* ai: list custom GGUF files in the Manage tab
A GGUF copied into ai/models/ by hand could be activated by editing
ai.yaml, but never appeared in the Manage tab because the catalog
endpoint only returned CDN/bundled models. The Active Model dropdown
builds from downloaded models, so a custom file was unselectable.
The catalog endpoint now merges any on-disk *.gguf that is not in the
catalog as a `custom` tier model with its real size and no URL. It is
discovered on every request, so a file appears the next time the page
loads; the file refresh is not involved and does not index models.
Catalog models are never duplicated and in-progress .tmp downloads are
ignored. Custom rows read "On disk" and the hint explains the copy-in
path, the tool-calling requirement, and the context_size fallback.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* ai: only list custom GGUFs llama-server can load; reset context on custom select
Review found two gaps in the custom-model listing.
Any *.gguf basename was listed and accepted, but start_llama_server.sh
reads active_model back out of ai.yaml through read_config_value.sh.
PyYAML dumps non-ASCII as escapes the reader prints literally, the
reader strips " #..." before unquoting, and the start script's
traversal guard rejects any ".." run. Those names got a 200 from the
API and then "Model file does not exist" at startup. A shared
is_loadable_model_name check now filters the catalog loop and rejects
the name in manage_settings, so the UI never shows an activation that
cannot start. Directories named *.gguf are skipped too.
Selecting a custom model after a catalog model kept the previous
model's explicit context_size (e.g. 16384 from the 4B) because the
adopt branch was skipped when there was no catalog default. The
default is now always written on model selection, so a custom model
resets to null and the conservative default applies. The hint text
says so instead of claiming the default is used unconditionally.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
A library with a few hundred channels returned 137k characters from one
list_collections call, far past any tool-result budget. The endpoint
now takes limit/offset (clamped like the other AI listings) and a
search_str name filter, and reports total and next_offset; the MCP tool
passes them through and tells the caller how to continue.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The per-kind MCP wrappers (search/search_videos/search_archives/
search_docs, get_video/get_archive/get_doc, get_video_captions/
get_video_comments/get_archive_text) duplicated search_files, get_file,
and read_content and cost an external client ten redundant definitions
on every request. They are gone. search_default_zims folds into
search_zim (zim_id optional), get_inventory_items into get_inventory
(slug optional), and read_file is exposed for the first time. 16 tools,
matching what the built-in assistant sees plus the read-only listings.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Research mode switches to nine kind-generic tools: search_files,
get_file, read_content, search_zims, get_zim_entry, get_inventory,
list_files, read_file, search_places. Tool definitions drop from 20
tools / 8,947 chars (~2,240 tokens) to 9 / 3,881 (~970) on every
request, a sixth of the 8k context the small tier runs with instead of
a quarter. The remaining descriptions are trimmed toward 300 chars.
The replaced per-kind endpoints are removed (four searches, three
getters, three readers, search_suggestions, the slug inventory route);
their tests move onto the generic routes. The research prompt, its
worked example, and the empty-search hint no longer tell the model to
list collections first: channels resolve by name from the library map
that is already in the prompt. list_collections, list_tags, and
get_map_overview stay as endpoints for the MCP wrappers but leave the
chat allowlist; the MCP search_suggestions tool goes since search_files
returns the same matches inline.
scripts/ai_eval.py expects search_files and treats "what channels do I
have" as answerable from context. On the 1.7B model against the dev
library both catalogs score 14/18, but the compact one answers with one
tool call where the old one needed up to five, no longer invents the
example's channel, and keeps kinds straight when asked for videos.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Small models carry every tool definition on every request, and the
per-kind trios (four searches, three getters, three readers) cost a
quarter of an 8k context while giving the model near-duplicate choices.
Three generic endpoints replace them:
- POST files/search: one search over videos, archives, and docs with an
optional kind; channel (name or id), domain, author, subject each
imply their kind and are ignored for other kinds. When the term also
names channels/domains/authors, they are returned as `matches` so the
model can narrow without a second tool. Channels resolve by name, so
the library map already in the prompt is enough: no list call first.
- GET files/{id}: detail by id whatever the kind; videos now report
has_captions/has_comments and the full description, archives keep
their snapshot history.
- GET files/{id}/content?part=text|comments: captions, page text, or
comments, paged; a wrong part for the kind says what the file is.
Inventories take an optional slug on GET /inventories (one tool lists
and fetches). Listings are slimmed: no derivable captions/text links,
no size/mimetype, duration as H:MM:SS, published as a date. The
library map gains the tag names.
wrolpi_mcp keeps every existing tool name; the wrappers now call the
generic routes, and search_files/get_file/read_content are exposed too.
The old per-kind endpoints remain until the modes switch over.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Five new read-only /api/ai endpoints, each with a matching wrolpi_mcp
proxy tool: list_tags (tags with what they are applied to, plus recent
names), search_suggestions (matching channels/domains/authors/subjects
and file/Zim result estimates), list_downloads (queue summary and lean
recurring/one-time entries with relative destinations; status filter;
errors keep their tail, which holds the exception message),
get_map_overview (map files, search-index state, subscriptions, pins
with map links), and search_places (place search with map links and
optional proximity ranking).
Research mode allowlists the library-facing four. Nothing here can
start, stop, or retry a download.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
New read-only GET /api/ai/files/list lists one directory of the media
directory (directories first, then files with size and guessed
mimetype), paged at 100 entries. Every entry carries its relative
path for list_files or read_file. Same guards as read_file via a
shared resolver: no escaping the media directory, and the config
directory is refused and hidden from the root listing.
The agent loop picks the operation up from the spec; library mode
allowlists it. wrolpi_mcp gains the matching list_files proxy tool.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The re-run showed "Is everything running?" parroting the example's
all-clear without checking. The system prompt now states that state is
never known from memory or examples — every health answer starts with a
fresh get_system_status — and the example's answer cites its own
example data (service names, version) so a verbatim parrot is
self-evidently wrong.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The eval's two remaining System misses were the model investigating a
fictional failed wrolpi-kiwix — planted by the few-shot example itself.
The example now shows a healthy status and an all-clear answer; it
still demonstrates status-first without seeding a fake diagnosis.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The first scored eval run (39/45) caught two example defects: a user
asking about the help example's own topic got the example's invented
UI detail parroted back without a search, so the example now carries
true content from the real certificates doc on a topic users rarely
lead with; and list-style answers dropped WROLPi links because the
research example only showed links in prose, so its answer now models
a numbered list with a link per item.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Response-quality work, all aimed at giving the model enough context to
answer accurately:
- Every chat's system prompt now carries a live library map (today's
date, counts, and the exact names of channels, domains, zims, and
inventories, capped and cached 60s). Exact spellings are in context
before the first token, and the list_collections round-trip — minutes
on slow hardware — is usually unnecessary.
- The request is budgeted against the model's real context: oldest
turns are trimmed first, the current question is never dropped, and
an over-budget turn gets a clear "start a new conversation" error
instead of a silently truncated prompt. cache_prompt reuses the KV
cache for the stable prefix, which is most of the latency on a Pi.
- context_size plumbing fixed: selecting a model adopts its catalog
default (the 4B was running at 8k instead of 16k), and llama-server
runs --parallel 1 so the single user gets the whole context.
- temperature is configurable in ai.yaml (default 0.3 — small models
call tools far more reliably near 0.2-0.3 than at server defaults).
- One worked tool-use example per mode is prepended to every
conversation; small models imitate far better than they obey.
- scripts/ai_eval.py scores the assistant against a live WROLPi on
mechanical checks (finished, called the right tool first, included a
WROLPi link, no external URLs) so prompt changes stop being validated
by anecdote.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
On the FX-8320 the first token took over a minute, and Sanic's default
60-second RESPONSE_TIMEOUT killed the stream before anything was sent.
Raise the timeout above the chat loop's own 600s hard ceiling, and emit
a status event before every model call so the stream is never silent —
for proxies and for the user staring at a slow machine.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The portable (GGML_NATIVE=OFF) x86-64 build still assumes AVX2, which
SIGILLs on older CPUs — found deploying to an AMD FX-8320. x86_64 now
builds all CPU-variant backends with runtime dispatch (GGML_BACKEND_DL
+ GGML_CPU_ALL_VARIANTS), the same scheme as llama.cpp's official
release binaries; arm64 keeps the baseline build, which runs on both
Pi 4 and Pi 5. The binary and its dlopen'd backends install together
in /opt/llama-cpp (the loader scans the executable's directory) with
$ORIGIN rpath, symlinked into /usr/local/bin.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Live testing surfaced three ways small models went wrong summarizing a
library: they gave up after one empty search (misspelled channel names
found nothing), so zero-result searches now carry a retry hint and the
Research prompt requires list_collections before channel/site questions
and search_videos(channel_id) before claiming a channel is summarized;
they presented the familiar YouTube source URL instead of the WROLPi
link, so listings omit the source URL (archive detail keeps it); and
they invented links for channels, so channel collections now carry
their real /videos/channel page link. Search descriptions state that
search_str is optional for browsing the newest items.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Three findings from running Qwen3-1.7B against the docker stack:
the llm image lost libgomp when the build tools were removed
(llama-server needs the OpenMP runtime); the Help prompt let the
small model answer from memory instead of the docs, so it now
requires search_help before any answer; and help search results
called the slug "doc" while get_help_doc's argument is "slug", so
the model invented slugs — the names now match.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
publish_ai_models.py produces the CDN artifacts for every AI_MODELS
entry — GPG-signed .meta4 sidecars (reusing the map trust chain) and
the signed ai/models.json manifest — and uploads them with s3cmd.
The bundled catalog now carries the real sha256/size of the two Qwen
GGUFs, which match Hugging Face's LFS records for the Apache 2.0
unsloth quants, and whose signed meta4s verify against the committed
public key.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Every MCP tool is now one /api/ai call with minimal text rendering;
the formatting, link building, and truncation logic it used to carry
untested lives (tested) in modules/ai. Tool names are unchanged for
external clients; a source-scan test guards the names and that the
proxy only reaches /api/ai. New lean blueprint endpoints back the two
tools that had no equivalent: GET archives/{id} (detail + snapshot
history) and GET videos/{id}/comments (paged author lines). Paged
content is read up to the old 50k-char bound. get_inventory_items now
takes the inventory slug (inventories are config-only; the old
numeric-id endpoint no longer exists). This also brings wrolpi_mcp
into version control (venv/ is already gitignored).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The app's first streaming client: sse.js reads the /api/chat POST body
with fetch + a stream reader (apiCall's 60s JSON path cannot stream),
parsing SSE events with abort support. AiChatContext holds the
conversation (browser-only, per design) so it survives visiting the
Manage tab; switching modes starts a fresh conversation. The Chat tab
has the three-button mode picker, per-mode suggestion chips, streamed
answers with clickable relative WROLPi links, tool-activity indicators,
a slow-hardware banner, and a "not ready" state pointing at Manage.
ChatMessage states are in the theme-sample gallery.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
New AI nav item and /ai route with Chat and Manage tabs (Chat is a
placeholder until the SSE chat lands). The Manage tab mirrors the
Map/Zim download UX: the model catalog with the RAM-based tier
recommendation and a Pi 4 performance hint, downloads through the
download manager (progress on the Downloads page), active-model
select, enable toggle, idle-unload minutes, and llama-server
start/stop through the Controller (compose service name in docker,
systemd unit natively). AiModelRow is in the theme-sample gallery.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Tool definitions are generated in-process from the /api/ai blueprint's
own OpenAPI operations (operationId = tool name, description = tool
description, dataclass bodies + declared parameters = one flat JSON
schema) — the spec IS the catalog, nothing to drift. Manage endpoints
never become tools. A mode is a server-enforced system prompt + tool
allowlist (Help / Research / System per ai.md); the executor can only
reach allowlisted /api/ai endpoints, path parameters are encoded and
traversal-rejected so a crafted argument cannot escape an endpoint.
POST /api/chat streams SSE (status/token/tool_call/tool_result/done/
error), starts llama-server on demand, validates tool args server-side
with instructive errors the model can retry on, caps tool results with
a paging hint, and bounds the loop by ai.yaml max_tool_calls plus a
hard timeout. Caddy passes the stream unbuffered (flush_interval -1)
in both deployments.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
install_llama_cpp.sh builds llama-server from the SHA-pinned v0.3.0
source tarball (single source of truth: on-device upgrade, Pi image,
Live ISO, and the new docker/llm container all call it; version-marker
makes re-runs no-ops). start_llama_server.sh refuses to start unless
ai.yaml enables AI and the active model file exists — which also covers
unmounted media. wrolpi-llm.service is installed but deliberately
never enabled at boot: the Controller starts it on demand (registered
in managed_services; the llm compose service follows the samba
pattern), and the API's perpetual idle-unload worker asks the
Controller to stop it after the configured idle minutes. llama-server
is localhost-only natively, internal-network-only in docker, and never
proxied by Caddy.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
ai.yaml is the source of truth (opt-in: enabled=False by default) with
active model, idle-unload minutes, context override, and tool-call
bound. The model catalog follows the map pattern: a bundled constant
keeps the Manage tab working offline, the CDN manifest (ai/models.json
+ detached GPG signature) wins when reachable. GGUF downloads clone
the map-extract flow — the per-file meta4 is GPG-verified once and the
same bytes handed to aria2c, so the verified hash is the enforced one;
unverified downloads are refused. Models land in <media>/ai/models/
(gated by the existing media_mounted download rules). Manage endpoints
(catalog with RAM-tier recommendation, settings) live under
/api/ai/manage/ which the future agent loop excludes from its tool
catalog.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Search the wrolpi-help markdown in-process: files are parsed on demand
and cached by mtime; title matches outrank headings outrank body. No
database table — the corpus is dozens of small files. Pages are read
as paged markdown. The docker api container gets a read-only bind
mount of the help docs at the native /opt/wrolpi-help location; a
missing help directory degrades to empty results with a message.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Proxy the Controller under /api/ai so the model sees one API: an
aggregate /status to start every diagnosis, lean service list, recent
service logs (lines clamped), and disks + SMART. One controller_client
helper resolves the Controller base URL (docker service name vs
localhost). Every proxy degrades to an `errors` entry when the
Controller is unreachable — never a 500.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Lean, flat, LLM-shaped read-only endpoints over the library: global,
video, archive, and doc search; paged captions, archive text, and zim
entries; collections, inventories, and a media-directory text reader.
Links are server-computed; big content is paged, never truncated.
This ports the formatting layer from wrolpi_mcp (untouched for now)
into tested code. First phase of ai.md.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* api: run the perpetual loops in one Sanic-managed process
The singleton background loops (file worker, downloads, switches, jobs,
bulk tag, flag checks, update check) now run in a dedicated process that
Sanic's Worker Manager owns: `run_perpetual_process` is registered with
`app.manager.manage()` from `main_process_ready`. The manager guarantees
a single instance and SIGTERMs it on shutdown. No server worker runs the
loops, so the owner-pid / heartbeat / watchdog claim scheme is deleted
rather than hardened.
`perpetual_signal` keeps its name and is now a plain registrar. A new
`per_worker_task` registrar covers loops every server worker must run for
itself: syncing its log level (previously only the owner did, so the other
workers never picked up level changes) and a health check that asks the
manager to restart the perpetual process when it is FAILED (Sanic marks a
dead managed process but does not restart it).
Downloads are started as background tasks instead of Sanic signals; the
perpetual process has no server, so no finalized signal router. The
update check and the upgrade-check endpoint run git in a thread via
`refresh_update_status`. /api/status lists the perpetual process next
to the servers.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* perpetual: survive a predecessor's death, exit cleanly, coordinate restarts
Review fixes for the managed perpetual process.
Shutdown: cancel_background_tasks / cancel_refresh_tasks gather with
return_exceptions, so a cancelled child's CancelledError no longer aborts
the shutdown that called them (the process exited 1 with a download in
flight).
Restart after a crash: reconcile_after_predecessor heals shared locks a
dead process left held, fails RUNNING Jobs owned by a dead pid, clears
processing_domains, renews pending Downloads, and resets the file-worker
status before the loops start. N server workers issue one restart
request per cooldown through a shared timestamp in shared_ctx.
The log-level sync now also runs in the perpetual process. Manage() is
registered through register_perpetual_process so it is testable; the
entry point refuses to run with no loops registered (spawn).
Tests exercise the real entry point in a forked process with a live
background task under SIGINT and SIGTERM, Sanic's real WorkerManager
marking a crashed managed process FAILED and reviving it on restart,
Sanic's process naming and ProcessState enum, the listener registration,
and the reconcile step.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* perpetual: import the jobs helper inside its reconcile step
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Discover/compare (find, TEMP table, SQL diff) now lives in
wrolpi.files.refresh so FileWorker is the queue plus apply/move/reorganize.
Existing `from wrolpi.files.worker import compare_file_groups` still works.
file_group.idempotency is unused since SQL compare. DROP COLUMN on SQLite
3.40 does not rebuild the table, so file_group_fts stays valid.
directory.idempotency is unchanged.
* files: pass find -printf \0 as an escape, not a Python NUL (#716)
* files: pass find -printf \\0 as an escape, not a Python NUL
'%h\\0%f\\0%T@\\0' in a Python string is an embedded null byte. subprocess
cannot pass that as an argument, so every GNU-find refresh on Linux CI
raised ValueError. find wants the two-character escape \\0 in the format
string and emits NULs itself.
* files: serialize stem backfill so a cancelled compare cannot lock SQLite
compare_file_groups runs ensure_file_group_stems in a thread. Cancelling
the task does not stop that thread, and a second compare then writes
wrolpi_kv on the same connection (tests share one session). A lock
plus a version re-check makes the second call a no-op.
* files: unique (directory, stem) after collapsing duplicate FileGroups
713 added stem and a non-unique index. This release backfills any NULL
stems, merges leftover duplicates (tags and collection membership move
onto the winner), and creates a UNIQUE index.
Upsert deletes sidecar FileGroups before INSERT so a new primary cannot
collide. Refresh deletes before upsert for the same reason.
from_paths looks up by stem so adding a sidecar does not insert a second
row. Same-stem files without a modeled primary stay one FileGroup.
* files: do not use the test SQLite connection from a worker thread
get_db_curs in pytest shares the test session connection. Running
ensure_file_group_stems and _upsert_files via asyncio.to_thread then
segfaults when the event loop commits that same connection (CI xdist).
Keep to_thread in production, where get_db_curs opens a dedicated
raw connection.
* files: keep tags when a sidecar joins an unmodeled group
choose_primary_file is the single fallback for unmodeled stem groups so
_upsert_files and _upsert_file_groups cannot pick different primaries
and DELETE the tagged row. FileGroup.move merges an existing
(directory, stem) group and flushes the delete before taking that stem.
refresh_files_directly deletes before upsert, matching the other two
callers.
* files: group by (directory, stem), not stem alone
group_files_by_stem mixed marauder/s3.bin and other/s3.bin into one
FileGroup because it ignored directory. Unique (directory, stem) then
kept only one of them, so flasher search missed a firmware file.
run_command_streaming awaits the stdout reader instead of cancelling it
when the process exits, so the last lines are not lost on a fast CI run.
* jobs: write the exception into the Job log on failure
test_jobs_failure only saw 'boom' in the log when JobLogHandler captured
logger.error(..., exc_info=e). Sanic/pytest dictConfig replaces root
handlers (especially under xdist on CI), so that path is empty even
though status and error are set. Log the exception on the Job itself
and store it in the same _update_job as FAILED.
* files: pass find -printf \\0 as an escape, not a Python NUL
'%h\\0%f\\0%T@\\0' in a Python string is an embedded null byte. subprocess
cannot pass that as an argument, so every GNU-find refresh on Linux CI
raised ValueError. find wants the two-character escape \\0 in the format
string and emits NULs itself.
* files: serialize stem backfill so a cancelled compare cannot lock SQLite
compare_file_groups runs ensure_file_group_stems in a thread. Cancelling
the task does not stop that thread, and a second compare then writes
wrolpi_kv on the same connection (tests share one session). A lock
plus a version re-check makes the second call a no-op.
* files: SQL FileGroup compare, stem column, drop unused discovery path
Refresh walked the tree twice (find, then stat every unchanged group),
and a two-directory refresh scanned the whole media directory. Compare
now diffs in SQL on a TEMP table that already has mtime, accepts multiple
roots, and groups by a stored stem that is versioned so a suffix-rule
change cannot delete-and-reinsert tagged files.
Deletes refresh_discover_paths and /api/files/refresh_progress, which
nothing in production called. Video/doc/zim modelers use the same
(no domain row OR indexed != True) gate.
* files: keep stem current on reorganize; NUL-safe find parse
Reorganize wrote the new primary_path but not stem, so the next refresh
treated the FileGroup as deleted+new and dropped tags. GNU find now
emits NUL-delimited records so a tab or newline in a filename cannot
abort a global refresh. Stem backfill skips the table scan once the
algorithm version matches, and runs off the event loop. An empty roots
list is a global refresh, not invalid SQL.
POST /api/videos/channels failed with "database is locked" on
`INSERT INTO collection` on the docker WROLPi while background workers
(config saves, refreshes) were writing. The handler wrote through the
deferred request session: `create_channel` reads (the URL/name/directory
conflict check) and then INSERTs, and SQLite refuses that mid-transaction
lock upgrade instantly when another connection wrote in between --
busy_timeout is skipped for upgrades.
Same shape and same fix as POST /api/download (daf2ba39): the handler
now uses `get_db_session(commit=True)`, which begins IMMEDIATE so the
concurrent writer waits its turn, and which commits before the refresh
is fired so the refresh does not wait on a lock the request still holds.
The regression test commits a write from a second connection between
the handler's conflict check and its INSERT. The existing
`write_lock_held_briefly` helper cannot reproduce this: the Sanic test
client takes over a second to start the app, so its lock is released
before the handler runs. Before the fix the test fails with the
production error verbatim.
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
* Jobs queue and video transcode from the video page
Add an in-memory FIFO Job queue (wrolpi/jobs.py): functions register by
name, callers enqueue by name with JSON kwargs, and the perpetual-tasks
owner runs them one at a time. Logging emitted inside a Job and the
output of its subprocesses are captured into a per-Job log; Jobs report
progress and can be cancelled (a running Job's subprocess is killed).
Jobs are not persisted; an API restart forgets them.
Add a transcode Job and POST /api/videos/<id>/transcode. ffmpeg reports
progress via -progress pipe:1. The existing FileGroup is pointed at the
new file before the refresh so tags and history survive a container
change.
Replace the video page Refresh button with an Edit dropdown (Refresh,
Transcode). The Transcode modal pre-fills from the preferred codecs in
the Videos settings and shows the Job's progress, log tail and Cancel.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* Video page: move Delete into the Edit dropdown
The Edit dropdown now holds Refresh, Transcode and Delete, so the action
row is Download / Edit / Add to Playlist. Delete keeps its confirmation
(a Confirm dialog) and the tagged-files prompt, and is disabled in WROL
Mode like the other items.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* Video page: Edit > Edit to change a video's title
PUT /api/videos/<file_group_id> writes the title (and fulltitle) into the
Video's info json, creating the file when the video has none, then
validates the Video again so its title and search text follow. Files
stay the source of truth: the edit survives a refresh, and a re-download
of the metadata replaces it (the modal says so).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* Video page: edit the description in Edit > Edit
PUT /api/videos/<file_group_id> now takes optional title and description
and changes only the fields given; an empty description clears it. The
modal sends only what changed and refetches the description after saving.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* Transcode modal: allow a container-only remux
Keeping both streams is no longer refused: ffmpeg copies them into the
chosen container with fast start. The container select starts at the
file's own container, the hint explains what a remux does, and the button
reads Remux when nothing is re-encoded.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* WROL Mode: refuse video edits and transcodes past the API
The routes already returned 403, but a transcode Job queued before WROL
Mode was enabled still ran, update_video could be called from anywhere,
and the Edit modal's Save button stayed live. The Job now fails before
ffmpeg starts, update_video carries wrol_mode_check, and the modal
disables Save with a notice.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* Transcode: verify the output before replacing the original
After ffmpeg exits, the output is probed and compared with the source
(container, presence of video/audio streams, codec per stream, duration
within 1s or 1%) and its last 30 seconds are decoded to a null output.
Any failure removes the output and keeps the original.
The tail decode replaces the download-time completeness check, which
grabbed a single frame near the end: one ffmpeg_video_complete (sync for
the downloader, async for Jobs) now serves both, and is stricter, since
every frame of the tail plus the audio must decode.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* Jobs: claim atomically, cap pending, own-session subprocesses; retarget before delete
Review fixes. The worker read a Job's status and wrote RUNNING as two
steps, so a cancel landing between them was overwritten and the Job ran;
_claim_job now does both under the lock. enqueue_job refuses past
JOB_PENDING_LIMIT pending Jobs. JobContext.run_command starts the
subprocess in its own session and raises CancelledError when the command
was killed, so a handler cannot mistake it for a normal exit.
The transcode Job now retargets the FileGroup (committed) while the
original file still exists, and deletes the original afterwards
(transcode_video_file gained keep_original). A crash in between leaves
both files with the database pointing at a real one.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* Video page: the Edit menu owns the transcode Job; polling survives Close and errors
Review fixes. The transcode Job's id and polling (useJob) move from the
modal to VideoEditMenu, so closing the modal no longer stops the page
from learning the Job finished (and refetching the video, whose path may
have changed). While a Job runs the menu shows Transcoding..., reopening
shows its progress instead of the form, and Delete waits. A failed poll
request is retried instead of ending the polling. getJobs checks the
response like its siblings. Design narration trimmed from comments.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* Transcode: audio files, and Remove video to turn a video into an audio file
video_codec 'none' (REMOVE_VIDEO) drops the video stream (-vn, audio
required); an audio-only source is audio-only output whatever the video
target says, and a video target on it is refused. Audio output gets an
audio container (m4a/ogg/mp3) checked against the codec it will hold, and
a video container asked for audio output is replaced by the codec's
natural one. faststart only for mp4/m4a. Verification learned the
audio-only shape and ffprobe's container names (mkv reports matroska).
The modal offers 'Remove video (audio only)', switches to audio
containers and follows the audio codec, reads Extract Audio, and opens
for audio files (no video codec select).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* Video page: yellow Edit button
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* Transcode: remove the temp output on cancel; tell the user when a closed-modal Job fails or is lost
Review fixes. The temp output was only removed on Exception, but a
cancelled Job's run_command raises CancelledError (a BaseException), so a
cancel left the partial file for a refresh to index. Closing the modal
during a Job that then failed showed nothing; it now toasts the error
(and a cancellation). Polling gives up on a 404 or five failures in a
row (the API restarted and forgot its Jobs), warning the user and
unlocking Transcode and Delete instead of reading Transcoding... forever.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* Jobs: _claim_job docstring states the invariant, not the old race
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* Transcode: fragmented mp4 option for very long videos
A whole-file mp4 index grows with the frame count; for a many-hour video
it reaches tens of MB, which a phone must fetch and parse before the
first frame, and every seek walks it. `fragmented` writes the file with
-movflags +frag_keyframe+empty_moov+default_base_moof+global_sidx and
10s fragments: a ~1 KB moov plus a global seek index, so playback starts
and seeks quickly. mp4/m4a only; a remux, so it takes seconds per GB.
Measured on an 11-hour h264/aac file: moov 40 MB -> 1.4 KB, seek index
2 x 120 KB, 10k fragments, duration and codecs unchanged, remux in 25s.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* Transcode modal: the Fragmented label says audio for audio output
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* Video edits keep the original: custom_title/custom_description in the info json's wrolpi section
Editing a title or description no longer overwrites yt-dlp's keys. The
values go into the info json's `wrolpi` section (custom_title,
custom_description); the modeler prefers them when deriving the Video, so
a refresh keeps the edit while the original title and description stay in
the file. An empty custom description means none.
Both re-download paths carry the wrolpi section over: the metadata-only
refresh merges it into the fresh info json, and a full re-download
snapshots it before yt-dlp overwrites the file and restores it (then
re-validates) in finalize_download. FileGroup.get_wrolpi_json reads it.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Refreshing a specific file (the Files page Refresh button) bypassed the
ignored_directories check that directory and global refreshes apply.
Refreshing a hard link inside the playlists directory therefore created a
second Video for the same file.
The direct-file path now drops existing files that live in an ignored
directory before comparing against the database. Deleted paths are kept
so stale records are still removed.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* Hotspot captive portal: welcome page with the WROLPi's address
Devices that join the hotspot probe a fixed plain-HTTP URL to check for
Internet access. NetworkManager's shared-mode dnsmasq now resolves those
probe hostnames (and `wrolpi`) to the hotspot address via an include the
Controller writes when the hotspot starts, the probes land on the Controller
(port 80), and are redirected to /portal: "This is your WROLPi, its address is
10.42.0.1, open WROLPi / open the Controller".
Only the probe hostnames are hijacked so the hotspot keeps forwarding
everything else when the WROLPi has an uplink. The include is corrected and
the connection re-activated once if NetworkManager picked a non-default
subnet. Writing the include is best-effort; the hotspot starts regardless.
`hotspot.captive_portal` (controller.yaml, default true) toggles the probe
redirects and the include; like the other hotspot settings it applies on the
next hotspot start, so saving while the hotspot is up reports
restart_required and both Controller UIs say so. The status endpoint gains
the hotspot address and portal URL, shown in both UIs. repair.sh only
creates the include directory so a disabled setting survives upgrades.
Native only: Docker has no hotspot.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* captive portal: address review
* React: re-read the hotspot address whenever the HotspotToggle starts or
stops the hotspot, matching the fallback UI's refresh-on-status.
* A WPA3 start that reverts to WPA2 leaves the hotspot up, so it now gets
the captive portal address correction too.
* The corrective `nmcli connection up` logs a warning on a non-zero exit;
dnsmasq would otherwise keep the old address silently.
* HotspotStatusResponse.captive_portal defaults to true like everywhere else.
* Trim design narrative from the module docstring and React comments.
* The include-before-activate test now observes the include at the moment
of the `nmcli ... hotspot` call instead of inferring from call order.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* captive portal: never forward any record type for probe hosts
`address=` only synthesizes A/AAAA answers. iOS also asks for the HTTPS
record type; dnsmasq forwarded that upstream, Apple's answer aliased to
captive.g.aaplimg.com, and the phone fetched the genuine "Success" from
Apple instead of the portal. Each probe host now also gets `local=`, so
dnsmasq answers every type itself, and the CDN alias is a probe host too.
Observed with dnsmasq query logging on a hotspot with an eth0 uplink.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* captive portal: let acknowledged clients finish so the OS offers "Done"
A phone's sign-in sheet only shows "Done" once the OS re-probes and gets
the genuine success reply. The Controller now remembers each client (by
hotspot IP) that has loaded the portal page and answers its later probes
with the reply each OS expects (Apple "Success" HTML, Android 204, Windows
"Microsoft Connect Test", ...). The set is cleared when the hotspot
starts so a new session shows the page again.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* captive portal: acknowledge on Continue, not on view
iOS fetches the portal page in the background before opening the sheet,
then re-probes after the sheet's own navigation. Acknowledging on view
made that re-probe return "Success" as the sheet's content. Phones only
re-check the network (and offer "Done") after a full-page navigation, so
the welcome page now has a Continue button leading to /portal/continue,
which is the only place a client is acknowledged. The Continue page says
to tap Done and links to WROLPi and the Controller.
Also hijack clients3.google.com (older Android probe host).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* help: bump wrolpi-help to the captive portal and Speed Test docs
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* captive portal: comments describe the Continue design; clear acks on stop
Three comments still described acknowledging on view, which c72a7f7b
removed. Acknowledged clients are now also cleared when the hotspot
stops, so a lease reused in a later session sees the portal page.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
* Speed Test calculator: ping, download, and upload against the API
Users asking whether the hotspot is fast enough, or how many people can
stream video at once, had no objective tool that works offline. This
adds one under Calculators > Tools & Reference.
Backend (wrolpi/speedtest.py, /api/speedtest):
* ping: tiny no-store JSON the browser times for round trip latency.
* download: streams slices of a pre-built random buffer for a clamped
number of seconds (0.1-15). Random so no proxy or compression can
flatter the result; slices so no bytes are generated per request.
* upload: streaming-body handler that reads and discards the body and
reports bytes and elapsed time. Requests over 8 MiB are refused.
* info: the client address Caddy forwarded plus how many tests are in
flight. Sanic now trusts one proxy hop (PROXIES_COUNT = 1) so
request.remote_addr is the browser, not Caddy. Nothing else read
that field before.
There is deliberately no one-test-at-a-time lock. Several people
testing at once is how a user learns how the link shares between
streamers, so the server only counts in-flight tests (a shared_ctx
Value created pre-fork so every worker sees the same number) and the
UI reports how many others ran alongside.
Both Caddyfiles route /api/speedtest/* unbuffered, as the SSE chat
route already is, so throughput is measured live rather than in bursts.
Frontend:
* speedtest.js holds the logic with no React in it (bucketed
throughput, steady-state average that drops the first second, ping
stats, stream-capacity tiers, hotspot address check) so Jest drives
it with fake fetch and streams.
* SpeedTestCalculator runs 20 pings, a 10 s download, and a 10 s upload
with a live figure and rolling trace, then shows Mbps and MB/s, ping
verdict, the client address with a hotspot badge, and a table of how
many 480p..4K streams the download speed supports. Upload is
sequential POSTs of one 4 MiB Blob because Safari cannot stream a
request body. Nothing is saved.
* Sparkline is a new inline-SVG ui component (no chart dependency),
added to /theme-sample and checked in all four themes. A missing
sample is a gap, not a plunge to zero.
The page states that it measures the network path, not the drive that
actually serves video files.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* speedtest: bound uploads in time, close the Start race, address review
Review on #708 found three bugs and some polish:
* Start could be clicked again while the first fetchInfo() was still
pending, because the phase only left `idle` after that await. A
second click made a second AbortController and a second concurrent
run, and Cancel only reached the later one. start() now refuses when
a controller already exists and enters a `starting` phase
synchronously, so the button is Cancel before the first await.
* The upload phase could overrun the advertised 10 s by however long
the in-flight 4 MiB POST took: on a slow hotspot, the very link this
tool exists for, that is tens of seconds. runUpload() now aborts the
in-flight request at the phase deadline (counting only completed
requests) and uses 1 MiB bodies so a few-Mbps link still completes
several inside the phase.
* The server upload sink was capped by size only, so a slow or paused
body could hold a worker until 8 MiB eventually arrived. It now also
enforces MAX_UPLOAD_SECONDS (15 s, matching download) and answers
408 when exceeded. Both caps together are what "cannot hold a
worker" actually requires.
Also: a NaN `seconds` now gets a NaN message instead of "must not be
negative"; the module, Sparkline, and gallery docstrings are trimmed
to purpose; and the upload counter is covered for the mid-stream 413,
the 408, and a broken body.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
* docs: browse walks the file_group date index instead of probing every doc
The unfiltered /api/docs/search drove its join from `doc`, so SQLite probed
every doc's file_group row for the sort key before returning the first page.
With 20k docs scattered through a 10 GB database on a spinning disk, the first
request took about a minute cold (0.07s warm).
Use the same CROSS JOIN shape as the archive browse: file_group is the outer
table, SQLite walks file_group_published_datetime_idx and probes doc's covering
unique index on file_group_id. Both sides stay index-only and the scan stops at
the page limit. Measured cold: ~60s -> 0.6s, identical rows.
Filtered searches and the size/title orderings keep the ORM query.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* docs: trim browse comments, stabler plan assertion
Review feedback: drop the benchmark narrative from docstrings, state only the
non-obvious constraint at the call site, and assert the doc covering-index probe
by shape rather than by autoindex name.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
* zim: recompute the outdated-Zims flag after startup and on Delete
The Dashboard shows the "Outdated Zim Files" banner from the persisted
`outdated_zims` flag, but the modal lists files from a live scan. The two
disagreed: the flag was restored from the DB on startup and never
recomputed once the outdated file had been removed, so the banner showed
with nothing to delete, and the modal displayed an endless placeholder.
- The startup recompute now runs after the DB is up and configs are
imported, so it overrides the value `init_flags` restored from the DB.
Previously it ran before the DB was up, and the restored value won.
- DELETE /api/zim/outdated always recomputes the flag, even when nothing
was deleted, so clicking Delete clears a stale banner.
- The modal re-scans when opened and says when nothing was found instead
of showing a placeholder.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* zim: enforce restore-then-scan ordering; no Kiwix restart on empty delete
- Startup calls flags.init_flags() then flag_outdated_zim_files() right
after db_up, before the DB config import, so the disk scan always
overwrites the DB-restored value in one local sequence.
- DELETE /api/zim/outdated recomputes the flag in a finally.
- remove_outdated_zim_files skips refresh and the Kiwix restart when no
file was deleted.
- Trim narrative comments to the constraint they enforce.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
COUNT(*) OVER() forced SQLite to visit every matching file_group
before LIMIT, which is a full table scan of the (large) file_group
table — ~22s cold on a Pi for videos ordered by download date.
Count separately and cache for 60s by filters so indexed orders can
stop after one page. Past-the-end pages now report the true total.
Videos/archives/docs pagers keep totalPages across page changes.
* videos: codec download preferences with transcode and strict-fail options
Add video_codecs/audio_codecs ordered preference lists to the video,
channel, and RSS downloaders, alongside the existing resolution
selector. Codec preference is codec-major: a preferred codec at a
lower resolution beats another codec at a higher resolution.
When 'transcode' is enabled, a downloaded file whose streams do not
match the preferences is re-encoded with ffmpeg after download (only
the mismatching stream; targets are extensible maps, currently h264
and aac/opus/mp3). When 'strict_codecs' is enabled and transcode is
not, a download with no matching format fails permanently instead of
falling back.
Empty codec lists produce a byte-identical yt-dlp command to before.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* videos: honor strict_codecs when transcode cannot produce the preference
Review fixes:
- strict_codecs is now suppressed only when transcoding can produce a
preferred codec for every non-empty preference list (new
transcode_can_satisfy_codecs); av1-only preferences with transcode
enabled fail under strict instead of silently keeping the file.
- No partial transcode: an unfixable video mismatch blocks an
audio-only conversion (and vice versa); the mismatch falls through
to strict/keep.
- fsync the transcode output before rename, and the directory after,
so power loss on a Pi cannot leave a truncated final file.
- mp3 yt-dlp pattern matches mp4a.6B as well as mp4a.6b.
- ffmpeg maps use 0:V:0 (excludes attached pictures) and 0🅰️0?
(tolerates missing audio).
- Strict toggle is no longer disabled by transcode in the UI; popup
wording corrected.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* videos: serialize transcodes machine-wide via shared_ctx lock
Sanic runs multiple worker processes, so an asyncio primitive cannot
serialize transcodes across download workers. A multiprocessing.Lock
in shared_ctx now guarantees only one ffmpeg transcode runs at a time
on the machine; waiting workers poll the lock asynchronously so the
event loop stays responsive while a transcode holds it for hours.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* videos: transcode-lock review fixes
- Document that a crashed lock holder requires an API restart; this is
accepted rather than adding owner-death recovery.
- Move the 'Transcoding' user notification to after the lock is
acquired, so it fires when work begins rather than before a
potentially hours-long wait.
- Soften the lock test docstring to what it measures (in-process
wait/release); trim the acquire docstring to the polling rationale.
- Give transcode_lock its own section header in contexts.py.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
test_search_videos_does_not_block_event_loop asserted a wall-clock
threshold (echo < 0.4s), which fails on loaded CI machines from
scheduler noise alone. Replace the timing with ordering: the mocked
search parks its thread on an Event, the echo is sent only once the
handler is provably inside the query, and the test asserts the echo
completes while the query is still held open. Events are only waited
on off the event loop, so a regression fails the ordering assertion
instead of deadlocking.
Verified both directions: passes repeatedly against the real endpoint,
and fails when the endpoint is changed to run the search on the loop.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
