Fix Codex GPT model selection

[AarushiShah-db]

Summary

• Select the highest semantic GPT version for Codex instead of relying on discovery order (before, we were auto choosing gpt-5 instead of gpt-5-5
• Convert Databricks GPT endpoint IDs like databricks-gpt-5-5 to OpenAI-style model IDs like gpt-5.5 in Codex config metadata so we don't get the warning signs

before:

after:

Tests

• uv run pytest tests/test_agent_codex.py

[rohita5l]

Good catch on the sort bug.

This adds a second GPT-version parser. _openai_model_id and the new _gpt_version_key are both pulling major/minor/patch out of a model id, but they do it differently (one's hyphen-only on the raw string, the other splits on / and takes . or -). They'll drift apart eventually. I'd pull the parsing into one helper and build both the OpenAI id and the sort key off it:

_GPT_RE = re.compile(r"(?:databricks-)?gpt-(\d+)(?:[.-](\d+))?(?:[.-](\d+))?(-.+|[a-z].*)?")


def _parse_gpt(model: str | None):
    """(major, minor, patch, suffix) from a Databricks- or OpenAI-style GPT id.
    minor/patch are None when absent; None for non-GPT ids."""
    if not model:
        return None
    m = _GPT_RE.fullmatch(model.split("/")[-1])
    if not m:
        return None
    major, minor, patch, suffix = m.groups()
    return (
        int(major),
        int(minor) if minor is not None else None,
        int(patch) if patch is not None else None,
        suffix or "",
    )


def _openai_model_id(model: str | None) -> str | None:
    parsed = _parse_gpt(model)
    if parsed is None:
        return model
    major, minor, patch, suffix = parsed
    version = str(major)
    if minor is not None:
        version += f".{minor}"
    if patch is not None:
        version += f".{patch}"
    return f"gpt-{version}{suffix}"


def default_model(state: dict) -> str | None:
    codex_models = state.get("codex_models") or []
    if not codex_models:
        return None

    def _key(mid: str):
        parsed = _parse_gpt(mid)
        if parsed is None:
            return (0, 0, 0, 0)
        major, minor, patch, suffix = parsed
        return (major, minor or 0, patch or 0, 0 if suffix else 1)

    return max(codex_models, key=_key)

Couple of other things while I'm here:

The try/except ValueError around max can't actually fire, you've already returned early on an empty list, and _key catches its own errors, so there's nothing left to raise. Safe to drop.

I think gpt-4o-style ids slip through. The suffix in the current regex needs a leading hyphen, so databricks-gpt-4o doesn't match and gets returned as-is, prefix and all. Codex would choke on that. The [a-z].* branch covers it. Might be worth a test since everything here is gpt-5 family.

Also default_model strips mid.split("/")[-1] but _openai_model_id doesn't, so a namespaced id would map in one place and not the other. Going through _parse_gpt evens that out, assuming ids can show up with a / at all ?