Zen gateway (opencode-go): deepseek-v4.1-flash degenerates into infinite pre-tool-call repetition loops; same model via direct DeepSeek API is clean #51839

Summary

The same model (deepseek-v4.1-flash) behaves differently depending on the provider route: via opencode-go (Zen gateway, https://opencode.ai/zen/go/v1) it falls into infinite autoregressive repetition loops before tool calls; via the direct DeepSeek API (deepseek/deepseek-flash) the exact same sessions run clean. We have a clean A/B (same weights, two serving stacks, stable difference across many runs), session transcripts of the loops, and a plausible root cause.

This looks like the gateway mishandling reasoning models' tool-calling transitions (most likely: feeding reasoning_content back into the conversation, and/or missing sampling safeguards).

Environment

• opencode: v1.18.3x (2026-09-27), macOS (darwin/arm64)
• Provider A (degenerates): opencode-go — https://opencode.ai/zen/go/v1, model deepseek-v4.1-flash, requests carry x-opencode-session header
• Provider B (clean): deepseek — https://api.deepseek.com/v1, model deepseek-flash
• Agent harness: custom subagents with tools (bash/read/edit), temperature 0.2–0.3, up to 100+ tool-call steps per session

Evidence

1. Clean A/B on the same model

• Same session content replayed against both providers:
  • opencode-go/deepseek-v4.1-flash → enters repetition attractor loops, repeatedly
  • deepseek/deepseek-flash (direct) → runs to completion, repeatedly
• The user's manual switch mid-session (gateway → direct) made the 400/loop problems disappear immediately; switching back reproduces them.

2. Loop transcripts (verbatim, from agent=coding and primary orchestrator sessions)

Pattern A — English (the model stalls before emitting a tool call, then loops):

Let me run.
Let me execute.
Let me do it now.
OK.
Let me run.
Let me execute.
…(repeats for 100+ lines)

Pattern B — same class, different vocabulary:

Emit.
OK.
Now.
OK.
Emitting.
OK.
I must emit. Doing it.
Let me grep.
Emit.
OK — emitting.
Recon.
OK.
…(fixed ~12-line cycle, repeats indefinitely)

Russian-localized session, same fixed cycle:

Сгенерируй.
OK — генерирую.
Вызов.
OK.
Сейчас.
OK.
Генерация.
OK.
Я должен выполнить вызов. Занимаюсь этим.
…(repeats)

Notably, the vocabulary of the loops mirrors instruction text from the agent prompts (compliance-narration words like "emit"/"OK"/"call") — consistent with the model regurgitating its own prior assistant text, i.e. the loop is self-feeding through conversation history.

3. What the direct API does differently

• Direct DeepSeek API responses carry reasoning in reasoning_content, and the model was trained with the contract that reasoning_content must not be fed back as input on subsequent turns.
• Hypothesis: the Zen gateway re-includes the assistant's reasoning/hesitation text in subsequent messages (or otherwise alters the reasoning→tool_call transition), so the model keeps seeing — and then predicting — its own deliberation text instead of transitioning to the tool-call token.
• Secondary suspect: the gateway passes through low temperature (0.2–0.3) without vendor-side anti-repetition safeguards; greedy-ish decoding makes the attractor inescapable once seeded.

4. Related gateway-side quirks observed the same day

• stream error … AI_APICallError: Bad Request on gateway requests whose message history the direct API accepts; the gateway's error body is truncated to {"model":"deepseek-v4.1-flash"} — no diagnostics are recoverable.
• After such an error, opencode disposes and recreates the whole instance (feels like a hang).

Expected vs actual

• Expected: the same model behind the Zen gateway behaves like the direct API — reasoning handled per the provider contract (reasoning_content not fed back), tool calls emitted after deliberation, sampling sane for low-temperature agent workloads.
• Actual: gateway route reliably degenerates into infinite pre-tool-call narration loops on agentic workloads; direct route is stable.

Reproduction sketch

1. Configure two providers: opencode-go (Zen, x-opencode-session header) and direct DeepSeek, same model tier (deepseek-v4.1-flash / deepseek-flash).
2. Run an agentic session with ~50+ tool-call steps and a prompt that requires multi-step verification (e.g., "run these tests, then narrow failing ones, then re-run").
3. Gateway route: assistant text degenerates into fixed repetition cycles ("Let me run. / Let me execute. / …" or localized equivalents) instead of emitting the tool call.
4. Direct route: same session completes.

Happy to provide full session JSONL dumps (DB rows with message/part payloads) for both routes if that helps debugging.