feat: preserve cache control breakpoints when injecting tools

[dannykopping]

Closes #69

When MCP tools are injected into a request that already has tools with
cache control breakpoints, we need to move the breakpoint to the final
tool in the list. This ensures prompt caching continues to work
correctly as documented by Anthropic.

The implementation:

• Captures the first cache control breakpoint found on existing tools
• Clears it from its original location
• Sets it on the final tool after injection

[dannykopping]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• feat: preserve cache control breakpoints when injecting tools #97 👈 (View in Graphite)
• fix: only disable parallel tool use when tools are injected #96
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[pawbana]

Instead of moving breakpoint could we change tool ordering moving injected tools to beginning of the list?

[dannykopping]

I mean... probably? It's the same outcome.

[dannykopping]

Hhmm ok maybe that's the more efficient way to do this, thanks.
I'll implement that tomorrow 👍

[pawbana]

Not exactly. If I understand caching/breakpoints correctly moving breakpoint could result in it not being usable.

Example:
Call 1: tools[A, B(breakpoint), C, Injected_1, Injected_2]
Call 2: tools[A, B, Injected_1, Injected_2] -> could be cached if breakpoint is not moved

If injected tools are moved to the front this example will work.

Downside is that changes to injected tools will break all breakpoints but I would imagine injected tools change less frequently then tools in requests.