pi-xai

Pi extension for xAI Grok Build — Responses API, native OAuth, 5 tools + automatic agentic mode.

Note: Full access to Grok Build requires a SuperGrok Heavy subscription (see below).

Features

• grok-build provider — /login grok-build (native Web PKCE OAuth recommended with browser + manual-paste fallback, or Device Code for headless; no binary), Responses API (openai-responses), full reasoning
• xai_generate_text — stateful conversations, structured JSON output, built-in tools (with advanced filters), reasoningEffort, previousResponseId, custom timeout
• xai_multi_agent — 4/16-agent research (reasoningEffort), live progress updates via onUpdate
• Agentic mode — when any grok-* model is active, the extension auto-injects web_search / x_search / code_execution; the model decides what to call (the "magic")
• Full OAuth implementation: Web PKCE (OIDC discovery, PKCE, callback server with CORS + manual input) + Device Code fallback, JWT expiry + per-key refresh lock, improved ~/.grok/auth.json parsing (canonical + legacy), optional import from Grok CLI

Models (via /model grok-build/...): grok-build (primary Coding Plan alias), grok-4.3, grok-4.3-latest.

Everything works with Grok Build OAuth or plain XAI_API_KEY.

Getting Grok Build Access

The grok-build provider and model require a SuperGrok Heavy subscription.

• SuperGrok has a free 3-day trial.
• When upgrading from SuperGrok, look for the limited offer to get SuperGrok Heavy for $99/month for the first 6 months.

Quick start

pi install npm:pi-xai
# then in Pi:
/login grok-build
/model grok-build/grok-build

Ask anything. Agentic mode handles research for you, or call any of the five explicit tools (xai_generate_text, xai_multi_agent, xai_web_search, xai_x_search, xai_code_execution) for precise/direct control.

Tools

| Tool | Description | | -------------------- | ---------------------------------------------------------------- | | xai_generate_text | Responses API text gen with reasoning, JSON schema, tools, state | | xai_multi_agent | Deep research (4 or 16 agents), progress, multi-turn | | xai_web_search | Web search via Grok (prompt simulation for current knowledge) | | xai_x_search | X/Twitter search via Grok (prompt simulation) | | xai_code_execution | Python code analysis/execution simulation via Grok |

Agentic mode config (optional)

~/.pi/agent/settings.json (or .pi/settings.json):

{
  "xai": {
    "text": {
      "agentic": true,
      "agenticTools": ["web_search", "x_search"]
    }
  }
}

Default = enabled with all three built-in tools.

Payload normalization modes (compatible vs aggressive)

The extension supports two payload normalization modes for grok-* models (via the openai-responses driver + our hook). Controlled by xai.payloadMode in ~/.pi/agent/settings.json (or project .pi/settings.json):

{
  "xai": {
    "payloadMode": "compatible"   // default — or "aggressive"
  }
}

compatible (default)

• 100% unchanged behavior for existing users.
• Applies a defensive, enhanced content-element normalizer (catches content: [], "", whitespace-only, arrays of only malformed/empty/garbage parts, while safely preserving pure input_image vision messages and mixed content).
• Protects against the common 400 "Each message must have at least one content element" after tool turns or certain histories.
• Sufficient for the majority of users and the rich xai_* tools.

aggressive

• Opt-in only (no default change ever).
• Runs the full rewrite in the provider chat/agentic path (before_provider_request for grok-*):
  • The complete normalizeForXai (content guarantee + Hermes-like reasoning-item stripping).
  • Relocates developer/system messages to top-level instructions.
  • Stricter function_call_output.output sanitization (arrays → text with [image] placeholders + JSON fallback to avoid 422s).
• Hermes parity (substantially): The aggressive path now brings over the highest-impact anti-400 patterns from the 2026-05-19 exploration of the Hermes agent clone (/tmp/hermes-agent-clone):
  • Proactive stripping of every type: "reasoning" item in input (directly ports the is_xai_responses=True logic in codex_responses_adapter.py:_chat_messages_to_responses_input that skips replay of codex_reasoning_items / encrypted_content for xAI, plus the transport's include:[] behavior).
  • Eliminates "reasoning without following content item" (missing_following_item) and replay of encrypted blobs that xAI OAuth/SuperGrok rejects.
  • Post-strip content fix ensures any leftover pure-reasoning follower assistants are valid [{type:"output_text", text:""}].
  • Matches the exact root cause of recurring 400s on high-reasoning (grok-4.3), long tool-using sessions, and news-search follow-ups that survived prior compatible hardening.
• Sibling extensions and direct /responses callers get the full benefit by calling the exported normalizeForXai(inputArray) helper (see JSDoc in index.ts for the exact usage recipe).

When to switch to aggressive:

• You hit 400s on complex/high-reasoning/tool-heavy conversations despite the default.
• Developing or using sibling packages (pi-xai-imagine, etc.) that do direct Responses calls.
• Long multi-turn agentic workflows or "imagine" workspaces with heavy reasoning.
• You want the closest experience to how Hermes (the reference) handles xAI paths.

Important limitations (transparent due to architectural constraints):

• The rich built-in tools (xai_generate_text, xai_web_search, xai_x_search, xai_code_execution, xai_multi_agent) and any code path that goes through the internal callXaiResponses (including many sibling direct calls during development) always use only the enhanced-compatible normalization, even if you set payloadMode: "aggressive". They never receive the reasoning strip, relocation, or extra tool cleaning.
• For full aggressive/Hermes guarantees on your own direct Responses calls, explicitly call normalizeForXai(...) on the input array (the helper is the single source of truth and the recommended path for power users / siblings).
• Only the normal provider-driven grok-* chat + auto-injected agentic tools path gets the complete aggressive treatment.

See the detailed JSDoc on normalizeForXai and the payloadMode block in index.ts for implementation notes, every (as any) rationale, and the full constraint history (smallest possible diff, no new internal helpers, single grok-build provider, in-place only).

Development

Zero-config modern stack (oxfmt default, oxlint zero-config, tsgo only):

npm install
npm run check      # tsgo --noEmit
npm run lint       # oxlint .
npm run format     # oxfmt --write .
npm test

Husky + lint-staged runs oxfmt --write, oxlint, tsgo --noEmit on every commit.

Exactly 4 source files. See CHANGELOG.md.