@benvargas/pi-openai-verbosity

Per-model text verbosity overrides for pi's openai and openai-codex providers.

Why This Exists

This extension was originally created because pi sent OpenAI Codex provider requests with the default Responses API text verbosity, which made some models, especially gpt-5.5, noticeably more verbose than the Codex CLI. The goal was to align openai-codex/gpt-5.5 with Codex CLI behavior by setting:

{
  "text": {
    "verbosity": "low"
  }
}

pi has since shipped an upstream fix: OpenAI Codex Responses requests now default to low verbosity when no explicit verbosity is provided.

That upstream change addresses the original gpt-5.5 issue, but it also means pi now defaults every openai-codex model to low. That may not be ideal for all model slugs. For example, the Codex CLI's own default for gpt-5.4-mini is medium, not low, and you may prefer high on a model you use for long-form writing.

This extension now provides the missing user-facing control: per-slug verbosity settings for pi's openai-codex provider.

pi 1.x supersedes openai-codex (now "OpenAI Codex (legacy)") with Sign in with ChatGPT on the openai provider. That provider sends no text.verbosity, so requests run at the Responses API default (medium) rather than the Codex CLI's per-model defaults. The extension applies the same Codex CLI defaults to openai/<model> keys, so ChatGPT-backed sessions keep their Codex verbosity.

Requires pi 0.74.0 or newer.

What It Does

The extension uses pi's before_provider_request hook to rewrite outgoing provider payloads for configured openai/<model> and openai-codex/<model> keys.

For matching models, it sets:

{
  "text": {
    "verbosity": "low | medium | high"
  }
}

Non-matching models are left unchanged.

Install

pi install npm:@benvargas/pi-openai-verbosity

Or try without installing:

pi -e npm:@benvargas/pi-openai-verbosity

Usage

Run pi with the extension enabled:

pi -e npm:@benvargas/pi-openai-verbosity --model openai-codex/gpt-5.5

Use /openai-verbosity status inside pi to report the configured rewrite for the current model. The command also reloads the config file.

Config

Config files follow pi's project-over-global pattern:

• Project: <repo>/.pi/extensions/pi-openai-verbosity.json
• Global: <agent-dir>/extensions/pi-openai-verbosity.json

The global path uses pi's agent directory — ~/.pi/agent by default, or wherever PI_CODING_AGENT_DIR points if you have relocated it.

Older versions of this extension always used ~/.pi/agent, even when PI_CODING_AGENT_DIR was set. If you have relocated the agent directory and it contains no config, a config left at the legacy ~/.pi/agent/extensions/pi-openai-verbosity.json path is still honored; it is read in place and never modified. A config in the relocated directory takes precedence once you create one there.

If no config exists at any of these locations, the extension writes a default global config on first run.

Example config:

{
  "models": {
    "openai-codex/gpt-5.6-sol": "low",
    "openai-codex/gpt-5.5": "low",
    "openai-codex/gpt-5.4-mini": "medium",
    "openai-codex/gpt-5.3-codex-spark": "medium"
  }
}

Settings:

• models: object mapping openai/<model-id> or openai-codex/<model-id> strings to low, medium, or high.

Project config overrides global config per model key. Any model not listed is left unchanged, which means pi's native default behavior applies.

Default Config

By default, the extension mirrors the Codex CLI's own per-model verbosity defaults for every model in pi's openai-codex catalog and for the matching current models on openai. That means low everywhere except gpt-5.4-mini, whose upstream default is medium (pi's blanket openai-codex fallback would otherwise force it to low):

{
  "models": {
    "openai/gpt-5.3-codex-spark": "low",
    "openai/gpt-5.4": "low",
    "openai/gpt-5.4-mini": "medium",
    "openai/gpt-5.5": "low",
    "openai/gpt-5.6-luna": "low",
    "openai/gpt-5.6-sol": "low",
    "openai/gpt-5.6-terra": "low",
    "openai/gpt-6-astra": "low",
    "openai/gpt-6-luna": "low",
    "openai/gpt-6-sol": "low",
    "openai/gpt-6.1-sol": "low",
    "openai-codex/gpt-5.3-codex-spark": "low",
    "openai-codex/gpt-5.4": "low",
    "openai-codex/gpt-5.4-mini": "medium",
    "openai-codex/gpt-5.5": "low",
    "openai-codex/gpt-5.6-luna": "low",
    "openai-codex/gpt-5.6-sol": "low",
    "openai-codex/gpt-5.6-terra": "low",
    "openai-codex/gpt-6-astra": "low"
  }
}

You can change any value to override the default for that model. If you have an existing config file from an older version of this extension, your file's entries win over these defaults, and entries for model slugs pi no longer exposes are simply ignored.

Why not samplingParams?

pi 0.84.0 added samplingParams, which lets you pass arbitrary OpenAI-compatible parameters through models.json, model overrides, and extension providers. Reading those release notes, it is natural to assume this extension is now redundant and that you can write:

{
  "modelOverrides": {
    "openai-codex/gpt-5.5": {
      "samplingParams": { "text": { "verbosity": "medium" } }
    }
  }
}

On openai-codex this silently does nothing. There is no error and no warning; the setting is simply never applied. samplingParams is explicitly scoped to OpenAI-compatible adapters — pi-ai documents it as "only applied by OpenAI-compatible adapters (completions, responses, Azure responses); other APIs ignore it" — and the Codex adapter is not one of them. It builds its own request and hardcodes text: { verbosity: ... }, defaulting to low.

So for the openai-codex provider, this extension remains the only way to set per-model verbosity.

If you use samplingParams on the plain openai provider

There it is applied, so samplingParams is a native alternative for openai models. However, the merge is a shallow top-level Object.assign. That means:

{ "samplingParams": { "text": { "verbosity": "medium" } } }

replaces the entire text object rather than merging into it, discarding any other text fields pi had already set. This extension deliberately does the opposite: it preserves existing text fields and replaces only text.verbosity.

Debugging

Pi does not currently expose a simple CLI flag to print the final provider request body. To verify this extension is matching and rewriting a request, set PI_OPENAI_VERBOSITY_DEBUG_LOG to a JSONL file path.

PI_OPENAI_VERBOSITY_DEBUG_LOG=/tmp/pi-openai-verbosity.jsonl \
  pi -e npm:@benvargas/pi-openai-verbosity \
  --model openai-codex/gpt-5.6-sol \
  -p "Reply in one short sentence."

Then inspect the last entries:

tail -n 5 /tmp/pi-openai-verbosity.jsonl | jq .

These entries include prompts, messages, tools, and the rest of the provider payload, so keep the file local and delete it when you are done debugging.

Notes

• This extension only changes outgoing provider request payloads.
• Existing text fields are preserved, and only text.verbosity is replaced.
• The openai and openai-codex providers are supported. pi's samplingParams does not reach openai-codex, so this extension remains the only per-model verbosity control for it — see Why not samplingParams?.
• On openai, the shipped defaults change requests from the API default (medium) to the Codex CLI defaults. Set a model to medium in your config to keep the API default for it.
• Older OpenAI models such as gpt-4.1 reject text.verbosity, so only add models that support it.

Uninstall

pi remove npm:@benvargas/pi-openai-verbosity

License

MIT