opencode-cache-manager: keep OpenCode's prompt cache warm, including through OmniRoute

[cris005]

I run OpenCode through OmniRoute, and every time I stepped away for more than 5 minutes the Claude prompt cache expired. The next prompt then paid 1.25x to rewrite the whole context instead of about 0.1x to read it. I wrote an OpenCode plugin to fix that:

https://github.com/noppu-labs/opencode-cache-manager

You pick a cache TTL per session: 5m, 1h or hybrid (on Claude, 1h on tools and the system prompt, 5m on messages). You can switch it with /cache-manager or set it per model variant. When a session goes idle, the plugin sends a hidden ping just before the cache expires, so it's still warm when you come back. A ping costs roughly one cache read, and it stops after a set number of pings (5 by default). The sidebar shows the cache status, how much of the prompt is cached, when the next ping goes out and what the pings have saved.

A few things specific to OmniRoute:

• It doesn't need any OmniRoute-specific setup. Whatever provider ID you register OmniRoute under, the plugin decides by model: a name containing claude or starting with anthropic/ is managed as Claude, GPT-5.6+ is matched by name, and you can map aliases with cacheProfiles. Other models are never rewritten or pinged; the sidebar just shows how much of their prompt was cached.
• It works on the Anthropic Messages wire, /chat/completions and /responses.
• In 1h and hybrid it adds extended-cache-ttl-2025-04-11 to the anthropic-beta header, merged with any values already there. Some routes need it before the 1h TTL applies.
• GPT-5.6+ through OmniRoute isn't fully working yet. The 12h breakpoint needs prompt_cache_breakpoint to get through the Chat Completions translator. I fixed that in fix(translator): keep OpenAI prompt_cache_breakpoint on content parts for openai/openrouter #18255 (merged 2026-10-27), but no release includes it yet; v3.8.52 is the latest. Until one does, use 5m for GPT sessions through OmniRoute. Claude isn't affected.

Install:

opencode plugin add @noppu-labs/opencode-cache-manager

There's a sample config for OpenAI-compatible proxies in docs/examples/. I'd like to hear from anyone on routes or combos I haven't tried.

[diegosouzapw]

@cris005 Nice work, thanks for sharing it. Your GPT-5.6+ note is correct. #18255 is on the release branch and goes out with 3.8.53, so 5m is the right setting until then.

[diegosouzapw]

@cris005 Update: 3.8.53 came out on 2026-11-06 with #18255 in it. The GPT-5.6+ 12h breakpoint should get through OmniRoute now.