This folder is a local reference and development environment for VirtualDJ skinning, pad pages, and VDJScript. VirtualDJ does not publish comprehensive developer documentation; this repo fills that gap.
README.md — human-facing project overview
TODO.md — active task queue: recovering the full verb and skin syntax
HISTORY.md — completed tasks and the dated progress log of open ones
RETIRED.md — the queue TODO.md replaced (2026-10-01); unfinished, not startable
INDEX.yml — topic-to-file routing map
docs/README.md — index of all reference docs, source label policy, current status
docs/VirtualDJ Reference.md — Quick Decisions guide: preferred methods, rationale, examples
docs/VDJScript Verbs.md — curated API reference: canonical names, aliases, surfaces
docs/Official VDJScript Coverage Audit.md — authoritative names-only parity audit and local-test gap count
docs/Completeness Roadmap.md — ordered backlog for closing behavior, fixture, and hardware gaps
examples/Pads/README.md — pad page inventory, status labels, and maintenance checklist
tests/README.md — documentation test harnesses and reproducible fixtures
Start a session with just doctor. It is the brew doctor of this repo, and it
answers in one screen the four things that decide what you can do today: which
interpreter is in use and whether it can import what the extraction tools need,
whether uv is there, which VirtualDJ is installed against the builds the
artifacts are anchored to, and whether the live probe channel answers — which
sorts the queue into desk work and live work before you pick a task. Only the
python section can fail. just install fixes that one (uv creates .venv from
requirements.txt); every recipe then runs that interpreter.
Run it before regenerating any binary-derived artifact. A missing numpy does not
fail those — it degrades them: extract_action_catalog.py caught the ImportError
alongside a genuinely absent app bundle and promoted 30 of the vendor's own
example placeholders into the probe worklist, and the only symptom was a
cross-check drift in an unrelated-looking artifact.
For a new binary investigation, run just binary-source before disassembly.
Start method and vtable identification with the preferred symbol-rich arm64
historical bundle it reports (just binary-source --path for scripts). Then
compare the identified path against /Applications/VirtualDJ.app when the
question concerns the installed build. Treat cross-build matches as leads;
the installed binary establishes current code paths, and live probes establish
behavior. Do not infer unnamed vtable slots from order alone while a named historical vtable is
available. The selector fails rather than silently returning a stripped image
when no named bundle is present. Existing artifact checks and live tools that
are explicitly anchored to the installed build keep their stated source.
Choose the evidence channel for the question. For names, argument candidates,
and code paths, start with the vendor's bundled descriptions and examples, then
the symbol-rich binary and existing extraction artifacts. A Ghidra 12.1.3
analysis of the symbol-rich 18.0.9246 arm64 build already exists: see
docs/Plugin SDK.md, docs/Runtime Memory Probe.md, and the saved
tests/ghidra-*-9246.json exports before starting a new decompilation.
Decompile a bounded function when extraction/disassembly leaves a specific
control-flow or data-flow question open; compare its path with the installed
build before claiming it applies there. Binary and decompiler findings identify
possible behavior, while a live observation establishes behavior.
For live VDJScript questions, use the local HTTP /query channel for fast,
repeatable request/response probes and the introspection plugin when native
value type or HRESULT matters. The Remote protocol
also evaluates VDJScript queries, but it is a separate device-to-desktop TCP
protocol with push subscriptions and typed updates. Reach for it when timing,
change notifications, Remote skin/control behavior, or transport differences
matter; do not treat an HTTP response as proof of Remote behavior. Use pad,
skin, or UI fixtures for surface-specific behavior. See
Evidence Standards before recording findings.
For “what should I do next?”, maintenance, documentation cleanup, or evidence-pass work, read TODO.md first. Treat TODO.md as the canonical active queue and start with the first Ready task unless the user names another task. just next-task prints that task; just task-queue lists every task with its state. Each task carries one machine-readable Status: line — one word from Ready/Blocked/Conditional/Parking lot/Done (Retired only in RETIRED.md), with the explanation in the Note: paragraph below it — and selection fails loudly rather than skipping a status it cannot read, so keep that shape when editing the queue. A task that lands moves, block and all, to HISTORY.md; just check fails on a Done task left in TODO.md. The queue's goal is syntax, not behaviour: live behaviour work lives in the retired blocks of RETIRED.md, to be revived into TODO.md under a new identifier when the user asks.
Context load is the dominant cost in this repo. These rules outrank thoroughness:
- Writing VDJScript? Read the summary of docs/VDJScript Grammar.md first. Its "Read this much" section is the whole language in about a dozen lines; the Contents list jumps to a section when you need detail. Grammar is the one thing lookup cannot save you on — the parser reports no syntax errors, so wrong script silently does something else. Not writing script (skin layout, XML structure, file format work)? Skip it.
- Never read the large docs end-to-end.
docs/VDJScript Verbs.md(~6,300 lines),docs/Skin SDK.md(~2,700 lines),docs/Effects Engines.md,docs/VirtualDJ Reference.md, anddocs/VDJScript Local Test Tracker.mdare section-addressed references. Read only the section a task names, or extract it with path-scopedrg/awk. - Start a "how do I do X" task with
just topic X. It aggregates, for one term, the matching verbs, effects, and skin/pad XML elements, plus the real example files that use them (grep-verified, ranked by how much of the topic each demonstrates), the topical docs, and known local-test quirks. A working example often answers the task outright, leaving only per-verb detail to look up.just topic sampler,just topic colorfx,--format=jsonfor structure. Drill in withget-verb/get-fx/get-xml-elementafterward. - Look verbs up through the tooling, not the monolith.
just verb <name>is the one-screen summary — store record, vendor description, real usages, argument shapes with return evidence, every tail candidate by source, vocabulary groups, probe state, each labelled with its evidence tier (--format=jsonfor structure).just get-verb <name>is the bare store record — it follows aliases and suggests near matches on a miss.just list-verbs <term>when you do not know the name. OpenVDJScript Verbs.mdonly at the specific lines a search hit points to. - Record a settled verb fact with
just put-verb, not by hand-editing tables. A per-verb conclusion — status, confidence, evidence — goes into the store once (just put-verb <name> test_status=Pass confidence=local_test evidence="…"), not promoted into the tracker and topical docs by hand. The tracker is for the run narrative that does not reduce to one verb (see the Live probe section); the store is the authoritative per-verb record. A tested status without evidence failsjust check.test_status=Disprovedis not a behaviour result: it marks a name the verb table shows is not a verb (Evidence Standards rule 1b). The record stays so the disproof is addressable, but it is excluded from totals, sections, default listings and the rendered reference;just list-verbs --status=Disprovedsurfaces them. Never file a disproved name asFail, which is the bucket for real verbs whose behaviour failed a test. - Every query above has an MCP twin.
tools/mcp_server.pyserves the same query layer (vdj_topic,vdj_verb,vdj_get_fx,vdj_element,vdj_grammar,vdj_controllers,vdj_sysicons,vdj_lintfor XML and script,vdj_query,vdj_execute, andvdj_restart) to any MCP client over stdio with zero dependencies, so an agent in another repo can author skins, pads and VDJScript without this checkout in context. Registration (claude mcp add --scope userto use it from any repo), tool list, read-only mode, and the restart opt-in: docs/MCP Server.md;just mcp-checktests every tool, over stdio too, without a client (--liveadds read-only live calls). When thevirtualdjMCP server is connected in a session, prefer its tools over shelling out tojust. - Effect controls come from the FX catalog, not prose tables.
just get-fx Echogives the full slider/button map with normalized defaults (spelling-tolerant:BeatGridresolves toBeat Grid,Shaderto its canonicalVisuals);just list-fx --category=video_fx,--has-button=quant,--min-sliders=6,--has-length,--format=jsonanswer the rest. Do not hand-transcribe the catalog into Markdown. - Introspect effects by name, without loading them. Every
get_effect_*helper accepts an effect name where the docs show a slot number —just vdj-query "get_effect_slider_count 'Echo'",get_effect_slider_default 'Echo' 3— so a live question about an effect needs noeffect_selectand changes no state.get_effect_title '<name>'returns'<Canonical> - Deck N'or'', which resolves spellings and probes existence in one call. - Query the store instead of reading listings.
just list-verbsfilters —--surface,--section,--tier,--status,--kind,--needs-test— and--format=jsonfor structured output. Ask for the verbs you need (just list-verbs --surface=SkinQuery --section=Sampler); do not pull a category listing and filter it yourself.just next-incomplete-verbgives the next active (non-hardware-blocked) work item,just verb-statsthe breakdown.just uncategorized-verbslists the non-alias records still missing asection, with the b9246 source module where one exists.just list-verb-categoriesprints the section vocabulary with counts, so a--section=filter never has to guess a name. - Planning docs are frozen.
docs/VDJScript Reference Consolidation Plan.mdanddocs/Completeness Roadmap.mdare design references, not active state. Do not refresh, reorder, or re-scope them; do not spend turns rewriting planning prose or reordering the task queue.TODO.mdis the only active planning state, and it changes when a task completes or the user asks. - Probe over HTTP first, fixtures second. When VirtualDJ is running with the network interface enabled (
just vdj-upto check), read values withjust vdj-queryinstead of pad readback. Pad/skin fixtures remain necessary only for surface-specific behavior (rendering, pad context, skin runtime). Batch every probe a live session can carry, and prefer dump-style sweeps over one-question rounds. - Delegate mechanical passes. Transcribing observed values into tables, promoting settled tracker rows, and lint fixes are cheap-model subagent work; keep the main context for evidence interpretation and ambiguous calls.
- Prefer
INDEX.yml,just next-task,just grep-verb-docs <name>, and path-scopedrgover repo-wide discovery; broaden only after the task route proves insufficient or contradictory. - Never write a bare count. Every number in this repo is either a measurement of one build or a total derived from a regenerable artifact, and each has its own rule. The test is: would this sentence become false if a binary were updated or an artifact regenerated? If yes, it is not written as a bare figure.
- Build-anchored counts are stamped, not refreshed. Write
1,032 records on build 18.0.9598 (arm64, extracted 2026-09-05), never1,032 records. A stamped figure is permanent evidence about that build and must never be silently updated — a newer build gets a new stamped line, and the old one stays as the record of what the old build held. Copy the stamp, never recall it:just verb-table-stampprints the exact phrase from the artifact's own summary. Recalling it is howREADME.mdcame to attribute one extraction's counts to a build the commit message disagreed with. - Derived totals are not written at all. Corpus size, fixture count, probe/form totals, cross-check set sizes: name the command and the field that answers (
just action-catalog --cross-check→documented_but_not_probe_confirmed), not the number. These have no build to anchor them and go wrong on the next regeneration — usually in a commit that never opened the file. - Dated observations keep their numbers, frozen. "A 32,000-query sweep on build 9598 established the two-token grammar" is evidence; it describes a past run and stays true. A date alone does not license a count —
(2026-07-27): the verb table is the complete verb set — 1,028 recordsreads as dated but asserts current state, and that is the shape to avoid. - The two counts in
docs/Official VDJScript Coverage Audit.mdare the exception:tools/check_reference_status.pyenforces them, so they may be written plainly.
- Build-anchored counts are stamped, not refreshed. Write
Direction of travel: verb facts live in a structured record store (docs/vdjscript-verbs.json) reached only through flat just data commands — get-verb, list-verbs, put-verb, next-incomplete-verb, verb-stats (and get-fx, list-fx, fx-stats for effects). The command name is always a command and the argument is always data, so a verb called search or get can never be mistaken for a subcommand. Treat the JSON layout as private — go through the commands so storage can change (e.g. to one file per verb) without breaking your usage. Reports are queries on stdout: do not add a generator that writes a Markdown copy of store data. The store is seeded from the existing index, coverage audit, and tracker via python3 tools/verbdb.py bootstrap.
examples/Pads/ — working/reference pad page XML files plus copied built-in pad pages
examples/Skins/ — skin source trees, reference skins, and copied built-in skins
examples/Mappers/ — real working controller/keyboard mapper XML (ground truth for the mapper format)
examples/Samplerbanks/ — copied built-in sampler-bank XML (third XML format)
tests/ — documentation test harnesses, pad XML fixtures, and the evidence artifacts
(verb-table, action-contracts, action-catalog, verb-arg-forms,
attested-tails, vdjscript-corpus — see tests/README.md)
docs/ — Markdown documentation
The repo has five channels into VirtualDJ, and they are not interchangeable. The binary is the map and the live app is the prover: under docs/Evidence Standards.md the binary can enumerate names, expose the tokens a verb parses, and disprove a name, but only the running app can prove behaviour. The biggest gains in this repo came from using the first to aim the second — the verb table, the action catalog, the shared vocabularies, the skin dispatch tables and the argument forms all came out of the symbol-rich binary, and nearly every settled per-verb fact was then confirmed over HTTP (store tallied 2026-10-17: of 1,038 tested records, 916 cited HTTP in their evidence, 61 a pad run, 40 the binary, 39 the plugin, none the Remote protocol). Pick the channel by the question, not by habit:
| Question | Channel | Route |
|---|---|---|
| Does this name exist? What tokens does it parse? What is the default? | Binary (symbols, languages.zip, Ghidra decompile) |
just verb, just action-catalog --get, just binary-vocab --verb; parser readers via Ghidra below |
| What does it return / do right now? | HTTP Control Interface | just vdj-query, just vdj-execute; MCP vdj_query / vdj_execute |
| Was the keyword recognised, or silently ignored? Is the value a double or text? | Introspection plugin (sees the HRESULT and the native type) |
just plugin-status, just plugin-probe <name> |
| How does a value change over time, at what rate, in what native type? | Remote protocol (push subscription, typed VALUE frames) |
tools/vdjremote_subscribe.py; docs/Remote Protocol.md |
| How does it render, or behave in pad/skin context? | Pad or skin fixture, or the agent driving the window | tests/, screenshot saved beside the capture |
- Ghidra has been done; reuse it, do not re-import. A Ghidra 12.1.3 headless project of
build 18.0.9246 (arm64) exists, with the scripts in
tools/ghidra/(VDJFindSymbols,VDJTailExport,VDJDrawDeckXrefs,VDJDecompileThree) and saved decompile exports intests/ghidra-*-9246.json. It underpins the Pad FX argument contract, the time-consumer analysis indocs/Runtime Memory Probe.md, the Plugin SDK caller trace and the skin tag dispatch work. The analyzed project itself lives outside the checkout in the sibling cache named underghidra:inINDEX.yml; open it-readOnly -noanalysis. If that path is absent on the machine, say so rather than re-analyzing: a fresh import is hours of work and produces a project no other session can find. Decompilation is targeted — the dispatch tables, per-verb argument readers and skin tag readers — not a whole-binary dump; the payoff is in the argument readers, because an unrecognised argument is silently ignored and only the parser code shows the recognised keyword set. Everything it yields is Tier 2: it narrows what to probe, it never replaces the probe. - HTTP and the Remote protocol answer the same VDJScript queries, so reach for HTTP
first. HTTP is request/response on
localhost, needs nothing but the Network Control option, and returns a string. The Remote channel is a push subscription: VirtualDJ dials out to a device you advertise over mDNS, then streams typed values as they change. Use it when the question is about timing, update rate or native type, where HTTP's polling and string-flattening lose the answer, and not for one-shot reads. - The MCP server can start VirtualDJ, not only restart it.
vdj_restart(andjust vdj-restart) goes throughtools/vdj_restart.py, whose not-running branch launches the app withopen -a, passeslaunch_argsthrough, and waits for HTTP to answer; itsstatusform is read-only. Bothvdj_executeandvdj_restartare off until the server is started withVDJ_MCP_EXECUTE=1/VDJ_MCP_RESTART=1(docs/MCP Server.md); if a task needs a write and the client was registered without the opt-in, say so instead of working around it.
When VirtualDJ is running with its network interface enabled, VDJScript can be executed and queried over plain HTTP on http://localhost/ — no pad fixture or manual readback needed. This is the preferred channel for local-test probes. Full contract, verified behavior, and gotchas: docs/HTTP Control Interface.md.
just vdj-up— reachability check; run it before planning any live-test work.just vdj-query 'get_effect_name 1'— evaluate any VDJScript query; exact result string back. Read-only, use freely for sweeps.just vdj-execute 'effect_active 1'— run an action; the body is the verb's owntrue/falseresult (not transport success —nothingreturnsfalse).- Unknown verbs return HTTP 200 with an
error:<code>body; check the body, not the status. - Execute only verbs the current task names; never
systemor file/database-touching verbs through this channel. - Before probing a verb's arguments, read what is already known.
just action-catalog --get <name>gives the vendor's own description and its documented parameters;just attested-tails --verb <name>gives tails Atomix wrote in shipped scripts and the argument shapes with their return evidence (fadeout DUR DUR `BOOL`, seen invisibility=), which is the only record for the 167 verbs whose arguments are values rather than keywords;just binary-vocab --verb <name>gives the shared enumeration the verb draws from (stem names, colours, settings pages) with the members no other source names;just verb-arg-forms <name>gives what has already been probed.get_song_event's two-token grammar was established by a 32,000 query sweep and had been documented in the app bundle the whole time. - Probing arguments needs prepared state and nonsense controls. An unrecognized argument is
silently ignored, so a verb answering proves nothing:
just fixtureslists the 10 named states, andtools/probe_arg_forms.pycompares every candidate against two junk tokens inside each. Separation is positive evidence; failing to separate means the state did not discriminate, not that the token is unreal. - Time-varying verbs manufacture false positives.
get_cpu,get_timeandrecord_vuseparate from a control by drift alone.--repeat Ncatches fast drift; only two independent runs catch slow drift. - Executing verbs to test them writes to a live instance.
tools/probe_execute_forms.pyshows the shape that is acceptable: an allowlist, a round-trip test before probing, restore and verify after every form, and abort on a failed restore.timecode_cd_modeis why — it can be set from script and only a restart clears it. - A second live channel exists: the read-only introspection plugin. Where HTTP flattens a
result to a string, the plugin sees the native call —
GetInfo→double,GetStringInfo→ text, and theHRESULTseparately from the value, which is the only way to tell a recognized keyword from a silently-ignored one, and it needs no prepared state.just plugin-status/just plugin-probe <name>read the existing captures (tests/plugin-introspection*.json);just plugin-buildthenjust plugin-collecttake a new one. Two method rules learned the hard way: anE_INVALIDARGfrom a load-time probe means "not available now", not "no such form" — re-take negatives from the delayed sweep (just plugin-collect-late); and VirtualDJ writes0.0to*resulteven on failure, so the HRESULT is the answer, not the value. Full evidence table: the tracker's "Plugin Channel (VDJIntrospect)" section. - A UI observation persists what it saw. When the evidence for a row is a screenshot of the running app (agent driving the window), save the screenshot under
tests/beside the capture and cite its path in the row'sevidence. A reading whose screenshot was not kept is an unverifiable assertion; if that has already happened, the capture must say so in ascreenshot_provenancefield rather than cite "a screenshot in this task". - Recording a result has two destinations, and they hold different things. The verb store is authoritative for per-verb conclusions:
just put-verb <name> test_status=… confidence=local_test evidence="…"— one verb, one settled fact, with the build in the evidence string. The tracker holds the run narrative for a session that does not reduce to a single verb: a fixture setup, a negative result, a multi-verb probe, cross-verb interactions. A tested status in the store with no evidence now failsjust check, so record the evidence at the same time as the status, not later. Note the channel (HTTP vs pad) in the evidence, since some behavior is surface-specific.
- One concern per commit. A capture and its tooling ship together; an unrelated doc note found on the way ships separately, even if it is one paragraph.
- Subject is a plain imperative sentence, no type prefix.
Settle sampler default-context playback contracts, nottest(sampler): …; the repo's history is mostly in this form and mixing the two makesgit log --onelineharder to scan. - A superseded capture is committed under its
-initial(or-baseline-*) name, never under the confirmed name. The confirmed file name is written once, by the run that produced the confirmed capture. - Task state moves, it does not accrete. When a task lands, its block moves from
TODO.mdtoHISTORY.md; while a task is open, run narrative goes into the tracker or the topical doc and the task'sNote:points at it in one sentence.
- VirtualDJ skins are XML. The scripting language inside attributes is VDJScript.
action=""attributes take VDJScript actions.query=""takes a boolean/value expression.&chains actions in XML attributes and must be written&inside XML.- Backtick-wrapped expressions (
`verb`) evaluate and return a value in string/color contexts. - Personal working pad pages live in
examples/Pads/Quarantine/(usage examples only, never evidence); copied built-in app-bundle pages live inexamples/Pads/Built-In/; documentation test harnesses live undertests/. Seeexamples/Pads/README.mdbefore choosing a reference page. Skins live inexamples/Skins/*/; copied built-in app-bundle skins live inexamples/Skins/Built-In/. - The pad-page container format is specified in
docs/Pad Page XML.md; skin waveforms indocs/Skin Waveforms.md; the mapper format indocs/Mapper XML.mdwith real mappers inexamples/Mappers/Local/.just element <name>is the one-screen summary per XML element — inventory row, doc section, reader vocabulary, live probe results (negatives included), real usage, and which of its attributes no doc explains; it is tojust get-xml-elementwhatjust verbis tojust get-verb.just inventoryrefreshesdocs/skin-xml-inventory.json;just list-skin-elementsshows skin/video-skin attributes and doc routes;--family=allincludes other XML families, and--undocumentedfilters element-name gaps. examples/Skins/GraveRaver/src/is intentionally minimal and demonstrates the build system only. Do not use it as a polished skin reference.- The official VDJScript appendix coverage and local-test gap are tracked in
docs/Official VDJScript Coverage Audit.md. - docs/Evidence Standards.md governs every claim in this repo — read it before recording a finding. Three tiers: only direct observation of the running app (network protocol, HTTP interface, live pad tests, agent driving the window) proves that something works; forums, unbacked binary analysis, and official example files are leads; everything else is not recorded. Existence, kind, and behavior are separate claims. A channel's own return value is not a result. The binary can disprove a name even though it cannot prove behavior.
- Source labels (
Official,Official forum,Community,Published skin,Built-in skin,Published pad page,Built-in pad page,Action catalog,Vendor script,Local test,Inference) appear throughout the reference docs; Evidence Standards maps each onto its tier. - The official verbs appendix ships inside the app.
Resources/languages.zip→English.xml→<Actions>is the same prose virtualdj.com publishes, so 816 verb descriptions — including parameter lists — are readable offline viajust action-catalog. Do not fetch the manual for something the bundle already answers. - Run
python3 tools/check_reference_status.pyafter changing coverage counts, fixture inventories, or local reference links.
- Slot FX:
effect_select <slot> 'Name',effect_slider <slot> <param> <value>,effect_active <slot>— not name-based toggling. - ColorFX:
filter_selectcolorfx 'Name'+filterfor the main deck filter knob.effect_colorfx <1-4>+effect_colorsliderfor extra dedicated controls. - Sampler (page-aware):
sampler_pad <n>,sampler_color <n> 'auto'; comparesampler_pad_pageto text ranges like"9 to 16"and use absolutesampler_loaded <slot>guards for empty checks (sampler_loaded 16for page 2 pad 8), sincesampler_loaded <n> 'auto'tested unreliable. - Sampler (fixed slot):
sampler_play <n>,get_sample_name <n>,get_sample_color <n>. - Panels:
<panel visibility="...">for query-driven;name=""+skin_panelgroupfor persistent manual switching. - Dynamic text color: one
<text color="action">, not per-state color attributes. - Dynamic border color: not supported (CTO confirmed). Use fill or background instead.
- Time mode:
display_time 'remain,elapsed'+get_time, not custom skin vars. - Computed arguments: transport-style verbs (
loop,beatjump,phrase_sync) ignore backtick-computed arguments even when the literal works — branch to literals (var_equal '$n' 16 ? loop 16 : loop 32) or chain params (get_var '$src' & param_multiply 2 & set '$dst');setdoes accept backticks. Tested rules:docs/VirtualDJ Reference.md§Tested Grammar Rules. - beatjump: argument must be signed (
beatjump +4; barebeatjump 4is a no-op).
| File | Status | Demonstrates |
|---|---|---|
| examples/Pads/Quarantine/Reference - Slot FX.xml | Quarantined (personal) | Slot-based audio FX pads |
| examples/Pads/Quarantine/Reference - ColorFX.xml | Quarantined (personal) | Filter + ColorFX selection |
| examples/Pads/Quarantine/SAMPLER READ ONLY.xml | Quarantined (personal) | Confirmed read-only multi-page sampler with absolute empty-slot guards |
| examples/Pads/Built-In/README.md | Built-in | Copied VirtualDJ app-bundle pad pages; semi-official executable examples |
| examples/Skins/Built-In/README.md | Built-in | Copied VirtualDJ app-bundle skins; semi-official executable examples |
| examples/Skins/ModularSkeleton/build/skin.xml | Unofficial (project-authored) | Minimal modular skin scaffold |
| examples/Skins/GraveRaver/src/skin.xml | Unofficial (project-authored) | Minimal XInclude build-system demo, not a skin design reference |
| examples/Pads/Quarantine/Reference - Page Aware Sampler.xml | Quarantined (personal, superseded) | Page-aware sampler labels, colors, actions — see examples/Pads/README.md before citing |
| examples/Pads/Quarantine/COLOR FX.xml | Quarantined (personal) | ColorFX selection with stems context |
No pad page in this repo is Canonical any more (2026-09-04): the three Reference - * pages were quarantined, so Built-in rows — Atomix's own shipped pages — are now the only pad source to copy patterns from or cite as evidence. Quarantined rows are real personal working files kept as usage examples only; never cite one as evidence that a verb exists or behaves a given way, because a personal page legitimately contains experiments that were never claimed to work. See examples/Pads/README.md and examples/Mappers/README.md for the full provenance tables.
For skin XML discovery, start with just list-skin-categories, then use
just list-skin-elements --category=<id>. just element <tag> --parents /
--children return observed vendor nesting with source locations; --parent=<tag> /
--child=<tag> filter the listing. Prefer these scoped queries over opening
tests/skin-xml-relations.json. Categories are editorial and nesting is Tier 2,
not a supported-child schema. Refresh nesting with just skin-relations after
source XML changes. See Skin Element Discovery.
| Path | Content |
|---|---|
~/Library/Application Support/VirtualDJ/Pads/ |
Installed pad pages |
~/Library/Application Support/VirtualDJ/Skins/ |
Installed skins |
~/Library/Application Support/VirtualDJ/Mappers/ |
Controller/keyboard mappings |
~/Library/Application Support/VirtualDJ/database.xml |
Main track database |
See docs/Application Internals.md for the full path map, database structure, and stem sidecar layout.