Skip to content

protocol-reference.md

Latest commit

 

History

History
282 lines (243 loc) · 11.2 KB

protocol-reference.md

File metadata and controls

282 lines (243 loc) · 11.2 KB
Raw

ChatGPT Web Protocol Reference — Autonomous Discovery

Autonomously captured: 2026-06-02 via CDP Runtime.evaluate + Bearer token auth Account: ChatGPT Plus (user@example.com) Success: 21/28 endpoints

Authentication

Access Token

GET /api/auth/session  →  {accessToken: "eyJ...", user: {...}, expires: "..."}
  • JWT token, ~1983 chars
  • All /backend-api/ calls need: Authorization: Bearer <token>
  • Token has expiry; refresh via same endpoint
  • credentials: 'include' needed for cookies (Cloudflare, session)

403 "Unusual activity" Block

  • POST endpoints (/f/conversation) return 403 when called from Runtime.evaluate
  • Requires valid sentinel token (Turnstile + PoW + so challenge)
  • Must go through the full sentinel flow first: prepare → solve PoW → solve Turnstile → finalize
  • This is the main blocker for autonomous message sending

Model Catalog

Slug Title Max Tokens Reasoning Tools
gpt-5-5 GPT-5.5 34,834 auto tools, tools2, dalle_3, search, canvas
gpt-5-3 GPT-5.3 34,834 auto tools, tools2, dalle_3, search, canvas
gpt-5-2 GPT-5.2 25,384 auto tools, tools2, dalle_3, search, canvas
gpt-5-1 GPT-5.1 17,384 auto tools, tools2, dalle_3, search, canvas
gpt-5 GPT-5 16,384 auto tools, tools2, dalle_3, search, canvas
gpt-5-3-mini GPT-5.3 Mini 34,834 none tools, tools2, dalle_3, search, canvas
gpt-5-mini GPT-5-mini 8,191 none tools, tools2, dalle_3, search, canvas
auto Auto 16,384 auto tools, tools2, dalle_3, search, canvas

Default: auto, Latest version: 5.5

Endpoints (Captured + Working)

Auth & User

Endpoint Status Response Size Notes
GET /api/auth/session 200 ~2KB Access token + user info
GET /backend-api/me 200 ~1KB User profile (id, name, email, country)
GET /backend-api/accounts/check/v4-2023-04-27 200 ~8KB Account status, plan_type=plus
GET /backend-api/settings/user 200 ~6KB User settings, announcements, preferences
GET /backend-api/user_segments 200 ~35B Feature flags
GET /backend-api/user_system_messages 200 — System messages
GET /backend-api/user_granular_consent 200 — Consent state

Models

Endpoint Status Response Size Notes
GET /backend-api/models 200 ~46KB Full model catalog
GET /backend-api/models/gpts 200 ~336B GPT-specific models

Conversations

Endpoint Status Response Size Notes
GET /backend-api/conversations?offset=0&limit=28&order=updated 200 ~4KB Conversation list
GET /backend-api/conversation/{id} 200 ~9KB Full conversation with message mapping
POST /backend-api/conversation/init 200 ~654B Returns limits, default model
GET /backend-api/conversation/{id}/stream_status 200 — Stream status
GET /backend-api/pins 200 ~2B Pinned conversations
GET /backend-api/calpico/chatgpt/rooms/summary 200 ~57B Rooms summary

Projects (Gizmos)

Endpoint Status Response Size Notes
GET /backend-api/gizmos/snorlax/sidebar 200 ~162KB All projects + GPTs
GET /backend-api/gizmos/{id} 200 varies Full project detail

Project Structure (gizmo detail)

{
  "gizmo": {
    "id": "g-p-6a1e3f6804588191902b398f5afcd6a7",
    "short_url": "g-p-6a1e3f6804588191902b398f5afcd6a7-c-project-shared",
    "gizmo_type": "snorlax",
    "display": {"name": "C-project-shared", "description": ""},
    "memory_scope": "global",         // "global" = shared, "project_v2" = dedicated
    "memory_enabled": true,
    "instructions": "",
    "context_stuffing_budget": 49152,
    "voice": {"id": "ember"},
    "tools": [],
    "files": [],
    "current_user_permission": {"can_delete": true, "can_write": true, ...}
  }
}

Memory scopes captured:

  • global — shared memory (uses global ChatGPT memory)
  • project_v2 — dedicated memory (project-specific)
  • global_enabled — on conversation level (inherited from project)

Images & Tasks

Endpoint Status Response Size Notes
GET /backend-api/images/bootstrap 200 ~288B Image count, thumbnail
GET /backend-api/tasks 200 ~62KB Background tasks (image gen, research)

Task Structure

{
  "task_id": "imagegen_6a1e7295...",
  "title": "Close-up of a black insect on gravel",
  "status": "completed",
  "conversation_id": "...",
  "created_at": "2026-06-02T06:05:09+00:00"
}

Sentinel Flow (Live-Captured + Solved)

Step 1: Prepare

POST /backend-api/sentinel/chat-requirements/prepare
Headers: Authorization: Bearer <token>
Body: {}
Response: {
  persona: "chatgpt-paid",  // "chatgpt-noauth" when unauthenticated
  prepare_token: "gAAAAABq...",
  turnstile: {required: true, dx: "<29KB encrypted blob>"},
  proofofwork: {required: true, seed: "0.559...", difficulty: "0689f6"},
  so: {required: true, collector_dx: "<17KB>", snapshot_dx: "..."}
}

Step 2: Solve PoW (SOLVED!)

// Algorithm: SHA-256(seed + counter) where first 3 bytes < parseInt(difficulty, 16)
const target = parseInt(difficulty, 16);
for (let i = 0; i < 10000000; i++) {
    const hash = await crypto.subtle.digest('SHA-256', new TextEncoder().encode(seed + i));
    const bytes = new Uint8Array(hash);
    const first3 = (bytes[0] << 16) | (bytes[1] << 8) | bytes[2];
    if (first3 < target) return i.toString();  // Usually solves in <100 iterations
}

Step 3: Solve Turnstile (Cloudflare CAPTCHA)

  • Required, 29KB encrypted dx blob
  • Must be solved in-browser via Cloudflare's Turnstile widget
  • Super Browser v2.0 has auto-solve capability for invisible Turnstile

Step 4: Solve so challenge

  • Required, encrypted collector_dx (17KB) + snapshot_dx
  • Unknown algorithm — likely Kasada-based
  • Super Browser v2.0 has detection but not solving (deferred to v2.1)

Step 5: Finalize

POST /backend-api/sentinel/chat-requirements/finalize
Body: {
  prepare_token: "<from step 1>",
  turnstile: "<solved token>",
  proofofwork: {seed, difficulty, answer: "<from step 2>"},
  so: {collector_dx: "<solved>", snapshot_dx: "<solved>"}
}

Full Sentinel Flow Captured from Real Session

The ChatGPT frontend automatically solves all 3 challenges when sending a message. The conversation/prepare endpoint returns a conduit_token (JWT) used for the conversation.

PoW Solver — VERIFIED WORKING

Seed: 0.559779845730002, Difficulty: 0689f6 → Answer: 30 (solved in 31 iterations)

The PoW is trivially solvable — typically requires <100 hash iterations.

Connectors

Endpoint Status Notes
GET /backend-api/aip/connectors/list_accessible 404 May need different params
GET /backend-api/system_hints?mode=basic 200 System capabilities (search)
GET /backend-api/system_hints?mode=connectors 200 Deep Research, GitHub, OpenAI Platform
GET /backend-api/apps/sources_dropdown 200 ~291KB! Full GPT marketplace catalog

Chat (BLOCKED by sentinel)

Endpoint Status Notes
POST /backend-api/f/conversation 403 Needs valid sentinel token
POST /backend-api/f/conversation/prepare ? Predictive pre-flight

Conversation Detail Structure

Captured from real conversation: "Model Inquiry" (gpt-5-5-thinking)

Message Types in Response

Content Type Role Description
text user User's message text
text system System prompts (hidden, empty parts)
user_editable_context user Editable context (usually empty)
model_editable_context assistant Model's editable context output
thoughts assistant Thinking model's chain-of-thought
reasoning_recap assistant Summary of reasoning process
text assistant Final text response

Thinking Model Message Sequence

1. [system] text: ""              # System prompt
2. [system] text: ""              # Additional context
3. [system] text: ""              # More context
4. [user]   user_editable_context  # User context
5. [system] text: ""              # Pre-injection
6. [system] text: ""              # Post-injection
7. [user]   text: "Hello..."       # USER MESSAGE
8. [system] text: ""              # Model=thinking
9. [assistant] model_editable_context  # Context injection
10. [system] text: ""             # Separator
11. [assistant] thoughts: []       # THINKING (empty = not captured)
12. [assistant] reasoning_recap: [] # REASONING SUMMARY
13. [assistant] text: "GPT-5.5 Thinking." # FINAL ANSWER

Full Conversation JSON

{
  "title": "Model Inquiry",
  "conversation_id": "6a1ed734-...",
  "gizmo_id": null,
  "default_model_slug": "gpt-5-5-thinking",
  "memory_scope": "global_enabled",
  "current_node": "<uuid>",
  "mapping": {
    "<uuid>": {
      "id": "<uuid>",
      "message": {
        "id": "<uuid>",
        "author": {"role": "user|assistant|system"},
        "content": {
          "content_type": "text|thoughts|reasoning_recap|model_editable_context",
          "parts": ["..."]
        },
        "status": "finished_successfully",
        "metadata": {"model_slug": "gpt-5-5-thinking"}
      },
      "parent": "<parent_uuid>",
      "children": ["<child_uuid>"]
    }
  }
}

User Info Captured

{
  "id": "ua-<opaque-user-id>",
  "name": "User",
  "email": "user@example.com",
  "country": "<country_code>",
  "region": "<region_name>",
  "plan_type": "<free|plus|team|enterprise>",
  "subscription_plan": "<subscription_plan_id>"
}

Key Findings

  1. Auth: JWT Bearer token from /api/auth/session (~1983 chars) — required for all /backend-api/ calls
  2. PoW SOLVED: SHA-256 hash comparison — seed + counter, first 3 bytes < difficulty. Trivially solvable (<100 iterations)
  3. Sentinel gatekeeper: 3 challenges (Turnstile + PoW + so) — PoW solved, Turnstile/so still need browser-based solving
  4. Projects ARE gizmos: gizmo_type: "snorlax", memory_scope differentiates shared (global) vs dedicated (project_v2)
  5. Message mapping is a tree: each node has parent + children, supporting branching conversations
  6. Thinking model content types: thoughts, reasoning_recap, model_editable_context, text
  7. Model slug with thinking: gpt-5-5-thinking (derived from base model gpt-5-5 + thinking mode)
  8. Message sending works: Via JS click with full MouseEvent sequence on send button
  9. BrowserFetch pattern: Dedicated about:blank scratch frame inherits browser auth (persona=chatgpt-paid)
  10. Conversation/prepare: Returns a conduit_token (JWT) for the session

Unblocked Path Forward

The main remaining blocker is the Turnstile + so challenges. Options:

  1. Intercept from real session: The page solves these automatically — intercept the finalize request
  2. Super Browser v2.0 Turnstile solver: Has auto-solve for invisible Turnstile
  3. Skip sentinel: Use the page's own message sending (type + click) which handles sentinel internally