[octo-patch]

Reason: Refresh the existing MiniMax text-agent rollout so MiniMax-M3.1-Flash is the first current preset and MiniMax-M3 remains the second choice for both global and China OpenAI-compatible endpoints.

Changes:

• Updated configs/agents/api_agents.yaml with Flash-first global and China presets while preserving MiniMax-M3.
• Kept explicit MiniMax-M2.7 assignments available through minimax-legacy-agents.yaml and the shared assignment definition.
• Preserved the HTTPAgent request contract, custom model IDs, saved configuration overrides, multimodal message history, and unrelated agent entries.
• Deferred the optional FastChat import until FastChat is actually used, so the HTTP path can run with the repository's lightweight dependencies.
• Documented Flash adaptive thinking defaults and the supported MiniMax request boundary.

Validation:

• env -i PATH="$PATH" PYTHONPATH=. python3 -m unittest src.client.agents.test_agent tests.test_minimax_agents -v — 5 passed.
• env -i PATH="$PATH" PYTHONPATH=. python3 -m unittest discover -s tests -v — 5 passed.
• env -i PATH="$PATH" PYTHONPATH=. python3 -m src.configs configs/agents/api_agents.yaml --output json — passed.
• env -i PATH="$PATH" PYTHONPATH=. python3 -m src.configs configs/assignments/definition.yaml --output json — passed.
• env -i PATH="$PATH" PYTHONPATH=. python3 -m compileall -q src tests — passed; one pre-existing SyntaxWarning remains in the Minecraft subtree.
• Focused tests mocked requests.adapters.HTTPAdapter.send after request preparation and verified serialized global/China payloads, history/image preservation, legacy IDs, custom IDs, and thinking/effort fields. No paid inference or credentials were used.

Limitations: Repository-wide environment validation was not run because the documented optional GPU/model stack is absent and the workspace is below the rollout brief's heavy-build disk threshold. Live CI was not configured on this repository at push time.