Explicit cache breakpoints are applied only to Anthropic-family models; every other family relies on implicit caching

Description

Explicit cache breakpoints only ever reach Anthropic-family models, so every other family runs on implicit prefix caching with no advancing anchor.

message() at packages/opencode/src/provider/transform.ts:471-484 calls applyCaching() behind a family gate (providerID === "anthropic", api.id/model.id containing anthropic or claude, npm @ai-sdk/anthropic or @ai-sdk/alibaba). A model on @ai-sdk/openai matches none of it, so the function never runs. The openrouter/bedrock/copilot keys inside its map are for Claude routed through those gateways, not for those providers' own models.

applyCaching() puts a breakpoint on the last two non-system messages, which moves forward each turn. Nothing else gets that. Once anything mutates the array mid-conversation, reuse freezes at a fixed offset for the rest of the session:

turn    input    cached
 #18    2,750    95,104     healthy
 #19   95,691     3,712     break
 #26   70,278    38,912     prompt grew 9k, cached frozen

Turn 21 ran 12 seconds after turn 20 and reused 38,912 of a ~100k prefix just written, so it isn't TTL. Measured over 10 days, the split follows model family rather than provider — ~32,000 Anthropic-family turns at 100.0% reuse, versus gpt-6-astra 92.1%, gpt-5.6-sol 64.9%, github-copilot/gpt-5.4-mini 80.1%, github-copilot/gemini-3.5-flash 91.3%. Astra alone re-sent ~39M tokens.

Note this is breakpoint placement, not cache keys — promptCacheKey = sessionID is already set at :1321.

@ai-sdk/openai@3.0.88 (pinned at v1.18.30) already supports the field:

function getPromptCacheBreakpoint(providerOptions) {
  return providerOptions?.openai?.promptCacheBreakpoint;
}

It has zero references in this repo. Part-level providerOptions already survive into content blocks via mapProviderOptions() at :447, so the placement path exists on the Responses API that Astra uses.

Suggested fix, in two parts, since they live in different places:

• Breakpoint — set promptCacheBreakpoint on content parts of the same trailing messages applyCaching() targets. That's the message path.
• Mode — prompt_cache_options is serialized model-level from openaiOptions.promptCacheOptions, so mode: "explicit" belongs in transform.options(), the same function already setting promptCacheKey at :1321. A message-path change alone sets breakpoints and leaves mode unset.

I'd avoid widening the existing applyCaching() gate — it also decides which families get cache_control blocks injected, and admitting @ai-sdk/openai risks sending those to openai-compatible endpoints that reject them.

Scope note: this covers the openai providerOptions namespace, which sdkKey() also maps @ai-sdk/amazon-bedrock/mantle onto. It does not reach Copilot — sdkKey("@ai-sdk/github-copilot") returns "copilot", so github-copilot/gpt-5.4-mini would stay at 80.1% and needs its own path. Worth treating as a separate change.

Related: #44727 (same mechanism, scoped to post-/compact), #43507, #39679.

Plugins

@cortexkit/opencode-anthropic-auth@1.22.0, @cortexkit/opencode-openai-auth@0.7.1, oh-my-opencode-slim@2.2.11, @cortexkit/opencode-magic-context@0.41.4, @cortexkit/aft-opencode@0.55.1, opencode-copilot-delegate@0.12.1, @fro.bot/systematic@3.16.5

OpenCode version

1.18.29, confirmed identical at v1.18.30

Steps to reproduce

1. Run a long session on any @ai-sdk/openai model.
2. Let the array be mutated mid-conversation (plugin injection or a compaction).
3. Compare usage.prompt_tokens_details.cached_tokens to total prompt tokens per turn.
4. Cached tokens stop tracking the prompt and pin at a fixed value, persisting across sub-minute gaps.
5. Repeat on an Anthropic model — reuse tracks continuously and recovers after compaction.

Screenshot and/or share link

n/a

Operating System

macOS 15.7.9 (arm64)

Terminal

Ghostty