Ooops amateur hour.

[Performance]: OpenRouter Anthropic path missing `cache_control` on system prompt — 80% unnecessary input cost on every turn

Summary

The maybeAddOpenRouterAnthropicCacheControl() function places a cache_control breakpoint on the last user/assistant message but not on the system/developer message. This means the ~18k-token system prompt (tool definitions, context files, skills, safety guardrails) pays full input price on every turn, even though it's identical across turns.

The native Anthropic path already caches the system prompt correctly. The OpenRouter path just needs parity — one additional cache_control: { type: "ephemeral" } on the last content part of the system message.

Impact

Measured with Langfuse tracing on anthropic/claude-opus-4.6 via OpenRouter:

Before fix: $0.115/turn input cost (system prompt uncached)
After fix: $0.012/turn input cost on turns 2+ (system prompt cached)
80% reduction in input costs, ~$0.095 saved per turn
For a 50-message/day user: ~$150/month savings

This affects every OpenClaw user on the OpenRouter + Anthropic path (likely the majority of users given OpenRouter is the recommended setup).

Root cause

In maybeAddOpenRouterAnthropicCacheControl:

The function finds the last user/assistant message and adds cache_control: { type: "ephemeral" } to its last content part ✅
It does not do the same for the system/developer message ❌
On turn 1, the cache writes for the entire request (system + conversation)
On turn 2, the user message changes → cache key changes → cache miss on the system prompt portion
The system prompt is re-processed at full price every turn

The native Anthropic path adds cache_control to the system prompt independently, so it survives across turns even when conversation messages change.

Proposed fix

Add cache_control: { type: "ephemeral" } to the last content part of the system/developer message in maybeAddOpenRouterAnthropicCacheControl(), same pattern as the existing user message breakpoint. This uses 2 of the 4 available Anthropic cache breakpoints — well within limits.

I have a working local patch with Langfuse traces confirming 97%+ cache hit rates on turns 2+. Happy to open a PR if this aligns with the project direction.

Discussion #1949 — "Burning through tokens" (general cost concerns)
Issue the tokens got burned by dragging a huge context forward #1594 — "The tokens got burned by dragging a huge context forward"
Issue Performance: Workspace file injection wastes 93.5% of token budget #9157 — "Workspace file injection wastes 93.5% of token budget"
Changelog: cacheControlTtl → cacheRetention rename (pi-ai 0.50.9) confirms caching is actively maintained

Environment

OpenClaw: latest stable
Provider: OpenRouter → Anthropic (anthropic/claude-opus-4.6, anthropic/claude-3.5-haiku)
Verification: Langfuse OTEL traces showing gen_ai.usage.input_tokens.cached

Uh oh!

[Bug]: fix(agents): add cache_control breakpoint to system prompt on OpenRouter Anthropic path #15151

Description

Metadata

Metadata

Assignees

Labels

Type

Fields

Projects

Milestone

Relationships

Development

Issue actions