fix(google-vertex): nonstandard SSE keepalive frames crash openai-compatible partner-model streams (xai/grok-4.6)

[benceferdinandy-signifyd]

Description

Streams from Vertex AI partner models routed through the OpenAI-compatible
endpoint (e.g. xai/grok-4.6 via @ai-sdk/openai-compatible) abort after
15–30s of reasoning with:

AI_JSONParseError: JSON parsing failed: Text: : keepalive.
Error message: JSON Parse error: Unexpected token ':'

Root cause: on a long time-to-first-token the Vertex partner-model proxy emits
keepalives every ~13–15s as SSE data events — data: : keepalive\n\n —
instead of spec-legal comment lines : keepalive\n\n. parseJsonEventStream
drops the latter correctly, but treats the former as a data event and hands
: keepalive to JSON.parse(), killing the stream.

Both google-vertex fetch wrappers pass the body through untouched, so both
paths are affected:

• packages/opencode/src/provider/provider.ts:542-552 (v1)
• packages/core/src/plugin/provider/google-vertex.ts → authFetch() (v2)

Suggested fix: when content-type is text/event-stream, pipe the body
through a TransformStream that drops lines matching /^data:\s*:/ before the
SDK's SSE decoder sees them (dropping content-length/content-encoding on
the rebuilt Response). I have a working implementation with unit tests and
can open a PR if that's wanted.

Related but distinct: #43519 / #43618 (chunkTimeout vs well-formed comment
heartbeats — that stream parses fine and hangs, this one crashes the parser);
#13914 (Vertex vertex_event frame on the Anthropic path).

Plugins

Local workaround only: a vertex-keepalive.ts plugin wrapping globalThis.fetch
to strip these frames.

OpenCode version

1.18.30

Steps to reproduce

1. Configure google-vertex with an @ai-sdk/openai-compatible partner model,
   e.g. xai/grok-4.6.
2. Send a prompt large enough that the model reasons for >15s before the first token.
3. The proxy emits data: : keepalive and the stream dies with the error above.

Operating System

macOS 26.7

Terminal

Ghostty (tmux)