llm 220

Floating Markdown note card for the DeepSeek Harness web GUI: per-session multi-note spaces, long-press marks with four colours, italic and underline and your own remarks, clickable task boxes, a repair flow for marks whose text moved under them, reading-

Cross-session and cross-project memory manager for dsh (silent turn-end sedimentation + native web client embedding)

Stop DeepSeek Harness compaction losing its summary: drop the tool schemas from the summarization call, because GenerateOptions has no toolChoice to say "don't call a tool".

dsh plugin: SYSTEM proxy (env) + LLM traffic-splitting proxy via a process-global undici routing dispatcher

OpenAI-compatible gateway adapter and dialect fixes for DeepSeek Harness

给 DSH 模型页补上官方适配器缺的那半：网关可达性探测、模型目录拉取、DeepSeek 余额常驻，以及 MiniMax / 智谱 / Moonshot / StepFun 的余额与配额卡。零运行时依赖，不改动 DSH 安装。

Rate-limit plugin for DeepSeek Harness: FIFO queue per provider/model bucket with minIntervalMs and maxConcurrent

DSH host plugin: retries EVERY model-gateway error (429, quota, auth, network, ...) with a linear 5s-step backoff capped at 5 minutes, covering agent-loop requests and direct ctx.llm.stream calls (compaction, session title). Max retries configurable in Se

A full-featured CodeBuddy (WorkBuddy) provider plugin.

DeepSeek Harness plugin: attaches a stable per-conversation session header to model requests routed to configured providers. Header name is configurable, so one plugin serves SMG (X-SMG-Routing-Key), OpenCode (x-opencode-session), and any other header-bas

VibeDev 账号与模型：在 DeepSeek Harness 里登录 VibeDev，与 DeepSeek 账号并存，使用 VibeDev 网关的模型（Claude、GPT 等），并生成图片、视频、音乐和播客 · VibeDev account and models for DeepSeek Harness: sign in to VibeDev next to the DeepSeek account and use the VibeDev gateway's models (Claude, GPT

DeepSeek Harness plugin: LLM cost metering with budget enforcement (deny/warn), peak-valley pricing and a live usage dashboard

Hand-declared pi-ai LLM routes for DeepSeek Harness whose configured request headers reach the wire, including the ones the harness reserves

Capability metadata for hand-declared gateway models, resolved at runtime by bare model name (context window, output cap, reasoning levels, vision) — matched against the official DeepSeek route's catalog, the installed pi-ai catalog, opencode and models.d

dsh plugin: bounded retry for transient gateway network failures the stock retry policy cannot classify (finish_reason: network_error, unexpected EOF, TLS alerts, stream read errors)

Side-by-side model comparison for the DeepSeek Harness Web UI: send one prompt to 2–4 models in sibling sessions forked from the current one, watch the answers stream in columns (tabs on phones) with latency, time to first token, tokens, and cost, then co

Richer model picker for the DeepSeek Harness Web composer: a searchable provider select (with logos, model counts, and key status) above a keyboard-first, fuzzy-searchable model list grouped by provider, with favorites, recents, context-window/vision/reas

The pi agent's provider catalog (~/.pi/agent/models.json) as DeepSeek Harness LLM routes — installable and removable with `dsh plugin`, no dsh source patch

OpenRouter provider for DeepSeek Harness with a live model list: every model OpenRouter serves, with context sizes, reasoning levels and image input read from its /models endpoint.

DSH web plugin: replaces the Settings → Models page with a one-stop panel — pick a preset or paste a baseURL + API key, fetch the model list from the endpoint, tick what you want, and save. Writes llm-pi-ai providers and credentials for you.

DeepSeek Harness plugin: learning agentic router — rule + UCB + LinUCB + k-NN experts under an EXP3 meta-selector, with a durable quality-reward flywheel, REAL model switching via the session request-header precedence chain, and STEP-LEVEL routing (v1.5.0

Independent team coordination, dynamic routing, usage ledger, and Web panel for DeepSeek Harness

AgentTeams for DeepSeek Harness: multi-agent team collaboration (captain, members, tasks with dependencies, messaging) driven by natural language, with a tree monitor in the web GUI

Per-provider requests-per-minute rate limiting for DeepSeek Harness, with its own Settings tab

Obsidian-compatible OKF v0.2 LLM Wiki plugin for DeepSeek Harness (DSH) — 16 tools, 16 slash commands, auto-rebuild guardrails, ambient recall & personal vault support

脑珊瑚 Coral Memory — 面向 LLM Agent 的热度感知记忆层（DeepSeek Harness 插件）：MCP 记忆工具（memory_search/insert/thread_* 等 16 个）+ 设置面板（统计/容量/检索/配额）。作者：Mr. Code Muggle (@Ne)。

DSH 回复可视化（图解）：把 AI 回复变成立即看得懂的单页图解——模型写单文件 HTML + 内联 SVG，右侧栏沙箱 iframe 渲染

Offloads large, low-judgement DeepSeek Harness tool results to small one-shot worker agents before they enter the main model context

ZenMux 供应商的 DSH 插件：对话内额度显示、推理等级配置、模型目录刷新与额度显式实施。

DeepSeek Harness plugin: role-based model routing — both roles default to deepseek-flash (V4.1 Flash, native multimodal) until V4.1-Pro launches.

Active Context Pruning (ACP) for the DeepSeek Harness — model-driven context management as a CompactionEngine backend.

Auxiliary-call reasoning policy for dsh: the harness marks its own background calls (compaction checkpoints, session titles) with a `purpose` and names no reasoning level for them, so `dsh-llm` fills the gap from the route's declared default and they thin

自动发现 Aperture 网关上的模型，并加进 DSH 的模型选择器。

DeepSeek Harness plugin: CodeBuddy (Tencent Agent SDK) as an LLM provider adapter (ctx.llm). WIP — port of pi-codebuddy-sdk.

Put DSH's automatic context compaction trigger where you want it, per provider/model. Stops DeepSeek Harness from condensing a conversation far earlier than the configured threshold on models with smaller context windows.

DSH model channel manager: roundrobin failover engine (host) + model config view tab (client).

OpenCode Go model provider and settings UI for DeepSeek Harness

DeepSeek Harness (DSH) plugin: LLM token usage & cost tracking with a settings dashboard, agent tools, HTTP API, persistent storage, multi-device cloud sync, session-log history import, official price sync and a multi-vendor price catalog.

AgentRouter 中转聚合：一条 pi-ai 路由承载多模型（设置卡片内可从中转站拉取、逐个编辑模型与参数）+ 国内/国际端点一键切换 + 出站请求 User-Agent/402-配额改写围栏。模型列表写入内核「设置 - 模型」页所编辑的同一份 llm-pi-ai 路由数据。

High-fidelity model calls for DeepSeek Harness: auto-adapts rejected optional parameters and keeps token usage accounting faithful

Stream ceiling for dsh: an absolute bound on one model call, mounted on the llm/stream waterfall. The harness's streamIdleTimeoutMs is re-armed by every SSE keep-alive comment (comments never enter the payload stream), so it bounds keep-alive silence rath

DeepSeek Harness (dsh) plugin that snapshots your workspace at every turn into a shadow git repository, so you can restore the files to any earlier turn. Your own .git is never touched, and plain directories work too.

Configurable model-request retry limit for DeepSeek Harness (Settings → General).

Configurable automatic retries and interrupted-session recovery for DeepSeek Harness

Background auto-fill of llm-pi-ai model capability fields from overrides → pi-ai catalogs → models.dev

DeepSeek Harness plugin that captures the raw HTTP request and response bodies of model provider calls, shows the request body as a collapsible size-annotated JSON tree in a Conversation View tab and on a loopback viewer page, and assembles SSE responses

Same-args circuit breaker for DeepSeek Harness: blocks identical tool calls repeated in a sliding window via ctx.tools.guard (agent loop prevention)

DSH (DeepSeek Harness) 插件：为每个 LLM 模型配置默认推理等级（reasoningEffort），并在请求进入底层调用链前按路由规则自动补齐

OpenCode on DeepSeek Harness — session affinity, Zen/Go gateway origin headers, free-tier tool fallback, and live Go quota display.

AgentTeams for DeepSeek Harness: durable teammates, task DAG, read-only role isolation