Claudian Plus

Claudian Plus embeds provider-backed coding agents directly into your Obsidian sidebar. Codex is the default; Claude, Kimi, OpenCode, Pi, Antigravity, and DeepSeek Harness plug into the same conversation model. When a session ends, its decisions do not evaporate — they are distilled into .claudian-plus/memory.md, a plain Markdown file inside your vault that you can read, edit, or delete.

🐋 DeepSeek Harness

DeepSeek Harness (dsh) is now a first-class provider, implemented directly against the official Harness HTTP Remote and WebSocket mux contracts rather than by adapting another agent's protocol.

• Sessions bind to native Harness history. A conversation survives a plugin reload and resumes after the harness process restarts, because replay reads the harness's own session files instead of a copy.
• What streams is what the harness sends. Text and reasoning streams, tool calls and results, one-shot tool approval, and an explicit Stop.
• Models come from the harness, not from a hardcoded list. Refresh the native catalog in Settings and the picker lists the provider/model pairs Harness actually advertises, plus the reasoning efforts it declares. The native default keeps following the native default instead of pinning a model. Local title generation spends no model credits.
• One harness process per workspace. Tabs claim separate native sessions against a single Web-profile process.
• Off by default, and quiet until asked. Rendering provider settings never starts Harness. First launch can take tens of seconds.

Requirements: the official npm CLI @deepseek-ai/dsh (npm install -g @deepseek-ai/dsh), with Node.js on PATH. The Desktop executable is not supported. In Settings, leave the CLI path empty to let discovery find the dsh shim, or point it at dsh.cmd/dsh.ps1 or its JavaScript entry. Set DSH_HOME in the provider environment if your profiles live outside the default location.

What it deliberately does not do: images, fork, rewind, steering, provider slash commands, shared MCP/skill management, and autonomous background subagents are not exposed. Unknown context capacity is reported as unknown rather than invented as a token count. The full boundary is in src/providers/dsh/README.md.

🌟 What's New in V3.2.2

A hardening release for the generated-UI layer, cold start, and plugin health. Nothing below adds a feature surface; every item was found by measuring the shipped build.

1. The generated-UI components are one design family now

• A timeline item without a timestamp no longer collapses. Grid tracks are driven by a row attribute instead of by how many cells a row happened to emit, and the connector is drawn by a rail that stretches with the row. The same content that rendered one glyph per line in a 10px column (1064px tall) now renders at full width in 161px.
• Obsidian's element styles stop leaking into widgets. The host styles button, input, select and textarea at a specificity that silently beat component classes: buttons were re-centred and repainted, text fields kept the host skin, and multi-line cards were pinned to the input height so their text spilled out of the border. Every button and form-control rule is element-qualified now, with a reset for the properties no component owns.
• One scale for radii, spacing, type and control height. Title sizes collapse from five to three rendered steps — md and lg used to render at 20px and 19.8px, and 2xl and 3xl were identical — and a panel nested inside another panel degrades to a tint instead of drawing a second border.
• One markdown reading style. MarkdownRenderer.render() does not add Obsidian's .markdown-rendered wrapper, so code blocks, tables, quotes and rules inside a widget had no reading polish at all. The answer body and the widget now share a single reading layer, and widget roots are excluded by class so a Table component never inherits markdown-table rules.
• The 361–479px sidebar tier exists. Three- and four-column grids degrade to two, table headers may wrap, and badges may break — that width range previously had no rules at all.

2. Faster cold start

• The production bundle is 3.42 MiB instead of 3.68 MiB, and median cold module evaluation drops from 286.8 ms to ~260–272 ms. Two build settings did it: emitting UTF-8 instead of \uXXXX escapes for the CJK locale bundles, and targeting ES2022 — the minimum supported Obsidian already ships a newer Chromium.

3. Plugin-health hardening

• No renderer storage. Onboarding state and the per-install device key live in the plugin data file now. A device key left by an earlier build is adopted when it is unambiguous; settings that travelled between machines are left alone rather than guessed at.
• Zero production dependency advisories. @modelcontextprotocol/sdk 1.32.2, smol-toml 1.9.1, zod 4.6.7 and the pinned hono override moved off the vulnerable ranges — npm audit --omit=dev reports none.
• Promise rejections carry Error reasons, and a duplicated CSS declaration is gone.

🌟 What's New in V3.2.1

A community-review compliance release. The 3.2.0 listing failed one automated check; this release fixes it.

• No direct inline style writes in the generative-UI renderer. All static styling now goes through CSS classes (is-bordered, is-align-*, is-italic, …) and dynamic values (gap, padding, colours) through Obsidian's setCssProps, clearing obsidianmd/no-static-styles-assignment for src/features/chat/rendering/ui/components.ts.
• The same rule now runs at error level in local CI, so a future release cannot ship a flagged assignment unnoticed.

🌟 What's New in V3.2.0

A feature release, headlined by Intelligent UI: agents can now answer with interactive native-DOM interfaces instead of plain markdown.

1. Intelligent UI (generative answers)

• Agents answer with real interfaces. A provider may reply with a fenced ```claudian-ui program that renders as native Obsidian DOM — never an iframe, never innerHTML — with byte-for-byte transcript replay on every provider.
• A compact declarative language. 54 components: layout (Col, Row, Grid, Card, …), text (rich Markdown, Icon, Math, Progress), data (sortable Table, multi-series Chart, Sparkline, clickable Diagram flows with edges), inputs (Input, Slider, Select, Switch, TextArea, NumberStepper, DatePicker, ChoiceCards, Form, …), structure (Tabs, Accordion, Steps, Timeline, Callout, Stat), learning (Flashcard, locally graded Quiz, Checklist), and Obsidian-native widgets (NoteLink, NoteCard, Mermaid, Tag, vault-relative Image). 24 pure @functions (@Sum, @Round, @Each, …) and 7 actions (@Set, @Reset, @Toggle, @ToAssistant, @OpenUrl, @Copy, @OpenNote) plus Action([...]) sequences. @ToAssistant sends the widget's context back to the agent as the next turn.
• Forms submit once. Form bundles every bound field into exactly one assistant turn; Quiz and Checklist grade or track progress locally without a model round-trip.
• Blocks polish themselves. Settled blocks fade in once, early streaming frames show a skeleton instead of blank space, and a hover toolbar copies the program, resets widget state, or expands the interface. Generated blocks also appear in the conversation outline (UI markers).
• Widget state survives reloads and replay. Slider positions, selections and checkboxes are persisted per conversation (16 KiB per block, 64 KiB per session) under a message-content key, so history replay and a vault reload both find the same state.
• Writing into a note is opt-in. @InsertToNote is off by default; when enabled in Settings → Appearance, every use still shows a preview and asks for confirmation before inserting at the cursor.
• Focus stays where you type. Rebuilds triggered by your own keystrokes — including mid-stream re-renders — restore focus and caret on the exact control.
• One collapsed row per turn. Every reasoning-and-tool run folds into a single Execution details line with an aggregate summary (tool count, names, failures, thinking time). Click to expand the full detail; no reading-mode toggle needed.
• Judge built in. The language is injected into every provider's system prompt (toggle in Settings → Appearance), the renderer degrades unrenderable programs to text instead of blanking them, and a model that forgets the claudian-ui tag on its fence gets a content-sniffing fallback.

2. Memory & consciousness V2

• The memory subsystem is rewritten as a single V2 pipeline (dream jobs, repository maintenance, claim keys, export views) with no new dependencies. Injection budgets are now token-aware.

3. DeepSeek Harness hardening

• Catalog refresh failures are visible and non-destructive. A native catalog that returns nothing or fails per provider now reports diagnostics instead of silently wiping your saved models.
• scripts/sync-dsh-profiles.mjs syncs model-provider config between DSH profiles (Desktop ⇄ web) without touching credentials.

4. Developer experience

• npm run build now deploys the built artifacts straight into a sibling vault checkout's .obsidian/plugins/claudian-plus/ (no-op outside that layout), so build-to-test is one command.
• Bundle budget raised to 3.85 MB for the new surface; the release build stays at 3.62 MiB.

🌟 What's New in V3.1.1

A hardening release. Every item below is a defect found by auditing the 3.1.0 tree, not a new feature.

1. Fixes

• A prompt could land in the wrong conversation. Sending to chat from the file menu, a command, or the quick-input modal while the tab limit was reached wrote the text into whatever live conversation was active. It now stops and explains the limit instead.
• Context usage was wrong for several providers.
  • Antigravity's Gemini models reported against a 200K denominator while running with a 1M window, so long sessions pinned the meter at 100%. The window is now resolved per model family (Gemini 1M; Claude- and GPT-OSS-hosted models stay conservative), and a model id it does not recognise keeps the safe default.
  • The context-capacity selector was a no-op for Antigravity: the value was stored under one model-id spelling and looked up under another, so the percentage snapped back on the next turn. Both spellings now resolve to the same setting.
  • A capacity you picked could not be undone. The meter menu gained Provider's default, which clears the override.
• "Obsidian tools unavailable" on every Claude turn. When the MCP SDK is not present in a build, the failure was retried and reported once per message. It is now reported once and only re-attempted when you explicitly reload MCP servers.
• A console window flashed for ! commands. Local ! command execution on Windows is now spawned hidden, like every other subprocess the plugin starts.
• Claude CLI detection could break entirely. A permission error while probing the npm global prefix escaped as an exception and aborted CLI resolution; it now degrades to "not found there" and keeps searching.
• Multi-step turns rendered as many collapsed blocks. A contiguous run of reasoning and tool calls is now one execution phase with a single summary line until expanded.
• Injected memory context appeared as your own words. Reference material injected into the user turn for ACP-based providers (Kimi, OpenCode, Pi), DeepSeek, Antigravity and Codex no longer shows up when a conversation is replayed from history.
• The header mark borrowed Claude's colour. The provider icon is now monochrome for every provider except Claude, which keeps its own brand colour.

2. Notes

• #11 (terminal console flash) is still open. 3.1.0's fix removed the plugin's own proxy-probe leak. The where.exe call seen in the bundle is not plugin code — it is inlined from the upstream Claude Agent SDK — and the remaining once-per-process flash needs one more data point from the reporter.
• Tab bar. The tab strip stays hidden by design; the new-tab button remains visible and disabled at the limit, and the limit is explained in a notice.

🌟 V3.1.0

DeepSeek Harness returns, rebuilt on the current architecture. It is joined by three fixes reported on the issue tracker.

1. DeepSeek Harness Provider

• Layered transport, not an adapter. DshHost owns the dsh serve process, DshSubprocess owns its stdio, DshNativeConnection speaks the harness' WebSocket mux and HTTP remote contract, and DshRemoteClient puts a typed surface on both. Nothing is translated through another agent's protocol.
• Sessions bind to the harness' own history. A reload, a tab switch, or a process restart resumes the same conversation with its full turn history.
• Models come from the harness. The model list is read from the native catalog and refreshed on demand from the settings tab — no hard-coded snapshot — and reasoning effort is exposed per model.
• One process per workspace. Every tab in the same vault shares one host, so a second conversation does not pay for a second cold start.
• Off until you turn it on. Selecting the provider does not spawn anything by itself; a host starts on the first turn.
• Boundaries stated up front. Images, fork, rewind, steering, slash commands, shared MCP/skill management, and background subagents are not exposed. Unknown context capacity is reported as unknown rather than invented as a token count.

2. Fixes

• Caret colour in dark themes (#9). The composer and the live-preview editor now inherit --text-normal for caret-color, so the caret inverts with the theme instead of staying dark.
• Scrambled history layout (#10). Opening a session from the conversation list no longer creates a second tab for the same id, and consecutive Antigravity assistant messages merge into one content block instead of one block each.
• Terminal flash on every turn (#11). Registry probing for the system proxy now passes windowsHide, removing the plugin's own per-turn console leak. A residual once-per-process flash is still tracked in #11 and is not fixed by this release.
• Codex fork on long threads. Forking passes an inclusive lastTurnId checkpoint to thread/fork instead of forking the whole thread and rolling back, which paginated threads reject.

🌟 V3.0.2

In V3.0.2, we introduced major architectural improvements, network resilience, and modern design enhancements:

1. Smart Tool Grouping & Auto-Collapse

• Eliminate Chat Stream Clutter: When an agent executes a sequence of multiple tool operations (e.g. 10+ consecutive file reads, bash commands, diff edits), tool items no longer overwhelm the screen vertically.
• Dynamic Live Expansion with Auto-Collapse: Tool calls expand smoothly during streaming execution so you can track live operations. Once the tool chain completes, the entire group automatically collapses into a compact ~28px summary bar.
• Full Status Overview: The summary bar displays the total call count (e.g., 13 calls), distinct tool names, and an aggregated health indicator (green for full success, red if any tool fails). One click re-expands the list to inspect parameter details and code diffs.
• Zero Overhead for Single Calls: Single tool invocations remain displayed as clean, single-line items without unnecessary collapse nesting.

2. Antigravity (Gemini) System Proxy Tunneling & Network Resilience

• Root-Cause Fix for Idle Disconnects (WSAECONNRESET 10054): Go CLIs (agy.exe) do not natively inspect the Windows Registry for Internet proxy settings. Under VPN / Clash TUN Fake-IP routing, idle TCP connections during extended multi-minute thinking turns were routinely severed by TUN idle timeouts after 4m 30s.
• Zero-Config Automatic Registry Proxy Detection: Claudian Plus now features an integrated Windows Registry system proxy sniffer (resolveSystemProxyEnvironment) that silently injects HTTP_PROXY into the child process. The CLI automatically routes via loopback tunnels to your local proxy, providing bulletproof connection stability without manual user configuration.
• Official Brand Assets & Full Gemini Spectrum: Shipped with official Antigravity SVG icons and comprehensive support for Gemini 3.1 Pro, 3.7 Flash, 3.8, and Thinking budget tuning.

3. Modern Minimalist Stop Button

• Flagship AI Interface Parity: Replaced the bulky, reddish rectangular Stop badge with a sleek, borderless circular stop button (28px × 28px) matching ChatGPT / DeepSeek design standards.
• Adaptive High-Contrast Dual Theme: High-contrast white circle with a dark rounded square in dark mode; solid dark circle with a white rounded square in light mode. Features tactile micro-motion (scale(1.06) hover, scale(0.94) active press) with native Stop generation (Esc) tooltip.

4. Layout & UI Polish

• Header Icon Alignment: Corrected the vertical drift on the provider icon in the header, preventing visual clipping at the top edge of the window.
• About & Community Hub: Integrated developer WeChat channel and issue feedback directly into the settings view.

01 / Watch it work

A 30-second animated overview of how the pieces fit together:

Three things worth noticing

• Context arrives as references, not pasted text. @note, @folder, and drag-and-drop resolve against the vault you already have open.
• Memory lands in a file you can open. No database, no sync service — just memory.md sitting in your file list next to your notes.
• The provider switches without the conversation breaking. Each provider keeps its own session history, so a session can move between Codex, Claude, Kimi, OpenCode, Antigravity, and DeepSeek Harness.

02 / Why it is different

Most AI plugins treat conversation history as disposable text. Claudian Plus treats a session as a working record: context enters, decisions distill into memory, and the result stays searchable weeks later.

03 / What this is not

• Not a cloud service. No telemetry or analytics code ships in this plugin. Network calls are made by the provider CLI you configure, from your machine.
• Not a chat window bolted onto a note app. Your memory is a Markdown file in your vault, not a row in someone's database.
• Not a promise that every provider behaves the same. Capabilities differ per provider and are declared explicitly, because the codebase is written so no feature can silently assume parity.

04 / How memory actually works

Conversations end; memory continues. The pipeline is deliberately boring — four steps, no timers, and nothing written behind your back.

• It is gated on real usage, not on a stopwatch. Distillation is evaluated when a turn completes, and only once a session has at least 3 exchanges. A separate hourly check consolidates older logs when a full interval has passed, and startup consolidates anything accumulated while Obsidian was closed.
• Input is capped before the model sees it. The transcript handed to the auxiliary query is truncated at 8,000 characters, so distillation cost stays bounded on long sessions.
• Nothing is written silently. Newly extracted preferences and decisions surface as a notice and wait in Settings for you to accept.
• Output is capped before it reaches your prompt. Approved memory is injected within a 1,500 character budget by default, split across a 350-character profile layer and a 500-character project layer. All three are adjustable in Settings → Memory & Consciousness.
• You can always pull the plug. Open memory file shows you exactly what has been distilled, and the file is plain Markdown.

Memory is opt-in. memoryEnabled defaults to on, but the consciousness and auto-memory switches — the ones that send a transcript to an auxiliary query — default to off on new installations, because distilled memory contains personal context. Turn them on in Settings → Memory & Consciousness when you want it.

05 / Providers and the dependency rule

Three layers, one direction of dependency: vault context feeds a provider-neutral core, which dispatches to provider adaptors. Feature code depends on core contracts, never on provider internals.

• Codex CLI — enabled by default. Default model gpt-5.6-sol, with native streaming output over codex app-server.
• Claude Code — enabled by default. Thinking-chain folding, permission modes, and native project history replay.
• Antigravity (Gemini) — opt in. Deep integration with automatic system proxy tunneling, multi-model selection, and official brand assets.
• DeepSeek Harness — opt in. Native HTTP Remote + WebSocket mux integration with durable history, provider-advertised models and reasoning efforts, one-shot tool approval, and one shared process per workspace. Needs Node.js on PATH; the Desktop build is not supported. Setup and the exact capability boundary live in src/providers/dsh/README.md.
• Kimi — opt in. ACP integration with dynamic model and command discovery, per-tool approval dialogs, and multimodal image attachments.
• OpenCode — opt in. ACP agent over a sandboxed sidecar process.
• Pi — opt in. RPC-mode sidecar that runs on Node built-ins and a bundled TypeBox, so it still works when external tool dependencies are missing.

06 / Installation

Method 1: Install from Release (Recommended)

1. Download main.js, manifest.json, and styles.css from the Latest Release (v3.0.2).
2. Create a folder named .obsidian/plugins/claudian-plus/ inside your vault.
3. Copy the three downloaded files into that folder.
4. In Obsidian, go to Settings → Community plugins, refresh the installed list, and toggle Claudian Plus on.

Requirements: Obsidian 1.11.4 or later, desktop only. Claudian Plus communicates with local agent CLI processes and requires desktop filesystem access.

Method 2: Build from source

Requires Node.js 24, plus at least one provider CLI installed locally: Codex, Claude Code, Antigravity, DeepSeek Harness, Kimi, OpenCode, or Pi.

git clone https://github.com/wuyifan-code/Claudian-plus.git
cd Claudian-plus

npm ci
npm run typecheck
npm run build

Tip: Set OBSIDIAN_VAULT=D:\Obsidian\My Vault in .env.local so npm run build copies the build output directly to your vault.

07 / Commands reference

08 / Privacy & Storage Layout

Everything the plugin knows lives inside your vault:

<your-vault>/
├── .claudian-plus/
│   ├── claudian-plus-settings.json    # Shared settings + provider configs
│   ├── memory.md                      # Distilled memory injected into prompts
│   └── sessions/
│       └── *.meta.json                # Session metadata
├── .claudian/                         # Legacy Claudian data, read-only migration
└── .obsidian/plugins/claudian-plus/
    ├── main.js
    ├── manifest.json
    └── styles.css

• No telemetry. There is zero analytics code under src/. Data leaves your machine only via the provider CLIs you configure.
• Data safety during migration. Existing .claudian/ directories are detected and migrated non-destructively; original files remain intact.
• Settings merge cleanly. Provider settings are merged rather than clobbered, preserving existing configs across updates.

09 / Development & Testing

npm run dev                  # Build CSS + start esbuild in watch mode
npm run typecheck            # TypeScript type check
npm run lint                 # ESLint checks
npm run test                 # Run all unit and integration tests
npm run test:watch           # Run tests in watch mode
npm run test:coverage        # Generate test coverage report
npm run test:architecture    # Architecture boundary enforcement
npm run check:performance    # Bundle size and startup performance check

10 / Upstream Projects

Claudian Plus is built upon two upstream open-source projects:

Original and Claudian-derived code is licensed under MIT; Codian-derived code retains AGPL-3.0 obligations. See LICENSE and NOTICE.

Community contributors

Diagnostics that changed the implementation, credited where the fix landed:

💡 Community & Follow

在递归中遇见更深的自己 (Discovering a deeper self within recursion)

If you encounter any issues, have questions, or want to explore feature ideas:

• File an Issue: GitHub Issues
• WeChat Official Account: 「递归识海」
• Follow & Updates: 关注公众号：递归识海