Repository navigation
War room
A web page for running the review and the schedule. clippa serve serves it together
with the background jobs (http://localhost:8765, [warroom] port); clippa warroom --open runs the page on its own. It
shows what is live, where each stream is in the pipeline, the review queue and the upload
schedule. Press ? on the page for every keyboard shortcut.
The page bar under the top bar has eight pages, one on screen at a time: Review (the queue, with the item open beside it; the page opens here), Schedule (Next up, the Week or the Timeline, and the tray of items with no slot), Pipeline (the streams and Health), Money (Costs, Modal and API usage), Stats (the analytics), Library (the cast library, and where faces are named), Research (News, Blinds and Investigations) and Settings (the keys, the keyboard shortcuts and the look). The browser remembers the page. Review counts what waits for you; Schedule counts the approved items with no slot. Library, Research and Settings open today's windows from a tile each for now.
The top bar is one row: the live shows, four widgets, Search, Chat and ⋯. When the live shows don't fit beside the rest (a narrower window, the chat open, a show gone live), they move to a row of their own under it, whole, rather than wrapping inside the bar.
-
Uploads: the last 24 hours against
max_per_24h, as a meter. Opens Schedule. - Spend: this month's model and Modal spend, the Costs panel's total. Opens Money.
- Pipeline: how many streams are recording (● live) or running, or idle. Opens Pipeline.
- The health light: green when serve and the bot are up and the numbers are fresh, amber past 45 s or when a job failed, red past 2 min, when serve or the bot is down or the Kitsu sync failed. Hover for what's wrong and how old the numbers are. Opens Pipeline.
- ⋯ has the actions that change things: Sync Kitsu, Replan, Undo…, the log, the keyboard shortcuts.
Ctrl K (Cmd K on a Mac), or Search in the top bar, opens
one box that reaches everything: the eight pages, an item by its title, show, kind or name
(s07), the ⋯ actions, the Cast, News, Blinds, Investigations and Keys windows, the other look,
the Clippa chat, a live show and the queue tabs. Type a few letters, pick with ↑
↓, Enter runs it, Esc closes it. It shows the best 8 items; the
queue's own search box finds the rest. On a phone it is Search… under More in the bottom bar.
G then a letter jumps to a page from anywhere but a box: R Review, S Schedule, P Pipeline, M Money, T Stats, L Library, F Research, , Settings. An item open on top of the page closes first (it asks about unsaved edits). The number keys still switch the queue tabs.
The page used to have a Layout button for dragging panels between columns and saving layouts
([paths] warroom_layouts); the fixed pages replace it (2026-11-18), and a saved-layouts file
left on the share is no longer read.
The Clippa messenger sits at the right of the page: tell Clippa what to do
without leaving the war room. It is the /chat page itself in a frame, so it is the same chat
(the same topics, history and cards) wherever you open it. Chat in the header, or
M, shows or hides it, and the browser remembers which; M also puts the
cursor in its box, and Esc there hands the keys back to the war room. ↗ opens
/chat on a page of its own. While it is hidden, the Chat button counts Clippa's new messages.
On a wide screen the page makes room for it; narrower, it lies over the right edge and the live
player moves left of it. On a phone it is the Chat tab at the bottom instead, so it never
covers the review cards.
The war room and its chat wear one look, picked on the Settings page (or Look in the
chat's buddy list): warroom, the default, or millennium, the alternate, the same two as the
Clippa messenger (retro and pager were retired on 2026-11-18; a
browser that had one picked goes back to the default). The browser remembers the pick and its
other tabs follow it, /chat included. A look is one file in clippa/web/themes/; see
Looks for what goes in one.
warroom looks like a video editor: dark panels edge to edge with flat title bars, one blue
accent for what is selected, a colour per show, and green, amber and red for status only. The
interface is set in Inter and the numbers, times and ids in JetBrains Mono, both shipped in
clippa/web/fonts/ (SIL Open Font License, the licence files beside them), so the page loads
nothing from outside. The icons are one line set, drawn in the page itself. While a panel waits
for its first numbers it shows a grey outline of its rows.
Click a live show's chip to watch it in a player on the page or on kick.com in a new window;
the chip also says whether the NAS is recording it.
Shows in [warroom] hide_shows (e.g. ["trish", "blade"]) are left off these chips and the
show filter; their streams and items still appear.
Click a card and the item opens in a pane beside the grid, about 40% of the width; drag the pane's left edge to make it wider (the browser remembers). J / K move it along the queue, opening it if it was closed, and Esc or ✕ closes it. Below 1100 px, on a phone, and from any other page (a schedule slot, say) the item opens on top instead.
Each card shows its show as a small tag in that show's colour, the same colour on every page
(--show-1 to --show-6 in each look, in [shows] order; a show not in [shows] is grey), and
the title without its "SHOW NAME |" prefix, which the tag already says (hover for the full title).
The tabs along the top are Waiting, Changes, Approved, Uploaded, Rejected, All and Retakes: the items with a retake, newest first, each card with its last note, when it ran, what it cost and what it did (the Retakes panel before 2026-11-18). When Waiting is empty it says "Nothing waiting" and when the next upload goes out.
It works from the keyboard:
| Key | Does |
|---|---|
| J / K | next / previous item |
| A | approve |
| R | reject |
| C | retake note |
| H | hold |
| Space, arrows | play, seek the video |
| O | pop the video out into its own window |
| B | Who's this? (a face on the paused video) |
| M | show or hide the chat |
| 1–7 | the queue tabs (7 is Retakes) |
| / | the search box |
| Ctrl K | search everything |
| G then a letter | jump to a page |
| Esc | close the item |
Each card shows the item's score (when it has one) and, under the hook, one line on why it was picked: the editor pass's reason for a clip or a Short (or the model's for a hand pick), the segment score's reason for a segment; items with neither, like a compilation, have none. Hover it for the whole line; the open item shows it in full under Why.
Tab reaches the card titles, schedule slots and live chips too; Enter or Space opens them. The card buttons name their key on hover.
The search box looks through every stream's titles, descriptions, topics and hashtags, and what is said in each video (its captions). Show, kind, status, order and dates sit behind Filters; each one that is on shows as a chip beside it (× clears it, "Clear all" clears them all), and they combine with the search. The queue is newest stream first; the order filter sorts by score, length or kind instead. The view, order included, is in the page's URL, so it can be bookmarked.
Tick items (shift-click for a range, "Fill the gaps" on a phone, or "Select shown") to approve, reject, hold, release, pin, reopen or send the same retake note to all of them, or to find and replace in their titles, descriptions, hashtags, tags or playlists: the dry run lists every line that would change before anything is written. A selection survives a change of tab or filter, but bulk actions only touch what is on screen: the bar says how many picks are left out.
The health light in the top bar shows the age of the page's numbers (amber past 45 s, red past
2 min). When a refresh fails it says Offline there (hover for why), with one
toast when it starts and one when serve is back; before the first answer the panels say
Loading… or why they are empty. A tab in the background stops polling and catches up when
you come back. A page whose sign-in has gone says Signed out there and asks for the ?key= link again.
A new title for an item that's already on YouTube renames the video there too (Scheduling and uploads): the item window says so above the title box and asks before it saves, and find and replace's preview marks which items it would rename. Its description and hashtags stay Kitsu-only.
Decisions, hold, pin, title/description/hashtag edits, find and replace and thumbnail words can be undone from their toast for 10 s or from "Undo…" under ⋯ (the last 20). Undo goes through Kitsu too, and an item already on YouTube keeps its live text (change it in Studio), except its title: undoing a title change made while it was up renames the video back on YouTube (one made before the upload is undone in Kitsu alone). Not on the list: the cover editor, faces, cast, and a retake already started (undo puts the item's status back; the retake still runs).
What costs money says so first. A retake note says what a retake costs (the average of the recent ones), a note sent to several items asks before it starts that many, and the Retakes tab shows each one's cost and this month's count and total. Cast's Find line says what its labels can cost.
A button sends once: a double tap is one request, and Sync Kitsu and Replan (under ⋯) stay greyed out until they finish. Closing an item, opening another card beside it, or moving on with J/K, a swipe or Approve, asks before throwing away a title, description, note or release time you typed but didn't save.
Titles, descriptions, hashtags and cover words are checked as you type ("hosts", unmasked swearing, YouTube's length limits with the show prefix), and cards with a problem carry a ⚠ badge. The badge counts the pre-flight checks' findings too (a face cut off, the wrong person on the cover), and the item's window lists them under Checks.
"Edit cover" in an item's window opens it: pick the main frame (shift-click for the inset)
from the candidate frames, drag and resize the face boxes, choose a photo or teaser photo
from [thumbnail] photos (a cast library person by label too: "Jer · angry"), change the words. Preview renders without saving; Save re-renders
with no model calls (every cover of the stream when the teaser changes) and re-attaches it in
Kitsu.
Who is who on screen (#54), kept in [paths] faces/known.json for every show.
-
A stream's groups: "Faces" in an item's window, or "faces" on a scanned stream in
the pipeline panel. Each group of faces
clippa facesfound has its contact sheet (first row its clearest faces, second a spread over the stream), its name, and whether that came by hand or by face (with the match, 0.7+ names a group). Type a name and press Name (this stream only) or Name + enroll (every stream; the stream's look-alike groups are named at once). "✓ Right" keeps a by-face name as a hand one; Clear removes a name. Filter by named/unnamed. - Who's this? (B): pause the video on a face, press B or the button, and drag a box round one face. Clippa takes that frame at full size, finds the face in the box (small ones too) and says who it looks like; Keep this face adds it to the known faces and names the stream's look-alike groups. For a face too rare or small to make a group: a guest, a costume, a picture-in-picture. Not on a phone (no inline video there).
- The × on a known face forgets that person (every sample). Hand names stay.
The command line does the same: clippa faces <stream> --name F03=Jer --enroll,
clippa faces <stream> --at 1:23:45 --box x,y,w,h --person Jer, clippa faces --forget Jer.
heat on a stream in the Pipeline panel, or beside the stream's name in an item's window, opens that stream's heat graph: the curve the signals stage makes from chat and loudness, over the whole stream. The picker at the top changes the stream.
- The shaded bands are the windows Clippa looked at (
windows.json), each with a dot at its peak; the dashed line is the least heat a window needs ([signals] min_heat). A grey stretch had no chat recorded. - Under the curve, each cut segment, Short and clip sits where it came from, coloured by where it is: not in review yet, waiting, changes asked, approved, scheduled, uploaded or rejected. The legend counts them. Opened from an item, that item is outlined.
- Point at the graph (or tap it) for the time, the heat, the window's chat, laughing, "clip" and loudness scores, and the items there. Click an item to open it; Tab reaches them too.
- The table lists the twelve hottest windows with what was said at the peak (when the transcript was there as signals ran) and what was cut from each, or "nothing cut here".
A long stream is wider than a phone's screen: the graph scrolls sideways. It shows the hottest second of each stretch (about 1,200 across the stream), so a short peak isn't averaged away. Nothing here costs anything or changes anything.
Headlines about the shows' people (clippa news, #110). News on the Research page opens it.
-
Stories from the last 1, 3, 7 or 30 days, newest first: the headline the judge wrote, strong
or maybe, who and which show, why it matters, and the outlets' links. "posted" means the bot
put it in Discord (
clippa servefinds them, the bot posts them); "in context" that it was added to a show's next stream. - Tonight's context… shows the text the stream would get and the show it goes to (change it in the menu); Add to its next stream queues it. Dismiss hides a story and the bot never posts it.
- Search asks Google News now (the last 7 days) and adds the stored headlines that match.
- The header line says how often serve checks and how many sources are failing (hover for why).
Gossip blind items with their clues read (clippa blinds, #111). Blinds on the Research page
opens it: each blind's text, its typed clues, the claim, and a box for your guess, which you give before
Clippa looks into anything. The whole of it is on Blind items.
The API keys serve itself has, which on the hub are the ones serve, the bot and the worker use.
Keys on the Settings page opens it. Each key shows its env var, where it was found ($ANTHROPIC_API_KEY,
or its file in secrets/), what uses it, and the last answer from its provider: ✓ works,
✗ rejected, or ? (no answer about the key, or not reachable). Check asks one provider again,
and Check all asks every one that's set. Each check is one free request (a model list or a
"who am I"); Brave, Pixabay, DVIDS and sync.so aren't asked, since there's no free way to. The
answers are kept in streams/.serve/keys.json. Streams the worker holds for a rejected key
(How the processes talk) are listed at the top.
No key's value ever reaches the page. To change one, edit deploy/hub/env.local (or its file in
secrets/), then run deploy.sh, which checks it against the laptop's own first
(NAS-hub). clippa keys prints the same table for the machine it runs on
(--check asks, --check Anthropic asks one; exit 1 on a rejected key).
The cast library (#41): a few head-and-shoulders photos of each regular in [thumbnail] photos/<Person>/, for covers and the weekly. Cast on the Library page opens it.
-
The list is everyone with photos or an enrolled face, with how many are kept and new;
"!" marks fewer than
[cast] min_photos. "Someone new" adds a person. -
A person: other names and channels (YouTube
@nameor Kick addresses), saved with Save; the channels are what Find > Channels reads. -
Find: Commons (Wikimedia, freely licensed), Channels (their avatar and recent video
thumbnails), Image search (Brave; needs
$BRAVE_API_KEY), a photo's web address, or a stream id (that stream's sharpest faces of the person, named in its Faces window). One runs at a time in the background; the window shows its progress and fills in. Each photo is checked against the person's known face, so other people are left out, and labelled by a small model. - New photos come first, blue-edged; ⚠ is the labeller's doubt (a designed thumbnail with text, a hidden face). Keep or Drop each; fix a label with its menu. Drop asks first: the photo is deleted and its source never fetched again, with no undo. Kept photos are what covers use: the cover model picks the label that fits the headline (angry for a meltdown, laughing for a gloat), else neutral; a stream, Commons or channel photo before an image search one. The weekly draft does the same for people it pictures from a photo, and everyone with a kept photo joins the prompts' known figures.
Photos also arrive by themselves (the worker after each stream, serve's daily cast auto);
the sure ones are already kept (not from image search, an enrolled face matched well, a label
with no doubt), so
look over Kept too and drop what's wrong.
The command line: clippa cast list|grab|find|add|keep|drop|label|set|auto (Cast library).
The Schedule page answers "what goes out next". Next up across the top shows the next five
uploads with their covers and how long until each (a red edge: its last upload failed). Under it,
Week shows seven days from today, hours down; the arrows page a week at a time out to the
scheduler's horizon_days. Each clip is a block in its show's colour with a stripe for its kind:
filled once it is uploaded (private until it goes public), outlined until then, faded once its time
has passed, with a pin when it is pinned. Each day's head shows its count against max_per_day, a
red line marks now, and the open slots are dashed. Hover a block for the whole title and its
state; click it to open the item. List shows the days one under the other instead, as before
(a phone always gets the List and the next three; see On a phone).
Timeline answers why a show got nothing this week and how close the upload cap is. It runs
from two days ago to seven ahead, a lane per show with a mark for each item at its time (the same
fills and outlines as the Week; Shorts are the short marks). The top lane is the recordings: a bar
in the show's colour for each, to now while it is still live (dashed when the show is live but not
being recorded), then a hatched stretch until kick_first_hours after it ended, when nothing of
it may go out; the same stretch shows in that show's lane. A lane with nothing planned ahead says
"nothing ahead" in amber. The Uploads lane marks each upload of the last two days and says how
many of [youtube] max_per_24h the last 24 hours used, the shaded band across every lane; a
dotted green line marks when the cap next has room, if it is full. Open slots are the small
dashed marks along the bottom; drop an item on one to pin it, as on the Week.
Ready to slot, beside it, lists the approved items with no slot that can still go out, newest stream first: the same number as on the Schedule tab. The ones the scheduler let go (over their stream's budget or stale, see Staying current) are one link under it, "N not going out", which opens them in the queue; a pin there (Pin or Next open slot in the item) sends one anyway. Drag an item or a block onto an open slot (or a day's head) to pin it, or onto "unpin"; the page warns before a drop breaks the spacing rules. Or tap an item in Ready to slot, then an open slot or a day's head: a bar at the bottom says which item you picked up (the open slots show while you do, even when they are hidden), and Cancel, Esc, another page or a second tap on the item puts it down.
Or let clippa pick the slot: the arrow-to-a-line button beside an item (or Next open slot in its window) finds
the first open slot of its kind that the scheduler's rules allow: the slot times, max_per_day,
min_gap_minutes, kick_first_hours after the stream and [youtube] min_lead_minutes from
now. Segments to open slots… does that for every approved segment without a slot, but
only into a slot before its stale_days cutoff, as the scheduler would. Both show a preview
first and pin through Kitsu ("Publish at"), like the Pin button, then replan; Undo takes the
pins back. A pin counts against its stream's stream_budget first, so pinning a segment from a
stream whose budget is full bumps that stream's weakest automatic one: the preview names it.
The bulk button leaves those streams out, since a pin there would only swap one for another.
Raise stream_budget to send more of a stream.
The Uploads panel under the tray shows uploads in the last 24 h against the cap, the last automatic run and its errors, and what goes next. Uploaded slots show views and a YouTube link.
Log on the Health panel (or ⋯ > the log) shows the end of the serve or bot log ([paths] serve_log,
bot_log), with a word filter and an "errors and warnings" switch. An error toast that says
"details in the serve log" has a Log button that opens it filtered to errors. Values of secret
environment variables and the page's key are blanked before the lines leave serve. A log set
to "" (console only) says to use docker compose logs instead.
The Money page has the first three, the Stats page the analytics.
-
Modal: the month's Modal bill, the spend rate, GPUs up, transcriptions
running and failed, and each archive's progress.
clippa modalprints the same. -
Costs: the month added up (the Spend widget shows its total). Every model call (logged as it is made, with its
stage and stream), Modal runs, and
[costs] monthlyfixed costs, by stage and show.clippa costsprints the same; revenue joins it once the channel is monetized. -
API usage: the month by the Claude API's and OpenRouter's own count,
today, by model, OpenRouter's week, limit and balance, each next to what our log recorded.
Claude's needs an Admin API key (Costs);
clippa costs --providersprints the same. -
Analytics: a month of YouTube. Views, watch time and
subscribers against the month before, views per day, where views came from (Shorts feed,
search, ...), and the first-7-day numbers of that month's uploads by show, kind, weekday and
hour, with the best and weakest.
clippa analytics --reportwrites the same asreports/analytics/<month>.mdplus a .csv of every upload, and ends with what predicts views, which the panel leaves out. serve refreshes the numbers every[analytics] auto_hours. A passing Google error (a 500, a rate limit) is retried on the spot; a video still failing is skipped and asked again next run, and only one failing three runs in a row fails the job. A failed run is tried again in 30 min, then less often (How the processes talk), not a day later.
The page bar moves to the bottom: Review, Schedule, Pipeline, Chat and More, which lists the other
pages. The top bar keeps the Uploads widget, the health light and ⋯ (Sync Kitsu and Replan
are in there, never the first thing to tap). The Review queue is one column, the first card on
the first screen; Filters opens as a sheet from the bottom (Done, or a tap above it, puts
it away). Items open full screen with the decisions at the bottom and swipe to the next one, and videos
play from a 360p copy (lowres/ beside each preview, about a quarter of the data). It can be
added to the home screen. Away from home it goes through Tailscale: clippa warroom --tailnet on (Phone access).
The Schedule is an agenda: Next up, then what is still to come by day. What already went out or was due folds into one row at the top (tap it for the last two days), days with nothing are left out, and the open slots show only while you pin: tap an item in Ready to slot, then a slot or a day (nothing is dragged on a phone). Pipeline, Money and Stats are one column of the same panels, and the windows (Cast, News, Blinds, Investigations, Keys, the log) open full screen.
The page is in clippa/web/ (warroom.html and what it loads). warroom.css is structure
only: every colour and the shape of windows, title bars, buttons and fields comes from the look's
file in web/themes/ (themes/warroom.css is the one place the default's colours live, and
declares its fonts from web/fonts/), loaded by theme.js, which the chat page shares. The icons
are <symbol>s in a sprite at the top of warroom.html, each with the id i- and its name;
markup points an <svg class="ic">'s <use> at one and scripts call ic("name") (js/core.js), and a test
checks that every name asked for is in the sprite and that no emoji stands in for one. js/chatdock.js shows and hides the chat frame. The server is the clippa/warroom/
package. httpd.py holds the route tables GET_API and POST_API, which map each /api/... path to
a function; data.py builds the page's data, review.py the decisions (through clippa/actions.py),
and each panel has a module of its own.
Home · Getting started · clippa doctor · These pages are built from docs/wiki/ in the repo, so edits made here are overwritten: how to edit them
Run it
- Overview
- Getting started
- Roles and services
- Configuration
- Pipeline
- Review
- Reviewing
- War room
- Clippa messenger
- Scheduling and uploads
- Costs
- Checkup
- Audit
- Backups
Install
- Install
- Install checklist
- Kitsu
- Discord app
- YouTube sign-in
- Windows
- macOS
- Transcribe on Modal
- NAS recorder
- NAS hub
- Claude on the NAS
- Phone access
- Model providers
- Adding a show
- YouTube channel
Features
- Pre-shows
- Fetching VODs
- Speakers
- Faces
- Covers
- Cast library
- B-roll
- Clip art
- Weekly issue
- X posts
- TikTok
- News watcher
- Character pieces
- Documentaries
Research
Working on Clippa