| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
fix(client): grok-4.7 catalog row and effort ladder; MiMo replays reasoning for every id (#6501) Finishes the interrupted follow-up on this branch. - grok-4.7 (https://docs.x.ai/docs/models/grok-4.7): bundled catalog row (500K context, text/image in, tool calls, structured output, effort low/medium/high/xhigh default high), tiered pricing 0.50/2.00/6.00 (2x at 200K), context window in model_registry and codewhale-models, picker list, agent static registry. Web search is not documented for 4.7, so it is not claimed. Default stays grok-4.6. - xAI reasoning_effort is driven by the catalog row's documented effort option; grok reasoning cannot be disabled so off -> high (https://docs.x.ai/docs/guides/reasoning). Drops the now-unused is_exact_xai_grok_4_6_route. - Xiaomi MiMo replays assistant reasoning_content for every model id (the API 400s on tool-call turns without it), unless effort is off. - Regenerated providers-export.golden.json (adds grok-4.7 to xai). - Docs: grok-4.7 in PROVIDERS/CONFIGURATION (en + zh_hans); zh_hans gains the reasoning_stream_style default note. sh scripts/with-hermetic-test-home.sh cargo test -p codewhale-tui --lib -- client::chat reasoning thinking stream -> 570 passed; 0 failed sh scripts/with-hermetic-test-home.sh cargo test -p codewhale-tui --lib -- grok xai pricing model_registry catalog mimo -> 399 passed; 0 failed; 2 ignored cargo test -p codewhale-config -> 692 passed + 9 + 1 (0 failed) cargo test -p codewhale-models -> 44 passed; cargo test -p codewhale-agent -> 70 passed python3 scripts/check-provider-registry.py -> passed cargo fmt --all -- --check -> clean Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> | 2 天前 | |
fix(models): retain DeepSeek Flash image input in bundled fallbacks Refs #6421 DeepSeek's current Vision guide documents image input for deepseek-flash; its Models guide also accepts deepseek-v4-flash and the existing deepseek-v4-flash-vision-exp alias. The two offline fallbacks still marked the first two names text-only, including the default native DeepSeek route. Correct those four entries without changing other providers, Pro, pricing, limits, catalog freshness or the live/cache/user-override precedence. Sources checked 2026-09-23: - https://api-docs.deepseek.com/guides/vision/ - https://api-docs.deepseek.com/quick_start/pricing/ Verification: - Targeted nextest, config/models: 6 passed, 0 failed; 731 not selected. Covers both new Flash alias regressions, image unknown/text-only states, bundled parsing and user > provider > bundled precedence. - Clippy for config/models, all targets/features, locked, exact CI warning flags: passed. - cargo fmt --all -- --check and git diff --check: passed. - scripts/catalog_models_dev.py snapshot --check: passed (21 providers, 2 canonical models, 99 provider offerings). - Before/after JSON proof: all 4 repaired entries fail the image predicate before the fix and pass after it; every other JSON fact is unchanged. - No web changes; web gates were not rerun for this metadata-only slice. No live provider request, Windows clipboard acceptance, installed upgrade or hosted-green claim. An explicit route/cache override can still supersede the bundled fallback; keep the report open until its exact route is verified. Signed-off-by: CodeWhale Bot <bot@codewhale.net> | 3 天前 |
| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
| 2 天前 | ||
| 3 天前 |