[Bug] OpenCode Go intermittently returns 400 [1210] "thinking cannot be disabled" and MissingSessionID when routing omen-alpha requests

Description

Description

When using omen-alpha via the OpenCode Go gateway (opencode.ai/zen/go/v1/chat/completions) from a third-party client (QwenPaw, an open-source agent framework using the OpenAI-compatible chat completions endpoint), requests intermittently fail with two distinct 400 errors returned by the "Console Go" provider:

Error 1 — [1210] thinking cannot be disabled

Error code: 400 - {'error': {'type': 'server_error', 'message': 'Error from provider (Console Go): Upstream request failed: [1210] This model always engages in thinking and cannot be disabled; please use low, high, or max'}}

Error 2 — MissingSessionID

Error code: 400 - {'type': 'error', 'error': {'type': 'MissingSessionID', 'message': 'Error from provider (Console Go): Request is missing x-opencode-session and cannot be routed efficiently. Please see https://opencode.ai/docs/go/#where-can-i-use-it'}}

Environment & Context

• Model: omen-alpha (Go subscription, chat/completions endpoint)
• Client: QwenPaw (open-source agent framework, uses the standard OpenAI-compatible endpoint; not yet in the Validated Clients list — it does not currently send x-opencode-session)
• Time window: 2026-09-08, 08:46–10:07 (UTC+8); intermittent — most requests succeed, some fail
• Observed pattern: burst failures on consecutive messages, then recovery — consistent with gateway-side routing/fallback decisions

Controlled experiment (same session, model-switched)

In the same session (identical context size, identical request pattern, rapid consecutive messages):

Model | Result -- | -- omen-alpha | 400 [1210] / MissingSessionID reproduced consistently DeepSeek v4 flash | zero failures, rapid consecutive messages all succeed

This isolates the failure to the omen-alpha routing path (not client concurrency, not channel queuing, not oversized context).

Analysis

Error 1: For the failing requests, the client does not send any thinking-related parameters (no reasoning_effort, no enable_thinking). The gateway appears to inject a "disable thinking" parameter when routing the request — but omen-alpha always engages in thinking and rejects it with [1210].

We also tried explicitly sending reasoning_effort: "low" (and "high") from the client via extra_body — the error still occurred, suggesting the gateway may override or drop client-side thinking parameters for these routed requests.

Error 2: QwenPaw does not send x-opencode-session (documented requirement: https://opencode.ai/docs/go/#where-can-i-use-it). Requests lacking the header fail when the gateway needs to make routing decisions.

Impact

In a chat/agent product this surfaces as intermittent "Internal error" replies to end users. Conversations fail in bursts (consecutive messages), then self-recover — making it hard to debug from the client side.

Workaround confirmed: switching the model from omen-alpha to deepseek-v4-flash (same session, same client, same request pattern) made all failures stop immediately — further confirming the issue is specific to the omen-alpha routing path.

Suggested remediation

1. Gateway: preserve or forward client thinking parameters when routing omen-alpha requests, instead of injecting a "disabled" flag (omen-alpha rejects it).
2. Gateway: make x-opencode-session degrade gracefully (log/flag) rather than hard-fail with 400, or provide a documented migration path for third-party clients.
3. Client-side (QwenPaw): we will add x-opencode-session support in an upcoming release; tracking in our repo.

Related

• #46378 — same [1210] error code but different root cause (GLM-5.3-Flash Unicode tokenizer vs. thinking parameter routing)

Additional info (compliance)

• OpenCode version: Go gateway API (opencode.ai/zen/go/v1/chat/completions),
  called via third-party client QwenPaw 2.2.0 (OpenAI-compatible provider)
• Steps to reproduce:
  1. Use omen-alpha via the Go chat/completions endpoint from a third-party
     client (no x-opencode-session header)
  2. Send rapid consecutive requests (burst pattern)
  3. Observe intermittent 400 [1210] "This model always engages in thinking
     and cannot be disabled" and/or MissingSessionID
  4. Switch model to deepseek-v4-flash in the same session → failures stop
     immediately
• Operating System: Linux (client-side server); issue is gateway-side
• Terminal: n/a (HTTP API client)

Start by tracing the OpenCode Go gateway's chat/completions handling for omen-alpha, focusing on the x-opencode-session header and thinking settings. Reproduce requests from QwenPaw and compare successful and failing cases. Done means the intermittent thinking-disabled and MissingSessionID 400 responses are resolved or clearly handled for this client pattern.