Skip to content

knowledge-gaps.md

Latest commit

 

History

History
106 lines (87 loc) · 4.22 KB

knowledge-gaps.md

File metadata and controls

106 lines (87 loc) · 4.22 KB
Raw

Knowledge Gaps — What We Still Need

What We Know (Live-Captured)

  • ✅ Message endpoint: POST /backend-api/f/conversation
  • ✅ Project scoping: conversation_mode.kind = "gizmo_interaction" + gizmo_id = "g-p-<hex>"
  • ✅ Message format: messages[].content.parts[]
  • ✅ Sentinel flow: prepare → finalize → req
  • ✅ Required headers (Authorization, oai-device-id, etc.)
  • ✅ Init endpoint: POST /backend-api/conversation/init
  • ✅ Prepare endpoint: POST /backend-api/f/conversation/prepare

What We DON'T Know Yet

1. Available Models (Critical for Implementation)

We only captured gpt-5-5-thinking. We need the full list of web model slugs. ChatGPT web likely uses different slugs than the API. We need to either:

  • Capture a model switch in the UI (switch model dropdown, observe network)
  • Scrape the model list from ChatGPT's frontend JS
  • Check the /backend-api/models endpoint (if it exists)

Known from previous reverse-engineering (may be outdated):

  • auto, gpt-4o, gpt-4o-mini, o1, o1-mini, o3-mini, o3-mini-high
  • Now also: gpt-5-5-thinking, possibly gpt-4.1, o4-mini

2. SSE Response Format (Critical for Parsing)

We captured request payloads but the SSE response body was truncated at 10K. We need to see:

  • The actual SSE line format (is it data: {json} or something else?)
  • How parts update incrementally (full text each time or deltas?)
  • The "done" signal format
  • How thinking/reasoning tokens are encoded (for gpt-5-5-thinking)

3. Non-Project Conversation Format

We only tested project-scoped. Need to verify:

  • What conversation_mode looks like for a normal chat: {"kind": "primary_assistant"}
  • Whether the request body structure is identical
  • Whether a brand new conversation (no conversation_id) works the same

4. Turnstile Solving (Blocking)

Turnstile IS required — the capture showed a large encrypted challenge. We need to figure out:

  • Can we solve it via the browser automatically?
  • Or do we need a CAPTCHA solver service?
  • Or can we bypass it by reusing the Turnstile token from the browser session?

5. PoW Algorithm Verification

We built an FNV-1a solver but haven't verified it works against the actual PoW. The captured seed was a float (0.6353036061864151), difficulty was hex (061a80). The actual algorithm is in sentinel/sdk.js — we should extract and verify it.

6. Brand New Conversation Flow

All our captures were continuing existing conversations (had conversation_id). We need to capture the flow for starting a brand new conversation in a project:

  • Is conversation_id: null sufficient?
  • Does the init step create the conversation ID, or does /f/conversation do it?
  • How is parent_message_id set for the first message?

7. AccessToken Lifecycle

We captured the Bearer token in headers but didn't capture:

  • The /api/auth/session response (token structure, expiry)
  • Token refresh behavior
  • What happens when token expires mid-conversation

8. Error Responses

We have no data on what happens when:

  • PoW solution is wrong
  • Turnstile fails
  • Token is expired
  • Rate limit hit
  • Model is unavailable

9. File Upload / Source Attachment

Projects support file uploads. We need:

  • The file upload endpoint flow
  • How files are referenced in messages
  • The selected_sources metadata format

10. client_version / build_number Stability

The captured values (prod-36401cb188ce4e77c4aeaf3e74996e3602a1410d, 7079867) change with every ChatGPT deployment. We need a strategy to handle this:

  • Hard-code and update manually?
  • Scrape from the page?
  • Does ChatGPT reject stale versions?

Recommended Additional Captures

Capture A: Model Discovery (5 min)

  1. Open ChatGPT in debug browser
  2. Run discover2.py
  3. Click the model dropdown, switch between all models
  4. Capture what model slug is sent

Capture B: SSE Response (5 min)

  1. Send a message while capturing
  2. This time save the FULL response body (not truncated)
  3. Parse the SSE format

Capture C: New Conversation (3 min)

  1. Start a brand new conversation in the project
  2. Capture the first message flow (no existing conversation_id)

Capture D: Normal Chat (3 min)

  1. Switch to non-project chat
  2. Send a message
  3. Verify conversation_mode is primary_assistant