fix(gateway): skip reasoning_content replay for MiniMax models

[venyon2k]

Problem

MiniMax models generate per-turn-independent reasoning_content. Unlike DeepSeek/Kimi which require it preserved in replay to avoid HTTP 400, MiniMax's reasoning is fresh each turn and adds only context bloat when replayed.

Fix

• _REASONING_BLOAT_PROVIDERS: minimax, minimax-cn, minimax-oauth
• Host-based detection for custom endpoints at api.minimax.io / api.minimaxi.com
• _is_reasoning_content_replay_bloat_model() — returns True for affected models
• _build_replay_entry() gets reasoning_bloat_model=True kwarg — skips reasoning/reasoning_content/reasoning_details when True
• DeepSeek/Kimi/Anthropic are explicitly NOT affected

Tests

• TestReasoningBloatModel: 10 cases (provider slugs, host detection, negative cases)
• TestBuildReplayEntryWithBloatModel: 4 cases (field suppression, non-reasoning field preservation)

Files

• gateway/run.py (+57/-2)
• tests/gateway/test_replay_entry_fields.py (+128/-2)

[daimon-nous]

Looks like a resubmission of #24958 (same author, same title, closed 2026-05-17). Keeping open for review — verify what changed from the prior version.

Related: reasoning_content cluster — #16500, #16892, #22839.

[venyon2k]

Problem

MiniMax models generate per-turn-independent reasoning_content — each turn produces fresh reasoning with no semantic dependency on prior turns. Unlike DeepSeek/Kimi which require reasoning_content preserved in replay to avoid HTTP 400, MiniMax's reasoning is stateless.

Without this fix, reasoning_content accumulates across turns in replay history and is re-sent on every API call. Since MiniMax reasoning output can be substantial, this causes:

• O(n²) token growth: each turn replayed includes all prior reasoning blobs, compounding with session length
• Accelerated context window exhaustion: session dies faster on longer tasks
• Wasted token budget: paying for reasoning content that serves no downstream purpose

Fix

• _REASONING_BLOAT_PROVIDERS: minimax, minimax-cn, minimax-oauth
• Host-based detection for custom endpoints at api.minimax.io / api.minimaxi.com
• _is_reasoning_content_replay_bloat_model() — returns True for affected models
• _build_replay_entry() accepts reasoning_bloat_model=True — skips reasoning, reasoning_content, reasoning_details when True
• Non-reasoning fields (finish_reason, codex_reasoning_items, codex_message_items) are preserved
• DeepSeek/Kimi/Anthropic are explicitly NOT affected

Tests

• TestReasoningBloatModel: 10 cases covering provider slugs, host detection, and negative cases for DeepSeek/Kimi/Anthropic
• TestBuildReplayEntryWithBloatModel: 4 cases covering field suppression and non-reasoning field preservation

Files

• gateway/run.py (+57/-2)
• tests/gateway/test_replay_entry_fields.py (+128/-2)

[venyon2k]

@teknium1 Hey — this is a clean single-commit PR. Happy to address any feedback. The O(n²) token growth issue is specific to MiniMax; DeepSeek/Kimi/Anthropic are completely unaffected.