Agent Identity Generation — Grok

Goal

On the launch wizard (and optionally My Agents draft edit), a connected creator can press Generate with Grok to produce a complete Agent identity suited to the selected template type, including a matching profile image. All xAI calls stay server-side; the browser never receives XAI_API_KEY.

User flow

1. Creator selects a template (Trading, Research, DeFi, Content, Social, Community, Gaming, Utility, Custom).
2. Optional seed fields: vibe keywords, niche, tone, ticker preference, “include meme energy”, etc.
3. Presses Generate with Grok.
4. Server runs a two-phase job (may stream progress to the client):

- Phase A — Identity text via Responses API + structured JSON output. - Phase B — Avatar image(s) via Grok Imagine (/v1/images/generations or Responses image_generation tool), using an image prompt derived from Phase A and the template style pack.

1. Wizard form is filled with editable results; avatar gallery shows 1–N candidates.
2. Creator can Regenerate all, Regenerate avatar only, edit any field, or fall back to manual upload.
3. Accepting an avatar stores it on FORGE object storage / local uploads (MVP: local disk or configured bucket); temporary xAI URLs are downloaded immediately and not relied on as permanent links.

Template style packs

Each template ships a code-owned style pack used in both phases:

Style packs include: system instructions, banned clichés, ticker naming rules, and a fixed visual art direction string (FORGE industrial / charcoal / forge-orange accents — original art, not third-party IP).

Phase A — structured identity

Model: XAI_MODEL_DEFAULT (e.g. grok-4.7).

Server prompt layers (immutable order):

1. FORGE identity-generation system harness (safety + JSON schema only)
2. Template style pack
3. Creator seed (untrusted user text, delimited)
4. Constraints: ticker length/charset, no financial guarantees, no impersonation of real living people / brands / copyrighted characters

Required structured fields (Zod-validated):

{
  name: string;                 // display name
  ticker: string;               // 3–8 alnum, no $
  shortDescription: string;     // <= 160 chars
  personality: string;          // backstory + voice, markdown-lite OK
  tags: string[];               // 2–5 public capability tags
  starterGreeting: string;      // first chat line suggestion
  avatarPrompt: string;         // image model prompt, 1:1 portrait
  capabilitySuggestions: string[]; // subset of template-allowed caps
}

Creator prompts cannot expand tool permissions beyond the template allowlist; capabilitySuggestions is filtered server-side.

Phase B — profile image

Preferred path for control: POST {XAI_BASE_URL}/v1/images/generations with:

XAI_IMAGE_MODEL=grok-imagine-image-2.0
XAI_IMAGE_ASPECT_RATIO=1:1
XAI_IMAGE_RESOLUTION=1k
XAI_IMAGE_QUALITY=auto
IDENTITY_AVATAR_VARIANTS=3

Prompt assembly:

{template.visualDirection}
{identity.avatarPrompt}
Square 1:1 character portrait suitable as a crypto agent profile image.
No logos of other brands. No readable wallet addresses or QR codes.
No real-person likeness requests. High clarity face/character focal point.

Generate IDENTITY_AVATAR_VARIANTS (default 3) candidates when budget allows; MVP may generate 1 and offer regenerate.

Alternate path: Responses API with tools: [{ type: "image_generation", action: "generate" }] after Phase A in the same conversation — acceptable if Imagine direct params are unavailable.

Persist selected image as avatar_url on the Agent draft after MIME sniff + size cap (e.g. 2 MB) + optional EXIF strip.

API surface

• POST /api/v1/agents/identity/generate — wallet session required; body { template, seed?, agentId? }; returns job id or inline result for MVP.
• POST /api/v1/agents/:id/identity/generate — same, bound to draft agent.
• POST /api/v1/agents/:id/identity/regenerate-avatar — rerun Phase B only with current draft fields.
• Existing POST /agents/:id/avatar remains for manual upload.

SSE optional progress events: identity.start → identity.text → identity.image.start → identity.image.candidate → identity.completed / error.

Rate limits & cost

• Per wallet + IP daily caps (IDENTITY_GEN_DAILY_LIMIT, IDENTITY_AVATAR_DAILY_LIMIT).
• Count against agent/compute usage (agent_usage_daily + generation-specific counters).
• Feature flag identity_generation_enabled; admin can disable without blocking chat.
• Soft timeout; on xAI failure return graceful error and leave form untouched.

Safety

• No celebrity / living-person likeness mode.
• Reject seeds that request CSAM, extreme gore, or clear IP impersonation (keyword + model refusal).
• Generated personality is still untrusted creator config when assembled into live Agent prompts.
• Never put infrastructure secrets, wallet keys, or admin prompts into generation context.
• Log model, token counts, image model, latency; do not log full seed if flagged sensitive — redact where practical.

UI placement

Launch wizard Step 1:

• Primary control: Generate with Grok (orange) next to template selector.
• Secondary: Regenerate avatar.
• Toggle remains: Generate with AI | Upload image (from UI references).
• Right-side preview updates as fields stream in.
• Disclosure: “Generated with Grok. You can edit everything before launch.”

Acceptance

• Trading template generate fills all identity fields + at least one avatar candidate.
• Switching template and regenerating changes tone and visual style.
• Manual edits after generate persist.
• Upload path still works without calling xAI.
• API key never appears in network responses or client bundles.
• Rate limit returns 429 with retry hint.