[Bug] OpenCode Go upstream 400 [1210] parameter error triggered by Unicode multiplication sign (U+00D7) and trailing quotes in GLM-5.3-Flash #46378

Description

When sending chat completion requests to glm-5.3-flash via the OpenCode Go gateway (opencode.ai/zen/go/v1), the API deterministically fails with an HTTP 400 error ([1210] API parameter invalid) whenever the request messages contain specific standard Unicode characters—most notably the multiplication sign × (U+00D7) and certain trailing quotation mark sequences like ” (U+201D).

This is not a rate-limit, context length, or token limit issue. It can be reproduced consistently with a single, minimal 1-turn text prompt.

Environment & Context

• Model: glm-5.3-flash (via OpenCode Go / Console Go)
• Endpoint: /zen/go/v1/chat/completions
• Error Response:

  {
    "error": {
      "type": "server_error",
      "message": "Error from provider (Console Go): Upstream request failed: [1210] API parameter invalid, please check documentation."
    }
  }

Minimal Reproducible Example (MRE)

1. Failing Request (U+00D7 Multiplication Sign)

curl -X POST https://opencode.ai/zen/go/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <OPENCODE_GO_API_KEY>" \
  -d '{
    "model": "glm-5.3-flash",
    "messages": [
      {"role": "user", "content": "Test: A \u00D7 B"}
    ]
  }'

Result: HTTP 400 / [1210] API parameter invalid

2. Working Baseline (ASCII x or Alternative Unicode ✕)

curl -X POST https://opencode.ai/zen/go/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <OPENCODE_GO_API_KEY>" \
  -d '{
    "model": "glm-5.3-flash",
    "messages": [
      {"role": "user", "content": "Test: A x B"}
    ]
  }'

Result: HTTP 200 OK

Comparative Test Matrix

Across alternating test rounds executed within the same timeframe, the outcome is 100% deterministic:

Real-World Impact on Multi-Turn Agents

In automated multi-turn AI coding agents or chat systems:

• If any historical turn (system prompt, tool output, resolution logs like 1920×1080, or arithmetic expressions like 7×24) contains the character \u00D7, the agent re-submits the entire conversation history on subsequent turns.
• This causes the entire conversation session to permanently fail with [1210], rendering the session unusable unless context compression or manual history sanitization is applied.

Suggested Remediation

1. Gateway Pre-processing / Normalization:
   • Sanitize or normalize incoming messages text on the OpenCode Go API gateway layer before forwarding to the tokenizer (e.g., auto-substitute \u00D7 with standard ASCII x).
2. Upstream Tokenizer Fix:
   • Inspect the GLM-5.3-Flash vocabulary and BPE tokenizer boundaries for U+00D7 and edge-case multi-byte quotation combinations (U+201D) to resolve the parser exception.