dsh-tool-bandit-search
A search tool for DeepSeek Harness that learns to pick between quick and thorough search strategies using a contextual bandit (Thompson sampling).
415 results
A search tool for DeepSeek Harness that learns to pick between quick and thorough search strategies using a contextual bandit (Thompson sampling).
MCP management panel for DeepSeek Harness: connection status (green dot), pip/npx local-server upgrade detection & one-click upgrade, add-server form writing cordis.patch.yml, connectivity probe, and usage help.
DeepSeek Harness plugin that enhances custom provider setup by auto-discovering models and auto-populating contextWindow, maxTokens, vision, and reasoning capabilities from models.dev
Standalone lightweight Chat for DSH Web: multi-turn LLM conversations with streaming, history, auto titles, rename and delete — no agent/session/workspace context.
Context compression for DeepSeek Harness: engineer-handoff-note summarization backed by dsh-compaction-basic.
DeepSeek Harness-native Roblox Studio MCP companion: Settings UI listing connected Studio instances, active-instance selection exposed to the agent as runtime context, plus live visibility into Studio's official Script Sync and agent guidance for file-first editing.
Claude Code-style `paths:` rule injection for DeepSeek Harness (DSH): inject rules from ~/.dsh/rules, <project>/.dsh/rules and <project>/.claude/rules into the model context.
TabNexus for DSH:轻量的本地任务、分类、网页资料与流程管理插件。
Per-directory context composition for DeepSeek Harness — .dsh/context.yaml declares systemPrompt + preload for every agent loop; memory is a baseline, not a preset
Focus board for DeepSeek Harness agents: durable, model-maintained notes in the session workspace that pin the objective, constraints, and decisions across compaction and sessions — with structured entry tags (kind/status), a grouped summary view, automatic context injection, an archive on clear, and an optional read-only web panel
DeepSeek Harness bundle that registers the session-summarize skill pack on ctx.skills.
Windows MCP Hub for DeepSeek Harness: scan MCP servers configured in Claude Code / Codex / CodeBuddy, probe real connectivity via stdio & streamable-http initialize handshakes, enable/disable from a /mcp popup panel or the settings page.
Web compatibility layer for the DeepSeek Harness UI: polyfills crypto.randomUUID on non-secure contexts (plain-http LAN access) so the official conversation UI no longer crashes when attaching images, and surfaces the context status in Settings
CodeBuddy-style deferred tool loading for DeepSeek Harness: keep tool schemas out of the model context until the model loads them on demand via tool_search / defer_execute_tool.
Right-click context menus for the DeepSeek Harness (DSH) web UI: open files, reveal in Explorer/Finder, open in a detected editor, copy paths; workspace/session context menus. 对话文件右键菜单插件。
MCP client hub (stdio / Streamable HTTP / SSE) with skill, command, memory and prompt management for the DeepSeek Harness Web GUI
Minimal epoch bootstrap, full-tool promotion, and hard handoff compaction for DeepSeek Harness
右侧面板插件:展示当前会话 Token 用量(tokenUsage/contextPressure 投影)与工作区文件树(host HTTP 路由 + client 面板)
High-fidelity, faithful, bilingual, recursive compaction backend for DeepSeek Harness — a drop-in upgrade over dsh-compaction-basic.
元压缩:模型自己决定何时、如何压缩自己的上下文——列 surface、选区间、以自写文本替换;替换走官方 compaction 事务(检查点、配对平衡、可重建)。Meta-compaction: the model directs its own context compaction through the official compaction seam.
Selection toolbar for the DSH web UI: quote (expandable markers in the composer), temp chat (in-page context-free panel), web ask (chat.deepseek.com popup with ?q= prefill)
Real-time context occupancy meter for DSH Web: draggable compact/detail panel, per-tool usage, settings-editable cost table (CNY/USD, flat or peak/off-peak), per-provider usage and token totals
A local DeepSeek Harness north-star guard with explicit AI indicator evaluation and task alignment context.
Hybrid memory plugin for DeepSeek Harness: L1 snapshot memory (MEMORY.md/USER.md, read on demand, zero context cost) + L2 searchable knowledge base (one-fact-per-file, SQLite FTS5) + L3 import from Hermes/Claude Code/Codex/WorkBuddy and approved Agent Hub documents. No automatic injection. Data lives on D drive, never C.