AI page: Quick Check vs Calibrate, host-specific verdicts, colour scheme
Documents that Calibrate is a superset of Quick Check rather than a
separate step people have to run afterwards, since two buttons side by
side read as alternatives. Explains why each candidate batch size has to
pass twice -- silent omission is intermittent, and one clean run is a
single sample of something that fails only some of the time.
Adds when Quick Check alone is the right call: auditioning models,
re-checking after a pull on unchanged hardware, a temporary host, or very
few indexers.
Adds the four-state colour scheme and the Host column, and records that a
verdict belongs to a (host, model) pair -- with the measured 436s vs 237s
difference for the identical model and batch size on two machines.
Corrects the stale claim that 8k context is the practical minimum. That
predates the batching work and contradicted the "Small context windows"
section further down the same page.
Documents the notification narration fallback, after a security-log model
returned a MITRE ATT&CK object that went out verbatim as a push
notification.
95a5bce
docs: model updates and context-window limits
The AI page gains two sections.
Model updates: what the badge means in each of its two states, that the check
never downloads anything to find out, and that auto_update_check is the one
thing that leaves your network.
Context windows: leads with the finding rather than the setting. Fitting the
window is not the whole story -- 25 indexers score, 30 makes the model echo
the input back with the window half full, and 37 in a window twice as large
returns well-formed JSON that silently leaves five out. That last one is what
someone choosing a model needs to know, because it looks exactly like
success.
8091bb2
docs: correct the pinned image tag
The release workflow strips the v, so the image is :2.0.0 while the git tag
is v2.0.0. Verified against the registry rather than inferred -- anyone
following the previous table would have hit manifest-unknown on their first
pin.
edea76c
docs: v2.0.0
Home leads with an upgrade section, because the single most useful thing to
tell someone already running v1.6 is that nothing changes until they change
it. The three settings actually worth turning on are named in one table
rather than left to be discovered across four pages.
Installation names the image tags properly -- :latest, a pinned :v2.0.0, and
:dev -- and says why you might pin: so upgrades are a deliberate act rather
than a side effect of docker compose pull.
Detection & Actions gained a pointer to operating modes at the end. Everything
on that page describes Automatic mode, which was true and unstated; a reader
following it in Monitor mode would have watched nothing happen and had no
idea why.
Home's feature list gained operating modes and replacement tracking, which
shipped after the page was last written.
de14e05
docs: state what the operating mode does and does not stop
The quarantine-mode row said "nothing is deleted automatically". A quarantine
timeout set to remediate deleted anyway -- the code is fixed now, but that
sentence was doing too much work on its own.
It names the timeout explicitly now, and a new section draws the distinction
the whole design rests on: the mode caps AUTOMATIC action (a scan, a timeout,
a retry), not your own decisions. The Delete button on the review page is
deliberately not capped, because a mode that blocked it would leave you with
a queue you could not empty.
720603b
docs: --test flag, and confirm hand-editing config.yaml still works
CLI Usage gains --test with its exit-code table. The codes are the reason
anyone would script this, so they are documented as three distinct meanings
rather than pass/fail: a config that will not parse is a different problem
from a service that will not answer.
Troubleshooting gains a short answer to a question the config cache will
otherwise cause someone to ask: yes, editing config.yaml by hand still takes
effect immediately. The cache revalidates against the file on every read.
Worth stating where a worried person will look, not only in the changelog.
7e2428c
docs: replacement outcome tracking
Documented on the Prowlarr scoring page rather than with quarantine and
severity: the point of this feature is what it tells you about an indexer,
which is why the UI puts it on the Indexers Stats tab too.
Leads with the distinction that motivates it -- an indexer whose mistakes get
replaced cleanly versus one whose mistakes are never replaced at all, which
the health score cannot separate -- rather than with the mechanism.
Also records why Lidarr is excluded, since 'unsupported for one app only'
invites someone to just switch it on. The reason is that each app needs a
different history endpoint and the wrong one returns HTTP 200 with unfiltered
data rather than an error, so guessing is silently destructive here.
46f6e7a
docs: operating modes
Documented on Quarantine-and-Severity rather than as a new page. Modes,
severity and quarantine are three layers of one question -- what happens when
a rule matches -- and splitting the answer across two pages would mean the
reader has to find both to understand either. The mode goes first because it
is the coarsest control: someone running Monitor does not need to read about
severity bands at all.
Covers the ceiling semantics explicitly (it does not rewrite your thresholds,
and can only ever reduce an outcome), why the banner exists, that Monitor and
--dry-run are the same behaviour with the flag winning, and how to tell
below_severity_floor apart from capped_by_operating_mode when working out why
something was not deleted.
Configuration gains the key in both the YAML sample and the reference table.
Sidebar renamed so someone looking for modes can find them.
Also fixed a self-link: the Quarantine review queue is a page in the web UI,
not a wiki page, and this document was linking to itself for it.
1b3a6db
chore: pin line endings to LF so wiki diffs are readable
core.autocrlf was true on the Windows clone and nothing overrode it, so every
checkout rewrote the working tree to CRLF. Two pages were sitting fully
modified against HEAD with no content change at all -- 204 lines removed and
204 identical lines re-added -- which meant a real one-line edit to either of
them would have been invisible in the diff.
Pinning eol=lf in .gitattributes fixes it for every clone rather than for this
machine only, and core.autocrlf is set to false locally so the working tree
stops flipping back.
3e69e7e
docs: repair the Docker compose example in Installation
Two paragraphs had been pasted inside the YAML fence, between
`environment:` and its `- TZ=America/New_York` value. Anyone who copied
that block got a compose file with a dangling `environment:` key and two
sentences of English where the value should be -- it does not parse, and the
error it produces points at the environment key rather than at the prose, so
it is not obvious what went wrong.
The prose is correct and worth keeping; it now sits after the closing fence
where it was meant to be.
Also in this file:
- seven em dashes were double-encoded (`a-EUR-`), left over from a
UTF-8 payload that was decoded as cp1252 somewhere upstream. This page
was the only one affected; every other page checked clean.
- line endings normalised to LF, matching the rest of the wiki. The file
had CRLF throughout, which made every diff against it show the whole
file as changed and hid real edits.
- the persistent-data table now names the tables added since v1.6.0
(inspections, quarantine holds, model validation results) rather than
stopping at grab attribution.
No content was rewritten beyond the compose block.
442e278
docs: document the AI master switch
Follows 4916959 / 44473b5, which added an Enable AI features toggle to
Settings -> AI and made it the single control for every AI feature.
AI-and-Model-Validation
New "The master switch" section up front: what stops when it is off, that
no Ollama request is made at all, and that turning it off erases nothing --
URL, model and system prompt stay on disk so switching back on restores
what you had. Notes the upgrade behaviour: an install that already had a
working URL stays enabled, only new installs start off.
Removed "Leave the Ollama URL blank and everything else works unchanged" --
that advice predates the switch and is now the wrong way to do it.
Configuration Added ollama.enabled to the example and the key table,
including the upgrade behaviour. Reworded the Ollama intro,
which said "when configured" and now also means "when
enabled".
FAQ "Do I need Ollama" now says it ships off and names the
switch, instead of telling readers to blank the URL.
Home AI scoring is called out as off by default.
Web-UI The AI pane row leads with the master switch.
Verified: no stale "leave the URL blank" phrasing remains anywhere, all 86
internal links across 17 pages resolve, every file is valid UTF-8.
94afd91
docs: bring the wiki up to date and add navigation
Structural
_Sidebar.md persistent navigation on every page, grouped getting-started /
how-it-works / using-it / help. GitHub renders this alongside
every page; without it readers had to return to Home to get
anywhere, and Home's hand-maintained index had already drifted
out of sync with the actual filenames.
_Footer.md repo / releases / issues links on every page.
New pages
Quarantine-and-Severity severity grading, the three bands, the review
queue and its three outcomes, timeouts. This was
previously undocumented and would otherwise have
been split across three pages.
AI-and-Model-Validation Ollama setup, the three validation tests, the
gate and the forced state, context requirements.
AI content had been scattered across
Configuration, Prowlarr scoring and Notifications.
API /api/health, /api/status, /api/scan, /api/logs.
The JSON API existed and was entirely
undocumented; Home did not mention it at all.
Fixed
Home [Docker](Docker) and [CLI](CLI) were broken -- no such
pages; the files are Installation and CLI-Usage. Rewritten
with a task-ordered getting-started path, corrected
Pushover -> Apprise (stale since v1.5), and a features list
that reflects quarantine, severity and model validation.
Detection The action sequence went straight from match to delete. It
now documents grading, the three bands, the evidence trail
and correlation IDs, and the client-delete failure path.
Also corrected Pushover -> Apprise.
Web-UI Added Quarantine and Settings -> AI. Corrected the Settings
pane list, which named panes that do not exist and omitted
several that do. Stats is a tab in the Indexers hub, not a
page, since v1.6.0.
Configuration Added the remediation block. Completed the ollama table --
system_prompt and update_check_hours were missing -- and
noted the measured context requirement.
Notifications Added the quarantine event and its fallback to `action`.
Noted that the Ollama toggles are gated on AI being set up.
FAQ "Will it delete good torrents" now covers quarantine. Added
two questions: whether AI is required (it is not) and how
to check a model before relying on it.
Troubleshooting Deleted-good-torrent now explains how to trace it via the
inspection record and three ways to prevent a repeat. Added
stuck-in-quarantine, nothing-ever-quarantined, and Ollama
returning nonsense.
CLI-Usage --daemon also sweeps quarantine timeouts.
Installation Re-encoded from CP1252 to UTF-8. It held 22 non-ASCII bytes
that GitHub renders as mojibake.
Verified: all 86 internal links across 17 pages resolve, and every page is
reachable from the sidebar.
b4bc410
docs: update notifications to Apprise, add multi-client torrent support
e54a2c6
wiki: update all pages for v1.4.0 features
b4d2547
Update wiki: v2 scoring formula, model selector, tasks/updates/backups pages, daemon mode, PUID/PGID, dev container
ebfc06b
docs: add Ollama AI scoring documentation
293e5e8
Merge wiki, keep local Home
d4d4f36