OpenCode Go: glm-5.3-flash returns zero prompt-cache reads while deepseek-v4.1-flash caches

Summary

glm-5.3-flash on OpenCode Go repeatedly returns zero prompt-cache reads through Claude Code + CC Switch, while deepseek-v4.1-flash returns substantial cache-read tokens through the same account and proxy path.

This appears model-specific rather than a general CC Switch monitoring problem.

Environment

• Date observed: 2026-09-11 (UTC+8)
• Service: OpenCode Go
• Route: https://opencode.ai/zen/go
• API format: OpenAI Chat Completions
• Client: Claude Code through CC Switch local proxy
• Model under test: glm-5.3-flash
• Control model: deepseek-v4.1-flash

Observed behavior

For consecutive successful glm-5.3-flash requests, CC Switch showed no cache read or cache creation tokens:

With deepseek-v4.1-flash on the same OpenCode Go account and CC Switch proxy path, repeated requests showed cache-read values around 42,368 to 44,800 tokens.

Expected behavior

After the first request warms the prompt prefix, repeated turns with a stable long context should return non-zero cache-read tokens for glm-5.3-flash, as they do for DeepSeek through the same setup.

Request

Could the OpenCode Go team please check:

1. Whether prompt caching is enabled for glm-5.3-flash in production.
2. Whether GLM requests can be routed across upstream providers without cache affinity.
3. Whether cache-read usage is omitted from GLM responses even when cache is used.
4. Whether the OpenAI Chat Completions route for GLM has a cache-specific limitation.

No API keys, account identifiers, or private logs are included. I can provide redacted screenshots and timestamps if useful.

Start by tracing the OpenAI Chat Completions route for glm-5.3-flash and comparing it with deepseek-v4.1-flash through the same proxy path. Confirm whether prompt caching is disabled, loses provider affinity, or is omitted from responses; done means identifying one cause with reproducible evidence or handing the report to the service team.