[FEATURE] Reduce unexpected cache write costs on Bedrock

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Real-world impact: Thousand dollar overnight bill from cache writes

Running Claude Code in an automated loop overnight on AWS Bedrock with Opus 4.5 resulted in a $1,000+ bill, primarily from repeated cache writes.

The problem:

• Prompt caching is ON by default for all backends
• For Max subscribers, this is harmless (Anthropic absorbs the cost or handles it differently maybe?)
• For Bedrock/API users, cache writes cost 1.25x standard input tokens
• In automated/headless loops, cache results in constant write penalties, few read benefits
• DISABLE_PROMPT_CACHING=1 is not prominently warned, until you see it in AWS cost breakdown surprisingly...

The current default assumes interactive use and flat-rate billing. Bedrock users have unlimited downside exposure and need safer defaults.

Proposed Solution

Proposed fixes:

1. Default OFF for Bedrock/API backends (or at minimum, default OFF in --headless/loop mode)
2. Startup warning when caching is enabled + Bedrock detected: "Caching enabled. For automated workloads, consider DISABLE_PROMPT_CACHING=1"
3. Cost guardrails: Optional spend limit or warning threshold (e.g., "You've spent $X this session")
4. Better documentation: Prominent warning in Bedrock setup docs about caching costs for automated use cases

Alternative Solutions

_No response_

Priority

Medium - Would be very helpful

Feature Category

Configuration and settings

Use Case Example

_No response_

Additional Context

_No response_