| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
feat: swap Ask AI to DeepSeek-V4.1-Flash DeepSeek-V4.1-Flash thinks by default at effort 50, so the Qwen-specific enable_thinking kwarg no longer applies; send the top-level reasoning_effort instead. Without it the model spends max_tokens reasoning and returns an empty answer. Log upstream failures so a rejected parameter and a wrong model name are distinguishable in production, where the client only sees a generic 502. Re-ran the 100-question evaluation against the new model. Mean held at 3.54 -> 3.56, and both misleading-assurance answers were eliminated, but five answers regressed by over-literal grounding -- including refusing to multiply two numbers the question supplied. Clarifying that the no-outside-facts rule restricts facts and not reasoning recovered all five without loosening any refusal boundary. Co-Authored-By: Claude <noreply@anthropic.com> Co-Authored-By: wackerow <54227730+wackerow@users.noreply.github.com> | 10 天前 | |
feat(data-layer): expose eth 24hr percent change extends fetchEthPrice to request usd_24h_change from coingecko, adds ethpricedata type with optional percentchange24h, surfaces it through the gas-eth-price api route and usegasethprice hook so consumers can render the indicator. | 5 个月前 | |
fix: move null-user guard to the data boundary + honest type Refines the render-layer guard: filter ghost-authored (user: null) issues in fetchGFIs so they never enter the cache, make GHIssue.user nullable so every consumer is forced to guard, upgrade the IssuesList filter to a type predicate, and guard the gfi-issues-webhook Discord embed. | 3 个月前 | |
migrate api endpoints | 1 年前 | |
Remove Torch webhook payload log | 3 个月前 |
| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
| 10 天前 | ||
| 5 个月前 | ||
| 3 个月前 | ||
| 1 年前 | ||
| 3 个月前 |