FMG-11: Add ChatGPT as a third usage provider #1
2026-09-11
Fix the app-server handshake against the documented protocol
The previous commit's messages were rejected by Codex. Checked against the app-server documentation rather than guessing again: - app-server omits the `"jsonrpc": "2.0"` header on the wire; we were sending it on every message. - `initialize` takes `params.capabilities.experimentalApi`, which part of the server's surface is gated behind. We sent only `clientInfo`. - The documented invocation is a bare `codex app-server`. The `-s read-only -a never` prefix was defensive but undocumented, and nothing here starts a thread or runs a turn for it to constrain. Also fixed alongside: - A PATH-only install was unreachable: `URL(fileURLWithPath: "codex")` resolves against the working directory, so the fallback now goes through `/usr/bin/env`. - stdout and stderr are drained as they fill instead of after exit, which removes both the pipe-buffer deadlock and the full timeout wait for a server that outlives its stdin. The request returns as soon as its response lands. - Failures now carry the exit status and the stderr tail instead of a bare "could not parse", so the next report says what actually happened. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Read ChatGPT usage through the Codex CLI only
The ChatGPT provider no longer touches credentials at all. Usage now comes from `codex app-server`, the CLI's stdio JSON-RPC interface: handshake, then `account/rateLimits/read`. Codex owns authentication end to end, so this works wherever it keeps its credentials — file, keyring or ephemeral — and the app never reads, stores or forwards a token. Removed accordingly: - Reading ~/.codex/auth.json and the CODEX_HOME environment variable. - The manual-token and Admin API key modes for this provider, along with the OpenAI organization usage/cost metrics. ChatGPT now offers Auto only, and Settings no longer queries the Keychain for a provider that cannot use it. - The shared JWT helper, which existed only to read claims out of Codex tokens; CursorAuthService is back to its original form. This also drops the HTTP path, whose endpoint was wrong: the earlier commit used /backend-api/codex/usage, while Codex actually polls /backend-api/wham/usage with a different response shape. The app-server command is documented as experimental and `account/rateLimits/read` is undocumented, so the decoder accepts the windows wrapped in `rateLimits` or inlined at the root, and an absolute `resetsAt` or a `resetsInSeconds` countdown. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Add ChatGPT as a third usage provider
Adds a ChatGPT provider alongside Cursor and Claude, following the same provider/auth/metric structure: - Auto mode reads the Codex CLI session from ~/.codex/auth.json (CODEX_HOME aware) and queries the ChatGPT usage endpoint for the 5-hour and weekly rate-limit windows. - Manual mode accepts a pasted ChatGPT access token; the account id and plan are recovered from the token's own claims. - Admin mode reports 7-day input/output tokens and cost from the OpenAI organization usage and costs APIs. Supporting changes: - Extract base64url JWT payload decoding into a shared JWT helper, reused by CursorAuthService. - Replace the duplicated provider lookup in UsageRefreshService and SettingsView with ProviderID.provider. - Build the Settings provider tabs from ProviderID.allCases. - Providers missing from previously saved settings now start disabled, so the upgrade does not add an unconfigured indicator to an existing menu bar. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>