fix(agent): clamp GLM-5.2 ultra reasoning effort

[shaoohh]

Summary

• normalize ultra to GLM-5.2's supported max reasoning effort before provider-specific request shaping
• cover canonical and relay GLM-5.2 model aliases, including the reported nvidia/z-ai/glm-5.2 route
• preserve the existing reasoning configuration for unrelated model families

Testing

• scripts/run_tests.sh tests/agent/transports/test_chat_completions.py -q (94 passed)
• .venv/Scripts/ruff.exe check agent/transports/chat_completions.py tests/agent/transports/test_chat_completions.py

Fixes #69855

[teknium1]

Thanks for the focused fix. The normal Chat Completions premise is confirmed on current main: agent/transports/chat_completions.py:25-32 only clamps ultra for GPT-5.6, while the Nous profile forwards enabled reasoning configuration at plugins/model-providers/nous/__init__.py:60-68.

Problems

• agent/chat_completion_helpers.py:2199-2206 builds session-summary Chat Completions requests directly and assigns agent.reasoning_config verbatim to summary_extra_body["reasoning"]; it bypasses ChatCompletionsTransport._reasoning_config_for_model(). A GLM-5.2 summary request can therefore still emit unsupported ultra after this PR.

Suggested changes

• Reuse the same model-aware normalization before constructing the summary reasoning body, and add a regression test for that direct summary path.

This is an automated hermes-sweeper review.

[teknium1]

Thanks for the fix. The GLM-5.2 ultra leak is resolved on main: PR #90330 clamped ultra → max unconditionally at the chat-completions chokepoint, and PR #90350 replaced the per-vendor guard lists with a canonical effort ladder + declared per-route vocabularies (GLM-5.2's two-level high/max set included, shared by the zai and opencode-zen profiles) — so this can't regress per-vendor again. Your model-name-token approach was a reasonable extension of the pattern that existed at the time; the class fix just went one level deeper. Closing with thanks.