model: glm-5.3-flash (SAIA) returns degenerate repetitive output on very large session context #62900

Summary

After continuing very large imported sessions, the model saia/glm-5.3-flash (openai-compatible provider) returns degenerate responses: incoherent German/English mixes in which the same sentence repeats dozens of times, truncated mid-sentence — delivered with finish: stop, so the model treats the garbage as a complete answer. The model's reasoning also misreads the user's request (it paraphrases a different question than the one asked). Reproduced in two separate sessions; in one of them the model itself notices ("Something is wrong with my output — I'm generating garbage").

Environment

• opencode version: 2.2.6
• OS: Windows_NT 10.0.26200 (win32 x64)
• Terminal: xterm-256color
• Shell: cmd.exe
• Install/channel: desktop app (ai.opencode.desktop), latest
• Active plugins: shared-chats-sync (local plugin at C:/Users/Public/opencode-data/shared-chats/plugin) — verified unrelated, sync logs clean
• Model: saia/glm-5.3-flash, variant medium (openai-compatible, baseURL https://chat-ai.academiccloud.de/v1)

Reproduction

1. Import a large session via opencode session import (~400 messages; the continued conversation reaches ~294k input tokens, ~20.8M cache-read tokens over the session).
2. Continue the conversation in a project and ask for a summary ("fasse zusammen, das ich einen neuen chat starten kann").
3. Observe the assistant reply.

Expected Behavior

A coherent (if abbreviated) summary of the session.

Actual Behavior

~2,722 output tokens of degenerate pseudo-reasoning — the same phrase repeated dozens of times, mixed German/English fragments, cut off mid-sentence ("... Let me check the stat"). Delivered with finish: stop after ~80 s. No tool calls were made. The reasoning block shows the model paraphrasing a request the user did not make.

Additional Context

• Affected session 1: ses_e43db6c1effdKqmP0dNCKq15Qu (~400 messages). Degenerate reply at 293,900 input tokens: ~2,722 output tokens of repetitive garbage, finish: stop, no tool calls. An earlier turn at 293,818 input tokens also produced a short garbled mixed-language text.
• Affected session 2: ses_e4fc6e594ffe9irbekHf0JqNuj (215M cumulative input tokens). Degeneration from ~595k input tokens onward: confused reasoning, three replies with finish: length and 0 output tokens, and one reply that burned the full output budget (in=704,475, out=32,711, finish: length) on the repetition loop. The model explicitly noted: "Something is wrong with my output — I'm generating garbage. Let me stop and think carefully." and "my previous responses have been degrading into nonsense".
• Earlier turns in the same sessions at ~130k input tokens produced normal output; degradation appeared at ~294k+ input tokens.
• Looks like a model-side degeneration loop on very large context rather than an OpenCode bug — reporting for visibility. If this belongs with the model provider (GWDG SAIA) instead, feel free to close/move.
• Possible OpenCode-side mitigation to consider: a warning (or automatic compaction/summarization) when a session's input context grows beyond a threshold where smaller models are known to degrade. (Workaround used locally: capping the model's limit.context in the provider config to 200k so auto-compaction triggers before the degeneration zone.)