enh: Refactor Event -> Message pipeline outside of CodeActAgent

[csmith49]

End-user friendly description of the problem this fixes or functionality that this introduces

• Include this change in the Release Notes. If checked, you must provide an end-user friendly description for your change below

---

Give a summary of what the PR does, explaining any non-trivial design decisions

This PR pulls the Event -> Message pipeline out of the CodeAct agent and into core.

Additionally:

• Improvements to truncate_content default behavior
• Moves Message conversion tests out of test_codeact_agent.py and into test_message_utils.py
• Refactor CodeActAgent._get_messages with two new private methods to encapsulate behavior

These changes enable us to:

• Move towards a uniform Event -> Message conversion
• Simplify the CodeAct agent implementation
• Compute message representations outside of the agent (useful for condensers and other "predictive" calculations)

Other agents don't use a similar pipeline: the browsing agents build messages manually from prompts, and the delegating agent just delegates.

---

Link of any specific issues this addresses

This PR is necessary to resolve #6707, as we need to compute the accurate message structure outside of the agent.

[csmith49]

This is also an opportunity to move message.py, message_format.md, and message_utils.py into a sub-module. I didn't do so here to limit the files touched but LMK if this is a good spot for the larger re-org as well.

[csmith49]

I'm not sure this function does anything substantial ATM, are there plans to change this?

[enyst]

Oh I think sometimes it matters a lot, because this is the function used to truncate the obs content.

max_chars is 30_000 by default ~= 9k tokens (seen some with anthropic tokenizer API)

This is because a single observation can be crazy large. A fun example is the result of an innocent ls -R /workspace on the Django repository. The observation has ~120k tokens. 😭 And we save them all in the stream.

I guess we could:

• in stream.py, return obs already truncated
• not save the full content in the first place (but we may want to keep more of it than 30k, idk). I am not sure about this one, simply conceptually... we do want the event stream to be "the source of truth" so we can find later whatever the underlying content was. The other side of the argument here though is that the full one is never used.

[csmith49]

Oh, you're absolutely right -- I was misreading the implementation and thought we were taking half = len(content) // 2 instead of half = max_chars // 2 🙃