tokentracker-cli 0.96.1 → 0.96.3

package/src/lib/devin-usage.js

@@ -0,0 +1,212 @@
+"use strict";
+
+// Devin CLI (Cognition — devin.ai) local history projection.
+//
+// Devin persists chat history in a SQLite database at
+// `$XDG_DATA_HOME/devin/cli/sessions.db` (default
+// `~/.local/share/devin/cli/sessions.db`). Token usage lives on assistant
+// message nodes inside the `chat_message` JSON document. The SQL below is a
+// deliberately narrow projection: request identity, generation model, the
+// original generation timestamps, the numeric usage metrics, and the owning
+// session's `working_directory`/`created_at` for canonical ordering and local
+// project attribution. Message bodies, prompt text, node `metadata`,
+// `sessions.cogs_json`, and any credential-bearing state are never selected.
+//
+// Verified against devin CLI 3000.10.21 (611c1cba): replay/fork/compaction
+// produces multiple `message_nodes` rows that share one
+// `metadata.request_id` with identical metrics and original timestamps, so
+// `request_id` — never `row_id` or `(session_id, request_id)` — is the
+// canonical usage identity across the whole local install.
+//
+// Token components are disjoint: the DB's `metrics.input_tokens` already
+// excludes `cache_read_tokens` and `cache_creation_tokens`, so each reported
+// counter is added once and caches are never subtracted from input.
+
+const DEVIN_TABLE_PROBE_SQL =
+  "SELECT name FROM sqlite_master WHERE type='table' AND name IN ('message_nodes','sessions')";
+
+const metricField = (jsonPath, alias) =>
+  `CASE WHEN json_valid(n.chat_message) THEN json_extract(n.chat_message, '${jsonPath}') ELSE NULL END AS ${alias}`;
+
+// `sessions` is LEFT JOINed only for `working_directory` and `created_at`;
+// older databases may not have the table, in which case the projection runs
+// without it (project attribution and session-age ordering degrade to nulls).
+function devinUsageSql({ hasSessionsTable = true } = {}) {
+  const join = hasSessionsTable ? "LEFT JOIN sessions s ON s.id = n.session_id" : "";
+  const sessionCreatedAt = hasSessionsTable ? "s.created_at" : "NULL";
+  const workingDirectory = hasSessionsTable ? "s.working_directory" : "NULL";
+  return `
+    SELECT
+      n.row_id AS row_id,
+      n.session_id AS session_id,
+      ${sessionCreatedAt} AS session_created_at,
+      ${workingDirectory} AS working_directory,
+      ${metricField("$.metadata.request_id", "request_id")},
+      ${metricField("$.metadata.generation_model", "generation_model")},
+      ${metricField("$.metadata.started_generation_at", "started_generation_at")},
+      ${metricField("$.metadata.created_at", "message_created_at")},
+      ${metricField("$.metadata.metrics.input_tokens", "input_tokens")},
+      ${metricField("$.metadata.metrics.output_tokens", "output_tokens")},
+      ${metricField("$.metadata.metrics.cache_read_tokens", "cache_read_tokens")},
+      ${metricField("$.metadata.metrics.cache_creation_tokens", "cache_creation_tokens")}
+    FROM message_nodes n
+    ${join}
+    WHERE json_valid(n.chat_message)
+      AND json_extract(n.chat_message, '$.role') = 'assistant'
+      AND json_extract(n.chat_message, '$.metadata.request_id') IS NOT NULL
+    ORDER BY n.row_id
+  `.trim();
+}
+
+function normalizeText(value) {
+  if (typeof value !== "string") return null;
+  const trimmed = value.trim();
+  return trimmed.length > 0 ? trimmed : null;
+}
+
+// Strict integer check: usage fields must be finite non-negative safe
+// integers. `null`/`undefined` cache counters are legal (no reported cache
+// contribution) — callers pass them through `optionalCount`; required
+// counters go through `requiredCount`.
+function toSafeCount(value) {
+  const n = Number(value);
+  return Number.isSafeInteger(n) && n >= 0 ? n : null;
+}
+
+function requiredCount(value) {
+  // A missing/NULL required counter is incomplete usage, not zero — pending
+  // assistant records carry no metrics block and must stay uncounted until
+  // the writer finalizes them.
+  if (value === null || value === undefined) return null;
+  return toSafeCount(value);
+}
+
+function optionalCount(value) {
+  if (value === null || value === undefined) return 0;
+  return toSafeCount(value);
+}
+
+// RFC3339 timestamps arrive with both `Z` and `+00:00` spellings; Date.parse
+// normalizes both to the same instant. Returns epoch ms or 0 when missing or
+// unparsable.
+function parseDevinTimestampMs(value) {
+  if (typeof value !== "string" || !value.trim()) return 0;
+  const ms = Date.parse(value.trim());
+  return Number.isFinite(ms) && ms > 0 ? ms : 0;
+}
+
+// session.created_at is Unix seconds; node row ordering uses ms.
+function epochSecondsToMs(value) {
+  const n = Number(value);
+  return Number.isFinite(n) && n > 0 ? Math.floor(n) * 1000 : null;
+}
+
+// Normalize one projected row into a canonical candidate, or null when the
+// record is pending/malformed: missing request identity, missing/invalid
+// numeric usage, or no usable original timestamp. Such rows are re-evaluated
+// on every scan, so a record that gains metrics or a timestamp later is
+// counted then.
+function normalizeDevinUsageRow(row) {
+  const requestId = normalizeText(row?.request_id);
+  if (!requestId) return null;
+
+  const input = requiredCount(row?.input_tokens);
+  const output = requiredCount(row?.output_tokens);
+  const cacheRead = optionalCount(row?.cache_read_tokens);
+  const cacheCreation = optionalCount(row?.cache_creation_tokens);
+  if (input === null || output === null || cacheRead === null || cacheCreation === null) {
+    return null;
+  }
+
+  // The original generation time decides the bucket; the node insertion time
+  // (n.created_at) is clone/replay insertion time and is never a fallback.
+  const tsMs =
+    parseDevinTimestampMs(row?.started_generation_at) ||
+    parseDevinTimestampMs(row?.message_created_at);
+  if (!tsMs) return null;
+
+  return {
+    requestId,
+    sessionId: normalizeText(row?.session_id) || "",
+    rowId: toSafeCount(row?.row_id) ?? 0,
+    sessionCreatedAtMs: epochSecondsToMs(row?.session_created_at),
+    workingDirectory: normalizeText(row?.working_directory),
+    model: normalizeText(row?.generation_model) || "unknown",
+    tsMs,
+    input,
+    output,
+    cacheRead,
+    cacheCreation,
+  };
+}
+
+// Pick the canonical retained record for one request_id. Copies created by
+// replay/fork/compaction carry identical original timestamps and metrics, so
+// the order is: earliest original generation time, then the earliest-created
+// owning session (a fork's copy lives in a younger session), then session id,
+// then newest row first so an in-place correction appended as a fresh node
+// inside the same session wins over the stale sibling.
+function compareCandidates(a, b) {
+  if (a.tsMs !== b.tsMs) return a.tsMs - b.tsMs;
+  const sessionA = a.sessionCreatedAtMs ?? Number.POSITIVE_INFINITY;
+  const sessionB = b.sessionCreatedAtMs ?? Number.POSITIVE_INFINITY;
+  if (sessionA !== sessionB) return sessionA - sessionB;
+  if (a.sessionId !== b.sessionId) return a.sessionId < b.sessionId ? -1 : 1;
+  return b.rowId - a.rowId;
+}
+
+// Build the deduplicated per-request event list. Conversation ownership and
+// session/project attribution are decided by the persisted request ledger in
+// parseDevinIncremental — the projection only reports each request's
+// best-evidence retained record; `conversation_count` stays 0 here and is
+// filled from ledger state at reconcile time.
+function buildDevinUsageEvents(rows) {
+  const byRequest = new Map();
+  let rowsSeen = 0;
+  let rowsSkipped = 0;
+  for (const row of Array.isArray(rows) ? rows : []) {
+    rowsSeen += 1;
+    const candidate = normalizeDevinUsageRow(row);
+    if (!candidate) {
+      rowsSkipped += 1;
+      continue;
+    }
+    const list = byRequest.get(candidate.requestId);
+    if (list) list.push(candidate);
+    else byRequest.set(candidate.requestId, [candidate]);
+  }
+
+  const events = [];
+  for (const [requestId, candidates] of byRequest) {
+    candidates.sort(compareCandidates);
+    const canonical = candidates[0];
+    const total = canonical.input + canonical.output + canonical.cacheRead + canonical.cacheCreation;
+    events.push({
+      requestId,
+      sessionId: canonical.sessionId,
+      workingDirectory: canonical.workingDirectory,
+      model: canonical.model,
+      tsMs: canonical.tsMs,
+      totals: {
+        input_tokens: canonical.input,
+        cached_input_tokens: canonical.cacheRead,
+        cache_creation_input_tokens: canonical.cacheCreation,
+        output_tokens: canonical.output,
+        reasoning_output_tokens: 0,
+        total_tokens: total,
+        billable_total_tokens: total,
+        total_cost_usd: 0,
+        conversation_count: 0,
+      },
+    });
+  }
+  events.sort((a, b) => a.tsMs - b.tsMs || (a.requestId < b.requestId ? -1 : 1));
+
+  return { events, rowsSeen, rowsSkipped };
+}
+
+module.exports = {
+  DEVIN_TABLE_PROBE_SQL,
+  devinUsageSql,
+  buildDevinUsageEvents,
+};

package/src/lib/local-api.js

@@ -3215,6 +3215,19 @@ function createLocalApiHandler({ queuePath }) {
     if (p === "/functions/tokentracker-usage-limits") {
       const { getUsageLimits, resetUsageLimitsCache } = require("./usage-limits");
       try {
+        // Devin quota is opt-in (Settings > Usage & Limits > Providers). An
+        // explicit devin=1 without local authentication is rejected before any
+        // cache reset or provider work — silently downgrading it to a disabled
+        // response would misreport an enabled client as a disabled provider.
+        // Authorization is evaluated unconditionally so the check never depends
+        // on the user-controlled opt-in flag itself.
+        const localAuthorized = isAuthorizedLocalMutation(req);
+        const devinParam = url.searchParams.get("devin");
+        const devinEnabled = devinParam === "1" || devinParam === "true";
+        if (devinEnabled && !localAuthorized) {
+          json(res, { error: "Unauthorized" }, 401);
+          return true;
+        }
         const refreshParam = url.searchParams.get("refresh");
         const forceRefresh = refreshParam === "1" || refreshParam === "true";
         if (forceRefresh) {
@@ -3227,6 +3240,7 @@ function createLocalApiHandler({ queuePath }) {
           // Punches through the Claude disk fresh-cache (but not the 429
           // cooldown) — an explicit user refresh should hit upstream.
           forceRefresh,
+          devinEnabled,
         });
         json(res, data);
       } catch (e) {

package/src/lib/passive-mode.js

@@ -89,7 +89,7 @@ function classifyWritableFailure(filePath) {
  *   every_code, opencode, openclaw, codebuddy, workbuddy, grok }
  * @returns {PassiveProvider[]}
  */
-function detectPassiveProviders({ home, hookStatus }) {
+function detectPassiveProviders({ home, hookStatus, env = process.env }) {
   const out = [];
 
   // Claude Code — logs at ~/.claude/projects/<name>/*.jsonl
@@ -128,6 +128,18 @@ function detectPassiveProviders({ home, hookStatus }) {
     settingsPath: path.join(home, ".codex", "config.toml"),
   }));
 
+  const acodeHome = env.TOKENTRACKER_ACODE_HOME || path.join(home, ".acode");
+  out.push(buildEntry({
+    name: "acode",
+    hookExpected: true,
+    hookInstalled: Boolean(hookStatus?.acode_notify ?? hookStatus?.acode),
+    logsDir: acodeHome,
+    logsPredicate: (_full, name, isDir) =>
+      (isDir && (name === "sessions" || name === "archived_sessions")) ||
+      (!isDir && name.endsWith(".jsonl")),
+    settingsPath: path.join(acodeHome, "config.toml"),
+  }));
+
   // Every Code — same shape as Codex, different home
   out.push(buildEntry({
     name: "every_code",

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

@@ -2,7 +2,8 @@
   "_meta": {
     "note": "Curated price overrides. Always wins over LiteLLM. Two reasons to live here: (1) self-defined alias names that LiteLLM will never carry (kiro-*, hy3-*, composer-*, kimi-for-coding, free-tier OpenRouter routes); (2) prices we want to pin even if LiteLLM has the model (e.g. cache_write fields LiteLLM often omits). Units: USD per million tokens. Edit this file to override pricing without redeploying.",
     "units": "usd_per_million_tokens",
-    "deepseek_v4_pricing_note": "Official DeepSeek V4 pricing verified on 2026-08-21. Peak hours are 01:00-04:00 and 06:00-10:00 UTC; all other hours are off-peak at half price. Exact entries store peak rates. computeRowCost applies the time-of-use discount from each row's UTC hour_start."
+    "deepseek_v4_pricing_note": "Official DeepSeek V4 pricing verified on 2026-08-21. Peak hours are 01:00-04:00 and 06:00-10:00 UTC; all other hours are off-peak at half price. Exact entries store peak rates. computeRowCost applies the time-of-use discount from each row's UTC hour_start.",
+    "iflytek_maas_pricing_note": "iFlytek MaaS inference prices supplied for AStudio on 2026-08-31. AStudio homepage: https://agent.xfyun.cn/. Official pricing source: https://maas.xfyun.cn/modelSquare. Source values are RMB per million tokens, converted at 7.2 RMB/USD, then rounded to two decimal places for billing. Models without a dedicated cache-hit price use the normal input rate; cache writes also use the normal input rate. iFlytek MaaS prices are source-scoped and do not inherit DeepSeek time-of-use discounts."
   },
   "exact": {
     "claude-fable-5": {
@@ -47,6 +48,13 @@
       "cache_write": 1.25,
       "note": "Canonical undated alias emitted by Cursor. Pin to the same rates as claude-haiku-4-5-20251001 so local cost does not depend on which dated LiteLLM aliases are present in the current cache."
     },
+    "gpt-6-astra": {
+      "input": 10,
+      "output": 50,
+      "cache_read": 1,
+      "cache_write": 12.5,
+      "note": "Official Standard USD/MTok rates verified 2026-09-07: https://developers.openai.com/api/docs/models/gpt-6-astra. Requests above 272K input tokens use 2x input/cache and 1.5x output rates; computeRowCost applies this only to observed long-context request subsets. Cloud aggregates use Standard short-context estimates because they do not retain per-request context or service tier. Reasoning-effort suffixes use the same base rates."
+    },
     "gpt-5.6-sol": {
       "input": 4,
       "output": 20,
@@ -413,7 +421,76 @@
   "alias": {
     "auto": "composer-1"
   },
+  "source_exact": {
+    "acode": {
+      "xopglm53": { "input": 1.11, "output": 3.89, "cache_read": 0.28, "cache_write": 1.11 },
+      "xopdeepseekv4pro0813": { "input": 1.25, "output": 3.75, "cache_read": 0.04, "cache_write": 1.25 },
+      "xopdeepseekv4flash0731": { "input": 0.14, "output": 0.28, "cache_read": 0.03, "cache_write": 0.14 },
+      "xopkimik27code": { "input": 0.90, "output": 3.75, "cache_read": 0.90, "cache_write": 0.90 },
+      "xopglm52": { "input": 1.11, "output": 3.89, "cache_read": 0.28, "cache_write": 1.11 },
+      "xopdeepseekv4flash": { "input": 0.14, "output": 0.28, "cache_read": 0.03, "cache_write": 0.14 },
+      "xopkimik26": { "input": 0.90, "output": 3.75, "cache_read": 0.18, "cache_write": 0.90 },
+      "xopdeepseekv4pro": { "input": 1.67, "output": 3.33, "cache_read": 0.14, "cache_write": 1.67 },
+      "xopqwen36v35b": { "input": 0.15, "output": 0.90, "cache_read": 0.15, "cache_write": 0.15 },
+      "xophunyuan7bmt": { "input": 0.07, "output": 0.28, "cache_read": 0.07, "cache_write": 0.07 },
+      "xoppaddleocrv16": { "input": 0.00, "output": 0.00, "cache_read": 0.00, "cache_write": 0.00 },
+      "xsparkx2flash": { "input": 0.14, "output": 0.28, "cache_read": 0.14, "cache_write": 0.14 },
+      "xopglm51": { "input": 1.11, "output": 3.89, "cache_read": 0.22, "cache_write": 1.11 },
+      "xsparkx2": { "input": 0.42, "output": 0.42, "cache_read": 0.42, "cache_write": 0.42 },
+      "xop35qwen2b": { "input": 0.03, "output": 0.06, "cache_read": 0.03, "cache_write": 0.03 },
+      "xopqwen35397b": { "input": 0.17, "output": 1.00, "cache_read": 0.17, "cache_write": 0.17 },
+      "xminimaxm25": { "input": 0.29, "output": 1.17, "cache_read": 0.29, "cache_write": 0.29 },
+      "xopglm5": { "input": 0.83, "output": 3.06, "cache_read": 0.17, "cache_write": 0.83 },
+      "xopkimik25": { "input": 0.56, "output": 2.92, "cache_read": 0.56, "cache_write": 0.56 },
+      "xopdeepseekv32": { "input": 0.14, "output": 0.21, "cache_read": 0.14, "cache_write": 0.14 },
+      "xop3qwencodernext": { "input": 0.35, "output": 1.39, "cache_read": 0.35, "cache_write": 0.35 },
+      "xopglmv47flash": { "input": 0.14, "output": 0.21, "cache_read": 0.14, "cache_write": 0.14 },
+      "xopglm47blth2": { "input": 0.56, "output": 2.22, "cache_read": 0.56, "cache_write": 0.56 },
+      "xop3qwen32bvl": { "input": 0.28, "output": 1.11, "cache_read": 0.28, "cache_write": 0.28 },
+      "xopdeepseekocr": { "input": 0.00, "output": 0.00, "cache_read": 0.00, "cache_write": 0.00 },
+      "xophunyuanocr": { "input": 0.00, "output": 0.00, "cache_read": 0.00, "cache_write": 0.00 },
+      "xop3qwen80bnext": { "input": 0.08, "output": 0.33, "cache_read": 0.08, "cache_write": 0.08 },
+      "xop3qwen235b2507": { "input": 0.17, "output": 1.67, "cache_read": 0.17, "cache_write": 0.17 },
+      "xop3qwen30b2507": { "input": 0.06, "output": 0.63, "cache_read": 0.06, "cache_write": 0.06 },
+      "xop3qwen235b": { "input": 0.17, "output": 1.67, "cache_read": 0.17, "cache_write": 0.17 },
+      "xop3qwen30b": { "input": 0.06, "output": 0.63, "cache_read": 0.06, "cache_write": 0.06 },
+      "xop3qwen32b": { "input": 0.17, "output": 1.67, "cache_read": 0.17, "cache_write": 0.17 },
+      "xdeepseekv3": { "input": 0.22, "output": 0.89, "cache_read": 0.22, "cache_write": 0.22 },
+      "xdeepseekr1": { "input": 0.44, "output": 1.78, "cache_read": 0.44, "cache_write": 0.44 },
+      "xdeepseekr1qwen32b": { "input": 0.22, "output": 0.67, "cache_read": 0.22, "cache_write": 0.22 },
+      "xopkimik2blth": { "input": 0.56, "output": 2.22, "cache_read": 0.56, "cache_write": 0.56 },
+      "xopkimik2blins": { "input": 0.56, "output": 2.22, "cache_read": 0.56, "cache_write": 0.56 },
+      "xop3qwen8breranker": { "input": 0.00, "output": 0.00, "cache_read": 0.00, "cache_write": 0.00 },
+      "xop3qwen8bembedding": { "input": 0.00, "output": 0.00, "cache_read": 0.00, "cache_write": 0.00 },
+      "xop3qwen0b6": { "input": 0.04, "output": 0.42, "cache_read": 0.04, "cache_write": 0.04 },
+      "xop3qwen4b": { "input": 0.04, "output": 0.42, "cache_read": 0.04, "cache_write": 0.04 },
+      "xqwen257bchat": { "input": 0.07, "output": 0.14, "cache_read": 0.07, "cache_write": 0.07 },
+      "xop3qwen14b": { "input": 0.14, "output": 1.39, "cache_read": 0.14, "cache_write": 0.14 },
+      "xop3qwen8b": { "input": 0.07, "output": 0.69, "cache_read": 0.07, "cache_write": 0.07 },
+      "xsparkprox": { "input": 1.11, "output": 5.56, "cache_read": 1.11, "cache_write": 1.11 },
+      "xspark13b6k": { "input": 0.28, "output": 0.83, "cache_read": 0.28, "cache_write": 0.28 },
+      "spark mini": { "input": 0.28, "output": 1.11, "cache_read": 0.28, "cache_write": 0.28 },
+      "spark mini instruct": { "input": 0.28, "output": 1.11, "cache_read": 0.28, "cache_write": 0.28 },
+      "spark tiny": { "input": 0.28, "output": 1.11, "cache_read": 0.28, "cache_write": 0.28 },
+      "internlm2.5_7b_chat": { "input": 0.28, "output": 1.11, "cache_read": 0.28, "cache_write": 0.28 },
+      "internlm2.5_1.8b_chat": { "input": 0.28, "output": 1.11, "cache_read": 0.28, "cache_write": 0.28 },
+      "qwen_v2.5_7b_base": { "input": 0.28, "output": 1.11, "cache_read": 0.28, "cache_write": 0.28 },
+      "xsqwen2d53b": { "input": 0.28, "output": 1.11, "cache_read": 0.28, "cache_write": 0.28 },
+      "qwen_v2.5_3b_base": { "input": 0.28, "output": 1.11, "cache_read": 0.28, "cache_write": 0.28 },
+      "qwen_v2.5_1.5b_instruct": { "input": 0.28, "output": 1.11, "cache_read": 0.28, "cache_write": 0.28 },
+      "qwen_v2.5_1.5b_base": { "input": 0.28, "output": 1.11, "cache_read": 0.28, "cache_write": 0.28 },
+      "qwen_v2.5_0.5b_instruct": { "input": 0.28, "output": 1.11, "cache_read": 0.28, "cache_write": 0.28 },
+      "qwen_v2.5_0.5b_base": { "input": 0.28, "output": 1.11, "cache_read": 0.28, "cache_write": 0.28 },
+      "xsqwenv2s1b5c": { "input": 0.14, "output": 0.28, "cache_read": 0.14, "cache_write": 0.14 },
+      "xsqwenv2s0b5c": { "input": 0.28, "output": 0.56, "cache_read": 0.28, "cache_write": 0.28 },
+      "xqwen14bchat": { "input": 0.28, "output": 0.83, "cache_read": 0.28, "cache_write": 0.28 }
+    }
+  },
   "fuzzy": [
+    {
+      "match": "gpt-6-astra",
+      "ref": "gpt-6-astra"
+    },
     {
       "match": "gpt-5.6-sol",
       "ref": "gpt-5.6-sol"

package/src/lib/pricing/index.js

@@ -25,10 +25,9 @@ const PI_SUBSCRIPTION_SOURCES = new Set([
 // server (including LM Link). Secure Cloud usage has a separate billing path
 // and is not present in these logs.
 const LOCAL_INFERENCE_SOURCES = new Set(["lmstudio"]);
-// Forward-looking: unreachable for Codex today. Every observed token_count
-// reports model_context_window = 258400, below this threshold, so no request
-// can exceed it (max raw input seen locally: 238853). The long-context
-// repricing stays in place for when OpenAI ships a larger Codex window.
+// OpenAI long-context pricing depends on a single request's raw input,
+// including cached tokens, never a session/day aggregate. Astra supports
+// a larger context window; only observed request subsets receive the premium.
 const OPENAI_LONG_CONTEXT_INPUT_THRESHOLD = 272_000;
 const SOURCES_WITH_AUTHORITATIVE_COST = new Set(["grok"]);
 const SEED_SNAPSHOT_PATH = path.resolve(__dirname, "seed-snapshot.json");
@@ -36,7 +35,6 @@ const DEEPSEEK_TIME_PRICED_MODELS = [
   "deepseek-v4-flash",
   "deepseek-v4-pro",
 ];
-
 // Sync seed load. Done at require-time so callers that haven't awaited
 // ensurePricingLoaded() (e.g. tests, vite mock startup, edge functions) still
 // get LiteLLM-backed pricing instead of all-zero. ensurePricingLoaded() will
@@ -164,6 +162,9 @@ function isDeepSeekOffPeak(row) {
 
 function getRowPricing(row) {
   const pricing = getModelPricing(row?.model, { source: row?.source });
+  // AStudio uses iFlytek MaaS fixed prices and does not inherit DeepSeek public API
+  // time-based discounts.
+  if (String(row?.source || "").toLowerCase() === "acode") return pricing;
   if (!isDeepSeekTimePricedModel(row?.model) || !isDeepSeekOffPeak(row)) return pricing;
   return {
     ...pricing,
@@ -193,7 +194,8 @@ function computeRowCost(row) {
     reportedCost > 0
   ) return reportedCost;
   const pricing = getRowPricing(row);
-  const reasoningIncludedInOutput = row.source === "codex" || row.source === "every-code";
+  const reasoningIncludedInOutput =
+    row.source === "codex" || row.source === "acode" || row.source === "every-code";
   const reasoningCost = reasoningIncludedInOutput
     ? 0
     : (row.reasoning_output_tokens || 0) * (pricing.output || 0);
@@ -207,9 +209,9 @@ function computeRowCost(row) {
   );
 
   const model = String(row?.model || "").toLowerCase();
-  const usesGpt56SolLongContextTier =
-    model === "gpt-5.6" || model.includes("gpt-5.6-sol");
-  if (!usesGpt56SolLongContextTier) return baseCost;
+  const usesOpenAILongContextTier =
+    model === "gpt-5.6" || model.includes("gpt-5.6-sol") || model.includes("gpt-6-astra");
+  if (!usesOpenAILongContextTier) return baseCost;
 
   const bounded = (value, total) => Math.min(
     Math.max(0, Number(value) || 0),

package/src/lib/pricing/matcher.js

@@ -1,6 +1,7 @@
 // Pure pricing-lookup logic. No I/O, no async. Tested in isolation.
 //
 // Resolve order:
+//   0. CURATED source exact match (source-specific pricing)
 //   1. CURATED exact match (self-defined aliases like kiro-*, hy3-*)
 //   2. LiteLLM exact match (mainstream claude/gpt-5/gemini)
 //   3. CURATED alias (e.g. "auto" -> "composer-1")
@@ -164,6 +165,15 @@ function normalizeWorkbuddyModel(model) {
   return model;
 }
 
+function normalizeIFlytekMaasModel(model) {
+  if (!model || typeof model !== "string") return model;
+  const trimmed = model.trim();
+  const lower = trimmed.toLowerCase();
+  // The default Agent slug maps to the Spark X2 service in the supplied price list.
+  if (lower === "xsparkx2agent") return "xsparkx2";
+  return trimmed;
+}
+
 // Unsloth Studio can mix local engines and paid API providers in one database.
 // The parser qualifies metered models with their provider and marks local,
 // subscription-backed, or ambiguous custom routes as unpriced. Returning a
@@ -182,6 +192,7 @@ function normalizeUnslothModel(model) {
 // raw model name is preserved for storage/display). Add a source here when its
 // model strings don't match the LiteLLM/curated keys verbatim.
 const SOURCE_MODEL_NORMALIZERS = {
+  acode: normalizeIFlytekMaasModel,
   antigravity: normalizeAntigravityModel,
   claude: normalizeClaudeModel,
   cursor: normalizeCursorModel,
@@ -242,6 +253,22 @@ function lookupPricing(model, { curated, litellm, source } = {}) {
   const lower = lookupModel.toLowerCase();
   const dotForm = buildDotRestoredModel(lookupModel);
 
+  // 0. CURATED source exact. Source-specific prices apply only to their source,
+  // preventing collisions with public prices for same-named models from other CLIs.
+  const sourceKey = typeof source === "string" ? source.toLowerCase() : "";
+  // AStudio does not disclose its routed model. Stop before generic aliases
+  // and fuzzy matching can turn an unresolved router into a priced model.
+  if (sourceKey === "acode" && (lower === "auto" || lower.endsWith("-auto"))) {
+    return { hit: false, source: "miss", value: null };
+  }
+  const sourceExact = curated.source_exact?.[sourceKey];
+  if (sourceExact) {
+    const sourceValue = lookupExactCaseInsensitive(sourceExact, lookupModel);
+    if (sourceValue) {
+      return { hit: true, source: "curated:source-exact", value: sourceValue };
+    }
+  }
+
   // 1. CURATED exact
   if (curated.exact && curated.exact[lookupModel]) {
     return { hit: true, source: "curated:exact", value: curated.exact[lookupModel] };
@@ -382,6 +409,7 @@ module.exports = {
   lookupPricing,
   stripReasoningSuffix,
   normalizeAntigravityModel,
+  normalizeIFlytekMaasModel,
   normalizeClaudeModel,
   normalizeCursorModel,
   normalizeZedModel,