feat(llm): cache recent Claude prompt context

[ousamabenyounes]

Fixes #279.

Summary

• Adds a second rolling Claude prompt-cache breakpoint on the previous message (index: -2).
• Keeps the existing system prompt and latest-message breakpoints.
• Keeps Bedrock within Claude's four-cache-control-block limit by using system + tool_config + previous message + latest message.

Test verification (RED -> GREEN)

RED on origin/main with only the updated expectations:

$ uv run pytest tests/test_inputs.py::test_make_model_settings_enables_prompt_cache_for_bedrock_claude tests/test_inputs.py::test_make_model_settings_enables_prompt_cache_for_non_bedrock_claude -q
AssertionError: ... {'location': 'message', 'index': -1} != {'location': 'message', 'index': -2}
4 failed

GREEN after the fix:

$ uv run pytest tests/test_inputs.py::test_make_model_settings_enables_prompt_cache_for_bedrock_claude tests/test_inputs.py::test_make_model_settings_enables_prompt_cache_for_non_bedrock_claude -q
4 passed

Fix-reverted proof:

$ git apply -R /tmp/strix-issue-279-fix.patch && uv run pytest tests/test_inputs.py::test_make_model_settings_enables_prompt_cache_for_bedrock_claude -q
AssertionError: ... {'location': 'message', 'index': -1} != {'location': 'message', 'index': -2}
1 failed

Validation

• uv run pytest tests/test_inputs.py tests/test_models.py -q -> 92 passed
• uv run ruff check strix/core/inputs.py tests/test_inputs.py -> pass
• uv run ruff format --check strix/core/inputs.py tests/test_inputs.py -> pass
• uv run mypy strix/core/inputs.py tests/test_inputs.py -> pass
• uv run bandit -q -c pyproject.toml strix/core/inputs.py -> pass

[greptile-apps]

Greptile Summary

Adds a rolling Claude prompt-cache breakpoint for the previous message.

• Defines constants for Claude cache-control locations, roles, and rolling message indices.
• Configures system, previous-message, and latest-message breakpoints for Claude routes.
• Retains Bedrock’s tool-config breakpoint while remaining within its four-block limit.
• Updates model-settings tests for supported Bedrock and non-Bedrock Claude models.

Confidence Score: 5/5

The PR appears safe to merge, with no actionable correctness or security issues identified.

The added previous-message breakpoint is ignored when no previous message exists and becomes effective once sufficient conversation history is available, while Bedrock remains within its four-cache-control-block limit.

Important Files Changed

Filename	Overview
strix/core/inputs.py	Adds the previous-message cache breakpoint while preserving route-specific Bedrock behavior and cache-block limits.
tests/test_inputs.py	Updates prompt-cache expectations for supported Bedrock, Anthropic, and OpenRouter Claude models.

_{Reviews (1): Last reviewed commit: "feat(llm): cache recent Claude prompt co..." | Re-trigger Greptile}