pi-telegram-notifier 0.1.6 → 0.1.7

package/README.md

@@ -40,6 +40,7 @@ The **quota/credit** field supports these Pi provider IDs (using the credential
 | Pi provider ID | Mode | Required credential | What is shown |
 | --- | --- | --- | --- |
 | `openai-codex` | Subscription | ChatGPT OAuth | Remaining 5h / 7d usage-window percentages; [official Codex client endpoint](https://github.com/openai/codex/blob/main/codex-rs/backend-client/src/client/rate_limit_resets.rs). |
+| `openai` | Subscription (Codex fallback) | ChatGPT OAuth **and** a separate `openai-codex` OAuth login | Codex account windows, explicitly labeled `(Codex login)`, not the new login's app-specific limit. The new ChatGPT token cannot read this endpoint. Log both providers into the same account; without the Codex login, `Quota unavailable`. API keys are not supported. |
 | `anthropic` | Subscription | Claude OAuth | Remaining 5h / 7d usage-window percentages; [private OAuth endpoint](https://github.com/openclaw/openclaw/blob/main/src/infra/provider-usage.fetch.claude.ts). |
 | `claude-bridge` ([pi-claude-bridge](https://www.npmjs.com/package/pi-claude-bridge)) | Subscription | Pi's Claude OAuth (`/login` → Anthropic) | Same 5h / 7d windows as `anthropic`: the bridge has no credential of its own, so the matching `anthropic` model's OAuth login is used. Log Pi and Claude Code into the **same** Anthropic account, otherwise Pi's account quota is shown. Without the login, `Quota unavailable`. |
 | `xai` | Subscription | Grok OAuth | Remaining weekly shared Grok credits percentage, **not** prepaid API credit; [official Grok client endpoint](https://github.com/xai-org/grok-build/blob/main/crates/codegen/xai-grok-shell/src/extensions/billing.rs). |

package/docs/providers.md

@@ -6,6 +6,7 @@ Inventory of the 41 built-in provider IDs in the installed Pi model catalog (`@e
 | --- | --- | --- |
 | `openrouter` | Account credit = `total_credits - total_usage` (USD), **only if Pi's credential is an OpenRouter management key**; normal inference keys receive HTTP 403 and show unavailable. | [Get remaining credits](https://openrouter.ai/docs/api/api-reference/credits/get-remaining-credits), `GET https://openrouter.ai/api/v1/credits` |
 | `openai-codex` | Remaining subscription quota for the primary and secondary usage windows (typically 5h / 7d), from `100 - used_percent`; Pi's OAuth token and ChatGPT account ID are required. This backend route is used by the official Codex client but is **not a stable public API**. | [Official Codex client usage request](https://github.com/openai/codex/blob/main/codex-rs/backend-client/src/client/rate_limit_resets.rs), [response fields](https://github.com/openai/codex/blob/main/codex-rs/codex-backend-openapi-models/src/models/rate_limit_status_details.rs), `GET https://chatgpt.com/backend-api/wham/usage` |
+| `openai` | With ChatGPT OAuth, reads the matching model's separate `openai-codex` OAuth login through Pi's registry and labels the result `(Codex login)`. Pi's new direct ChatGPT token cannot read WHAM usage and does not expose the legacy account ID. Account equality cannot be verified: sign both providers into the same account. Missing Codex login, API keys and proxy endpoints show unavailable. | Same official Codex endpoint as above; [Pi ChatGPT OAuth implementation](https://github.com/earendil-works/pi/blob/main/packages/ai/src/auth/oauth/openai-chatgpt.ts). |
 | `deepseek` | Available balance per currency (CNY and/or USD); no currency conversion. | [Get User Balance](https://api-docs.deepseek.com/api/get-user-balance), `GET https://api.deepseek.com/user/balance` |
 | `moonshotai` | Available balance, USD. | [Balance](https://platform.moonshot.ai/docs/api/balance), `GET https://api.moonshot.ai/v1/users/me/balance` |
 | `moonshotai-cn` | Available balance, CNY. | [Balance](https://platform.moonshot.cn/docs/api/balance), `GET https://api.moonshot.cn/v1/users/me/balance` |
@@ -15,9 +16,11 @@ Inventory of the 41 built-in provider IDs in the installed Pi model catalog (`@e
 | `claude-bridge` (extension provider from [pi-claude-bridge](https://github.com/elidickinson/pi-claude-bridge)) | Same Claude subscription windows as `anthropic`. The bridge registers a placeholder key and runs Claude Code with its own login, so the quota is read with Pi's `anthropic` OAuth credential for the same model ID; without that login it shows unavailable. Claude Code's own credential store is deliberately not read, so the quota belongs to Pi's Anthropic account: it matches the bridge turn only when Pi and Claude Code are logged into the same account (the usage response does not identify the account). | Same endpoint as `anthropic` |
 | `xai` | Remaining **weekly Grok subscription** credits with Pi OAuth only. The per-product percentages are not the shared pool; xAI API prepaid credit is a different quantity requiring a management key. | [Official Grok billing client](https://github.com/xai-org/grok-build/blob/main/crates/codegen/xai-grok-shell/src/extensions/billing.rs), `GET https://cli-chat-proxy.grok.com/v1/billing?format=credits` (private endpoint); [API management billing](https://docs.x.ai/developers/rest-api-reference/management/billing) |
 | `amazon-bedrock`, `azure-openai-responses`, `cloudflare-ai-gateway`, `cloudflare-workers-ai`, `google-vertex`, `radius`, `vercel-ai-gateway` | Cloud/gateway billing belongs to separate accounts/credentials; no verified remaining plan balance endpoint with Pi's inference credential. | Provider-specific cloud billing tools, not Pi model API keys. |
-| `openai`, `github-copilot` | Subscription balance not exposed through a verified quota endpoint accessible with Pi's resolved inference credential. Admin usage/spending APIs are **not** plan remainder. | [OpenAI organization costs](https://platform.openai.com/docs/api-reference/organization/costs) |
+| `github-copilot` | Subscription balance not exposed through a verified quota endpoint accessible with Pi's resolved inference credential. Admin usage/spending APIs are **not** plan remainder. | [OpenAI organization costs](https://platform.openai.com/docs/api-reference/organization/costs) |
 | `fireworks`, `together` | Documented billing usage / quotas are past spend or request limits, not an available credit balance verified for inference keys. | [Fireworks billing summary](https://docs.fireworks.ai/api-reference/get-billing-summary), [Together billing usage](https://docs.together.ai/reference/billing-usage) |
 | `ant-ling`, `baseten`, `cerebras`, `google`, `groq`, `huggingface`, `kimi-coding`, `meta`, `minimax`, `minimax-cn`, `mistral` | No verified official endpoint returning remaining allowance with Pi's inference credential. | Check provider billing documentation before adding integrations. |
 | `nvidia`, `opencode`, `opencode-go`, `qwen-token-plan`, `qwen-token-plan-cn`, `qwen-token-plan-individual`, `xiaomi`, `xiaomi-token-plan-ams`, `xiaomi-token-plan-cn`, `xiaomi-token-plan-sgp` | No verified official endpoint returning remaining allowance with Pi's inference credential; management/billing credentials, if available, are distinct. | Check provider billing documentation before adding integrations. |
 
+The new `openai` ChatGPT login has no verified numerical quota endpoint. OpenAI documents [public Responses API inference](https://developers.openai.com/siwc/token-sharing-open-source/models-and-inference) and [usage-limit errors](https://developers.openai.com/siwc/token-sharing-open-source/errors-and-recovery), directing users to ChatGPT settings → Usage. App-specific limits can apply even when the account has remaining quota, so the Codex fallback does **not** measure the new login's app allowance. Live checks returned 401 for WHAM and `/v1/usage`, and 404 for `/v1/usage/limits`, `/v1/rate_limits`, and `/api/codex/usage` on `api.openai.com`. Other projects report the same limitation: [pi-minimal-footer PR #50](https://github.com/ogulcancelik/pi-extensions/pull/50), [the-last-harness #707](https://github.com/diegopetrucci/the-last-harness/issues/707). No browser cookies or Codex CLI credential files are read by this extension.
+
 An HTTP 401/403, invalid response or missing credential always yields `Quota unavailable` without suppressing the notification. OpenRouter's `/api/v1/key` reports a **per-key** limit, not remaining account credit: deliberately not used. This inventory is for the Pi catalog present at implementation time; new providers require fresh verification.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-telegram-notifier",
-  "version": "0.1.6",
+  "version": "0.1.7",
   "description": "Telegram notifications for final Pi agent responses and provider balances",
   "type": "module",
   "license": "MIT",

package/src/quota.ts

@@ -23,6 +23,20 @@ function remainingPercent(used: unknown): number | undefined {
 export async function getQuota(model: Model | undefined, modelRegistry: Pick<ModelRegistry, 'find' | 'getApiKeyAndHeaders' | 'isUsingOAuth'> | undefined, fetchImpl: typeof fetch = fetch): Promise<string> {
   // pi-claude-bridge has no credential of its own; it draws on the same Claude subscription as Pi's Anthropic OAuth login.
   if (model?.provider === 'claude-bridge') model = modelRegistry?.find('anthropic', model.id);
+  if (model?.provider === 'openai') {
+    try {
+      if (new URL(model.baseUrl).origin !== 'https://api.openai.com' || !modelRegistry?.isUsingOAuth(model)) return UNAVAILABLE;
+      const auth = await modelRegistry.getApiKeyAndHeaders(model);
+      if (!auth.ok || !auth.apiKey || (auth.baseUrl && new URL(auth.baseUrl).origin !== 'https://api.openai.com')) return UNAVAILABLE;
+      // shortcut: reports legacy Codex quota, replace when SIWC exposes a supported quota endpoint.
+      const codex = modelRegistry.find('openai-codex', model.id);
+      if (!codex || codex.provider !== 'openai-codex' || !modelRegistry.isUsingOAuth(codex)) return UNAVAILABLE;
+      const quota = await getQuota(codex, modelRegistry, fetchImpl);
+      return quota === UNAVAILABLE ? quota : `${quota} (Codex login)`;
+    } catch {
+      return UNAVAILABLE;
+    }
+  }
   const provider = model && providers[model.provider];
   if (!model || !provider) return UNAVAILABLE;
   try {