@y2zyyr/dsh-model-retry-settings
Configurable model-request retry limit for DeepSeek Harness (Settings → General).
71 results
Configurable model-request retry limit for DeepSeek Harness (Settings → General).
@dsh-external/llm-ratelimit
API/接口设计:Hyrum's Law、稳定接口、REST/GraphQL/模块边界/组件 props 设计原则。受 addyosmani/agent-skills(88k★ MIT)启发。
自主研究 agent:给 AI agent 一个小而真实的 LLM 训练设置,让它过夜自主实验——改代码→固定 5 分钟训练→检查 val_bpb 是否改善→保留或丢弃→重复(约 12 实验/小时、一晚约 100 个);单文件修改、program.md 超轻量 skill。受 karpathy/autoresearch(94k★ MIT)启发。
调试与错误恢复:系统化定位根因、错误分类、恢复策略与防回归。受 addyosmani/agent-skills(88k★ MIT)启发。
文档与架构决策记录:记录为什么而非只是什么,ADR 格式与文档策略。受 addyosmani/agent-skills(88k★ MIT)启发。
个人投研方法论(A股/美股/港股):五维分析框架(估值/资金面/财报质量/行业景气/事件催化与风险)、多空辩论(事实底稿→多方/空方→中立主持归纳分歧)、反思审计(揪出无数据支撑的脑补结论)。受 simonlin1212/Vibe-Research(2.1k★ MIT)启发。
LLM 评估:输出质量评估、幻觉检测、基准测试、回归守护。受 wshobson/agents(38k★ MIT)启发。
计划与任务拆解:把目标拆成可验证的小任务、排序依赖、估算工作量、追踪进度。受 addyosmani/agent-skills(88k★ MIT)启发。
误食催吐判断
Compile Windows .bat/.cmd launchers into real .exe shells inside DeepSeek Harness: renders the icon, emits the C# wrapper, compiles with csc.exe (or dotnet), and installs Start Menu / Desktop shortcuts.
提示词工程模式:结构化推理(CoT/ToT)、few-shot、模板变量、生产级提示词优化。受 wshobson/agents(38k★ MIT)启发。
Patch DSH to use opencode User-Agent instead of deepseek-harness to avoid rate limit discrimination on opencode.ai
DSH tool plugin: generate images through ToAPIs async GPT-Image-2 API (submit task, poll, download).
加班时长合规校验
工资扣除限额
无固定期限合同判断
发布上线:发布检查清单、灰度、回滚预案、上线后监控。受 addyosmani/agent-skills(88k★ MIT)启发。
Markdown view: markdown rendering for DSH client plugins, exposed through the client service mdView. A thin facade over the harness MarkdownText primitive (GFM, math, footnotes, highlighted code) plus README-ism handling: raw <img> tags and document-relative targets resolved against the repo.
源码驱动开发:每个框架决策都要官方文档背书,不凭记忆实现,验证并引用来源。受 addyosmani/agent-skills(88k★ MIT)启发。
任务完成约束自动化插件:监听 AI 回合结束,检测回复是否含 ✅ 任务完成,按「推送许可」开关自动 commit+push 交付物。带 HTTP API(status/config)与 agent 工具(task_completion_status / task_completion_config)。
Patient auto-retry for DeepSeek Harness: honor the upstream gateway's retry_after_seconds on a 429 capacity cooldown instead of giving up after two fast retries, and wait out the mislabeled modality 400 a saturated gateway replays from its client-error circuit. Optional floating countdown badge included.
Triage a scattered backlog into a bounded queue of decisions: every item gets a disposition (now / next / scheduled / waiting / reference / drop) with a stated reason, dropped items leave the list instead of haunting it, and anything that changes a commitment goes out as one approval batch.