@buyong/pi-codex-fast-mode

English | 한국어

A Pi extension that requests an OpenAI service tier while keeping the same conversation model and endpoint. Choose Standard, Fast, or Ultrafast per model, with a global speed for models without an override. Session speed takes precedence over model settings, then global speed. The extension adds service_tier only when its local compatibility checks pass; initial Standard without an explicit preference leaves the payload unchanged.

The selected tier applies to requests. Response tiers can differ; the Codex backend has returned default for Fast and Ultrafast requests. A default response alone does not disable the selection or cause a failure. Check processing speed and billing separately. See the OpenAI Fast and Ultrafast guides.

Requirements

Use Node.js 22.19 or later and Pi 1.0.0 or later with configured model access. This repository builds against Pi 1.3.0. For Fast, select an available model in the local allowlist; Ultrafast has additional API, authentication, and endpoint requirements. The list does not supply account access or establish backend availability.

Quick start

If you already use @buyong/pi-codex, it includes this extension; do not install or enable it separately. Otherwise, install a published release:

pi install npm:@buyong/pi-codex-fast-mode
pi list
pi

pi list shows npm:@buyong/pi-codex-fast-mode and its installed path under User packages:.

In Pi:

1. With /model, select a model in the local Fast allowlist.
2. Run /codex-fast. The view opens on the session row, so press ↓ to reach your current model's row, select Fast under Speed with → or Tab, press Enter to save, and press Esc to close the view.
3. Run /codex-fast status to confirm that requests for the desired tier are enabled on the selected model.
4. Send a short prompt to make a model request, for example:

Reply with READY only. Do not use tools.

Run /codex-fast status again. OpenAI requested service tier identifies the selected request tier, Request enabled means the extension's local checks passed, and Last payload injection shows priority after a matching request. This record shows the field the extension added to the request; it does not confirm the tier the server actually used. While the choice is enabled, Pi's footer also shows the model and tier, such as gpt-6.1-sol fast. To disable acceleration for this model, save Standard as its model-specific speed.

Select a tier

/codex-fast shows Session speed (not saved) first, then the available models selected in /scoped-models that can request Fast or Ultrafast (models with speed controls), followed by Global speed. Without a configured scope, it uses all available models, matching Pi's behavior. Models are sorted by Astra, Sol, then Luna family, descending numeric version within each family, then model key; other models follow in model-key order. The current model has no special sort priority. Models outside the scope are not added, even if currently selected. Under each model, Speed offers one choice among Standard, Fast, and Ultrafast. Models without an override show global beside their name and inherit the global speed.

After changing /scoped-models, reopen /codex-fast to see the updated list. Speed settings for models hidden from the list are preserved.

• ↑/↓: select session, model, or global speed. Long model lists scroll with the selection.
• ←/→ or Tab/Shift+Tab: select a speed. Brackets mark the session selection or saved model/global value; ›…‹ marks keyboard focus.
• Enter: apply the session row without saving, or save a model/global speed. The view stays open.
• R: clear session speed on the session row, or restore global inheritance on a model row.
• Esc: close. Choices not applied with Enter are discarded.

Unavailable model speeds are dimmed, marked ×, and skipped during selection. The session row offers all three speeds, retaining the choice even when the current model cannot use it. Narrow terminals show the saved speed and pending choice in a compact form.

Commands with arguments also work without a UI. Replace <provider/id> with an actual model ID such as openai-codex/gpt-6.1-sol. Speed changes wait for the current agent run to become idle. status does not save settings or make a model request.

In RPC mode, /codex-fast first asks whether to change session, model, or global speed, then opens a speed picker. The session picker includes a clear option, and model pickers include global inheritance.

Starting Pi with --fast uses global Fast instead of the saved global speed, without saving it. Model settings still take precedence.

Session speed overrides model settings and global speed (including --fast). It stays in memory only, regardless of persistState, and survives model switches with compatibility recalculated for each model. It is cleared on process exit, /new, /resume, /fork, or /reload; session startup returns to saved settings and any startup flag. Use /codex-fast session clear to return without changing sessions. While set, the footer includes · session (and unsupported when acceleration cannot be requested), and status explains that saved speeds are overridden. Global/model commands and TUI saves still save normally, but their result messages note that session speed takes precedence.

Without session speed, switching models recalculates the model-specific or inherited global speed. An unsupported accelerated preference is retained without adding a tier. Model-specific commands check the target model's compatibility before changing its setting.

Settings and persistence

agentDir honors PI_CODING_AGENT_DIR and normally is ~/.pi/agent. The project layer overrides the global layer. Saving a global or model speed, from the /codex-fast view or a command, writes to the project file if it exists, otherwise to the global file. Session speed never writes either file.

{
  "serviceTier": "standard",
  "modelServiceTiers": {
    "openai-codex/gpt-6.1-sol": "ultrafast",
    "openai-codex/gpt-6-astra": "fast"
  },
  "persistState": true,
  "notifyOnModelSwitch": true
}

Each file's global speed is resolved on its own first, so legacy booleans in the project file can override serviceTier in the global file. Model settings merge by ID, with project values taking precedence for the same model. A global-file model override also takes precedence over a project's global speed. A project null entry masks the global-file model override and restores global-speed inheritance.

Global changes write serviceTier together with the legacy booleans: desiredActive is true for Fast or Ultrafast, and active records whether the tier is active for the current model. Within a file, serviceTier still takes precedence. Model changes update only the target ID while preserving unrelated fields. Restoring inheritance removes that model's entry from the global file, or writes null for it in the project file. Concurrent saves to the same file are serialized across extension instances and Pi processes, preserving other models and unrelated fields. If a save cannot acquire the lock within two seconds, it reports an error and restores the previous choice; retry after the other save finishes. Other failed saves also restore the previous choice. Model switches recalculate speed and activation without writing settings.

After manual edits, run /reload. Invalid or unreadable settings report Could not load Fast settings: …; the extension then starts with defaults and keeps later changes in memory only, without writing either settings file.

Supported models and payloads

This is the extension's default local Fast allowlist, not a promise that your provider/account offers every listed model or service tier:

Ultrafast can be requested for gpt-6-astra or gpt-6.1-sol through these routes:

Use the default HTTPS port or 443, without URL credentials, query, or fragment. A trailing / is accepted. Activation on the Codex route means the extension can prepare the request. The Codex response tier alone does not establish account support or the processing tier actually used.

The extension adds service_tier only to a request payload that is an object whose payload.model matches the current model ID. Auxiliary requests for a different model are not changed. Unsupported cases are left unchanged; the extension does not change endpoints, select another model, or verify the backend's actual tier.

The Sol service-tier PoC (in Korean) checks the final HTTP body produced by Pi and a completed response, recording request tiers, raw response tiers, and tokens per second (TPS) separately. A response-tier difference alone does not fail the run; success does not guarantee the processing tier or billing.

When the selected tier is not active

Migrating from 0.1.0 commands

Version 0.1.1 and later register only /codex-fast. The 0.1.0 commands /openai-tier and /openai-settings were removed, not kept as aliases, so update scripts that call them. Use /codex-fast status for diagnostics, model commands for model overrides, and speed commands for global speed. on/off and the named-mode shortcuts still change global speed. Settings saved by 0.1.0, including legacy booleans and storage paths, still load.

Build and load from source

Run from the pi-packages repository root:

pnpm install
pnpm --filter @buyong/pi-codex-fast-mode build
pnpm --filter @buyong/pi-codex-fast-mode exec pi -ne -e ./dist/index.js

This loads the extension for one invocation without registering a permanent package source. Rebuild after source changes.

To check a source build without a model request:

1. From apps/pi-codex-fast-mode, run pnpm run typecheck, pnpm run lint, and pnpm run build. All three should exit successfully.
2. Inside Pi, run /codex-fast session fast, then /codex-fast status. The status shows the session override; no settings file changes.
3. Run /codex-fast session clear or /reload, then status again. It shows Session speed: not set.

The repository release process describes release preparation.

License

Apache-2.0. The npm archive includes LICENSE.