llm_anthropic: support Anthropic prompt caching

[joshuadarron]

Problem

The llm_anthropic node instantiates ChatAnthropic with no cache configuration:

# nodes/llm_anthropic/anthropic.py:110
self._llm = ChatAnthropic(
    model=model, api_key=apikey, temperature=0, max_tokens=self._modelOutputTokens
)

.rocketride/schema/llm_anthropic.json exposes model / modelTotalTokens / apikey only. There is no way for a pipeline author to opt into Anthropic prompt caching.

In a 28-minute coding-agent run, every claude-opus-4-6 call re-shipped 2–5 KB of stable system prompt plus an accumulating message history. With ~180 LLM calls across the run and large stable prefixes, cache_control: {"type": "ephemeral"} would cut input-token latency on cached blocks by ~85% and cost by ~90% per Anthropic's published numbers. Estimated savings on that run alone: ~470 s (~30%).

Proposed fix

1. Add an optional caching boolean (or finer-grained struct: { system: bool, history: bool }) to services-catalog.json profiles for llm_anthropic.
2. When enabled, attach cache_control: {"type": "ephemeral"} to system-prompt content blocks (and optionally to the most recent stable message tail) before passing to ChatAnthropic. Modern langchain-anthropic accepts this either via message content blocks or model_kwargs.
3. Surface usage.cache_creation_input_tokens / cache_read_input_tokens from responses into the flow trace (see companion issue on token-usage emission) so users can verify cache hits.

Acceptance

• A pipeline with caching: true on an llm_anthropic profile shows non-zero cache_read_input_tokens on the second and later calls in the same session.
• Default behavior (no caching field) is unchanged — backwards compatible.
• Schema + docs updated.

Suggested labels

enhancement, performance, cost, nodes/llm_anthropic

[github-actions]

[mithileshgau]

Hi @joshuadarron, need advice on how to proceed

The current implementation marks the entire prompt as one cached block, which only helps on exact-repeat calls. Real per-wave savings require placing the cache_control breakpoint between the static prefix (role + instructions + tool schemas — same every wave) and the dynamic suffix (scratch + previous tool results — grows every wave).

The blocker is that chat.py:_chat(prompt: str) receives a flat string — the Question object that knows this boundary is flattened by ChatBase.chat() before it gets there. Three ways to fix it:

Sentinel in the string — Question.getPrompt() emits a magic delimiter; _chat splits on it. Fast to ship, but it's a hidden in-band protocol and forces planner.py to know about a provider-level caching detail (layer violation).

getPromptParts() — Add a parallel method on Question returning (static, dynamic). Problem: Question is already discarded by the time _chat is called, so _chat can't call it without storing state on self.

Pass Question through to _chat — Change _chat(self, prompt: str) to _chat(self, question: Question). Base implementation calls question.getPrompt() — identical behavior. Anthropic overrides to call question.getPromptParts() and emit two blocks. Gemini/Watson need trivial signature updates only. planner.py untouched.

or some other approach.