_read": 0.175,
+      "note": "Codex spark capacity tier; not a separate LiteLLM key. Local (LiteLLM reverse-substring) and the edge matcher both bill it at the base gpt-5.3-codex rate ($1.75/$14, $0.175 cache read, https://developers.openai.com/api/docs/pricing); the exact entry pins that resolution on both sides (issue #929)."
+    },
     "kiro-agent": {
       "input": 3,
       "output": 15,
@@ -265,6 +329,12 @@
       "cache_write": 0,
       "note": "Google Gemini 4 Argon introductory list rate (through 2026-11-30) matching Cursor Other Models pool: $2/M input, $0.10/M cached input, $10/M output. Effort suffixes (-high/-medium/-low/-minimal) are stripped by normalizeCursorModel. Sources: https://cursor.com/docs/models-and-pricing and https://ai.google.dev/gemini-api/docs/pricing"
     },
+    "gemini-3-flash-preview": {
+      "input": 0.5,
+      "output": 3,
+      "cache_read": 0.05,
+      "note": "LiteLLM entry (sourced from https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing, cached 2026-10-31): $0.50/$3 per MTok, $0.05 cache read, no published cache-write rate. Same rate the edge table already pins; promoted so curated/edge parity guards the key (issue #929)."
+    },
     "MiniMax-M2.7": {
       "input": 0.3,
       "output": 1.2,
@@ -579,6 +649,30 @@
       "output": 0,
       "cache_read": 0
     },
+    "mimo-v2.5": {
+      "input": 0.4,
+      "output": 2,
+      "cache_read": 0.08,
+      "note": "Xiaomi MiMo v2.5 (queue rows report the bare id; LiteLLM keys are provider-qualified). Rates match the deepinfra entry — the lexicographically smallest provider key the matcher deterministically picks: $0.40/$2 per MTok, $0.08 cache read, no published cache-write rate. The edge table pinned this exact rate after the kimi-k2.6/mimo profile-modal incident; curated pins the same value so local billing no longer depends on LiteLLM prefix-strip (issue #929)."
+    },
+    "mimo-v2.5-pro": {
+      "input": 1,
+      "output": 3,
+      "cache_read": 0.2,
+      "note": "Xiaomi MiMo v2.5 Pro (deepinfra is the lexicographically smallest LiteLLM provider key, matching the edge table): $1/$3 per MTok, $0.20 cache read. Exact entry required alongside mimo-v2.5 so the contained-exact fallback cannot bill -pro at the base v2.5 rate (issue #929)."
+    },
+    "muse-spark-1.2-contributor": {
+      "input": 0.1,
+      "output": 0.2,
+      "cache_read": 0.002,
+      "note": "Paid Muse contributor SKU. LiteLLM provider-qualified entries agree on $0.10/$0.20 per MTok, $0.002 cache read (meta and opencode_go both list these rates, cached 2026-10-31). Previously LiteLLM prefix-strip-only, so cloud account views showed it unpriced; promoted so cloud prices it and curated/edge parity guards the key (issue #929). The -free SKU below stays $0."
+    },
+    "muse-spark-1.3-contributor": {
+      "input": 0.1,
+      "output": 0.2,
+      "cache_read": 0.002,
+      "note": "See muse-spark-1.2-contributor — same paid contributor SKU at the same verified rates (issue #929)."
+    },
     "muse-spark-1.2-contributor-free": {
       "input": 0,
       "output": 0,