fix(llms): strip cache_breakpoint from LiteLLM messages

[warren-run-bot]

Summary

fix(llms): strip cache_breakpoint from LiteLLM messages

Agent notes

Move type: ignore[list-item] comment to the line with the spread
operator to properly suppress mypy error about list item type mismatch
when spreading cleaned_messages. This maintains the same number of mypy
errors as before the cache_breakpoint fix.

Run

• Warren run: run_zp2t6219v894
• Agent: pi
• Cost: $2.47 (681 in / 26.3k out / 5.8M cache-r)

Commits (4)

• b2ce1dd fix(llms): strip cache_breakpoint from LiteLLM messages
• e57a946 test(llms): add tests for cache_breakpoint stripping in LiteLLM path
• dc6561e fix(llms): add type annotations for cleaned_messages
• f670f9e fix(llms): adjust type ignore placement for mypy

Files changed

...159Z_01a05990-9abf-7109-9253-0d50244d14e9.jsonl | 211 +++++++++++++++++++++
 .warren/agent.json                                 |  22 +++
 lib/crewai/src/crewai/llm.py                       |  36 ++--
 lib/crewai/tests/llms/test_prompt_cache.py         | 103 +++++++++-
 4 files changed, 360 insertions(+), 12 deletions(-)

Prompt

Show prompt

Work on CrewAI GitHub issue #6789: with the native Mistral API (e.g. model mistral/mistral-large-latest), a direct llm.call() works, but Agent execution through crew.kickoff() fails with litellm.BadRequestError — the request body carries a `cache_breakpoint` key inside message objects (both system and user), and api.mistral.ai rejects it with 'extra_forbidden: Extra inputs are not permitted'. The bug is in the Agent execution path (message preparation for call_llm_native_tools), which injects prompt-cache metadata that only some providers accept. Fix message preparation so provider-unsupported keys like cache_breakpoint are never sent to the native Mistral API — strip or gate them by provider capability — while preserving current behavior for providers that do support prompt caching. Add focused tests proving Mistral-bound messages carry no cache_breakpoint and cache-supporting providers keep theirs. The package source lives under lib/crewai/src/crewai/ and its tests under lib/crewai/tests/. COMMIT DISCIPLINE (critical): commit your work with git as soon as each coherent change compiles, BEFORE running any validation, and keep committing incrementally — never leave completed work uncommitted while a long command runs. VALIDATION DISCIPLINE (critical): validation must be scoped and bounded. Run only the focused test files for the code you changed with 'uv run pytest <path> -x -q' — never repo-wide commands: no full pytest run, no pre-commit run --all-files, no repo-wide type-checker sweep. Give every shell command an explicit timeout of at most 600 seconds; if a scoped test cannot finish in that budget, record that in your summary instead of waiting. Follow the repository's contribution rules strictly: the smallest possible diff, behavior-focused tests, Conventional Commits style for every commit message (e.g. fix(llms): ...), no refactor-only changes, no unrelated edits, and do not touch docs/v*/, docs/images/, uv.lock, pyproject.toml, .github/, or LICENSE. In your final summary, record exactly which commands you ran and their results as evidence.

--- BEGIN AGENT GUIDANCE (repository policy agentGuidance v1) ---
The repository policy binds the following contribution norms. They are binding for this run:
1. Produce the smallest possible diff that resolves the issue; scope changes to what the reviewed issue actually requires.
2. Write behavior-focused unit tests for the fix: assert observable behavior, never implementation details.
3. Never stringify LLMMessage.content with str(); collapse message content to text with the message_content_text helper from crewai.utilities.agent_utils.
4. Never modify files under docs/v*/ or docs/images/ — those are frozen release snapshots managed by devtools; current docs live under docs/edge/en/ and register in docs/docs.json.
5. Use Conventional Commits for every commit message and the PR title (feat|fix|docs|refactor|perf|test|chore with a scope), for example fix(llms): map default context window.
6. Run the repository's own gates before finishing: uv run pytest on the touched test files, plus the configured pre-commit hooks; keep changes DRY and YAGNI.
7. State unprovable claims as known gaps rather than omitting them; silent omissions are harder for review to catch than declared uncertainty.
--- END AGENT GUIDANCE ---

---

🤖 Opened by warren run run_zp2t6219v894