@bumpyclock/pi-codex-compat

One Pi extension that consolidates OpenAI/Codex-focused behavior:

• /fast priority service tier (service_tier: "priority") for verified OpenAI Codex Responses requests
• Optional /ultrafast service tier for supported OpenAI/Codex models (/utrafast alias), off by default
• OpenAI native compaction replay for openai and openai-codex Responses APIs
• Named OpenAI Codex OAuth accounts (/accounts)
• Live Codex subscription quota (/usage)
• Historical Pi/Codex CLI cost report (/usage-history)
• Token throughput status (all providers)
• Unified /codex manager menu

Minimum tested Pi release: 0.82.1 (@earendil-works/*).

Install

pi install npm:@bumpyclock/pi-codex-compat

Or from a local checkout:

pi install /absolute/path/to/pi-codex-compat

Important migration notes

Do not load this package together with:

• local openai-codex-fast-mode.ts
• local openai-compaction
• local usage.ts / token-rate-status.ts
• pi-accounts / pi-codex-accounts
• pi-usage / pi-codex-usage

Duplicate extensions can refresh the same rotating credential or register conflicting commands.

Disable or remove those local/extension copies before enabling this package.

Commands

Print and JSON modes throw a clear unsupported-mode error for these commands because Pi 0.82.1 has no observable extension-command output channel there. RPC receives a notification for TUI-only menus.

Settings

Canonical user settings:

~/.pi/agent/pi-codex-compat.json

Trusted project overrides (non-secret only):

<workspace>/.pi/pi-codex-compat.json

Private Codex account credentials:

~/.pi/agent/pi-codex-compat-accounts.json

Example:

{
  "fastMode": { "enabled": false, "tier": "priority" },
  "openaiNativeCompaction": {
    "enabled": true,
    "debug": false,
    "redactSensitiveData": true,
    "supportedProviders": ["openai", "openai-codex"],
    "supportedApis": ["openai-responses", "openai-codex-responses"]
  },
  "tps": {
    "enabled": true,
    "notifyOnComplete": true
  },
  "auth": {
    "disableApiKeyWhenCodexAuthenticated": false
  }
}

Ultrafast mode

Select gpt-6-astra or gpt-6.1-sol, then run /ultrafast on (or /utrafast on). Run /ultrafast off to disable it, or /ultrafast status to inspect the selected mode. Commands and the /codex menu share validation, persistence, and cost warnings.

Ultrafast injects service_tier: "ultrafast" only for these exact model IDs on openai / openai-responses or openai-codex / openai-codex-responses. Other providers, APIs, and models are unchanged. Unsupported selections cannot enable Ultrafast. Switching to an unsupported model after enabling it leaves the preference saved, but marks Ultrafast inactive in status and leaves requests unchanged. Returning to a supported model resumes Ultrafast requests.

Speed modes are mutually exclusive. /fast on replaces Ultrafast with priority mode; /ultrafast on replaces priority mode. Each command's off disables only its own mode. Alt+Shift+F continues to toggle priority mode, replacing Ultrafast when pressed. Neither mode changes reasoning effort, selects a different model, switches accounts, or falls back to API-key billing.

Settings stay backward compatible. Existing fastMode.enabled values retain their meaning, and an absent tier defaults to "priority". Enabling Ultrafast saves "fastMode": { "enabled": true, "tier": "ultrafast" } in the canonical user file. enabled: false disables both tiers. Recognized fields inherit through the existing trusted-project precedence rules. Explicit on and off selections persist even when trusted project settings already make the effective mode match. Turning off the other active tier is a no-op. Commands change the current session immediately after a successful user-setting write; trusted project overrides can take effect again on a settings reload or new session.

Eligibility and cost. OpenAI currently lists GPT-6 Astra and GPT-6.1 Sol Ultrafast for Pro $500 and eligible Enterprise/Edu workspaces when signed in through ChatGPT. Plus, Pro $100, Pro $200, and Business do not include it at launch, even with purchased credits. Pro $500 consumes included allowance at 8x Standard, then purchased credits at 6x Standard. Enterprise/Edu billing follows the workspace agreement. API-key access has separate Ultrafast pricing and rate limits. These are usage multipliers, not speed guarantees.

Pi 0.82.1 does not apply an Ultrafast service-tier cost multiplier. Pi cost estimates and historical reports can understate charges. Use OpenAI usage/billing for actual cost.

This extension requests a tier; it does not grant plan access or verify entitlement from plan-name strings. Provider access/rate-limit errors remain visible through Pi, with no automatic tier or authentication fallback. A successful command means the preference was saved, not that OpenAI accepted an Ultrafast request.

OpenAI supports HTTP and recommends persistent WebSockets for tool-heavy agents. No transport change is required to request the tier. Live acceptance and latency with an eligible Codex OAuth account have not been verified here.

Official references:

• Ultrafast API guide.
• ChatGPT Work and Codex eligibility and usage.
• Fast mode and Ultrafast FAQ.

Auth cost-safety policy

auth.disableApiKeyWhenCodexAuthenticated (default false) is an opt-in guard:

• When on and a Codex OAuth login exists (default Pi openai-codex OAuth in auth.json, or a named account selected by this extension), every exact provider openai model is hidden — including built-ins and models.json additions.
• Stale sessions still selecting openai are blocked on each turn, compaction, and summarizing tree navigation. Non-summarizing tree navigation stays allowed.
• openai-codex and unrelated providers are unchanged. The policy never auto-switches models and never mutates OPENAI_API_KEY or credential files.
• When auth storage cannot be read, presence is treated as unknown and the policy fails closed while enabled.
• Toggle from /codex → Settings. Status shows Direct OpenAI API key: allowed|blocked while Codex is logged in.

Precedence

Defaults → old package-local compaction settings → legacy global settings.json → canonical user settings → legacy project settings.json → canonical project settings → PI_OPENAI_NATIVE_COMPACTION_* environment overrides.

Canonical values win over legacy values at the same scope. Environment overrides win over all files. Validated recognized values are migrated into canonical files; legacy sources are retained for rollback.

Legacy sources consulted when canonical values are absent:

• ~/.pi/agent/settings.json → extensionSettings.openaiCodexFastMode
• ~/.pi/agent/extensions/openai-compaction/settings.json
• openaiNativeCompaction in global/project settings.json
• PI_OPENAI_NATIVE_COMPACTION_*
• pi-accounts.json → only the openai-codex provider section
• pi-codex-accounts.json / codex-accounts.json

Security and failure behavior

• Account activation failures fail closed for openai-codex (no silent fallback credential).
• When auth.disableApiKeyWhenCodexAuthenticated is enabled, unknown/corrupt Codex auth storage fails closed for direct openai API-key use (hide/block with a redacted error).
• The auth.json presence reader is extension-owned and read-only: it never creates or writes auth.json, and never logs token fields.
• Compaction failures fail open to Pi's normal compaction.
• Quota requests send credentials only to the official ChatGPT origin (https://chatgpt.com) and use bounded response reads.
• Debug artifacts redact authorization headers, API keys, JWT-like values, and exact credential strings.
• Credential files use private directory permissions and mode 0600 on POSIX.
• Pi extensions run with full user permissions. Only install packages you trust. See Pi's unsandboxed extension trust model.

Development

npm install
npm run format
npm run check
npm run pack:dry
npm run smoke:pi

Releases

Merges to main update a Release Please pull request. Merging that release pull request creates a GitHub release and publishes the matching package version through .github/workflows/release.yml.

The workflow uses npm trusted publishing with GitHub Actions OIDC and provenance; it does not require a long-lived npm token. Manual runs default to a dry run. Choose release only when package.json contains the version to release.

License

MIT. See LICENSE and NOTICES.md for adapted upstream attribution (Jordy Van Domselaar; narumiruna).