AI Linzi

简体中文说明

AI Linzi connects Alina's business coaching service to a local knowledge Vault. It provides sidebar chat, local search, note-aware content workflows, AI illustrations, long-document processing, and WeChat formatting and publishing.

Features

• Chat with AI Linzi in a sidebar. A current-note request uses the Markdown tab that is still open as the primary source; closing the tab revokes that exact-note context.
• Open independent parallel chats in separate Obsidian tabs. Each tab owns its session, running state, and stop signal; opening the same historical conversation focuses its existing tab instead of creating two writers for one session.
• Select text in both user and assistant messages with the mouse and copy it from the context menu.
• Every compatible text-chat turn gives the currently selected model the same on-demand Vault tools. A named note or folder is searched first for speed, but it is not a hard read boundary; if needed, the model can continue searching the rest of the current Vault. No model can access files outside the Vault through these tools.
• Search Markdown, TXT, text-based PDF, DOCX, HTML, PPTX, and XLSX files locally.
• Ask for a whole-Vault workspace or directory review. The plugin sends the model one bounded metadata-only inventory (folder paths, file-type counts, sizes, modification times, and a short recent-file list), then reads only the few documents whose content is actually needed.
• Add .xlsx workbooks to the main chat from the Vault or directly from the computer. Computer workbooks are converted to bounded worksheet text on-device; the original file is never uploaded or saved in chat history. Legacy .xls files must first be saved as .xlsx.
• Pull images embedded in a .docx handout into the generated deck, placed where they appear in the source (parsed and compressed on-device).
• Build a presentable HTML slide deck from one explicitly selected Markdown, TXT, text-based PDF, or DOCX note: the server returns only a structured outline, the plugin assembles fixed local templates, and images are embedded on-device (three color themes; print to PDF from the browser).
• Run sales reviews and customer consultation briefs from one explicitly selected Markdown, TXT, text-based PDF, or DOCX transcript; the original file stays local. After a brief PNG is generated, its text can be revised directly in the main chat. The plugin keeps only the latest structured draft in local conversation data, sends it only to the brief-revision endpoint when requested, creates a new PNG without overwriting the prior image, and does not add a Markdown source file to the Vault.
• Send only a bounded set of relevant excerpts to the service.
• Explicitly select complete documents for authorized long-document tasks.
• Create new notes and folders after an in-app confirmation.
• Ask in natural language to save a reply as a new note, replace the still-open current note, save that note to the AI Linzi knowledge base, or persist an explicit first-person fact. Success is reported only after the requested write has actually completed.
• Search for a transcript in the main chat, read the selected file in bounded local steps, refine the result over multiple turns, then create, append, or precisely update a confirmed Markdown change set. Up to 12 notes can be included; every target is previewed and version-locked, and a failed step rolls the current change set back instead of leaving a partial result.
• Run an explicitly requested batch file task for up to 36 tool rounds. The plugin can checkpoint up to 500 candidate paths, page cursors, file fingerprints, and failures without storing document text; after stopping or restarting, return to that conversation and say “continue” to re-read the required content safely. Unread or failed files are never counted as completed synthesis.
• Move explicitly requested files to the Trash/Recycle Bin after a separate confirmation. Any file type is eligible (including folders and attachments), and several targets can be batched into one confirmation; an explicit request to delete the current note uses the note path locked when the message was sent, so later tab changes cannot retarget it. Permanent deletion is never performed — items always go to the system Trash/Recycle Bin, and deletion requests must come from you in the current message.
• Generate or revise article illustrations and save successful images locally.
• Format WeChat articles and send them to a configured WeChat draft box.
• Connect a personal WeChat inbox through WeChat's official iLink service. While Obsidian is running, direct messages sent to the connected bot can be saved locally as text, WeChat-provided voice transcripts, images, or ordinary files up to 25 MB. Raw voice/audio and video are not saved in this version.
• Turn a WeChat article into a publish-ready Xiaohongshu note with three title choices, 300–800 Chinese characters of copy, hashtags, and local 3:4 cards that mix the original images with surrounding text.
• Review a local four-platform publishing matrix, five-stage creation pipeline, account growth, and per-post performance for WeChat, Xiaohongshu, Channels, and Douyin. Platform screenshots are analyzed only after an explicit selection and require confirmation before local metrics are saved.
• Review local content activity and authorized account data in the one-person-company cockpit.
• Create, test, import, export, and update portable personal Skills in Skill Studio or by asking in the main chat. Imported and generated Skills use whole-Vault on-demand reading by default; there is no install-time folder permission choice, and a note or folder named at run time only narrows that run's search priority. Main-chat routing distinguishes creating a reusable Skill from asking an existing Skill to generate a business deliverable; when creation versus update or the exact installed target is unclear, the plugin asks before proceeding. Screenshots can accompany an update only as transient read-only context. A main-chat request that asks for both a reusable Skill and the current task can install and start it with one confirmation. Every update locks one installed Skill, shows all text changes and deletions, and writes only after one confirmation. The plugin does not create a separate Skill version history; if a write fails mid-update, it restores the affected files from an in-memory copy. Existing Skills can progressively read referenced files and—when the user separately enables local execution—request per-step confirmation for Node.js, Python, FFmpeg, or FFprobe actions. Shell command strings are not accepted.
• Skill context and Vault permission are separate. A Skill can still search the current whole Vault on demand, while its cloud prompt preloads only the context needed for that task: exact-source processing excludes unrelated profile, memory, methodology, and persona data; content creation may use the user's relevant knowledge and style; Vault dashboards use Vault/task data; explicit business-coaching workflows may use the full coach context. Existing Skills without this optional manifest field remain compatible and are classified conservatively at run time.
• Generate a new HTML, DOCX, PDF, or PPTX deliverable through the same preview-and-confirm flow, or create an editable Markdown workspace whose cards refresh locally when Vault files change. Dynamic workspaces use metadata only, run no generated scripts, and never overwrite an existing file.

Installation

Install AI Linzi from the Obsidian community plugin directory:

1. Open Settings → Community plugins → Browse.
2. Search for AI Linzi.
3. Select Install, then Enable.
4. Open the AI Linzi settings tab and paste a connection key created at chat.alinalinzi.com.

AI Linzi is currently desktop-only and requires Obsidian 1.11.4 or later.

Privacy and network access

Local search remains dormant until the current model actually invokes a file-search or read tool. This capability belongs to the plugin and does not depend on a particular model provider. At that point the plugin builds a local, persistent search index and keeps it current from Obsidian Vault events. The index stores file metadata and a compact term-membership bitset, not document text, and is never uploaded. Search hits are re-read from the Vault before use. A user-named note or folder is treated as the first place to search, while the rest of the current Vault remains available on demand. For a normal chat request, the plugin sends only an explicitly requested still-open note, explicitly authorized documents, or a bounded set of locally matched excerpts. When the user requests a whole-Vault workspace, dashboard, or structural review, the plugin may additionally send one bounded metadata-only inventory: up to 240 folder paths, aggregate file-type/size counts, and up to 40 recently modified file paths with sizes and timestamps. The inventory contains no file content, is excluded from cloud chat history, and is used only for that tool loop. The plugin never uploads the whole Vault's contents, and the model never receives direct filesystem access.

The only supported user document read outside the Vault is an .xlsx workbook that the user explicitly selects or drops into the chat. It is converted to bounded worksheet text on-device; the original workbook is not uploaded, copied into the Vault, or stored in chat history. The desktop-only Article to Video workflow may also inspect installed-program metadata and its npm cache, and use short-lived operating-system temporary files, as described below; it never treats other files outside the Vault as user content.

The plugin reads or writes Vault files only for user-triggered actions such as including a note, saving a generated result, creating or updating confirmed notes, creating a folder, inserting an image, or publishing a WeChat draft. Vault organization is plan-first and confirmation-gated: it can create folders, move, or rename without overwriting. A cross-file Markdown change set is limited to 12 exact paths, shows every complete addition, replacement, or local edit, locks each existing target version, and rolls back completed writes if a later operation fails. When the user explicitly asks to delete files, a delete-only plan lists every target (any file type, folders included, batched within one confirmation) and a separate confirmation moves them to the operating-system Trash/Recycle Bin or Obsidian .trash; folders containing protected paths are rejected as a whole, and the plugin never permanently deletes anything. Move/rename actions have a local undo log. Connection keys and WeChat AppSecrets use Obsidian SecretStorage and are not written to plugin settings or logs.

Cloud features require an AI Linzi account and connection key. Some AI features require a paid entitlement or credits managed on the AI Linzi website. User-triggered requests connect to https://chat.alinalinzi.com; account-scoped chat, entitlement, credit, and reliability records are handled under the AI Linzi privacy policy, and the service-side prompts and orchestration are closed source. WeChat draft publishing connects to the official WeChat API. After the optional WeChat inbox is connected, the desktop plugin receives automatically over an abortable long-poll connection to https://ilinkai.weixin.qq.com; there is no separate receive switch. It downloads selected image/file bytes only from validated official *.cdn.weixin.qq.com HTTPS URLs. The iLink token stays in Obsidian SecretStorage; received content and the polling cursor stay on the current device/Vault and are not sent to AI Linzi. Obsidian must be running to receive in real time. After reopening, the plugin resumes from the saved WeChat cursor when messages are still available for replay, but it does not promise unlimited offline retention. This local transport does not invoke AI or consume AI Linzi credits. The optional “check public IP” settings action connects to https://myip.ipip.net and falls back to https://api.ipify.org. When a user explicitly selects Fish Audio for Article to Video narration, the desktop plugin sends only the confirmed narration text, selected voice ID, and the user's own API key directly to https://api.fish.audio; the key stays in Obsidian SecretStorage and is never sent to AI Linzi. The plugin does not include client-side telemetry, advertising SDKs, or an independent updater.

Local program execution is disabled by default. If enabled, every action is shown for confirmation with its program, arguments, working directory, network declaration, timeout, expected output files, and whether a bounded terminal-output excerpt will be shared with the model. Terminal output is not shared by default and is never written to cloud chat history. The executor does not use a shell, rejects inline Node/Python code, remote FFmpeg inputs, and declared-output overwrites, and only runs Node/Python scripts that the active Skill has explicitly referenced and fully read. Scripts still run with the same operating-system permissions as Obsidian, so users must run only trusted Skills. Generated Vault files can be moved to the operating system Trash/Recycle Bin when they have not changed since generation.

The built-in Article to Video workflow is available only on desktop. After the user confirms the reviewed script, it checks for existing Node.js 22+, FFmpeg/FFprobe, and HyperFrames 0.8.15 installations and invokes those local programs without a shell to render files under AI霖子输出/文章转短视频/. The Community-directory build never installs, downloads, or updates these dependencies. If one is missing, the plugin stops and shows official links and copyable terminal commands for the user to run outside Obsidian, then rechecks only after the user asks it to continue. HyperFrames telemetry is disabled for every invocation. Local narration uses the macOS or Windows system speech service; temporary text and audio files are created in the operating-system temporary directory and deleted after each segment. Environment detection may read executable metadata and the local npm cache outside the Vault, but it does not read user documents there.

Generated images and local conversation cards remain in the Vault or Obsidian-managed local plugin data. Cloud history contains text only and excludes local paths and image data. User-edited titles for Obsidian-plugin conversations are synced as account metadata so the same title can appear on another device; titles are not added to model prompts.

An installed Skill update uses one complete preview and one confirmation. The plugin rechecks the locked Skill immediately before writing and keeps the original affected bytes only in memory for automatic rollback during that operation. It does not create or retain plugin-managed Skill history. Normal Obsidian File Recovery, Trash, sync, or user backups remain available according to the user's own setup.

Security and implementation notes

• The public repository is a thin client. Private prompts, model routing, billing, account data, and service-side orchestration remain on the AI Linzi service.
• Vault enumeration powers user-enabled local search and bounded read-tool results. Tool calls and local paths are excluded from cloud chat history.
• Dynamic workspace rendering and refresh run entirely inside Obsidian from a fixed, validated JSON specification. The generated Markdown cannot execute scripts, load remote code, or bypass the normal confirmation and no-overwrite checks.
• Clipboard access occurs only after an explicit user copy action.
• PDF.js is bundled for local extraction of text-based PDFs. Runtime code compilation is disabled during the production build, so PDF.js uses its built-in interpreter; the release scan requires zero dynamic <script>, eval, or new Function findings. No remote script or CDN is used.
• Streaming chat uses fetch because Obsidian's requestUrl API buffers the complete response and does not expose the response stream.
• The optional WeChat inbox uses Obsidian's network API with an abortable local lifecycle for the official long-poll protocol. It accepts only the connected owner's direct messages, validates API/CDN hosts, bounds media to 25 MB, writes deterministic no-overwrite attachment paths, and stores connection credentials only in SecretStorage.
• The official release contains only main.js, manifest.json, and styles.css. GitHub Actions generates build-provenance attestations for these assets.

Development

npm ci
npm run build
npm test
npm run check:marketplace

Every release tag must exactly match the version in manifest.json. The release workflow builds from source, verifies the version, creates provenance attestations, and publishes the three files supported by Obsidian's plugin installer.

License

MIT