Skip to content

codex.md

Latest commit

 

History

History
267 lines (244 loc) · 19.4 KB

codex.md

File metadata and controls

267 lines (244 loc) · 19.4 KB
Raw
summary Codex provider data sources: OpenAI web dashboard, Codex CLI RPC, credits, and local cost usage.
read_when
Debugging Codex usage/credits parsing
Updating OpenAI dashboard scraping or cookie import
Changing Codex CLI RPC or diagnostic PTY behavior
Reviewing local cost usage scanning

Codex provider

Codex has three automatic usage data paths (OAuth API, web dashboard, CLI RPC) plus a manual CLI PTY diagnostic parser and a local cost-usage scanner. The OAuth API is the default app source when credentials are available; web access is optional for dashboard extras.

Data sources + fallback order

App default selection (debug menu disabled)

  1. OAuth API (auth.json credentials).
  2. CLI RPC through codex app-server.
  3. If OpenAI web extras are enabled and a matching OpenAI web session is available (Automatic or Manual cookies), dashboard extras load as a separate follow-up refresh and the source label becomes primary + openai-web.

Usage source picker:

  • Preferences → Providers → Codex → Usage source (Auto/OAuth/CLI).

CLI default selection (--source auto)

  1. OpenAI web dashboard (when available).
  2. Codex CLI RPC through codex app-server.

OAuth API (preferred for the app)

  • Reads OAuth tokens from ~/.codex/auth.json (or $CODEX_HOME/auth.json).
  • CodexBar never publishes refreshed native tokens into auth.json; when native credentials are stale, the explicit OAuth path delegates recovery to the Codex CLI, which owns that file. If the CLI is unavailable, the OAuth error is surfaced instead of mutating the shared file.
  • Calls GET https://chatgpt.com/backend-api/wham/usage (default) with Authorization: Bearer <token>.
  • The app reads reset-credit inventory once per refresh with a best-effort GET https://chatgpt.com/backend-api/wham/rate-limit-reset-credits using the same account-scoped OAuth context; the CLI requests it only when optional credits are included.
  • The menu and provider settings list every still-available expiry, while the optional credits setting controls nearing-expiry notifications. CodexBar does not redeem or modify reset credits.
  • rate_limit.primary_window / secondary_window map to the session/weekly lanes.
  • Suspicious weekly resets keep the last trusted usage while confirmation is pending. A successful refresh for the same account and workspace clears stale connectivity errors even when the reading is withheld; failed, cancelled, or superseded refreshes do not clear them. Cached usage, credits, and other accounts remain unchanged.
  • Credits-only updates preserve pending weekly-reset evidence in memory and account-snapshot storage, including when published credits are cleared. Candidate admission, expiry, boundary tolerances, and account guards remain unchanged; preserving evidence does not make an otherwise incompatible reset eligible for publication.
  • Debug logs in codex-weekly-reset-publication include fixed reason codes for delayed-candidate creation, pruning, revalidation, and account-scoped storage requests. They distinguish source/confidence, timing, boundary, identity/plan compatibility, and credit-inventory failures without logging account or credit identifiers. Codes describe the first rejected prerequisite; they do not relax confirmation policy or prove the cause of a past stale reading. storeRequested means the file-store call was made, not that a disk write succeeded.
  • additional_rate_limits[] (model-specific limits such as GPT-5.3-Codex-Spark) map to named UsageSnapshot.extraRateWindows entries. Spark uses stable codex-spark / codex-spark-weekly ids and Codex Spark 5-hour / Codex Spark Weekly titles. When the field is absent, the snapshot is unchanged.
  • Preferences → Providers → Codex → Show Codex Spark usage hides only the Spark rows in menus and the provider preview. It does not change fetching, history, notifications, widgets, credits, or other extra limits.

Optional external OAuth sources (off by default)

  • External Codex OAuth sources is a provider setting that must be enabled explicitly before CodexBar reads another application's OAuth file. It is off by default because this is a cross-application credential boundary.
  • Without an explicit $CODEX_HOME, native Codex auth wins first, followed by legacy ~/.config/codex/auth.json, then OpenCode's ~/.local/share/opencode/auth.json (or the equivalent XDG_DATA_HOME path).
  • An explicit $CODEX_HOME remains isolated; it never borrows credentials from those external locations.
  • External fallbacks accept OAuth token structures only; API-key entries are ignored. Usage probes never refresh or publish OAuth token material into a shared auth.json without a cross-writer publication contract. Stale native credentials can delegate to the CLI recovery path, while stale external credentials fail closed in every mode. Automatic mode also suppresses unscoped CLI fallback whenever a managed workspace is selected. Explicit managed-account workspace selection is stored in CodexBar's private managed-account metadata; it never edits the source auth.json or publishes an account_id change back to another application's credential file.
  • Stacked account refreshes retain each managed account's selected workspace through usage publication and menu matching, even when its auth file names a different default workspace. Changing the selected workspace while a refresh is running discards the old workspace's result.
  • Reusing OpenCode OAuth enables remote account quota, not OpenCode session token/cost ingestion. See OpenCode with Codex or OpenAI for the current history boundary.

Advanced profile-home accounts

  • Managed Codex accounts remain the default multi-account path.
  • Advanced users can add existing Codex homes to ~/.codexbar/config.json with providers[].codexProfileHomePaths.
  • Each configured path must be absolute or start with ~/, and point at a Codex home that contains auth.json.
  • CodexBar reads identity from the configured home, exposes it in the Codex account switcher, and scopes remote Codex fetches with CODEX_HOME.
  • Profile homes are not copied, reauthenticated, or removed by CodexBar.

Example:

{
  "id": "codex",
  "codexProfileHomePaths": [
    "~/.codex-work",
    "~/.codex-personal"
  ]
}

Same-email workspace labels

Account settings, the System Account picker, and the menu switcher retain the workspace name when it is available. If the same email and workspace label would appear more than once (including missing names or the “Personal” fallback), CodexBar adds a stable eight-character hash of the workspace identity. The hash stays the same when selecting or promoting that workspace and never exposes the full provider identifier. This is display-only; stored account metadata and credential selection are unchanged. Separate profile homes for the same workspace also include a hashed source identity, so their labels stay distinct without exposing paths. Compact switcher buttons keep the discriminator visible when space is limited, using additional rows when needed.

OpenAI web dashboard (optional, off by default)

  • Enable it in Preferences -> Providers -> Codex -> OpenAI web extras.
  • It exists for dashboard-only extras such as code review remaining, usage breakdown, and credits history.
  • It is intentionally opt-in because it loads chatgpt.com in a hidden WebView and can materially increase battery or network usage.
  • OpenAI web battery saver is a separate toggle. When enabled, routine background/settings-driven refreshes are reduced, but explicit manual refreshes still run.
  • OpenAI web battery saver currently defaults to off.
  • Preferences → Providers → Codex → OpenAI cookies (Automatic or Manual).
  • URL: https://chatgpt.com/codex/settings/usage.
  • Uses an off-screen WKWebView with a per-account WKWebsiteDataStore.
    • Store key: deterministic UUID from the normalized email.
  • WebKit store can hold multiple accounts concurrently.
  • Each WebView acquisition keeps ownership across asynchronous page preparation. Explicit store eviction invalidates that store's pending preparations, so stale success, failure, or timeout retry cannot displace a replacement view. Evict-all invalidates all pending preparations; ordinary lease release does not invalidate concurrent temporary views.
  • Leases retain their cleanup owner independently of the cache and release only once. Validated pages still support the brief reuse handoff; other releases schedule the existing deferred WebKit cleanup, including temporary views.
  • Cookie import (Automatic mode, when WebKit store has no matching session or login required):
    1. Safari: ~/Library/Cookies/Cookies.binarycookies
    2. Chrome/Chromium forks: ~/Library/Application Support/Google/Chrome/*/Cookies
    3. Firefox: ~/Library/Application Support/Firefox/Profiles/*/cookies.sqlite
    • Domains loaded: chatgpt.com, openai.com.
    • No cookie-name filter; we import all matching domain cookies.
  • Cached cookies: Keychain cache com.steipete.codexbar.cache (account cookie.codex, source + timestamp). Reused before re-importing from browsers.
  • Manual cookie header:
    • Paste the Cookie: header from a chatgpt.com request in Preferences → Providers → Codex.
    • Used when OpenAI cookies are set to Manual.
  • Account match:
    • Signed-in email extracted from client-bootstrap JSON in HTML (or __NEXT_DATA__).
    • If Codex email is known and does not match, the web path is rejected.
  • Web scrape payload (via OpenAIDashboardScrapeScript + OpenAIDashboardParser):
    • Rate limits (5h + weekly) parsed from body text.
    • Credits remaining parsed from body text.
    • Code review remaining (%).
    • Usage breakdown chart (Recharts bar data + legend colors).
    • Credits usage history table rows.
    • Credits purchase URL (best-effort).
  • Errors surfaced:
    • Login required or Cloudflare interstitial.

Codex CLI RPC (automatic CLI source)

  • Launches local RPC server: codex -s read-only -a never app-server.
  • JSON-RPC over stdin/stdout:
    • initialize (client name/version)
    • account/read
    • account/rateLimits/read
  • RPC reads are bounded: initialization has a longer startup budget, and normal requests have a shorter per-method timeout. On timeout, CodexBar closes the child codex app-server process's stdin and escalates from SIGTERM to SIGKILL after a bounded grace period, so the stdout reader unwinds and unresponsive children cannot linger.
  • Provides:
    • Usage windows (primary + secondary) with reset timestamps.
    • Credits snapshot (balance, hasCredits, unlimited).
    • Account identity (email + plan type) when available.
  • App-server errors are terminal for the CLI strategy, except when Codex includes a recoverable wham/usage JSON body in the error text.
  • If macOS blocks or quarantines the codex executable, CodexBar records the launch failure and skips background CLI launches for 30 minutes. Use a manual refresh after reinstalling or unblocking codex to retry immediately.
  • CodexBar also discovers the Codex CLI bundled with current ChatGPT and legacy Codex desktop apps, even when codex is absent from the shell PATH.
  • If managed Codex account login still reports a missing executable, turn on Show debug settings in Settings > Advanced, then check Settings > Debug > CLI Paths. When no Codex binary appears there, confirm codex --version works in Terminal, check which -a codex for stale duplicate installs, then run npm install -g --include=optional @openai/codex@latest before retrying Add Account.

Codex CLI PTY diagnostics (/status)

  • Manual/debug parser only; automatic background refresh and CodexBarCLI usage --source cli do not launch bare Codex TUI.
  • Kept for explicit diagnostics/parser coverage because bare codex TUI can start interactive auth and open browser tabs.
  • Parses rendered /status output:
    • Credits: line
    • 5h limit line → percent + reset text
    • Weekly limit line → percent + reset text
  • Detects update prompts and surfaces a "CLI update needed" error.

Account identity resolution (for web matching)

  1. Latest Codex usage snapshot (from RPC, if available).
  2. ~/.codex/auth.json (JWT claims: email + plan).
  3. OpenAI dashboard signed-in email (cached).
  4. Last imported browser cookie email (cached).

Credits

  • Web dashboard fills credits only when OAuth/CLI do not provide them.
  • CLI RPC: account/rateLimits/read → credits balance.
  • CLI PTY diagnostics can still parse Credits: from saved/manual /status output.

Cost usage (local log scan)

  • Menu source selection:
    • By default, a selected managed account keeps its own CODEX_HOME session history.
    • Local session cost estimates is a Codex-only opt-in that instead scans this Mac's ambient $CODEX_HOME (or ~/.codex) independently of quota, OAuth, web-dashboard, and administrator access.
    • Regular menu cost refreshes publish local session estimates even when global cost tracking is off. This does not enable other providers' cost scans; results still require the same provider configuration and history/account scope.
    • The local-only mode never makes a network request or uploads session content. It uses an existing local models.dev cache when available, then the bundled CostUsagePricing rates.
  • Source files:
    • Native Codex logs:
      • ~/.codex/sessions/YYYY/MM/DD/*.jsonl
      • ~/.codex/archived_sessions/*.jsonl (flat; date inferred from filename when present)
      • Or $CODEX_HOME/sessions/... + $CODEX_HOME/archived_sessions/... if CODEX_HOME is set.
    • Supported pi-compatible sessions:
      • ~/.pi/agent/sessions/**/*.jsonl
      • ~/.omp/agent/sessions/**/*.jsonl
  • Scanner:
    • Native Codex logs parse event_msg token_count entries and turn_context model markers; when both are present, turn_context is authoritative for the model bucket.
    • pi and OMP sessions count assistant-message usage rows and attribute openai-codex assistant usage to Codex.
    • pi-compatible assistant usage is bucketed by assistant-turn timestamp, so mixed-model sessions can contribute to multiple days/models correctly.
    • Matching assistant entry IDs within the same session are counted once across roots; distinct turns are retained.
    • Native conversation rows reuse the corrected cached per-file totals and existing pricing tables. They are hidden when pi-compatible usage joins the aggregate because the native-only rows would not reconcile with the merged total.
  • Cache:
    • Native session store: ~/Library/Caches/CodexBar/cost-usage/cost-usage.sqlite
    • pi-compatible session cache: ~/Library/Caches/CodexBar/cost-usage/pi-sessions-v8.json
    • Catch-up status reads progress metadata without loading historical usage JSON or replay bodies. Cached reports retain row-level pricing evidence and project/session details, but omit raw token snapshots, accumulator state, and replay bodies. File cursor metadata, including JSONL resume state, remains available for progress tracking. A native scan loads complete state once and carries a single-use receipt to save. The store reuses decoded rows only while the same connection, database identity and SQLite change observations remain valid, checking again under the writer lock. Filesystem/anchor and catch-up reconciliation still run at comparison time; a concurrent database change requests a rescan. Fresh database opens retain integrity validation.
    • Saved day/model aggregates group each file's usage rows in one pass per aggregate build. Packed token totals, authoritative costs (including zero), and standard/priority estimation buckets retain their existing meanings.
    • Priority trace scans resume after ordinary log pruning when enough distributed content anchors still match; changed source rows, replaced databases, or insufficient matching anchors require a fresh scan. Temporary trace-database failures retain the last validated report pricing and leave scan freshness unchanged for retry. Successful historical queries update their own pricing window independently of the live scan cursor, including results with no priority turns; validated pricing outside that window remains intact.
  • Window: configurable 1-365 day rolling history.
  • App cadence: regular timer-driven local-history refreshes have a 15-minute minimum (30 minutes in Low Power Mode). Manual disables the recurring refresh timer, not all scan activity: startup refreshes and pending Codex catch-up can still scan local history. Faster provider refreshes still update quota/status. The scanner's default 60-second debounce is a separate internal limit, bypassed by forced scans and catch-up passes; it is not the app's refresh cadence.
  • Usage & Spend catch-up remains inactive after a no-progress or error pause until you choose Refresh in the dashboard toolbar or catch-up panel. Opening the dashboard or receiving background updates does not retry those terminal pauses. Low-power and thermal pauses can still recover automatically; this retry policy does not change cached history or token accounting.
  • Inline cost charts preserve a slot for every day in that window, using the selected cost-bucket time zone and the snapshot's date. Missing days are zero only after history coverage is established; unscanned days and entries without prices remain unknown. Long windows fit within the menu width without dropping dates.
  • Hide personal information replaces project/source names with numbered labels and hides their paths in the cost-history submenu; Usage & Spend also masks project names. Costs, tokens, grouping, and stored history are unchanged, and disabling the setting restores the original labels. This is display masking, not data deletion or export sanitization.
  • While a bounded refresh catches up with new session history, established totals remain visible only for the same account, history window, and bucket time zone. An incomplete first scan never borrows another account's totals.
  • Pending local-history files receive a turn before fresh work, within the existing byte and duration limits. Unfinished files rotate behind waiting work, and the queue survives restarts without rebuilding compatible caches.
  • Parent-session discovery also resumes within those limits after the requesting fork files leave both scan roots. Stale pending path associations are reconciled in the existing cache; surviving forks with missing parents still retain their unresolved usage instead of being counted as complete.

Usage & Spend account rows

Settings → Usage & Spend performs a separate fixed 30-day scan for every visible Codex account. Each request freezes the account source, exact Codex home, authentication fingerprint, and cache identity before scanning. A missing or invalid home is omitted; it never falls back to ambient ~/.codex or to the global Codex token snapshot.

These account rows intentionally exclude pi and OMP sessions because their history is machine-local rather than owned by one Codex account. The normal Codex cost menu and CLI scan continue to include supported pi-compatible history. The dashboard labels its values as local estimates and keeps currencies separate.

Key files

  • Web: Sources/CodexBarCore/OpenAIWeb/*
  • CLI RPC + diagnostic PTY parser: Sources/CodexBarCore/UsageFetcher.swift, Sources/CodexBarCore/Providers/Codex/CodexStatusProbe.swift
  • Cost usage: Sources/CodexBarCore/CostUsageFetcher.swift, Sources/CodexBarCore/PiSessionCostScanner.swift, Sources/CodexBarCore/PiSessionCostCache.swift, Sources/CodexBarCore/Vendored/CostUsage/*