pi-websearch

A standalone Pi package that registers a small host-side web_search tool. The tool calls the configured codex-local OpenAI Responses endpoint with the native Responses web_search tool, then returns the nested search result to the parent Pi model.

The parent model decides the final answer format from the user's prompt. The plugin does not impose [1] citations, a Sources: section, a fixed list layout, or a title/URL template.

Architecture

Pi Agent / parent model
   ↓
pi-websearch web_search tool
   ↓
codex-local Responses endpoint
   ↓
Native Responses web_search
   ↓
Nested Codex output_text
   ↓
Tool result returned to the parent model
   ↓
Parent model writes the final answer

This is intentionally a thin adapter. It does not reimplement Codex citation formatting or final-answer rendering.

Responsibilities

Plugin

• Register web_search.
• Reuse the active codex-local model's endpoint and authentication.
• Send the original user request together with the search query.
• Use native Responses web_search.
• Extract response.output_text or message output_text content.
• Return that text unchanged as the tool result.
• Report request failures or empty results.
• Provide /web-search on, /web-search off, /web-search stream, /web-search display, and /web-search status.

Parent model

• Decide whether a search is needed.
• Decide the language and level of detail.
• Decide whether to use paragraphs, lists, tables, Markdown links, title/URL lines, [1] citations, or Sources:.
• Produce the final answer for the user.

What the plugin does not do

The plugin deliberately does not:

• insert or renumber [1], [2], etc.;
• generate a Sources: section;
• parse or reposition citation spans;
• rewrite Markdown links;
• deduplicate or relabel visible sources;
• append a synthetic source fallback;
• clean or transform the nested result text;
• render a separate TUI-only final entry;
• terminate the parent model's turn;
• provide /web-search format numbered;
• patch pi-ai, pi-tui, pi-coding-agent, or node_modules.

The nested result is intentionally passed through unchanged so the parent model can use it as research context and follow the user's original instructions.

Install

For the current local development package:

pi install npm:@lyd123qw2008/[email protected]

The package is enabled through Pi's settings package list. It does not require copying an extension into ~/.pi/agent/extensions/.

Configuration

The persistent settings control whether the plugin-owned search tool is enabled and whether the nested Responses request uses SSE streaming. Streaming defaults to true when omitted:

{
  "nativeWebSearch": {
    "enabled": true,
    "stream": true,
    "statusDisplay": "switch"
  }
}

statusDisplay defaults to switch, which keeps the persistent footer short: web-search: on. Other choices are mode, verbose, and hidden.

Inside interactive Pi:

/web-search on
/web-search off
/web-search stream on
/web-search stream off
/web-search stream status
/web-search display
/web-search display switch|mode|verbose|hidden
/web-search status

With stream: true, the plugin parses nested Responses SSE events and sends search progress through Pi's tool onUpdate callback. The final nested result is still returned unchanged to the parent model. With stream: false, the plugin waits for one complete JSON Responses payload.

The package expects the active model to be:

provider: codex-local
api: openai-responses

The base URL and authentication are taken from Pi's normal model configuration. No API key, proxy URL, session, or model credential is stored in this package.

Nested request

The plugin sends a small instruction block with the original request and the search query:

Use the native web search tool to answer the original user's request.
Return only the search result for the parent model.
Follow the original user's language, scope, count, and requested output format.
Preserve exact page titles and complete URLs when the user asks for them or when they are useful.
Do not mention this nested search call or add planning commentary.
Do not invent sources or URLs.

There is no fixed citation or output template.

Runtime behavior

The tool uses a normal Pi tool result and does not use terminate: true:

user request
  ↓
web_search tool execution
  ↓
(optional) nested Responses SSE progress updates
  ↓
search result tool message
  ↓
parent model response
  ↓
normal Pi rendering

The persistent footer shows only web-search: on by default. Use /web-search display to open a selector for the short switch-only view, the switch-plus-mode view, a verbose implementation view, or no footer status. /web-search status always shows the detailed configuration regardless of the selected footer display. The tool call renderer shows the actual query. When nested streaming is enabled, Pi can also show progress such as 正在搜索网页：...; those updates are UI/tool-execution updates, not additional user messages.

TUI, text, JSON, and RPC modes therefore share the same basic behavior.

Security and boundaries

• Only codex-local with openai-responses is targeted.
• Authentication is resolved through Pi's public ModelRegistry API.
• Native Responses web_search is used inside the nested request.
• The plugin does not use bash, Python, curl, browser tools, DuckDuckGo, or external search APIs as a fallback.
• The plugin does not modify Pi installation files or generated dependencies.
• If the nested request fails, the tool reports the failure instead of inventing a result.

Tests

npm test
node --check src/extract-responses-text.mjs
git diff --check

Tests cover the minimal response-text extraction and SSE paths:

• canonical response.output_text;
• message-content fallback;
• ignoring search-call and reasoning items;
• empty responses;
• SSE chunk boundaries and CRLF;
• multiline SSE data and the Responses [DONE] sentinel.