fix(transports+agent): GLM reasoning_effort ultra + drop empty tool_calls key (#70058, #70126)

[ms-alan]

Closes #70058
Closes #70126

Two fixes:

1. reasoning_effort: "ultra" rejected by GLM API (HTTP 400), triggers silent fallback to Claude Sonnet #70058 - GLM reasoning_effort ultra:
   only checked for when normalizing
   to the wire-compatible . GLM models (glm-4, glm-4-plus,
   etc.) use the same parameter but were never normalized, causing the
   raw string 'ultra' to be sent instead of the expected integer.

2. [Bug]: sanitize_api_messages dedup writes tool_calls: [], breaking Qwen-series tool-calling sessions on v0.18.2+ #70126 - drop empty tool_calls key: Step 3
   deduplication replaced with even when
   was (every tool_call had a duplicate id). Strict
   OpenAI-compatible providers (Qwen-series) reject
   with HTTP 400. Fix: when is empty, drop the
   key entirely instead of setting it to .

[alt-glitch]

This was generated by AI during triage.

Related: #69917 covers the focused GLM-5.2 alias policy and #64345 removes the empty tool_calls key. This PR broadens the GLM policy, while its empty-list hunk leaves the original key unchanged when no calls remain; maintainers should choose the intended consolidation.

[teknium1]

Thanks for combining two real provider-compatibility reports.

Problems

• The agent/agent_runtime_helpers.py:2875 hunk does not remove the key when every call was deduplicated. if kept_tcs: leaves the original duplicate calls unchanged in that case, instead of preserving deduplication and dropping tool_calls. Current main's dedup contract is at agent/agent_runtime_helpers.py:3142-3167.
• The GLM change is broader than the existing GLM-5.2 policy in plugins/model-providers/zai/__init__.py:49-82, and it misses the direct summary path: agent/chat_completion_helpers.py:2199-2206 forwards agent.reasoning_config without calling the transport helper.
• The PR adds no regression tests; existing transport coverage only checks GPT-5.6 at tests/agent/transports/test_chat_completions.py:20-35.

Suggested changes

• On a changed dedup list, retain non-empty calls or remove the tool_calls key when none survive; add the all-duplicate regression case.
• Limit GLM normalization to verified GLM-5.2 aliases and apply the same normalizer to summary requests, with tests for both paths.

Automated hermes-sweeper review.

[teknium1]

When every call is a duplicate, kept_tcs is empty and this leaves msg unchanged, including the original duplicate IDs. Please retain the changed-list check and remove the tool_calls key when no calls survive, so the sanitizer does not trade tool_calls: [] for duplicate-ID validation failures.

[teknium1]

Please scope this to verified GLM-5.2 aliases rather than every glm- model. The existing Z.AI profile only maps native reasoning effort for GLM-5.2 (plugins/model-providers/zai/__init__.py:49-82), and the direct summary path also needs the same normalization.

[teknium1]

Thanks for the report-and-fix. This added glm- next to the existing gpt-5.6 guard, but the bug class was broader than the two vendor prefixes — every model on the OpenAI-compatible transport rejects the internal ultra level. PR #90330 fixes it by translating ultra → max unconditionally at the chokepoint (plus sibling sites), which covers GLM without a per-vendor list. The bundled sanitize_api_messages change was unrelated to this bug and would need its own PR with its own justification if still relevant. Closing in favor of #90330.