ox-alpha-free (Go free tier): long generations severed server-side without finish_reason — strict clients throw "Stream ended without finish_reason" #44385

Description

Model: ox-alpha-free (Go plan free tier, "Ox Alpha Free (Unlimited)")
Endpoint: https://opencode.ai/zen/go/v1/chat/completions (OpenAI-compatible SSE, stream: true)
Client: Pi coding agent (earendil-works/pi v0.84.2) — strict OpenAI-SSE parser
OS: Windows 11

On long generations, the gateway severs the stream server-side without emitting finish_reason and without data: [DONE]. Strict OpenAI-compatible clients therefore throw Stream ended without finish_reason and discard the whole partial response. This is 100% reproducible with a complex prompt, and is the ox-alpha-free instance of the same gateway bug class as #43379 / #43882 (muse-*) and #44148 (mid-stream cutoffs on long-reasoning requests).

Why the complex prompt triggers it

ox-alpha-free burns its entire budget on reasoning. A max_tokens=8000 request produced 3134 reasoning tokens and 0 visible content tokens. Complex prompts (e.g. "write a complete interactive single-file HTML") push the reasoning phase past the gateway's patience window (3–7 min), at which point the stream is killed. Simple prompts finish reasoning quickly and complete normally — which is why the failure only shows up on long tasks.

Reproduction (direct HTTPS probes, same API key, no agent in the loop)

Prompt used:

我需要一个解释洛伦兹变换的可视化科普设施，你最好用二维例子，x-y-t立体网格坐标，
附带交互组件。让我完全理解，你必须保证这个科普设施能让我完全理解为啥(ct)^2 - x^2 - y^2是不变量。
使用单html文件构建

The kill time varies (211s / 403s across runs), so it looks like server-side preemption or a resource cap rather than a fixed client-visible limit.

Client-side error in strict clients

When the stream ends without finish_reason, pi's OpenAI parser throws:

Error: Stream ended without finish_reason

and discards the partial response (per-spec: the final chunk must carry a non-null finish_reason followed by data: [DONE]).

Additional framing issue

In one captured stream, 141 of ~2,378 events contained raw newlines embedded inside JSON strings (the reasoning_content delta spans multiple physical lines), splitting a single SSE event across several data:-free lines. Line-based SSE parsers see these as truncated JSON events.

Expected behavior

When the gateway severs a generation (time cap, preemption, quota), it must still close the OpenAI-compatible stream per spec: a final chunk with finish_reason (e.g. "length") followed by data: [DONE]. The 8k-cap probe shows the gateway can do this — it's specifically the long-run kill path that omits the terminal chunk.

Related issues

• [Go plan] Empty completions (finish_reason=other) and mid-stream cutoffs on long-reasoning requests — reproduced via direct API calls across multiple models #44148 — mid-stream cutoffs on long-reasoning requests (paid models, same signature)
• Streaming responses for muse-* models on the Zen gateway never send finish_reason — strict OpenAI-compatible clients enter a retry loop #43379 / muse-spark-1.2-contributor-free: stream always ends without finish_reason — every request fails in strict OpenAI-compatible clients #43882 — muse-* models never send finish_reason
• bug: Hermes Agent — gpt-5.6-luna via opencode-go provider returns finish_reason:null (no [DONE]) #40420 — gpt-5.6-luna missing finish_reason
• Ox Alpha Free (Unlimited) not working it will start thinking and dont give any output #44332 / Provider finish_reason: network_error with Ox Alpha Free (unlimited) #44379 / ox-alpha-free via zen/go/v1: any tools payload returns network_error/503 since 2026-08-23 (was working 02:55-06:11) #44382 — other ox-alpha-free outage reports