# CURRENT.md — session state

Last updated: 2026-09-16

This file tracks the work state. Read it first. Update it before you stop.

## Current focus (broad)

The text layer is done. Voice actions and audio work. The local LLM harness (Veskslop agent, Go +
Eino) is scaffolded and smoke-tested; the daemon-side task push is next. The web UI has a first
mobile draft at 390px.

## In progress

- Media archive project, branch `media-scrape-project` on `veskslop-t`. Design approved 2026-09-16
  (docs: [media.md](docs/media.md), [media-scraper.md](docs/media-scraper.md),
  [slopbot.md](docs/slopbot.md), [media-mcp.md](docs/media-mcp.md)). Decisions: scrape the channel
  the command is sent in; slopbot (discord.js bot core) for guild channels and the puppet core
  bridge for group DMs; one self-contained SQLite + media dir per scrape under `media-scrape/db/`
  (gitignored); links indexed by default, downloads optional, third-party embeds never downloaded;
  30 GiB default limit with a 90% warning and 5% no-delete overrun, usage bar and clean button in
  Settings → Bot; Media and Bot top-level tabs; MCP hooks designed for C8. The slopbot scaffold
  (config loader, invite helper) and its discord.js dependency were added 2026-09-16. C5 smoke
  test: `/api/media/status` and `/api/media` answer on a daemon with no core.

  Checkpoints, push after each:
  - [x] C1 design docs, AGENTS/INDEX pointers, `.gitignore`
  - [x] C2 core message shape: embeds, forwards, attachment metadata
  - [x] C3 media-scrape store and puppet walker, tests
  - [x] C4 slopbot backend, tests
  - [x] C5 daemon wiring: `scrape media` command, MediaService, `/api/media/*`
  - [x] C6 Media and Bot tabs in the Web UI
  - [x] C7a storage limits, retention, warning/confirm flow, storage and clean API
  - [x] C7b Settings group: limit field with info hover, usage bar, clean modal
  - [x] C8 MCP service hooks (`scrape_media`, scrape reads, storage, clean; reads in `run_commands`)
  - [x] C10 downloader: attachment and sticker bytes, hash-addressed, serial, size-capped, retries;
    embeds excluded in SQL; local files served at `/api/media/<scrape>/<item>/file` and preferred
    by the Media tab. Live run 2026-09-16: scrape `...-20260916083401` fetched the 784 KB
    attachment; storage counted actual bytes.
  - [x] C11 preview caching: Discord-hosted image/GIF previews (WebP 320px, hash-addressed, 5 MiB
    cap, one attempt) run after every walk in both modes. Live run `...-20260916090722`:
    `previewed: 2`; thumbs served over `/api/media/<scrape>/<item>/thumb` (200).
  - [x] C62 VOX prosody controls. The last open Veskslop-side item from the MicroslopSam design:
    `tts.vox` and `tts.voxClean` preferences (per-line overrides on `/api/audio/speak` and MCP
    `speak`), `VESKSLOP_TTS_VOX` / `VESKSLOP_TTS_VOX_CLEAN` exports plus `{vox}` / `{vox_clean}`
    placeholders for the command provider, and a Radio-kit row in the TTS panel with the checkboxes
    and a help tooltip. VOX disables word mode in the panel; the wrapper keeps its documented
    precedence. No other MicroslopSam work is outstanding: the engine branch is already an ancestor
    of main, and the remaining contract note (HTTP `/voices` discovery) is not needed for the
    command-provider path in use.
  - [x] C61 Media browsing performance. The grid's merge step rescanned every scrape database per
    request (about 300 ms for a page, 250 ms for stats, both refetched on every scroll append).
    `MediaService` now keeps a candidate cache keyed by a scrape revision plus the normalized
    filter shape (paging and empties dropped): one scan per filter set, then ~5 ms per page, and
    stats share the scan. The client skips stats when the filters are unchanged (60 s TTL), updates
    only the pager instead of rebuilding the toolbar on page loads, and renders tiles with
    `content-visibility`/`decoding=async`/low fetch priority. Measured: 302 ms first page, 5 ms
    later pages; smoke test streamed 2,480 tiles with no console errors.
  - [x] C60 Incremental re-scrapes and archive cleanup. A new scrape of a channel that already
    has a completed walk now starts after that walk's newest message (the best completed scrape by
    message count), so it only covers the newer gap instead of re-fetching history; the previous
    newest is stored as `afterId` and shown as the from-date. Live-verified: a re-scrape of
    `559354230444654605` started after `999789325183090769` from the 2026-09-17 scrape. Deleted the
    four pre-pagination-fix test scrapes (superseded by full walks) and the truncated DM scrape
    with no complete counterpart; `pnpm verify:scrapes` now reports 10 complete, 0 superseded,
    0 incomplete.
  - [x] C59 Scrape names and the completeness check. Scrape listings (History, Clean media,
    Bot-tab job/lane cards) now lead with the first two words of the channel or group-chat name
    plus the scrape id. New `poc/scripts/verify-scrapes.ts` (`pnpm verify:scrapes`) checks every
    index: done state, no lastError, messages non-zero, messages vs the Discord search total, and
    the oldest message against the source's snowflake creation time (60-day grace). Scrapes a
    later full walk supersedes are reported as `old`, not failures. Result: 10 complete, 4
    superseded, 1 genuinely incomplete — the early DM test scrape
    `805255536223715370-20260916215657` (10 messages, pagination stopped before the fix, never
    re-scraped).
  - [x] C58 Titles from Discord's embed cache. Gone YouTube videos have no oEmbed answer, but
    Discord keeps the embed title (and, with the rebuilt core bundle, the author/channel). The
    title lane now asks the source message's embed first and only then the provider oEmbed; the
    core maps `authorName` from `embed.author.name`, extraction stores it as `title_author`, and
    the viewer shows a prominent title/channel heading plus "Video unavailable (removed or
    private)" when a YouTube row has neither. The 783 title errors were reset and the lane
    backfilled them: the reported example (message 703788394337009754) now reads "Gun Test: Firing
    Empty Shell's Handheld XM556 Microgun". The channel name populates after the Vesktop client
    restarts on the new bundle (the client is externally started).
  - [x] C57 Page mode. The media grid has an Infinite/Page flippy toggle in the bar. Page mode
    renders one `pageSize` slice via `offset = (page-1)*pageSize`, shows Prev / editable page
    number / total / Next, persists the page in localStorage, and scrolls to the next page one at
    a time with a 2 s cooldown after each load. Same filter + page is stable while the archive is
    unchanged. Infinite mode keeps the stream buffer; prefetch, buffer fill, and the buffer are
    disabled in page mode. Page mode pre-loads the next page's JSON and preview images, so
    jumping one page renders instantly.
  - [x] C56 Recovery completed. First run after the C55 fix recovered 1,117 of 1,240 Discord-CDN
    previews (90%); the last 123 were confirmed genuinely gone (refresh + direct fetch and `/stream`
    both fail) and are marked `unavailable after recovery` so later runs skip them. The recovery
    query now only covers HTTP errors on refreshable URLs, so proxy `no preview source` rows are
    not swept. Archive: 26,725 cached previews, 7 pending, 1,621 terminal provider failures, and the
    recovery task reports 0 to retry.
  - [x] C55 Recovery fix. The first recovery run parked live files because refreshed embed URLs
    land in `play_url` while preview resolution kept using the stale `url`; the engine now resolves
    embed previews from `play_url` first and re-resolves from the refreshed URL on retry, and both
    refresh branches canonicalize `url` to the direct media link. The task also counts rows that
    park as suspects as `deferred` (not gone), tries each row once per run instead of looping, and
    initializes the slopbot so guild scrapes refresh through the bot. The 192 suspects from the
    broken run were converted back to unverified so they are retried.
  - [x] C54 Message links, suspect retries, and preview recovery. Items now carry `messageUrl`
    (from `media_sources` + scrape guild), the viewer shows/copies it instead of the expiring CDN
    link and opens a fresh stream for View full, YouTube items offer an **Open on YouTube**
    fallback, and the metadata gained Channel/Scrape/Title (title includes the channel kept from
    oEmbed, `title_author`). Preview and download HTTP failures on Discord hosts first refresh the
    URL from the source message and retry; if that fails the row parks as a **suspect** with one
    delayed retry (50 s, runs after the batch; a batch two-thirds suspect doubles the cooldown,
    60 s floor). Still-failing rows are **unverified** and skipped by normal passes; the Bot tab
    shows suspect/unverified counts and a segmented progress bar (cached blue, pending gray,
    failed red). **Recover Discord previews** is a paused one-shot task that re-fetches every
    Discord-CDN row closed with an HTTP error across all scrapes (mixed bot/puppet via the
    existing refresh path) and lands in the task history when done. All 1,208 Discord 404 rows
    sampled 15/15 recoverable.
  - [x] C53 Title catch-up as its own lane. Titles no longer share the preview batch: a dedicated
    `TitleLane` walks every scrape with a pending title, 100 rows per batch with a 10 s pause
    (hard-coded; it is a one-time catch-up and future titles arrive with the scrape), and takes no
    media-traffic lock because it only calls YouTube's oEmbed. New `titleBatch` on the engine,
    per-scrape `titlePending/titled/titleFailed` counters, `/api/media/titles` + start/pause
    routes, and a Bot-tab card with progress, throughput, ETA, and Start/Pause. The lane
    auto-starts on boot; with nothing pending it exits immediately. Preview backfill goes back to
    previews only. Unit tests for `titleBatch` and a lane catch-up run; smoke covers the card.
  - [x] C52 Wildcard title search. `*` now maps to a LIKE wildcard in both title modes: a leading
    `*` drops the word anchor (`*car` finds *scar* and *discard*), in-token stars match between
    letters (`c*r*`), and a trailing star matches the plain prefix. `%` and `_` are escaped with an
    `ESCAPE` clause so they only match themselves. Unit tests cover prefix vs wildcard vs in-word
    matches; the drawer hint documents the syntax.
  - [x] C51 YouTube Shorts normalization. Shorts links broke the embedded player and View full
    because the viewer needs `watch?v=`. Extraction now rewrites `youtube.com/shorts/<id>` (and
    `m.`) to the watch form, the viewer also accepts `/shorts/<id>` defensively, and a one-time
    `PRAGMA user_version` migration rewrites existing rows. The live archive had 45 shorts rows;
    one collided with an already indexed watch URL and folded into it (preview fields merged,
    message links moved) — the migration is idempotent and unit-tested, including the collision.
    45 → 0 shorts rows, and the user's example `uAf18d7mrnU` is now
    `watch?v=uAf18d7mrnU` with `embed/uAf18d7mrnU` as the play URL.
  - [x] C50 Title backfill in the queue. YouTube-like rows (provider or URL matching
    youtube/youtu.be) without a title are now a second work type in the backfill queue: resolved
    through the provider oEmbed, stored as `title`/`title_key`, with `title_attempts` and
    `title_error` so deleted videos close out after the tries setting. Titles and previews share a
    batch (up to half for titles while both remain); `previewCounts().titlePending`,
    `BackfillMember.pendingTitles`, and the job totals carry the work, and the Bot tab member line
    shows `+ N titles`. Unit tests cover selection, resolution, and the give-up cap. Live: the
    user's queue (batch 40, 30 s, auto off) started once to verify — 133 titles resolved in two
    minutes, dead videos counting toward the cap.
  - [x] C49 Searchable video titles. Embed rows now store Discord's `title` plus a normalized
    `title_key` (lowercase, punctuation collapsed). The Advanced drawer has a title field with an
    **Exact match (quotes)** toggle: default is order-free multi-word prefix search ("Ma Ca Ju"
    finds *Man who jumps over car*), exact mode is one phrase; rows without a title drop out while
    the query is set. Store, `/api/media`, `/api/media/stats`, and MCP carry `title`/`titleExact`;
    the API result and card tooltip expose the title. Unit tests cover prefixes, mid-word misses,
    exact phrases, and punctuation folding; smoke checks the field and toggle. Note: titles only
    exist for rows scraped from now on — pre-existing rows need the backfill queued in TODO.md.
  - [x] C48 Expanded providers and the Must include mirror. The list grew to 19 entries from the
    user's picks: reddit, twitter, tiktok, instagram, streamable, facebook, gyazo, a merged
    `giphy,tenor` (one checkbox, two substrings), 2ch.hk (stand-in for 4chan; the CDN hosts stay
    behind the default URL exclusion), funnyjunk, and internet archive. The drawer gained a
    **Must include** block mirroring **Do not include**: provider/source-kind checkboxes
    (`includeProviders`, `includeSources`), file-type chips, and the size/length bounds (moved here
    since they are inclusions, not exclusions). Store, `/api/media`, `/api/media/stats`, and MCP
    all carry the two include params. Live-verified: `includeProviders=youtube` returns only
    YouTube rows, `giphy,tenor` includes/excludes the same rows complementarily; smoke checks both
    groups against the API list (22 = 19 + 3 kinds).
  - [x] C47 Provider list from the dataset. Aggregated the `provider` column over all 11 scraper
    databases (distinct URLs): discord 13631 (folding cdn.discordapp.com and media.discordapp.net),
    youtube 4840, tenor 352, twitch 196, twitter 111, steam 83, medal 59, sketchfab 31, gfycat 30,
    imgur 23. Those ten now live in `PROVIDER_FILTERS` (media-scrape), are served in
    `/api/media/status` as `providerFilters`, and the Advanced drawer renders the source checkboxes
    from that list (4chan stays out; the default URL host exclusion covers it). Unit test for the
    list shape and a smoke check that the checkbox count matches the API list (13 = 10 + 3 kinds).
  - [x] C46 Sort and exclusion filters. The Advanced drawer gained a sort select (newest, oldest,
    user, size, length; global over the merged URL-dedupe stream) and a **Do not include** group
    with nested collapsibles: sources (Twitch, YouTube, Discord, Tenor, Giphy, embeds, attachments,
    stickers), file types, size (less/greater than with B/KB/MB/GB units), and length (seconds).
    Store, `/api/media` + `/api/media/stats` (shared parser), and MCP `query_media` all carry the
    new params; size and length ranges drop rows with unknown values, and the ext matching now
    coalesces NULL content types so a `NOT (...)` cannot swallow rows. Also hid the media filter bar
    and drawer on the Bot tab. Unit tests for every exclusion, all five sorts (nulls last), and
    smoke checks for the hidden bar on the Bot tab.
  - [x] C45 Streaming media grid. The grid rebuilt every card on each appended page, which made
    filtered browsing flash, and the `Load more` trigger only worked once the grid could already
    scroll, so short filtered pages dead-stopped at ~11 cards. Now pages append incrementally
    (existing images are never rebuilt), a stream buffer capped at 1.5 viewport heights keeps the
    scroll alive while the next page loads, auto-fill cascades while the buffer is on screen, the
    buffer note reads `Loading more…` / `Scroll for more` / `End of results`, and an elastic nudge
    replaces the hard stop at the bottom. Smoke test: first load reaches 120 cards via auto-fill,
    scrolling appends 120 → 440 with no broken images.
  - [x] C44 One traffic pipeline and split ETAs. Walks and backfill batches now share a
    media-traffic lock: a walk holds it for its whole run, a batch for one batch, so the pacing
    settings bound total traffic instead of each action. Starting then immediately stopping a
    scrape now fetches nothing (tested). Backfill ETAs split into `etaStartMs` (waiting behind
    earlier jobs), `etaRunMs` (own work), and `etaTotalMs`; the Bot tab shows only the total and
    puts `starts in X · runs for Y` in the tooltip, same for the queue totals. Tests: lock
    ordering, stop semantics, ETA ordering.
  - [x] C43 Bot-tab backfill control room and media-tab strip rules. The engine gained
    `previewBatch`/`rateLimitStatus`/`updateOptions`; a new `BackfillQueue` groups preview work per
    chat (paused and resumed scrapes of a chat combine), runs one job at a time in batches, and
    reports pending/cached/failed per kind, throughput, measured and settings-based ETA, and the
    live cooldown. `MediaService` drives it and exposes `GET /api/media/backfill` plus start/pause;
    `/api/preferences` now carries `media.tunables` (Backfilling / Scraping / Timers, persisted and
    applied live). The Bot tab shows totals with a Bot/Puppet split, per-chat job cards with
    progress and Start/Pause, **Start backfill** / **Pause all**, the red `API timed out, increase
    timers` banner, and the three settings groups. The media strip shows running plus one paused
    scrape, dismisses a paused one on Stop without stopping it, turns red while limited, and shows
    a session-only `N scrapes completed` note. The old `VESKSLOP_MEDIA_*` tuning env vars are gone.
    New unit tests cover batching, the cooldown, live retuning, and queue grouping/finish/pause;
    the Playwright smoke test covers the Bot tab. Also fixed a real stall the queue exposed:
    `pendingPreviews` never returned thumbnail-less embeds, so they sat pending forever and stalled
    the job; the store now includes them and the resolver handles Discord CDN unfurls through the
    proxy frame, with a content-type guard so the og:image fallback never reads media bytes. Live:
    daemon restarted, pending falling (6018 → ~5900), cached rising, dead CDN unfurls closed as 404,
    jobs queue smallest-first, settings round-trip verified (21.4 → 37.5 rows/min at 30 s cooldown).
  - [x] C42 Preview attempts cap. The restart log showed the backfill retrying dead hosts (Gfycat
    DNS is gone, Streamable answers 400) forever as "transient". Rows now carry
    `preview_attempts`; after `VESKSLOP_MEDIA_PREVIEW_MAX_ATTEMPTS` (3) the row is marked
    `unreachable after N attempt(s)` and leaves the pending queue. Unit-tested.
  - [x] C41 Restart keeps its logs. A button restart spawned the replacement with `stdio: ignore`,
    so anything after the restart vanished. The replacement now appends to `daemon.log` beside the
    token file (`VESKSLOP_DAEMON_LOG` overrides); documented in [docs/daemon.md](docs/daemon.md).
  - [x] C40 Stale scrape states. Two `general_chut` scrapes showed `running` because the daemons
    that owned them were killed mid-job; the state file kept the last written value. On daemon
    start the engine now reconciles `running`/`queued` scrapes to `paused`, and History has a
    per-scrape **Delete** button (clean with a single ID) next to Resume. Unit-tested.
  - [x] C39 File-type filter and the 4chan default. The Advanced drawer gained extension chips
    (video mp4/webm/mov/mkv, image png/jpg/gif/webp/avif, audio mp3/ogg/wav/m4a/flac) matched on
    URL or content type, plus a default-on **Hide 4chan links (usually dead)** toggle. Backend,
    API (`ext`, `excludeHosts`) and MCP support; live: video count 7696 → 7383 with the default
    exclusion, `ext=mp4`/`webm` return only those files. Also confirmed the mr-game-memorial group
    has 924 attachment videos with 772 previews cached and 0 preview errors — the rest are just
    pending in the 2,140-item backfill queue.
  - [x] C38 Filter layout swap. The common filters (filetype chips, source kind, server/DM
    origin, multi-select People popover) now sit in the always-visible bar; the Advanced search
    drawer holds aspect, resolution, channel, date range, cached-only, and the card-size slider.
    Apply/Reset stay in the bar. Fixed a dropped `defaultFilters` helper found by the Playwright
    smoke test, which passes on the new layout.
  - [x] C37 Advanced search drawer, adaptive paging, and zoom. Filters moved into an **Advanced
    search** drawer with Apply/Reset (no auto-update): filetype chips, source kind, origin
    (server/DM), multi-select People, and an Advanced section (aspect, resolution, channel, dates,
    **cachedOnly**). A **Card size** slider scales tiles from 100 px (2/3 of current) to one image
    per row; page size = full rows about 1.5 screens tall and follows zoom/resize. Infinite scroll
    preloads the next page's JSON and previews. Backend gained `authorIds`, `origin`, `cachedOnly`
    on `/api/media`, `/api/media/stats`, and MCP `query_media`. Verified: `origin=dm` and
    `cachedOnly=1` return the expected rows.
  - [x] C36 Daemon restart, preview diagnostics, browse-time caching, and rate-limit pauses.
    Settings → Server has a **Restart daemon** button (`POST /api/daemon/restart`: re-execs the
    daemon, keeps the client running). The preview pass logs every failure with scrape/item/kind/
    provider/reason plus a per-pass summary, and `GET /api/media/preview-failures` lists them.
    Browsing no longer fetches Discord at all: the grid uses cached stills only (placeholders for
    missing ones), GIF hover plays the cached animated GIF (kept as `anim_hash`, served from
    `/anim`), and streaming only starts on explicit play. `429`s from previews/downloads raise a
    cooldown (`VESKSLOP_MEDIA_RATELIMIT_PAUSE_MS`, default 60 s, or `Retry-After`), leaving rows
    pending and pausing both walker and previews.
  - [x] C35 Walk/preview cycling. A live scrape now alternates: walk 5 pages, preview up to one
    batch (25, with the cooldown), walk on. Tiles fill in during a long walk instead of only at the
    end, and preview traffic is spread through the walk. Configurable with
    `VESKSLOP_MEDIA_WALK_CYCLE_PAGES` (0 defers previews to the end). The final pass still drains
    whatever remains.
  - [x] C34 Resume from cursor. A paused scrape can continue where it stopped instead of
    re-walking the top: `POST /api/media/scrape/resume`, MCP `resume_media_scrape`, and a Resume
    button in History. Previews never needed the walk — backfill fetches them straight from the
    stored URLs; the walk is only for discovery. Live: the paused `gernal_chut` scrape resumed at
    cursor `365256427247239181`, fetched its final page, and entered the preview pass.
  - [x] C33 Puppet history pagination. The core's `message.fetch` read the client store, which
    keeps only the newest messages plus the channel's first — every page was 10 messages and the
    walk stopped after one page (a 10-message, 4-media scrape of an ancient group with thousands
    of messages). It now fetches through the client's REST layer first (contiguous 100-message
    pages, `transport: "rest"`), store as fallback. Rebuilt and restarted the managed client;
    live re-run of channel `591000745164406784`: 21+ pages / 2,100+ messages and climbing, with
    hundreds of images/videos being indexed.
  - [x] C32 Group-chat clarity and puppet refresh. Group DM scrapes already worked (live-verified
    against the running daemon: `zıbb & fred` and a 1:1 DM, backend `puppet`, done, previews
    cached); the Group chat tab had shown an empty list while the core was offline. It now says so
    explicitly and offers Retry. Expired attachment URLs in DM/group-DM scrapes refresh through
    the core bridge (`message.fetch` with `before: id+1`, which returns exactly the target).
  - [x] C31 Batched previews. The preview pass now runs in batches (default 25) with a cooldown
    (default 60 s), configurable via `VESKSLOP_MEDIA_PREVIEW_BATCH` and
    `VESKSLOP_MEDIA_PREVIEW_BATCH_PAUSE_MS`; batching counts rows, not successes. Live log:
    `preview batch of 25 done, pausing 60000 ms`.
  - [x] C30 Preview completeness. Audio attachments and voice messages get an ffmpeg waveform
    strip; oversized animated GIFs fall back to an ffmpeg-read first frame instead of `too large`;
    transient network failures stay pending for the next backfill instead of becoming permanent.
    Live: a scrape with 58 audio items backfilled 13 waveforms in 40 s, serving `image/webp`.
    Remaining non-cached: Lottie stickers/custom emoji (not indexed as media) and embeds whose
    pages refuse bots (marked PRIVATE/GONE by their 403/404).
  - [x] C29 Attachment video previews. Discord's image proxy renders a frame for video files
    (`?width=320&height=320&format=webp`), so `previewUrlFor` now resizes attachment mp4/webm in
    links mode too and `pendingPreviews` includes non-embed videos. Live: paused-scrape videos
    backfilled to `image/webp` thumbs, pending queue drained to 0.
  - [x] C28 Media grid usability. Infinite scroll (240 px threshold) with a Load more fallback,
    `previewError` exposed so tiles show PRIVATE (403) or GONE (404/410) with the reason in the
    tooltip, and the scrape strip lists only queued/running jobs — paused scrapes move to History.
  - [x] C27 Non-media stream guard. Rows awaiting backfill fell back to first-frame `<video>` with
    a provider page URL (YouTube embed, redirectors), and the stream route piped the HTML through
    (`text/html` errors). Now only real video files (mp4/webm/m4v) are framed, the stream route
    returns `415` for non-media upstreams, the viewer's still-frame is gated the same way, and
    "Open in new tab" prefers `play_url`. Live: YouTube item → 415 JSON; backfill reached
    184/271 previews during the run.
  - [x] C26 YouTube preview root cause. Dedupe picked the newest row per URL; a paused scrape
    (`gernal_chut`, stopped mid-walk) had no previews yet and hid an older scrape that did. Fixes:
    dedupe prefers the richer row (downloaded > preview > newer), paused walks still run the
    download/preview passes for collected pages, daemon start backfills previews for scrapes that
    paused early, and preview gaps are now 200-600 ms. Live: backfill cached 24 previews in the
    paused scrape within 45 s and a YouTube item then served `thumbUrl`.
  - [x] C25 Live scrape progress. The media tab polls `/api/media/scrapes` every 4 s while open,
    so the running row counts update; pages show `X/Yp (~Z%)` from `estimatedMessages` when the
    search total is available (some channels deny the bot search with Missing Access, so it stays
    optional). Live: pages/messages incremented 2→10 during a walk on `gernal_chut`.
  - [x] C24 Embeddable items open straight into their player: clicking a YouTube/Vimeo/Twitch/
    Streamable card shows the iframe immediately, no extra play press. Non-embeddable videos keep
    the still + play icon.
  - [x] C23 Grid and viewer source fixes. The grid now prefers the cached still (or local file) for
    every kind, and only falls back to a first-frame video, served through the same-origin stream
    endpoint instead of hotlinked CDN/redirector URLs (fixes fxtwitter/YouTube previews that ORB
    blocked while the cached thumb sat unused). The viewer shows a muted first frame for videos
    without a still, and stream refreshes the signed URL only when the stored one fails (4xx). Live:
    stream returned 206 video/webm for a thumbnail-less video; YouTube rows keep their cached thumb.
  - [x] C22 Media grid polish: per-kind status row
    (`(Images/Videos) | Downloaded I/V | Cached I/V | Loaded I/V`), GIF tiles with a translucent
    GIF label that play the animated original on hover (cached previews are static ffmpeg frames),
    Tenor oEmbed preview resolution, and first-frame/inline playback for embeds that carry
    Discord's `videoUrl` (Tenor/gifv now classify as video with `play_url`). `remotePreviewUrl`
    exposes the animated source to the UI. Live: stats per kind correct; Tenor embed stored with
    `play_url`.
  - [x] C21 Playback of non-downloaded and expired media. `play_url` stores the embed's direct
    video (gifv/Tenor mp4); a new `/api/media/<scrape>/<item>/stream` route serves the local file
    or re-fetches the source message through slopbot to refresh expired signed CDN URLs and streams
    it with Range support. The viewer and **View full** use it, so Discord attachment videos play
    without a prior download and old links no longer dead-end. Live: 2.5 MB webm streamed 200 with
    a refreshed URL; play_url populated for a media.discordapp.net gifv mp4 and provider
    redirectors.
  - [x] C20 Cross-scrape dedupe and viewer polish. `queryMedia` and `stats` deduplicate by URL
    (newest scrape wins), so overlapping scrapes no longer duplicate cards or counters (live: 17
    items / 17 unique). Viewer: big play icon top-right; play embeds a provider player or a direct
    mp4/webm/m4v, otherwise opens the external warning; **View full** opens the embedded
    media/player in a new tab while the link button opens the original page; duration is stored
    (Discord attachment duration, ffprobe for downloads, 0 normalized to unknown); favicon added.
  - [x] C19 External embed previews and MIME fixes. Embeds with no thumbnail in Discord's data are
    resolved via provider oEmbed/fxtwitter and an `og:image` fallback; preview bytes keep their
    real extension (`preview_ext`), so JPEG thumbnails serve as `image/jpeg` instead of mislabeled
    WebP (the browser "Content-Type image/webp is not supported" errors). The grid no longer uses
    embed page URLs as media, and first-frame `<video>` is limited to playable direct files
    (mp4/webm/m4v) with an on-error placeholder. Live: fxtwitter/vxReddit/TikTok previews cached;
    7 jpg + 4 webp thumbs served with the right types.
  - [x] C18 Video previews and the status row. Uploaded videos have no Discord thumbnails: the
    grid renders a first-frame `<video>` for them, and downloads generate a poster frame with
    ffmpeg into `thumbs/` (live: both video attachments in the test channel got posters). The
    status row is now `(Images/Videos) | Downloaded A/B | Cached C/B (excl Downl) | Loaded L/B`
    with B = total for the active filters.
  - [x] C17 Media grid and viewer. Embed videos now render their cached thumbnail with a play
    icon; a Metadata dropdown picks banner fields (User, Type, Date, Source, Resolution, Length);
    nothing shows until Load media; the status row reads `downloaded A/B | cached C/D (excludes
    downloaded) | N loaded` from `/api/media/stats`; clicking a tile opens a viewer with
    wide-below / vertical-right metadata, full link with an external-site warning, View full, and
    an embedded player for YouTube/Vimeo/Twitch/Streamable when possible.
  - [x] C16 Message counting and date-bounded scrapes. Discord search (`total_results`, one guild
    request, slopbot only) powers `POST /api/media/count` and MCP `count_media_messages`; the
    confirm dialog has From/To dates plus a **Check Message Count** button; the walk honors
    `afterId`/`beforeId`; `estimatedMessages`, `fromDate`, and `toDate` land in the scrape meta and
    History. Live: count 17 all / 17 after 2026-09-16 / 0 before, DM returns a reason, and a
    from-dated scrape stored 17 messages with `estimatedMessages: 17`.
  - [x] C15 Scrape UX: custom download destination (absolute path per scrape; files there,
    previews local, embed thumbnails copied to both, custom files not served), confirmation popup
    before Links/Download, New scrape/History right-aligned with New scrape furthest right, modal
    X and backdrop close, active-only progress strip, and a Bot error feed with a red unseen badge
    that clears on open plus a Clear button. Live run `...-20260916095029`: 6 files and 7 embed
    thumbnails in `/tmp/opencode/media-dest/<id>/`, 11 local previews, 0 local files, `localUrl`
    null; error record and clear verified over the API.
  - [x] C14 Links-mode embed thumbnails, item metadata, filters, and the scrape popup. Metadata:
    poster (`author_id`/`author_name`), `posted_at`, `width`/`height`, `aspect_ratio`, with an
    `ffprobe` fallback after downloads. Filters: poster, aspect preset (vertical/square/4:3/16:9/
    21:9/other), resolution preset (SD/HD/FHD/QHD/4K+/unknown, by shortest side), and a
    posted-date range. UI: scrape popup with Server and Group-chat trees (server -> channels, each
    with Links/Download), History popup (time, duration, pages/messages, media, bytes, files,
    previews), live progress pulse, full-window view. Live run `...-20260916094120`: 11 previews
    in links mode (embed thumbnails included), poster zzib, a 9:16 short stored as 0.563, and the
    16:9 filter matched.
  - [x] C13 Embed thumbnails in download mode: embed rows cache their stored thumbnail URL;
    the full third-party media is never fetched. Live run `...-20260916091458` on the test
    channel: 6 files downloaded, 11 previews cached (YouTube, Reddit, Tenor, TikTok, X). Also
    fixed the state transition: the job stays `running` through download and preview passes so
    clients do not see `done` early.
  - [x] C12 Media tab filters and full-window layout: kind chips, source filter
    (attachment/sticker/embed), source-channel filter, and the view now fills the window below the
    header and call strip. API/MCP gained `source` and `channelId` on `query_media`. Smoke-tested
    over `/api/media?source=` and `?channelId=`.
  - [~] C9 live verification. Slopbot path verified live on 2026-09-16: login Slopbot#5775, guild
    `177931418704281600`. Test channel links scrape: 17 messages, 4 images + 13 videos, 0 errors;
    download scrape fetched the 784 KB attachment. Large channel `gernal_chut`: 11 pages, 1088
    messages, 96 images + 1 GIF + 203 videos, 0 errors (no 429 at the current pacing).     Puppet path
    verified live on a group DM (`zıbb & fred`) and a 1:1 DM: backend `puppet`, done, 0 errors.
    The puppet field check is done too: the group DM returned attachments plus mapped embeds
    (YouTube, Instagram), and the large slopbot channel produced YouTube, Gfycat, Imgur, Steam,
    and Twitch embeds alongside attachments. Still owed: a stress walk on the second guild
    (`559346566754729995`, channel `559346566754729997`).
- Flextalk voice profile, end to end. The profile window is a movable panel that recenters on each
  open; it carries the character sliders with engine defaults and +/-2 snapping, five named
  presets, radio-kit effects (high-pass 40-400 Hz, echo 0-100, word mode with a gap in ms), and
  test lines (Attention, Laboratory, Critical, Health, Radiation, plus two custom lines in
  `localStorage`) that post the current profile straight to `/api/audio/speak`. Fixed the test
  buttons reading a stale expression from saved state; they now read the slider like Speak does.
- Root cause of the profile never reaching the engine (2026-09-15): the wrapper wrote the Flextalk
  `.reg` with `echo`, and dash's `echo` treats `\2` in `...\Watson\2.0\...` as an escape, which
  corrupted the key so `wine reg import` was a silent no-op (it still returns 0). The wrapper now
  uses `printf`, validates that every profile value is numeric, and rejects a `.reg` containing
  control characters. Documented in TTS-DESIGN.md.
- The Veskslop agent lives in `agent/` (Go 1.25+, Eino v0.9.19, remote llama.cpp over the
  OpenAI-compatible API, daemon tools over MCP Streamable HTTP). It is disabled by default
  (`VESKSLOP_AGENT_ENABLED=1`), takes tasks on `POST /task`, and logs every tool call. Smoke-tested
  on 2026-09-14: 28 tools loaded, one ReAct round, final message returned. The daemon does not push
  tasks to it yet. Plan and env vars: [docs/agent.md](docs/agent.md).
- Remote agents use the daemon's MCP endpoint directly (query state, send notifications). A
  notification tool set and a documented LAN/Tailscale bind are planned.
- Mobile web UI first draft: four tabs (Users, Chat, Call, Activity), one pane at a time, 44px
  targets, tab badge, call bar pinned, Activity becomes the full console. Verified headless at
  390x844. Next: transcript and TTS box in the Call tab, per the 1c design.
- Fixed a pre-existing web UI bug: `loadDms` called the nonexistent `renderDirect`, so the DM rail
  only rendered from other paths. It now calls `renderDmRail()`.
- Fixed the speaking marks: the core turned Discord's numeric `speaking: 0` into `true`, so the
  equalizer bars never stopped. The bars now fade in on start and decay out on stop (500 ms), with
  a 2.5 s idle timeout for missed stops; a mute or channel change ends the state at once. Muted
  members show no bars.
- Browser audio: **Client audio** now runs the browser unlock in the click (microphone permission,
  AudioContext created and resumed, output sink applied) before it starts listening. Verified
  headless: 52 PCM frames played through a running context. The button label no longer wipes the
  icon markup.
- Capture routing: nothing moved the client's playback streams into `veskslop_capture`, so the
  browser monitor streamed digital silence. The audio manager now adopts matching streams
  (`VESKSLOP_AUDIO_CLIENT_MATCH`, default `vesktop,electron,chromium,discord`) on device setup,
  monitor start, and every 5 s while listening or capturing.
- LAN audio verified end to end on 2026-09-14: a tone stream on `veskslop_tts` was moved to
  `veskslop_capture` (daemon log: `moved stream 81 to veskpop_capture`) and the browser, connected
  to `https://192.168.0.178:18801/`, played 256+ PCM frames with a peak of 4096 (a stream already
  on the capture sink peaked at 29274).
- Speaking bars are always in the layout now: `scaleY(0)` and transparent when idle, they scale up
  with the fade-in at start, run the equalizer animation while speaking, and stay on until an
  explicit stop, a mute, or a channel change. Live data showed repeated `speaking: true` pulses and
  no stop events inside an 8 s window, so the idle decay is removed: bars stick rather than
  disappear. The fade runs the equalizer inside a shrinking container, so the amplitude drops
  smoothly instead of the block collapsing at full height. The core treats a missing `speaking`
  flag as a start pulse and accepts `user_id`. Core-side changes need a rebuild plus a client
  restart.
- Added `poc/scripts/audio-doctor.sh` (`pnpm audio:doctor`): default sink, sink inputs, audio
  status with adoption counters, a 5 s `volumedetect` on `veskslop_capture.monitor`, and a live
  event peek. Run it in a real call to trace a silent browser monitor.
- Live doctor output (2026-09-14): the `vesktop`/`electron` streams sit on `veskslop_capture`
  (sink 40) already, `adopted: 0`, and `VOICE_SPEAKING` events arrive. Client audio routing is
  therefore in place. Bot TTS works. Capture was off (`capturing: false`), so the transcript was
  empty; Vosk transcribes correctly when invoked with the model env.
- Live call audio works now (the user confirmed TTS is audible and the silent call was partly a
  per-user volume set to zero for some members, plus another issue). The `volume` read returns
  `null` because this client build exposes no `UserVolumeStore`; the setter reports
  `NOT_SUPPORTED`. Volume is not a usable control here.
- External TTS engine support (external project, first target is a command wrapper): the contract
  lives in `TTS-DESIGN.md`. Veskslop now has nullable `tts.speed` (0.5-2.0) and `tts.pitch`
  (-12..+12) preferences, per-line overrides on `/api/audio/speak` and MCP `speak`, `{speed}`/
  `{pitch}` placeholders, `VESKSLOP_TTS_SPEED`/`_PITCH` exports, and speed/pitch sliders in the TTS
  panel. A custom command advertises its voices through `VESKSLOP_TTS_VOICES`; the espeak list is
  hidden while a custom command is set. Verified with a wrapper that logged argv and env:
  preferences and overrides both arrive. Reading the engine's `GET /voices` for the HTTP provider
  remains future work.
- The MicroslopSam engine project is now part of the tree in `microslopsam/` (copied from the
  separate project; its `build/` output and voice data stay ignored). The wrapper was renamed
  `setup/vekslop-tts.sh` -> `setup/veskslop-tts.sh` and every reference and configuration export
  was fixed. The engine also corrected the Flextalk pitch finding (REG_DWORD bug), so Keith Bell
  and Kerry Watson now support pitch and carry no `no-pitch` flag.
- Implemented the Flextalk Expression profile parameter end to end: `VESKSLOP_TTS_VOICES` marks the
  two voices as `expression=59` / `expression=100`; the wrapper writes the registry value before
  synthesis; Veskslop has `tts.expression` (0-100 nullable), the `{expression}` placeholder,
  `VESKSLOP_TTS_EXPRESSION`, an HTTP/MCP field, and an Expression slider that appears only for
  those voices (default shown until moved, null leaves the registry untouched). Verified: 0/59/100
  produce distinct audio, the registry reaches 0x14 for 20, and the UI shows 25/100/default states.
- Fixed the null-value coercion in the preference normalizers (`Number(null) === 0` turned a
  cleared speed/expression into 0.5/0); unset now stays null.
- Renderer-down handling polished: the bridge logs one warning per outage instead of a stack per
  polled command, and the `navigator.languages` request catches the new timeout rejection instead
  of raising an unhandled promise warning. (The `exitCode: 9` renderer kills in the live log were
  client restarts; the daemon's retry handled the startup race as designed.)
- The TTS panel has speed presets (0.5x, 0.75x, 1x, 1.25x, 1.5x) under the fine-adjust slider, and
  the pitch control is hidden entirely for voices that report `supportsPitch: false`. (Flextalk
  voices no longer report that; see the pitch correction above.) Verified: preset clicks update the
  slider/label and save.
- TTS playback has a lead-in of silence before each line (`VESKSLOP_TTS_LEAD_MS`, default 150 ms,
  max 2000, added with `adelay` in `ffplay`; status reports `ttsLeadMs`) so Discord's voice
  pipeline is settled before the first syllable. Verified on the sink monitor: first sound at
  0.02 s with 0 ms and 0.62 s with 600 ms.
- Speech settings have a **Start engine** button that warms the selected engine
  (`POST /api/audio/tts/warm`, MCP `start_tts_engine`): it runs one throwaway synthesis so the
  command engine's Xvfb/Wine/host are up. Verified with MicroslopSam: cold 1732 ms, warm 505 ms,
  and the UI shows "MicroslopSam ready in 835 ms". The warm text is a full sentence because their
  two-pass `loudnorm` fails with `inf` on very short input.
- Startup and crash robustness: `sendRendererCommand` now rejects immediately when the client
  renderer is destroyed or crashed, and times out after 15 s, so a dead frame no longer logs
  Electron's "Render frame was disposed" stack once per poll. Device auto-connect retries up to
  10 times every 2.5 s, because the bridge socket is ready before the renderer's modules load
  (the live error was `Reflect.get called on non-object` from a lazy lookup at startup). Verified
  with the mock: attempt 1 failed, attempt 2 set the devices.
- TTS engines are selectable: the Speech settings and the TTS panel list espeak-ng, the command
  engine (`VESKSLOP_TTS_NAME`, default "custom engine"), and the HTTP API; unavailable entries are
  disabled and Auto resolves to the command engine first. Selecting an engine switches the voice
  list (`/api/audio/tts/voices?engine=`) and hides the language row for single-language engines.
  The TTS panel now has a voice select next to the speed/pitch sliders. Fixed a slider bug where a
  null speed preference (`Number(null) === 0`) clamped the slider to 0.5 and then persisted it.
- TTS controls: the TTS panel has a working **duck on voice** toggle (stored in preferences,
  reported as `duckEnabled`). Ducking was verified live with the mock: the `veskslop_tts` sink
  drops to 30% while TTS plays and someone speaks, and returns to 100% after. The Transcript tab
  in the console has **Start** and **Stop** buttons; the status line shows `capturing`/`capture
  off`. The espeak-ng voice list now includes the `m1`–`m7`, `f1`–`f5`, `croak`, and `whisper`
  variants (`<voice>+<variant>`), so each language has about 15 voices instead of one.
- Speaking state is now authoritative: the daemon polls the core's `voice.speaking` action
  (`ChannelRTCStore.getSpeakingParticipants`) every second and synthesizes the missing stop/start
  events; dispatcher `VOICE_SPEAKING` hints feed the same set so stale users clear. This fixes the
  bars sticking for everyone and the transcript attribution, which credited the first member
  because speaker intervals never closed. Verified with the mock: a `speaking: true` event cleared
  within one poll.
- The client streams sit uncorked, unmuted, and at 100% on `veskslop_capture`, yet the monitor is
  silent. Gateway `VOICE_SPEAKING` events do not prove inbound RTP, and the TTS path only proves
  outbound, so the remaining suspect is the client's inbound voice state (stuck deafen or a broken
  voice connection), not the sink routing. `/health` now returns `selfId` so the bot's own member
  flags can be picked out of the voice member list; the doctor prints it.
- Client output is still silent at the sink (`-91 dB`) although the streams are uncorked, unmuted,
  and at 100%, and a headless Chromium tone into the same sink measured `-7.4 dB`. Added per-user
  volume reading and writing to find Discord-side silencing: `user.volume` / `user.setVolume` in
  the core, `/api/voice/volume` (GET and POST), MCP `get_user_volume` / `set_user_volume`, and
  `volume` in every voice member. A volume of 0 (or an output volume of 0) silences a user even
  when Pulse shows a healthy stream. `voice.join` also clears persisted self-deafen now.
  The doctor prints output and per-user volumes.
- Voice audio is implemented: daemon-managed virtual devices (`veskslop_tts`, `veskslop_mic`,
  `veskslop_capture`), TTS providers (local command + HTTP API), queued playback through `ffplay`,
  ducking on `VOICE_SPEAKING`, and WAV capture. The daemon starts `pulseaudio` when `pactl` cannot
  reach a server and reports install hints. Verified live on 2026-09-13 after installing
  `pulseaudio`/`ffmpeg`: engine ok, devices created, PCM streamed.
- The Web UI serves self-signed HTTPS (`openssl`, stored under the state dir) so browser media
  APIs have a secure context. `WEBUI_TLS=0` disables it.
- Capture and transcript are implemented: ffmpeg reads the client output, a VAD cuts it into
  utterances, each becomes a WAV plus a transcript entry with the dominant speaker, and a
  pluggable STT hook (local command or HTTP API) fills the text. Live events feed the console
  Transcript tab; `get_transcript` and `/api/transcript` read the store. `VESKSLOP_CAPTURE_AUTO=1`
  records calls automatically. Verified end to end with real PulseAudio on 2026-09-13.
- STT uses Vosk (CPU, light): `poc/scripts/stt-vosk.py` plus a small model in `poc/models/`
  (auto-detected). Verified with a TTS→capture→Vosk round trip. Other providers remain available
  via `VESKSLOP_STT_CMD`/`_API_URL`.
- TTS defaults to `espeak-ng` when installed; engine and voice live in Settings → Client → Speech
  (daemon preferences, so all browsers share them).
- Preferences moved to `<state dir>/preferences.json` and no longer depend on the core, so guild
  visibility saves even while the client is down. API responses send `Cache-Control: no-store`.
- Browser audio has an **Enable browser audio** button that requests the microphone permission,
  which browsers require before listing output devices.
- Browser microphone puppeting (talking through the bot) is next; it needs a mix sink, a `pacat`
  input, and echo handling.
- The web UI follows the Nocturne design handoff (`poc/WebUIMockup.zip`). The 7a shell is
  implemented and rendered headless for review: header, call strip, three-column body, composer,
  and the agent console.
- Takeover is a client-side read-only gate. Off, the composer is locked. Manual sends are logged.
- Core process control is implemented and verified live: the daemon starts the client under
  auto-Xvfb, the bridge connects, and stop exits cleanly.
- Web QR login is implemented. The session reached `pending` against the real gateway; a full
  phone approval is pending.

## Done

- Validated core bridge: protocol, actions, validation, structured errors, hard outbound rate limit.
- Auth persistence: pinned `DATA_DIR`, token file, preload seeding, `VESKSLOP_HEADLESS`.
- Daemon on the core-bridge transport. CDP stays as a development fallback.
- Trusted-only triggers: DMs and direct mentions. Trust tiers: user, trusted, admin.
- Context: 15 messages or 4000 characters, time stamps, file names only, incremental fetch, tmp
  cache with a size limit, fetch cooldown.
- Continuation window: a rushed follow-up joins the request context.
- Response policy: origin channel or DM. Sensitive commands default to DM.
- Persistent action log: triggers, commands, replies, errors, rotation, and secret redaction.
- `poll` is paused (`POLL_ENABLED = false`). The parser and its tests stay.
- Human send pacing: first send of a burst waits 30-200 ms, later sends 500-2800 ms, and the burst
  resets after a random 5-20 s of quiet. Reads take 100-300 ms and do not reset the burst.
- `user.search`: cache and guild-member lookup. `voice.get` searches all guilds without a `guildId`.
- `voice.overview`: occupied voice channels and members for a guild.
- `VOICE_STATE_UPDATE` events: normalized and pushed. Mock-verified.
- Voice actions: `voice.join` (muted), `voice.leave`, and `voice.move` in the core, with daemon
  routes (`/api/voice/*`), MCP tools, and Web UI wiring. Voice actions are paced 250-750 ms and
  written to the action log. The internal lookups were probed on the live client.
- Web UI voice rail: every voice channel is clickable. Clicking joins (muted); clicking the active
  channel leaves. Verified with the stateful mock.
- Web UI startup: a **Start server** button appears left of Takeover while the bridge is down and
  hides when it connects. State reloads retry while the client boots, so guilds appear without a
  manual refresh. Core requests now time out after `VESKSLOP_REQUEST_TIMEOUT_MS` instead of
  hanging forever on a missing reply.
- Settings: "Visible guilds" stays top level; Audio sits under a collapsible "Client" group;
  Core process, Account login, and Action log sit under a collapsible "Server" group.
- Web UI timestamps use the browser timezone by default; Settings → Client → Display can override
  it with any IANA zone. The choice is per browser (`localStorage`).
- Web UI preferences: visible guilds persist server-side in the daemon state file
  (`/api/preferences`), so every browser shares them. The Settings panel has an explicit **Save**
  button; selections filter the UI live and persist only on Save.
- DM list: recipients resolve from IDs through the user store, so names and avatars load.
- Guild menu: fixed the light default button background.
- Web UI: Nocturne design tokens in `poc/public/nocturne.css`, the 7a debug shell, live guild and
  channel rail, messages, DMs, voice membership, agent console with a live event tab, and a
  Settings slide-over for core process, login, visible guilds, and the action log. Roadmap
  surfaces show as placeholders.
- Web UI data: opening a chat fetches the latest 50 messages into the client cache; attachments
  hotlink Discord CDN URLs and are never stored; avatars and guild icons load from the CDN with a
  monogram fallback; the page reloads itself when the bridge connects.
- Web UI drafts: composer and TTS text persist per channel in `localStorage` and survive
  re-renders, reconnects, and page reloads.
- Visible guilds: the Settings panel filters the guild switcher and picker.
- Takeover: client-side read-only switch with `localStorage` persistence; manual sends are written
  to the action log with `type: "manual"`.
- `dm.list`: cached DM and group DM channels, exposed over MCP as `list_dms`.
- Core process control: the daemon starts, stops, and restarts the client, reconnects the bridge
  after a restart, auto-starts Xvfb when `DISPLAY` is missing, and stops what it started on exit.
  Override with `VESKSLOP_CORE_CMD`/`ARGS`; disable Xvfb with `VESKSLOP_XVFB=0`.
- Web QR login: the Settings panel runs the remote-auth session and renders the QR. The token file
  is written on approval and a running managed core restarts with the new token.
- Unit tests cover the pacing policy. `pnpm lint` and `pnpm testTypes` pass at the root.

- Web UI navigation: Guild/DM scope tabs, a DM rail with unread badges (purple/orange, grey while
  Takeover is off), last-message times, favourite stars, and a People list of users seen in voice.
  Favourited users get presence and call tracking, grouped by call in **In voice**. Speaking shows
  level bars and bold names (2c marks for mute/deaf/video/share). The body dims while the bridge
  is down; the Start button becomes Refresh when the daemon is unreachable.
- TTS settings are split into language and voice, grouped by base language with the requested
  languages pinned on top.

- Text notifications are parsed: edits update the context cache, deletes drop from it, trusted
  reactions are logged, and replies are kept in the request context. A reply to a puppet message
  counts as a trigger without a mention. The web UI marks edited messages and replies.
- Voice auto-leave is implemented: when the owner (zzib) leaves the bot's call, the bot leaves
  after `voiceLeaveDelaySeconds` (default 30). A rejoin cancels it. Verified with the mock,
  including cancellation. `stay @user` is deferred.
- Audio verified end to end on 2026-09-14: TTS played into the virtual mic (recording
  `veskslop_mic` peaked at -4.1 dB) and the browser monitor streamed PCM over the TLS WebSocket.
- Next: per-user limits and the collector account. The agent attachment (remote llama.cpp, MCP
  transport, event delivery) waits for a design discussion; the draft prompt has open issues to
  settle first.

## Next (in order)

1. Live verification: `voice.join`/`voice.leave` end to end, `user.search`, `voice.overview`,
   `VOICE_STATE_UPDATE`, and a full web QR login approval on the real client.
2. Agent: daemon task push for free-form triggers (commands stay local), tool allowlist,
   per-conversation serialization, and tests against a fake model.
3. Collector account: read-only modules for voice membership fallback.
4. Per-user rate limit and command cooldowns.
5. Mobile web UI: transcript and TTS in the Call tab; verify on a phone.

## Open discussions (need the user)

1. **Web UI auth.** Accepted: no auth, LAN or Tailscale only. Revisit before any wider exposure.
2. **MCP event delivery.** Decided for the local harness: the daemon pushes tasks to the agent's
   `POST /task`. The transport details (retry, dedupe IDs, queue) are still open. Remote agents
   poll or call tools directly; they do not receive pushes.
3. **Untrusted logging.** Untrusted interactions are ignored at this time. Confirm whether to log
   untrusted DMs and mentions again.
4. **Privacy pass.** Only token-shaped and `mfa.` strings are redacted, and only in the persistent
   action log. A fuller policy (retention, cache clearing, message redaction) is deferred.
5. **Live checks.** `reaction.add`, `user.search`, `voice.overview`, and the voice state event
   normalization are mock-verified only. The QR login session reached `pending` and `cancel`
   against the real gateway; the approval path needs the phone. Confirm on the first real-client
   run.

## Blocked / needs input

- Slopbot: the bot is invited to the test guild and the intent is enabled. The token lives in
  `poc/.env` (gitignored, mode 600) and will be replaced when the project goes to prod. The daemon
  must be restarted to pick it up.
- The original repo `zib/Veskslop` has a broken server-side git setup. Fetches and clones fail with
  `bad pack header` because a script (`/home/git/.sys_health_s3`) runs on the Veskslop read path.
  The admin must fix it. Use `zib/Veskslop-t` (remote `veskslop-t`) until then. That remote has no
  stored credential; use the one-off token URL in `GITEA.md`. Rechecked 2026-09-16: the original
  still fails with `bad pack header`.
- `main` on the original server is `68fc8bb` (zib, "Upload files to docs/images"). It diverged from
  the local `main` (`24447f5`). Reconcile after the server is fixed.

## Where to look

- Backlog: `TODO.md`
- Design documents: `docs/INDEX.md`
- Design handoff: `poc/WebUIMockup.zip` (screen 7a is implemented)
- Agent rules: `AGENTS.md`
- Daemon code: `poc/src/`
- Web UI code: `poc/src/webui.ts`, `poc/public/`
- Core bridge code: `src/shared/puppetProtocol.ts`, `src/renderer/puppet/`,
  `src/main/puppetBridge.ts`