feat: add session-start prompt-cache preload for crew kickoff (#5921)

[devin-ai-integration]

Summary

Implements the session-start prompt-cache preload feature requested in #5921. Adds opt-in cache_preload and cache_preload_strategy parameters to the Crew class that fire lightweight 1-token cache-warming probes against each agent's system prompt at crew.kickoff() time.

This warms the provider's prompt cache (Anthropic prompt caching, OpenAI prefix caching, Gemini context caching) before the first real task runs, reducing first-step latency and cache-write costs for multi-agent crews.

Changes:

• BaseLLM.preload_probe(system_prompt) — new method on the base LLM class that temporarily sets max_tokens=1 and temperature=0, then delegates to self.call(). This works with all LLM implementations (native providers + LiteLLM). Failures are logged as warnings and do not block execution.
• Crew.cache_preload (bool, default False) — opt-in flag to enable cache warming at kickoff.
• Crew.cache_preload_strategy (Literal, default "parallel") — three strategies:
  • parallel — probes fired concurrently via a thread pool
  • sequential — probes fired one-by-one in agent order
  • shared_prefix — detects the common system-prompt prefix across agents; if ≥ 1024 chars, warms it once before per-agent suffixes; falls back to parallel otherwise
• Crew._preload_caches() — internal method called during kickoff() after prepare_kickoff() completes (agents are fully initialized) but before the process runs. Only activates for crews with 2+ agents.
• Crew._get_agent_system_prompt(agent) — helper that builds the exact system prompt for an agent using Prompts.task_execution().
• Crew._common_prefix(strings) — utility to find the longest common character prefix.

Usage:

crew = Crew(
    agents=[a1, a2, a3],
    tasks=[t1, t2, t3],
    cache_preload=True,                    # opt-in
    cache_preload_strategy="parallel",     # or "sequential" / "shared_prefix"
)
crew.kickoff()

Review & Testing Checklist for Human

• Verify the BaseLLM.preload_probe method correctly sends a 1-token completion via self.call() and does not raise on API errors (review lib/crewai/src/crewai/llms/base_llm.py)
• Verify _preload_caches is called at the right point in kickoff() — after prepare_kickoff() but before process execution (review lib/crewai/src/crewai/crew.py)
• Verify the shared_prefix strategy correctly falls back to parallel when the common prefix is < 1024 chars
• Test with a real multi-agent crew (cache_preload=True) against Anthropic/OpenAI to confirm cache-warming reduces first-step latency
• Confirm that cache_preload=False (default) does not change any existing behavior

Notes

• Feature is fully opt-in — cache_preload=False by default, no behavioral changes for existing users
• Single-agent crews skip preloading entirely (no-op)
• preload_probe is defined on BaseLLM so it works with all LLM implementations (native OpenAI, Anthropic, Gemini, Azure, Bedrock, and LiteLLM fallback)
• 22 new tests added covering all strategies, field defaults, kickoff integration, and edge cases
• All 126 existing crew tests and 56 LLM tests continue to pass

Link to Devin session: https://app.devin.ai/sessions/cd612f749eca4b80af1ebea64e832a8f

[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