Fix #5808: Add support for Claude 4.7 Opus (no assistant prefill, drop temperature)

[devin-ai-integration]

Summary

Fixes #5808 — Claude 4.7 Opus (and Claude 4.6+ models) broke CrewAI because:

1. These models reject requests where the last message has the assistant role (no assistant message prefill support)
2. These models reject the temperature parameter

Changes

BaseLLM (lib/crewai/src/crewai/llms/base_llm.py):

• Added supports_assistant_prefill() method defaulting to True so existing providers are unaffected

AnthropicCompletion (lib/crewai/src/crewai/llms/providers/anthropic/completion.py):

• Added supports_assistant_prefill() override using _is_no_prefill_model() regex-based detection (Claude 4.6+ / 5.0+)
• Temperature parameter is now conditionally dropped for models that don't support it

LLM (litellm fallback, lib/crewai/src/crewai/llm.py):

• Added supports_assistant_prefill() using litellm.get_model_info() with fallback to name-based heuristic

CrewAgentExecutor (lib/crewai/src/crewai/agents/crew_agent_executor.py):

• Added supports_prefill property using getattr() for backward compatibility with custom LLM adapters
• Modified _append_message() to split observation into a user-role message when the model doesn't support prefill, ensuring the conversation never ends on an assistant turn

agent_utils.py (lib/crewai/src/crewai/utilities/agent_utils.py):

• Modified handle_max_iterations_exceeded() to use user role for the forced-answer message when the model doesn't support prefill

Tests

24 new tests in lib/crewai/tests/agents/test_claude_opus_4_7_support.py covering:

• BaseLLM default behavior
• Anthropic model detection (claude-opus-4-7, claude-sonnet-4-6, claude-3-5-sonnet, etc.)
• Temperature dropping for no-prefill models
• LiteLLM model_info detection and heuristic fallback
• Message splitting logic (with/without observations)
• Backward compatibility for custom LLM adapters
• handle_max_iterations_exceeded role selection

Review & Testing Checklist for Human

• Verify Claude 4.7 Opus works end-to-end with a real agent task (the core issue from [FEATURE] Support for new Anthropic model: Claude-4.7-Opus #5808)
• Verify older Claude models (3.5 Sonnet, 3 Opus) still work correctly with assistant prefill
• Check that the regex claude.*?(\d+)[.-](\d+) correctly matches all Anthropic naming patterns
• Verify temperature is properly omitted for claude-opus-4-7 but kept for older models
• Test with a custom LLM adapter that doesn't have supports_assistant_prefill to confirm backward compatibility

Notes

• Supersedes Fix #5808: Add support for Claude 4.7 Opus (no assistant prefill, drop temperature) #5810 which had merge conflicts due to the recent repository refactoring (src/ → lib/crewai/src/)
• The fix uses getattr() with a fallback for the prefill check to ensure backward compatibility with custom LLM adapters that may not implement supports_assistant_prefill()
• The model detection regex is future-proof: it will also flag Claude 5.0+ models

Link to Devin session: https://app.devin.ai/sessions/91e01d2a89e04f188e77f8cb39edb2d8

Summary by CodeRabbit

• New Features
  • Added support and automatic detection for models that don't accept assistant-message prefill (e.g., Claude 4.7+).
• Bug Fixes
  • Adjusted conversation handling to avoid ending on an assistant turn for no-prefill models and to omit incompatible parameters (e.g., temperature) where required.
• Tests
  • Added tests covering prefill detection, parameter gating, and conversation/continuation behaviors.

[devin-ai-integration]

Prompt hidden (unlisted session)

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment and CI monitoring

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: ce9b7390-e194-4bbb-86c0-e6d97e64e423

📥 Commits

Reviewing files that changed from the base of the PR and between 7c68e0e and 6c26461.

📒 Files selected for processing (3)

• lib/crewai/src/crewai/llm.py
• lib/crewai/src/crewai/llms/providers/anthropic/completion.py
• lib/crewai/tests/agents/test_claude_opus_4_7_support.py

🚧 Files skipped from review as they are similar to previous changes (3)

• lib/crewai/src/crewai/llms/providers/anthropic/completion.py
• lib/crewai/src/crewai/llm.py
• lib/crewai/tests/agents/test_claude_opus_4_7_support.py

---

📝 Walkthrough

Walkthrough

Adds model-aware assistant-prefill detection and handling: BaseLLM gains a capability method; Anthropic provider detects Claude 4.6+ and strips temperature; LLM queries LiteLLM metadata with a Claude-version fallback; executor and utilities rewrite assistant messages to avoid final assistant turns; tests added.

Changes

Claude 4.7 Opus prefill support

Layer / File(s)	Summary
Base capability contract in BaseLLM lib/crewai/src/crewai/llms/base_llm.py	BaseLLM adds a supports_assistant_prefill() method defaulting to True, providing an overridable interface for providers to signal whether assistant-role message prefill is accepted.
Anthropic model detection and parameter adjustment lib/crewai/src/crewai/llms/providers/anthropic/completion.py	AnthropicCompletion detects Claude 4.6+ models via regex-based version parsing, implements supports_assistant_prefill() and _is_no_prefill_model() helpers, and conditionally omits the temperature parameter for detected no-prefill models.
LiteLLM model info integration lib/crewai/src/crewai/llm.py	LLM adds supports_assistant_prefill() that queries litellm.get_model_info() for provider-specific metadata and falls back to a Claude-version heuristic (Claude 4.6+ treated as no-prefill) when lookup fails.
CrewAgentExecutor message restructuring lib/crewai/src/crewai/agents/crew_agent_executor.py	CrewAgentExecutor exposes supports_prefill and updates _append_message to restructure assistant messages for no-prefill models: split Observation: into a separate user message or append a continuation user prompt so the conversation doesn't end on an assistant turn.
Forced final answer role selection lib/crewai/src/crewai/utilities/agent_utils.py	handle_max_iterations_exceeded now queries llm.supports_assistant_prefill() and appends the forced final-answer message as role="user" when prefill is unsupported (or assistant when supported).
Comprehensive test suite lib/crewai/tests/agents/test_claude_opus_4_7_support.py	Adds tests for BaseLLM default, Anthropic model parsing, temperature removal for no-prefill models, LiteLLM metadata integration and fallback, _append_message restructuring, and handle_max_iterations_exceeded role selection.

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~20 minutes

Suggested labels

size/M, enhancement

Suggested reviewers

• greysonlalonde

Poem

A rabbit hops through Claude's new halls, 🐰
Where Opus 4.7 now answers calls,
No prefill? No problem — split what it said,
Assistant then user, the flow keeps ahead,
Hopping on code, we keep conversations tall!

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 18.52% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately reflects the main changes: adding support for Claude 4.7 Opus with fixes for assistant prefill and temperature parameter handling.
Linked Issues check	✅ Passed	The PR comprehensively addresses both requirements from issue #5808: it implements assistant message prefill detection and handling across all LLM components, and conditionally omits the temperature parameter for no-prefill models.
Out of Scope Changes check	✅ Passed	All changes are directly scoped to addressing Claude 4.7 Opus compatibility: model detection logic, temperature parameter handling, message role management, and comprehensive test coverage for the feature.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch devin/1778762278-fix-claude-4.7-opus-support

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 2

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@lib/crewai/src/crewai/llm.py`:
- Around line 2295-2302: The fallback heuristic that inspects Claude model IDs
(the model_lower logic using re.search r"claude.*?(\d+)[.-](\d+)") can
accidentally treat date suffixes as the minor version; change the regex to
require the minor version be a short numeric token (e.g., use
(\d+)[.-](\d{1,2})(?!\d) or otherwise assert the minor part is not followed by
more digits) so date-like sequences (e.g., -20250514) are not captured, then
keep the existing parsing/branching (assign major, minor from the groups and
return False when (major == 4 and minor >= 6) or major >= 5) using the updated
match logic on model_lower.

In `@lib/crewai/src/crewai/llms/providers/anthropic/completion.py`:
- Around line 457-459: The regex that parses Claude model strings is treating
date-stamped names like "claude-opus-4-20250514-v1:0" as major=4,
minor=20250514; update the parsing used by _is_no_prefill_model (the function
that decides whether to drop temperature) so the minor version is only captured
when it follows a dot (e.g., match major with something like (\d+) and minor
with (?:\.(\d+))?), then treat missing minor as None/0 when deciding "4.6+";
ensure the temperature-dropping check (the branch that sets
params["temperature"] only if not self._is_no_prefill_model()) uses the
corrected parsed values so date-stamped models are not misclassified.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 70ad8940-51b5-437b-830e-65a10971b9b1

📥 Commits

Reviewing files that changed from the base of the PR and between c36827b and 7c68e0e.

📒 Files selected for processing (6)

• lib/crewai/src/crewai/agents/crew_agent_executor.py
• lib/crewai/src/crewai/llm.py
• lib/crewai/src/crewai/llms/base_llm.py
• lib/crewai/src/crewai/llms/providers/anthropic/completion.py
• lib/crewai/src/crewai/utilities/agent_utils.py
• lib/crewai/tests/agents/test_claude_opus_4_7_support.py