OpenCode Go: gpt-5.6-luna stream degradation (repeats, mid-stream cuts, junk tails); Codex path clean

Description

gpt-5.6-luna on OpenCode Go is reachable and returns completions, but the stream/output quality is degraded compared to the same model via OpenAI Codex.

Observed symptoms on OpenCode Go (opencode-go / gpt-5.6-luna):

1. Repeated final answers — exact phrase echoed multiple times in one completion
2. Mid-stream cut / continue loops — client sees network-cut continuation behavior (previous response was cut off by a network error mid-stream) even on short prompts
3. Garbage / wrong identity tails — after a finished answer, extra junk text appears (e.g. model claiming to be a different model, trailing “Fine.” / completion noise)
4. Control/count tasks get polluted — e.g. “count 1..5” returns the numbers plus an unnecessary trailing remark

Same model ID via OpenAI Codex (openai-codex / gpt-5.6-luna) on the same machine/day is clean: single exact answer, no repetition, no junk tail.

This looks like a Go relay / Responses streaming issue for Luna, not a general Luna model failure.

Related (not identical):

• #39831 — Zen/Go Luna/Terra hard-fail with Upstream request failed / Error from provider (Console Go)
• Our case is softer: requests often succeed, but stream integrity / final text quality is bad

Docs reference

Per https://opencode.ai/docs/go , GPT 5.6 Luna is served on:

• Model ID: gpt-5.6-luna
• Endpoint: https://opencode.ai/zen/go/v1/responses
• Package: @ai-sdk/openai

Plugins

N/A (reproduced via OpenAI-compatible client / Hermes Agent using OpenCode Go API key against the public Go endpoint). Not using OpenCode TUI plugins.

OpenCode version

N/A for TUI — reproduced against the public Go API:

• Base URL: https://opencode.ai/zen/go/v1
• Provider label: OpenCode Go
• Model: gpt-5.6-luna

Client side: Hermes Agent using the OpenCode Go key / Responses path for Luna.

Steps to reproduce

1. Authenticate with a valid OpenCode Go API key.
2. Call Luna on Go (Responses path as documented):
   • POST https://opencode.ai/zen/go/v1/responses
   • model: gpt-5.6-luna
3. Minimal prompts that reproduced degradation today (2026-07-31, Europe/Rome):

Reply with exactly: LUNA_OK_OPENCODE

Observed: answer repeated multiple times (LUNA_OK_OPENCODE × N), plus mid-stream continue behavior in the client session log.

In Italian, one line only: say who you are and your model id if known. No repetitions.

Observed on Go: wrong/noisy identity text (including claiming another model) and trailing junk.

Count from 1 to 5, one number per line, nothing else.

Observed on Go:

1
2
3
4
5(The count was already complete.)

4. Control comparison on the same day, same machine, same model ID via OpenAI Codex:
   • Prompt: Reply with exactly: LUNA_OK_CODEX
   • Result: clean single-line LUNA_OK_CODEX
   • Natural one-line identity prompt: clean, correct, no junk

Expected behavior

• Single clean final answer
• No repeated completion text
• No mid-stream cut/continue loops on short prompts
• No trailing garbage after a finished response
• Behavior parity with Codex for the same model ID on simple deterministic prompts

Actual behavior

On OpenCode Go, Luna often returns usable tokens but with degraded stream/finalization quality (repetitions, cut/continue, junk tails, wrong identity noise). Codex path remains clean.

Screenshot and/or share link

No TUI screenshot. Session evidence from 2026-07-31 CEST:

• Go smoke: repeated LUNA_OK_OPENCODE
• Go natural prompt: polluted/wrong identity + trailing noise
• Go count prompt: numbers + trailing remark after completion
• Codex control: clean exact responses

Happy to provide redacted request/response traces (no API keys) if useful.

Operating System

Linux (server-side client against public Go API), Europe/Rome

Terminal

Headless API client (Hermes Agent CLI), not OpenCode TUI

Additional notes

• UI label (2x) on GPT 5.6 Luna in some clients appears to be the temporary 2× usage limits promo, not a separate model SKU. This report is about quality/stream integrity of gpt-5.6-luna on Go, not quota labeling.
• Antigravity/OpenCode TUI usage of Luna may appear “fine” depending on transport/path; the degradation was consistently reproducible through the Go API Responses path from an external agent client.
• Please advise if Go Luna currently requires a specific Responses contract (stream flags, previous_response_id, WS vs SSE, etc.) that third-party clients must mirror.