What's new in 2.5.0 - Long conversations open faster: recent messages load first, with more history added as needed. Scrolling back no longer duplicates loaded history or new messages. - Images in history load as previews first. Tap an image to fetch the original, reducing the wait on slower connections. - More reliable connections: fewer unnecessary reconnects on slow networks. When you return to the same local network, Pairlet switches back to a direct connection while idle, keeping the current connection during transfers and replies. - A new + button at the top of a conversation lets you start another session in the current project. - Fixed Follow System always using dark mode. Appearance now follows your system setting. - More reliable approval and question notifications, even when another device is online. - Model lists appear from cache and refresh in the background, reducing waits and preventing older lists from replacing newer models. Update Pairlet on your computer too for the full connection and history improvements.

What's new in 2.4.0 - Dictation: a slow connection no longer shows a failure first. Uploading and waiting are shown separately, a transcript that arrives late is still used, and retrying does not transcribe the same recording twice. - Reopening a session you just looked at shows what you already had right away and then adds what is new, so long sessions do not reload their whole history each time. - Approvals: a card ignores taps for a moment after it appears, so a double tap cannot approve the next request, and its buttons only act on the request they show. - While a photo is still being compressed the send button waits, instead of sending the message without the photo. - Settings > About now shows this device's fingerprint and your computer's, so you can check that you paired with your own computer. - Removed a few rarely used features: review requests, handing a session over to a collaborator, sharing a folder with a guest, and the unencrypted direct-connection option. Normal pairing and sessions are not affected. - After you also update Pairlet on your computer: Kimi sessions run with the model and permission mode you chose, and a device you revoked on the computer stays revoked after a restart.

What's new in 2.3.0 - The name is now Pairlet everywhere, and the icon no longer carries the Pocket label. It is the same app: pairing and data are kept, and there is nothing to reinstall or pair again. - Tapping a project now shows "Opening…" and offers a retry if the computer does not answer; projects with many sessions no longer fail to open. - When Claude asks a question, tap the card's title to collapse it to one line and read what came before; your selections are kept. The "Answer required" notice at the top now takes a single line. - The Codex model list is no longer limited to three built-in models, and it says why when only the built-in ones can be shown. - After you also update Pairlet on your computer: Kimi background tasks no longer stay "running" after they end, ZCode 3.14 sessions can send messages again, and DeepSeek sessions started on the phone show up on the computer too.

What's new in 2.2.0 - New experimental feature, Voice memo to tasks: turn it on in Settings > General, then talk through a few things. The recording is sent end-to-end encrypted to your own computer, transcribed there locally, and organised into to-dos by the Claude or Codex tool on that computer; without an organiser the transcript is kept for you to split by hand. Organising never runs anything, and each to-do is sent to the session you choose only after you review and confirm it. Before the first recording, the app explains where the data goes and asks for your consent. - The back button in the top-left corner is now the same control on every page, at the same position, and screen readers announce where it leads.

What's new in 2.1.5 - Fixed a crash when previewing PDF, Word, Excel or PowerPoint files from a session. - Tapping a document now opens it straight in the system preview. A small progress card appears only if loading takes a moment, and you can cancel it. When a file cannot be previewed, the viewer explains why. - Previews wait until the app is unlocked and in the foreground.

What's new in 2.1.4 - The agent's steps now show up the moment the first one starts, with a live line for the current step, so the reply area no longer jumps while work is in progress. Failed and not-returned steps are marked separately in the block title. - Chat layout refresh: Copy message sits on the sender line, spacing between turns and paragraphs is clearer, and the header and composer use fewer boxes. - Links in replies: tap to open, long-press for Copy link, Open file and copy path. Markdown links open their real destination. - Copy a file's path with one tap from the file viewer. - The Full access confirmation appears only the first time you start such a session.

What's new in 2.1.3 - Sessions with a very long history open reliably again instead of reconnecting in a loop. - Notifications: one alert per session, no early "complete" while background agents are still working, and alerts clear when you open the session, send a message or answer a question. - Finished Codex tool steps now fold away like Claude's, keeping the final reply in view. - Quoted passages in agent replies render as proper quotes with one-tap copy. - Fixed typing a full-width punctuation mark occasionally erasing the text before it. - When a DeepSeek Harness install is incomplete, you can repair it with one tap and your message is resent automatically.

What's new in 2.1.2 - Clearer separation between your messages and agent replies. - Conversation summaries now stay collapsed until you expand them, with an option to copy the full summary. - Context usage refreshes correctly after conversation compression. - Corrected push notification signing configuration and improved notification diagnostics. When diagnostic sharing is enabled, you can copy push diagnostics from Settings to share with support.

What's new in 2.1.1 • Expand long questions in conversation history to read the full text, then collapse them again. • Improve notification registration recovery after reconnecting or returning to the app. • Improve scrolling in HTML attachment previews. • Improve reliability when starting and handing off supported coding sessions.

What's new in 2.1.0 • Collapse finished tool steps in a conversation so the final reply stands out; failed tools, approvals and questions stay visible. • Managed session list: only sessions you created in Pairlet are listed, and sessions started elsewhere can be imported explicitly. • Completion, error and quota notifications now reach your phone even while the desktop app is online; repeats within 30 seconds are merged. • Context usage shows a clear status when no data is available yet, and history sessions from more backends report their usage. • New sessions start in your default mode, and the mode you picked is kept when switching agents. • Phone landscape opens a session in a full-width single column; the project list defaults to a flat view; the desktop sidebar folds older projects. • Project pins sync per computer, iPad multitasking keeps the session state, and image attachments are forwarded to more backends.

What's new in 2.0.0 • CC Pocket is transitioning to Pairlet. Look for CC Pairlet in the App Store and on your device during the transition, with a one-time notice explaining the change. • Updated website and help links in the app for easier access to guides and support. • Added a usage and diagnostics sharing setting. When a diagnostic record is available, you can copy its ID in Settings to help report a problem. • Fixed conversation text sometimes appearing as an oversized heading, improving readability.

What's new in 1.9.8 • DeepSeek Harness sessions work again with the latest dsh: its 0.1.2 release replaced the local API cc-pocket was talking to, so sessions failed with "could not reach the dsh local API". The daemon now drives dsh over its official ACP interface — update dsh to 0.1.2-rc.1 or newer. • DeepSeek Harness no longer offers agent presets or session rename; the official automation interface exposes only model and reasoning effort. • ZCode 3.9 and newer open again: the model is now passed to the runtime directly instead of through a config file it stopped reading. • HTML files preview inline in a sandboxed browser view, with the source a tap away. • Images in the file viewer zoom: pinch, pan within bounds, and fit back to the page. • Desktop: Markdown documents open in a preview pane beside the conversation. • iPad: a wider session list so "New session" no longer truncates, and the quota bar drops its reset time when the row is too narrow to hold it.

What's new in 1.9.7 • Screenshots and images returned by tools now show up on the phone: tap a thumbnail to view full screen with pinch-to-zoom, live and when reopening a session. • Codex accounts show their remaining weekly allowance next to Claude's, on the phone and in the desktop sidebar. • DeepSeek Harness sessions can pick an agent preset and a model when starting, and switch models mid-session. • A separate Thinking switch next to reasoning effort — turn it off for gateways whose models reject thinking content. • iPad and Android tablets get a two-pane layout: session list and conversation side by side. • macOS gets a full application menu bar; Settings and shortcuts are reachable even before connecting. • Fixed the back button occasionally bouncing back into the session list, and gateway model lists for plain-http gateways.

What's new in 1.9.6 • The desktop window can now show several conversations side by side, each with its own input, and the session sidebar can be collapsed to give them the room. • Clicking a session that is already open in a column now focuses it instead of leaving the main area stuck on "Opening…". • A column keeps working after the daemon restarts, and a question raised in a column can now be answered right there. • Buttons inside a column — stop, approvals, "always allow" — now act on that column's conversation instead of the focused one.

What's new in 1.9.5 • DeepSeek Harness sessions now show the model that actually answered and their real context usage, instead of a placeholder "default". • A question nobody answered is now labelled as such in the transcript, and a question the app cannot answer says so instead of leaving the turn waiting forever. • Windows tray: the icon now has a right-click menu, and its flyout keeps proper rounded corners with a clearer primary action. • The usage page now names which agent its cache-hit rate covers and prints the token counts the percentage came from. • The subscription allowance strip no longer disappears after a few hours without running Claude.

What's new in 1.9.4 • Codex resumed and branched sessions now follow the latest rollout file, so turns added after branching remain visible when the session is reopened. • ZCode no longer shows internal compact summaries and reminders as user messages, while genuine prompts remain intact. • Agent and session default mode choices in Settings now share the same grouped-card visual structure.

What's new in 1.9.3 — everything from 1.9.2, plus Windows companion fixes • Claude plan allowance stays visible on phone and desktop, with 5-hour and 7-day usage, exact reset times, and a detailed refreshable view. • Codex sessions are found and resumed more reliably across macOS, Windows and Linux, with native fork, compact and review controls plus clearer failures instead of silent hangs. • Worktrees now show an explicit creation receipt, can open a session in the new checkout immediately, and keep the main checkout first with newer linked worktrees easy to find. • Pairing-day reconnects, duplicate resumed history, sheet gestures and bottom-edge controls are more stable and predictable. • The Windows desktop companion now renders collaboration QR codes reliably, checks for updates correctly, restores dsh tool steps in history, and adds a draggable scrollbar to wide code blocks.

What's new in 1.9.0 — Git, worktrees, rewind, and dsh answers • Git panel: finish your review on the phone — status, stage, commit, push, pull and branch switching, with two-step confirms for anything destructive. • Worktrees: see every checkout of a repo, create one for a new or existing branch, and clean them up safely (live sessions are protected). • Rewind & fork: long-press one of your messages to retry it with different wording, or fork a parallel session from that point. Files are never rolled back — only the conversation. • dsh sessions: the agent's questions and permission requests now show up on the phone and can be answered there; unanswered requests time out safely instead of hanging the turn.

What's new in 1.8.1 New • A guided first run: install cc-pocket on your computer with per-platform steps, then pair — no more guessing • DeepSeek usage now appears in usage stats alongside every other backend Improved • Pairing failures explain what happened and what to do next — a stale code, network trouble, or a link that isn't a pairing link each get their own guidance • The setup guide and support chat are reachable before the first pairing succeeds Fixed • All agent-scoped requests respect the computer's supported backends when connected to an older daemon

What's new in 1.8.0 New • DeepSeek joins Claude, Codex, OpenCode, Kimi, and ZCode as a session backend • Start a task directly from the project list with a clearer, cross-device composer • Usage can now be filtered by agent, including Kimi and ZCode Improved • Prompt images remain visible when a session is replayed • Project loading, recent sessions, voice input, and prompt delivery provide clearer progress and recovery states • Feishu routing, threaded replies, and command controls are more reliable Fixed • Agent capability downgrades no longer corrupt persisted defaults or send unsupported frames • Relay and LAN gateway handling is hardened against malformed or spoofed connections

What's new in 1.7.7 New • ZCode joins Claude, Codex, Kimi, and OpenCode as a session backend • Choose a default model and reasoning effort for each supported agent • Feishu bridge sessions now understand their group, sender, and visible-context limits Improved • Voice input stays available with an existing draft and while a turn is running, then appends the transcript for review • Recently completed sessions remain easy to find, and switching through All Projects no longer wedges a reopened chat • Feishu reviewed and full-auto modes use clearer, request-scoped approval rules Fixed • Windows daemon updates and bridge gateway authentication are more reliable • Malformed connections and concurrent device-limit races no longer disturb healthy sessions

What's new in 1.7.6 New • Codex sessions now accept image attachments — send error screenshots or design mocks to Codex just like Claude Redesigned • Settings groups its five categories into one clear container, separated from the utility destinations • Bridge cards split identity, facts, and actions into zones; action buttons wrap as whole controls instead of squeezing, so labels never break vertically on narrow screens or at 200% text Improved • Consecutive tool calls in Chat share one source label instead of repeating it between every card • A streaming turn no longer pins a redundant Running banner — the composer note and Stop control already say it • Session lists stop printing an untitled session's first prompt twice; the Bridges screen names the computer it controls Fixed • Feishu group sessions failed to start on Windows daemons — clean-room launch arguments now survive the Windows shell

What's new in 1.7.5 Redesigned • Mobile UI 2.0 rebuilds Computers, Projects, Sessions, Chat, Approvals, Reviews, and Settings around clearer hierarchy and truthful status • Chat now uses full-width conversation entries, grouped context, persistent running/queue/upload notes, and 48 pt interaction targets • Quick Actions groups session settings, tools, and context; Handoff is an ordinary peer action without promotional treatment Improved • Long session titles, paths, and tool commands wrap without horizontal clipping, while critical controls remain reachable at 200% text • Desktop session running indicators stay fresh as background work starts and stops Fixed • Kimi ACP conversations preserve replay, tool cards, mid-turn queued messages, approvals, and prompt settlement more reliably • Session info and the composer remain pinned instead of making the entire Chat screen scroll • Tool payloads keep their compact two-line preview and expand in full on tap

What's new in 1.7.0 Added • Full-auto mode can now expire after a chosen duration, falling back to ask-first instead of staying on forever • Kimi Code CLI backend (preview): open and drive Kimi sessions from your phone • New global accent color preference (CC Pocket / Codex) • Usage dashboard now counts OpenCode session tokens Improved • Voice input lands in the composer for confirmation instead of sending immediately • Faster, more reliable downloads and self-updates in mainland China via a new mirror Fixed • Background session announcements no longer hijack the session view you are reading • Gateway server errors are attributed by evidence instead of suggesting /clear

What's new in 1.6.0 Added • Session handoff: a session running on one device can be handed to another and continued there, with context and approval state carried over • Collaborator links: invite someone into a specific session with a one-time credential — read-only or allowed to drive — and the offer lands directly on their device • Approvals, upgraded: permission prompts can wait for you indefinitely ("keep waiting for me") instead of auto-denying after 10 minutes, and every approval lands in one reviewable history • Archivable sessions: tuck away sessions you rarely open; browse and unarchive them all from one place • Project entry points: pin projects, start a new session right from the recents list, and pick the model when creating one • Version visibility: check the daemon's version, get update reminders on the phone, one-tap update on desktop/Android • Feishu groups gain Smart Review (/review), between per-request approval and full trust: an AI screens each request against the group's declared purpose first — clearly low-risk ones run directly (still confined to in-project tools), everything else still asks you. Off by default; enable it on the bridge first, /trust-status to inspect