Skip to content

LEGAL_OAUTH_WORKAROUNDS.md

Latest commit

 

History

History
121 lines (95 loc) · 6.49 KB

LEGAL_OAUTH_WORKAROUNDS.md

File metadata and controls

121 lines (95 loc) · 6.49 KB
Raw

Legal OAuth Workarounds

This is the practical map after the extra source search and live probes.

Boundary used here:

  • Use only this machine's own Hermes/Codex/ChatGPT OAuth token.
  • Use OpenAI-owned public clients, official docs, and the public openai/codex source as route evidence.
  • Do not bypass another user's data, workspace permissions, billing, rate limits, or admin boundaries.
  • Do not run write/installation/enrollment routes automatically unless the user explicitly asks for that specific side effect.

Best Working Substitutes

Feature wanted OAuth-only route that works now Current evidence
Responses / chat POST /backend-api/codex/responses direct text pass
Vision input POST /backend-api/codex/responses with input_image vision pass in matrix
Image / drawing Codex Responses image_generation tool with gpt-image-2 PNG artifacts generated
TTS / voice output Realtime WebSocket response audio PCM16 artifact generated
WebRTC voice call setup POST /v1/realtime/calls multipart with SDP + session 201, answer SDP returned
STT POST /v1/audio/transcriptions pass
Realtime transcription POST /v1/realtime/client_secrets with session.type="transcription" pass
Embeddings POST /v1/embeddings pass
Files / uploads POST /backend-api/files + signed upload + finalize returns sediment://file_...
Vector search / RAG local vector store + OAuth embeddings compatibility layer smoke passed
Evals local eval runner + Codex text compatibility layer smoke passed
Connectors / tools Apps MCP through POST /backend-api/wham/apps tools/list returned 111 tools
Public GitHub search Apps MCP tools/call github_search_repositories safe public search passed
Connector discovery GET /backend-api/connectors/directory/list* 892 apps in latest run
Plugin discovery/export GET /backend-api/plugins/list, featured, export/curated pass
ChatGPT account/user shape GET /backend-api/me, GET /backend-api/accounts/check/... pass, sensitive values redacted
Browser Use site status GET /backend-api/aura/site_status pass
Codex usage GET /backend-api/wham/usage pass
Codex cloud tasks GET /backend-api/wham/tasks/list pass
Codex environments GET /backend-api/wham/environments pass
Repo environments GET /backend-api/wham/environments/by-repo/... pass
Agent identity JWKS GET /backend-api/wham/agent-identities/jwks pass

The older POST /v1/realtime/transcription_sessions path now returns 400 Unknown beta requested: 'realtime' here. The current working transcription setup is the client_secrets route with a transcription session object, which matches the current OpenAI Realtime transcription docs.

New Finding

/v1/realtime/calls is no longer just "auth accepted but invalid payload". With Codex-style multipart form-data and a realistic WebRTC offer, this account's Codex OAuth token got:

  • HTTP 201
  • Location: /v1/realtime/calls/calls/<call-id>
  • answer SDP body, 40 lines

That makes WebRTC call setup a real OAuth-capable path here. The wrapper exposes it as OpenAIOAuthAccess.realtime_webrtc_call_offer(offer_sdp, ...). For a real browser app, pass the SDP from RTCPeerConnection.createOffer().

Source-Backed But Not Auto-Run

These are legal own-account routes found in the current openai/codex source, but I did not execute them because they mutate real workspace state.

Capability Route Why not auto-run
Create a Codex cloud task POST /backend-api/wham/tasks creates a real task
Remote app-server control POST /backend-api/wham/remote/control/server/enroll; WSS /backend-api/wham/remote/control/server enrolls/controls an environment
Share/update workspace plugin POST /backend-api/public/plugins/workspace/upload-url; POST /backend-api/public/plugins/workspace publishes workspace plugin share
Update/delete workspace plugin share POST /backend-api/public/plugins/workspace/{remote_plugin_id}; PUT /backend-api/ps/plugins/{remote_plugin_id}/shares; DELETE /backend-api/public/plugins/workspace/{remote_plugin_id} changes workspace share state
Install/uninstall plugin POST /backend-api/ps/plugins/{plugin_id}/install; POST /backend-api/plugins/{plugin_id}/uninstall changes installed plugin state
Add-credit nudge email POST /backend-api/wham/accounts/send_add_credits_nudge_email; POST /api/codex/accounts/send_add_credits_nudge_email sends real account email
Codex analytics POST /backend-api/codex/analytics-events/events telemetry write, not app capability

If you want to test any of those, do it as an explicit one-off because they will change the account/workspace.

Still Not Legally Replaceable With This OAuth Token

These are account/project/admin server resources, not just "model capability":

  • real OpenAI Platform fine-tuning jobs
  • OpenAI-hosted /v1/files, /v1/uploads, /v1/vector_stores
  • Admin/Usage organization endpoints
  • hosted ChatKit session creation
  • OpenAI containers/video resources
  • official /v1/responses, /v1/chat/completions, /v1/images/*, /v1/audio/speech with the exact public endpoint
  • official codex-responses-api-proxy as OAuth. It exists, but its own README says it reads OPENAI_API_KEY from stdin and forwards /v1/responses.

For app features, the substitutes above cover most of the behavior. For Platform resources inside an OpenAI API organization, API key/Admin key is still the legitimate boundary.

Local /v1 Compatibility Proxy

For local apps that expect an OpenAI-shaped base URL, this repo now includes:

src/oauth_openai_compat_server.py

It is not an OpenAI-hosted endpoint. It runs on 127.0.0.1 and maps:

  • /v1/responses -> Codex OAuth text
  • /v1/chat/completions -> Codex OAuth text over serialized messages
  • /v1/embeddings -> OAuth /v1/embeddings
  • /v1/images/generations -> Codex OAuth image generation

Smoke test on 2026-05-11: /health, /v1/responses, /v1/chat/completions, and /v1/embeddings all returned 200.

Source Links

  • Codex OAuth authentication: https://developers.openai.com/codex/auth
  • OpenAI API authentication boundary: https://platform.openai.com/docs/api-reference/authentication
  • Realtime WebRTC: https://platform.openai.com/docs/guides/realtime-webrtc
  • MCP and connectors: https://platform.openai.com/docs/guides/tools-remote-mcp
  • Public source searched: https://github.com/openai/codex