openai-params

OpenAI fast-mode, long-cache, and verbosity settings for pi.

What it does

This extension adds /openai-params, which opens a small settings screen for fast mode, long cache retention, and verbosity.

• Ctrl+S saves
• Esc cancels

Use the list to:

• toggling fast mode
• toggling 24-hour prompt cache retention
• setting verbosity to low, medium, high, or the default unset state

When enabled, it patches provider requests right before send:

• fast mode → service_tier=priority
• long cache → prompt_cache_retention=24h
• verbosity → text.verbosity=<level>

Behavior

• Fast mode is applied to any model whose ID starts with gpt- on the official openai and openai-codex providers when they use a compatible API:
  • openai-completions
  • openai-responses
  • openai-codex-responses
• Models from GitHub Copilot, local servers, and custom proxies are not patched merely because they use an OpenAI-compatible serializer.
• Long cache retention is applied only to the official openai provider with openai-completions or openai-responses.
• OpenAI Codex subscription models are excluded because their endpoint rejects prompt_cache_retention.
• Verbosity is applied only to OpenAI Responses-family APIs:
  • openai-responses
  • openai-codex-responses
  • azure-openai-responses
• The default verbosity setting is unset, so the extension does not send any text.verbosity field unless you choose one.

Config

Config uses the same project-over-global pattern as the fast-mode package:

• project: .pi/extensions/openai-params.json
• global: ~/.pi/agent/openai-params.json

If neither file exists, the extension creates the global file on first run.

Default config:

{
  "fast": false,
  "longCache": false,
  "verbosity": null
}

Older supportedModels settings are ignored because fast-mode support is detected from the active model provider, ID, and API.

Integration

This extension emits its current state on pi's extension event bus over pi:openai-params with:

• source
• cwd
• fast
• longCache
• verbosity

That lets other extensions, including status, show the active non-default fast/cache/verbosity settings for the current workspace.

When long cache retention is actually added to a request, the extension also emits pi:prompt-cache-retention with the effective TTL and request start time. Cache-aware extensions can therefore observe the applied behavior without duplicating OpenAI model support rules.

Notes

This extension combines the behavior of:

• @benvargas/pi-openai-fast for service_tier=priority
• Pi's long OpenAI cache retention for prompt_cache_retention=24h
• pi-verbosity-control for text.verbosity

based on OpenAI GPT-5 / Responses API parameter docs.