tokentracker-cli 1.5.16 → 1.5.18

package/src/lib/dsh-session.js

@@ -8,6 +8,7 @@
 const fs = require("node:fs");
 const path = require("node:path");
 const crypto = require("node:crypto");
+const { createDshPerformanceCollector, normalizePerformance } = require("./session-performance");
 const {
   isDshSessionLogName,
   parseDshVersion,
@@ -101,6 +102,48 @@ function eventTimestampMs(entry) {
   return Number.isFinite(time) && time >= MIN_DSH_TIMESTAMP_MS ? time : null;
 }
 
+function firstStreamTokenTime(stream) {
+  // DSH v4 stores raw chunks and packed delta runs together. Only non-empty
+  // deltas (or a name-bearing tool delta) mark a token; block-start does not.
+  for (const record of stream) {
+    if (record?.type === "chunk") {
+      const chunk = record.chunk;
+      const token = ["text-delta", "reasoning-delta"].includes(chunk?.type)
+        ? typeof chunk.text === "string" && chunk.text.length > 0
+        : chunk?.type === "tool-call-delta" && (
+          (typeof chunk.argumentsDelta === "string" && chunk.argumentsDelta.length > 0)
+          || (typeof chunk.name === "string" && chunk.name.length > 0));
+      const time = eventTimestampMs(record);
+      if (token && time !== null) return time;
+    } else if (["text-chunks", "reasoning-chunks", "tool-call-chunks"].includes(record?.type)) {
+      const tool = record.type === "tool-call-chunks";
+      const fragments = tool ? record.args : record.texts;
+      let time = eventTimestampMs({ time: record.time0 });
+      if (time === null || !Array.isArray(fragments)) continue;
+      for (let index = 0; index < fragments.length; index += 1) {
+        if (index > 0) {
+          const gap = record.dt?.[index - 1];
+          if (typeof gap !== "number" || !Number.isFinite(gap)) break;
+          time += gap;
+        }
+        if ((typeof fragments[index] === "string" && fragments[index].length > 0)
+          || (tool && typeof record.name === "string" && record.name.length > 0)) return time;
+      }
+    }
+  }
+  return null;
+}
+
+function messageTiming(data, step, end) {
+  if (!step || step.failed || data?.interrupted || step.turn !== data?.turn || step.step !== data?.step) return null;
+  const stream = Array.isArray(data?.stream) ? data.stream : null;
+  if (stream) {
+    const finish = stream.findLast((record) => record?.type === "chunk" && record.chunk?.type === "finish");
+    if (!finish || !finish.chunk.reason?.kind || ["error", "aborted"].includes(finish.chunk.reason.kind)) return null;
+  }
+  return { start: step.ts, end, firstToken: stream ? firstStreamTokenTime(stream) : null };
+}
+
 function directoryEntryIs(entry, fullPath, kind) {
   const check = kind === "dir" ? entry.isDirectory() : entry.isFile();
   if (check) return true;
@@ -279,6 +322,7 @@ async function parseDshTranscript(filePath) {
     events: [],
   };
   const lines = text.split("\n");
+  let openStep = null;
   for (let index = 0; index < lines.length; index += 1) {
     const line = lines[index];
     if (!line) continue;
@@ -292,6 +336,33 @@ async function parseDshTranscript(filePath) {
     const data = entry.data && typeof entry.data === "object" ? entry.data : null;
     const seq = Number.isFinite(entry.seq) ? entry.seq : index;
     switch (entry.type) {
+      case "step/start": {
+        openStep = Number.isInteger(data?.turn) && Number.isInteger(data?.step)
+          ? { turn: data.turn, step: data.step, ts: eventTimestampMs(entry), failed: false } : null;
+        break;
+      }
+      case "assistant/attempt":
+      case "assistant/chunk": {
+        // A retry's step time includes failed work/backoff, so it cannot be a
+        // clean request sample even if the eventual message reports usage.
+        // v4 records the failed try as assistant/attempt. v1 has no attempt
+        // event: the failed try ends as a top-level assistant/chunk whose
+        // chunk is a finish with an error or aborted reason.
+        const chunk = data?.chunk;
+        const failedFinish = entry.type === "assistant/chunk"
+          && chunk?.type === "finish"
+          && ["error", "aborted"].includes(chunk.reason?.kind);
+        if ((entry.type === "assistant/attempt" || failedFinish)
+          && openStep && openStep.turn === data?.turn && openStep.step === data?.step) {
+          openStep.failed = true;
+        }
+        break;
+      }
+      case "step/end":
+      case "turn/start":
+      case "turn/end":
+        openStep = null;
+        break;
       case "session": {
         transcript.header = entry;
         break;
@@ -329,6 +400,8 @@ async function parseDshTranscript(filePath) {
         const ts = eventTimestampMs(entry);
         const model = modelName(message?.source?.model) || transcript.fallbackModel || "unknown";
         const delta = dshUsageDelta(data?.usage);
+        const timing = messageTiming(data, openStep, ts);
+        openStep = null;
         if (delta) {
           transcript.events.push({
             kind: "usage",
@@ -336,6 +409,7 @@ async function parseDshTranscript(filePath) {
             ts,
             model,
             delta,
+            timing,
           });
         }
         break;
@@ -405,6 +479,7 @@ function buildRow(transcript, owners) {
   const times = new Set();
   const agentTypes = new Map();
   const modelsForEdits = new Map();
+  const performanceCollector = createDshPerformanceCollector();
   let turns = 0;
   let edits = 0;
   let retries = 0;
@@ -451,6 +526,9 @@ function buildRow(transcript, owners) {
       addUsage(modelRow, delta);
       modelRow.usage_events += 1;
       modelUsage.set(model, modelRow);
+      if (event.timing) performanceCollector.consumeUsage({
+        model, outputTokens: delta.output_tokens + delta.reasoning_output_tokens, ...event.timing,
+      });
       if (event.ts != null) {
         const bucketTime = new Date(Math.floor(event.ts / 60000) * 60000).toISOString();
         const key = `${model}\u0000${bucketTime}`;
@@ -478,9 +556,11 @@ function buildRow(transcript, owners) {
     const gap = orderedTimes[index] - orderedTimes[index - 1];
     if (gap > 0 && gap <= DSH_ACTIVE_GAP_MS) activeMs += gap;
   }
+  const performance = performanceCollector.finish();
   const models = [...modelUsage.values()].map((row) => ({
     ...row,
     edit_turns: modelsForEdits.get(row.model) || 0,
+    performance: performance.byModel.get(row.model) || normalizePerformance(),
   }));
   // A session that only handshook (no user turn, no usage) has nothing to say.
   if (!tokens.total_tokens && !turns) return null;
@@ -507,6 +587,7 @@ function buildRow(transcript, owners) {
     started_at: orderedTimes.length ? new Date(orderedTimes[0]).toISOString() : null,
     ended_at: orderedTimes.length ? new Date(orderedTimes.at(-1)).toISOString() : null,
     active_ms: activeMs,
+    performance: performance.performance,
     provenance: {
       source: "local-session-log",
       confidence: "observed",

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

@@ -47,12 +47,61 @@
       "cache_write": 0,
       "note": "Cline's own gateway namespaces (`cline-free/*` free tier, `cline-pass/*` flat-rate Cline Pass). Cline's shipped model catalog declares pricing 0 for every id in both namespaces, so a turn there has no marginal per-token cost — same convention as PI_SUBSCRIPTION_SOURCES. The fuzzy entries at the top of `fuzzy` route both prefixes here BEFORE the model names can match a public rate (`cline-pass/glm-5.3` must not bill GLM-5.3 list price, `cline-free/deepseek-v4.1-flash` must not bill DeepSeek). Verified against Cline 0.0.32 (/Applications/Cline.app model catalog + `usageDelta` cost, which stays 0 for both namespaces). A Cline turn that DOES report a positive per-call cost is honoured first because `cline` is listed in SOURCES_WITH_AUTHORITATIVE_COST in pricing/index.js. Remove or re-point these entries when Cline publishes per-token pricing for either namespace."
     },
+    "claude-fable-5-2": {
+      "input": 8,
+      "output": 40,
+      "cache_read": 0.2,
+      "cache_write": 10,
+      "note": "Claude Fable 5.2 standard tier, verified 2026-11-10: https://platform.claude.com/docs/en/about-claude/pricing and https://platform.claude.com/docs/en/models/fable-5-2/overview. $8 / $40 per MTok input/output. cache_read is the published cache-hit rate ($0.20 / MTok, 0.025x base input \u2014 not the usual 0.1x). cache_write is the published 5-minute cache-write rate ($10 / MTok, 1.25x); the 1-hour write is $16 / MTok and cannot fit this single field. Channel prefixes (anthropic/, bedrock/, vertex_ai/, azure_ai/, openrouter/, github_copilot/, \u2026) do not change the rate. Fast mode is a separate SKU at $16/$80 \u2014 see claude-fable-5-2-fast. Floating -latest aliases are NOT pinned here."
+    },
+    "claude-fable-5.2": {
+      "input": 8,
+      "output": 40,
+      "cache_read": 0.2,
+      "cache_write": 10,
+      "note": "Dotted spelling of claude-fable-5-2 (OpenRouter anthropic/claude-fable-5.2, github_copilot/claude-fable-5.2). Same standard-tier rates, verified 2026-11-10: https://platform.claude.com/docs/en/about-claude/pricing. Its own exact key because normalizeClaudeModel only hyphenates sonnet/opus/haiku, so a dotted fable id never reaches claude-fable-5-2."
+    },
+    "anthropic/claude-fable-5.2": {
+      "input": 8,
+      "output": 40,
+      "cache_read": 0.2,
+      "cache_write": 10,
+      "note": "OpenRouter id for Claude Fable 5.2 (https://openrouter.ai/anthropic/claude-fable-5.2). Same standard-tier rates as claude-fable-5-2, verified 2026-11-10. Kept exact so a lookup without the claude source normalizer (which is what strips provider prefixes) still hits."
+    },
+    "anthropic/claude-fable-5.2-20261109": {
+      "input": 8,
+      "output": 40,
+      "cache_read": 0.2,
+      "cache_write": 10,
+      "note": "OpenRouter dated snapshot of anthropic/claude-fable-5.2 (canonical_slug anthropic/claude-fable-5.2-20261109, released 2026-11-09). Same standard-tier rates, verified 2026-11-10: https://platform.claude.com/docs/en/about-claude/pricing. The Anthropic model page itself publishes no dated snapshot id."
+    },
+    "anthropic/claude-fable-5-2": {
+      "input": 8,
+      "output": 40,
+      "cache_read": 0.2,
+      "cache_write": 10,
+      "note": "Hyphenated provider-qualified id for Claude Fable 5.2 (models.dev anthropic/claude-fable-5-2). Same standard-tier rates, verified 2026-11-10. Channel prefix does not change billing."
+    },
+    "claude-fable-5-2-fast": {
+      "input": 16,
+      "output": 80,
+      "cache_read": 0.4,
+      "cache_write": 20,
+      "note": "Claude Fable 5.2 fast mode, a separate SKU (Foundry deployment id claude-fable-5-2-fast; also speed=fast on the Claude API). Input/output are the published fast rates, 2x standard: $16 / $80 per MTok, verified 2026-11-10 at https://platform.claude.com/docs/en/about-claude/pricing and https://platform.claude.com/docs/en/build-with-claude/fast-mode. Those pages publish no fast-mode cache dollar amounts; they say prompt-caching multipliers apply on top of fast-mode pricing, so cache_read/cache_write use the same Fable 5.2 multipliers as the standard tier (0.025x \u2192 $0.40, 5-minute 1.25x \u2192 $20) rather than a published fast cache price. The 1-hour fast write (2x \u2192 $32) cannot fit this single field. Do not bill this id at the $8/$40 standard rate."
+    },
+    "claude-fable-5.2-fast": {
+      "input": 16,
+      "output": 80,
+      "cache_read": 0.4,
+      "cache_write": 20,
+      "note": "Dotted spelling of claude-fable-5-2-fast (vercel/anthropic/claude-fable-5.2-fast, anthropic/claude-fable-5.2-fast). Same fast-mode rates as claude-fable-5-2-fast, verified 2026-11-10. Its own exact key because the dotted form is not rewritten to the hyphenated one."
+    },
     "claude-fable-5": {
       "input": 10,
       "output": 50,
       "cache_read": 1,
       "cache_write": 12.5,
-      "note": "Fable 5 \u2014 Anthropic's top tier above Opus. Not yet in LiteLLM. $10/$50 per MTok; cache_read 0.1x, cache_write 1.25x. Remove once LiteLLM carries it."
+      "note": "Fable 5 \u2014 Anthropic's top tier above Opus. Not yet in LiteLLM. $10/$50 per MTok; cache_read 0.1x, cache_write 1.25x. Remove once LiteLLM carries it. Fable 5.2 is a cheaper separate SKU \u2014 see claude-fable-5-2."
     },
     "claude-opus-5": {
       "input": 5,
@@ -935,6 +984,22 @@
       "match": "haiku-5.5",
       "ref": "claude-haiku-5.5"
     },
+    {
+      "match": "fable-5-2-fast",
+      "ref": "claude-fable-5-2-fast"
+    },
+    {
+      "match": "fable-5.2-fast",
+      "ref": "claude-fable-5.2-fast"
+    },
+    {
+      "match": "fable-5-2",
+      "ref": "claude-fable-5-2"
+    },
+    {
+      "match": "fable-5.2",
+      "ref": "claude-fable-5.2"
+    },
     {
       "match": "claude-opus-5-fast",
       "ref": "claude-opus-5-fast"

package/src/lib/pricing/matcher.js

@@ -104,8 +104,8 @@ function normalizeClaudeModel(model) {
     .replace(/^-+|-+$/g, "")
     .replace(/-{2,}/g, "-");
 
-  if (/^claude-(sonnet|opus|haiku)-\d+\.\d+/.test(m)) {
-    return m.replace(/^(claude-(?:sonnet|opus|haiku)-\d+)\.(\d+)/, "$1-$2");
+  if (/^claude-(sonnet|opus|haiku|fable)-\d+\.\d+/.test(m)) {
+    return m.replace(/^(claude-(?:sonnet|opus|haiku|fable)-\d+)\.(\d+)/, "$1-$2");
   }
   if (/^(sonnet|opus|haiku)-\d+[.-]\d+/.test(m)) {
     return m
@@ -345,6 +345,17 @@ function lookupPricing(model, { curated, litellm, source } = {}) {
     };
   }
 
+  // Floating Fable aliases (`claude-fable-latest` and every provider's
+  // `…/claude-fable-latest`) retarget on each release. Checked 2026-11-10: the
+  // Anthropic model overview and the Fable 5.2 pages do not document the alias,
+  // so nothing pins it to 5.2. Stop here, after exact lookups, so the
+  // fable-5.2 fuzzy needles and the broad fable matcher cannot price it, but
+  // an explicit exact entry added later still wins. Re-check on the next
+  // Fable release before pointing it anywhere.
+  if (/(?:^|\/)claude-fable-latest(?:$|[^a-z0-9])/.test(lower)) {
+    return { hit: false, source: "miss", value: null };
+  }
+
   // Kimi Code K2.8 Preview aliases must not inherit fuzzy `kimi` → kimi-k2.5 rates (#766).
   // Parser stores bare `k2.8`; raw `kimicoding` / `kimi-k2.8` / `kimicoding/k2.8` may still
   // appear in lookups. `kimi-for-coding` keeps its curated exact entry (resolved above).
@@ -382,10 +393,25 @@ function lookupPricing(model, { curated, litellm, source } = {}) {
       if (!match || !ref) continue;
       const needle = match.toLowerCase();
       if (!curated.exact[ref]) continue;
-      if (lower.includes(needle)) {
+      // The Fable 5.2 needles are version-bounded. A plain substring lets the
+      // dot-restored form of a dated Fable 5 id through: claude-fable-5-20250901
+      // becomes claude-fable-5.20250901, which contains "fable-5.2". fable-5-20
+      // contains "fable-5-2" the same way. The minor must end the id or be
+      // followed by a non-digit, so -20261109 and -20 stay off the 5.2 rate.
+      const bounded = /^fable-5[.-]2(?:-fast)?$/.test(needle)
+        ? new RegExp(`${needle.replace(".", "\\.")}(?:$|[^0-9])`)
+        : null;
+      const matches = (value) => (bounded ? bounded.test(value) : value.includes(needle));
+      if (matches(lower)) {
+        return { hit: true, source: "curated:fuzzy", value: curated.exact[ref] };
+      }
+      if (dotForm && matches(dotForm)) {
         return { hit: true, source: "curated:fuzzy", value: curated.exact[ref] };
       }
-      if (dotForm && dotForm.includes(needle)) {
+      // Display names ("Claude Fable 5.2", "Fable 5.2 Fast") only reach this
+      // scan when no source normalizer ran. Fold spaces for the pinned 5.2
+      // needles only — a general fold would price unrelated display names.
+      if (bounded && bounded.test(lower.replace(/[^a-z0-9.]+/g, "-"))) {
         return { hit: true, source: "curated:fuzzy", value: curated.exact[ref] };
       }
     }