fix(llm): prevent infinite recursion and params overwrite in call()/acall()

[eren-karakus0]

Summary

Fixes two critical bugs in LLM.call() and LLM.acall() that occur when a provider permanently rejects the stop parameter with an Unsupported parameter: 'stop' error.

Bug 1: Infinite Recursion (RecursionError)

When the stop parameter is unsupported, the error handler adds "stop" to additional_drop_params and recursively calls self.call() / self.acall(). However, if the provider still rejects the parameter (e.g., due to a provider-side bug, or if the drop_params mechanism fails to strip it), there is no guard to prevent re-entry. The method keeps calling itself until Python hits RecursionError after ~1000 frames.

Before (crashes):

# No check if we already tried dropping 'stop'
except Exception as e:
    if "Unsupported parameter" in str(e) and "'stop'" in str(e):
        self.additional_params = {"additional_drop_params": ["stop"]}
        return self.call(...)  # infinite loop if error persists

After (fails cleanly after one retry):

except Exception as e:
    if "Unsupported parameter" in str(e) and "'stop'" in str(e):
        existing = self.additional_params.get("additional_drop_params", [])
        if isinstance(existing, list) and "stop" in existing:
            raise  # already tried, don't recurse again
        # ... append "stop" and retry once

Bug 2: Additional Parameters Overwritten

The else branch replaces self.additional_params entirely:

self.additional_params = {"additional_drop_params": ["stop"]}

This destroys any existing keys (extra_headers, seed, top_k, etc.) that were previously configured. The fix appends to the existing dict instead:

self.additional_params["additional_drop_params"] = ["stop"]

Scope

Both bugs exist in the sync call() method and the async acall() method. Both are fixed with identical logic.

Test Plan

Added 7 tests in TestUnsupportedStopRetryGuard:

• test_retries_once_then_raises_on_persistent_stop_error - Verifies exactly 2 calls (not infinite), then raises
• test_preserves_existing_additional_params - Confirms extra_headers and seed survive the retry
• test_appends_to_existing_drop_params - Verifies ["another_param"] becomes ["another_param", "stop"]
• test_non_stop_exceptions_are_not_retried - Non-stop errors propagate immediately (1 call only)
• test_acall_retries_once_then_raises_on_persistent_stop_error - Async version of recursion guard
• test_acall_preserves_existing_additional_params - Async version of params preservation

Reproduction

from crewai import LLM

llm = LLM(model="some-provider/model", extra_headers={"X-Key": "val"}, seed=42)

# Before fix: RecursionError after ~1000 frames
# Before fix: extra_headers and seed are lost
# After fix: clean exception after 1 retry, params preserved
llm.call(messages=[{"role": "user", "content": "hello"}])

---

Note

Medium Risk
Touches core LLM invocation error-handling and retry logic in both sync and async codepaths; incorrect conditions could change retry behavior or error propagation across providers.

Overview
Fixes LLM.call() and LLM.acall() retry behavior when a provider errors on the stop parameter by retrying at most once (detecting when stop is already in additional_drop_params) and by adding stop to additional_drop_params without clobbering other additional_params.

Adds a focused test suite (TestUnsupportedStopRetryGuard) validating single-retry semantics, non-stop errors not being retried, appending to existing drop params, and preservation of pre-existing additional_params for both sync and async paths.

^{Written by Cursor Bugbot for commit 03890d7. This will update automatically on new commits. Configure here.}