iobroker.ai-usage 0.13.0 → 0.14.0

package/README.md

@@ -84,7 +84,7 @@ ai-usage.0.
 │   ├── limitReached           — this account is at 100 % (bool)
 │   ├── info.unreach           — account is not delivering (bool) — drives the connection icon
 │   ├── info.error             — why there is no data; empty while all is well, "Unknown" while the adapter itself has nothing to report
-│   ├── info.lastUpdate        — time of the last successful read
+│   ├── info.lastUpdate        — when the current values were fetched
 │   ├── limits.<window>.*      — percent, reset time and whether this window is the limit in force
 │   └── credits.*              — where the provider reports a balance
 └── <name>-api                 — one node per key-based account
@@ -100,9 +100,14 @@ the reset time simply empties while no window is running, and a window or model
 tree only when the provider stops reporting it entirely.
 
 **The connection icon** sits next to each account, green while it delivers. A throttle keeps it
-green — the last values stay valid while the adapter waits. A rejected sign-in or a broken service
-switch it off at once, an unreachable service after three attempts, so a hiccup does not make it
-flap. `info.error` names the cause whenever the provider gave one.
+green — the last values stay valid while the adapter waits. A rejected sign-in, a broken service or
+an answer that cannot be processed switch it off at once, an unreachable service after three
+attempts, so a hiccup does not make it flap. `info.error` names the cause whenever the provider gave
+one, in the provider's own words where it sent any.
+
+**Signing out clears that account's alarms.** Its measured values stay in the tree, but `warning`,
+`limitReached` and every total that counts them drop it — an account nobody watches must not keep an
+automation waiting.
 
 **Which window is in force** is shown per window: with Claude the provider says so itself, elsewhere it
 is the window that speaks for the account. It tells you what you will run into next — a model window
@@ -148,9 +153,26 @@ so instead of pretending to be connected; signing in again is all it takes.
 
 <!--
     Placeholder for the next version (at the beginning of the line):
-    ### **WORK IN PROGRESS**
 -->
 
+### 0.14.0 (2026-09-15)
+
+- Fixed: A failed write of the token file after a refresh lost the sign-in for good — the provider had already rotated them, so the next poll reported a rejected sign-in
+- Fixed: Signing out no longer comes undone by itself — a sign-out that landed during a background token renewal could leave the account signed in
+- Fixed: The last-update stamp moved forward on a tolerated connection failure, dating values the round had never fetched
+- Fixed: An answer the adapter cannot process is reported as a service fault at once, instead of claiming for three rounds that the service is unreachable
+- Fixed: A failed cleanup of vanished windows or models no longer discards the round — the values were in the tree, but the account called them unstored and the totals froze
+- Fixed: A connection failure counter that a throttle or a rejected sign-in had interrupted no longer adds up to "not reachable"
+- Fixed: An account you signed out of drops its warning and limit alarms and leaves the adapter-wide totals — its measured values stay in the tree
+- Fixed: The reset time of the Claude session and week windows is filled from the plan-wide block when the window entry itself carries none
+- Improved: Window reset times, the next voucher expiry and the credit ceiling are written only when they change — announced facts, not measurements, so their timestamp stops moving every poll
+- Improved: Where a provider sends a reason of its own, `info.error` now says it ("invalid API key") instead of a bare status number
+- Improved: An access token the provider invalidated early is refreshed once and the request repeated, instead of reporting a rejected sign-in until it would have expired
+
+Only the Claude subscription runs against a real account here. The token-file fix, the early-refresh
+retry and the sign-out behaviour are covered by tests but were not seen on a real ChatGPT, Google,
+OpenRouter, DeepSeek, OpenAI or Anthropic account.
+
 ### 0.13.0 (2026-09-12)
 
 - Fixed: The costs of an Anthropic organisation account were a hundred times too high — the provider counts them in cents, the adapter read them as dollars
@@ -204,15 +226,6 @@ so instead of pretending to be connected; signing in again is all it takes.
 - Improved: Monthly cost reports can no longer be cut short in silence — a report that does not fit is reported in the log instead of producing a figure that is too low
 - Changed: "Highest account utilisation" says what it always measured — the fullest limit window **or** the account's remaining budget
 
-### 0.10.0 (2026-09-01)
-
-- Fixed: The reset-time datapoint of a limit window no longer disappears and reappears — it stays and simply empties while no window is running
-- Fixed: The settings page no longer shows the sign-in screen to a signed-in account, and its rows load without waiting for the credential storage scan
-- Improved: Claude usage is read with far fewer rejections — the query now identifies itself the way the endpoint expects
-- Changed: New Claude sign-ins request only the profile permission — the stored access can no longer create API keys or run models
-- New: ChatGPT accounts show their purchasable limit-reset credits — how many are available and when the next one expires
-- Improved: An unreadable provider answer is now reported as a service fault instead of a missing connection
-
 [Older changelogs can be found there](CHANGELOG_OLD.md)
 
 ## Support

package/admin/i18n/de.json

@@ -51,7 +51,6 @@
   "nameModels": "Verbrauch je Modell",
   "nameModel": "Modell %s",
   "nameModelTokens": "%s Token heute",
-  "nameModelCosts": "%s Kosten heute",
   "nameAvailable": "Guthaben reicht für Anfragen",
   "nameWindowSession": "Sitzung (5 Std.)",
   "nameWindowWeek": "Woche (alle Modelle)",

package/admin/i18n/en.json

@@ -51,7 +51,6 @@
   "nameModels": "Per-model usage",
   "nameModel": "Model %s",
   "nameModelTokens": "%s tokens today",
-  "nameModelCosts": "%s costs today",
   "nameAvailable": "Balance sufficient for calls",
   "nameWindowSession": "Session (5 h)",
   "nameWindowWeek": "Week (all models)",

package/admin/i18n/es.json

@@ -51,7 +51,6 @@
   "nameModels": "Uso por modelo",
   "nameModel": "Modelo %s",
   "nameModelTokens": "%s tokens de hoy",
-  "nameModelCosts": "%s costes de hoy",
   "nameAvailable": "Saldo suficiente para las llamadas",
   "nameWindowSession": "Sesión (5 h)",
   "nameWindowWeek": "Semana (todos los modelos)",

package/admin/i18n/fr.json

@@ -51,7 +51,6 @@
   "nameModels": "Utilisation par modèle",
   "nameModel": "Modèle %s",
   "nameModelTokens": "%s jetons aujourd'hui",
-  "nameModelCosts": "%s coûts aujourd'hui",
   "nameAvailable": "Solde suffisant pour les appels",
   "nameWindowSession": "Session (5 h)",
   "nameWindowWeek": "Semaine (tous les modèles)",

package/admin/i18n/it.json

@@ -51,7 +51,6 @@
   "nameModels": "Utilizzo per modello",
   "nameModel": "Modello %s",
   "nameModelTokens": "%s token di oggi",
-  "nameModelCosts": "%s costi di oggi",
   "nameAvailable": "Saldo sufficiente per le chiamate",
   "nameWindowSession": "Sessione (5 h)",
   "nameWindowWeek": "Settimana (tutti i modelli)",

package/admin/i18n/nl.json

@@ -51,7 +51,6 @@
   "nameModels": "Verbruik per model",
   "nameModel": "Model %s",
   "nameModelTokens": "%s tokens vandaag",
-  "nameModelCosts": "%s kosten vandaag",
   "nameAvailable": "Tegoed voldoende voor aanvragen",
   "nameWindowSession": "Sessie (5 u)",
   "nameWindowWeek": "Week (alle modellen)",

package/admin/i18n/pl.json

@@ -51,7 +51,6 @@
   "nameModels": "Zużycie na model",
   "nameModel": "Model %s",
   "nameModelTokens": "%s tokenów dzisiaj",
-  "nameModelCosts": "%s koszty dzisiaj",
   "nameAvailable": "Środki wystarczają na zapytania",
   "nameWindowSession": "Sesja (5 godz.)",
   "nameWindowWeek": "Tydzień (wszystkie modele)",

package/admin/i18n/pt.json

@@ -51,7 +51,6 @@
   "nameModels": "Utilização por modelo",
   "nameModel": "Modelo %s",
   "nameModelTokens": "%s tokens hoje",
-  "nameModelCosts": "%s custos hoje",
   "nameAvailable": "Saldo suficiente para chamadas",
   "nameWindowSession": "Sessão (5 h)",
   "nameWindowWeek": "Semana (todos os modelos)",

package/admin/i18n/ru.json

@@ -51,7 +51,6 @@
   "nameModels": "Использование по моделям",
   "nameModel": "Модель %s",
   "nameModelTokens": "%s токенов сегодня",
-  "nameModelCosts": "%s расходы сегодня",
   "nameAvailable": "Баланса хватает для запросов",
   "nameWindowSession": "Сессия (5 ч)",
   "nameWindowWeek": "Неделя (все модели)",

package/admin/i18n/uk.json

@@ -51,7 +51,6 @@
   "nameModels": "Використання за моделями",
   "nameModel": "Модель %s",
   "nameModelTokens": "%s токенів сьогодні",
-  "nameModelCosts": "%s витрати сьогодні",
   "nameAvailable": "Балансу достатньо для запитів",
   "nameWindowSession": "Сесія (5 год)",
   "nameWindowWeek": "Тиждень (усі моделі)",

package/admin/i18n/zh-cn.json

@@ -51,7 +51,6 @@
   "nameModels": "各模型用量",
   "nameModel": "模型 %s",
   "nameModelTokens": "%s 今日 Token",
-  "nameModelCosts": "%s 今日费用",
   "nameAvailable": "余额足以发起调用",
   "nameWindowSession": "会话（5 小时）",
   "nameWindowWeek": "一周（所有模型）",

package/build/lib/http.js

@@ -32,14 +32,15 @@ async function request(url, init, authOn400 = false) {
   } catch (e) {
     throw new import_provider.FetchError("network", e instanceof Error ? e.message : String(e));
   }
-  if (response.status === 401 || response.status === 403 || authOn400 && response.status === 400) {
-    throw new import_provider.FetchError("auth", `HTTP ${response.status}`);
-  }
-  if (response.status === 429) {
-    throw new import_provider.FetchError("rate-limit", "HTTP 429");
-  }
   if (!response.ok) {
-    throw new import_provider.FetchError("service", `HTTP ${response.status}`);
+    const detail = await errorDetail(response);
+    if (response.status === 401 || response.status === 403 || authOn400 && response.status === 400) {
+      throw new import_provider.FetchError("auth", `HTTP ${response.status}${detail}`);
+    }
+    if (response.status === 429) {
+      throw new import_provider.FetchError("rate-limit", `HTTP 429${detail}`);
+    }
+    throw new import_provider.FetchError("service", `HTTP ${response.status}${detail}`);
   }
   try {
     return await response.json();
@@ -47,6 +48,29 @@ async function request(url, init, authOn400 = false) {
     throw new import_provider.FetchError("service", `invalid JSON: ${e instanceof Error ? e.message : String(e)}`);
   }
 }
+const MAX_DETAIL_CHARS = 200;
+async function errorDetail(response) {
+  var _a, _b;
+  let text;
+  try {
+    text = await response.text();
+  } catch {
+    return "";
+  }
+  let message;
+  try {
+    const body = JSON.parse(text);
+    const error = body.error;
+    message = typeof error === "string" ? error : (_b = (_a = error == null ? void 0 : error.message) != null ? _a : body.message) != null ? _b : void 0;
+  } catch {
+    message = text.trim().length > 0 && text.trim().length <= MAX_DETAIL_CHARS ? text.trim() : void 0;
+  }
+  if (typeof message !== "string" || message.trim().length === 0) {
+    return "";
+  }
+  const trimmed = message.trim();
+  return ` \u2014 ${trimmed.length > MAX_DETAIL_CHARS ? `${trimmed.slice(0, MAX_DETAIL_CHARS)}\u2026` : trimmed}`;
+}
 async function getJson(url, headers) {
   return request(url, { headers });
 }

package/build/lib/http.js.map

@@ -1,7 +1,7 @@
 {
   "version": 3,
   "sources": ["../../src/lib/http.ts"],
-  "sourcesContent": ["import { FetchError } from \"./provider\";\n\n/** Per-request timeout (ms). */\nconst REQUEST_TIMEOUT_MS = 15000;\n\n/** The JSON-GET seam the provider modules use \u2014 injectable for tests. */\nexport type JsonFetch = (url: string, headers: Record<string, string>) => Promise<unknown>;\n\n/**\n * Per-call options of the two POST seams.\n *\n * `authOn400` used to be baked into `postJson`/`postForm`, which made EVERY post\n * read a 400 as a rejected sign-in. That is right for the OAuth token endpoints\n * and wrong everywhere else: the ChatGPT device-code poll takes an auth failure\n * as \"the user has not confirmed yet\" and would have waited out the whole\n * fifteen minutes on a 400, and Google's Code-Assist call skipped its second\n * host because it thought the sign-in was gone. The flag belongs at the call.\n */\nexport interface PostOptions {\n  /** Extra request headers. */\n  headers?: Record<string, string>;\n  /** True only for OAuth token endpoints, which answer a dead code or refresh token with 400. */\n  authOn400?: boolean;\n}\n\n/** The JSON-POST seam \u2014 one definition for every provider module. */\nexport type JsonPost = (url: string, body: Record<string, unknown>, options?: PostOptions) => Promise<unknown>;\n\n/** The form-POST seam (OAuth code redemption). */\nexport type FormPost = (url: string, form: Record<string, string>, options?: PostOptions) => Promise<unknown>;\n\n/**\n * Run one request and turn its outcome into the shared failure classification:\n * 401/403 (and 400 where the provider answers a rejected grant that way) become an\n * auth error, 429 a rate-limit error, any other bad status or unparsable body a\n * SERVICE error (the service answered and is broken), and only a throw \u2014 refused\n * connection, DNS failure, timeout \u2014 a network error. The poll engine turns that\n * split into \"the AI service is down\" versus \"this host has no connection\".\n *\n * @param url the request URL\n * @param init the request options (method, headers, body)\n * @param authOn400 true when a 400 means a rejected grant rather than a service fault\n * @returns the parsed JSON body\n */\nasync function request(url: string, init: RequestInit, authOn400 = false): Promise<unknown> {\n  let response: Response;\n  try {\n    response = await fetch(url, { ...init, signal: AbortSignal.timeout(REQUEST_TIMEOUT_MS) });\n  } catch (e) {\n    throw new FetchError(\"network\", e instanceof Error ? e.message : String(e));\n  }\n  if (response.status === 401 || response.status === 403 || (authOn400 && response.status === 400)) {\n    throw new FetchError(\"auth\", `HTTP ${response.status}`);\n  }\n  if (response.status === 429) {\n    throw new FetchError(\"rate-limit\", \"HTTP 429\");\n  }\n  if (!response.ok) {\n    // 5xx = the service answered and is broken; anything else unexpected is treated\n    // the same way, because the service DID answer \u2014 only a throw above means we\n    // never reached it.\n    throw new FetchError(\"service\", `HTTP ${response.status}`);\n  }\n  try {\n    return await response.json();\n  } catch (e) {\n    throw new FetchError(\"service\", `invalid JSON: ${e instanceof Error ? e.message : String(e)}`);\n  }\n}\n\n/**\n * GET a JSON document.\n *\n * @param url the request URL\n * @param headers request headers (Authorization etc.)\n * @returns the parsed JSON body\n */\nexport async function getJson(url: string, headers: Record<string, string>): Promise<unknown> {\n  return request(url, { headers });\n}\n\n/**\n * POST a JSON document.\n *\n * @param url the request URL\n * @param body the JSON body\n * @param options extra headers, and whether a 400 means a rejected grant\n * @returns the parsed JSON response\n */\nexport async function postJson(\n  url: string,\n  body: Record<string, unknown>,\n  options: PostOptions = {},\n): Promise<unknown> {\n  return request(\n    url,\n    {\n      method: \"POST\",\n      headers: { \"Content-Type\": \"application/json\", ...options.headers },\n      body: JSON.stringify(body),\n    },\n    options.authOn400 === true,\n  );\n}\n\n/**\n * POST a form-encoded body and return the parsed JSON answer.\n *\n * OAuth code redemption uses form encoding while token refresh often uses JSON \u2014\n * ChatGPT/Codex needs BOTH, so the two shapes are separate helpers rather than one\n * guessing wrapper.\n *\n * @param url the request URL\n * @param form the form fields\n * @param options extra headers, and whether a 400 means a rejected grant\n * @returns the parsed JSON body\n */\nexport async function postForm(url: string, form: Record<string, string>, options: PostOptions = {}): Promise<unknown> {\n  return request(\n    url,\n    {\n      method: \"POST\",\n      headers: { \"Content-Type\": \"application/x-www-form-urlencoded\", ...options.headers },\n      body: new URLSearchParams(form).toString(),\n    },\n    options.authOn400 === true,\n  );\n}\n"],
-  "mappings": ";;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,sBAA2B;AAG3B,MAAM,qBAAqB;AAyC3B,eAAe,QAAQ,KAAa,MAAmB,YAAY,OAAyB;AAC1F,MAAI;AACJ,MAAI;AACF,eAAW,MAAM,MAAM,KAAK,EAAE,GAAG,MAAM,QAAQ,YAAY,QAAQ,kBAAkB,EAAE,CAAC;AAAA,EAC1F,SAAS,GAAG;AACV,UAAM,IAAI,2BAAW,WAAW,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC,CAAC;AAAA,EAC5E;AACA,MAAI,SAAS,WAAW,OAAO,SAAS,WAAW,OAAQ,aAAa,SAAS,WAAW,KAAM;AAChG,UAAM,IAAI,2BAAW,QAAQ,QAAQ,SAAS,MAAM,EAAE;AAAA,EACxD;AACA,MAAI,SAAS,WAAW,KAAK;AAC3B,UAAM,IAAI,2BAAW,cAAc,UAAU;AAAA,EAC/C;AACA,MAAI,CAAC,SAAS,IAAI;AAIhB,UAAM,IAAI,2BAAW,WAAW,QAAQ,SAAS,MAAM,EAAE;AAAA,EAC3D;AACA,MAAI;AACF,WAAO,MAAM,SAAS,KAAK;AAAA,EAC7B,SAAS,GAAG;AACV,UAAM,IAAI,2BAAW,WAAW,iBAAiB,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC,CAAC,EAAE;AAAA,EAC/F;AACF;AASA,eAAsB,QAAQ,KAAa,SAAmD;AAC5F,SAAO,QAAQ,KAAK,EAAE,QAAQ,CAAC;AACjC;AAUA,eAAsB,SACpB,KACA,MACA,UAAuB,CAAC,GACN;AAClB,SAAO;AAAA,IACL;AAAA,IACA;AAAA,MACE,QAAQ;AAAA,MACR,SAAS,EAAE,gBAAgB,oBAAoB,GAAG,QAAQ,QAAQ;AAAA,MAClE,MAAM,KAAK,UAAU,IAAI;AAAA,IAC3B;AAAA,IACA,QAAQ,cAAc;AAAA,EACxB;AACF;AAcA,eAAsB,SAAS,KAAa,MAA8B,UAAuB,CAAC,GAAqB;AACrH,SAAO;AAAA,IACL;AAAA,IACA;AAAA,MACE,QAAQ;AAAA,MACR,SAAS,EAAE,gBAAgB,qCAAqC,GAAG,QAAQ,QAAQ;AAAA,MACnF,MAAM,IAAI,gBAAgB,IAAI,EAAE,SAAS;AAAA,IAC3C;AAAA,IACA,QAAQ,cAAc;AAAA,EACxB;AACF;",
+  "sourcesContent": ["import { FetchError } from \"./provider\";\n\n/** Per-request timeout (ms). */\nconst REQUEST_TIMEOUT_MS = 15000;\n\n/** The JSON-GET seam the provider modules use \u2014 injectable for tests. */\nexport type JsonFetch = (url: string, headers: Record<string, string>) => Promise<unknown>;\n\n/**\n * Per-call options of the two POST seams.\n *\n * `authOn400` used to be baked into `postJson`/`postForm`, which made EVERY post\n * read a 400 as a rejected sign-in. That is right for the OAuth token endpoints\n * and wrong everywhere else: the ChatGPT device-code poll takes an auth failure\n * as \"the user has not confirmed yet\" and would have waited out the whole\n * fifteen minutes on a 400, and Google's Code-Assist call skipped its second\n * host because it thought the sign-in was gone. The flag belongs at the call.\n */\nexport interface PostOptions {\n  /** Extra request headers. */\n  headers?: Record<string, string>;\n  /** True only for OAuth token endpoints, which answer a dead code or refresh token with 400. */\n  authOn400?: boolean;\n}\n\n/** The JSON-POST seam \u2014 one definition for every provider module. */\nexport type JsonPost = (url: string, body: Record<string, unknown>, options?: PostOptions) => Promise<unknown>;\n\n/** The form-POST seam (OAuth code redemption). */\nexport type FormPost = (url: string, form: Record<string, string>, options?: PostOptions) => Promise<unknown>;\n\n/**\n * Run one request and turn its outcome into the shared failure classification:\n * 401/403 (and 400 where the provider answers a rejected grant that way) become an\n * auth error, 429 a rate-limit error, any other bad status or unparsable body a\n * SERVICE error (the service answered and is broken), and only a throw \u2014 refused\n * connection, DNS failure, timeout \u2014 a network error. The poll engine turns that\n * split into \"the AI service is down\" versus \"this host has no connection\".\n *\n * @param url the request URL\n * @param init the request options (method, headers, body)\n * @param authOn400 true when a 400 means a rejected grant rather than a service fault\n * @returns the parsed JSON body\n */\nasync function request(url: string, init: RequestInit, authOn400 = false): Promise<unknown> {\n  let response: Response;\n  try {\n    response = await fetch(url, { ...init, signal: AbortSignal.timeout(REQUEST_TIMEOUT_MS) });\n  } catch (e) {\n    throw new FetchError(\"network\", e instanceof Error ? e.message : String(e));\n  }\n  if (!response.ok) {\n    // The status alone reaches the user as `info.error`, and \"HTTP 401\" tells them\n    // nothing they can act on. OpenRouter, OpenAI and Anthropic all answer with\n    // `{ error: { message } }` \u2014 reading it turns the datapoint into \"invalid API\n    // key\". Best effort in the strictest sense: the status decides the class, the\n    // body only decorates the text, and a body that cannot be read changes nothing.\n    const detail = await errorDetail(response);\n    if (response.status === 401 || response.status === 403 || (authOn400 && response.status === 400)) {\n      throw new FetchError(\"auth\", `HTTP ${response.status}${detail}`);\n    }\n    if (response.status === 429) {\n      throw new FetchError(\"rate-limit\", `HTTP 429${detail}`);\n    }\n    // 5xx = the service answered and is broken; anything else unexpected is treated\n    // the same way, because the service DID answer \u2014 only a throw above means we\n    // never reached it.\n    throw new FetchError(\"service\", `HTTP ${response.status}${detail}`);\n  }\n  try {\n    return await response.json();\n  } catch (e) {\n    throw new FetchError(\"service\", `invalid JSON: ${e instanceof Error ? e.message : String(e)}`);\n  }\n}\n\n/** Longest provider message taken over into the error text. */\nconst MAX_DETAIL_CHARS = 200;\n\n/**\n * The provider's own words for a failed request, ready to append.\n *\n * Never throws and never rejects: a body that is missing, unreadable, not JSON or\n * shaped differently simply yields \"\" and the caller keeps the bare status. The\n * body of a failed response is consumed here either way, so nothing is left\n * dangling (measured 2026-09-12: the \"socket leak\" this used to be blamed on does\n * not exist).\n *\n * @param response the failed response\n * @returns \" \u2014 <message>\" or an empty string\n */\nasync function errorDetail(response: Response): Promise<string> {\n  let text: string;\n  try {\n    text = await response.text();\n  } catch {\n    return \"\";\n  }\n  let message: unknown;\n  try {\n    const body = JSON.parse(text) as { error?: unknown; message?: unknown };\n    const error = body.error;\n    message =\n      typeof error === \"string\" ? error : ((error as { message?: unknown })?.message ?? body.message ?? undefined);\n  } catch {\n    // Not JSON: a short plain-text body is still better than nothing, a long one\n    // (an HTML error page from a proxy) is noise.\n    message = text.trim().length > 0 && text.trim().length <= MAX_DETAIL_CHARS ? text.trim() : undefined;\n  }\n  if (typeof message !== \"string\" || message.trim().length === 0) {\n    return \"\";\n  }\n  const trimmed = message.trim();\n  return ` \u2014 ${trimmed.length > MAX_DETAIL_CHARS ? `${trimmed.slice(0, MAX_DETAIL_CHARS)}\u2026` : trimmed}`;\n}\n\n/**\n * GET a JSON document.\n *\n * @param url the request URL\n * @param headers request headers (Authorization etc.)\n * @returns the parsed JSON body\n */\nexport async function getJson(url: string, headers: Record<string, string>): Promise<unknown> {\n  return request(url, { headers });\n}\n\n/**\n * POST a JSON document.\n *\n * @param url the request URL\n * @param body the JSON body\n * @param options extra headers, and whether a 400 means a rejected grant\n * @returns the parsed JSON response\n */\nexport async function postJson(\n  url: string,\n  body: Record<string, unknown>,\n  options: PostOptions = {},\n): Promise<unknown> {\n  return request(\n    url,\n    {\n      method: \"POST\",\n      headers: { \"Content-Type\": \"application/json\", ...options.headers },\n      body: JSON.stringify(body),\n    },\n    options.authOn400 === true,\n  );\n}\n\n/**\n * POST a form-encoded body and return the parsed JSON answer.\n *\n * OAuth code redemption uses form encoding while token refresh often uses JSON \u2014\n * ChatGPT/Codex needs BOTH, so the two shapes are separate helpers rather than one\n * guessing wrapper.\n *\n * @param url the request URL\n * @param form the form fields\n * @param options extra headers, and whether a 400 means a rejected grant\n * @returns the parsed JSON body\n */\nexport async function postForm(url: string, form: Record<string, string>, options: PostOptions = {}): Promise<unknown> {\n  return request(\n    url,\n    {\n      method: \"POST\",\n      headers: { \"Content-Type\": \"application/x-www-form-urlencoded\", ...options.headers },\n      body: new URLSearchParams(form).toString(),\n    },\n    options.authOn400 === true,\n  );\n}\n"],
+  "mappings": ";;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,sBAA2B;AAG3B,MAAM,qBAAqB;AAyC3B,eAAe,QAAQ,KAAa,MAAmB,YAAY,OAAyB;AAC1F,MAAI;AACJ,MAAI;AACF,eAAW,MAAM,MAAM,KAAK,EAAE,GAAG,MAAM,QAAQ,YAAY,QAAQ,kBAAkB,EAAE,CAAC;AAAA,EAC1F,SAAS,GAAG;AACV,UAAM,IAAI,2BAAW,WAAW,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC,CAAC;AAAA,EAC5E;AACA,MAAI,CAAC,SAAS,IAAI;AAMhB,UAAM,SAAS,MAAM,YAAY,QAAQ;AACzC,QAAI,SAAS,WAAW,OAAO,SAAS,WAAW,OAAQ,aAAa,SAAS,WAAW,KAAM;AAChG,YAAM,IAAI,2BAAW,QAAQ,QAAQ,SAAS,MAAM,GAAG,MAAM,EAAE;AAAA,IACjE;AACA,QAAI,SAAS,WAAW,KAAK;AAC3B,YAAM,IAAI,2BAAW,cAAc,WAAW,MAAM,EAAE;AAAA,IACxD;AAIA,UAAM,IAAI,2BAAW,WAAW,QAAQ,SAAS,MAAM,GAAG,MAAM,EAAE;AAAA,EACpE;AACA,MAAI;AACF,WAAO,MAAM,SAAS,KAAK;AAAA,EAC7B,SAAS,GAAG;AACV,UAAM,IAAI,2BAAW,WAAW,iBAAiB,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC,CAAC,EAAE;AAAA,EAC/F;AACF;AAGA,MAAM,mBAAmB;AAczB,eAAe,YAAY,UAAqC;AA3FhE;AA4FE,MAAI;AACJ,MAAI;AACF,WAAO,MAAM,SAAS,KAAK;AAAA,EAC7B,QAAQ;AACN,WAAO;AAAA,EACT;AACA,MAAI;AACJ,MAAI;AACF,UAAM,OAAO,KAAK,MAAM,IAAI;AAC5B,UAAM,QAAQ,KAAK;AACnB,cACE,OAAO,UAAU,WAAW,SAAU,0CAAiC,YAAjC,YAA4C,KAAK,YAAjD,YAA4D;AAAA,EACtG,QAAQ;AAGN,cAAU,KAAK,KAAK,EAAE,SAAS,KAAK,KAAK,KAAK,EAAE,UAAU,mBAAmB,KAAK,KAAK,IAAI;AAAA,EAC7F;AACA,MAAI,OAAO,YAAY,YAAY,QAAQ,KAAK,EAAE,WAAW,GAAG;AAC9D,WAAO;AAAA,EACT;AACA,QAAM,UAAU,QAAQ,KAAK;AAC7B,SAAO,WAAM,QAAQ,SAAS,mBAAmB,GAAG,QAAQ,MAAM,GAAG,gBAAgB,CAAC,WAAM,OAAO;AACrG;AASA,eAAsB,QAAQ,KAAa,SAAmD;AAC5F,SAAO,QAAQ,KAAK,EAAE,QAAQ,CAAC;AACjC;AAUA,eAAsB,SACpB,KACA,MACA,UAAuB,CAAC,GACN;AAClB,SAAO;AAAA,IACL;AAAA,IACA;AAAA,MACE,QAAQ;AAAA,MACR,SAAS,EAAE,gBAAgB,oBAAoB,GAAG,QAAQ,QAAQ;AAAA,MAClE,MAAM,KAAK,UAAU,IAAI;AAAA,IAC3B;AAAA,IACA,QAAQ,cAAc;AAAA,EACxB;AACF;AAcA,eAAsB,SAAS,KAAa,MAA8B,UAAuB,CAAC,GAAqB;AACrH,SAAO;AAAA,IACL;AAAA,IACA;AAAA,MACE,QAAQ;AAAA,MACR,SAAS,EAAE,gBAAgB,qCAAqC,GAAG,QAAQ,QAAQ;AAAA,MACnF,MAAM,IAAI,gBAAgB,IAAI,EAAE,SAAS;AAAA,IAC3C;AAAA,IACA,QAAQ,cAAc;AAAA,EACxB;AACF;",
   "names": []
 }

package/build/lib/poll-engine.js

@@ -63,6 +63,7 @@ class PollEngine {
         authNotified: false,
         authRejected: false,
         storageFailed: false,
+        sweepFailed: false,
         serviceOnline: false,
         state: provider ? "no-connection" : "not-signed-in",
         // Nothing known until the service says something; the skeleton writes
@@ -216,6 +217,7 @@ class PollEngine {
       return;
     }
     const wasFailing = runtime.state !== "ok";
+    let delivered = false;
     let fetched;
     try {
       fetched = await runtime.provider.fetch();
@@ -246,7 +248,8 @@ class PollEngine {
         }
         runtime.status.snapshot = fetched;
         runtime.storageFailed = false;
-        if (wasFailing) {
+        delivered = true;
+        if (wasFailing && runtime.firstPollDone) {
           this.deps.log.info(`${config.name}: delivering again`);
         }
       } catch (e) {
@@ -257,7 +260,7 @@ class PollEngine {
       }
     }
     runtime.status.reachable = isDelivering(runtime.state);
-    this.writeAccountInfo(runtime);
+    this.writeAccountInfo(runtime, delivered);
     this.writeTotals();
     runtime.firstPollDone = true;
     this.reportFirstRoundOnce();
@@ -292,16 +295,21 @@ class PollEngine {
       return;
     }
     for (const write of writes) {
-      if (write.indicator) {
+      if (write.compare) {
         void this.deps.setStateChanged(write.id, write.value);
       } else {
         this.deps.setState(write.id, write.value);
       }
     }
-    await this.removeVanished(
-      runtime,
-      writes.map((write) => write.id)
-    );
+    try {
+      await this.removeVanished(
+        runtime,
+        writes.map((write) => write.id)
+      );
+      runtime.sweepFailed = false;
+    } catch (e) {
+      this.handleSweepFailure(runtime, e);
+    }
     if (this.stopped) {
       return;
     }
@@ -370,6 +378,29 @@ class PollEngine {
       `${runtime.config.name}: the values were fetched but could not be stored \u2014 the object database rejected the write (${message})`
     );
   }
+  /**
+   * The values are in the tree, the cleanup did not run.
+   *
+   * Not a storage failure: nothing the provider delivered was lost, so the round
+   * counts as delivered and the snapshot is adopted. What did not happen is the
+   * removal of a window or model the provider no longer reports — it retries next
+   * round, because `deliveredIds` is left untouched and the first round after a
+   * start compares against the DATABASE by design.
+   *
+   * @param runtime the account's runtime
+   * @param error the thrown error
+   */
+  handleSweepFailure(runtime, error) {
+    const message = error instanceof Error ? error.message : String(error);
+    if (runtime.sweepFailed) {
+      this.deps.log.debug(`${runtime.config.name}: the cleanup of vanished entries still fails (${message})`);
+      return;
+    }
+    runtime.sweepFailed = true;
+    this.deps.log.warn(
+      `${runtime.config.name}: the values were stored, but vanished entries could not be cleaned up (${message})`
+    );
+  }
   /**
    * Classify a fetch failure.
    *
@@ -388,7 +419,15 @@ class PollEngine {
     const { config } = runtime;
     const message = error instanceof Error ? error.message : String(error);
     if (error instanceof import_provider.FetchError && error.kind === "no-credentials") {
+      runtime.failCount = 0;
       runtime.serviceOnline = true;
+      if (runtime.state !== "not-signed-in") {
+        runtime.status.snapshot = void 0;
+        runtime.status.warning = false;
+        runtime.locked = false;
+        void this.deps.setStateChanged(`${config.id}.warning`, false);
+        void this.deps.setStateChanged(`${config.id}.limitReached`, false);
+      }
       runtime.state = "not-signed-in";
       runtime.error = message;
       if (!runtime.authNotified) {
@@ -398,6 +437,7 @@ class PollEngine {
       return;
     }
     if (error instanceof import_provider.FetchError && error.kind === "auth") {
+      runtime.failCount = 0;
       runtime.serviceOnline = true;
       runtime.state = "unauthorized";
       runtime.error = `Sign-in rejected \u2014 ${message}`;
@@ -414,6 +454,7 @@ class PollEngine {
       return;
     }
     if (error instanceof import_provider.FetchError && error.kind === "rate-limit") {
+      runtime.failCount = 0;
       runtime.serviceOnline = true;
       runtime.state = "rate-limited";
       runtime.error = `Throttled by the provider \u2014 retrying in ${Math.round(runtime.backoffMs / 6e4)} min, last values kept`;
@@ -434,6 +475,16 @@ class PollEngine {
       runtime.error = `The AI service reports a fault \u2014 ${message}`;
       return;
     }
+    if (!(error instanceof import_provider.FetchError)) {
+      runtime.failCount = 0;
+      if (runtime.serviceOnline) {
+        this.deps.log.warn(`${config.name}: the answer could not be processed (${message}) \u2014 values kept`);
+      }
+      runtime.serviceOnline = false;
+      runtime.state = "service-down";
+      runtime.error = `The answer could not be processed \u2014 ${message}`;
+      return;
+    }
     runtime.failCount++;
     this.deps.log.debug(`${config.name}: fetch failed (${message}), attempt ${runtime.failCount}`);
     if (runtime.failCount >= MAX_NETWORK_FAILURES) {
@@ -455,12 +506,18 @@ class PollEngine {
    * called "last successful update" must never do (measured 2026-09-07 while
    * writing its description).
    *
+   * It hung on `runtime.state === "ok"` alone, and the network branch only sets the
+   * state on the THIRD strike — so the first two tolerated failures left the state
+   * on "ok" and re-dated values the round had never fetched. The caller now says
+   * whether this round actually brought a snapshot into the tree.
+   *
    * @param runtime the account's runtime
+   * @param delivered whether this round wrote a fetched snapshot
    */
-  writeAccountInfo(runtime) {
+  writeAccountInfo(runtime, delivered) {
     const { config } = runtime;
     void this.writeAccountStatus(runtime);
-    if (runtime.state === "ok") {
+    if (delivered) {
       this.deps.setState(`${config.id}.info.lastUpdate`, new Date(this.deps.now()).toISOString());
     }
   }