Buffer Commit, day 1

The boss is registering the Buffer OAuth app now. Buffer's form asks for a name, a logo and a public privacy policy URL.

1. Put up tap.bananafest-destiny.com, a new Worker (demo/tap), with the home page placeholder and /privacy. The privacy page says only what the build will do. Both pages carry the pulse tag.
2. Make the logo: 400×400 PNG, at /logo.png.
3. Tell the boss what to put in each field: name "Tap", the privacy URL, confidential client (not public), the permissions, and the redirect URL.

Done when both URLs return 200 and the boss has the field values.

Build step 1: the tap page, end to end

What the boss's account shows: three bio-mode networks (Instagram, TikTok, YouTube), and posts that each name their own site in the text (puzzlepress..., sizecurve...). So each post gets its own destination.

• After sign-in: a sign-in cookie for the owner only (random id → account in KV, 30 days; HttpOnly, Secure). Visitors get no cookie. Then /me.
• Posts are copied from Buffer (sent, newest 50: text, link, thumbnail, YouTube/TikTok title) when /me is opened and the copy is older than six hours. Never on a public page view. The access token is refreshed when it is within a minute of expiring, and the new refresh token is kept at once.
• Each post's destination defaults to the first web address in its text. The owner can change it, or set one for all.
• /u/<name>: the public bio page. One tile per post (thumbnail, title or first line, network), and each tile goes to /p/<post>. A skip link goes to the main destination.
• /p/<post>: counts, then 302s to the destination. It counts a person only if sec-fetch-mode is navigate or sec-fetch-dest is document, the accept header asks for HTML, and there's an accept-language. Anything else is counted as a fetcher. No IPs are kept.
• /me shows each post with its people and fetcher counts. Zero shows as 0, with one line on what is and isn't counted. It also has the bio link to copy, and Disconnect, which deletes the tokens, posts, counts and sessions.
• Privacy page: say there's one sign-in cookie, for the owner, and that metrics are read only if Buffer allows them.
• Known limit: KV counts are read-then-write, so two taps in the same second can come out as one. Fine at demo volume; a Durable Object if it matters.

Build step 3: Keep this, and Post it again

Checked on the boss's account first (made, read back, deleted at once): createPost with saveToDraft makes a draft on Instagram (reel), TikTok and YouTube with the original's video URL and a tag. Buffer refuses a video thumbnailUrl, and YouTube needs a categoryId; use the original's category. Templates hold text only, with no media.

• Each post on /me gets a "Keep this" button. It makes a Buffer template: the post's text with its site address turned into {{link}}, titled with the result ("YouTube, 3 people, 27 Oct"). It also makes a tag, "Tap: <first words>". Both show up in Buffer's own composer.
• A "Kept" list above the posts. Each kept post shows the number of posts (the original, plus any post carrying its tag, whether Tap made it or the owner did by hand), the people they sent, and the best network. Sorted by people, so winners rise and the rest sink.
• "Post it again" on a kept post makes drafts in Buffer, never sent, for each connected channel it hasn't run on yet. Each draft has the same text and video, the tag, and the network's required fields. The owner reviews and sends them in Buffer. When they're sent, they appear as tiles and count toward the kept post.
• "Forget" removes it from Tap's list and deletes the template and tag in Buffer.

Tap is up for Buffer's app form

tap.bananafest-destiny.com is live from demo/tap: a placeholder home page, /privacy (what Tap reads, writes, keeps, and how it is deleted) and a 400×400 logo. Both pages carry the pulse tag. / and /logo.png return 200; /privacy redirects to /privacy/, which returns 200. Gave the boss the values for Buffer's New Client form.

Sign in with Buffer, and the first check

The boss created the OAuth client (name Tap, confidential; account:read, posts:read, posts:write, ideas:read, ideas:write). The client id and secret are in .git-credentials and set as Worker secrets; a bogus-code call to the token endpoint gave invalid_grant, so the pair is right. Tap's Worker now has /auth/start (PKCE S256, state kept in KV for 10 minutes) and /auth/callback. The callback exchanges the code, keeps the tokens in KV under the Buffer account id, and runs one check: account, channels, sent posts, post metrics, read templates and tags, and create-then-delete one template and one tag. Buffer's auth page accepts the request (303 to its login). Waiting for the boss to sign in once.

What the sign-in showed

The boss signed in. Buffer granted account:read, posts:read, posts:write, ideas:read, ideas:write and offline_access, with a refresh token and a one-hour access token. The account has one organization and three channels (Instagram, TikTok, YouTube); five sent posts read back, each with its link. Reading templates and tags worked, and so did creating and deleting a template and a tag, so Keep this and Post it again can use real Buffer templates and tags.

Metrics did not: every post's metrics field errors with "Insufficient scope. Required: insights:read". Asking Buffer's auth page for insights:read returns invalid_scope, "requested scope is not allowed", for this client. The check's page said "no metrics, 0 of 5"; the real reason is the scope. Buffer's developer docs and changelog say nothing about third-party apps getting it; its roadmap has open requests for analytics in the API. Build step 2 (views and clicks) waits; taps are ours and need none of it.

Step 1, the tap page, is live

Signing in now leads to /me rather than the check page; the check is gone from the code, since it wrote to the account. /me reads the newest 50 sent posts from Buffer (at most every six hours), shows each one with its people and fetcher counts, and lets the owner set where each post sends people. The default is the first web address in the post's text. The bio link, /u/<name>, shows one tile per post; each tile goes through /p/<name>/<post>, which counts and then redirects. /go/<name> is the skip link. Disconnect deletes the tokens, the posts copy, the counts, the name and the sessions. Tested locally with seeded data: a browser-shaped tap counted as a person, a bare curl as a fetcher, and the skip link was counted. A javascript: destination was refused, a POST from another origin got 403, and after Disconnect the bio page 404s. Deployed; /, /privacy/, /me (302 to sign-in) and /auth/start answer as expected. The home page has the sign-in button. The privacy page now mentions the owner's one sign-in cookie, and says metrics aren't readable yet.

The boss's first look

The boss signed in and the page was built from their 50 posts, all at 0. Two faults on their screen: the "People" header wrapped onto two lines, and most thumbnails showed as broken images in their browser, though every one returned 200 to a plain fetch from here. Thumbnails now go through Tap (/t/<name>/<post>, fetched from Buffer once and cached for a day), so visitors' browsers never touch Buffer at all; all 50 return image/jpeg. The counts column no longer wraps.

Step 3: Keep this, and Post it again, are live

Built as planned. The posts copy now also holds each post's text, media URLs, tags, channel and YouTube category, plus the account's channels. Each post on /me has a "Keep this" button. Kept posts sit above the posts list, sorted by people. "Post it again" drafts for channels the post hasn't run on, and "Forget" deletes the template and tag. Results show as one line at the top of /me.

Tested live on the boss's account through a 15-minute test session, deleted afterwards. Keep this on a YouTube post made the template "YouTube, 0 people, 27 Oct" and the tag "Tap: In the middle of". Post it again made two drafts, Instagram (reel) and TikTok, with the tag and the same video, and then offered nothing more. I deleted both drafts through the API; Forget deleted the template and the tag. Nothing from the test is left in the boss's Buffer.

Drafts Tap made by hand aren't removed when the owner presses Forget; they belong to the owner. A post's sends count toward its kept post only after it is sent and shows up in Tap's copy of the posts.

Studio step 1: one link to a video in Buffer (live)

• tap.bananafest-destiny.com/make/ (signed-in owners only; linked from /me). Paste a link, and Tap reads the page: title, description, images, main text. It suggests how many videos (Comic Cover: "8 substantial points (3 with numbers) and 1 image", so 3), and the user picks 1 to 10.
• Scripts come from gpt-5.4-mini as JSON (about 5 seconds for two). Each scene line is editable, and its image can be changed. Caps are 20 runs a day per account and 200 a day overall, until the boss sets a budget.
• The video is rendered in the browser: 1080×1920 at 30 fps, slow zoom on the page's images with a blurred fill behind, big captions a few words at a time, and an end card with the address. WebCodecs + mp4-muxer 5.1.5. H.264 first. VP9 in MP4 is the fallback where the browser has no H.264 encoder; headless Chromium is one, so that is what was tested here: a 21.5-second video, 14 MB, in about a minute.
• Upload goes to KV (30 days, served with ranges at /v/<id>.mp4). Drafting makes one createPost per ticked channel.
• Live test on the boss's account: uploaded, then drafted to TikTok and YouTube. Both returned ok. Both drafts were deleted through the API, and the test video and the 15-minute session were deleted.
• Not yet seen: an H.264 render in real Chrome, and Instagram. The privacy page now covers fetched pages, OpenAI and hosted videos. Disconnect deletes videos too.
• The boss asked for more than one link ("you can put the main page and the feature page"). Now up to 5, one per line. The first is the main page and the videos send people there. Lines repeated across pages (menus, footers) count once. A page that fails is listed and the rest are used. Tested locally with three links, one of them a 404, and rendered in the browser. Deployed.
• Instagram: the test video was drafted as a reel and Buffer accepted it. The draft, the video and the session were deleted. Buffer may check the video itself only when the post is sent, so a real send is still unseen.
• The boss: "we need to be able to upload music to a clip". Added to each video: a song file from the device, a start point, volume, "Use on every video", and "Render again". The song is mixed in the browser (0.3 s fade in, 1.5 s fade out, cut to the video's length) and encoded as AAC, or as Opus where the browser has no AAC encoder. Headless Chromium is one of those. Tested with a 40-second file started 5 s in: the MP4 has two tracks, the audio is 21.5 s like the video, and the levels show the fades. The song never goes to Tap on its own, only inside a drafted video. Deployed.
• Step 2 started (plan in the studio plan). Voiceover: each scene line is read by gpt-4o-mini-tts, choosing from 6 voices per video. A scene stretches to fit its line, and music under a voice drops to 30%. Capped at 150 lines a day per account and 1,500 overall. Render all / Draft all: one channel choice drafts every rendered video with its own caption and title. Tested with two videos, one with voice and music: both rendered in 136 s, and the voiced one is 25.8 s. Speech-to-text on the finished MP4 gives back the script ("Free, print size, no sign-up. Put anyone on the cover…"), so the voice is clear over the music. Deployed.
• Captions per network, a tag per video. The script writer now returns hashtags. Each network gets the caption with as many as suit it: Instagram 5, TikTok 4, YouTube 3 plus #Shorts, X 1, Bluesky none. Each video gets a Buffer tag "Tap: <angle>" when it is first drafted, so its posts on every network can be found together later. Lines are written to be spoken, and the last one says the host name without "slash" paths. Live test: an Instagram draft with "…\n\n#ComicCover #Test" and the tag "Tap: test delete me". The draft, tag, video and session were deleted. Deployed.

Step 3: the Claude Code MCP server (live)

• tap.bananafest-destiny.com/mcp: stateless JSON-RPC over POST with Bearer auth. initialize, tools/list, tools/call and ping all work; there's a 401 without a key and a 405 on GET.
• Keys: "Make a Claude key" on /me shows the key once, with the claude mcp add line. Tap keeps only its SHA-256 hash. A new key replaces the old one, and disconnect deletes it.
• Tools: make_build_video (notes, an optional link, screenshots and a count → scripts in a build-in-public voice, saved 7 days, returns /make/?job=<id>), list_channels, draft_post (text networks only; the boss has none, so it says so), and queue_status (drafts and scheduled posts per network).
• The review page loads the job straight into the script cards. Screenshots are served at /i/<id> for 7 days. In a headless test, the cards loaded and a render finished in 58 s.
• Fixes on the way: - The first scripts called an old feature "shipped today". A new rule says only the maker's notes are new. - A worker can't fetch its own hostname (522). A link that can't be read now falls back to the notes and still sends people to it.
• Dogfood: a key was made on the boss's account through the real button (with a temporary session, then deleted) and added to this repo's Claude Code at local scope, so the key is outside git. claude mcp list shows it connected. The video of this step was made through the live endpoint: 2 scripts, with the link https://tap.bananafest-destiny.com/make/?job=fYC6PzHcxEhuVv (7 days), waiting for the boss to render and draft.
• Not done: "Write the scripts" again on a job page uses the plain voice rather than the build voice.

Step 4a: store mode (live)

• Reading a link now also asks the site for Shopify's public /products.json (or a collection's) and /meta.json (store name, currency). These fetches are cached for 30 minutes at Cloudflare, because stores rate-limit (429).
• From Cloudflare, live: Death Wish Coffee, Allbirds, Kith and ColourPop each returned 50 products with their currency. Gymshark failed: its home page wouldn't load either (530).
• The studio shows the products with picture, price and type. 5 are ticked to start, up to 10 at a time. "Write N videos, one per product" writes one script per product from that product alone. Each video carries its own page, so its pictures and end card belong to the product and it sits in the same batch for Render all and Draft all.
• Headless test: 2 products got 2 scripts, each with its own pictures. One render took 86 s.
• Fixed on the way: - A review link from Claude Code with no link no longer crashes the end card. - "Write the scripts" again on a review link keeps the build-in-public voice.
• eBay and Whatnot: no public catalogue like this, so their listings are pasted as links for now.

Step 4b: the planner (live)

• "Plan a month" on the studio: posts a week (2–7), weeks (1–4) and a start day. Tap sends the page, the store's products and the connected networks to the script model. It returns one dated item per slot (video or text, angle, hook, notes, and the finished text for text networks). It counts as one script run.
• Each video item has "Write this video", which writes a script for that angle into the batch below. Ticked items save to Buffer as Ideas (title, hook and notes, the date, and the networks).
• Local headless test: a 2-week plan for Comic Cover came back in 5 s with 6 distinct angles from the page, and "Write this video" wrote one.
• Live on the boss's account: one Idea, "Test idea from Tap, delete me", dated 2 Nov, was accepted. Buffer's API has no way to delete ideas, so the boss has to delete it in Buffer.

Step 5: long video to clips (live)

• "Or cut a long video into clips" on the studio: pick an MP4 or MOV (up to 60 minutes) and a number of clips (1–10; it suggests one per 2 minutes). The video stays in the browser. mp4box.js reads only its index (ftyp and moov, wherever moov sits) and WebCodecs decodes only what's needed.
• Sound: decoded in the browser, turned into 16 kHz mono WAV in 10-minute pieces, and sent to /api/transcribe, which passes it to whisper-1 for timed words and sentences. Cap: 60 minutes a day per account, 600 overall (underCap now counts by amount). /api/moments gives the timed sentences to the script model, which picks stand-alone moments of 20–45 s with title, caption and hashtags (one script run).
• Clip edges move to the longest pause within a few seconds, because whisper's words carry no punctuation and the first try started mid-sentence ("photo never leaves your phone…").
• Each clip is a card: the words it contains, From/to in seconds, Fit (whole frame on a blurred copy, title above) or Fill (crop the middle), captions on or off, title, caption, hashtags. Render, Render all, Draft and Draft all work as for the other videos. Over 25 MB can be downloaded but not drafted.
• Tested in Chrome for Testing 156 (Playwright's Chromium can't decode H.264) on a 94 s H.264/AAC talk made here with a clock on screen. Two clips found in 9–13 s. A 24 s clip rendered in 88 s and a 41 s clip in 149 s (headless, software encode). The clock checks out: clip 0.2 s shows 0:23.5, 12.0 s shows 0:35.3, 23.5 s shows 0:46.8. Captions follow the speech three words at a time.
• Bug found in the test: every caption was one word, because the gap check mixed whole-video and clip times. Fixed.
• Sound in this test came out as Opus, because Chrome on Linux has no AAC encoder. Chrome on Mac and Windows makes AAC, as before. Not checked: iPhone HEVC files (Chrome decodes HEVC only where the device does; the page says so) and Safari.
• Lost and restored on the way: an edit to make.js dropped the drafting helpers. The local server refused to build, and they were put back from git before deploying.
• Privacy page: a section on what is sent (the sound only) and that Tap keeps neither the sound nor the words.

The name is Dailies

• The boss: "It is your product". Cutroom is already an AI vertical-video editor and Spool an iOS music-video editor; no AI short-video product called Dailies turned up, so Dailies it is.
• Every visible "Tap" is now "Dailies": the landing page (rewritten for what the app does now), /make, /me, privacy, the MCP server (name "dailies") and its messages, and new Buffer tags. "Tap the one you saw" on the bio page stays, as a verb.
• Still on tap.bananafest-destiny.com; keys keep tap_, and an existing claude mcp add keeps working. Deployed and checked live.
• Left for the boss: the app's name on Buffer's sign-in screen.

iPhone (HEVC) videos

• The boss, on converting every video: "I don't know, theres no automated way to do it without asking someone". There's no need to ask everyone: the clips screen now checks, the moment a video is picked, whether this browser can decode it. Most can, and nothing is shown. Only when it can't, and the video is HEVC, it says so with what to do (Safari, or Chrome on a Mac or PC, or Settings → Camera → Formats → Most Compatible). Before, the error came only after the speech-to-text had run.
• Tested in Chrome for Testing on Linux, which has no HEVC: an HEVC file gets the message and the button stays off; the H.264 talk opens as before.

Keep this / Post it again, in the studio

• Each draft from /make now leaves a record (title, caption, hook, lines, link, its tag), newest 100, outlasting the 30-day video file.
• /me is the Dailies home: "Made with Dailies" lists each video and where it went out; "Keep this" on a sent one makes a Buffer template and keeps the video's own tag (Forget then deletes only the template). Kept videos get "Make more like this" and "Post it again". Bio link and counts moved into a section below, open only if anyone has used it.
• "Make more like this" opens /make with the kept link read; the scripts follow the kept video's manner on new angles. Tested locally with a seeded account: the page loaded the kept video, read Comic Cover, and wrote 2 new videos (a no-sign-up angle and a gift angle) in the same short style.
• Not tested live: Keep this on a real Dailies post, because none of the boss's Dailies drafts has been sent yet.
• Privacy page: says what the records hold; dropped "metrics" (we have none) and the Stripe line (not built).

Polish pass 1

• The boss: polish, then keep adding inside what the app does.
• /make: "Make shorts", with two equal starts side by side (one above the other on a phone): "From a link or a store" and "From a long video". The top link goes back to "Your shorts".
• Text boxes grow to fit (on a phone the on-screen lines were cut off). The planner's controls wrap in pairs, and empty status lines take no room.
• Captions in the videos break into even lines: "PUT ANYONE / ON A", not "PUT ANYONE ON / A". Clip titles too.
• Home: the links under Sign in on their own line. /me: "Make shorts", and a proper heading for "Bio link and counts".
• Checked with screenshots at 390 and 1280 px wide (no sideways scroll), a script written for Comic Cover, and the long-video box opening an H.264 and an HEVC file.

The UI plan, steps 1–7

• The boss: "go ahead with the plan" (the audit and plan are in plan/2026-10-28-buffer-commit-studio.md).
• Steps 1–5 went in as one commit, not one each as the plan said: they all rest on the new /ui.css and touch the same files, so splitting them would have meant committing half-styled pages.
• One look: /ui.css holds the colours, type, buttons (gold primary, plain, red danger), fields, cards and badges, light and dark. Page styles are layout only. One header on every page (logo, Dailies, Make shorts · Your shorts · Privacy, current page marked).
• /make: video cards have a heading and grouped fields (On screen, Post, Sound); every network tick shows the network's mark and name; file pickers are real buttons, with drag and drop for the long video; the plan reads as dated rows. Clips: the quote is two lines (tap for all), start and end fit a phone, plainer size warning.
• /me: each draft from /make now saves a small still (216×384 JPEG) with its record, shown on the "Made with Dailies" rows, served only to its owner and deleted on Disconnect. Kept badge; Disconnect is red, apart, with a confirm.
• Audit item 14 (timecode and yellow bar on a clip) is dropped: they were in the test video itself, not added by Dailies.
• Found while doing it: the sticky preview column grew taller than a laptop screen after rendering and hid Draft (no longer sticky); "Draft every rendered video to" showed before anything was rendered (now hidden until then).
• Step 6, home hero: a real Dailies short beside the headline, made locally from the Comic Cover page (one of three it wrote; no voiceover), cut to 540×960 H.264 with no sound, 0.9 MB, plus a still. It plays muted on a loop; with reduced motion it stays still with a play button. The feature list became five cards.
• Rendering it showed a bug in any video that says a web address: one word wider than the frame (VIBECIDER.BANANAFEST-DESTINY.COM) ran off both edges. The caption now shrinks its type to fit instead. Seen fixed in the three new renders.
• Step 7, live: the bio page on tap. at phone width, light and dark, and /make at desk width look as they do locally. The key page needs the boss's sign-in, so it was checked locally against the live origin only.
• Screenshots: every page at 390 and 1280 px, light and dark, after steps 1–5; the home page again in all four after step 6. No sideways scroll.
• After deploying step 6: the live video answered a byte-range request with the whole file, and iPhone Safari won't play video from a server like that. Static files skip the worker, so /home/* now goes through it (run_worker_first) and the worker cuts the range. Live: 206 with the right bytes, 416 past the end; the home page plays at both widths, light and dark, and stays still with reduced motion.

now.md, every product

• The boss (via OptiServer): now.md carries every product. It now lists the VibeCider bot (headline), Dailies, Comic Cover and Round Robin, each with status, url, blurb and since.
• Stripe, checked for the status lines: one completed checkout on the account, 29 Sep, for another bot's product. No VibeCider or Comic Cover sale yet.