[octo-patch]

Reason: The shipped registry omits MiniMax-M3.1-Flash, and ordinary listings still offer older MiniMax entries. MiniMax tool continuations also need reasoning fields preserved through node results and agent history.

Add verified Flash metadata and pricing. Ordinary MiniMax listings show Flash first, then M3; explicit legacy lookup, include_legacy=True, saved model IDs and custom registrations remain supported. All 26 existing metadata entries retain their original values apart from three listing-only legacy flags.

On official MiniMax endpoints, forward Anthropic-compatible thinking/effort controls per request without registering global model aliases or mutating shared clients. Ignore M3 effort before SDK validation so xhigh neither rejects the call nor enables thinking. Other endpoints, legacy/custom models and registry-sync opt-out retain their behavior.

Preserve OpenAI reasoning_content and Anthropic signed thinking blocks through shared response handling, message serialization and agent tool history. For official MiniMax Anthropic streams, keep thought text once and concatenate signature fragments within the same thinking block. New thinking deltas and redacted blocks preserve separate block boundaries, including adjacent blocks with identical text. The implementation keeps LiteLLM's existing HTTP, message, response and streaming handling; the small request wrapper delegates to its pooled client.

Validation (Python 3.13.14, frozen lockfile, LiteLLM 1.103.0):

• Full offline unit suite, SDK/HTTP and agent-history regressions, plus 40 independent re-audit cases: 4,372 passed, 1 skipped. Two workers, credential-free environment.
• 70 repository audit regressions cover confirmed SSE and non-stream transport, async controls, exact reasoning/signature preservation through agent tool continuation, images, tool results, message serialization, same custom-node requests before/after official requests, alias preservation, opt-out and ignored M3 effort. The 32 newest cases verify split signatures, distinct identical-text blocks, redacted blocks, sync/async and both current models.
• All 40 independent re-audit cases additionally passed in a fresh registry-opt-out process with sockets blocked. The unchanged split-signature probe reproduces the original complete thought/signature in the subsequent HTTP payload.
• Repository-pinned hooks on the follow-up files: passed, including darker, flake8 and bandit. Earlier registry/JSON hooks also passed.
• One existing test, tests/unit/nodes/prompts/test_manager.py::test_file_tools_persist_across_resets, is explicitly deselected because it makes an unmocked placeholder-key API call. Its timeout was previously reproduced on unchanged base b62e66b93e17139f2bc16f9052b3ac6faba48302.
• Original audit probe replay confirms response reasoning/signatures, preserved custom alias values and no rejected effort combinations. New assertions additionally verify the real agent history and exact stream payloads.

The offline fixtures test protocol edge cases; they do not establish their frequency in production. No paid inference or local Docker build was performed. Remote CI is tracked separately for the latest commit; earlier-head results are not claimed for the follow-up.

Sources: model capabilities, pricing, OpenAI-compatible requests, Anthropic-compatible requests.

[octo-patch]

I corrected the MiniMax China text route to use the documented api.minimaxi.com host and pushed a new commit. I ran the focused registry and MiniMax transport/history tests (156 passed) plus the scoped pre-commit checks.