Bug: 5.28 transport refactor regressed prompt caching for both Anthropic and OpenAI-compatible providers

Bug type

Regression (worked in 5.18, broken in 5.28)

Summary

The 5.28 transport layer refactor (src/llm/providers/) lost two critical cache optimizations that were present in 5.18, causing prompt cache hit rate to drop significantly for both Anthropic-compatible and OpenAI-compatible (Completions) providers.

Evidence

• 5.18: Anthropic-compatible provider ~95% cache hit rate; OpenAI-compatible provider ~95% cache hit rate
• 5.28: Both dropped to ~75-80% under similar usage patterns (interactive sessions with tool calls)

Root Causes

1. Anthropic: CACHE_BOUNDARY splitting lost

5.18 (anthropic-transport-stream.ts) called applyAnthropicPayloadPolicyToParams which used splitSystemPromptCacheBoundary to split the system prompt into two content blocks:

• Stable prefix block with cache_control: { type: "ephemeral" } → cached by Anthropic
• Dynamic suffix block without cache_control → not cached, changes don't invalidate cache

5.28 (anthropic-BND0seXX.js, buildParams at ~L588) no longer calls applyAnthropicPayloadPolicyToParams. The entire system prompt is placed in a single content block with one cache_control marker. When dynamic content after <!-- OPENCLAW_CACHE_BOUNDARY --> changes between turns (e.g., Runtime section with activeProcessSessions, group chat context, messaging hints), Anthropic detects the entire cached prefix as changed and invalidates all cached tokens.

2. OpenAI Completions: prompt_cache_key condition narrowed

5.18 condition (openai-transport-stream.ts):

compat.supportsPromptCacheKey && cacheRetention !== "none" && options?.sessionId

5.28 condition (openai-completions-ghIO5MDN.js buildParams ~L346):

model.baseUrl.includes("api.openai.com") && cacheRetention !== "none" 
  || cacheRetention === "long" && compat.supportsLongCacheRetention

Non-OpenAI providers using openai-completions API (e.g., DeepSeek, OpenRouter) have baseUrl that does NOT contain "api.openai.com", so prompt_cache_key is never sent. These providers rely on prompt_cache_key for prefix-based prompt caching. Without it, every request with a dynamic system prompt suffix misses cache.

3. stripSystemPromptCacheBoundary not called

5.28's OpenAI completions transport (openai-completions-ghIO5MDN.js) does not call stripSystemPromptCacheBoundary on the system prompt before sending. The <!-- OPENCLAW_CACHE_BOUNDARY --> marker text is sent to the API as part of the system message, which is a behavior change from 5.18.

Steps to reproduce

1. Configure any Anthropic-compatible provider (MiMo, Claude via proxy, etc.)
2. Start a session and make multiple turns with tool calls (exec, web_search, etc.)
3. Observe that prompt cache hit rate is significantly lower than on 5.18

Same for any non-OpenAI OpenAI-compatible provider using completions API.

Expected behavior

CACHE_BOUNDARY should be respected by the transport layer:

• Anthropic: system prompt split into cached stable prefix + uncached dynamic suffix
• OpenAI Completions: prompt_cache_key forwarded based on compat.supportsPromptCacheKey, not hardcoded URL check

Suggested fix

Fix 1: In anthropic-BND0seXX.js buildParams, apply splitSystemPromptCacheBoundary to system prompt, creating two content blocks (or restore applyAnthropicPayloadPolicyToParams call).

Fix 2: In openai-completions-ghIO5MDN.js buildParams, change prompt_cache_key gate from model.baseUrl.includes("api.openai.com") back to compat.supportsPromptCacheKey.

Fix 3: Call stripSystemPromptCacheBoundary on system prompt in OpenAI completions transport.

OpenClaw version

2026.5.28 (commit e932160)

Operating system

macOS

Install method

npm global