[iOS] Two failures with OpenCode Go: built-in provider blocked before the model test, custom provider streams empty (both work via a reverse proxy)

Author: BiberpelzCreated Sep 13, 2026Updated Sep 15, 2026

Correction (2026-09-15): the provider configuration described below is wrong. The desktop side was using the built-in OpenCode Go provider, the iOS side the same endpoint configured as a generic custom provider. A second, separate failure of the built-in provider on iOS is also not covered below. See the correction comment for the full test matrix, the proxy captures and the transport evidence.

Summary

Chatbox Desktop (Electron) sends an x-opencode-session header on requests to OpenAI-compatible custom providers. The iOS (Capacitor) app does not. The OpenCode Go endpoint (https://opencode.ai/zen/go/v1/chat/completions) rejects requests without this header with HTTP 400 MissingSessionID. Because the iOS app swallows the 400 on the SSE stream, the chat appears completely silent — no error message, no response.

Related issues

#3891 (closed as completed, reported fixed in 1.23.2) asked for exactly this header when calling OpenCode Go / OpenCode Zen. The fix is effective on desktop 1.23.2, but the iOS app of the same version still sends no x-opencode-session — the mobile code path appears not to be covered by it. This report is meant as a follow-up for that gap, not as a duplicate.
#3516 (open) reports iOS failures for custom OpenAI-compatible providers and points at src/renderer/utils/request.ts: on mobile, requests are only routed through handleMobileRequest when useProxy is true, and otherwise fall back to a plain fetch() in the web view. That is consistent with the runtime/browser User-Agent captured below and may be the same root cause.

Environment

Desktop: Chatbox 1.23.2, Windows (Electron 35.7.5) — works
Mobile: Chatbox iOS 1.23.2, iPhone, iOS 18.7 (Capacitor web view) — fails silently
Provider config (both apps): OpenAI-compatible custom provider, API host https://opencode.ai/zen/go, model glm-5.3-flash, OpenCode Go API key

Steps to reproduce

Create a custom OpenAI-compatible provider pointing to https://opencode.ai/zen/go with a valid OpenCode Go API key and model glm-5.3-flash.

Desktop: chat works. Header dump of a desktop request shows:

x-opencode-session: 00000000-0000-0000-0000-000000000000   <- redacted; the real value is a per-conversation UUID
User-Agent: Mozilla/5.0 ... xyz.chatboxapp.app/1.23.2 Chrome/134.0.6998.205 Electron/35.7.5 Safari/537.36
-> HTTP 200 OK

iOS: model list loads fine (GET /v1/models → 200; this endpoint does not require the header), but sending a chat message fails. Header dump of an iOS request (captured via a local logging proxy on ):
```
POST /v1/chat/completions
Accept: text/event-stream
Authorization: Bearer sk-D3...[redacted]
User-Agent: ai-sdk/openai-compatible/2.0.62 ai-sdk/provider-utils/4.0.40 runtime/browser
-> HTTP 400 BadRequest
```
No x-opencode-session header is present.

Direct API test against the endpoint (bypassing Chatbox):

# without header:
HTTP/1.1 400 Bad Request
{"type":"error","error":{"type":"MissingSessionID","message":"Error from provider (Console Go): Request is missing x-opencode-session and cannot be routed efficiently. Please see https://opencode.ai/docs/go/#where-can-i-use-it"}}

# same request with -H "x-opencode-session: test-123":
HTTP/1.1 200 OK  (full chat completion response)

Injecting the header on the mobile request via a local proxy makes the iOS chat work immediately (200 OK, model responds).

Actual behavior

iOS: provider setup and model list load successfully, but every chat message silently produces no output (the 400 on the SSE stream is swallowed).
No error is surfaced to the user.

Expected behavior

The iOS/Android apps should send a stable x-opencode-session header (e.g. a UUID generated per conversation, as the desktop app already does) on requests to OpenAI-compatible providers — or at least allow configuring custom HTTP headers per provider so users can satisfy provider requirements like OpenCode Go's.
Additionally: 4xx errors on streamed chat requests should be surfaced as an error message in the chat instead of failing silently.

Likely cause

(Inferred from the captured headers above; not verified against the source.)

The mobile apps use a different network stack (Capacitor + ai-sdk/openai-compatible/2.0.62) than the desktop app (Electron). The x-opencode-session header support that exists in the desktop code path is missing in the mobile path. OpenCode Go documents the requirement here: https://opencode.ai/docs/go/#where-can-i-use-it ("Send a stable session ID in x-opencode-session for each conversation").

Suggested fix

Generate a stable per-conversation UUID and send it as x-opencode-session on all provider chat requests in the mobile apps (parity with desktop), and/or
add a "custom headers" field to custom provider settings on all platforms, and/or
surface HTTP errors from streamed requests in the chat UI instead of failing silently.

Workaround

A local/VPS reverse proxy that injects x-opencode-session makes the iOS app work (verified). This is a workaround only; native support would remove the need.

Source: chatboxai/chatbox

View original on GitHub View discussion on GitHub