Cancip

Obsidian Cancip AI is a right-side AI chat panel shaped toward a mobile-first agent workflow for Obsidian.

Cancip is a lightweight prototype for managing an Obsidian vault from a mobile-friendly AI panel:

• Session completion and failure notifications use the installed Ntfy Notifications hub first, so its selected default Channel, recent-conversation routing, review, quiet hours, receipts, and error handling apply consistently. Cancip's legacy direct ntfy topic remains only as a compatibility fallback when the hub plugin/API is unavailable.
• Process-record raw blocks keep Copy and Wrap controls in a fixed header above an independent scroll layer. Hidden process/tool blocks share the same controls, so long lines stay horizontally scrollable by default and can be wrapped on demand.
• Separate model sources and models: each source stores its own Base URL, key, and API mode, while each model ID is explicitly bound to one source and a single default model drives AI modules unless they override it.
• Multilingual UI with auto device-language detection for Simplified Chinese, Traditional Chinese, English, Uyghur, Turkish, Russian, Japanese, Korean, Spanish, French, German, and Arabic; missing low-frequency strings fall back to English, and Arabic/Uyghur use RTL layout hints.
• Automatic OpenAI Responses and OpenAI-compatible Chat Completions support.
• <vault.configDir>/plugins/cancip/data/config.json as the authoritative vault-level config (normally .obsidian/plugins/cancip/data/config.json).
• Two execution access modes only: confirmation mode reads freely and queues write-like actions for approval; full-access mode executes implemented actions directly. Access is controlled only by the UI or the Cancip data config, not by conversation text.
• Current note, selection, dynamic @ mentions, visible long-term/core memory, and universal hard-first search across notes, sessions, memories, paths, configs, PDFs, images/OCR sidecars, Office files, archives, and other Vault files.
• cancip-action JSON tool blocks for validated vault-relative actions, including read/write/append/patch/mkdir/rename/copy.
• Agent-style tool runs: approval mode queues action blocks under the assistant message with Run/Reject controls, while Full access executes and records results.
• Tool result continuation loop: after tools finish, Cancip can feed results back into the model and continue for a bounded number of iterations, closer to local agent runs.
• Outcome verification closes the execution loop inside Obsidian: cancip.outcome.observe/verify compare the active view, DOM/layout, file readback, plugin state, and workspace leaves with explicit expectations; capture adds active-region PNG evidence only when structured checks are insufficient, and exportPdf uses an installed exporter for PDF evidence. Failed checks trigger bounded difference-only correction, while reports and screenshots stay under the Cancip data evidence/ folder with direct review links and no persisted Base64 payloads.
• Structured command bus actions for Obsidian internal commands, Cancip built-ins, and GitHub CLI-equivalent REST API commands.
• Dedicated Mobile Webviewer integration through cancip.webviewer.help/status/read/selection/open/tabs/tab/bookmark/reading/send. Cancip reads the page already open in Mobile Webviewer through its versioned plugin API, preserving URL, title, selection, edited reader text, images, and links; generic plugin routing and web.fetch remain compatibility fallbacks.
• Native File Explorer pinning: pin/unpin files and folders from the normal context menu, keep mixed pinned siblings above ordinary items, reorder with explicit up/down buttons or drag handles, and persist folder-local order in the Cancip data file-pins.json without replacing Obsidian's normal unpinned sort or file drag/move behavior.
• Universal document workbench: every Vault file has an explicit workbench action in the native file menu without changing its normal default view. Text, HTML/MHTML, CSV, JSON, PDF, DOCX, XLSX, PPTX, media, ZIP-based documents, and unknown binaries receive format-aware previews. One export menu converts the current workbench document to Markdown, HTML, PDF, DOCX, PNG, or PPTX, verifies the non-empty output, and opens it in the workbench; Markdown/HTML use Cancip's semantic converters, while PDF/DOCX/PNG/PPTX use the installed Mobile PDF Exporter. Binary originals remain protected.
• Obsidian-native Markdown rendering for chat messages, including Obsidian-supported HTML.
• Cancip chat and rendered Markdown note code blocks keep Obsidian's copy action beside one global wrap toggle. Unwrapped horizontal scrolling is the default; one toggle updates current and future chat/note code blocks and persists across restarts.
• Long-term/core memory defaults to visible AI/Cancip/Memory/ and is included in every model interaction.
• Vault orientation keeps the Cancip-owned AI/Cancip/Memory/VAULT_OVERVIEW.md current after startup and material file/plugin changes. It is a shallow programmatic map of top-level folders, file kinds, recent user-facing files, and installed Obsidian plugins, so later turns can pick the right folder/plugin to inspect on demand without sending the whole vault; a user-authored replacement without the Cancip marker is never overwritten.
• Full-vault search is not attached by default. Cancip should first use long-term memory and necessary short-term/session context, then decide whether to run cancip.searchVault and read only the necessary matched files.
• Model calls use a payload policy: trivial chat stays lightweight, informational turns add only targeted context, and implementation/self-repair turns include the full tool protocol and compact memory.
• Skills and experience recipes are routed on demand: memory/rule/preference, OB plugin, command, attachment, and self-optimization tasks can auto-select relevant Skills, query cancip.skills.* / cancip.experience.*, and harvest repeatable successful workflows into the Cancip data skills/generated/ folder.
• Full session export from the chat header to Markdown and JSON under visible AI/Cancip/Exports/.
• Lightweight project session history stays under the Cancip data sessions/ folder, opened from the compact history button beside the new-chat button.
• Every session now keeps a normalized timeline: immutable creation time, latest activity, first start, and available completion/stop/failure times. History rows expand to exact local time with seconds and no timezone suffix; session commands, event audit, parent/child rows, and Markdown/JSON exports expose the same canonical ISO timestamps. Legacy sessions recover only timestamps that can be determined from stored data or a valid session ID.
• The Review Gate view opens immediately on desktop and mobile, recovers deferred or stale leaves, reuses an already rendered review page, and avoids scanning the same package again when the requested package and file have not changed. Machine review packages stay under the Cancip data review-gates/ folder; startup also imports legacy .cancip/review-gates/ packages without deleting the legacy backup. Paired review panes synchronize vertically on mobile and horizontally on desktop.
• Review data now uses one deduplicated parsed snapshot shared by the status bar, pending count, file list, and Review view. The snapshot is invalidated by Review changes, prewarmed before the view is revealed, and reused for ten seconds, eliminating repeated package scans and empty-list flashes. Status attention refreshes are coalesced and DOM updates are skipped when counts are unchanged.
• Compact context chips live inside the rounded composer/input box: the current active file is shown automatically with its extension, and source/context chips no longer occupy a separate panel.
• Context chips can be opened directly: file chips open the file in a tab, folder chips reveal the folder in the file navigator, and the small x removes only that context chip.
• Search is deterministic first: Cancip uses a persistent metadata/Bloom index, rereads every candidate before returning it, and includes source-backed hard-search excerpts in model context. If those results are insufficient, the active model may call cancip.searchAll once with 2-4 explicit softQueries; Cancip does not make a hidden second model call solely for query expansion. Config/session excerpts are redacted before indexing and return.
• Rounded composer with floating upward icon trays for context, access mode, and model selection; trays overlay from their buttons and close after selection or outside taps.
• Header controls include a single Plan button. The Plan button opens the planning/todo panel; it is not a chat mode and does not change read/write permission.
• The header includes an OB Review Gate button wired to a programmatic TypeScript builder adapted from arias007/ob-review-gate-skill: it scans selected vault files, writes review data under AI/Cancip/Review/, and opens it inside a native Cancip audit panel with file lists, structure changes, diffs, old text, and new text.
• Structured Plan todos are available as cancip-action tools, so the agent can set/add/update/remove/list/clear the visible Plan panel during an agent run instead of only describing a plan in prose.
• The composer keeps the access selector visible and wider for mobile tapping, with a paperclip attachment button beside it for quickly adding file/folder context.
• Settings expose common settings and every advanced category as peer horizontal pages. The tab strip scrolls on mobile, only the selected page is rendered, and changing an option no longer collapses an outer advanced section.
• The settings page can show two payment QR codes embedded directly in main.js; no extra image files are required, and the images are not included in prompts or JSON exports.
• Built-in model presets include GPT, Claude, Gemini, DeepSeek, Qwen, and Kimi-style names while still allowing a custom model string.
• Hierarchical @ picker for sessions, Skills, automations, files/folders, plugins, Cancip functions, command bus entries, and real Obsidian commands. Empty @ first shows human-readable categories with Session history, Skill, and Automations fixed at the top; selecting a category loads and filters only that list. Stable session/automation mentions survive title changes, and an automation mention plus follow-up text runs that task directly with a one-time supplemental prompt.
• Mobile composer geometry follows the active Obsidian WebView visual viewport, keeping the input and @ picker above the Android keyboard. The Review detail shell reserves the Obsidian mobile status bar and keeps structure changes directly beside the content diff.
• Mobile keyboard positioning measures overlap between the stable chat host and the active WebView visualViewport while the composer is focused. Review bottom clearance uses the measured Obsidian status-bar height on every mobile viewport width instead of relying on a fixed narrow-screen media query.
• Button settings include Send button info to Cancip again, using the existing actionable button-context route without changing the button's normal short-press behavior.
• Running conversations use one first-level process record. Explicit readable model progress stays visible as numbered step headlines while API profiles, raw sent/received payloads, routing audits, tool JSON, tool results, and file-action details remain collapsed. Attribute-bearing <details> audit blocks are parsed structurally, empty audit blocks are omitted, and markup/JSON/truncation text cannot leak into step headlines. The process record opens while the request is running and collapses after the final answer.
• Compact live plan and changed-file summaries sit beside the Cancip title and in the composer status strip. The composer shows a short current status plus real +added/-removed totals and completed/total plan progress; the metadata links open Review Gate and Plan directly.
• New chat renders and focuses immediately, then serializes the previous session and bootstraps the new session in the background. Session history reads before replacing visible rows, keeps the previous list during refresh, lazy-loads in batches, puts pinned sessions first, and shows loaded/total counts in the header.
• Sessions unused for 30 days move out of the hot session set into reversible cold storage under the Cancip data archive/ folder. Pinned, running, actively requested, and currently open sessions are protected. Cold sessions stay visible/searchable and are copy-verified before the hot file is removed; opening or restoring one verifies it back into the data sessions/ folder. Old event/experience history is deduplicated into month bundles, while user-authored long-term memory is never moved or deleted merely because it is old.
• Completed multi-step workflows are harvested only after at least two successful tool actions. The verified action sequence and final result feed the generated experience Skill so similar future tasks can route to a concrete prior workflow instead of repeating broad discovery.
• Built-in Skill routing now includes vetted Cancip adaptations of ADHD divergent exploration and Caveman response compression. High-value open decisions can fan out through isolated Cancip sub-sessions, while concise/low-token and memory-writing requests select only their matching compact Skill context instead of loading every Skill.
• Completion now requires a real user-visible final answer after process/tool work. Missing or process-only closure retries up to five times, then remains resumable instead of being stored as a false completion. Running and completed sessions broadcast disk-backed updates to every open view of the same session so background/foreground transitions preserve their correct shape.
• PrimeTTS is the single default playback route for both Chinese and English. Each sentence starts with a tiny character/word chunk for fast first audio, ramps through punctuation, then uses stable sentence-bounded chunks with look-ahead; paired source/spoken chunks keep original digits for display/highlight while applying language-aware number conversion.
• Button rules store view, command, icon, stable target identity, creation time, and modification time. Recently changed buttons sort first in settings. Menu insertion applies rules only to the newly mounted menu subtree and runs one short subtree-only trailing pass; full refreshes cache repeated scope/selector queries. Observers and timers follow each live Obsidian document/WebView, ignore unrelated body/container class churn, and react only to actual editable buttons/menu/status items/tab headers; internal DOM writes never re-enter the queue, and stable menu-group/name identity prevents adjacent-item cross-application.
• Lightweight local versioning under the Cancip data versions/ folder: manual commits and one daily auto snapshot, without native git and without per-edit history.
• Built-in local automation templates are seeded once from source and remain permanently user-editable. They include precomputed personalized new-chat greetings, evidence-backed diary assistance, review-gate generation, local version snapshots, GitHub status, vault index refresh, one combined daily Vault/memory/Skill/Wiki care task, and guarded new-file curation. News brief templates remain available for explicit user setup instead of being treated as universal defaults. Deleted defaults stay dismissed instead of being recreated on restart.
• Automations run in a dedicated focus-neutral background view: starting a scheduled or manual task never reveals or switches the user's active leaf/session. Hidden runners skip message DOM repaint and cold foreground restore/index work, session snapshots are persisted at a lower bounded frequency for recoverability, and idle runners are released after a short reuse window.
• Startup and foreground loading follow a warm/cold lifecycle: the visible shell renders first, latest-session restore follows asynchronously, small high-value indexes warm during browser idle time, and startup maintenance yields between tasks so mobile interaction remains responsive.
• New-file curation runs visibly in an isolated session with a stable minimal prompt prefix. Most nonempty user-created Markdown notes receive one bounded organization pass; explicit opt-outs, templates, frequently referenced notes, plugin syntax, and generated files are protected. Changes keep per-file action allowlists, Review Gate capture, backups, and readback verification.
• TTS is provider-routed by language. English defaults to Web Speech / system TTS and does not need a local model package. Chinese can auto-download and use the current compact PrimeTTS Chinese/English ONNX package. Other languages use system/Web/custom URL unless a compatible local PrimeTTS package is installed under tts/<package>/ with a manifest.

3.0.32

• Automation tasks now expose only a task-specific model selector. Task-level API profile settings are hidden and ignored; model source, URL, and key stay in the global model-source/model binding settings. Manual automation run buttons now give immediate started feedback, suppress duplicate started notices, and use guarded mobile tap handling so a tap reliably launches the task.

3.0.31

• Separate automation silence from notification policy. Personalized new-chat greeting refresh now migrates to silent background execution with notification policy set to Never; Silent only prevents opening or switching to the automation session, while Notify controls notices independently.

3.0.30

• Stop background universal search from chasing a complete index in repeated follow-up batches. Scheduled indexing now does one small maintenance pass per trigger, skips unchanged shard writes, and relies on on-demand local search for deeper content so desktop sync churn does not slow the mobile-first runtime.

3.0.29

• Treat mobile as the primary runtime for universal search: background indexing now keeps only the most useful recent items per category, sessions and mobile binary files index title/path first, unchanged shards are not rewritten, and full content is read only when a search actually needs it.

3.0.26

• Keep status-bar chat and review badges stable while background counts refresh. Existing blue/red attention markers stay visible until fresh disk counts are ready, then update atomically.

3.0.27

• Reduce mobile background load by avoiding repeated full universal-search rewrites, excluding machine-generated review/version/archive/config-backup data from the search index, and polling only key Cancip state files instead of large data folders.

3.0.28

• Make mobile search self-sufficient and on-demand: universal search writes small per-kind shards instead of one large document list, mobile indexing no longer waits for desktop-synced full indexes, and config/attachment content is read only when the query or user option asks for it.

3.0.22

• Add compact composer links for real changed-line totals and completed plan progress; tap them to open Review Gate or Plan.
• Personalization now prefers evidence from the last 24 hours, then falls back through 72 hours, seven days, and the latest reliable older clue with age-correct wording.
• New-file curation is visible and active for most new notes while respecting explicit opt-outs and protected plugin/template/generated content. Daily care runs at 00:00, catches up later the same day without duplicating a completed run, and inventories Skills, memory, Vault, plugins, commands, automations, and verified recipes for ongoing maintenance.
• Information retrieval stays minimal first, then searches targeted Vault notes, Skills, automations, session history, memory, plugins, commands, and experience before external web facts or model prior knowledge. Verified outcomes update, merge, downgrade, or prune stale internal assets.

3.0.21

• Compact composer status messages and clear stale insertion/generation hints as soon as the user edits the input.
• Restrict personalized greetings to files modified and sessions updated within the last 48 hours; synced old files are no longer presented as recent changes.
• Make greeting and new-file curation automations silent by default, repair first-run event races, and accept verified no-op curation decisions without retry loops.

3.0.20

• Process records now show the real sent payload instead of input-size statistics. Before the API call exists, the folded sent section contains the exact assembled system instructions, context, turn prompt, original prompt when distinct, and actual model input; after the call it prefers the exact request body.
• Ordinary model answers and execution/status text stay folded in process details. Only an explicit provider-supplied reasoning/thinking summary is expanded as readable progress; raw response text, commands, JSON, and audits remain folded.
• Note autocomplete no longer starts from editor construction, focus, file switching, or caret movement alone. A real user edit in the current focus cycle arms completion; focus loss or an empty document invalidates queued work and clears the suggestion.

3.0.19

• Automations now have a separate Silent run option. Silent tasks keep their full result in a stable dedicated session while suppressing start/completion notices and never revealing the response session.
• Session history reflects live request state immediately: running conversations spin as soon as the request starts and stop as soon as it finishes. Process records open during execution, persist raw sent/received API contents with runtime verification, survive session switches, and collapse only after the final answer.
• Sidebar and note autocomplete now require a real focused caret. Focus loss invalidates debounce, rotation, queued generations, and stale results, while focus gain schedules completion without requiring another character.
• Final completion is gated against the original request and verified tool results. Generic completion filler, avoidable questions, prose-only approval requests, read-only false completion, and unverified changes are rejected; short results stay natural and larger results use compact structure.
• Process records keep readable execution summaries expanded while commands, raw API payloads, JSON, and other technical details stay folded by default. The record footer shows only compact plan progress and changed-file line totals, with links to the full Plan and changed-file views.

3.0.18

• More now keeps Skill and automation subpanels open while their real lists load; load failures stay visible instead of leaving a permanent spinner. The attached-context editor closes reliably after Save or Cancel without removing the sidebar composer.
• Automation commands that already return a verified programmatic result no longer make a second model call. Daily memory and Vault maintenance are merged into one dedicated task, failed schedules retry after a 20-minute cooldown, dedicated sessions compact old process payloads, and a saved session file repairs a temporarily stale history index.
• Low-capability model action JSON accepts tool as an alias for type, including standalone write objects. New-file curation consumes an actionable file only after a captured target mutation; stale event paths are filtered against the pending queue, and a one-time state migration recovers recent candidates lost by the old false-success rule.
• Implementation turns suppress avoidable “provide a file / should I continue” questions, fall back to deterministic target discovery when a model still emits no action, and keep fallback final answers to the concrete result, changed target, or first exact blocker.

3.0.17

• Process details keep long raw lines horizontally scrollable by default, with stable wrap/copy controls. Mobile context-file editors are protected from accidental popup dismissal, and an empty sidebar composer no longer starts autocomplete requests or shows a loading indicator.

3.0.16

• Process tool results now use the same raw-content controls as sent/received details: long lines stay horizontally scrollable by default, with explicit wrap and copy buttons.

3.0.15

• Mobile Cancip popups now distinguish a tap from a vertical gesture: buttons execute on an unmoved pointer-up, while a swipe keeps scrolling and cannot close or activate the popup accidentally.

3.0.14

• Context-file edit dialogs are isolated from mobile button/first-touch handlers so editing an attached file no longer closes the dialog unexpectedly. Process details keep long raw lines intact for horizontal scrolling and now provide copy and wrap controls.

3.0.13

• Workbench HTML editing now writes through Obsidian's binary Vault API and verifies the resulting bytes. Preview edits carry a stable DOM selector, so text spanning nested HTML tags can fall back to an exact element writeback instead of reporting a false save failure.
• Automation status is evidence-based: missing model configuration and missing final answers fail, empty new-file scans are marked skipped, approval/review waits are marked pending, and only a stored visible result can become successful. Manual @ invocation passes following text as a one-run supplement without changing the saved task.
• The @ picker is hierarchical and fast: Session history, Skill, and Automations are the first three categories; category selection expands the matching list, supports continued filtering, and inserts stable references. Session, Skill, and automation management lists now include a Send to Cancip action.
• Normal model turns include up to six recent non-process messages without duplicating the current prompt. Implementation requests get a hard second action retry and can no longer be completed from prose-only instructions. New-file tracking no longer drops files that share a creation timestamp, rename state migrates by path, and the Cancip-owned Vault/plugin overview refreshes as the environment changes.

3.0.12

• Composer access, More, Skills, and automation menus now measure the active Obsidian window's visual viewport and open above the composer, keeping the complete menus visible above the mobile keyboard.
• Common model settings are split into three clear layers: collapsed model-source credentials, a collapsed source-bound model list, and a visible default-model selector. Automation and autocomplete inherit the default model/source unless explicitly overridden.

2.14.19

• 自动补全默认每批 2 个候选；设置为 1 时只显示一个并自然停止轮换。预加载在当前批次显示后立即后台请求，和轮换并行，轮换只切换已返回候选，不会额外触发模型请求。
• 记录真正应用的笔记补全选择，不记录单纯轮换：保存候选位置、Tab/按钮/菜单入口、是否手动切换、是否自动轮换后接受、下一层是否已预加载和脱敏短前缀。后台生成本地低置信度偏好摘要，后续补全仅在上下文相关时参考，不增加模型调用。

2.14.18

• 修复笔记补全后台分支超时后继续逐项补发、再次重试造成的长期请求和状态栏持续旋转。现在每层严格只批量请求一次：首层 3 个一次，显示后对应 9 个一次；响应不完整或超时也不会在同一层自动追加模型请求。
• 自动补全网络活动使用 12 秒共享截止时间，逻辑超时会同步结束状态栏转圈，即使底层网络 Promise 迟迟不返回也不会留下活动计数。应用候选后通过专用分支交接事件同步显示已缓存的对应 3 个，只启动这 3 个的下一层预加载，不再重新生成或替换当前批次。

2.14.17

• 笔记自动补全改为低 token 递归树：首轮只请求 3 个一级并立即显示轮换；一级显示后后台请求这 3 个一级各自的 3 个下一级，共 9 个。选择其中一个时优先立即交接已缓存的 3 个，再为这 3 个请求下一层 9 个，依次递归，未选而输入新内容则开始新的一级 3 项请求。
• 状态栏转圈只绑定真实自动补全网络请求 requestUrl 的生命周期。记忆读取、模型冷却、JSON 解析、缓存命中、Promise 等待、UI 更新和后台无网络处理均不转圈；已在生成中的分支会复用同一 Promise，完成后自动接上。

2.14.16

• 修复自动补全返回 3 个候选后被重定位逻辑错误裁成 1 个的问题，模型候选现在会完整显示并按设置轮换。
• 二级候选预取改为静默后台任务，不再让状态栏图标长时间旋转；批量响应缺少部分分支时并发补齐缺项。选择候选后会接管同一个进行中的预取 Promise，已缓存则立即显示对应下一批，仍在生成则完成后直接交接，不重复发请求，也不会把迟到结果写到已移动的光标。

2.14.15

• 笔记自动补全改为先显示、后预加载：当前层按设置生成 N 个候选后立即显示和轮换，不再等待下一级；候选出现后再用一个后台批量请求准备每项各 N 个下一级，默认先显示 3 项、随后预加载 9 项。选中候选时直接交接已就绪的对应分支，并继续后台准备再下一层。
• 自动补全设置新增预加载开关、2 至 30 秒轮换间隔和 1 至 5 个每批候选数；候选数同时控制当前层、下一级和聊天输入推荐。手动选择“上一个/下一个”后仅暂停当前批次，输入变化、应用候选或换批后恢复自动轮换。

2.14.14

• 编辑器自动补全改为真正的前瞻流水线：首批三个轮换候选只有在各自三个二级候选全部进入缓存后才显示，当前批次始终先备齐 3+9 共 12 项；选中一级后同步切换到对应三个二级候选，不再临时等待模型。
• 一级与二级改为一次紧凑的 3×3 树请求，首个响应直接缓存完整 12 项；显示下一层时再用一次批量请求准备新的九项，并按完整文件、前缀、模型和补全偏好复用进行中的 Promise。重复刷新不重复消耗请求，强制换批同样先通过 12 项就绪检查，生成期间保留现有候选。

2.14.13

• 默认开启编辑器自动补全；只要光标仍在当前活动编辑器，即使焦点短暂转移到工作台或其他 UI，也会继续准备候选。首批 3 个候选先快速返回，3×3 二级候选随后后台预取；个性化缓存刷新只更新记忆，不再取消正在返回的补全。新会话启动立即后台预热个性化问候；没有可靠资料时只显示自然的时段问候，不声称“没有信息”或猜测近况。
• DOCX、XLSX、PPTX、HTML/MHTML 和其他未被占用的常见文档默认进入文档工作台；PDF 保持 Obsidian 核心视图默认打开，也可手动送入工作台。扩展冲突逐项隔离，不再中断 Cancip 启动。XLSX 按 OOXML 工作簿关系解析工作表，兼容大小写/相对路径差异，并在关系异常时按工作表顺序回退，保留共享字符串、空列位置、公式和布尔值。

2.14.12

• Keeps recursive 3-by-3 editor completion prefetch, but makes it lighter: compact complete candidates, smaller relevant context, lower output budget, shorter debounce/cooldown, and a cached working Responses/Chat Completions route avoid repeated failed transport probes.
• Selecting a completion synchronously adopts its prefetched children at the new cursor, including multiline candidates. If recursive prefetch finishes after selection, the inserted text is matched and its children are handed off without losing the branch.
• Editor ghost text now participates in the same inline text flow as inserted text instead of moving as one atomic flex box, so wrapping and explicit newlines begin at the real cursor position and match the applied result.

2.14.11

• Mobile buttons inside the keyboard-lifted composer footer now remain owned by the Cancip view after the footer is portaled to the document body, so recommendations, menus, queue controls, model/access actions, and Send execute on the first touch instead of only dismissing the keyboard.
• The autocomplete apply/settings control preserves its long-press menu while handling a short touch on pointer-up and suppressing the delayed duplicate click.

2.14.10

• Restores document-workbench leaves in place after plugin hot reload, preserving the same tab, file, mode, pin state, and active state without detaching leaves or creating blank tabs.
• Reuses one workbench tab by default when opening different files. A dedicated Workbench settings page controls default Preview/Markdown/Edit mode, compact headers, file metadata, and single-tab reuse.
• Compact workbench layout keeps the title, mode selector, and fixed action buttons in a short two-row header on narrow sidebars; metadata becomes a single scrollable line and document content begins near the top instead of the lower half.

2.14.9

• Mobile touch now activates Cancip recommendations and conversation controls on the first press while suppressing the delayed duplicate click and preserving autocomplete long-press controls.
• Live process records use a stable turn key: untouched records open while running and collapse on completion, while a user's manual open/closed choice survives streaming DOM rebuilds and the final answer.
• Process audits are reduced to at most three independently expandable groups: exact sent source, exact received source, and compact runtime information. Redundant split prompts and visible-answer copies are omitted when raw payloads exist.
• Composer todo/plan rows are hard-bounded to the sidebar and floating visual viewport, preventing long tasks from overflowing left or right.
• Markdown completion rotates every five seconds and generates a three-option tree with three prefetched continuations per option. Applying a branch shows its cached next choices immediately and starts prefetching the following level.
• Button editing and autocomplete now have independent settings pages. Built-in automation controls use localized labels, templates choose Chinese or English fallback content from the active UI language, and user-edited tasks are never overwritten by startup migration.
• Personalized greeting refresh and diary writing assistance are visible built-in automation tasks. Their prompts, schedules, model/profile routes, notifications, and enabled state remain editable in the Cancip data automations.json through the normal automation settings UI.

2.14.4

• Moved Copy and Wrap into a fixed action layer outside horizontal code scrolling, eliminating the visible move-then-snap effect in chat and rendered notes.
• Restored live elapsed-time headlines and classified real fenced API audits into sent, received, runtime, and other details without exposing redundant progress prose.
• Settings category redraws now retain the horizontal tab-strip position as well as the visible vertical option.
• Markdown editor autocomplete now keeps one model request in flight, rebases a matching response onto continued input, retries only after a mismatch, and cancels stale work after a newline, completed sentence, block delimiter, selection change, or focus loss.

2.14.3

• Removed the precomputed diary button and file-list insertion flow. Diary-writing requests in an active daily note now use the current model with a bounded record of today's real prompts and verified tool outcomes, while normal review still governs note writes.
• Restored independently expandable Chinese process-audit fields, consecutive previous-prompt navigation, stable TTS session UI, and sentence-change source-highlight recentering.
• Automation task settings are collapsed by default and default tasks migrate to a relevant prior session when one is available. Chinese composer autocomplete now triggers earlier and preserves local suggestions when model completion fails.
• Copy and wrap controls remain fixed at the code-block viewport edge while chat and note code scroll horizontally.
• Settings redraws preserve the visible option instead of jumping the page. Add current file now creates the same removable file snapshot used by Send to Cancip, including parsed document or image context when applicable.
• The active model now provides quiet line-end completion inside Markdown editing as well as the Cancip composer, with local candidates first, bounded nearby context, stale-result rejection, and Tab or an inline icon to apply.
• Preferred name, locality, weather, and restrained caring cues are inferred from direct user-related evidence by Cancip itself. Manual name/location fields are optional corrections; ambiguous or unverified details stay absent.

2.14.2

• New chats now select a stable greeting from several background-generated, evidence-backed variants. Greetings can use an explicitly configured or reliably labeled preferred name, optionally include cached live Open-Meteo weather, and offer two or three concrete next-action buttons without delaying chat creation.
• Composer autocomplete now returns one compact JSON result containing the gray suffix, up to three input-specific action buttons, and short reusable workflow steps. The suggestions float without changing footer height, while scroll clearance keeps the final chat content reachable above the keyboard.
• Cancip records recommendation-button use locally. Repeated choices create a native Review Gate sorting proposal at three uses; priority changes only after approval. Repeated successful multi-tool workflows may become generated Skills, while one-off actions are no longer promoted.
• The core agent prompt now explicitly completes clear low-risk next steps on the user's behalf while keeping writes and high-impact operations behind the existing access mode and audit flow.

2.14.1

• Reworked outcome-capture and hidden-button styling through Obsidian DOM/CSS helpers, removing the blocking static-style review errors without changing capture behavior.
• Replaced native element creation with popout-safe Obsidian helpers, adopted the active Vault configuration directory for universal search, and added the Obsidian 1.13 settings-definition entry point while preserving the complete legacy settings UI.
• Cleaned review warnings for unsafe values, redundant assertions, regular expressions, cross-window element checks, unused code, code-block controls, and document-preview selectors.

2.14.0

• Added a single mobile-safe document workbench with Preview, Markdown, and Edit modes. Every Vault file now has Open in document workbench and Open as Markdown actions in the normal file menu.
• Added semantic DOCX conversion with headings, emphasis, tables, headers/footers and notes; XLSX conversion preserves sheet names and cell column positions; PPTX conversion keeps slide boundaries, titles, text, and speaker notes.
• Added isolated HTML/MHTML preview, MHTML embedded-media decoding, quoted CSV/TSV tables, native PDF/media preview, ZIP-based document extraction, and safe metadata/embed fallback for unsupported binaries.
• Text-family files can be edited in place with readback verification. Office, PDF, media, archive, and unknown binary originals are never rewritten; edited output is saved as a new Markdown or standalone HTML conversion.
• Added cancip.documents.help/open/convert command-bus routes and fixed the release workflow to publish normal numeric tags while keeping release assets limited to main.js, manifest.json, and styles.css.

2.13.0

• Native File Explorer pinning now keeps files and folders mixed at the top of their original folder, with inline up/down/unpin controls and no separate pinned-items panel.
• New chats use a short background-precomputed greeting from the current time, recent files, compact memory, and recent session signals; date-based notes expose an enabled-by-default personalized diary action.
• The composer now offers quiet gray context completion from local candidates first and a compact model fallback only after an idle delay. Tab or the inline check applies it, while long-press opens enable, regenerate, and guidance controls.
• Chat and Markdown-note code blocks reserve fixed non-overlapping 26px slots for Copy and Wrap across narrow and desktop layouts.

2.12.0

• Added structured post-action verification for active view/file text, DOM selectors and geometry, file contents/JSON paths, plugin versions, workspace leaf counts, layout stability, and overflow.
• Added active-region PNG evidence with non-blank pixel checks, optional one-image model review, local JSON reports, and visible evidence links in tool runs. Screenshot Base64 remains memory-only and is never stored in session JSON.
• Failed verification now continues with the same loop id and incremented attempt for measured, minimal correction only. The loop stops at its declared limit and hands the saved evidence to the user instead of retrying indefinitely.
• Added installed-exporter PDF verification and a real Obsidian regression fixture covering pass/fail/attempt-limit behavior, PNG integrity and pixels, report schema, evidence UI, image handoff, and cleanup.

2.11.0

• Process-record raw sent/received fields and inline tool results now load only when the first-level process record is opened. Folded sessions keep the full auditable source on disk without immediately laying out tens of thousands of hidden DOM characters.
• Mobile keyboard clearance now follows the actual overlay inset, so the final chat record can scroll above both the keyboard and floating composer and docks cleanly after keyboard dismissal.
• Fixed temporary hidden-button reveal restoring explicit hide rules, flattened mobile menu groups and cross-section sorting, and made native File Explorer pin tests open and restore the mobile left drawer explicitly.
• Exact-answer prompts now use one zero-context model call, explicit read-only memory/plugin questions take narrow programmatic routes, and completed tool-backed answers retain a completed session state instead of being misclassified as failed.

2.10.2

• Extended the persistent code-block wrap toggle to rendered code blocks in Markdown note reading view, Live Preview widgets, and note embeds. Non-Markdown leaves and raw editor text are untouched, and unloading Cancip restores native note DOM.
• Plugin reload now restores an existing Cancip placeholder leaf in place, preserving its sidebar location without detaching the leaf or creating another tab.

2.10.1

• Added a persistent code-block wrap button beside Copy. The default keeps code unwrapped with horizontal touch scrolling; toggling it updates all current and future Cancip code blocks.

2.10.0

• Added reversible 30-day cold archiving for inactive sessions, session events, and machine-generated experience history. Archive indexes self-recover from the cold files, and age alone never deletes data.
• Added persistent universal search with hard search first, optional model-supplied semantic expansions second, source rereads to reject Bloom false positives, secret redaction, archived-session coverage, and focused image handoff.
• Added cancip.searchAll, search/archive status and maintenance commands, cold-session restore, and lastOpenedAt session tracking.
• Automations can now combine manual/hourly/daily schedules with debounced new-file triggers, comma-separated glob filters, per-task model/profile selection, and inherit/always/failure/never notification policies. The built-in new-file curation task batches Markdown creation events for 60 seconds to avoid one model call per file.
• Completion notifications are deduplicated and reduced to compact task/result/session lines. Programmatic/background completions use the same format and still honor Obsidian/ntfy switches.
• Model replies that report length, max_output_tokens, or incomplete, or end with strongly unfinished Markdown/sentences, are continued up to two times and overlap-merged. A reply that remains incomplete is kept resumable instead of being marked complete; actionable final answers receive local recommendation-button fallbacks without a hidden model call.
• Up to three recent sessions that were still running when Obsidian or Cancip exited are recovered before scheduled automation scanning starts. Interrupted runs of the same automation are deduplicated by task id/title so restart does not launch a replacement session; manual stops and ordinary failures remain user-controlled.
• Mobile send submission now happens before keyboard blur can move the floating composer, so one tap sends instead of only dismissing the keyboard.
• File Explorer pinning now supports mixed files/folders, explicit up/down controls, drag sorting, and reliable unpinning from the normal context menu.
• Native Review now includes AI changes to .cancip/config.json, top-level Obsidian JSON config, and installed plugin main.js, manifest.json, styles.css, and data.json, while sessions, indexes, logs, caches, and Review packages remain excluded.

Build

npm install
npm run check
npm run build

Build output is written to:

outputs/cancip/

Run the normal development gate before committing:

npm run verify

Install

Copy these files into an Obsidian community plugin folder named cancip:

manifest.json
main.js
prime-tts-worker.js
styles.css
versions.json
README.md

Then enable Cancip in Obsidian.

TTS packages

Cancip supports multiple TTS routes:

• auto: picks a route by language and availability.
• web-speech: browser/WebView speech synthesis, useful for English when Android WebView exposes voices.
• android-system: native/system bridge route, used only when the Obsidian mobile environment exposes one.
• builtin-prime-tts: optional local ONNX package route.
• custom-url: trusted local/private TTS bridge route.

English is intentionally lightweight by default: auto tries Web Speech, Android/system, and custom URL before the local package. A local package is not required for ordinary English reading.

Chinese uses the current default downloadable local package when available:

tts/prime-tts/acoustic_encoder.onnx
tts/prime-tts/acoustic_decoder.onnx
tts/prime-tts/vocoder.onnx
tts/prime-tts/meta.json
tts/prime-tts/symbol_table.json
tts/prime-tts/ort/ort-wasm-simd-threaded.mjs
tts/prime-tts/ort/ort-wasm-simd-threaded.wasm

The official review-clean plugin release ships only the core plugin files. Model/WASM assets are not bundled in the core release. Cancip downloads the default package only when the local install command/button is used, or when builtin-prime-tts is selected and the package is missing.

The default downloadable package is:

https://github.com/arias007/obsidian-cancip-ai/releases/download/prime-tts/prime-tts.zip

The zip may contain the files at the archive root or inside one top-level folder such as prime-tts/ or prime-tts-zh-en/.

Compatible packages for other languages can be installed as:

tts/<package-id>/acoustic_encoder.onnx
tts/<package-id>/acoustic_decoder.onnx
tts/<package-id>/vocoder.onnx
tts/<package-id>/meta.json
tts/<package-id>/symbol_table.json
tts/<package-id>/ort/ort-wasm-simd-threaded.mjs
tts/<package-id>/ort/ort-wasm-simd-threaded.wasm
tts/<package-id>/manifest.json

Manifest example:

{
  "engine": "prime-tts",
  "id": "prime-tts-ja",
  "label": "PrimeTTS Japanese package",
  "languages": ["ja-JP", "ja"],
  "notes": "Compatible ONNX PrimeTTS package."
}

When Cancip language is Japanese, Russian, Turkish, Uyghur, Korean, Spanish, French, German, Arabic, or another supported UI language, auto first checks whether an installed compatible local package declares that language. If not, Cancip falls back to system/Web/custom URL instead of pretending the bundled Chinese package can speak that language.

Use cancip.tts.probe to inspect the current route, selected package, installed compatible packages, and fallback reason. If the folder is missing or incomplete, Cancip reports that directly.

Config

Cancip reads and writes:

<vault.configDir>/plugins/cancip/data/config.json

With the normal Obsidian config folder this is .obsidian/plugins/cancip/data/config.json. On first startup after upgrading, Cancip copies missing files from legacy .cancip/ into this data directory without overwriting files already synced from another device. The legacy directory remains as a read-only rollback backup.

On startup, the Cancip data config.json wins over plugin data.json and the settings UI. Settings changed in the UI are saved back to both places.

API settings are stored in apiProfiles; the active profile is mirrored to the legacy apiUrl, apiKey, apiMode, and model fields for compatibility.

Do not commit or share the Cancip data config.json if it contains an API key.

Session Export

Use the history icon in the chat header to reload recent conversations. Cancip saves lightweight JSON session history under:

<vault.configDir>/plugins/cancip/data/sessions/

Use the download icon to export the current conversation as a handoff/archive snapshot.

Cancip writes two files under:

AI/Cancip/Exports/

The Markdown export is for reading and handoff to another agent. The JSON export preserves structured data:

• message ids, roles, timestamps, and full message text
• per-turn system prompt and context text actually sent to the model
• source hits attached to user messages
• current mode and access mode
• active API profile name/mode/model
• draft context currently attached to the composer
• key configuration booleans such as whether an API URL/key or GitHub token is configured
• tool run status, action summaries, results, and errors

The export intentionally does not include plaintext API keys or GitHub tokens.

Command Bus

2.5.5

• Daily automations created after today's scheduled instant wait for the next scheduled day; opening Obsidian later can still catch up a genuinely missed run.
• Every model response carries hidden continue or final control metadata. Cancip no longer treats ordinary prose or a preparation plan as a final answer.
• Simple tasks may finish with a direct short answer. Complex tasks keep process updates folded, continue through actions and verification, and finish with a compact total-detail-total answer.
• Ordinary chat, tool continuation, approval continuation, and prompt/model-backed automations share the same structured completion gate. Missing or invalid control metadata is retried with a bounded correction prompt.
• Elapsed time, token usage, changed-file links, and verification facts remain programmatic; the model controls the useful visible conclusion and optional recommendations.

2.5.4

• Keeps the mobile composer and @ panel attached to the Android visual keyboard viewport.
• Expands @ discovery to installed plugins, automation instances, attachments, Skills, tools, functions, and Obsidian commands.
• Keeps the review panel above the mobile status bar and renders structure changes as a compact type plus old -> new route.
• Treats only manually created Markdown files as automatic new-file curation candidates; Cancip-created review items are excluded programmatically.
• Adds reciprocal links only for strongly related, currently unlinked notes, writes short property-only link summaries when useful, and skips clean notes that do not need curation.

Cancip treats command execution like a CLI surface inside the vault, but the executable layer is structured and auditable instead of raw string eval.

Example:

```cancip-action
{"actions":[{"type":"command","command":"obsidian.listCommands","args":{"query":"file","limit":20}}]}
```

Plan todo example:

```cancip-action
{"actions":[{"type":"todo","op":"set","items":[{"text":"Inspect relevant files"},{"text":"Apply the patch"},{"text":"Run verification"}]}]}
```

Currently supported command names:

• obsidian.listCommands: list Obsidian internal command ids from app.commands.commands.
• obsidian.resolveCommand: resolve a fuzzy command name, translated label, or id into exact Obsidian command candidates.
• obsidian.execute: execute an Obsidian command by exact id or high-confidence fuzzy name, for example {"id":"app:open-settings"} or {"query":"open command palette"}.
• obsidian.js.help, obsidian.js.probe: inspect the Obsidian JS bridge before writing glue code. The probe reports active file/view, loaded plugin ids, command count, and helper methods.
• obsidian.eval: execute explicit Obsidian app/workspace/vault/plugin JavaScript in the current WebView. Aliases obsidian.js, js.eval, javascript.eval, and browser.eval route here. Use args.code, args.script, args.js, args.body, or args.expression. Available variables include app, workspace, vault, metadataCache, activeDocument, window, args, plugins, activeFile, activeLeaf, activeView, and helpers.plugin/api/runCommand/openPath/notice/query/click/input/sleep/snapshot.
• obsidian.files.pins, obsidian.files.pin, obsidian.files.unpin, obsidian.files.movePin, obsidian.files.reorderPins: inspect and change the native File Explorer's folder-local pinned-file order by full Vault path.
• cancip.pluginCapabilities: inspect installed plugin capability routes by plugin name or feature words, including commands, runtime API surface, plugin files/settings, UI/API/config/web route hints, e.g. {"query":"notedraw 涂鸦 高亮"}.
• cancip.pluginRoute: generic plugin auto-adapter discovery for current or newly installed plugins. It summarizes commands, public API methods, settings/files, UI routes, and exact pluginAction examples.
• cancip.pluginAction: execute a plugin command or public API method after discovery, e.g. {"pluginId":"plugin-id","commandQuery":"open panel"} or {"pluginId":"plugin-id","target":"api","method":"methodName","params":[]}. Access mode controls approval/full-access execution.
• cancip.annotate.help, cancip.annotate.note, cancip.annotate.pdf: programmatic note/PDF annotation routes for NoteDraw/Pdftion-style highlighting, drawing, text, covers, exports, and active selection operations.
• cancip.study.help, cancip.study.review: spaced-repetition routes for review queue, active-note flashcards, all due cards, cram review, and note-review ratings.
• cancip.rebuildIndex: refresh Cancip's lightweight vault index.
• cancip.searchAll: search all indexed Vault/Cancip sources. Hard search always runs first; optional args include softQueries, includeArchived, includeConfigs, and limit. cancip.searchVault remains a compatibility alias to the same engine.
• cancip.search.status, cancip.search.rebuild: inspect the persistent universal-search index or rebuild it. Full rebuild is effectful and can take time on large Vaults.
• cancip.archive.status, cancip.archive.run, cancip.archive.restore: inspect cold-storage counts, run the 30-day maintenance pass, or restore a cold session by sessionId.
• cancip.reviewGate: programmatically build native Cancip audit-panel data. Example args: {"paths":["Folder/Note.md"],"maxFiles":20} or {"items":[{"path":"Note.md","old_text":"...","new_text":"..."}]}.
• cancip.reviewGate.list: list recent review data packages under AI/Cancip/Review/.
• cancip.previewVaultSearch: preview local Vault Search results.
• cancip.localVersionCommit: create a manual lightweight local version commit.
• cancip.vaultDailyReport: generate a read-only Vault maintenance and merge-candidate daily report.
• cancip.automation.templates: list built-in local automation presets.
• cancip.automation.addTemplate: add a built-in preset, e.g. {"id":"auto-review-gate-current-vault"}.
• cancip.automation.addVaultDailyReport: add or refresh the daily Vault maintenance report automation.
• cancip.automation.addVaultCuration: add or refresh the unified Vault curation automation. It keeps new/recent Markdown notes and old/specified-scope notes in separate lanes, then runs beautify/refactor, properties/tags/summaries/links, and file renaming as needed. New notes are scanned by the plugin before the model call and passed as a concrete candidate pack; specified files/folders become a strong-scope lane that the model must read and act on instead of doing vague full-vault scanning. Cancip also installs skills/vault-curation-specified-scope.skill.md under its data directory as a built-in strong Skill for explicit file/folder curation.
• todo action type: maintain the current session's visible Plan todos. Supported operations are set, add, update, remove, list, and clear.
• github.help: list mobile GitHub command targets.
• github.repo: show repository status from GitHub REST.
• github.issues: list issues, e.g. {"state":"open","limit":10}.
• github.pulls: list pull requests.
• github.releases: list releases.
• github.workflowRuns: list GitHub Actions runs.
• github.branches: list branches.
• github.file: read a repository file or directory via Contents API, e.g. {"path":"README.md"}.
• github.createIssue: create an issue with {"title":"...","body":"..."}; requires a GitHub token.

JS boundary: this is an Obsidian WebView/API bridge, not an OS shell. obsidian.js.help/probe are read-only. obsidian.eval and aliases are effect-capable and follow Cancip access mode: confirmation mode queues Approve/Reject, while full-access executes directly.

GitHub settings live in the advanced Command bus group and mirror to the Cancip data config.json:

{
  "githubApiBaseUrl": "https://api.github.com",
  "githubOwner": "arias007",
  "githubRepo": "obsidian-cancip-ai",
  "githubToken": "",
  "autoContinueAfterTools": true,
  "maxToolIterations": 3
}

Use the official API or a trusted self-owned relay; do not send GitHub tokens through public accelerators.

Cancip CLI and local Agent bridge

Desktop Cancip can connect bidirectionally with Codex, Claude Code, and other MCP-compatible local agents. Cancip remains the Obsidian interface and execution layer: it supplies Vault search/read/open tools, displays the conversation and progress, and applies its existing confirmation/full-access and Review rules to every write-like action. The local agent can be selected as Cancip's text-model brain without gaining a direct write route around Cancip.

Open Cancip settings -> Advanced -> Agent bridge and Cancip CLI, enable the bridge, then copy the generated link command. The CLI discovers the open Vault and reads its private bridge credential locally; it never prints that credential. Typical commands are:

node "/path/to/.obsidian/plugins/cancip/cli/cancip-cli.mjs" status
node "/path/to/.obsidian/plugins/cancip/cli/cancip-cli.mjs" link obsidian cancip
node "/path/to/.obsidian/plugins/cancip/cli/cancip-cli.mjs" connect codex
node "/path/to/.obsidian/plugins/cancip/cli/cancip-cli.mjs" connect claude
node "/path/to/.obsidian/plugins/cancip/cli/cancip-cli.mjs" send "Summarize the active project"

link obsidian cancip detects installed Codex/Claude Code CLIs and registers Cancip with each available agent; connect codex and connect claude provide explicit routes. Cancip also writes cli/CANCIP_AGENT_CONNECT.md beside the installed CLI so an agent can discover the exact local command after the user says “link Obsidian Cancip”. doctor reports CLI availability and real MCP registration state. Other MCP clients can start the same CLI with the mcp command.

Available Codex and Claude Code CLIs appear directly in Cancip's model selector and can be used by normal chat, automation, planning, subagents, Review, and final-answer flows. For local models, the model-source presets include Ollama (127.0.0.1:11434), LM Studio (127.0.0.1:1234), and vLLM (127.0.0.1:8000); arbitrary OpenAI-compatible local/private endpoints can be added as normal model sources. Local/private sources may omit the API key. “Refresh local models” discovers both Ollama /api/tags and OpenAI-compatible /v1/models catalogs.

The HTTP bridge listens only on 127.0.0.1, requires a random Bearer credential on every route, limits request/output sizes and request rate, and is disabled on mobile. When Cancip invokes a local Agent as its brain, Codex keeps the user's existing authentication/model provider but runs read-only with MCP and project rules disabled for that child invocation; Claude Code runs with tools disabled. This prevents recursive bridge calls or direct writes without reading or copying Agent credentials. Confirmation mode queues writes for approval; Full access executes them through Cancip and records Review data as usual.

autoContinueAfterTools controls whether completed tool runs are sent back to the model for another reasoning step. maxToolIterations caps the loop so a bad prompt cannot run forever.

Roadmap

• Product target: mobile-first AI agent for an Obsidian vault, not a whole-device remote-control assistant.
• Keep the core boundary vault-scoped: Cancip can control vault files, its plugin data config, configured project workspaces, GitHub, and plugin build/install workflows, but should not control the entire device.
• Absorb Smart Composer's Obsidian-native chat UX: file chips, current-file context, vault chat, tool visibility, and compact mobile controls.
• Keep a clear frontend interaction model: command bar, access-mode selector, project/session list, plan feature, and action transparency.
• Build the backend toward a local-first agent runtime with code actions, OpenClaw-style tool routing, and Hermes-style memory/workflow patterns.
• Support GitHub management from mobile: status, issues, branches, commits, pushes, PRs, releases, workflow results, and a safe API acceleration layer with credential redaction.
• Expand the command bus into the main backend interface: Obsidian commands, Cancip tools, plugin/skill tools, GitHub CLI-equivalent commands, and the Obsidian JS bridge should all connect to the AI through named, reviewable command actions.
• Continue expanding lightweight local versioning: restore/diff UI, retention, and GitHub sync from the Cancip data versions/ folder.
• Support Obsidian plugin building/adaptation: source-first TypeScript/CSS changes, build/package/install loops, and built-JS patching only as a fallback.

See docs/VISION.md for the detailed product boundary and architecture direction. See docs/GITHUB_AND_VERSIONING.md for the mobile GitHub and local versioning design.

2.6.8

• Review now filters Cancip's own review/export artifacts out of pending visible-note review lists, keeps the mobile review tree/detail panels scrollable above the status bar, and avoids stale internal files inflating review counts.
• Mobile chat keyboard handling uses both visual viewport shrinkage and direct footer overlap, so the composer and @ panel follow the Android keyboard more reliably.
• Session history omits created/last-activity rows from the list details, while individual chat messages show exact timestamps for task replay.
• Process records now auto-expand only for the currently running turn, fold structured raw details, and filter Cancip-generated progress filler such as generic "continue from result" messages.

2.6.9

• Internal review-gate JSON/JSONL now writes to the hidden .cancip/review-gates store by default instead of the visible AI/Cancip/Review note tree.
• Startup migrates legacy visible Cancip review JSON and session export JSON into hidden internal folders, then removes only the old internal JSON files when safe.
• Review package lookup accepts old visible paths but canonicalizes them to hidden paths, so old sessions and pending approvals keep working without recreating visible JSON.
• Session export keeps the human Markdown export visible and stores the machine JSON snapshot under .cancip/exports.

2.6.10

• Chat/session display times now show local seconds without the +08:00 suffix.
• Mobile composer footer records visual-viewport bottom/side offsets and fixes the input box above the Android keyboard while focused.
• Mobile Review Gate constrains the outer shell above the Obsidian status bar and lets the file tree, diff body, and detail rail own their vertical scrolling.

Plugin compatibility API

Plugins can register structured capabilities through app.plugins.plugins.cancip.api.v1.registerPluginAdapter(), or ship a declarative cancip-plugin.json beside their manifest. Cancip falls back through Obsidian commands, public plugin APIs, and verified visible UI actions for plugins without a dedicated adapter. See the compatibility guide and descriptor schema.