Changelog

Added

• LAMPO_OPENAI_APPS_CHALLENGE: the token ChatGPT's app directory checks at /.well-known/openai-apps-challenge when you list your server's MCP connector there.
• Files for every project, folder and the House. A Files tab beside a project's or folder's Videos and Playbook (the House's in Settings → Files) holds the material the work is made from: footage, music, fonts, logos, After Effects, Cinema 4D and Photoshop files, LUTs, briefs. Drop files and folders anywhere on the tab, or use Add files: before a byte moves, one sheet says what will happen — how much, what is already in Lampo (nothing to upload for it), which files become a new version of one already there, what is left out (.DS_Store, ._ files, .git/, node_modules/ …), and whether it fits the plan, with the ways out when it doesn't. Uploads go on in the upload tray, grouped by folder, resumable, and a file someone changed while yours was on its way never overwrites theirs: the tray says who changed it and when, and offers Keep both or Replace; a file you made a day's versions of already says when it takes the next, and offers Save as a copy. Every file keeps its versions — who made each, an agent's marked as such — and any of them comes back with Restore; trashed files wait up to 30 days in the trash (each says when it goes) and don't count toward the plan, nor do older versions. Open a file to look at it (pictures, video, sound, PDF and text right in the browser), download it, rename or move it (into another project's files too), or Copy for an agent: the one lampo files pull line that fetches exactly those files. A folder shows what it inherits from its project and the House, folded. Search, the kinds and the order narrow the list in place; ↑↓, ↵, Space and ⌫ (with Undo) work on it. Settings → Billing and the plan's sheet say how much of the storage is videos and how much files. Reviewers and review links never see files.

Changed

• Safe zones and the phone view are two separate choices. The safe-zones button (G) draws only the zones, the phone view (V) only the phone and an app's interface around the picture, and each button names its own choice; pick both to see an app's buttons and caption inside its zones. What you had set keeps showing as it did.
• MCP tools now say which of them change things: reading tools are marked read-only, so clients can let them run without asking, and the three that replace or move what is there (filing a video elsewhere, saying a version's source anew, drafting over a post) are marked destructive, so clients ask before each one. Every tool has a title.
• At a plan's limit an agent hears a plain sentence — what can't be done, and that a person decides in Lampo — instead of the plan's own words; people in the app still see the plan that would fit. API tokens get the same, with the refusal's reason and numbers kept.
• Chat apps (Claude, ChatGPT) connected over MCP are told the loop without commands to run: upload answers hand them the URL to PUT, with the app's page as the way out.
• An app that connects over OAuth is asked to allow what Lampo's MCP tools use (reading, notes, acting on feedback, drafting posts), no longer the project files' scopes, which no tool uses yet; an app may still ask for them.
• The review card loads its frame itself, so show_review's answer stays small for the model (one downscaled frame), and the card declares that it loads nothing from anywhere.
• The command is now lampo. lampo login, lampo render, lampo push, lampo watch, lampo open, lampo mcp config <client> and every other command, lampo-mcp for the stdio MCP server; settings are LAMPO_*, the MCP resources lampo://inbox and lampo://review/<slug>, and the npm package @lampo-vr/lampo. Nothing you set up before breaks: vr and vr-mcp keep working and print the same (only a person typing vr in a terminal is told the new name), VR_* settings are still read wherever the LAMPO_ one is unset, the vr:// addresses still answer, and a login saved by vr login is still read (lampo logout forgets it too and asks its own server to revoke its token; a server that doesn't confirm within 5 s is named, so you can revoke the token there). To switch: npm run link once more, also on a machine linked before (it puts lampo beside vr in ~/.local/bin), lampo mcp config <client> for a new MCP setup (it starts bin/lampo-mcp and reads $LAMPO_TOKEN), and rename VR_ to LAMPO_ in your env files when convenient. The login moves to ~/.config/lampo/ and downloads to ~/.cache/lampo/; the store stays where it is (~/.video-review, data/).
• Settings in both spellings. LAMPO_SERVER and LAMPO_TOKEN are taken as a pair, both LAMPO_ when either is set, else both VR_, so one server's token never goes to another. A server that finds a setting in both spellings with different values says so once at start, by name only; an empty LAMPO_ setting, which leaves the VR_ one in force, is named too. The Docker setup takes the domain in either spelling (LAMPO_DOMAIN, or VR_DOMAIN from an .env from before; the proxy says so when neither is set) and hands LAMPO_MEDIA_DOMAIN (or VR_MEDIA_DOMAIN) to the proxy, so the config for running behind a CDN needs no extra line in docker-compose.yml.
• Agents named by their MCP client (claude-code · Mia) read as the agent's name (Claude Code · Mia) on the Connect page, in the agent menus, Settings → Connected agents and Insights. A name someone gave (codex-cuts) stays as it was given, and Insights shows an agent's kind beside its name only where the name doesn't say it already.

Fixed

• ChatGPT can sign in with its own client metadata document: a client that may be public (it lists none among its token methods) is no longer refused because it prefers another one.
• Send all on a busy connection no longer makes the notes not sent yet vanish for a moment before they leave: they stay where they are, sending, until they go with their motion (and so does a draft sent on its own).
• Pictures show in a file's preview on servers with their own media host, and a PDF opens in a tab of its own (with Download beside it) instead of a frame that stayed blank.
• A big drop of files no longer fails part way: their uploads are prepared a few at a time as the files go, one that waited too long is prepared again by itself, and a busy server makes them wait and go on instead of saying "no more versions today".
• Signing out stops the uploads under way and forgets them, and what the browser kept to resume them; the next person in that browser starts with an empty tray.
• A file name with ! or a quote in it no longer breaks the line Copy for an agent hands out.
• A link to a file that isn't there any more says so in the file's place, with the way back to the files.
• lampo login <server> signs in with a token from the environment only when no other server is named beside it (LAMPO_SERVER with LAMPO_TOKEN): a token one server issued is never sent to another.
• Settings → Users names the server's own command for a lost password only to whoever runs the server (on the machine itself, or a hosted server's operator), not to someone signed in to a machine's app over the network.
• A failed publishing kit says why in a sentence in every list of posts too, not only on its own page.
• A reference a visitor adds with words through one review link no longer shows on another link that shows only its own visitors' notes.
• An agent connected over MCP can't be spoken for by anyone else's check-in, however its id is spelled and however long it has been quiet.
• An agent that waited for one video's notes and then waits again picks up after the last note it was handed, instead of skipping past everything it looked at.
• Sent notes leave the "Not sent yet" area in one motion: the area no longer blinks out and back while they go.
• One person's many versions of a project file no longer hold up anyone else. Each account has its own daily share of new versions of a file, and saving as a copy always works. An upload that can't become a version any more is refused before its bytes are sent, saying when to try again.
• On a server reached over https, a browser is signed in only by the session cookie no other site can set: the older cookie name is no longer read there, so a browser that still holds one signs in once more.
• When a server's storage refuses a profile picture (or another account change fails on the server's side), the person gets a sentence with a reference, never the storage's own answer.
• Routes a server's billing module adds are held to the role they declare for every kind of request.
• Settings → Billing's "talk to us" about a Business contract opens a message to Lampo's address, as the limit sheet's does; before, it opened an empty one.
• Settings → Users names the server's own command for a lost password only to whoever runs the server, not to every workspace's owners and admins.
• A publishing kit that couldn't be made tells the team so in a sentence: what went wrong inside (the encoder's own output, file paths on the server) is kept for whoever runs the server.
• Starting an upload when the plan can't be checked answers with a sentence and a reference, not the server's own error.
• A review link on the sample video no longer counts as the workspace's first link, and its opening no longer takes the place of the first link a visitor opens.
• On a server with several workspaces, a workspace's billing emails and an agent's link to a post open that workspace, not whichever one the browser used last.
• A picture or clip a visitor was about to add through a review link isn't taken once the link has a new password: they open the link again with it.
• Billing's refusals (a plan you can't switch to, a card that can't be removed, too many tries) read in German for German users too, in the checkout, paying, cancelling and the plan picker.
• Signing out tells the browser to drop its cached copies of the account's videos and pictures too, not only the session.
• Renaming a workspace and removing or testing a webhook are a person's, like adding one: an API token is refused.
• One video whose saved state can't be read (cut short by a crash, edited by hand) no longer takes the whole library, search, status and vr ls down with it: the others show, and the library says one is left out.
• Moving a store: an import stopped the hard way right after a video went in no longer loses that video's history; running the import again brings it.
• Someone removed from a workspace, or whose token is revoked, stops receiving its live updates at once, not up to half a minute later.
• An agent's check-in speaks only for the account that sends it: nobody can make someone else's agent look like it is listening, or list an agent under another person's name.
• vr help names repair-folders among the admin commands that take --workspace, no longer says vr sessions lists only this machine's agents, and vr mcp config names its --json.
• Going live behind a CDN: deploy/Caddyfile.cdn is a working proxy config for it (only the CDN reaches the app, the visitor's own address, a media host of its own), linked from the go-live guide.
• The relaxed window policy an app's sign-in popup needs stays with the screen where you allow the app: any other screen reached from there loads with the usual protection.
• A review link that shows only the newest version no longer shows the screenshots of notes made on older versions, and no review link hands out a note's unmarked frame.
• A reply a visitor writes through one review link no longer shows on another link that shows only its own visitors' notes: each link's visitors read their own replies and the team's.
• When a session ends elsewhere (it ran out, a sign-out everywhere, a new password), the browser that held it stops getting the account's notifications too, not only when you sign out there.
• An agent waiting for one video's notes no longer misses a note written in the same second as one on another video, nor a request made in the second the video was assigned to it.
• A note about the whole video reads as one everywhere an agent reads notes ("Copy for an agent", INBOX.md, review.md, vr open), not as a note on frame 0 with screenshots it doesn't have.
• A server whose publishing endpoints setting can't be read says so in one line when it starts, before it changes anything, instead of stopping with a stack trace.
• Settings → Billing thanks you for a payment only in the browser tab that made it; a link that merely names a payment shows nothing.
• Footage search: its model's download can no longer hang or grow past its files, a damaged setting keeps it off instead of turning it back on, and a video it couldn't take its turn for is indexed later instead of never.
• The plan this browser keeps between visits holds no card details.
• Taking an invite counts against the same per-network limits as signing in, for IPv6 networks too.
• A video holds only a few pieces of work that no agent has begun yet: sending another stops the one that waited longest (its notes stay open).
• Chat webhooks (Slack, Discord) name someone on a review link "via review link" instead of "client", write versions as V2, and post to Discord as Lampo.
• Over HTTP, attach_preview and attach_reference say the inline limit that holds there (about 700 KB, the request is at most 1 MB; bigger files go through an upload URL) instead of 8 MB, which only holds over stdio.
• lampo help lists render --verbose and admin import --dry-run, --people and --no-derive.

Added

• Project files, in the API (the agents' tools come next). Footage, music, fonts and project files kept per House, project and folder, so a team and its agents work from the same material on any machine: a folder sees its own files and everything above it, paths stay as uploaded, each workspace keeps the same bytes once. Every change is a version with who made it (and which agent); a push that would replace a version it didn't see is refused or kept beside it as a copy; the trash and replaced versions are kept 30 days. Files count toward the plan's storage once per workspace, the trash doesn't. Reviewers and review links don't see files; apps need the new files:read or files:write scope. /api/files… (docs/files.md, docs/api.md).
• An agent's work on a video is kept, from start to end. When you send notes to the video's agent (Send, Ask, a nudge, your answer to its question, Try again), Lampo keeps what it does with them as one piece of work: the notes you sent as its plan (which it is on, which it fixed, asked about or left), what it is doing now, how long it worked (not counting the time it waited for you), whether it needs you, and the version it handed back. It costs the agent nothing: it comes from the calls it makes anyway. A version an agent made says so, notes you send while it works join the same work (the agent hears of them with its next answer), and an agent that goes quiet shows as not heard from instead of quietly reading as idle. Only team members who may work with agents start one; reviewers' notes and review links never do. For tools: GET /api/runs?slug=, GET /api/runs/:id, each video's run in the library, the live event run, and AGENT RUN … lines in vr watch --all (docs/api.md, docs/agents.md).
• See what your agent is doing, while it does it. Wherever a video has an agent, one line under the notes panel's head (above the controls on a phone) says what is happening: "Claude Code · fixing 3 of 6 · editing Logo.tsx", "rendering V4 · 42 % · about 1 min left" with a thin edge that fills, "needs you · a question" with Answer, "V4 is ready · 5 fixed · 1 asked · 12 min" with Check fixes, a failure in its own words with Log and Try again, "No word from Claude Code for 22 min" with Nudge. It keeps its place in every state, so nothing on the page moves. Pressing it opens the new Agent view beside Notes and Transcript: your notes as the plan (next, on it, fixed, asked you), what it is doing now with what it said to you, quoted, the steps it took, how it ended and what came before, and Tell it… to send it a few words. Stop is at once. Reviewers see all of it, without the controls.
• When an agent needs you, the Inbox says so, and so does your phone. Right after agents' questions, the Inbox now lists work that failed — in its own words, with the last lines the tool printed and Try again — and work that waits for a permission it doesn't have, with the exact rule to add to the agent's settings, Copy and Send again (Lampo never allows anything itself). An agent you haven't heard from in a while, or notes sent and never picked up, appear under Stalled with Nudge and Stop. None of it needs a "Got it": a failure leaves once you opened it or tried again, a permission once the agent goes on or is stopped. In the player, "How to allow it" opens the same rule. Only team members who may work with agents see these. Notifications gain Agents that stop or wait (on): "Claude Code stopped — the render of promo.mp4 failed", "Claude Code is waiting for your OK to render promo.mp4"; and Agents gone quiet (off until you turn it on), after 30 minutes without a word. Starting, rendering and progress never ping you, and the app's badge counts what needs you.
• Stop reaches agents that listen. When you stop an agent that works through Lampo's tools, its next call to Lampo tells it to stop, once — "The person stopped this work on promo.mp4: stop now, render nothing, mark nothing, and say you stopped." — and until then the line says "Stopped · it will notice at its next step". An agent Lampo started on your machine is asked to finish its turn first (as Ctrl-C would), then stopped for good if it doesn't.
• Your notes say where the agent is with them: a quiet line under each note it was sent — "on it", "fixed · in V4", "asked you" — and the timeline shows the note in hand as an hourglass.
• Who made each version: the version picker says "Claude Code · 12 min · 5 fixed · 1 asked", from which of your notes, and Steps opens what it did; a version still rendering shows there as "V4 · rendering 42 %" until it lands.
• Archive a project. A project that is done can be put away from its ⋯ menu (owners and admins), with Undo: it leaves the sidebar, All videos, the board, Recent, the Inbox and Insights' lists of what waits now, and the sidebar's Archived row (with how many) opens the archived projects; ⌘K finds them under Archived. Opened, an archived project and its videos say Archived · Restore where Share or the next step stands, and are read only: you watch, read and download, but nothing new goes in — no version, note, reply, decision, playbook change, post, move into it or review link — until it is restored, in one click and as it was. Owners and admins can still take a video out, in the app. Its review links play watch only meanwhile and take notes and decisions again once it is back; embeds keep playing. Agents get one sentence for any write into it, and list_videos, list_folders, vr ls and vr folders leave it out unless asked (archived). Storage counts as before.
• Suggestions waiting in a playbook further down are no longer out of sight. A project's playbook (and the House's) says when agents' suggestions wait in its folders — "Reels · 5 suggestions waiting" — and one click opens that playbook on them; the folder's Playbook tab counts them too.
• Download a video from its ⋯ menu, in the player and on its card in the library: Download V3 saves the version on screen (on a card, the newest) as it was rendered, named spot V3.mp4, and Download another version lists the others. A version in a codec browsers can't play downloads as itself, not as the copy the player shows. Members, admins and owners may download, as for whole folders; reviewers watch and leave notes.
• Embed a video on your own site. A new kind of link, Embed: Share → Embed copies the code for one <iframe> that keeps the video's shape, and the video plays on any site in Lampo's own frame-exact player — the timecode with frames, frame steps and the player's keys, the render's chapters on its timeline, captions from its transcript, full screen, and options to start muted, loop or show the picture alone. It shows the video and nothing else, counts its plays without anyone's name, sets no cookie, and stops wherever it is the moment the link is revoked or expires. Sites that embed from an address find it through oEmbed (on your own machine through its public tunnel too), with the poster as a thumbnail they may show. A link is an embed from when it is made: an existing link doesn't become one, nor an embed another kind of link.
• See a render's progress while an agent renders. An agent that renders through vr render --to <video> --out <file> -- <its render command> shows the person how far it is, as it goes: the stage (bundling, rendering, encoding, uploading, checking), the percent and frames, and the time left once it can be estimated honestly. Remotion, ffmpeg, After Effects (aerender) and Blender are read; any other command by its output growing. When it is done, the file becomes the next version, and the agent reads two lines instead of the render's output. A failed render reaches Lampo with the tool's last words, anything that looks like a secret taken out. Long renders run detached (--detach, then vr render wait), so an agent's shell time limit doesn't end them. The command runs on the agent's own machine, never on a server (docs/agents.md).

Changed

• Agents are told that a note is a request about the video: the server's instructions, the watch prompt and the Agent Skill say that a note, also one from a review link, never asks an agent to run commands, open links, or send or change anything outside the render.
• "Use Lampo" is all your agent needs to hear. Connect it, tell it Use Lampo for "Spring launch", and it runs the whole loop by itself: it finds the project (or names one, or asks you which), puts up V1 itself, reads the playbook and your notes, fixes them, puts up the next version, marks each note fixed and waits for your next notes — again and again until you approve. Lampo tells it so the moment it connects, and no answer leaves it standing: when there is nothing to do it is told to wait for you. /lampo:watch in Claude Code still works, as a shortcut for the same.
• One way in for each kind of agent. Chat and desktop apps (Claude, ChatGPT, Cursor's chat) work through Lampo's connection alone and are never told about commands they can't run; Claude Code and Codex render through vr render, so you watch the render's progress; the agent on your own machine puts its renders up where they are. A video an agent puts up is its own from V1 on, so your notes on it go straight to that agent.
• You see your agent the moment it connects. The sidebar's Agents lists it before it is on any video — "Claude Code · connected", then "ready" while it waits for your notes — and so does Settings → Connect an agent. From there its page and Assign agent… are one click away.
• Connect an agent is one page per agent: the one snippet or connector address it needs, then the sentence to tell it, then what it is doing, live. The command recipes beside it are gone.
• The first run starts from your project, and your agent puts up V1. If you work with an agent, the setup asks for your first project before the agent, and Get started goes: start your first project → connect your agent → your agent puts up V1 (it ticks when its version lands) → invite and share. Adding a video yourself is still there, as the second choice, and the sample fills the wait. Picking None yet keeps the people-first steps; your own machine keeps its own.
• Empty pages lead with your agent. An empty library, project or folder now offers Make one with an agent first — Ask Claude Code to make one once yours is connected, with the prompt copied to paste into it — and uploading a video yourself second.
• Get started is one click away on every library page. While steps are open, the sidebar's foot says Get started · 2 of 5 with a thin line of the steps. A click opens them right there: connect your agent, make a review link, invite someone or link a file without leaving the page you're on; the sample and uploading take you where they live. A step ticks the moment it's done, and when the last one is, it says You're set and goes. Closing the big card now keeps the row; Hide for good puts both away, with Undo, and the account menu still brings them back. On phones and tablets the row sits at the foot of the menu.
• Telling an agent something while it works joins what it is doing: your words become part of that work, beside the notes it is working through, instead of starting something new. It hears them as before.
• A Claude Code run Lampo starts on your machine now stops after 30 minutes without a sign of the agent (its output, or a call to Lampo), and after 3 hours at most, instead of after 30 minutes whatever it was doing: a long render no longer cuts it off.
• The library says what agents are doing in one line. A card on the board, in the grid or in the list says the agent's work in the same words as the player — rendering with its percentage and a thin edge along the poster's foot, failed, needs you with Answer where the next step stands — instead of the status chip with its spinner; the sidebar's agents say where they stand in a word or two, and Being fixed counts the agents at work.
• Several suggestions for the same part of a playbook (the brief, the rules, one skill) stand together, the newest first and marked, with a line that says how they relate. Once one is accepted, or someone changed that part by hand, the others say so straight away and their diff shows what accepting them now would replace; Accept anyway does it on purpose, where before they could only be rejected. The inbox's preview says the same.
• The skill dialog is a place to write. The instructions start twelve lines high and grow with the text until the dialog is full, then scroll inside while the name, the files and the actions stay put; Write and Preview are one box, so switching moves nothing; on a phone the dialog takes the whole screen. Import and Copy sit together under SKILL.md. An open brief or rules editor starts with room for a paragraph.
• The main button in the note composer is always the one on the right. On a video with an agent, Save sits on the right as the main action (⌘↵) and Send on its left (⇧⌘↵); without an agent it stays Save, then Send.

Fixed

• A project file's recent versions stay restorable. New versions, by anyone, never drop a version replaced the same day; a file's new versions per day are bounded and the next is refused with when it may come. A file deleted close to the plan's limit can be restored: what it already counted is no longer counted twice.
• What you just trashed stays restorable. A file just trashed, a folder just deleted or a version just replaced is never taken early, even when the trash holds more than its share of the plan; meanwhile what is over counts toward the plan, and each trashed file says when it really goes.
• What an agent says in its own words no longer gets cut off where there is no room for it. A status like "rendering v2 (notes and b-roll from the content folder, then the end card)" used to stand, cut short, in the player's agent button. The button, a poster's chip in the grid and the sidebar's Agents rows now say where the agent stands in a word or two, with how far beside it: "fixing 3 of 6", "rendering V2 · 42%", "needs you". The run strip and the board's card line still show the agent's words after that, cut with an ellipsis; hovering shows them whole, and a click opens the Agent view, whose Now shows them whole and wrapped (up to the 200 characters a status keeps, where before it stopped at 80), under "working" rather than "nothing yet" when they are all the agent has done so far.
• The sample video takes no versions but its own: "Upload new version…" is gone from it, an upload aimed at it is refused, and a file named like it becomes a video of its own. Your renders always count as your workspace's, and "Remove the sample" can't take one of them with it: a sample that holds one stays.
• Safe zones look the same for every app. Instagram Reels drew its top bar, icons and caption as dashed boxes while TikTok, Shorts and Stories used red stripes for the same thing; now every app's interface is red stripes (keep text and logos out), and dashed lines are only margins such as title and action safe.
• Project files' trash and older versions are bounded on every server. What the trash and replaced versions keep is held to its share of the plan, and to what the workspace's files count when no plan says; past it the oldest goes at once. Files uploaded and not yet committed count toward the plan until they are. Bringing a file back from the trash or an older version is checked against the plan like an upload.
• A file keeps a bounded number of older versions, and bringing back the bytes it already has makes no new version.
• An agent's work kept with a video now stays small whatever an agent sends: a video and each person's agents hold only so many pieces of work at once, and what is kept of finished work is bounded, so the server stays quick.
• An agent's work is its account's: another person's agent can no longer take over, join or be stopped in place of it by naming itself like it, and an account's name can no longer contain "·".
• Reviewers see where an agent's work stands, but no longer what it touched in the project: its steps, commands and the last lines of a failed render are for people who work with agents.
• Stopping an agent this computer started is done from this computer itself, as before for its own agent controls.
• Your data export now includes what you asked of agents.
• A phone hears fewer agent pings in a row: work that fails again and again is said once for a while, and a device gets only a few pings a minute.
• A failed render's last lines leave out more kinds of keys and passwords before they reach Lampo.
• The operator's Accounts page shows when each person was last active, not only their last sign-in, which accounts signed in before it was recorded never had.
• The app no longer opens on a blank page when the browser can't keep a copy of its code (its storage full or busy, a private window): what arrived from the server is used either way, and the copy is kept when it can be.
• Setting up a server from a phone: the command that creates the owner (and the one that sets a new password) breaks at a space when the screen is too narrow for it, instead of running past the edge of its box.
• Checking a video or a download without fetching it (what browsers and download managers ask before they start) no longer makes the server read the whole file first: the player, downloads, review links, embeds, the publishing kit and your data export answer with the details alone.
• A publishing kit's ZIP or your data export cut short (a closed tab, a lost connection) now ends on the server too: before, it went on waiting for the viewer for good, with the kit's files kept open.
• Video playback, downloads and review-link files that were cut short (a seek, a closed tab) no longer leave the file open on the server; and a version's file says when it was made, so a browser can resume a download.
• A video whose file name held a broken character no longer breaks the library for the whole workspace: names are made well-formed when they come in.
• The same holds for folders: a folder name with a broken character, or a long one cut in the middle of an emoji, no longer breaks the inbox, search, For you and the library. Everything people and agents send in — folder and link names, display names, notes — comes in well-formed, names are shortened between characters, never through one, and a store that already holds such a name reads it whole.
• The skill dialog's instructions field was one line high until you typed.
• A playbook opened on its suggestions (the inbox's Open the playbook) put the focus on Reject, so Enter opened the reject form instead of accepting: Accept has it now, on the suggestion the inbox showed.
• A suggestion for a skill no longer ends its diff with an empty line taken out.
• A download that breaks off and picks up again — a video from the library, or a file from a Delivery link — now gets the whole rest of the file, not just the next piece of it.
• An account deleted, disabled or given a new password while someone signs in to it gets no session and sends no sign-in email: the answer is the same as for any wrong sign-in. The same holds while someone takes an invite with it: it joins no workspace, and nobody is signed in.
• Save and Send in the note composer keep their size when the video's agent is assigned, starts waiting or is unassigned while you write: the pair never grows, shrinks or shifts beside the rest of the toolbar.
• On a phone, a long note keeps the line you are typing above the pinned Save and Send, and that bar sits flush with the bottom of the notes sheet instead of letting the note show beneath it.
• Auto-check works again on videos longer than about 50 seconds with current FFmpeg releases (5.1.9, 7.1.4, 8.0.2 and later), which refused how it picked the frames to read; it still reads exactly the same frames.
• When Auto-check can't read a version, the player and the inbox say so, with Run again, instead of "Checking…" for good, and it is no longer started again every time someone opens the video.

Added

• Export my data (Settings → Profile): one zip of plain files — your account, your workspaces and roles, your API tokens, apps and devices (never a token itself), and per workspace the notes and replies you wrote, your drafts, unsent recordings with their audio, your approvals and change requests, the review links you made, what you watched and what you uploaded. Never anyone else's words: a reply on someone else's note names that note by its id only.
• Delete my account (Settings → Profile, a hosted server), confirmed with your password: the workspaces only you work in go with it, the others you leave. The last owner of a workspace others work in is told to hand it over or delete it first. Delete workspace (Settings → Workspace, its owner): its name typed; everyone in it is emailed. On the server: vr admin delete-account, delete-workspace (a dry run without --yes), export-account, erasures.
• The operator takes a workspace down (the operator's Workspaces page): Suspend makes it read-only for its people — they read and download, nothing changes, its agents write nothing, its review links stop — and tells them; Lift gives it all back; Delete removes it with everything it holds once its name is typed.
• The payment form waits for the person's say in the cookie settings, as on the website (CookieConsent 3.1.0, vendored; the same cc_cookie, 182 days, the same words, buttons of equal weight, in the app's look). Stripe.js with its fraud cookies and the address search with Google's suggestions are a category of their own, The payment form, off until allowed; nothing is requested from Stripe or Google before. Refused, the checkout says so where the form would be, with Allow the payment form and Cookie settings; the choice is kept, and Settings → About opens the settings again. None of it loads before a payment form opens.
• "Powered by Lampo" stays on every review link by default, on every plan; a paid workspace may hide it. Its owners and admins find Hide the Lampo badge in Settings → Review links (where a billing provider runs); a review link's answer says badge: false only while the plan may hide it and the admins did, so a lapsed plan shows it again by itself. The source offer and the legal pages stay on the page either way. The plans and the in-app moments no longer promise "no badge"; a trial shows the badge like Free.
• A hosted server links its operator's imprint, privacy policy and terms where people meet it: under the sign-in screens' server line, at the foot of every review link's page (imprint and privacy policy, under Powered by Lampo) and in Settings → About (with the withdrawal information too). New settings VR_IMPRINT_URL, VR_WITHDRAWAL_URL and VR_CANCEL_URL beside VR_TERMS_URL and VR_PRIVACY_URL; each must be an http(s) URL. VR_SIGNUP=open now refuses to start without VR_TERMS_URL and VR_PRIVACY_URL: strangers who sign up accept the terms and are told how their data is used before they have an account.
• Elements maps: a renderer that knows what it put where sends it with the render — each named element's box, frame by frame ({v: 1, fps, size, elements: [{id, name, kind, keys: [[frame, x, y, w, h]], runs?}]}, docs/agents.md). vr push --elements map.json, vr elements <video> <map.json> [--v N], MCP track_video's elements (a file on the machine) and PUT /api/review/:slug/versions/:v/elements. Checked whole (500 elements, keys sorted and at most max(2000, the version's frames) per element, runs sorted and apart, 1 MB, names on one line), scaled to the version, thinned to the keys a straight line needs, kept per version beside the review.
• Every note says what its drawing points at in its version's map: · on #title, #card +2 on its line in vr open, vr show, INBOX.md, get_open_notes and get_note ( · near #card for a drawing over empty space), the names once in the header (elements: #title "Launch day", …), elements per note in vr open --json and in get_open_notes' new structured content. The rules come from a renderer's own review loop: the closest match under a box, an arrow by its tip, a range across its frames, the full-frame background only under a large box. GET /api/review/:slug/elements for agents on another machine.
• A part's room per note: where a person allowed a part render, the note's line ends · part f96–f188 and vr open --json / get_open_notes carry part_ok: {from, to} — the frames vr push --part-at checks.
• vr <command> --help prints that command's usage on stdout (vr push --help names --elements).
• docs/agents.md: the elements map, and "Three directions before one render" (three agents, three directions, a 4-second motion test each, one vr ask --options to pick from).
• Whoever runs a hosted server sees every workspace and account on it, at #/operator/workspaces and #/operator/accounts beside the funnel (one frame, its tab line Funnel · Workspaces · Accounts), reached from Operator in their account menu. Workspaces lists each one's owner, members, plan and state (Free, Solo, Team, a trial and its end, grace, read-only, complimentary), storage used of the plan, videos, when it was made and last active — searched by name or owner, filtered by where the plan stands, ordered by activity or age. A workspace opened shows its facts and members, and its plan set by hand through the billing module: complimentary on any plan (no limits, never billed), the trial run to a day, or back to normal billing, each with a reason, all of it in the workspace's log (who, when, what, why). Accounts lists each one's workspaces and roles, last sign-in and whether it is disabled; Disable signs it out everywhere and stops its tokens at once, Enable lets it in again.
• LAMPO_OPERATOR names who runs the server (account addresses or ids); unset, the owners of the first workspace, as before. Anyone else, and every API token, is answered as if there were no such page. Accounts now remember their last sign-in (kept from now on, shown only on the operator's page).
• An empty library offers making a video, not only uploading one: beside Upload video, Make one with an agent copies a prompt for the person's own agent — make a short video (what it's for, who it's for, how long), put it up for review with vr push (at the machine vr track), read the notes with vr open and fix them, and if vr isn't set up, ask to be connected. Under it: "No video yet? Your agent can make one and put it here for review." with Connect an agent. It replaces the line that only showed a vr push command.

Changed

• "Allow all" and "Necessary only" both wear the primary orange in the cookie settings, as on the website; Settings and Save stay plain, all one size. An accept that stands out beside a plain refusal would steer the choice.
• Reverse charge is a switch of the billing provider (BillingInfo.reverseCharge), on only where the seller has a VAT ID, which a reverse-charge invoice must name. While it is off, a business from another EU country pays VAT like everyone else, the checkout asks a business's name and VAT ID in fields of its own (the VAT ID goes on the invoices only, and is said wrong at its field), and nothing in the app says reverse charge — the plans' VAT line, the business box and the order's tax row included. With it on, the checkout is as before.
• A consumer's yearly plan after its first year may end with one month's notice (the terms' § 6 (1)): Cancel contracts here offers it where the provider says so, with the day it would end and what goes back for the time paid for after it; the contract keeps one height whichever way is chosen. What was received says the refund, the confirmation email states it, and the plan says it until the end; keeping the plan owes nothing. Businesses and plans in their first year cancel as before.
• A removed account leaves nothing of its person behind: its picture, push devices, drafts and unsent recordings in every workspace, app connections, account links and its copy in the backups of the move to workspaces go with it; what it watched stays in the team's numbers under no name. Its notes stay with the team, signed with its name. Someone removed from one workspace loses their drafts and recordings there. Deleted accounts and workspaces are written down by id (data/erasures.jsonl), so a restored backup can delete them again.
• A workspace's admins see who its members are, not what their accounts keep: no settings, first runs or other workspaces' dates, no address waiting to be confirmed.
• The checkout takes consumers, not only businesses. I'm purchasing as a business is optional and off; ticked, it asks for the business's name and VAT ID (needed only for reverse charge). A consumer ticks a box above the order button asking for the plan to start at once, before the 14-day withdrawal period ends (§ 356(4) / § 357a BGB), with the terms line right under it linking the operator's terms, privacy policy and withdrawal information; the button says Buy now (Zahlungspflichtig bestellen), in a trial and in a limit's sheet too. What was agreed is recorded with the billing provider before the order goes to Stripe.
• Prices in Settings → Billing are shown with VAT where the billing provider names the rate a consumer pays (PAngV): "€14.28 a month · €171.36 billed yearly, incl. VAT", and the checkout works out the tax for the billing address.
• Cancel contracts here (Verträge hier kündigen, § 312k BGB) replaces Cancel plan…: it leads to a confirmation step of its own (at the period's end, or for an important reason with that reason; who, which plan, until when) whose Cancel now (Jetzt kündigen) is confirmed by email at once with when it was received and when the plan ends. The sign-in's foot and Settings → About link it too: to VR_CANCEL_URL when set (a page without signing in), else to that step.
• A question an agent asked on a project or folder before its first version now leads that page, above the videos or the empty project, instead of a small chip beside the Videos · Playbook tabs: who asks and when, the question in its own words (two lines, then cut), what there is to compare ("3 clips", "Narrator: 3 sounds · Look: 3 pictures"), "n more questions in the inbox" when several wait, and one action, Compare and pick. It is the page's one orange: while it waits, the empty project's Upload a video to … and the top bar's Add video step back to neutral, and the empty state takes its quieter size. The question comes with the inbox's list the bell already keeps, so the block is there with the page (on a first visit the content waits for that list, so nothing moves when it arrives). The inbox and the bell's count already listed such questions; unchanged.
• Compare and pick uses the room it has, and every picture or clip opens large in it. With pictures or clips to look at the dialog is up to 1600 px wide and as tall as a dialog may be; tiles share the row (three clips across it, up to four a row) instead of staying 220 px. A click on a tile, or its Expand, shows that one large inside the dialog — as much of it as fits at 16:9 with the clip's own controls, its group and "2 of 3" above it, its pick under it —; ←/→ (or the arrows) go to the next of its group at the same moment, playing if it played; A goes back to the one before; 1–9 pick there too; Esc or All options returns to the grid, which kept its place. The grid and the large view share one box, so switching never resizes the dialog. A clip in the grid is a way into it (no controls of its own); Play together plays them all side by side and becomes Pause while they do; A/B for sounds is unchanged.
• Everything that is the server's rather than a workspace's goes by that one rule now, never by a role in the first workspace: the server's setup, its health check and test mail, the speech engine's model path and last error, and making workspaces (VR_WORKSPACE_CREATE=owners, and no limit with anyone). Someone invited into the first workspace as an owner or admin works there and runs nothing; with LAMPO_OPERATOR set, the first account too unless listed. An admin of the first workspace no longer sees the server's health check (the owners still do while the list is unset).
• Every setting of a link is a row with its name and its switch, in the share dialog and when a link is changed: Leave notes · Approve or request changes · See notes from other links · All versions, to switch and compare, each a switch; Downloads a menu (Off · Preview · Original); Expiry date a switch whose button says the day; Password a switch with its field in the row. No Customise to open first, and nothing opens under a row: the settings keep their height while they change. Review · Watch only · Delivery set the switches; what depends on notes is shown off and can't be changed for a link that only watches.
• The expiry date is one picker: its button opens 1 day · 7 days · 30 days and a calendar for any day (a sheet on a phone); the chosen day has the focus.
• A password is made up for you: turning it on fills its field with one that reads well (three words and a number), Generate makes another, and it can be typed over. Made with a password, the link and its password are copied together.
• A link is one line, in the dialog and in Settings → Review links: its name (a lock with a password, Expired when it ended), what visitors may do, until when and what came of it; Copy, and a ⋯ menu with Change link, Activity and Revoke. It was a card with the address, chips, a summary bar and five buttons. On a phone Copy is its icon.
• Change link opens the link as its own page in the dialog (its name, the kind, the same rows, Revoke link, Save), in the same height as the list; from Settings → Review links it opens there too.
• The dialog's links have a heading with their count and All links, which leads to Settings → Review links.
• No QR code in the share dialog.
• German says the expiry's days in German ("in 7 Tagen", it said "in 7 days").
• An agent hears notes only while it waits in wait_for_feedback, so every hand-off now says "wait now". track_video, mark_fixed and wont_fix (once none of the video's notes is open; until then "2 notes still open on this video.") end with Now call wait_for_feedback with since "<cursor>": the person's notes arrive together when they press Send. — the cursor from that moment, so a note written before the agent's next call is still heard. request_upload's PUT (and the GET after it) answers with cursor and next. vr track, vr push, vr fix and vr wontfix end with how to listen there: vr watch. Appended lines only; nothing an agent parses changed.
• A wait that ends with nothing new says what is going on: after No new feedback in 50 s. and its cursor, one line says the person's notes arrive together when they press Send, and to call again now.
• After 30 minutes of nothing but that, the agent is told to stop and to tell the person it stopped listening and how to start it again (the watch prompt, /lampo:watch), with stop: true in structuredContent. Counted per agent (over HTTP the connected agent, else its token or app), in bounded memory; a pause of more than two minutes between waits, or any real answer, starts the count again.
• On a video with an agent, notes go to it together. The agent starts on the first note it gets, so the composer's main action there keeps the note (Save, ⌘↵) and "Not sent yet" sends them all with one "Send 3 to <agent>"; the composer's own Send (now quiet, ⇧⌘↵) still sends this note and the others at once. Replies and answers to an agent's question go at once, as before; videos without an agent and review links are unchanged.
• The player says when the agent waits: "<agent> is waiting · gets your notes when you send", one quiet line in the composer or under "Not sent yet" (in place of "Only you see these…" while it waits), and the toast after a send says "<agent> got your 3 notes". While notes wait to be sent, + Note steps down so their Send is the one raised action.
• vr login <url> signs in through your browser instead of asking for an email and a password, the way gh and gcloud do. It opens the server's consent screen (after you sign in there, if you aren't): the same page apps see, naming vr and this machine, the API token it gets (vr on <machine>, --expires days or until revoked), the workspace on a server with several, and what the token may do. Press Allow and the browser hands a one-time code back to a port on this machine (127.0.0.1, PKCE); vr trades it for the same API token --email makes, and the page says "Back to the terminal". The terminal prints the address too, says it is waiting (Ctrl-C cancels, 5 minutes at most) and then who you are signed in as, in which workspace. Over SSH it prints the address to open on any device; after Allow, paste the address that browser ends on. BROWSER names another browser.
• --email (the password prompt) and --token - stay for CI, containers and scripts. A script that piped the email and the password into a bare vr login <url> now says --email. vr logout revokes a token from either way.
• Settings → Agents says how vr login signs in now.
• The public repository runs every CI job on GitHub's ubuntu-latest. A private copy can send check, unit and the browser suites to a self-hosted runner for pushes to main and pull requests from its own branches; a pull request from a fork or from Dependabot, and the Docker build, always run on GitHub's ubuntu-latest. A job on a self-hosted runner installs nothing: it checks for ffmpeg, tesseract, hunspell, fonts and Chrome's libraries and names what is missing (scripts/ci-runner.sh), keeps temp files and the npm cache in the runner's work dir, holds no token on disk and stops whatever it left running.

Fixed

• Settings → Voice notes speaks to whoever reads it. On a hosted server it says the notes are written down on the server the workspace is on, by Lampo's own speech model, and that the recording goes to no other service. It no longer mentions a Mac's graphics chip. The engine, the model and the server's settings (config.json, VR_STT_*) are shown only to whoever runs the app: the person at their own computer, or a hosted server's operator.
• Comparing two versions streams them instead of holding every frame of both in memory, so a long or high frame rate upload no longer weighs on the server; the results are the same. A comparison reports at most 1,000 changed sections of each kind.
• A render that brings the server down can't do it on every start. A background job whose process died under it isn't started again after the second time, and the comparison says why there is none; a graceful stop is no crash.
• Uploads of extreme shapes are refused (thinner than 8:1, or under 32 pixels on a side; VR_MAX_ASPECT allows more for LED boards and the like), and the analyses never look at a picture taller than four times its width.
• Uploads hold room only while they move: the disk's reserve and a workspace's plan count what unfinished uploads still have to send, an upload that stops sending gives its room back, and no one account or workspace can hold all of it.
• Pictures made while someone waits take turns across the server, and each workspace gets its share, so one team's requests never hold back another's notes; a note is saved even when its screenshots have to wait. Frames stay exactly the frames they were.
• Invites, sign-ins started by apps and decisions through a review link are bounded, per account, per address and per day, so no one can fill a file, the server's memory or a video's record, or crowd out a real client.
• A password reset goes before every other mail, so a flood of sign-ups can't hold it back; forgot password for a sign-up nobody confirmed sends its confirmation, which lets you in with a new password too, and whoever asks hears the same as for any address.
• Footage search's work follows a video's length, not how it is cut.
• A watch-only link shows no one else's notes, names or decisions, as its dialog says, links made before included, and the dialog saves it that way.
• After a link was made, an expiry or password opened for it stayed open on the empty form, the expiry saying "Never expires".
• Settings → Review links opened directly (before any share dialog) showed its lines without their styles.
• The server's setup (name the workspace, check the server, invite the team, connect agents) is its own owner's alone: an account that joined through an invite gets an invited teammate's setup, whatever role the invite gave it.
• The thank-you after a decision lies at the picture's foot, centred on it, instead of over the drawing tools; a delivery link shows its Download on a phone; a link without notes ends with its foot at the bottom of a phone's screen; no "No notes yet" under an open composer; the folder room fills wide screens and its cards on a phone lose the empty band under them.
• The check that keeps a module off the app's own routes sees every route: a router mounted under a path, whose routes it couldn't read with their full paths, now stops the app from starting.
• An imported partial render keeps to its own posters and analysis: what names its derived files is made from the part's bytes, the version it patches and where, as an uploaded part's is.
• vr admin import's report can't be altered by the bundle it reports on: every string it prints from the bundle is one line of plain text, and vr's errors drop control characters.
• An extension module can't take over one of the app's own routes: a module that names a route the app answers is refused when the server starts, in one line, and the route walk covers a module's routes too.
• Retry asks the plan's gate as Publish does. An agent's token changes and deletes drafts only: a failed or cancelled post — the inbox's work for who may publish — is a person's to change or delete, and a post that went out before is never deleted, by anyone.
• Publishing's smaller points: an upload's progress is saved every ten seconds and tells the post's page, not the library; only https links from a platform are kept; draft_post takes the API's caps; video names in get_posts and vr post are one line; the seal's workspace binding has a test; docs/publishing.md says that on your machine any local program is the owner.
• Publishing's smaller points, too: adding a posting-API key or changing one counts against the checks' limit; what a platform says about its accounts is kept one line per field, and bounded; the damaged secret.key message and go-live's backup section say that a new key costs every publishing connection and any unfinished upload; docs/publishing.md says what reviewers read of posts, on purpose.
• A folder path has a limit. A folder path named for a write is now at most 12 levels and 400 characters, refused with a 400 (never cut to fit) everywhere a folder comes in — POST, PATCH and PUT on folders, POST /api/asks, upload metadata and upload URLs, the MCP tools move_video, track_video, request_upload and ask_options, and vr move, track, push and ask before they send anything; a rename can't carry a tree past them either. Making a folder and saving folders.json take one pass. A folder a store holds from before still lists, opens, downloads, is renamed and deleted.
• An invite sign-up held from before the invite fix is confirmed only from the invited address's own inbox: the server drops a pending change of address for such a sign-up at start.
• Whether an address has an account no longer shows after Forgot password?, a sign-up, Send it again or taking an invite, in what the server answers or in how long it takes.
• An invite link made out to an address keeps that address to itself: it can't be used to find the address out.
• A sign-up waiting for its address to be confirmed learns nothing of other teams' people: its name is told apart from theirs in a workspace only when it joins it.
• A newest-only review link shows no picture of an older version, also where another link's note refers to a moment of one: the moment shows, its picture doesn't.
• /oauth/authorize sends nobody to an app's address before they decided on the consent screen: every problem before the decision shows Lampo's page "The app can't connect", in Lampo's own words.
• vr playbook skill --files and vr playbook export write only inside the folder they write to: every skill and file name is checked before the first file is written, and one that isn't plain stops the command with nothing written.
• On a hosted server with Bunny or S3, a review link's video stops playing soon after the link is revoked, expires or gets a password: the signed storage URLs its visitors get are short-lived, the server checks the link every time it hands one out, and the player fetches a fresh one when an old one stops working and plays on from the same frame. The bytes still come straight from the CDN or bucket.
• A sound offered as an option is checked before anything decodes it (at most 192 kHz and 8 channels; a clip's sound and references on notes too), every ffmpeg run on a reference's or an option's file has a time limit of its own, and a question's files go through the server's job queue one at a time. A hosted workspace whose queue is full answers 503 with when to try again, and keeps nothing of the question.
• Questions asked on folders are bounded: their number and the file's size have a limit, answered ones go first, a question keeps at most 50 answers (on a video too) and 5,000 characters of text, and the file is read once per change. A file of the wrong shape is no longer read as empty and written over; asking waits until it is put right, and other people read a sentence instead of its path.
• Storage keys stay inside the store: a render's key names its video by an id (as fix previews and references did already).
• A device's push service is reached only at a public address, through the same guard as webhooks and OAuth metadata; a refusal is logged once and not retried.
• On a hosted server, what one member's agent reports it did shows under that member's agent and at about the time it was posted: lines posted to the live monitor are listed as <agent> · <account>, like an MCP client's connection.
• Revoking or refreshing an app's token acts only on a token one of that app's connections holds: only then is anything revoked, written down or told to the open MCP waits and listens. Signing out everywhere an account that isn't there, and the clean-up of unconfirmed sign-ups when there is none, tell nobody either.
• A workspace's MCP waits and listens are shared out by role, and a part is kept for owners and admins, so their own agents always find room (docs/mcp.md).
• The /api/events live streams a person holds at once on a hosted server are bounded; one more answers 429. A stdio MCP server signed in to a hosted server no longer reads the server's recent event log (2,000 events, ~888 KB) once a second for each wait_for_feedback: it sleeps until the live stream it already follows says something happened, then asks only for what is new since its cursor.

Security

• Review links name videos by opaque ids, keep their tokens hashed (and sealed, so their owner can copy them again), and are rate-limited per visitor and per link, with every record they keep bounded.

Added

• A limit reached is a sheet, not a toast: where a billing provider runs, an owner or admin who hits the plan's storage, member or video limit (a 402) sees a sheet instead of a red toast: what was asked for and why there is no room, with the workspace's own numbers (for storage a meter with the file hanging over the plan's notch); what fits by the smallest step (Solo for room on Free, one more terabyte on a paid Team, Team for a second person, billed for at least two, said in the first sentence); Yearly · Monthly with the saving said exactly; a now → then table; paying in place (the billing page's checkout in one column, or the card on file with €0 today and the rest prorated onto the next invoice); "Or make room" with the workspace's final videos; and "Not now". Members and review-link pages keep a toast, without a way to plans they can't choose. Agents and vr keep the 402 sentence.
• An upload waits for room instead of failing: refused for room, it stays in the tray as "Waiting for room · 1.8 GB" (with See what fits and Try again) and starts again by itself once the plan holds it.
• An invite beyond the plan says what brings them in: on a subscription with a card on file, typing an address the plan has no room for shows the plan that fits, its price for the people there will be, and "Today: Nothing", with the prorated difference on the next invoice; the button becomes "Switch to Team and invite Ben". On Free the invite opens Team's sheet, and the invite goes out once it's paid.
• Insights on a plan without them explain themselves: what Insights would show from the workspace's own work, an illustration marked as one, and See Team (the feature's sheet); the sidebar's row says which plan brings them.
• The trial, visible all along: where a billing provider runs, the library's sidebar ends with "Team trial · 10 days left" over a ruler of the trial's days (today the playhead, the end a keyframe that turns half and amber in the last three days, never red). A click opens its card: the end date, what Team includes with the workspace's own numbers (members, storage, videos under review, roles, Insights and webhooks in use, review links in your name), that nothing is deleted afterwards, and one way on — Choose a plan, for owners and admins (everyone else reads who chooses it). A sheet on a phone. The account menu carries the same line and Billing with a quiet "Trial"; after the trial the line counts the week of grace. Paid plans show nothing there.
• The trial's end at the library's head, one line until clicked: three days before the date, and opened what Free would mean for the workspace (its members, storage and videos against Free's) and what happens if nobody chooses (a week of grace, then read-only, nothing deleted); on the last day the hour; after it the week of grace and read-only with three honest ways on — keep the plan, fit into Free (exactly what that takes, who would have to leave), or do nothing (what still works, what waits). A failed payment's banner leads to Settings → Billing.
• Read-only locks Add video instead of failing later: a neutral button with a lock (A and U too, and the empty library's button) whose popover says new videos wait, what still works, and the two ways on; a drop says it waits until there's room.
• Moments of value: the first fix checked on a video of the workspace's own (never the sample) adds a line under the player's stage while the trial runs ("That's the loop, on your own video", with the price for the workspace's people and Keep Team…); the workspace's first review link opened shows its maker a card in the library's corner with a picture of what the link shows and the video's chip turned "opened". Each shows once per workspace; Not now puts it away for 14 days on every device, with Undo.
• The operator's funnel (#/operator/funnel, the owners of a hosted server's first workspace only): how sign-ups become paying workspaces — four figures, the eight steps with the biggest drop before paying flagged (each step's detail on hover or Tab), each sign-up week's share per step with the weeks still in their trial marked, 4 / 8 / 12 weeks, and what is counted and what never is. Before anything was counted it says so; it never shows example numbers. Anyone else gets "There's no page here".
• First-party counts only where a billing module runs: eight steps once per workspace (signed up, setup done, first video, first review link, a link's first opening, first fix checked, active at the trial's end, paid) with the day and the plan, and what the conversion moments did per week — never names, addresses, IPs, devices, videos or anything a visitor types; no third-party trackers. Kept 13 months, then weekly counts. A self-hosted server counts nothing.
• "Not now" holds on every device: a conversion moment put away is kept with the account for that workspace (and comes back with Undo); the first loop on a video of the workspace's own and the first review link opened wait a day for the one person they are for, told live.
• A plan's 402 says more for the limit sheet: what was asked for (needed: the upload's size, or the address of the member being added), the room the workspace could make (room: its final or archived videos and their bytes) and the smallest step that fits (fits: a plan or addon:storage_tb); PlanRefusal in lib/types.ts. The sentence agents and vr read is unchanged.
• The app shows whether an agent hears new notes: listening (it waits in wait_for_feedback, or follows with vr watch), working (on what a wait handed it), not listening or not connected — from the waits it holds open, not from being connected. In the video's agent menu and card, where you assign one, in Settings → Connected agents and Connect an agent. Only an agent that listens or works reads as "An agent is on it".
• One command starts it: /lampo:watch in Claude Code. The MCP server's new watch prompt has the agent work the notes on the videos assigned to it, then keep calling wait_for_feedback until you say stop; any other agent gets the same as a sentence to paste. Settings → Connect an agent (a fourth step), the setup, the agent menu and Connected agents show it to copy, and the server's instructions and the Agent Skill tell an agent to offer it.
• When a note goes to an agent that isn't listening, the player says once how to start it.
• One log line per MCP tool call (mcp: <workspace> <agent session id> <tool> <outcome> <time>, a wait when it starts and when it ends): never note text, names, file names or tokens. VR_MCP_LOG=off turns it off.
• Footage search: B-roll for agents (docs/footage.md). Lampo indexes the newest version of every video in the background — its shots (Auto-check's cut rule), the camera's move per shot, keyframes, text in the picture, what its transcript says, one image embedding per keyframe — and answers a request in words with shots, each with its exact in and out frames: vr footage find "product close-up on white, slow push-in, ≥ 2 s, 9:16, no text" [--json] [--sheet], the MCP tool find_footage (one line per shot, and on request one labelled contact sheet: ≈ 210 + 480 tokens instead of tens of thousands), and GET /api/footage/find. The filters in the words (aspect, length, move, "no text", words on screen or said) are read as the research measured them; on its test set a right shot comes first 93 % of the time and in the first five always.
• On by default on your machine; on a hosted server, per workspace by its owners and admins (vr footage on, PUT /api/footage/settings); footage: "off" (or VR_FOOTAGE=off) turns it off everywhere. vr footage status says how far the index is; vr footage index indexes now without the app; vr footage sheet <ids> makes a sheet.
• The model downloads on first use: SigLIP B/16 int8 (Apache-2.0, 213 MB) into the cache, from one pinned revision with every file's SHA-256 checked, and runs in a process of its own (onnxruntime-node, an optional dependency). Indexing is the job queue's least urgent work, a minute of video at a time; each workspace's index is a file in its own cache. The render's path on this disk comes with a shot for the machine itself only.
• Visitors compare two versions side by side, or as a wipe. On a review link that shows every version (All, to switch and compare in the share dialog), a video with two or more has Compare beside the version switcher (in ⋯ on a phone; B on a keyboard). A is the version on screen — what notes, drawings and the decision are about —, B the reference: by default the newest older version the visitor wrote notes on, else the one before. Either side is picked in the compare bar, ⇄ swaps them on the same frame, Esc or × closes it, and a reload keeps it open. Each picture says which it is ("A V3", "B V1 · reference"), playing, stepping and scrubbing keep both on the same frame, and B's picture holds its place before its video comes. On a phone two landscape pictures stand one above the other (two upright ones side by side). Watch-only and delivery links that show every version compare too; a newest-only link never does and receives no other version's media. B comes from a new guest route that only such links answer.
• A new account starts with a short setup (#/welcome, every step skippable), each step's picture on a light table beside it that follows what is typed and picked. Lampo Cloud: the workspace's name, who the videos are for (several picks, "Something else" in a few words), the agent with its connect block in place and its status live ("Waiting for Claude Code…" → "Connected · Claude Code from …" the moment it calls), and the team row by row (a pasted list splits, each row checks itself, the role follows who the videos are for). Someone invited gets Welcome (the workspace, who invited them) and the agent. A self-hosted server's owner gets a health check — the public address, the storage, the mail relay with Send a test mail, the speech engine — before the team and its agents; without a relay the invites become links. The machine finds where its exports land and links them, finds the agent installed, and ends at the sample. Phones get the picture as a band on top and the actions as a bar at the bottom.
• Get started replaces the first run's start and strip: a card above All videos, the steps on a keyframe track beside the selected one at work (connect the agent, link or upload a video, make a review link, invite someone, right there; an accordion on a phone). It folds to one line, × puts it away with Undo, a plan picked on the website's pricing is offered ("You picked Team · add a card any time"), and the machine offers vr export for later.
• The sample opens in check mode on its fix, its agent named after the one picked in the setup; checking the fix and answering its question says That's the loop, with the way back to the library.
• No UI text calls the people who review through a link "clients" any more (the new-workspace dialog and Settings → Workspace say "another team or brand"; German "Kunde" is gone too); stage-words.test.ts now fails on any UI string that does, apart from the software kind (MCP, OAuth, API client).
• A self-hosted server checks itself for its owner. GET /api/server/health says whether the public address is set and safe, writes to and reads back the storage (with the free space), names the mail relay by host and port (never its password) and reports the speech engine; POST /api/server/mail-test sends one test mail now to the asker's own address (the same as vr admin mail-test). For the owners and admins of the server's first workspace only, and only signed in as themselves.
• Who a workspace's videos are for (PUT /api/workspaces/current/persona: for other brands, for our own brand, for my channel, something else in a few words) is kept on the workspace; Get started's order follows it.
• The plan picked on the website rides with a sign-up (plan on POST /api/auth/signup: cloud-solo, cloud-team, cloud-business; anything else is ignored) through the confirm link, for Get started to offer.
• The machine's setup finds what is there: folders with renders in a few likely places (GET /api/onboarding/folders) and the agents installed, by looking and never by running them (GET /api/onboarding/agents); both only for the machine itself.
• GET /api/onboarding adds the website's plan, who invited you, and the sample's note to check and its question; PUT /api/onboarding keeps the setup's end and the agent picked in it. New accounts carry setup_due until the setup is over.

Changed

• Billing is one page in three parts: the plan as one panel (its name and state, the price, one sentence with the dates, the trial's or a grace period's days as a ruler, what Free would mean in a trial's last days, extra storage, three meters with a dashed track where there is no limit, and the one thing to do); the plans as tiles chosen like radio buttons whose prices and lists line up in any language, with one bar for the cost and the one orange action; and the account as one panel: cards with their brands (Remove with Undo, a new card added in place), the billing details edited in their row with the VAT ID said wrong at its field and checked by VIES, and the invoices as a table with the next one estimated. One 1040 px column, centred on wide screens. A failed renewal is fixed in a sheet (what still works, a new card, paid); a complimentary workspace reads its own words; a member sees the plan without prices and can let the owner and the admins know.
• Stripe's card and address fields look like the app's own: Instrument Sans, the app's field heights, borders, focus ring and error colour, in both themes.
• Checkout is a step of its own (#/settings/billing/checkout?plan=…&interval=…&currency=…): the plan for the workspace as its title, the billing address, "I'm purchasing as a business" with the business name and VAT ID, and the card on the left; the order on the right with its lines, VAT from the address, a promotion code above the total, and what is due today while the trial runs on. Stripe's fields keep their room while they load, so nothing moves; on phones what is due and the button stay in a bar at the bottom. Paid, a calm panel says the plan, the first payment, the card and where invoices go. A new address for receipts can be given there and is saved once paid.
• The billing address can be searched: type a street and pick a suggestion, and the street, postal code and city fill in (Stripe's address search, its suggestions powered by Google); "Enter address manually" stays.
• Adding a card and paying an open invoice use the same fields: the card alone (no wallets, no Link box), its room kept while it loads, Cancel and the one orange action at the field height.
• A declined card is said in our words: what happened, that nothing was charged and the way out (an expired card, a wrong security code, a bank that declined), never the payment provider's own text; English and German.
• Paid during a trial, the trial's ruler fills in one sweep and its end turns solid green on the done panel.
• The agent menu reads calmer. Its head is the agent and, in plain words, whether it listens (with the command that starts it when it doesn't); the live log shows three lines — what it does now and the two actions before —, then Show all; what people ask most are quiet choices above the field, with the quick check as a smaller option under them, and Copy for an agent · Assign agent… sit quietly at the foot. No orange sparkles.
• What an agent did never shows its note ids: "Fixed “End card rebuilt…”" in the words of its fix, cut at a word, or "Reading the note at 00:15:08" — in the menu and on the agent button; the sidebar and board cards say "a note".
• A review link's notes look as they do in the app: a row each under the name of whoever wrote them, the selected note (and every fix waiting for the visitor's check) opened into the app's card — the timecode chip, Change or Idea, the version, Fixed · V3 in green, the marked frame at the app's size, the thread as messages and one-line status changes, Looks right and Still wrong together under their question, Reply…. An agent is Editor (Schnitt in German), and German pages no longer say "on V1".
• The share dialog says what All does: All, to switch and compare.
• The sample is two versions of Lampo's own brand film, frame for frame the same, with the title "EVERY MILE, ON THE RECORD." over the car in V1 and on the hill in V2: Alex's must ("The title covers the car…") fixed by the agent in V2 and waiting for a check, an idea on the opening with the agent's answer, and the agent's question with "Fade it out" / "Hold it" (in German too). The two files are committed in lib/sample-film/ (about 320 KB each; their origin in its README, made again with node scripts/sample-film.ts), replacing the generated gradients.
• It is in the library from the start: made in the background when an account starts in a workspace of its own (an open sign-up's confirmation, a server's setup page, the machine's first start), in the person's language. Invited people find the workspace's. VR_ONBOARDING_SAMPLE=off makes it only on request; POST /api/auth/verify takes the page's lang for it.
• Get started's steps are the sample first, then the agent, a video, a review link and the invite, ordered by where the app runs and who the videos are for (an in-house team invites before it uploads, a channel alone invites nobody, a self-hosted server adds its first video before agents and invites, the machine links before it connects). "Leave a note" and "Name your workspace" left the list (the sample and the setup teach them); "Try the sample" ticks once its fix is checked or its agent's question answered.
• Paying for a plan happens on Settings → Billing itself, no longer on the provider's own page: the billing address, a VAT ID, the card, a promotion code and the total with its tax in one form, with the card fields in the provider's frames (card details never reach the app). Only a bank that wants its own approval page is left for, and it comes back. The provider's script loads only when a payment form opens.
• The billing account is on the page too, instead of Manage billing: payment methods (add one, choose the one renewals use, remove one), the name, address and VAT ID on the invoices, and the invoices with their PDFs; an open invoice after a failed payment is paid there (Pay now). Switching plans says first what the next invoice will be; Cancel plan and Keep plan are on the plan's card.
• A plan picked on the website is marked when Settings → Billing opens with ?plan=cloud-solo|cloud-team| cloud-business (anything else is ignored).
• A billing module may name the origins its payment form loads from; the hosted app's own pages allow exactly those in their Content-Security-Policy, review links and a person's own machine never (server-mode.md).
• Every screen in front of the app is one split page. A review link's password, a link that expired or isn't available, and the owner's setup, sign-in, sign-up, "Check your inbox", the confirm and reset links, an invite and an app asking to connect (and its error) show the form on the right and, on the left, a panel with the logo and Lampo's brand film: 60 real frames of its own demo footage (F 0266–0325 at 24 fps, around 00:12:07) in slow motion, each dissolving into the next while their strip glides through a lit gate under a playhead and a readout says the frame and the timecode. It is calm, never reacts to typing, and is still on phones and with reduced motion; narrow windows and phones stack a slim band of it above the logo and the form. The film slate, the graph editor behind it and the viewfinder look (?look=b) are gone.
• The form reads quieter. Who shared the review or who invites (picture or initials, name · team, one line), the title and one sentence, then the fields (40 px, the eye in every password), the line kept for what went wrong (the field it is about shakes and turns red), the orange button without a chevron, and one short line of fine print. What that line leaves out sits behind a "?" (whose address to use, how an existing account joins an invite); the agents' vr login command sits behind Sign in an agent as a light line with Copy; the setup token's hint carries vr admin create-user the same way. Screens without a form say what they are with a small sign (an envelope for "Check your inbox", a broken link for a link that can't be used).
• Nothing of a protected review shows before its password, as before: the film is the same footage on every server and loads as images after the form has painted (the still first, the frames in an idle moment, each decoded before it moves; about 0.7 MB, and none of it on phones or with reduced motion beyond the still's 90 KB). The first-paint JavaScript stays within its budget (181.0 of 183 KB, +0.3 for three icons).
• The zoom has a place of its own and never covers the timeline. It sits in the transport row beside the speed and the sound, in the row's own buttons (the level keeps the width of its longest words, so the buttons never move while zooming; where the row is short, the level gives way), and on a phone in a quiet row above the ruler. It used to sit on the film strip and the ruler at the timeline's top right, against its end. The first-time tip is a small popover in the tooltips' material pointing at the control, ⌘/Ctrl as a key cap (it was a purple line inside the control); it takes no room, and waits while check mode's card is over the picture's foot. zoom.mjs and share.mjs check by bounding boxes that nothing of it touches the film strip, the ruler or the timeline at 390–1920 in both themes.

Fixed

• A tab left open across a server restart (a deploy) went deaf: its live stream was refused once while the server started, browsers never try a refused stream again, and the app reopened it only after a sign-in. Everything that changed afterwards stayed out of that tab until a reload — a fix checked in another tab stayed in its inbox with "Check 1 fix". The stream now opens again by itself (1 s, then longer, at most 30 s; at once when the network or the tab comes back), and once it is back every tab asks again for what it shows, since what was said meanwhile is lost.
• An agent in a chat app's sandbox no longer has to guess when its upload is refused. Claude and ChatGPT run code in a sandbox that reaches only the domains it allows; request_upload's URL is on the media host, so the sandbox's own proxy answered 403 and the render never reached Lampo. The tool now says what to do then — ask the person to allow that host in their client's network settings, or hand them the app link — and its answer ends with that link: the folder in the library for a new video, the video's page for its next version, where the person uploads it.
• Settings → Connect an agent → Claude names the domain to allow on a hosted server with a media host, with its Copy: Claude's allowed domains are in Settings → Capabilities (on Team and Enterprise plans the organization's owner adds it). The machine and a server without a media host show nothing new. GET /api/info names the media host (media_origin) to someone signed in.
• Settings → Connect an agent: every agent tile is one size, its name and line the same distance apart (a line that wrapped in "Other client" made its row taller and spread ChatGPT's and Claude's apart). "Other client" reads "VS Code, Zed and more" on one line, the clients themselves behind it as before, and a narrow panel shows two columns instead of cutting "Claude Co…".
• Playing a section right after a jump or a scrub (before that seek had landed, on a loaded machine or with hosted media) skipped the section and left the picture on its last frame: the frame the earlier seek put up was taken for where playback was. A frame shown while the player is still seeking no longer ends, loops or records a section.
• Notes written while an agent wasn't listening reach it when it starts. An agent connected over MCP (Claude Code, Codex …) acts only when prompted, and wait_for_feedback without a cursor started from now: an agent told to "wait for my notes" after they were written heard nothing. Its first call now answers at once with what is assigned to it and waiting — the videos with open notes or a request it hasn't been told of — each thing once, then waits as before.
• list_videos({session: "me"}) works over HTTP: the videos assigned to the agent the app lists for this connection. It refused ("only works when the server runs inside a Claude Code session").
• The way in loads as itself. Invite, reset, consent and confirm draw their own column at once (a separate skeleton then the form 238 px higher before); the sign-in no longer shows the sign-up-off sentence and then swaps it, "Forgot password?" holds its place until the server says it sends mail; the column is centred once and what grows in it (the fine print, "Sign in an agent", a two-line error) grows downward; the foot line no longer slides. The brand film has its own tones from the first paint (no black box before the still), a tablet keeps a 440 px column, a wide screen draws the film at most 1.2×. An invite's "Check your inbox" offers Send it again; the consent's scopes are sentences; the sign-in in front of an app's request says so; phone tap areas reach 44 px; a review link's name longer than 80 characters ends at a word.
• Words people read are never --faint (under AA contrast): password rules, terms, counts, dates, timecodes, hints and labels are --muted; a browser check keeps it so. A chip's count is a quiet second word (the UI face, a step under its label) in the tags, the lanes and the tabs.
• Settings stands in the middle of its room on a wide screen; an agent tile wraps its line instead of cutting it.
• A phone card's foot has only the lines it holds (no empty row between the stage and the agent); a narrow desktop window grows its fields and selects with its buttons; a mouse click on a native select draws no keyboard ring; the toast's × loses the browser's padding.
• Library and inbox: an empty workspace's first visit no longer jumps from a sidebar skeleton to the empty state (a browser that never saw the library draws the top bar alone); an inbox row cuts its words, never the timecode; on a wide screen the list's columns share the width; the phone's bell count sits on the bell; the item in the preview doesn't look picked beside the picked rows; the search field keeps its words at 1024; "Changed Changed…" says it once; the list's Open column one glyph; a group of cards says no age of its own; a folder's own videos are "In this folder"; Insights' facts wrap between facts.
• Player: check mode asks Looks right · Still wrong once (its card over the picture; the panel's card shows the state); the zoom tip goes with its control (it stood half off the screen over the title with the notes sheet up); German "1 kleinerer Fund"; the speed's word stands against its chevron; the dock's foot has a soft edge and its Δ chips come first so late words move nothing; a reply edit says its keys once.
• "New workspace…" in the account menu for whoever may make one, with one workspace too (it was only at the foot of Settings → Workspace).
• Compare: after a pause, B stops on A's frame. It stayed where playing left it, a frame or two off (the owner's compare and check mode too); a B that comes while A plays now plays along instead of standing still.
• Paying isn't set up / the payment provider refused read as "something went wrong (ref …)" after an earlier change; a billing module may now mark exactly those two answers as meant for people, and a declined card says why.
• The payment form keeps its room while Stripe loads, and Pay appears only once the card field is there.
• Stripe's form shows a card and nothing else: no test-mode assistant over the page, no Link box, not every payment method the account has switched on.
• An applied promotion code can be removed, and a declined card is said in the error colour right under the card field.
• An open payment form follows a light/dark switch.
• The library no longer drops when the billing banner arrives: its room is held from the first paint (after a first visit).
• The owner of the operator's own workspace reads that there is nothing to pay, not that owners choose the plan.
• A billing module's routes are held to the role they declare: each says the lowest workspace role that may call it and whether only a person may; the role table enforces it whatever the module checks itself, and a route that declares nothing stops the server at start.
• A billing module's server errors answer like the app's own: one sentence and a ref, the module's text only in the server log.
• A billing module's payment sources can't widen the page's policy past its provider: no IP addresses, and a wildcard only over the provider's own domains.
• Paying during the trial says when it ends the trial: a plan the trial doesn't carry into (Business, or with too few days left) is billed from today, and the checkout says so.
• The setup's health check names the data folder only on the machine itself: a phone through the LAN link and a hosted server's admins read "Local disk · writable · 27 GB free" (a bucket's or zone's name stays).
• A person's first run is theirs alone: an API token can no longer end the setup, put Get started away or pick the agent the sample names (PUT /api/onboarding is people only, 403 for a token).
• The sample can't be made and removed in a loop: a few times per workspace in a short while (then 429 with when to come back), and never started while a hosted server's job queue for the workspace is full (503).
• Several clicks on "Try it with a sample" make one sample, said once: one answer says it was made, the others get the same sample; it is warmed up once, and removing it while it warms up leaves nothing in the log.
• Agents never take the sample for work: vr ls, vr folders, MCP list_videos, list_folders and the review resources leave it out; asked for by name, vr open, get_open_notes, vr show and get_note say "SAMPLE: … not work" first, and there are no notes to work through.
• A section plays from its first frame, also right after the playhead was moved. Seeks that come while one is on its way wait in line, and playing a section (a note's ▶, its bar on the timeline, Play changes, a finding) didn't end that line: a seek still waiting landed after the section's own and playback went on from there. Pressing End and then Home just before ▶ played a section that starts at 0:01 from the first frame of the video. Where playback starts or jumps to (a section, the start of a loop) now goes first, and what was waiting is dropped.
• A long imported history no longer hides the note made just before it. What a waiting agent, INBOX.md and vr inbox read is the newest 2,000 to 5,000 events within 4 MB of the log; the history a vr admin import brought in was skipped there but still took its room, so a note made just before an import of more than 2,000 events never reached them. Imported history now goes into a file of its own, events.imported.jsonl beside events.jsonl, which only what reads history reads (For you, a person's part opt-ins, the folders' repair, vr export). A store imported into before keeps its log as it is: the readers pass those lines by before they count.
• Only an upload may take hours to send: a render's body (a tus piece, a one-time upload URL's PUT) may take up to 6 hours, every other request Node's own time limit, and an upload refused before its body is in is let go soon after.
• A menu no longer wiggles the first time it opens. Buttons shrink a little while pressed, and the first time a menu's code arrived mid-press it was placed from the shrunken button, then jumped 3 px when the button let go. Buttons that open a menu, popover or select keep their size while pressed: what opens is the feedback.
• Typing in the library's search while the page still loads keeps going. On a first visit with a slow answer from the server, the loading page took the first letters and the real page then replaced it without the focus: the next letters went nowhere. The real field now takes over the focus and the caret.
• A box, arrow or stroke ends where you let go. On a busy machine several pointer moves (and the release) could arrive before the page redrew, and each was built on the last drawn copy: a quickly drawn box came out short, a freehand stroke lost points. Every move now builds on the one before, and the release point is part of the shape.
• A note's card no longer jumps when its marked frame arrives. The thumbnail took no room until the picture had loaded, then pushed the thread and Reply… down: a click aimed at Reply… at that moment landed on the line above it. It keeps its frame's shape from the start now.
• On a review link, a fixed note's Looks good and Not yet no longer jump when its marked frame arrives. The same thumbnail on the review link's page took no room until its picture had loaded (an upright video's pushed the buttons 34 px down), so a visitor's tap could miss them. It keeps its frame's shape from the start now: the link's versions carry their frame size (optional width and height in GuestReviewResponse.versions).
• A team's Download all link on the media host stops working for someone who lost access: the media host checks who asked for it when the zip starts, as it does for one-time upload URLs, and a zip holds only what the folder held when it was asked for (a folder that changed: start the download again). URLs made before this change are refused once; asking again gives a new one.
• A one-time upload URL takes one upload, also when two arrive at the same moment; a refusal before any bytes arrive gives it back.
• A whole render in one request no longer breaks off after 5 minutes. Node's own request timeout cut a one-time upload URL's PUT (and a tus piece on a slow line) whose body took longer than 5 minutes to send; a render's body may now take up to 6 hours.
• An imported store's history is no news to agents: the history a vr admin import brought in is skipped by wait_for_feedback, INBOX.md, vr://inbox and vr inbox, as it already was by vr watch, the live stream, webhooks and push, and a bundle holding times from the future is refused, so real notes are always handed out.
• An imported playbook can't write over another playbook's files: it gets new ids (its own, its skills' and its suggestions'), nothing is put where a file is kept already, and its files go in only after the videos are in, taken back if the playbook can't be written.
• What an import brings in passes the checks the app runs when it makes the same thing: link references, a version's facts (read by ffprobe on the server, not taken from the bundle) and playbooks (only for the House or a folder name the server takes, within a playbook's normal limits).
• An import that fails, is killed or meets an upload leaves nothing behind. A video's history now goes into the log only once the video is in (it went first, and stayed for a video that didn't come in, to be inherited by a later upload of the same name); a video id is marked for the import under the same hold an upload takes, after checking it is free, so an upload of the same name under way is never written over or taken back (the import stops before writing anything); one import runs in a workspace at a time; and what a killed run of another bundle left is taken back by the next import.
• An import's limits and room check hold at their edges. A bundle's history may be 500 MB; the room an import needs counts its notes and history, on the disk of the store's data as well; and vr admin import on a hosted server no longer writes an INBOX.md the server never reads.
• Text at a normal margin is no longer "cut by the 52 px side crop". The text reader's box runs wider than the letters (about 20 px before an opening quote mark at 1080 px), which is the whole margin to Instagram's crop, so a Reel with its text 65–70 px from the edge got a should-fix it didn't have. Auto-check now measures where the letters start and end on the frame it read for every line near the crop, and judges the crop by that; text over a picture, where the ground can't be told, keeps the reader's box. On five agent-made Reels the five false crop findings are gone and the three real ones (a Send button under the icon column) stay.

Added

• Antigravity connects like the other agents. vr mcp config antigravity and Settings → Connect an agent (under Other client) give its setup: its own file (~/.gemini/config/mcp_config.json, or .agents/mcp_config.json in a project) with the address as serverUrl, which is the only key Antigravity reads for a server it reaches over the network. It signs in by itself, as Lampo lets clients register on their own. The app recognises it by its name and shows it with letters in a tile (Simple Icons has no mark for it); docs/mcp.md has the setup.
• vr export <out.tar> [--folder <name>]… packs your machine's reviews for a server; vr admin import <bundle.tar> --workspace <id> --owner <email> unpacks them into a workspace there (docs/moving.md). Notes keep their frames, ranges, drawings, replies and statuses, approvals keep their history, every version keeps its bytes, and screenshots, voice clips, references, fix previews, folders, playbooks, the team's watching and the history (inbox, Insights) come along. No path of the machine travels: a video tracked from a file becomes an upload in its folder, under that new id everywhere it was named. Your machine's account and the names you wrote under become your account on the server; --people "Name=email" gives others theirs; agents and clients keep their names. The history is marked as imported: no agent, webhook, push or mail hears of it. --dry-run says what would happen; a second run brings nothing new; nothing on the server is ever written over (a video id it has is skipped, a note id it has is renamed). The bundle is checked whole before anything is written (sizes and checksums, a strict format, every version's bytes against its hash) and works with renders on disk, Bunny or S3 alike.
• Edit and delete your own replies. A ⋯ on your own reply edits it in place (⌘↵ saves, Esc cancels, then "edited") or deletes it with Undo. Only its author may — by account, not a role; with an API token as with the person's own notes — and only plain words change: status changes, picks and fix previews stay as they happened. New routes PATCH / DELETE /api/comments/:id/replies/:n (api.md); agents read the reply as it is now (↳ Sam: … (edited) in vr and MCP get_note), the change as EDITED REPLY, a deletion as DELETED REPLY. Replies now record the writer's account (by_id). Review links' visitors can't edit notes, so not replies either.
• Settings → Billing, where a billing provider runs (a module behind the extension point; a self-hosted server has none and shows nothing of it): the workspace's plan and where it stands — a trial and its days, the free plan, paid, a grace period, read-only — what it uses of what the plan holds (members, storage, videos under review), and for owners and admins the plans monthly or yearly in the currencies offered. Choosing one opens the provider's own page and comes back to the new plan; Manage billing opens the provider's page for invoices, the payment method and cancelling. In English and German.
• A banner above the library when a trial is in its last three days, a payment is overdue or the workspace is read-only; a plan's refusal (a 402) says why in one sentence, in the page's language, with See plans.
• The extension point: a module may answer sign-ups (it places the person as the server would, then starts what its plans give a newcomer), email a workspace's owners and admins in its own words through the server's mailer and layout (links only into the app), say it provides billing (/api/info), give its 402 sentence in more languages, and stop its timers with the server (server-mode.md).

Changed

• A watch-only review link with no notes to show has no notes panel. The video gets the width (on a phone the page ends after the controls) and "Powered by Lampo · Source" moves under the controls. A link that shows notes but takes none lists them read only, without a composer ("Read only: this link doesn’t take new notes."). Every kind of link (review, watch only, delivery) is checked at 390, 768, 1440 and 1920 in both themes.
• The review link speaks to anyone it is sent to, a colleague, a producer or a client. The line under the video's name is the link's own name when it was given one, else the team's name, else "Shared by {name}", then the version ("Client · V9" is gone). A link nobody named is now called "Review link" (it was "Client", or "Review" for a folder); visitors never see a default as a name: not under the title, not as the password page's title, not in the room. The share dialog's first kind is Review (it was Client review), its notes setting Notes from all links (was All client notes), and it says "Visitors see it’s from {name}."; German alike ("Notizen von allen Links", "Besucher sehen …", "Geteilt von {name}").
• The app no longer calls whoever reviews through a review link "the client": a link is for a colleague, a producer or a client alike. The stage reads Out for review (was Client reviewing; German In Abstimmung) and Approved V3 via link (was Client approved V3; V3 per Link freigegeben), what it waits for Waiting for their decision / Waiting for the link to be opened, the next step Send out for review, the stepper's fourth step Shared; the one-line detail says "Mia approved V3 via review link" and "(Mia via review link)" in English too. The history marks those decisions Link, the filter and the list's column are Review link (Opened, Changes requested, Approved), the inbox's group From review links ("Mia · via review link"), Insights' card Review links with Out for review and Send a reminder, the viewers list "via For Mia", notifications and webhooks Review link activity, and the settings' and first run's sentences "people on review links", "everyone else". Agents keep their words (with_client, client_approved, "Waiting for the client's verdict", client: Mia).
• The share dialog's limits read as two options. Expiry date and Password are buttons until added; an added one is a row with its name, what it does ("The link stops working after that day.", "Anyone with the link needs it."), its field and a × that takes it away again (the expiry's "Never" is that ×). The password is a labelled field typed behind dots, with an eye to show it; one that is too short (or empty) says so under the field and no link is made until it is right, also when editing a link (Set · Change, × removes it on save).
• The timeline's zoom is a control you see, in the player and on review links. Zoom out · the level ("Fit", "2×", "30 frames") · zoom in, whose tooltips name the keys (− = and Z / ⇧Z on the level); a click on the level fits again, or zooms to the marked section (else around the playhead). It was a faint "− Fit +" in the ruler's type and absent on touch screens: tablets now have it with finger-sized tap areas, and a phone shows zoom in alone until the timeline is zoomed. The first time a video longer than two seconds opens, a tip says how to zoom without it (⌘/Ctrl + scroll or a pinch; on a touch screen, a pinch), until it is dismissed with × or the timeline is zoomed — once per browser. ⌘/Ctrl + wheel zooms more evenly (a mouse notch about 1.8×, a trackpad pinch smoothly) and never the page; Safari's trackpad pinch, which arrives as gesture events and used to zoom the page, now zooms the timeline.
• The inbox's selection is one calm system. The item in the preview and the rows picked for the bar share one flat fill a step deeper than the row under the pointer — no lifted card, no orange. "n selected · Done · Later · ×" floats over the list's foot in the float material, centred (also in the bell's popover and sheet), and keeps a room of its own at the list's end: the last row and the prompt under the list always scroll clear of it, and a list whose end was in view moves on by that room as the bar comes, so it covers nothing; the keys' list (?) sits in the inbox view's header beside By video · By kind; the "Get a ping" prompt is a quiet footnote under the list, still dismissible.
• The upload dialog can make a new project. Its picker is one control — the search in a hairline box over the list, one highlight, a check on the current choice — and ends in New project, a row that becomes a name field in place (↵ creates it, / for a folder in it); a name typed in the search that isn't there is offered as "Create “name”". Closed, it is one select-like field (also in Add video). ↑↓ ↵ pick, Esc folds the list without closing the dialog, ⌘↵ uploads.
• The upload progress is a quieter float. The tray wears the app's float material without the ruler strip; the whole upload is a ring in its head, each video a thin line with "14 MB of 21 MB · 2 MB/s · 3 s left" in the UI's face; a click on its head leaves no focus ring behind.
• Review link on phones: keyboard shortcuts ("Add note ⌘↵") are gone wherever nothing hovers or the pointer is a finger, and a note's Reply rides in its head with the state instead of a row of its own.
• Setup, sign-in, sign-up, "Check your inbox", every confirm link, forgot and reset, invites and the new-workspace dialog look like the password page clients get — the same parts, shared in code (ui/EntryForm.tsx), so the two can't drift apart again: boxed fields that take the brand's light on focus, an eye on every password, one line kept for what went wrong (a wrong password shakes its field and stamps the take NG), helper text under its field, and the orange way in with a chevron. The button is never grey while the form is empty: a press says what is missing ("Type your email first.") and puts the cursor there. The sentence under the title runs the slate's width; the setup's 01/02/03 numbering is gone.
• The top is the gate's too: the logo and one theme button. Which server this is, its version and its source move to a quiet line under the slate, with room below it and a calm patch behind its words, so it never sits on the backdrop's ruler; "Invited? Create your account…" and the command agents sign in with (vr login <url>, copyable, never broken) sit in a quiet row under the way in.
• The slate's cells say the same things on every screen, like the gate's: where (the server, or the workspace an invite is for), the scene (SIGN IN, SETUP, SIGN UP, CONFIRM, PASSWORD, INVITE, APP), the take — or where it stands (SENT, CONFIRMED, EXPIRED …) — and the day (an invite: until when it works). The NG stamp sits beside the take's number.
• Opened, the strip above All videos is the first-run start in small: its panel, the steps on their track from the left and the sample card at its own width pinned to the right edge (never stretched over half the page; under the steps on a phone), the next step's action a quiet button. Its buttons show focus with the soft ring instead of a hard black outline. Closed, it is the same one line as before.
• One decision row on a fix waiting to be checked: Looks right · Still wrong, the reason and its microphone opening in place — the same component as check mode's card, so both ask and send the same. The reply field moves below it, full width: "Reply without changing the status". Typed there, the words go with Send as reply (⌘↵, the fix still waits to be checked) or Send and reopen, both named, so a reply is never taken for a reopen. What agents read for a reply or a reopen is unchanged.
• A review link on a phone is laid out like an app's screen. One compact bar that stays at the top while the page scrolls — the title, Approve, and ⋯ for Request changes, versions, downloads and the theme — the picture across the whole width with the drawing tools in a strip under it instead of over it, one row of playback controls (step, play, step, timecode, speed, sound), then the notes. Tablets name their buttons again (Download, Request changes) instead of showing an undo arrow.
• The name is asked when it is first needed, not in a card above the notes: in the composer as the first note goes (↵ sends both), in a small dialog (a sheet on phones) when Approve, a reply or a check comes first, and the notes' head shows who is writing (tap to change). Where it is asked, the page still says who sees the notes.
• The notes say what to do once, in the empty list (without "press C" on a touch screen), not again above it; the notes' head shows how many there are. A note's Reply keeps its quiet size on a phone. Clients no longer see the picture's size in pixels under it.
• "Powered by Lampo" sits quietly in the middle at the foot of the room and the video page (the notes column's foot on a computer), linking to lampo.video, with Source beside it for the AGPL-3.0 §13 offer (source_url); German "Bereitgestellt von Lampo" · "Quellcode". Neither link sends the review link as referrer.

Fixed

• The play button's glyph stands in its middle. The pause bars were half a pixel off the circle (a 15 px glyph in a 38 px button lands between pixels, and the browser snaps it one way): the glyph now has the button's parity (16 in 38, 20 in 58) and no browser padding moves it; the play triangle stands a pixel right of the middle (--ic-play-nudge), where it looks centred. On the review link, the player's transport and the phone's; share.mjs and mobile.mjs measure it.
• Settings → Notifications: a webhook's failed delivery ("failed 2 min ago: HTTP 404") was English in German.
• The upload tray's Open button ran off the row next to a long file name.
• Check mode and the notes panel no longer disagree. A fix reopened, checked or closed on its card in the panel (or by someone else, or an agent) leaves check mode's queue at once: the check card keeps the fix it is on, or moves on to the next one when that one was settled, and the reopened fix never comes up. A fix answered in check mode shows on its card in the panel at once. The queue is read from the notes themselves, not copied when check mode starts.
• On an iPhone, the page scrolled under the status bar and its clock ("Notes" behind the time): the bar now stays at the top and makes room for the status bar and the notch (env(safe-area-inset-top) with viewport-fit=cover), the page's foot clears the home indicator, and the thank-you after a decision hangs under the bar instead of over the menus that rise from the bottom.
• A post that went out is never posted twice by Lampo. A posted or scheduled post whose later look failed for good (a revoked Google sign-in or posting key) turned "failed", and Retry uploaded it a second time; a post the posting API kept working on past six hours did the same, and a stop while YouTube's cover was set uploaded the final again on the next start. Now a look that fails keeps where the post stands and marks the connection; once a send reached the request that makes the post, a failure without an answer is sent, not confirmed (sent) for a person to look, instead of a new try; YouTube's upload session is kept until the video's id is; Retry asks the platform again about a post it holds, and sending one that went out before needs the person to say so (again). None of it depends on the posting API honouring its idempotency key.
• What goes out is what the person saw: the publish confirmation carries the post's digest (every outgoing field, as shown); anything edited since is a 409 to look at again, and every edit is in the post's history with who made it.
• One team's posts don't hold another's back on a hosted server: teams take turns in the publish queue like the job queue's; the file and cover a send needs are made before its turn; every platform request has a deadline; a kit's encode is a job under the workspace's cap (503 when it is full) instead of owed work.
• A deleted video's publish kit and cover are no longer served, and its posts don't attach to a later video of the same name (posts are bound by the video's id). The cover frame is held inside the version and one cover is kept per post. The Instagram and Facebook encode carries none of the render's own metadata.
• Reopening a final with a YouTube post that YouTube holds as scheduled asks first and says that schedule stays live (take it back in YouTube Studio), in the player and on the board; the docs said reopening paused it.
• Publishing's UI follows: Sent, not confirmed posts wait in the inbox; Retry says Try again, Check again (the platform holds it) or Post again (behind its own confirm); the publish confirmation carries the post's digest and shows an edit made meanwhile; Delete is hidden for a post that went out; edits show in the history.
• Never twice, the rest of the ways out of a send: a post whose tries ran out, or that a person cancelled or a reopened final paused, after its upload's last bytes or the posting API's post call went out was failed or cancelled, so Retry or a new publish uploaded it a second time; it is now sent, not confirmed. A kept YouTube session YouTube no longer knows after the last bytes went out is no reason to upload again. The platform's clear "no" to the post call (a slow-down, say) no longer turns the next try into sent, not confirmed. Publishing without the digest of what the person saw is refused in the store too, not only by the route, and the confirmation shows the post's visibility and time as confirmed.
• Teams take turns with posts published mid-pass too: the queue looks at what is due before every turn, so a post published while another team's uploads run waits behind one more of them at most, not the rest of that team's backlog; a send whose file is still being made takes its turn when the file is ready.
• Switching between English and German in Settings → Appearance changes every word in place: no reload, no empty page or loading states, nothing on screen moves; the section stays open, dates and numbers follow, other tabs switch too. Before, the page reloaded, and a status kept from the last visit could switch it back and forth several times (the theme too): the account's choice is now followed only as the server says it.
• Compare and pick: the row the keys act on lights up like a hovered row; its square outline sat on the list's rounded border.
• The safe-zone menu names TikTok and YouTube Shorts without "(approx.)": their zones are measured from the apps, as the Reels icons are.

Added

• Frame-exact review. The frame the player shows is the frame ffmpeg decodes (tested against ffmpeg's decode in Chrome and in Safari's engine; screenshots at 23.976, 25 and 30 fps); any codec plays through a proxy with the same frame numbers, long renders scrub at once, and safe-zone overlays and a real-size phone view come with it.
• Notes on a frame, a section or the whole video: box, arrow and freehand drawings in video pixels, tags, severities (must, should, nice, idea) and kinds (feedback, question, info), threads with replies, references (pictures, clips, links, moments of any version), and a clean and a marked screenshot of the exact frame. A section is one drag on the timeline, or I and O while playing; the timeline zooms down to single frames. Save keeps a note as a draft only you see; Send sends the video's drafts as one batch.
• Talk instead of typing. Voice notes (hold T), recorded feedback (talk, point and draw while the video plays; everything said becomes a draft note on the frame it was said at) and spoken reasons for "Still wrong". Speech is heard by transcribe.cpp where the app runs, or by an OpenAI-compatible server you name.
• Transcripts of every version, word by word on its frames: a click goes to the word, "Change the words" writes a note about exactly those frames, SRT and WebVTT captions download, and versions compare by their words.
• Versions. Every re-render or upload is the next version and open notes carry forward. What changed between versions (picture regions, sound, retimes), compare side by side, as a wipe, a difference or an onion skin, and check mode for every fix before and after.
• Auto-check on every version: typos in on-screen text, safe zones, flash and black frames, freezes, clipping, silence and loudness. Each finding says what, where (with its frame), why, and whether it looks intended.
• Where every video stands. One stage per video from to review to final, approvals per version by the team and the client with their history, one next-step button, and a board where moving a card is the decision (with Undo). A final version is locked for agents.
• Fix previews and partial renders. An agent shows a fix as a still or short clip before it renders, or renders only the shots a note allows; the next full render is compared with what was approved.
• Options before a render. An agent offers takes, lines or looks to audition side by side, sounds at the same loudness; the picks go back as one line.
• Playbooks: a House playbook and one per folder (brief, rules, references, Agent Skills), inherited down the folders and stamped on every version. Agents read them and suggest changes; people decide.
• Agents: the vr CLI and an MCP server (stdio, and Streamable HTTP at /mcp on the 2026-07-28 spec, 2025 clients still served) that hand over notes with their marked frames; wait_for_feedback and live subscriptions; the review as an MCP App (show_review); an Agent Skill; ready configs for Claude Code, Codex, Cursor, VS Code, Antigravity, Windsurf, Gemini CLI and Zed (vr mcp config), and the steps for Claude and ChatGPT (docs/mcp.md). Agents ask questions with answers to pick, answer per note (fixed, won't fix) and read the reviewer's taste.
• Few tokens: a lean tool set, pictures only for drawn notes, answers that name only what changed (bench/tokens/).
• See your agents. Each agent with its kind and mark, what it is doing now (from what it already sends Lampo, at no token cost), and on your own machine a Claude Code session that isn't running can be started with the notes.
• Review links for one video or a whole folder, without an account: notes and approval, notes only or watch only; which versions; original or preview downloads and Download all as one zip; an expiry and a password. Owners see who opened what and how far they watched. Webhooks to Slack, Discord or signed JSON.
• The library in four layouts (grid, compact, list, board) with filters, hover-scrub posters, recent videos and ⌘K search across videos, folders and notes.
• The inbox: everything that is your turn, by video or by kind, with a frame-exact preview beside the list and the actions on every row (done, looks right, still wrong, later), with Undo.
• Insights: who watched how much of what, where the time goes (on you, on agents, on clients), how agents do, and the notes that keep coming up, a click from a playbook rule.
• On the phone: an installable app, Web Push without a third-party push provider, a touch player and timeline, and bottom sheets for notes and menus.
• One app, on your machine or hosted. On your machine you are signed in at the machine itself, renders stay where they are, and Claude Code sessions, a LAN link and a tunnel are a click away. Hosted (VR_MODE=server): accounts with owner, admin, member and reviewer roles, invites, API tokens, OAuth for MCP clients, resumable uploads, storage on disk, Bunny or S3, Docker and compose, /readyz and a graceful shutdown.
• Workspaces: many teams on one hosted server, each seeing only its own work; an account can belong to several. Open sign-up gives everyone a workspace of their own.
• Email: invites, password resets, address confirmation and account notices through any SMTP relay, or into an outbox folder without one.
• A first run that teaches by doing, ticked off by what you really did, with a sample video to try.
• Going live: docs/go-live.md (one server with Docker and Caddy) and scripts/smoke.ts to check an instance.
• English and German, light and dark, one design system, loading states that keep the page's layout, and the Lampo identity (docs/brand/).
• VR_MEDIA_ORIGIN: video on a second host name of the same server, for an app host behind a CDN proxy that must not carry video (Cloudflare's terms, a 100 MB request limit). The player, review links, downloads, fix-preview and reference clips and Download all zips are redirected there with signed URLs, as with Bunny or S3: they name no video or folder (AES-GCM sealed), live 6 hours for the team and 5 minutes on a review link, and the same file gets the same URL within a minute or an hour, so the browser keeps its cached ranges. The media host answers nothing else (no page, no API, no cookie). One-time upload URLs point there too, so a whole render never meets the proxy's request limit. A link's zip is asked again when it starts (revoked or expired: refused). The setting is refused at start when it isn't an origin of its own, plain http off this machine, or the public URL's host. scripts/smoke.ts --media checks it (server-mode.md, go-live.md).
• The page, the live stream and every answer that carries a secret say Cache-Control: no-transform: a CDN in front neither compresses what the app leaves uncompressed on purpose nor rewrites the page its CSP pins.
• Test only what a change can affect: npm run test:changed [base] runs the unit tests that import a changed file (directly or through local imports, through a helper they call, or by naming it as a path), and npm run test:e2e -- --changed [base] the browser suites that cover it. Each suite's first line says what it tests (// covers: …), and the import graph does the rest; the data contract, the shell, the UI building blocks, the tokens and the dependencies run everything, and the run names the file that chose each suite. The base is main (else origin/main), plus what isn't committed. CONTRIBUTING.md "Testing what you changed".
• VR_TEST_JOBS=<n> npm test runs n test files at a time.
• Publish a final video to YouTube, Instagram and Facebook (docs/publishing.md): one post per platform per final version — title, description or caption, tags, a cover on an exact frame, visibility, a time, YouTube's category and made for kids, Instagram Reel or feed, the Facebook Page, and a required answer on realistic AI-made content; each platform's limits said as you write. Agents draft (MCP draft_post, vr post draft); only a person (owner or admin, in the app) publishes, after a confirm that names the platform and account.
• Connections in Settings → Publishing, per workspace, with your own keys: YouTube through your own Google Cloud OAuth client (resumable upload, publishAt, the cover; uploads from a project that hasn't passed YouTube's API audit stay private, and Lampo says so), Instagram and Facebook through Zernio. Keys and tokens sealed at rest, never shown or logged; every request through the address guard.
• The publish queue: posts go out at once or at their time, passing failures are tried again, a reopened final pauses what waits, an upload resumes after a dropped connection or a restart; status and history on the video (StageInfo.published), failed posts in the inbox, post events for webhooks and push, vr post and MCP get_posts to read them.
• The publish kit, no connection needed: each platform's encode, an SRT from the transcript, the cover frame, the copy as text, and all of it as one ZIP.
• Publishing in the app: Publish… is a final video's next step (and in the video menu); the composer has one tab per platform with its mark, saves as you type, picks the cover on the player, and shows the post's state, link, history and kit. Settings → Publishing adds and checks connections; the stage line says where it was published, failed posts wait in the inbox with Try again, and a Posts push switch joins Notifications. Loaded on demand, in English and German, from phone to desktop.

Changed

• The phone view looks like the phone and the app. Each phone is drawn from its public dimensions: iPhone 15/16, 16 Pro and 16 Pro Max with the Dynamic Island, Apple's 9:41 status bar, the home indicator, their corner radii and side buttons; the iPhone SE with its home button, chins and classic status bar; the Pixel 8 with its punch hole, Android's status bar and the gesture bar. One menu chooses what the phone shows — Off · Full height (the video over the whole screen, top to bottom, under the status bar only) · Instagram Reels · TikTok · YouTube Shorts · Stories — and the phone, with Show safe zones to draw the preset's zones over the app. Each app is its current interface in generic glyphs (no logos): the top bar, the rail with its counts, the name, Follow or Subscribe, two lines of caption with "more", the audio line, the tab bar; Stories' progress, header and reply field. The app is the safe-zone preset (one choice, so the two never disagree); V brings the last choice back, G cycles the presets as before, and a phone's More has the same choice. The drawings load with the phone view, not before it.
• Safe zones corrected against the apps as they are now (web/src/player/zones.ts, each with its source): Reels' top bar 14 % (Meta's guidance) and its rail down to the audio's cover, TikTok's and Shorts' top bars under the status bar, their rails as wide as the buttons, Shorts' rail to the bottom. Off the phone view the stage keeps the plain zones; Reels' old drawn mock moved into the phone view, where every app has one.
• A browser's Do Not Track or Global Privacy Control signal no longer changes what a review link records: every client's visit, the name they type, how far they watched and their downloads are recorded alike, so the team sees who watched (decided 2026-10-03). The team's own viewing is unchanged.
• The tests hold on a Linux server as on a Mac. The browser suites' Chrome is a desktop with a mouse everywhere (hover: hover, pointer: fine, also after a resize; headless Chrome on a server without input devices reported neither, and the app showed its no-hover layout). A unit file with a top-level after() keeps its imports above its first test (Node 22 ran the hook early; test-files.test.ts). Tests no longer assume the temp dir lies outside $HOME or that two mails sent at random moments arrive in one order.
• Insights asks why a video takes so many versions to approval, and what would cut that. It answers in a sentence ("7.8 versions to approval, up from 5.1. SFX notes cause most of the rounds."), then four tiles (versions to approval with its median and the target ≤ 3, the rounds the top topic came up before, right the first time, turnaround per round and whom it waited on) and cards that each answer first: What causes the rounds (each new version that followed notes, by the notes' topics, with Make it a rule — the playbook opens with the rule drafted — or "Rule exists — still 6 rounds"), versions to approval by project against the target, What came back as still wrong by topic and agent, agents right the first time, Where it's stuck now with the one thing to do (nudge the agent, remind the client, open it), and the clients who watched through review links. Every figure against the period before; with too little history the page says what shows after how many approvals. The team's own watching left the page (the API keeps it). GET /api/insights adds toApproval, causes, stillWrong, turnaround and firstTime (lib/insightsRounds.ts), stuck[].agent and a client viewer's version, link, plays and stretch watched again; nothing is renamed. The page loads as its own chunk (the first paint: 183.1 → 178.3 KB).
• The notes panel reads calm at 25 notes and more: one row per note — timecode, its severity as a keyframe glyph, the first line, where it stands when that is news (fixed, a question, carried over), replies, a section's length — with who wrote it and when said once above each person's sitting, like a chat; the picture comes under the pointer. The selected note (a click, ↑ / ↓, or the playhead stopping on it) opens in place into its card — picture, thread, references, actions — one at a time; no red outline on every must. The notes' tags are chips with counts under the filters (picked: the list narrows, with Open / Mine / Closed / All). While playing, the note at the playhead is marked and the list follows it unless you scroll or type. Phones: the same rows at a finger's height.
• Auto-check is one chip in the notes panel's head instead of a card above the notes: quiet "✓ Auto-check · 5 minor", "◆ 2 to check" in the problem colour, "Checking…" while it runs; it opens the findings in a popover (a sheet on a phone) — what looks like a problem first, the minor ones folded under them, Run again in its head, what the spelling check read in its foot. The timeline's diamonds open it on their finding.
• Auto-check's safe zones are the phone view's: text is flagged under today's Instagram Reels zones — the top 14 % (269 px of 1920, was 220) and the rail as it is since its repost button (890–1028 across, 1100–1900 down, was 905–1000 × 1165–1750) — from the same numbers the Reels preset and the phone view draw (lib/zones.ts). The finding's words and zone names (ig-topbar, ig-icons, ig-caption, ig-crop) are unchanged.
• CI is green on its first GitHub run and says what it didn't check. Every job has a time limit; every suite runs and every failure is listed, the run failing at the end; the browser suites run in two parts of equal time on the cached chrome-headless-shell; macOS (unit + WebKit) runs only on demand (macos label or Run workflow); the styleguide's pixel comparison reports skipped (a warning, never a pass) until Linux baselines are committed from the run's artifact, and perf's time budgets are warnings on CI's runners (counts and the bundle stay strict).
• One runner for the browser suites (test/e2e/lib/suites.mjs): every test/e2e/*.mjs on the harness is a suite (no list to keep); npm run test:e2e runs all of them one after another, test:e2e:parallel side by side, both listing every failure, a hung suite stopped after 15 minutes.
• The unit suite holds under load. API tests start their app through one startApp (test/lib/app.ts) with production's keep-alive — Node's 5 s default reset sockets when a synchronous ffmpeg or vr blocked the test's event loop on a busy machine (ECONNRESET); waits are for states, not times. Generated clips are encoded once per machine (<tmp>/vr-test-media/). New guards: the refusals and limits no test reached, vr admin, and a frozen store in the oldest data format (test/fixtures/store-v0/) whose agent-facing text must not change.
• Every way access ends is held to cutting open MCP waits and listens (test/unit/access-ends.test.ts): signing out, signing out everywhere, a new password, a removed account, an app disconnected by an admin, and an app revoking its own refresh token. Each fails if that way stops calling accessEnded() (then the next note reached the wait).
• The docs say what the app does today, and say it shorter. Every page was checked against the code: the agents, MCP and API pages, review links, where a video stands, playbooks, the inbox and the phone, speech, taste, configuration (every setting the code reads, none it doesn't), Docker, hosting, the data format and the architecture. Each starts with what it does and the shortest way to do it; steps are numbered, settings are tables, deep detail moved to a closing Details section, and code stays for what you type or copy. Client setups are one short block per client, checked against each client's own documentation. No line in a code block is wider than a page of docs (a test holds them to it), so code keeps its columns and never scrolls sideways.
• The docs show the app. Twenty pictures of the real app where a page explains a screen — the board, the inbox, an agent's question, check mode on a fix preview, the live agent menu, recorded feedback, the transcript, the share dialog and a link's activity, a client's review room, settings, a playbook, the MCP review card, a hosted server's setup and an app's consent screen — and the README's pictures taken again from today's UI, each in the dark and the light theme. npm run screenshots makes them all again from the synthetic demo (--only for some). How Lampo fits together is a diagram now instead of a text drawing.
• Settings → Connect an agent and docs/mcp.md name the chat apps' menus as they are now: Claude's Customize → Connectors → + Add → Add custom connector, and ChatGPT's developer mode under Settings → Security and login with the app made at chatgpt.com/plugins.

Fixed

• A section dragged on the timeline ends on the frame where it was let go. On a busy machine the release could reach the timeline before its last moves were drawn, and the section ended a frame or two short (on a review link too).
• Publishing's warning on a YouTube video over 15 minutes says what YouTube asks for — "YouTube allows it only on an account confirmed by phone" — instead of "the YouTube account must be verified" (the UI never says "verify": a fix is checked).
• Auto-check said "Spelling checked in Norwegian Bokmål" (or Danish, Italian) for German or English on-screen text, and checked the words in that language: macOS's language guess on a few words of captions and names was taken as it came. The language now comes from the video's transcript, your voice-note languages and the server's speech languages, decided by the words themselves; a language nobody expects only for a long text the guess is sure of. Older results name only German or English.
• The agent menu shows the latest run of a session when two started within the same second (a run stopped and the next started at once): it showed the older one, so a finished run could read as stopped (found by wake.mjs).
• A command or token to copy in Settings (a fresh API token's setup, Connect an agent, vr login) breaks lines at spaces only. A token is base64url and the browser broke a full line after one of its -, so on a phone or a tablet the token in the Claude Code command could split across two lines, depending on the token drawn; a URL could break after a / the same way. Each run of non-spaces now stays whole and a line too long scrolls sideways. The browser check measures a token with a - every few characters at 390 to 1920 px, so it no longer passes by luck.
• The confirm page's Choose your password screen no longer lists the workspaces the account "joins": a new password leaves the invites taken with the old one behind, so it now names them as left behind (who invited, which role), and the list of joins stays on the screen that asks for the original password, where it is true.
• A confirm link never hands someone else's password to the address's owner. A held account's link confirms at once only in the browser that chose its password (or signed in as it). Anywhere else the page asks for that password — the same person on another device is in, signed in there — or for a new one, which leaves the invites taken with the old password behind and ends whatever it could have made; it says which workspace and inviter confirming would join. Signing up with a held address and another password sends a reset link instead of the link; with VR_SIGNUP=invite asking sends the invites made out to the address too.
• Signing up a second time in the same browser (another address, or the same one again) no longer stops the first confirm link from signing that browser in: the browser keeps one mark for all its sign-ups and taken invites.
• Two workspaces asking for the first run's sample within the seconds it takes to make one each get their own: the second was told its sample was made and had none (and the first's was warmed up in it).
• VR_WEBHOOK_ALLOW_PRIVATE no longer opens the server's own network to every workspace: it lets the first workspace's webhooks (the operator's own team) reach private addresses, and every other workspace's stay on public addresses — when they are saved, tested and sent. A hook another workspace saved to a private address before stops being sent.
• Only the owners of workspace #1 make workspaces now, unless the instance says anyone may (VR_WORKSPACE_CREATE defaults to owners; it was anyone), so the server's work stays shared fairly between the teams on it. With VR_WORKSPACE_CREATE=anyone, each account makes at most VR_WORKSPACE_CREATE_LIMIT (3; the workspace its sign-up gave it counts; #1's owners have none), counted on the server, and New workspace… is offered only while it may. Sign-up (VR_SIGNUP=open) still gives each person a workspace of their own.
• An API token is told of its own workspace only: /api/auth/status, /api/auth/me and /api/workspaces no longer list the names, ids and sizes of the person's other workspaces to an agent given a token for one of them. The switcher in the browser still lists them all, and vr login still names the workspace a token works in.
• Someone who signs up with your address and never confirms it no longer keeps you from moving your account to it: the link goes to that inbox, and confirming it there removes their sign-up.
• A sign-up confirmed with a link asked for late in its week is no longer removed before the link can be used: the week counts from the newest link sent to its address.
• One address is one account. Another spelling of an address (full-width letters, another Unicode form, a domain written in Unicode or with a trailing dot) is that address, and letters from another script that look like ours, invisible characters and addresses no email can reach are refused for sign-ups, invites and new addresses. Accounts made before sign in as they did.
• A plan's member limit (Lampo Cloud) holds when someone confirms an invite they took earlier: with no room left, the confirm link is refused and stays good for when there is.
• Two people invited into one workspace under the same name no longer both join as it: the one confirming second is asked for another name on the confirm page (names tell people apart on notes).
• A held account whose invite someone else took first is refused every time its confirm link is opened (This invite can't be used): a second try used to confirm it into no workspace at all.
• Disabling someone on a server with workspaces shuts them out of that workspace only. Disable suspends the membership — no role in that workspace, its tokens and apps of theirs end — and the person keeps their account: they sign in, reset their password and join other workspaces as before. The admin can let them in again; a person in several workspaces can be disabled in one without the others hearing of it.
• Links from outside the app name their workspace on a server with several: a notification, a chat webhook's link and an agent's "open in the player" link open the video in the workspace it belongs to (the app switches there first) instead of whichever the browser was in, where a video of the same name can be another team's. Someone who isn't in that workspace is told so and lands in their library.
• An app is connected (OAuth) to the workspace its consent screen named, never another: when the session switched workspace in another tab between seeing the screen and pressing Allow, the answer is refused, the screen shows where the app would work now, and the person answers again. POST /api/oauth/requests/:id takes the shown workspace.
• A review link's page no longer tells a client that only people with this link see their name: a link to the same video that shows all client notes shows theirs too. It now says who does see the notes: the team, and anyone with this link or another that shows all client notes.
• The UI keeps to its words in a few more places, English and German: version where it said "render" (Auto-check's summary, the playbook's brief and rules, Final on a partial version, linking a file on this machine), write notes and check fixes in what reviewers do, Writing as on review links where it said "Commenting as"; the list layout's open-notes count says "3 open notes, 1 must-fix" in the UI's language instead of "3 open, 1 must".
• npm run link works on a machine without ~/.local/bin (a stock Mac): it makes the folder, links vr there, and says which line to add to the shell's profile when the folder isn't on the PATH.
• Review links offer the app's source (AGPL-3.0 §13): the room and the video page end with one quiet line linking to source_url (the link's answer carries it as source, before the password too). The Docker image carries OCI labels: its source (SOURCE_URL build argument, the repository by default), licence and the commit it was built from (REVISION; CI passes it).
• A review link names no folder above what it shares (folder names are often client and project names): a video link names no folder, a folder link its own name and where each video sits below it, and a locked link none.
• Writing on a client's note says the client reads it: what the team writes there (a reply, a reason) and what an agent writes (its fix note too) reaches the client word for word. The reply field and check mode's reason on a client's note say "Mia can see this on the review link", and agents read CLIENT: they read your replies and fix note as written in the note's head line (vr open, vr show, the MCP notes; SKILL.md says to keep paths and team remarks out).
• Signing out leaves nothing of the account in the browser's storage: the sidebar's open folders, collapsed projects, the player's last video and zoom per video, the machine's last folder, Insights' remembered layout and the tab's last library view went with the session's other data only on "Sign out everywhere"; now every sign-out clears them. The theme and language chosen on the device stay.
• The installed app's service worker no longer keeps every build's code forever: each version has its own cache and deletes the older ones when it takes over, and icons (whose names don't change) are refreshed in the background, so a new icon shows on the next load.
• Over https the session cookie is __Host-vr_session: the browser takes it only from this exact host, Secure, for the whole site, so a sibling subdomain can't set or overwrite it. Nobody is signed out by the change: a browser holding the old vr_session keeps its session, and its next request moves it to the new name. Over plain http (the app on your own machine) the name stays vr_session. The health checks, robots.txt and the 404 of a path spelled another way now carry the same protective headers as every other answer.
• An app that runs the sign-in in a popup and listens on window.opener (as browser and desktop MCP clients do) now hears the answer after Allow: the consent page and the redirect to it no longer cut the popup from its opener (Cross-Origin-Opener-Policy: unsafe-none there, same-origin everywhere else). /oauth/authorize sends the person to /?consent#/oauth/<request>.
• Query strings are checked like request bodies: a version that isn't a positive whole number (?v=0, ?v=abc, ?v=1&v=2), an array where one value belongs and similar junk answer 400 naming the field, on review links and for the team alike, instead of being read as the newest version or as "unknown".
• npm install reports no vulnerabilities: the browser suites' puppeteer-core is 25 (it was 23, whose @puppeteer/browsers pulled in extract-zip and basic-ftp with 7 high advisories); the app's runtime dependencies never were affected.
• TRADEMARKS.md covers the name Lampo and its logo only; the technical name video-review is not claimed. NOTICE.md no longer lists OpenAI's mark among the current Simple Icons release (it is from 15.22.0, the last that carried it).
• On a person's own machine, a page the owner opens on another site can no longer make the owner's requests. A request another site makes the browser send (and this machine on another port) is refused unless anyone may ask it: the app's pages, sign-in, OAuth and review links. Renders, frames, posters, screenshots and API answers also tell browsers that only this origin may load them (both modes). Agents, vr and the LAN link work as before.
• The MCP tools take the HTTP API's limits on what they write: a note, reply, fix note or reason over 20,000 characters, a tag over 60, a status over 200, a render source, caption, link or playbook suggestion over the API's caps is refused instead of going into review.json. Both ways in share one set of schemas. A reference or fix preview sent inline over MCP, or an upload URL for one, now asks the workspace's plan like the HTTP API does, so a read-only workspace refuses it with its sentence. The tool list doesn't grow: the limits are checked, not announced.
• An option's file the server fails to store — the bucket refuses it, ffmpeg dies on it — tells its details only to the machine's owner at the machine: everyone else reads a sentence with a reference (the details are in the log under it), and a store's failure answers 5xx, the server's fault rather than the file's.
• A question may show at most 8 moments of renders, and they are grabbed in the job queue like its files, not inside the request. More than 8 is refused before anything is grabbed.
• A question whose project was deleted (it waits on no folder) reads as a folder's question in vr watch, wait_for_feedback and INBOX.md (ANSWERED c_… folder - by …), no longer as a video's line with undefined in it.
• An option's file that arrives through its upload URL after the person answered or closed the question is refused (409) instead of changing an item they already picked from without anyone hearing of it.
• A question whose upload URLs are refused (too many open) is no longer stored and announced without them: the URLs are handed out first. Over MCP an item of words alone no longer takes an upload URL; ask for one with upload: true (ask_options items that had neither a file nor a link used to get one each).
• ask_options checks what it is sent the way POST /api/asks does — the question ≤ 5,000 characters, the prompt, labels, ids, links and items within the API's bounds — before anything is read. The bounds are checked, not announced: the tool list doesn't grow.
• An option's file sent inline with ask_options asks the workspace's plan first, like POST /api/asks and the references and fix previews sent over MCP: a read-only workspace refuses it with its sentence and no question is made (an upload URL's file is asked when it arrives).
• An option's sound is read only in the codecs sounds come in (PCM, MP3, AAC, ALAC, FLAC, Vorbis, Opus, AC-3); another (MS ADPCM, WavPack …) is refused from its header before anything decodes it, and a file that is no picture, clip or sound says so with the formats to send (a raw FLAC was told it was "not an image or a video").
• A stored sound or clip keeps none of the sender's tags or chapters. Loudness is read only from the meter's own summary, and a true peak that can't be read is unknown — such a sound is no longer raised as if it had room.
• vr ask against a hosted server sends a question's files inline only while the request stays under the server's limit, the rest through upload URLs (four 6 MB takes were refused as "request entity too large"); ask_options over HTTP says what fits inline there (≤ 700 KB, /mcp takes 1 MB a request) instead of 8 MB.
• Deleting a video, a sample or a folder's files on S3 storage no longer leaves an object behind whose key holds an escaped character such as < spelled out, or a control character: the listing's keys are read as XML text in one pass.
• The HTTP API no longer takes a version the video doesn't have for the newest one: vr note --v 5 against a server with only V1 recorded the note on V1 without a word, while the same call on the machine said "no v5". Notes, render sources, transcripts, frames, waveforms and the playable file now answer 404 {error: "no v5"} for it; naming no version still means the newest.
• A hung ffmpeg no longer keeps /readyz from answering: the check gives up after 5 s and says not ready. On the machine a hung claude CLI no longer holds the library's first load (up to an hour): the agents list waits a second, then comes when it comes, and claude agents gets 10 s before the session files are read instead. A tool stopped at its deadline whose children still held its output open is let go too.
• On a machine other people sign in to, they can no longer read what your agents did: the logs of runs Lampo started (an agent's whole transcript) are 0600 in a 0700 folder, the live activity file is 0600, and vr's download cache of a hosted server's screenshots is a 0700 folder. Ones made by an older version are closed on next use.
• vr login http://<another machine> no longer sends your password or token, and then every note, unencrypted without a word: it refuses and says why, unless you add --insecure (then it warns); this machine's own http://localhost and 127.0.0.1 work as before. vr mcp config claude --with-token warns that the command it prints holds your token and will stay in your shell's history.
• Access ended outside the server's own process (vr admin reset-password, a token revoked by vr admin, a restored file) or by expiry (an API token's, an app's hour-long access token, a session's) also ends open MCP waits and listens: they ask again whenever the files access is decided by have changed, and never remember a yes past their credential's own end, so the next note is held back from them.
• On the machine, each phone or tablet with the LAN link now has its own 4 MCP listens and waits; they shared one connection's 4 between them.
• On a Linux machine other people sign in to, their programs can no longer use the local app as you: a request from localhost counts as the machine's owner only from your own OS account (or root); another account's is signed out. On macOS and Windows the system can't tell accounts apart, which SECURITY.md now says. A new store is made readable by you alone, and the app warns at start when the store's folder can be opened by other accounts.
• An API token's "last used" no longer moves while an agent merely holds a wait, a listen or /api/events open with it: asking again whether it still gets in isn't a use (it rewrote the account file about once a minute per token).

Security

• A hosted server treats every request as hostile: zod on every input, no paths from clients, ffmpeg only on accepted formats and with time limits, routes matched exactly, a deny-by-default guard and a permission table that every route is in (a test walks them all, in every spelling).
• Error text by audience: only the machine's owner at the machine reads internal errors; everyone else gets a sentence and a reference to the log.
• Agents never sign off, and API tokens never touch credentials, roles, members, invites, tokens, apps, webhooks, workspaces or review links: that is a person's, signed in in the app.
• Sign-in limits that can't lock a person out, sessions that end on sign-out and when idle, the same answer for every address, one-time links stored as hashes; a new password ends sessions, devices and connected apps.
• What people write reaches agents on its own line, so a note can't forge another; names are cleaned.
• Workspaces are isolated: a test walks every route, MCP tool and file URL as a member of one workspace asking for another's things.
• Webhooks and OAuth fetches reach public addresses only; a hosted server refuses to start with unsafe settings; the container runs read-only and without capabilities.

Project

• AGPL-3.0-only with a Contributor License Agreement; the source link the AGPL asks of network services; third-party notices travel with the build.
• TypeScript that Node ≥ 22.18 runs directly; React 19.
• Tests: unit, API, server mode, storage adapters against mocks, vr and MCP end to end, browser suites in both modes (Chrome and WebKit), layout and quality checks from phone to wide screens, speed budgets (bench/perf/); CI on Linux for every push and pull request, macOS on demand.
• npm run demo and npm run screenshots on synthetic footage.