[octo-patch]

Reason: The built-in MiniMax catalog and TUI defaults still select retired M1/M2 models, and Flash requires adaptive thinking instead of the older enabled/disabled request format.

• Offer MiniMax-M3 and MiniMax-M3.1-Flash on the existing international and China MiniMax endpoints; default the TUI's MiniMax selection to Flash.
• Remove 30 retired MiniMax entries from the vendored catalog and regenerate the matching snapshot/provider entries, while preserving M3 gateway entries and all non-MiniMax definitions. The importer also excludes retired MiniMax models on later refreshes.
• Record the documented 1,000,000-token context, 524,288-token output maximum, text/image input, and current Standard pricing. Leave unpublished cache-write pricing unset and keep cost calculation finite.
• Send adaptive thinking and supported effort levels for Flash, preserve M3's on/off behavior, and omit manual cache-control markers for MiniMax. Preserve exact supported effort choices so the shared effort selector does not collapse Flash's max setting into xhigh.
• Document that Devo cannot represent video input yet and that its flat M3 price estimate does not model the published long-input and Priority tiers.

Validation:

• npm ci --ignore-scripts --no-audit --no-fund succeeded in apps/tui.
• node --import tsx --test src/minimax-models.test.ts src/model-catalog-refresh.test.ts src/normalize-agent-model.test.ts lib/coding-agent/src/modes/interactive/interactive-mode-model-candidates.test.ts passed: 19 tests. Request tests intercept payloads before network access.
• npx --no-install tsc --ignoreConfig --noEmit --module Node16 --moduleResolution Node16 --target ES2022 --strict --skipLibCheck lib/ai/src/models.ts lib/ai/src/providers/anthropic.ts lib/ai/src/cache-pricing.ts passed.
• Catalog verification confirmed reproducible MiniMax provider entries and unchanged non-MiniMax definitions. Existing manually maintained Codex endpoint/header values were preserved.
• Full TUI npx --no-install tsc --noEmit fails with diagnostics in unrelated auth, daemon/session, and native bridge code, including auth-storage.ts PrimeTeam fields and snapshot.ts AgentSession fields.
• cargo test -p devo-provider anthropic::messages::tests --lib passed: 15 tests, including adaptive thinking and Flash effort serialization.
• cargo test -p devo-core model_catalog --lib passed: 11 tests.
• cargo test -p devo-protocol reasoning_effort --lib passed: 17 tests, including preservation of distinct max/xhigh choices.
• Changed Rust files were formatted; git diff --check and the configured credential scan passed.

Sources:

• https://platform.minimax.io/docs/api-reference/text-chat-anthropic.md — exact model IDs, modalities, thinking/effort behavior, and the max_tokens range [1, 524288] for both M3 and Flash.
• https://platform.minimax.io/docs/pricing/overview.md — current input/output/cache-read prices, M3's literal 512k tier boundary, and Priority multiplier.

[octo-patch]

I updated the branch with the latest main and resolved the model catalog conflict while keeping the MiniMax-M2.7 entry and regression coverage. I ran cargo fmt --all --check and cargo test -p devo-core builtin_minimax_catalog_contains_both_target_models.