feat(openai): update to support GPT-5.6 and require openai>=2.45.0

[solvemproblr]

Summary

Replaces the temporary openai==2.44.0 pin with an intentional openai>=2.45.0 floor and adds GPT-5.6 support (HUD-1923).

OpenAI SDK upgrade

• openai==2.44.0 → openai>=2.45.0. The pin was an emergency response to 2.45.0 narrowing ResponseFunctionCallOutputItemListParam; this upgrade absorbs that change properly.
• Empty function-call outputs now use ResponseInputTextContentParam (the content-item type every other branch of format_openai_result already appends) instead of the no-longer-valid ResponseInputTextParam.
• Regression test: a contentless tool result still emits exactly one input_text content item.

GPT-5.6 support

• OpenAIConfig default model: gpt-5.5 → gpt-5.6 (upstream alias for gpt-5.6-sol).
• hud eval interactive presets: GPT-5.6 added, GPT-5.5 retained.
• Codex cookbook: gpt-5.6 added to the native shell/apply_patch allowlist (per its model page).
• Parameter audit: no agent changes needed — reasoning, text, and tool_choice pass through untyped, and openai 2.45.0 already types the new effort levels (xhigh, max) and reasoning.mode (standard/pro).
• Docs: default-model table and changelog updated.

Testing

• uv run pyright — 0 errors on openai 2.45.0
• uv run pytest -q — 677 passed on Python 3.12 and 3.11
• uv run ruff format --check / ruff check — clean

---

Note

Medium Risk
Touches the core OpenAI agent path and bumps the OpenAI SDK floor, which affects all Responses API tool loops; changes are narrow (defaults, typing, empty-output guard) with test coverage.

Overview
Adds GPT-5.6 as the default OpenAI Responses agent model and surfaces it across the SDK: OpenAIConfig and docs default to gpt-5.6, hud eval gets a GPT-5.6 preset (GPT-5.5 kept), the Codex cookbook allowlist includes gpt-5.6, and the changelog documents the family and reasoning controls.

Dependency: openai moves from a pinned 2.44.0 to openai>=2.45.0, matching narrowed Responses API types for function-call outputs.

Compatibility fix: Empty tool results in format_openai_result now emit a single ResponseInputTextContentParam instead of the removed ResponseInputTextParam, so the API never gets an empty output list. A regression test locks that behavior.

^{Reviewed by Cursor Bugbot for commit 912e194. Bugbot is set up for automated code reviews on this repo. Configure here.}

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
hud	🟢 Ready	View Preview	Jul 15, 2026, 2:55 AM

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
hud	🟡 Building	–	Jul 15, 2026, 2:55 AM