Skip to content

Latest commit

 

History

History
1559 lines (1256 loc) · 100 KB

HISTORY.md

File metadata and controls

1559 lines (1256 loc) · 100 KB

History

Completed tasks, moved here from the task queue when they landed, plus the dated progress log of open tasks whose narrative outgrew a Note:. Each completed block keeps its ### <id>. <title> heading and Status: Done line so identifiers stay unique across both files; just check enforces that and that nothing here is anything but Done, and that nothing Done remains in TODO.md, the queue since 2026-10-01 (its predecessor, TASKS.md, is kept as RETIRED.md). Blocks are the record of what was established and where it was recorded; they are not startable and just next-task never reads this file.

Syntax recovery (TODO.md, 2026-10-01)

K3. Recover template, class, include and condition expansion

Status: Done

Note: The factory host never sees these rules because the skin loader applies them first. The Offline Skin Host now runs CSkinEngine::load through loadEnd(), so the loader stage can be observed. Recover the rules for the template and class mechanism, includes, and conditional constructs, including plain group turning into a panel when it has visibility. Each rule needs paired fixtures in that host: one expanded form and one control.

Done when every expansion directive the 9246 loader handles has a stated rule and a fixture, or is listed as unresolved. Whether a condition evaluates to true is a behaviour question and stays out.

Done 2026-10-02. The rules are in Skin Expansion, read from CSkinClass::add/apply/applyXml/applyPlaceholders and the condition functions on 9246, and checked against paired native fixtures in the composition host (just skin-expansion check, capture). Every prediction held beside its control. There is no include mechanism on this build. Rules without a fixture (named colours, mixer-order pseudo-decks, placeholder attribute removal, deck context for conditions) are listed under Open in that doc. Tier 2.

Accepted next sequence (2026-09-05 review)

R1. Repair queue selection

Status: Done

Note: 2026-09-05. The convention, the parser, and the regression fixtures are in place; every task in this file now carries a readable status.

What the selector used to do: just next-task was an awk pass matching ^Status: Ready$ exactly. Of the 25 tasks here, 22 carried a decorated status — Status: **Ready.**, Status: Ready — reframed, Status: Ready, but low expected yield, Status: DONE (2026-07-26, HTTP). A bank is… — so the selector saw three tasks, would have gone silent the moment R1-R3 closed, and had no way to tell a finished task from an invisible one.

What replaced it:

  • A machine-readable status line, specified in Queue Rules above: one state word from the closed vocabulary Ready / Blocked / Conditional / Parking lot / Done, first line under the heading, explanation moved to the Note: paragraph below it. All 22 decorated lines were rewritten this way, each one's prose preserved verbatim in its Note:.
  • tools/task_queue.py as the single parser behind just next-task, just task-queue, just status, and just check. It refuses to select when anything in the file is malformed, so an unreadable status stops the run instead of removing a task from the queue. It rejects a missing status, two status lines in one block, a status buried below narrative, a state outside the vocabulary, and a duplicate task identifier.
  • Regression fixtures in tests/todo-status/, run by just check: legacy-decorated.md holds the four real decorated forms this file carried (all four must now error, none may be skipped), normalized.md proves selection skips a Done entry and returns the first Ready one, and structural.md covers the four structural faults including the duplicate 13.

States assigned by reading each task's completion narrative rather than its old wording — the rule is that the state describes the remaining work. That reclassified several: task 0 and R4 are Conditional (both move only when task 10 names a record the store cannot hold), task 5 is Blocked (needs unrecognized hardware or a virtual MIDI port), and tasks 1, 3, 10a and 10b are Ready on their remainders even though their first pass landed. Identifiers were already unique after the 13/13b split; just check now enforces it.

Not done here: the states are one agent's reading of each narrative, so a task whose remainder is finer-grained than its heading still needs its own pass. The vocabulary is deliberately small — add a state only when a task genuinely does not fit one of the five.

R2. Panel and group parser pilot

Status: Done

Note: 2026-09-05, build 18.0.9598 (arm64) + live deck-skin fixture. The reader inspection, the candidate diff, and one candidate carried to a confirmed local test. Full evidence: the tracker's "Skin Reader Vocabulary And The clickthrough Attribute" section.

The inspection. Three readers located and recorded, each anchored by strings only it compares, each window the tightest stretch of __text holding an xref to every anchor: skin_object_base (0x10037c54c-0x10037cebc) — the attributes every skin object reads, <panel> and <group> included; element_dispatch (0x10037dfd4-0x10037ebf8) — the element-name switch; panel_builder (0x1007959a4-0x100795f40). Packaged as tools/extract_skin_readers.py (just skin-readers, just skin-reader <name>, just skin-candidates) over tests/skin-reader-vocabulary.json, so the diff is regenerable rather than transcribed. --check is build-anchored: it skips itself on a CFBundleVersion change rather than failing on addresses that were never expected to survive a bump.

Boundary, stated as R2 asked. String references only, one call level, no call target followed, no branch coverage measured, no disassembly of control flow, and no function names — the build is stripped and none is guessed. A name a reader passes to a helper is invisible to this method. <group> has no builder of its own in this extraction, which locates the next question rather than proving none exists.

The wall this pass hit, and it is not a helper-resolution wall. Attributes shipped skins use heavily — sourcecolor (460 uses), textaction (678), panelname, swapdeck, firstvisible, textwidth, dblaction — are absent from the binary in any case. They are not reader vocabulary at all: they are class="" template placeholders (the binary carries [TEXTACTION], [ACTION1], [bordercolor] as placeholder tokens), which is why lint_skins.py already skips attribute checks on elements with class="". So absence from the binary is evidence a name is not reader vocabulary, and says nothing about whether a skin may use it. No R4 trigger came out of this task — nothing here was a finding the current record could not hold, and no extractor failed to resolve a reference it needed.

The candidate taken live: clickthrough. Read by every skin object, compared against the single value pass, present in no shipped skin and no SDK doc. Five generated deck skins in tests/Skins/clickthrough-probe/, identical apart from one attribute on one element: two buttons on the same rectangle, each writing its own global, so the answer is read over HTTP instead of judged from a screenshot.

Variant Attribute on the top button $ct_top $ct_bottom
baseline (none) 1 0
visible-off visibility="param_equal 'no' 'yes'" 0 1
pass clickthrough="pass" 1 1
value-control clickthrough="qzqzqz" 1 0
attr-control zzclickthrough="pass" 1 0

Two independent runs, variant order reversed in the second, identical both times. clickthrough="pass" makes an element fire its own action and let the click continue to what is underneath — additive, not a redirect. Both controls separate, so the attribute name and the value each carry the behavior. visible-off is the calibration with a known working attribute: it proves in the same fixture that the click was over the bottom button and that attributes on the top button are honored, so the negatives are about clickthrough and not about aim. Documented in docs/Skin SDK.md under "Attributes Every Skin Object Reads".

Two facts the setup established, both needed by anyone repeating this. A skin folder with no image beside the XML is refused — a modal "Impossible to open skin <name>" while load_skin still returns true, so the channel's own result says nothing. And the skin list is not cached: a folder created while VirtualDJ runs loads immediately, verified by copying a known-good skin to a new name. load_skin also turned out to be its own restore oracle — in query position it returns the current skin identity — and is recorded on the verb.

Still open, deliberately. Only pass was tested here, because it is the only value that window compares. (2026-09-07: the historical-installer excavation read the boolean parser the loader falls back to, and the yes/TRUE fixture variants then confirmed a third state live — drawn but transparent to clicks, own action not fired. Recorded in Skin SDK.) clickthrough on a container rather than a button, and whether the pass-through reaches more than one layer, are untested. just skin-candidates still lists applyfx, setdeck, song_pos, foldersearch and the forceshow values as untested leads, and the element switch knows multibutton, resizepanel, keyboardmap, rack, onexit, os and darkmode — existence only. Task 10a's waveform questions are untouched: this pilot shared no fixture with them and closes none of them.

R3. Known-position fixture and get_time discrimination

Status: Done

Note: 2026-09-06, HTTP, build 18.0.9598, deck stopped. The fixture exists and proves itself, and each of the three targets has a reproducible position relationship across both runs. Full evidence: the tracker's "Known-Position Fixture And get_time Discrimination" section and tests/get-time-positions.json.

The fixture, which was the prerequisite. known_positions in tools/fixtures.py — deck 1 stopped, cue 1, a loop start and a loop end at three distinct positions — plus tools/probe_known_positions.py (just known-positions), which establishes it, proves it, probes it and restores it. Every position is read by an oracle that is not get_time: cue_pos 1 mseconly, get_loop_in_time on, get_loop_out_time on, and get_position × get_time 'total' for the playhead. Nothing assumes the numbers it asked for — quantize moved set_cue 1 15000ms to 14496 — so all four are read back and the run aborts rather than probe if any two coincide. Restoration is verified: cue deleted (has_cue 1 → no), loop exited, deck returned to the contents found at start, display_time never touched.

The result. Two runs, the fixture torn down and rebuilt between them with different numbers and the form order reversed; two phases each; all four phases agreeing.

  • get_time 'cue1' equals cue_pos 1 mseconly exactly, and tracked the cue when it moved (14496→70496, 24496→60496) while the loop endpoints were held fixed.
  • get_time 'loopout' equals get_loop_out_time exactly, in all four phases.
  • get_time 'loopin' equals get_loop_in_time only while a loop is active. With the loop exited it returns the loop-out value while the oracle still reports the in point. Reproduced in both runs, so it is recorded as a conditional rule, not as instability.

The controls did the work R3 said they would. An unrecognized tail falls back to elapsed, not to the bare form — bare followed the operator's display_time 'remain' throughout — so bare-versus-argument separation was never recognition, and the three targets are confirmed by matching their own oracle and by tracking a moved position. short alone behaves as an unrecognized tail. absolute keeps the display_time mode, so it is a modifier rather than a mode. cue is not cue1: it reads the active cue, 0 until one is activated.

Recorded elsewhere: get_time, cue_pos, get_loop_in_time, get_loop_out_time, loop, set_cue and goto all now carry local-test evidence in the store, and docs/VirtualDJ Reference.md gained the tested rule with both traps. Two setup facts worth their own line: loop N lays an N-beat loop ending at the playhead, and a bare signed number to goto is beats.

Unresolved, explicitly. to_lyrics returned 0 in every phase — the fixture has no lyrics, so this says nothing about the tail. Whether cue2, cue3, … exist as tails was not probed. loop_out after loop_in on a stopped deck produced a 4-beat loop unrelated to either point; loop N was used instead and that observation was not chased. Nothing was tested on a playing deck, deliberately, so that no result could be drift. No R4 trigger: the existing capture contract held this fixture without change, and no extractor failed to resolve anything.

R5. Repair next-incomplete-verb, which fails the way the queue selector did

Status: Done

Note: 2026-09-07. Both pools now report; the selftest fixtures hold the two silent failures. just verb-stats prints active_incomplete: 2 with incomplete_pools: {audit: 2, contract_gap: 224}, and once the audit pool empties the selector hands back a task-10 worklist item labelled with the pool it came from.

  • Arithmetic by inclusion (audit_pool): a record is active if it is needs_test, not blocked, and not already Pass. The old needs_test - blocked - tested_and_needs subtracted a set that only partly overlaps — records are blocked without ever having been flagged needs_test — so the figure crossed zero into -9.
  • Fall-through (contract_gap_pool, select_next): with the audit pool empty, the pick comes from records with no row in tests/verb-return-types.json and none in tests/verb-arg-forms.json, resolving aliases to their canonical and skipping HARDWARE_BLOCKED. Both the picked record's pool and the pool sizes are reported; the split is computed at read time and never written into the store, since the artifacts regenerate.
  • Regression fixtures in tests/verbdb-selection/, run by verbdb.py check: blocked-not-needs-test.json is the store shape that went negative, audit-pool-exhausted.json must yield a contract-gap pick rather than silence, and both-pools-empty.json keeps real completion distinguishable from a stalled selector.

The diagnosis this task was written from, kept as the record of the defect:

Added 2026-09-07 from the review's metric finding ("the store's incomplete selector follows an assigned needs_test flag, not every missing contract"). The finding has since turned from a limitation into a defect: just verb-stats reports active_incomplete as a negative number, and just next-incomplete-verb reports one active item while task 10 has open contracts on most of the store. Same failure shape R1 repaired for the task queue: a selector that goes quiet is read as completion.

Cause, measured on the store as of 2026-09-07: active_incomplete is needs_test - blocked - tested_and_needs, but eleven records are blocked without being needs_test (the Rane, NS7 and motorwheel names), so the subtraction crosses zero. The selector then draws only from the audit's needs_test set, which is nearly exhausted, and never from records with no return-type or argument-form evidence — the very records task 10 exists to close.

Do, in tools/verbdb.py and nothing else:

  1. Fix the arithmetic: count a record active only if it is needs_test and not blocked and not already Pass. The figure can never be negative.
  2. Let the selector fall through: once the needs_test set is empty, select the next record with no Tier-1 contract evidence (no return-type row in tests/verb-return-types.json, no argument-form row in tests/verb-arg-forms.json), skipping HARDWARE_BLOCKED. Report which pool the pick came from so a reader can tell "audit gap" from "contract gap".
  3. A selftest fixture in the style of tests/todo-status/, run by verbdb.py check: a store with a blocked-not-needs-test record must not go negative, and an exhausted needs_test pool must yield a contract-gap pick rather than nothing.

Acceptance: just verb-stats prints a non-negative active_incomplete with the pool split shown, and just next-incomplete-verb returns a task-10 worklist item after the audit pool is exhausted. No prose, no schema change; this is the worklist source for 10/10b and belongs before the next 10b sweep.

R6. Make just xml-stats name its own blind spot

Status: Done

Note: 2026-09-07. just xml-stats now prints reader_vocabulary_unused beside undocumented in totals: the reader-vocabulary names no shipped file writes, joined from tests/skin-reader-vocabulary.json at generation time.

  • Build-anchored the way extract_skin_readers.py --check is. The join carries a status: current lists the names with the build they were read off; a CFBundleVersion mismatch between that artifact and the installed app reports stale and lists nothing, since a reader vocabulary read off one build says nothing about another; unverified (no app installed) and unavailable (vocabulary not extracted) are distinguished. --check prints a non-fatal notice when the committed join no longer matches a re-join — a build bump is not a broken inventory, but the blind spot must not read as current when it is not.
  • The extractor docstring and the docs/README.md inventory entry now say what the count measures: mentions of the elements shipped files happen to use, with attributes and behavior contracts out of scope. Neither carries a number.
  • Regenerating the inventory also picked up R2's tests/Skins/clickthrough-probe/ fixtures (21 → 27 skin files scanned); element sets are unchanged, only use counts moved.

The task as written:

Added 2026-09-07. The review named the limitation ("XML inventory documented means an element mention, not a full attribute/behavior contract"); R2 then proved it — clickthrough is read by every skin object, was in no shipped skin and no doc, and undocumented: 0 never moved. It still reads undocumented: 0 across all elements, and the next agent will read that as completeness. The fix is a data-command change, not a warning paragraph.

Do:

  1. In tools/extract_xml_inventory.py, have the stats output carry a second field alongside undocumented: the reader-vocabulary names no shipped file writes, joined from tests/skin-reader-vocabulary.json (the same list just skin-candidates prints). Build-anchored like that artifact's own --check: on a CFBundleVersion change the field says the vocabulary is stale rather than listing it.
  2. One sentence in the extractor docstring and in the inventory entry of docs/README.md: the count measures mentions of elements shipped files happen to use; attributes and unused features are out of scope, see extract_skin_readers.py.

Acceptance: just xml-stats shows the candidates beside undocumented, and neither doc line carries a number. Cheap-model delegable; can ride with R5.

R7. Carry the skin-reader candidates R2 left untested to live tests

Status: Done

Note: 2026-09-07. Ran on build 18.0.9598 against a live VirtualDJ, deck-skin surface. Fixture: tests/Skins/reader-candidates-probe/ — generate.py renders every variant from one template, run.py loads each one, resets the globals, clicks a named point with a CGEvent helper it compiles itself, and reads the globals back. Six series, each run forward and reversed with identical rows both times. Full tables in the fixture README; the run narrative and its method notes are in the tracker's "Skin Reader Candidates Taken Live 2026-09-07" section; the confirmed behavior is in docs/Skin SDK.md.

Per candidate, as the task asked:

Candidate Outcome
r Confirmed: the <mousecircle> radius in skin units, with a nonsense-attribute control separating it; the default without it is the element's half-height, and x/y turned out to be absolute skin coordinates, not element-local
onexit Confirmed: <onexit action="…"> runs when the skin is replaced; nonsense-tag control never fired
multibutton Existence + behavior: builds an object that absorbs a click while drawing nothing, and does not build child elements
pannel Confirmed as a container: a button inside one is built and clickable, as inside <group>
song_pos Negative: behaves like the nonsense control where <songpos> takes the click — very likely the verb name referenced as that element's default action, not an element spelling
foldersearch Negative, calibrated against <folderlist> in the same placement, which does take the click
resizepanel, rack, keyboardmap, os, darkmode Negative in a deck-skin panel and at the skin root: no hit area, no children built
8pads (and the whole forceshow vocabulary) Not reached: a lone panel shows whatever it forces, and a two-panel group= kept showing the same member under both skin3FxLayout/skin6FxLayout states and through effect_3slots_layout. Boundary in the tracker
applyfx, setdeck Not reached: no attribute or element to attach them to, and click routing cannot discriminate them. Needs the H4-style read of the panel builder

Knock-on changes the run forced, all small: probe fixtures are excluded from the XML inventory (they carry nonsense tags by design); extract_skin_readers.py now also counts a backtick-quoted token as documented, since its {2,} pattern could never see a one-letter name like r, which would have stayed a "candidate" forever after being confirmed. Writing r, song_pos and foldersearch into the SDK doc shrank summary.candidates to the forceshow values plus applyfx/setdeck — the intended reading of that field. setting and effect_3slots_layout picked up local-test records along the way.

The task as written:

Added 2026-09-07. This is the review's "systematic Skin SDK discovery" gap, sized to what R2 already extracted rather than to a new sweep. just skin-candidates lists the reader vocabulary that no shipped skin and no doc names — applyfx, setdeck, song_pos, foldersearch, r, and the forceshow value 8pads on build 18.0.9598 — and R2's note adds seven element names the dispatch switch knows only by existence (multibutton, resizepanel, keyboardmap, rack, onexit, os, darkmode). Needs just vdj-up.

Method is the clickthrough series in tests/Skins/clickthrough-probe/ and H1: one generated deck skin per candidate, identical apart from the attribute or element under test, each variant writing its own global so the answer is read over HTTP rather than judged from a screenshot; a nonsense-attribute and nonsense-value control per candidate; the reader window (just skin-reader <name>) read first to size the value set, as H1 did with getBoolParam. A candidate the window compares against a fixed set gets that set; one that takes free text gets the argument shapes attested tails show for the nearest verb.

Deliver per candidate one of: confirmed behavior recorded in docs/Skin SDK.md with the fixture path; existence-only (parsed, no observable effect in the fixture) recorded as such; or not reached, with the reason. A negative needs the R2 boundary — which reader, which values tried, which branch was not exercised. Do not extend to the waveform questions in 10a; they share no fixture with this.

MCP authoring tools (2026-09-28)

M1. Make The Skin Linter Check Well-Formedness

Status: Done

Note: Added 2026-09-28, found alongside the pad-lint pass-through fixed in 0b84148. tools/lint_skins.py scans tags with a regex (scan_tags) and never parses the file, so a malformed skin produces vocabulary warnings and "Skins lint passed" — the MCP vdj_lint tool told an agent a truncated file was fine. Parse each file with xml.etree first and report a parse error as an error (non-zero exit, as lint_mappers.py already does), keeping the regex scan for vocabulary. Check examples/Skins/Built-In/ still passes, since shipped skins may rely on entities or XInclude the stdlib parser rejects; if they do, scope the parse check to explicit paths. Add a malformed-skin case next to the pad one in mcp_server.py's self_check.

Landed 2026-09-28. The planned stdlib parse was dropped on evidence: it rejects 50 of the 64 files under examples/Skins — raw & in values (43), case-differing close tags (<Tooltip>…</tooltip>, the two DDJ-WeGO skins), duplicate attributes (built-in Lite.xml among them) and a stray ? before the root (AFX_beatgrid.xml) — all of which ship, so the parser tolerates them. lint_skins.py now runs a quote-aware tag-balance check that permits exactly those (duplicate attributes as a warning) and errors on unterminated tags, comments or quoted values, unmatched close tags, unclosed elements and a missing root. It reports zero errors across all 166 skin files in examples/Skins and tests/Skins. The MCP self_check now lints a truncated skin and expects a failure. Tolerances recorded in tools/README.md.

M2. A Static VDJScript Linter For vdj_lint

Status: Done

Note: Added 2026-09-28. The MCP validates XML but not the script inside action="" / query="", and VirtualDJ reports no syntax error — wrong script silently does something else — so this is the largest gap in the authoring toolset. Add tools/lint_script.py and a kind: "script" for vdj_lint that takes a bare script string, plus a pass over action/query/backtick attributes when linting pad, skin and mapper files.

Every rule must cite a settled record, per Evidence Standards: flag unknown verbs against the verb table (a Disproved record is a hard error; a name merely absent from the store is a warning), and encode only rules with local_test evidence — the Tested Grammar Rules section is the source: unsigned beatjump, backtick-computed arguments to loop/beatjump/phrase_sync, sampler_loaded <n> 'auto', and filter_selectcolorfx in a query (already in lint_pads.py; move it here). Do not make raw & in an XML attribute an error: M1 found it in 43 shipped skins that load, so it is at most a style note. Tokenise per VDJScript Grammar, not with ad-hoc regex. Each finding prints its rule's evidence source. Validate against the vendor corpus (tests/vdjscript-corpus.json): shipped scripts should raise no errors, and any that do are either a rule bug or a finding worth recording.

Landed 2026-09-28 as tools/lint_script.py, just lint-script, and vdj_lint kind: "script"; pad, skin and mapper kinds now also lint the script inside the files. Three levels — error, warning, note — each finding naming its source; rule list in tools/README.md, one case per rule in tools/test_lint_script.py. just check runs --repo (no errors; the warnings are the deliberate invalid-verb and tested-bad forms in probe fixtures) and the tests.

Calibration against the vendor corpus (factory, built-in, add-on and wiki snippets, 2026-09-28) changed three rules. Quotes and backticks nest ('pitch_slider'), and parentheses group a branch (? (a ? b : c) : (…)), so neither is an error. effect, sampler and get take sub-keywords rather than verbs and are skipped, as are skin define placeholders ([ACTION]). The truthiness-trap rule is a warning only for query-only and slider verbs; for a dual verb such as effect_select or video_transition the sweep read false at rest, Atomix uses them as conditions, so it is a note. The filter_selectcolorfx-in-query rule moved here from lint_pads.py as a note: it is repo convention, and the vendor corpus uses that form.

Fourteen vendor scripts still raise errors, and each is a real mistake in a shipped file rather than a linter bug: stray quotes and backticks (deck all effect_stems 'rhythm'' in the Astro Audio DJ KONTROL 4 mapping, 'Filter'`` twice in the Traktor Kontrol MX2 mapping, '$decksubassign[SUBDECK]'' in the Denon SC5000 Screen skin), and empty ternary branches (shift ? loop 32 ? : offin the Numark 4Trak mapping, trailing… : zoom -2% :in the Denon SC5000 mapping,? : off/: :ladders in the Pioneer DDJ-800, XDJ-AERO and Traktor S8 mappings). Most sit on rarely taken branches, which is how they ship. The unknown-verb warning also surfaces vendor typos:lopp_roll, ar_equal, veffect_3slots_layout, deck1; the deck-target warning, once placeholders were skipped, deck masster. just lint-script --corpus` reproduces the list.

Open overlap: lint_mappers.py already resolves each action's leading verb against the doc-derived verb index, so a mapper passed to vdj_lint can report an unknown verb twice (once per source). Retire one when M3 touches vdj_lint again.

M3. Let vdj_lint Take Content As Well As Paths

Status: Done

Note: Added 2026-09-28. An agent working from another repo usually holds a draft, not a file this server can read. Add an optional content string to vdj_lint (mutually exclusive with paths), written to a temp file and linted with the same scripts so output is identical. Do it after M2 so kind: "script" is content-first from the start. Cover both forms in self_check.

Landed 2026-09-28. vdj_lint takes content in place of paths for pad, skin and mapper drafts (written to a temp file, reported as <content>:<line>); kind: "script" already took it. The M2 overlap is closed: lint_mappers.py no longer checks verbs, and the script inside mapper actions is lint_script.py's alone, against the verb table. Retiring the old check mattered: it read the doc-derived verb index, which still lists none, browser_filter and browser_search — names the binary disproves — so it passed quarantined user mappers that call them (AKAI APC Mini MK2 - Custom Mapping.xml, KEYBOARD - DeathDisco Keybindings v2025.12.xml). lint_script.py --repo now includes examples/Mappers/Official-Addons, which raises no errors; Quarantine stays out of just check. Self-check covers a pad and a mapper draft.

M4. Expose Controller Vocabulary Through The MCP

Status: Done

Note: Added 2026-09-28. Mapper authoring is a stated MCP use case, but the server can lint a mapper without answering which controls a device has. Wrap tools/controller_schema_inventory.py (just controllers --device <id> / --path /device/slider) as a read-only vdj_controllers tool. Its output is shipped syntax, Tier 2 — the tool description must say so, as the script's own help does. Offline only; add a self_check call.

Landed 2026-09-28 as vdj_controllers. Wrapping the existing queries was not enough: --device returned the device's metadata and the local-vs-shipped mapper comparison, not the control names a mapper's <map value=""> needs. controller_schema_inventory.py gained --controls (--match to narrow), which reads the device definition from the extracted vendor/controllers/ tree — the manifest's build when extracted, otherwise the newest — and reports which build it read. The tree is gitignored; without it the tool says to run just controllers-vendor. The MCP tool defaults to the control listing, with compare for the old comparison and path for attribute usage on a definition path.

M5. Expose The Sysicon Atlas Through The MCP

Status: Done

Note: Added 2026-09-28. Skins and pads name built-in icons, and an agent currently guesses the names. Wrap tools/sysicon_atlas.py (--cell, --unnamed, --format json; never --extract) as a read-only vdj_sysicons tool, carrying each cell's evidence tier through. Lower value than M1–M4 because icon use is rarer than script or mapper authoring.

Landed 2026-09-28 as vdj_sysicons, taken off the parking lot when the user asked for the remaining MCP work. The atlas could filter only by cell, so sysicon_atlas.py gained --search over descriptions, wiki keys, tested keys and binary candidates — the question an author actually has is "which key draws the sideview icon", not "what is in H6". Each row keeps its key_status, so a binary candidate is never presented as a tested key.

Historical-installer follow-ups (2026-09-07 review)

H1. Clickthrough value matrix

Status: Done

Note: 2026-09-07, build 18.0.9598, deck-skin surface. The named CXMLNode::getBoolParam disassembly (captured for 9.0.5308, 9.0.7607 and 18.0.9246 as bool-param in tests/build-history-2026-09-06/) accepts only yes/true/no/false, case-insensitively, and returns the caller's false default otherwise — so the proposed six-value test collapsed to one unobserved state. yes and TRUE variants on the overlapping-button fixture, two reversed-order runs: boolean true is a third state — the element stays drawn, its own action does not fire, the click reaches the element underneath — beside additive pass and the swallowing default. Recorded in Skin SDK clickthrough, the tracker's second-pass table, and the fixture README.

H2. Wrapper construction paths

Status: Done

Note: 2026-09-07, same build. Twelve more generated skins: the top button wrapped in a <panel>, a plain <group>, and a <group visibility="…true">, each with no attribute, pass, and yes; plus three-layer stacks. Panels and visibility-bearing groups honor clickthrough exactly as a button does; a plain group ignores it in both values, the live counterpart of the historical CSkinPanel::loadChildren split that tests the presence of visibility/novisibility. pass carries a click exactly one layer down. Recorded in Skin SDK <group> (two construction paths, constant-true visibility as the workaround) and clickthrough. Not run: a constant-false novisibility wrapper. Not probed, deliberately: whether the recorded <group class="…"> crash is a plain-group artefact — a hypothesis only, and a crash on a live instance is not a fixture.

H3. Historical vendor corpus diff and store-visible compatibility history

Status: Done

Note: 2026-09-07. Both deliverables landed as artifacts and joins, not prose: tools/diff_vendor_history.py (just vendor-history-diff) writes tests/build-history-2026-09-06/vendor-text-diff.json; summary.json gained verb_history; joined_view and just verb show both per verb. Three unofficial verbs (setting_if_unchanged, get_pad_page_name, pad_page_favorite) received their 9.0.5308 appendix descriptions via put-verb. Findings and the probe worklist they open (old parameter spellings such as get_saved_loop 'len', video_source 'shader') are in the excavation doc's "What the older vendor text and skins still say". The historical skin archives were not copied into examples/. Original brief follows.

Offline, delegable to a cheap model, no runtime claims. Two deliverables from the three older payloads (~/Downloads/install_virtualdj_2020_b5308_mac.pkg, ~/Downloads/VirtualDJ_2023_b7607_mac.pkg, ~/Downloads/install_virtualdj_2026_b9246_mac.pkg, expanded with pkgutil --expand-full; the memory notes hold the same paths):

  • Description and example diff. Compare each historical languages.zip → English.xml → <Actions> and each shipped skin/pad archive against the current vendor corpus. Every verb name in the old appendices is already in the store (checked 2026-09-06), so the target is text: descriptions or parameter explanations that were shortened or dropped, and script-bearing XML that no longer ships. Output is a provenance-stamped list (archive path, member, build, exact source text), deduplicated against unchanged material. Historical vendor usage is a lead about current behavior, never a supported-form promotion.
  • first_seen_build and alias transitions in the store. summary.json → transitions already holds the sampled name additions and primary-to-alias flag changes (goto_bar → goto_beat_in_bar between 18.0.9246 and 18.0.9583). Carry them into the verb store through the extractor/bootstrap path so just verb <name> prints them — never by hand, and never as a release date: the samples bracket table appearance, nothing finer.

H4. Runtime argument parsing from the named IAction::create

Status: Done

Note: Closed 2026-09-19 by a desk-only reconciliation that added no evidence. It closed with named limits; it did not complete. just runtime-grammar --audit still reports completion_claim: false, because tools/runtime_grammar_audit.py has no closed-with-limits state and hardcodes the field. The three named limits are in Runtime Argument Grammar Tests and in each obligation's limit record:

  • remote-entry. A b9246 lead scoped to Remote-skin and skin-load contexts, not a finding. It is blocked until the isRemote flag can be established independently.
  • list-conversion. Reachability of IAction::getListParam is unresolved. No direct reference does not prove dead code.
  • editor-structure. Editor token spans and guard hints are blocked on a safe targeting method, and coordinate clicking is banned.

What the closure did:

  • It labelled every H4 rule in VDJScript Grammar as repeated, one capture, or matches controls. Rules resting on one capture or on control-matching readings were reworded as unresolved.
  • It folded the unmapped-symbol triage into the obligations. review-with-existing-family symbols are now navigation-only related_symbols, and support-only symbols are out of scope.

These stay named open items, not blockers of this closure:

  • the context-fixture-needed triage groups (dispatch/source context and deck-state selection, which covers multiple-playing and automatic-master transitions);
  • editor-fixture-needed;
  • consumer-specific;
  • the cause of the mixerN permutation.

Per-verb behaviour inside constructed actions passes to task 10, which now points at the b9246 caller scan as a lead list.

Progress: every dated pass is summarised in HISTORY.md and reported by just runtime-grammar (per capture), just runtime-grammar-editor and just frontier-closure; the rules that survived two suites are promoted in VDJScript Grammar and the method in Runtime Argument Grammar Tests. Still open: editor token spans and guard hints, multiple-playing and automatic-master transitions, the mixerN permutation cause, and per-verb behaviour inside constructed actions.

Original note: Was Conditional on task 10b's harness landing; that trigger fired on 2026-09-02 (fixtures and the argument prober both shipped), and the status was flipped on 2026-09-11. Every rule this walk recovers is a 10b test and nothing else. Folded into task 10 as its first sub-step; listed here so the lead is not lost. Bound it to the questions that change how a probe is constructed — delimiters, quoting, fallback when a token is unrecognized — and stop there; it is not a mandate to reverse every action implementation. Order: 10c's tooling pass first (it is what H4's findings get recorded against), then this. IAction::create(char const*, char const**, int) is named on 18.0.9246 (x86_64 entry 0x100596f1c; resolve it again with nm before use). Follow argument consumption, delimiter handling and fallback branches, and contrast with the separately documented editor parser. Every recovered rule becomes a discriminating test for the 10b harness, not a documented rule on its own — editor acceptance and parser tolerance cannot substitute for an observed result. This has the highest ceiling of the excavation's leads because VDJScript grammar is the repo's stated cliff and the parser reports no errors.

H4 progress log

Progress 2026-09-17 (generic evaluator caller): retained the direct-caller scan and bounded bodies in runtime-parser-branch-routes.json. A verified b9246 ACTION_setting::onQuery call supplied the read-only parser_setting_eval fixture on build 9598. Computed boolean comparisons and the outer-quoted missing-final- backtick forms discriminated. Computed numeric one and text on failed their true predictions; literal/type controls retained and explained the exact consumer comparison without generalizing to other settings. No settings were changed. getFloatParamEval caller bodies are captured, but its live fixture is still open.

Progress 2026-09-17 (backtick evidence routing): split the broad backtick family into exact consumer/case links. The existing constant, get_text and param_add results follow distinct captured readers; they no longer stand in for generic getParamEval or getFloatParamEval coverage. The audit validates recorded call edges against the manifest and hashed assembly and rejects unknown selected case IDs. No live result or grammar prediction was changed by this correction.

Progress 2026-09-17 (unmatched-quote chains): froze and ran the parser_constants suite in tests/runtime-grammar-unmatched-quote-cases.json on build 9598. Unmatched opening quotes and opposite closers preserved the numeric prefix instead of applying the following addition; a later matching closer allowed the outer addition. Both quote styles and numeric baselines matched their frozen predictions and separated from their controls. The audit links the capture without claiming internal cursor behavior or universal consumer fallback.

Progress 2026-09-17 (empty-quote consumer): the frozen parser_constants suite tests/runtime-grammar-quote-consumer-cases.json ran through the existing argument prober on build 9598. Single/double empty operands and both marker/value baselines matched the predictions. Explicit omission contrasts distinguish the positional cases; those cases still match unequal-string nonsense controls and retain that label. The quoted-argument audit links the new capture without claiming a universal empty-argument rule or closing unmatched-quote behavior. The same-length follow-up also held and separated from both nonsense-value controls: equal nonempty triples and an empty last operand selected the equal result, unlike an empty first or second operand. This tests the argument-count alternative directly.

Progress 2026-09-17 (editor span calibration): aborted before candidate tests after a cropped-dialog coordinate click reached a cue pad behind the editor and started deck 1. The safety pause was independently verified; exact position restoration was impossible without a pre-click baseline. The failed calibration is retained in tests/runtime-grammar-editor-spans-calibration-aborted-9598.json; no span or guard finding was recorded. Further coordinate-based editor tests require a verified targeting method that preserves the active dialog.

Moved from the H4 task block on 2026-09-16; the block itself followed on 2026-09-19, when H4 closed.

Progress 2026-09-17: repeated the frozen parser_editor_help predictions on build 9598 with forward/reverse UI screenshots saved under tests/, plus original and reopened-restoration images and a fresh paired HTTP capture. Candidate predictions held again; both nonsense controls again displayed zoom help, preserving the failed no-help prediction. The new capture is separate from the historical run whose images remain unrecoverable. runtime_grammar_editor.py --http … --ui … selects a capture pair and validates saved screenshot paths and hashes. Token spans, guard hints, and the independently established remote-mode fixture remain open.

Progress 2026-09-12: bounded parser/editor capture and exact-script HTTP candidate suites landed; see Runtime Argument Grammar Tests. just runtime-grammar reports the completed build-9598 confirmation capture. H4 remains open: factory/consumer coverage is not exhaustive, live editor comparison has only a visual spot check, and the interrupted deck-context case has unproven exit causation. The continuation adds guarded zoom/beatlock/all-deck execution fixtures, selected-deck query comparisons, untrimmed output tests, restoration journals and a regenerable static frontier. Inspect each capture through just runtime-grammar --artifact <path>; the report distinguishes complete and interrupted runs, and derives a separation field so a prediction that held against a blank result is not mistaken for a discriminating one. The rules that survived two suites and two baselines were promoted into VDJScript Grammar on 2026-09-12 (keyword quoting, unit-suffix case/adjacency, comma decimals, signed-vs-unsigned numbers, the malformed-number reset, backtick inertness on execute), with local_test store records for zoom and beatlock. Single-capture and editor-side candidates stay unpromoted.

Progress 2026-09-12 (later): the asymmetric-scope item is half closed. parser_master_scope (just runtime-grammar-master) pins selection and master to different decks, which the earlier suites never did, and the capture settles that deck master and deck active both track the master deck while an unwrapped verb and deck default track the selection — promoted into VDJScript Grammar with local_test store records for get_deck, masterdeck and masterdeck_auto. It also found playing and mixer1-mixer4 to be recognized targets the wiki does not list, with mixerN resolving to a deck that is not N (cause untested, recorded as an observation only). That stopped fixture left asymmetric playback untested; the 2026-09-16 playing-scope result below now supplies that comparison. Button press/release lifetime is now done via the mapper surface: a virtual CoreMIDI button read over HTTP between note-on and note-off settles that a button action runs on press only, that while_pressed saves and RESTORES the prior value rather than clearing, and that it binds its own statement rather than the chain. See VDJScript Grammar.

Progress 2026-09-13: the static frontier is closed (just frontier-closure, gated in just check). All 30 queued indirect sites resolve to virtual dispatch (26, of which 18 are refcount releases), _actionFactory calls (3) or a disassembly artifact (1) — no argument consumers. _actionFactory is indexed by verb id, proven by two fixed-entry calls that store the same number at object+0xc. The structural finding: IAction::create finishes the argument loop into vector<SActionParam> before calling the factory, so arguments are lexed centrally and only then dispatched per verb. The static route to argument grammar therefore ends here; what remains is per-verb behavior inside the constructed action. Variable scope is also settled, including @ persistence across a real restart. The paired editor-help pass on 2026-09-16 is captured by just runtime-grammar-editor: candidate HTTP/help predictions held, but the frozen no-help prediction for two unknown-head controls failed (both showed zoom help). This is appearance evidence only; matching editor token spans or guard hints remain outstanding. The 2026-09-16 boundary-placement suites now isolate whitespace position from operator adjacency, retain the failed skip-whitespace predictions, and confirm the narrower original-value result with changed values and quoted forms; inspect the boundary captures through just runtime-grammar. The asymmetric playback gap is also covered: --grammar-playing uses verified digital silence on initially empty decks 3/4, preserves loaded decks 1/2, swaps the sole playing deck against a pinned master, and verifies restoration. The completed build-9598 capture separated active/master from playing and default; the initial selection-changing calibration is retained as incomplete. Multiple-playing, automatic-master transitions and the mixer permutation remain outside that fixture.

Hazard 2026-09-12 (superseded, kept for the reasoning): the reported VirtualDJ "crashes" were a minimized window — live process, live HTTP, no window, cmd-tab unable to restore it. /query is inert; /execute minimize is what does it, and every skin has a minimize button. See Runtime Argument Grammar Tests. The original note read: These exits leave no crash report, so a clean artifact directory is not evidence a run was safe; the one recorded instance had deck master constant 37 pending right after deck sandbox constant 37. The shared factor with the asymmetric-master run is unusual deck-wrapper tokens (sandbox, playing, mixer1-mixer4). Do not probe deck targets on an instance in use, and do not treat one clean completion as clearance — see the hazard section in Runtime Argument Grammar Tests.

H4 floating evaluator progress (2026-09-17)

Recovered and hashed the b9246 Pioneer display query entry route to the captured floating-evaluator consumer. Added lossless binary HTTP comparisons to the existing argument prober, a read-only empty-deck fixture, pre-candidate literal calibration, and frozen questions. Two independent build 18.0.9598 captures each recorded 10 held predictions and one failed computed-text prediction; a held leading-space result matched controls and remains null evidence. Preserved both captures and joined their exact cases to the branch audit. This does not close H4: editor token spans, remote-mode establishment/restoration, reachability and final branch reconciliation remain open. Details: docs/Runtime Argument Grammar Tests.md.

H4 floating conversion follow-up (2026-09-17)

Two independent build 18.0.9598 read-only captures each recorded 13 held and two failed predictions against frozen display-frame oracles. Direct numeric/unit forms, computed beats, and longer expressions separated from controls. Zero-frame cases matched controls. Chained constant-prefix predictions failed with ASCII constants, so they do not establish the inherited-parameter path. Original expectations and both captures are retained; see the Runtime Argument Grammar Tests follow-up.

H4 evaluator branch classification (2026-09-17)

The structured audit now partitions the conditional jumps of the two captured b9246 evaluator bodies into grammar questions, caller context, cache lifecycle, string storage, ownership and caller-interface groups. It checks exact assembly-derived membership and retains a question/limit per group. This narrows future fixture work without claiming live branch coverage or whole-parser completeness; cache lifecycle, incoming parameters and nonzero relative input remain explicit evidence limits.

H4 arithmetic reader comparisons (2026-09-17)

Both operand positions and two asymmetric constant pairs were tested in the existing read-only prober. Two build 18.0.9598 runs each recorded 36 held predictions and eight failures, with 12 held cases matching controls. Raw quoted actions and computed beats produced sums; direct beats returned error:1; computed numeric text yielded second- then-first concatenation. Trailing-backtick-only predictions failed. Frozen suites, both captures and exact consumer/conversion call edges are preserved and audit-linked.

H4 text interpolation comparisons (2026-09-17)

Added sentinel-controlled get_text boundary, escape and formatting questions to the existing read-only prober. Two build 18.0.9598 captures each recorded 20 held and two failed predictions; three held cases matched controls. Fractional and beat formatting predictions failed with exact outputs retained. The audit links these observations to the distinct interpolation route and excludes outer-token escape, internal type, skin-cache and universal-formatting claims. No live state writes were needed.

H4 unmapped-symbol triage (2026-09-17)

Reviewed the captured b9246 symbols lacking primary branch-family mappings and recorded dispositions, representative assembly anchors and next actions in the existing obligation checklist. just runtime-grammar --triage exposes this focused view; totals come from disposition_counts, with unreviewed additions visible in symbols_without_triage. Priorities are optional-cache boolean evaluation, incoming-parameter selection and the distinct typed/float pair evaluators. Existing family gaps remain visible and no live coverage was promoted. Anchor/duplicate/category regressions and the runtime grammar checks passed. The full just check stopped at the unchanged binary-vocabulary artifact's re-extraction mismatch against installed build 18.0.9628; historical data was not rewritten.

H4 boolean cache consumer follow-up (2026-09-17)

Extended the existing historical caller scan to getBoolParam, identifying the selected-slot and named-effect effect_active routes as cache-supplying leads on b9246. Reproduced the bounded assembly capture and added checked argument-setup classification without claiming indirect-call completeness or live cache reuse.

Extended the reversible-action runner with the prepared parser_effect_boolean Phaser fixture on HTTP build 18.0.9628. Initial and independent confirmation captures preserve predictions, off/on baseline readbacks, controls, journals and restorations. The confirmation adds shape-matched quoted-text/unclosed-expression controls and opposite-valued forms: their null results prevent interpreting selected-slot activation or malformed-input toggling as expression evaluation. Paired-backtick boolean/integer cases separate from controls. Exact results are joined by boolean-cache-consumer in just runtime-grammar --audit, with per-verb evidence available through just verb effect_active. Cache reuse and native type/branch coverage remain open; the unmapped-symbol triage now starts with incoming-parameter selection.

Runtime-grammar, capture, restoration-guard and reference-status checks passed. Full just check again stopped at the unchanged historical binary-vocabulary artifact's mismatch against installed build 18.0.9628; no historical binary data was re-anchored as part of this consumer test.

H4 incoming parameter consumer follow-up (2026-09-17)

Mapped IAction::getParam to the captured b9246 ACTION_zoom::onExecute index-zero call and exact build-18.0.9628 HTTP execute/readback comparisons. The initial query fixture failed its contrasts, and direct constant-to-zoom chains did not propagate the source. Completed pipeline and independent confirmation runs show that param_cast float supplies the omitted zoom operand; explicit decimal/zero values override it and explicit relative values use the current zoom baseline. Same-source direct-chain controls separate the bridge from bare default behavior.

Both interrupted pipeline attempts remain aborted captures with verified restoration and uncertain writes excluded from subsequent suites. Each mutation now starts on a fresh HTTP connection without retrying a lost response; the original timeout cause remains unproved. Frozen suites, exact readbacks, controls, journals and restoration are joined through incoming-parameter-selection in just runtime-grammar --audit; per-verb evidence is in just verb zoom. Second incoming slots, flag branches and other consumers remain open. The unmapped triage's next priority is the distinct typed/float getValues overloads.

Corrected the grammar summary's blanket XML-only backtick claim using the already verified effect consumer captures. Runtime-grammar, incoming evidence, transport, allowlist and reference-status checks passed. Full just check stopped at the unchanged historical binary-vocabulary artifact's re-extraction mismatch against installed build 18.0.9628; historical data was preserved. For efficient follow-up, use the scoped capture query in the run report rather than opening full journals.

H4 pair-reader consumer follow-up (2026-09-18)

Extended the b9246 bounded caller scan to both exact getValues overloads. Verified typed param_add and float param_multiply associations are now separate obligations; a false-positive byte candidate remains explicitly unverified. just runtime-grammar --callers <exact-symbol> provides a compact, checked lookup and rejects symbols that were not scanned, avoiding full disassembly context loads.

Read-only build-18.0.9628 param_multiply comparisons and an independent confirmation preserve frozen predictions, both operand positions and shaped controls. Numeric, raw-action-text and paired-backtick cases discriminate. Percentage, millisecond and computed-beat outputs retain units; direct beats return error:1. These fail the historical scalar-result predictions and do not prove execution of the b9246 float route on the newer build. Malformed/numeric-text zero outputs match controls and are not evaluation proof. just verb param_multiply records the bounded query result; execute, incoming input and cache reuse remain unmeasured.

The pair symbols leave the unmapped triage only because each now has an explicit consumer association, not because its branches are closed. Next: establish the current-build multiplication route and partition the overloads separately. The runtime-grammar, capture, caller-validation and reference-status checks passed. Full just check again stopped at the unchanged historical binary-vocabulary re-extraction mismatch against installed build 18.0.9628; historical data was kept.

Ready Tasks

0b. Topic Search Across Every Corpus

Status: Done

Note: 2026-09-06. The metadata layer landed, and with it the two problems it was meant to fix plus one it exposed.

The gap was worse than described. just topic waveform returned four hardware helpers (get_numark_waveform, gemini_waveform_zoomlevel, …) and not one skin element — the entire waveform surface (wave, scratchwave, rhythm, rhythmzone, zoomed, songpos, grid, gridlines, beattunnel, blockwave) was unreachable, and only the doc pointer saved it.

docs/topic-tags.json is the fix: 15 topics and 23 aliases, the only hand-maintained input to just topic. Tagged hits are marked + in the report and the topic prints why it needed a tag. just topic "color fx" now redirects to colorfx, "beat grid" to waveform, and waveform returns the eleven elements plus the built-in skins that use them.

Not a topics: [] field on each verb record, deliberately. Topics are cross-cutting, so a per-record field scatters one topic's membership across a thousand records and makes it unreviewable; it would also put a navigation aid inside the store that is authoritative for per-verb evidence. One map, reviewed whole, is the maintainable shape. The reason is recorded in the file's own _meta so the next agent does not re-litigate it.

Tags cannot rot. just check fails on a tag naming a verb, element or doc that does not exist, on an alias pointing at an undefined topic, and on a topic with no stated reason. A tag that names nothing is worse than no tag: it silently promises reach it does not have.

The bug the work exposed. "EXAMPLE FILES — real usage, grep-verified" was dominated by tests/*.json — the verb table, the sweeps, the corpus. A verb name appears in those because it exists, not because anything uses it, and they were crowding out the working examples the section exists for. The evidence artifacts are now excluded from that grep, so just topic sampler leads with real pad XML instead of tests/verb-table.json.

Left open. Tag coverage is the 15 topics where the gap was demonstrable, not a taxonomy — add a topic when a search visibly misses something, and state why in the entry. Effects are still matched by a whole-record substring search, which is generous and occasionally noisy.

2. Characterize FX Bank Save And Load

Status: Done

Note: 2026-07-26, HTTP. A bank is a rack of effect SELECTIONS for slots 1-6 — not active state, not slider values, and global across decks. effect_bank_load returns true/false as a bank-populated probe. Recorded in the tracker and on effect_bank_save/effect_bank_load (just get-verb effect_bank_save).

Start here:

Read first:

Record results in:

Promote to:

Done when:

  • Restored effect names, active states, slider values, and deck scope are recorded.

4. Keep BeatGrid effect_command Plugin-Specific

Status: Done

Note: 2026-07-26, HTTP. Confirmed plugin-instance-scoped (targets the BeatGrid slot), with a bare form and an unquoted-slot-number form; get/set/cur are BeatGrid's own vocabulary. Recorded as BeatGrid-specific, not generic. See just get-verb effect_command.

Start here:

Read first:

Record results in:

Promote to:

Done when:

  • Confirmed effect_command examples are documented as BeatGrid-specific rather than generic plugin control advice.

5. Author And Load-Test A Minimal Custom Device Definition

Status: Done

Note: Custom device-definition loading and mapper firing completed 2026-09-12 on build 18.0.9598 using paired virtual CoreMIDI ports in the existing SIMPLE_MIDI_0_0 context. The app displayed the authored device description; named note/CC controls fired, with wrong-address/channel controls and independent live variable-window readback. Fixture, exact bytes, results and cleanup.

The archive reader also recovers every original XML member from bundled controllers.dat, including definitions, factory mappings and audio presets. The earlier assumption that the compiled definition layer was opaque is corrected in Mapper XML and Compiled Controller Definitions. just controllers queries the generated inventory; just controllers-extract emits XML. Extraction is Tier 2; the live fixture validates only its tested MIDI behavior.

The earlier physical DDJ-GRV6 mapper-firing result remains in the tracker. The historical browser_filter/browser_search/none naming questions are covered by the verb table and existing disproof records; they are not prerequisites for this now-completed definition test.

6. Continue Hidden Button Editor Candidate Probes

Status: Done

Note: 2026-09-06, HTTP, build 18.0.9598. Both done-when clauses are met: every one of the 38 flags == 256 names is now recorded Pass, Partial or Fail, or explicitly blocked, and none is left as active untested work. Full narrative: the tracker's "Editor-Hidden Verbs: Behavior, Not Existence". State restored — pad page back to 1 CUE, deck volume back to 1, decks empty.

The Flip family closed entirely, and it needed no gate. The candidates doc had deferred all six pending "Flip availability and how recording is gated"; a loaded track was enough. Driven end to end on a disposable generated fixture with the deck volume at zero: flip_record is a toggle whose first press reads Rec Standby and whose recording begins on the first cue press, the status counts up, a second press stops it, and flip_load flips no → yes because a flip now exists; flip_play jumps to the flip start and plays it; flip_loop and flip_arm toggle cleanly. The reusable find is that flip_get_status is a text query absent from the catalog — '' / Rec Standby / Rec MM:SS / Play MM:SS, the display string a Flip control wants.

effect_beats_sliderindex verified against an independent oracle. It takes an effect name and returns the 1-based index of that effect's beats/length slider: BrakeStart → 1, Backspin/Echo/VinylBrake → 2, Beat Brake/Reverb → 0, matching the FX catalog (built by a separate sweep) on three distinct answers. A nonsense name also returns 0, so 0 conflates "no beats slider" with "unknown effect" and cannot probe existence.

A trap in the pad-page trio. get_pad_page_name <n> is index-only and follows padsPagesOrder; pad_page_favorite <n> answers only for 1-4, so it addresses a four-slot favourites bank rather than a per-page flag; and pad_page_insplit '<name>' takes a name, not an index — but it tracks the current page. Switching to 2 SYNC moved the yes with it. With no split layout configured, "part of a split" and "is the current page" cannot be separated, so its catalog meaning stays unconfirmed and no split indicator should be built on it yet.

all_decks and combine_query are not query-position verbs — error:-2147467259 bare and deck-scoped, while every other hidden verb in the sweep answered. Consistent with script-structural prefixes that only parse in execute position, which was not tested.

Two names deliberately not executed: crash, now marked blocked — verb-table membership already proves the name and that is all this repo wants from it — and browser_colorfilter_edit, which opens a modal dialog (query position gives error:-2147467263, a different code from the two above, matching its action-only kind).

Ten hardware-gated names marked blocked, under two different gates — the first labelling flattened them and was corrected the same day. controllerscreen_action (needs a controller with a screen) and assign_related_controller need only an attached controller, which the operator has, so they become testable whenever one is plugged in; get_controller_name returned '' during this pass and is the presence oracle to check first. The five rane_* names, ns7_get_drift and motorwheel2/3 need specific vendor hardware that is not here and cannot be substituted.

Fifteen more are recorded Partial as return shapes only — is_colorfx, masterbpm, pad_pressure_switch, sampler_inputgain, send_nothing, shoutout, stem_volume, timecode_no_jump, load_security_shown and the two that require arguments — each with the state that would move it named in its evidence, rather than dressed up as a behavior claim. stem_volume in particular still wants a stem-analysed track, which is the candidates doc's own first-probe row.

8. Characterize The VirtualDJ Remote App Wire Protocol

Status: Done

Note: 2026-07-27 — transport, wire format, subscriptions, and actions are all verified in both directions; only minor open questions remain (see end of this task)

Settled with a live session (socket watcher + dns-sd + per-connection nettop deltas; recorded in the tracker, docs/HTTP Control Interface.md, and docs/Application Internals.md Remote Skins):

  • Remote does not use the Network Control HTTP channel; port 80 saw no Remote traffic.
  • Discovery is inverted from the obvious guess: the phone advertises Bonjour type _vdjremote8._tcp (SRV → phone, port 4243 observed) and listens; VirtualDJ connects out to the phone as the TCP client, one persistent connection.
  • Semantics are event-driven push: idle seconds carry 0 bytes on that connection; a deck load pushed ~249 KiB desktop→phone in one second with no inbound request; unload ~1.4 KiB; otherwise only sub-KB keepalives.

Wire format also DONE (2026-07-27) — see docs/Remote Protocol.md. Framing is 8JDV + u32 total length + u16 type; the device opens with subscription frames carrying ordinary VDJScript queries by id, and VirtualDJ pushes typed values (val float32 / txt / fail) plus browser folder XML, settings, and selected-folder state. Replaying a captured opener is enough to hold a session — no pairing token. Capture tool: python3 tools/vdjremote_dial.py <device-ip>; reference capture at tests/vdjremote-opener.bin.

Subscriptions also DONE (2026-07-27): the vocabulary is all of VDJScript, not a fixed schema. Verified by substituting synthetic SUBSCRIBE frames into a replayed opener — get_version, get_effect_name 1, deck 3 get_bpm, and a full ternary all resolved, and push-on-change was measured (a load pushed title/artist/BPM/path within the same second; get_position streamed at 33-34 Hz while playing, silent when paused). KIND is a hint, not a request; fail means "no value now", not "bad query". Tool: python3 tools/vdjremote_subscribe.py tests/vdjremote-opener.bin 'left:get_title' 'get_clock' paired with a dns-sd -R advert.

Remaining:

  • Mid-session subscribe/unsubscribe is untested; only opening-burst registration has been exercised. A client that switches views needs it.
  • Map or bypass the 0x02 control id space. Only four ids are known, all from one device skin (0xc6 play, 0xc7 cue, 0x41 crossfader, 0x36 volume). Whether the space is global or skin-defined is open — but 0x31 may make it moot for third-party clients.
  • Mid-session subscribe/unsubscribe is untested — only opening-burst registration has been exercised. A client that switches views needs it.
  • Undecoded types: device→desktop 0x09, 0x0c, 0x27, 0x29, 0x34; desktop→device 0x2b, 0x3b. Sessions work without understanding them (replay reproduces them), so this is lower priority.
  • Waveform data has not been located in any frame; check inside the 0x25 ZIP payloads.

Record results in:

Promote to:

Done when:

  • Action frames are catalogued with example payloads, and the subscription vocabulary is characterized as either "any VDJScript query" or a documented subset.

9. Map Verbs To Button Editor Categories

Status: Done

Note: 2026-07-27 — verb set, aliases, hidden flag, and categories all extracted

The exact verb set is settled: tools/extract_verb_table.py extracts VirtualDJ's own verb table (1,028 sorted 16-byte records {const char *name; uint32 id; uint32 flags} at 0x10402d020), covering 1,007/1,007 HTTP-proven names. Existence and non-existence are now a membership test (just verb-table <name>), aliases are read off shared ids, and flags == 256 marks the 37 Button-Editor-hidden names. 61 alias groups are written into the verb store.

Categories are solved too: a const char *[38] name array after the verb table, plus a non-decreasing uint8[956] in __TEXT,__const giving the category per verb indexed by id + 1. Both located structurally, no pinned addresses. Confirmed by the live Button Editor list (read off screen, flow..video with defines skipped), and reproducing docs/Button Editor Taxonomy.md's per-category counts exactly — all 37 categories, total 1,028. See docs/Undocumented VDJScript Candidates.md §Categories.

Note (2026-07-29) that the count agreement is a reproduction, not corroboration: extract_vdjscript_taxonomy.py reads the same three structures, only by pinned address instead of by anchor. Rule 1c2 in docs/Evidence Standards.md now says so. The live Button Editor remains the only independent check.

Four-verb residue: RESOLVED as doc error. mute, silent_cue, stems_split, and loaded were listed under different categories in the taxonomy doc's example column than the mapping gives. That column is hand-written, not extraction output: the extractor emits visible_examples[:8] over an alphabetically sorted table, yet every doc row has exactly 4 entries and four rows are not even alphabetical (play, pause, stop, silent_cue; mic, linein, aux_volume, mic_eq_high; leftdeck, rightdeck, masterdeck, pfl; beatjump, clone_deck, dualdeckmode, mute). The counts column is extraction output, so the two columns have different pedigree and only the counts carry weight. Mapping stands; the doc's example column is corrected in place.

9b. Remaining Verb-Name Structure Notes

Status: Done

Note: 2026-07-29 — both "done when" conditions are met: every HTTP-proven name is accounted for by a named structure (the verb table covers 1,007/1,007), and every structure-found name absent from the store has been added (the 35 hidden verbs, 2026-07-29). Kept because the corroborating sources are still wired.

Three structures are extracted so far (tools/extract_binary_verbs.py, 1,019 names): 954 ACTION_ implementation classes, 812 language-catalog entries, and the parser's 967-entry alphabetically sorted name table. Their union covers 998 of the 1,007 names the HTTP sweep proved real.

Four names prove at least one more structure exists: browser, config, preview, and volume are real verbs, are ≥4 characters (so strings sees them), and appear in none of the three sources. (jog, no, off, on, yes are also missing but are simply under the strings length floor — use strings -n 2 for those, and note a low minimum breaks the sorted-run detection, so probe them separately.)

Why this is worth doing: the app must match input against a complete set of names, so the sets exist in the binary. Recovering all of them yields the exact verb set — including any undiscovered verbs — which is the strongest possible outcome for this topic, and it lets the disproof drop its conservative string leg (see docs/Evidence Standards.md rule 1a).

Start here:

  • Locate browser / config / preview / volume occurrences and inspect their neighbourhoods for table structure, as was done to find the sorted name table.
  • The sorted table runs action_deck … zoom_vertical and omits load, loop, cue, hot_cue, nothing — so it is a subset dispatcher, and whatever holds those five is another candidate structure worth characterizing even though the union already covers them.
  • Both older extractors are stale on this build (extract_vdjscript_symbols.py returns 0; extract_vdjscript_taxonomy.py is address-pinned) — re-anchoring the taxonomy tables would recover the Button Editor's own category structure, which is a likely home for the strays.

Done when:

  • Every HTTP-proven name is accounted for by a named structure, or the residue is explained.
  • Any name found in a structure but absent from the store is swept (just verb-probe) and recorded as a candidate.

10c. Make The Contract View Say What It Can Vouch For

Status: Done

Note: Added 2026-09-11 from a two-round review of task 10. The instruments exist; what is missing is a per-verb view that answers what can I rely on, within what scope, and what is still open — and closure rules in just coverage that do not claim more than their test established. Do this before collecting more evidence: it is offline, and it is the stopping-point and next-test that every later batch will be recorded against. Do not add a reporting command or a Markdown document; extend just verb and just coverage.

Landed with the review (2026-09-11), so not on the list below:

  • probe-arg-forms / probe-execute-forms no longer shell-redirect into the artifact. The tools take --out FILE and write atomically only on a real run; --dry-run used to truncate tests/verb-arg-forms.json to zero bytes and --merge read it after the shell had already emptied it.
  • just verb prints the execute-position row it was already loading (verdict, recognized tails, how many tails the readback could not separate from nonsense, or why it was skipped).
  • H4 flipped to Ready; 10b's opening line no longer says the sweep is unrun.

Landed 2026-09-12 (items 1-4 below): assess() in tools/coverage_report.py is the one per-verb assessment and just verb renders it as a Contract: block; CLOSED is settled only, no_known_candidates replaces the old closing label, execute reads the row's verdict (partial / undiscriminated / no_observable), and a verb with no capability row is unknown and counted open. A state-limited live result that had lived only in the tracker (get_time_hour elapsed/remain, 0 on a 2:26 track) now reaches the assessment through the catalog tool's LOCAL_UNDISCRIMINATED table and the documented_but_undiscriminated_here cross-check bucket, so just verb get_time_hour says undiscriminated and names the long-track fixture rather than never probed. Coverage after the change reads lower on arguments and execute; that is the labels telling the truth, not lost evidence.

Review follow-up (2026-09-12): regression tests now gate the shared assessment. Failed or unstable position captures stay open; an unmeasured shape is not closed by recognized keywords, and query confirmations cannot close execute evidence. Known-token coverage is labelled vocabulary_covered, not a full argument contract. Aborted execute runs preserve the previous output and save partial results separately with a failing exit. The view still cannot supply missing build stamps or infer complete value/overload coverage from legacy captures; those need explicit observations in later batches. Next is 10d, which doubles as the validation run.

Follow-up correction (2026-09-12): the blanket vocabulary_covered residual above was too strong and is superseded. Arguments close when all known obligations are resolved; remaining obligations must name a catalog value/multiple-argument form, attested shape, unresolved token or unaccounted binary argument-demand lead. The BPM-transition artifact is joined directly as execute-with-independent-BPM-readback evidence, so a toggle-only capture's nondiscrimination cannot reopen its confirmed landing-BPM forms. Capture-time build provenance is part of 10d below, not a separate task.

Follow-up 2026-09-12, the evidence_in_prose bucket. It held 37 verbs and two kinds of thing. Eighteen were behaviour-only prose (deck_has_error stayed off) that said nothing about arguments; the assessment now asks whether the evidence names a form or the argument before using the label, and those read unprobed with the note that behaviour evidence exists. Nine were the effect-introspection verbs whose NAME-form result had a real artifact all along — tests/fx-introspection-dump.json reads every installed effect by name — so the dump is joined directly through the sweep's own NAME_FORM_VERBS declaration, the same way the BPM-transition capture is, and those verbs close with the dump's provenance (product version only; it predates build stamping, and the claim says so). What is left in the bucket is genuinely prose-only: get_date's strftime value form, get_effect_slider_name's plugin-channel result, the two get_video*_name verbs whose argument is ignored (a negative finding), is_releasefx's undiscriminated slot/name forms, the grammar wrappers deck / all_decks / nothing, and browser_colorfilter_edit's E_NOTIMPL. Locate and structure each existing observation first, including the plugin capture for get_effect_slider_name. Run a new probe (for example, a get_date position probe) only if the recorded evidence is insufficient. Check the current count with just coverage; do not copy it here.

Review correction: the FX join now checks each verb's own result field, including the conditional slider/button rows, before claiming a measured form. It excludes error and missing results, retains the legitimate blank skip-length label, and reports a concrete effect/index/value. Legacy count zero is ambiguous because the sweep converted errors to zero; positive count measurements support that form. The prose filter is only a routing heuristic: a miss means no explicit argument reference was detected, not "never probed".

Checklist as landed, each line with the command that proved it:

  1. One per-verb assessment, shared. tools/coverage_report.py already builds a per-verb dictionary (dimensions, recognized/unresolved tails, positions probed). Lift it into a function just verb imports, so the two commands cannot drift. Proof: no second copy of the closure logic in tools/verb_summary.py.
  2. Loosen three closure rules in coverage_report.py, in the same change:
    • argument_less fires on probed and not candidates, but probed is also true for a verb that only appears in the positions or execute artifact. Rename to no_known_candidates, require an actual tail-prober row, and drop it from CLOSED. Exhaustive absence is not claimed.
    • execute is settled on row presence. Read the row: has-execute-tokens with the tails it recognized is settled for those tails; tail-ignored-in-execute and skipped are open with the reason carried. auto_bpm_transition is the type case — its toggle readback cannot see a destination-BPM argument, and the row says so.
    • A verb with no contract row gets queries/executes false and both dimensions n/a. Make that unknown so missing capability data cannot shrink the denominator. Proof: just coverage shows the new labels, just check passes, and no sentence in this file or the coverage audit quotes the old label (rg argument_less).
  3. A contract block at the top of just verb, rendered from the shared assessment: Contract: settled | partial | open, then per dimension one line each — what was observed, on which channel and build, and the smallest test that would close what is open. Every unresolved item carries one of five reasons: not measured (unswept, unprobed), measured but the observable did not discriminate (untyped, tail-ignored-in-execute), state or surface unavailable (blocked), conflicting observations (the disputed list), evidence only in prose (evidence_in_prose). These are the existing labels; the change is the unit, from per-dimension to per-form claim. Proof: just verb crossfader_curve reads as partial — bare float over HTTP, six names settled by HRESULT on the plugin, no execute evidence — and just verb get_time_hour names the long-track fixture as its next test.
  4. Write task 10's completion criterion as one sentence, in task 10's note: every verb either has an evidence-backed contract for the stated build and surfaces or a named open item with one of the five reasons above; just coverage --frontier is the test. "Every undocumented overload found" is not finite and is not the criterion.

Done when the four proofs hold and just check passes — they did on 2026-09-12. Then 10d.

10d. Long-Track Time Fixture: get_time_hour And Its Neighbours

Status: Done

Note: Completed 2026-09-12 on live build 9598 over HTTP. tests/long-time-forms.json records independently established runs, reversed form order, phase readbacks and verified restoration. just long-time-forms get_time_hour reports the result; just verb get_time_hour now closes return type, arguments and behavior using this stamped capture. elapsed is accounted for as the documented elapsed fallback, not a token separated from nonsense. get_time_sign remains undiscriminated in positive-time states. The final fixture uses the exact midpoint 3750s and fractional position 4207.125s; scaling get_position before HTTP serialization provides the needed precision. A focused runner preserves phase controls and restoration rather than merging into the generic prober's incompatible fixture set.

Original plan, added 2026-09-11. The first live batch after 10c, chosen because the expected differences are already worked out and the state is cheap to manufacture. The recipe is in 10b's note ("get_time_hour wants a track longer than an hour"): generated 2h05m track, playhead at 1h10m so elapsed/remain/total read 1/0/2, read once with display_time on elapsed and once on remain, then place the playhead where pitched and unpitched remaining time straddle an hour boundary (about 3,700 s in at +12%) to separate absolute. Do not restate the recipe here; extend tools/fixtures.py with the state and run tools/probe_arg_forms.py --verbs … --fixtures <new state> against every time reader that the state can discriminate in one session — the get_time* family and display_time readback, batched, not one question per round. Record per-verb conclusions with just put-verb … confidence=local_test evidence="… build …, HTTP", the fixture and any non-discriminating readings in the tracker, and check the artifact in with its stamp. This run doubles as the validation of 10c's view: after it, just verb get_time_hour must show the dimension closed with the observation, not Untested.

Prerequisite: just vdj-up (reachable on 2026-09-11) and ffmpeg for the generated track.

Capture checklist: both argument and execute probers read get_build from the running instance before fixture writes, and include that value, query, HTTP channel and UTC time in the new capture's summary and observations. Do not copy the installed binary's stamp or apply a new stamp to legacy rows during a merge. Check just verb's per-claim provenance after recording 10d; older unstamped observations must remain explicitly unknown.

11. Build The Verb Index From The Artifacts, Not From Prose

Status: Done

Note: 2026-09-06, offline. The generator is inverted, the schema is unchanged, both consumers still pass, and the reconciliation — the actual deliverable — is recorded in the tracker's "Verb Index Inversion: What The Prose And The Artifacts Disagreed About".

The defect was live and reproducible. lint_mappers.py reads the index, and before the inversion it called flip_record, flip_play and get_pad_page_name unknown verbs — all three real, all three characterised the day before. Every one of the 37 editor-hidden names was missing because the prose never listed them. The same file now lints with 0 verb warnings.

The whole diff was 36 entries (959 of 995 shared entries differed only by an empty "aliases": [] the old generator emitted), splitting exactly into the two buckets this task predicted:

  • Artifacts correcting the prose — 30 alias facts. 17 pairs the prose missed (browser is an alias of browser_zoom, preview of prelisten, goto_bar of goto_beat_in_bar, the four eq_*_slider spellings, …) and three claims the table contradicts outright: the prose made pitch2 canonical for pitch2_slider when all four pitch* names share id 722 under pitch, and made scratch_wheel_touch canonical when all four share id 153 under touchwheel_touch. No name was contradicted on existence — the prose never invented a verb, it mis-ranked aliases and omitted the hidden set.
  • Curated facts with no artifact home — 2, repaired rather than dropped. auto_bpm_transition and auto_bpm_transition_options carried richer prose than the store, so the store was updated to hold it before switching. Every other alias row that lost a description had only Official alias of X boilerplate, which the canonical field now states structurally.

The schema gap it exposed. Six store names are absent from the verb table, which is silent on them rather than negative: ONINIT, while_pressed and deck are structural keywords rather than verbs, and browser_filter, browser_search, none are names this repo disproved. Index rows carry not_in_verb_table: true rather than promoting or dropping them. The store field this names — a kind meaning "structural keyword, not a verb" — is not added; it wants a second case before being designed, and step 3 of the plan is deliberately left open rather than guessed.

Also fixed in passing: lint_mappers.py raised ValueError on any path outside the repo, so linting a scratch file produced a traceback instead of a lint.

Not reasons to do this, recorded so they are not resurrected

The trigger for this task was "should the monolith become verbs/<category>/<verb>.md?" — assessed 2026-08-11 and mostly not worth it:

  • Per-verb files: no benefit. The only problem a file tree solves is finding one verb without loading 6,300 lines, and just get-verb already solves it better — it joins five artifacts at read time, where a checked-in page can only be a stale copy of that join. 955 files, nothing gained. Per-verb pages belong in the HTML export (Parking Lot): generated at build time, never committed.
  • Family files: a small benefit, and not a retrieval one. The one thing just genuinely cannot do is narrative — it returns records, not "how the transport verbs relate to each other". That kind of writing wants related verbs on one page. Real, but a nice-to-have; the frozen consolidation plan already chose families over per-verb, so nothing needs redeciding. Do it if the prose becomes hard to write, not on a schedule.
  • Dropping the 1,175 catalog rows is conditional, not a driver. They duplicate the store and can drift, but they are harmless while nothing reads them and they only become removable after step 1. The ~1,000 lines of curated prose stay hand-authored regardless — that is the part no artifact can produce.

Ordering note: 10b adds a fifth artifact. Doing this first means it joins one generator; doing it after means reconciling a Markdown parser against it too. That is a convenience, not the justification.

Done when just verb-index reads no Markdown, just check is green, and the step-2 reconciliation is recorded.

14. Run The Corpus As A Parse-Regression Set

Status: Done

Note: Expanded corpus regression shipped in 675addf.

The 2026-09-05 review verified 1,610 snippets through just check. The counts and results below describe the earlier 2026-09-03 run, not the current corpus. Query the current artifact/check for current results; this task is not startable work.

Historical setup (2026-09-03): the first check that could falsify a grammar claim rather than extend one.

Every one of the 1,427 snippets in the corpus is a form the vendor considers valid — 1,131 from shipped Built-In XML, 269 quoted in the catalog's own descriptions, 41 from the wiki. Send each through the read-only query channel and any that fail to parse contradicts something this repo believes about the grammar.

Design constraints, learned the hard way this session:

  • Query only. Corpus snippets contain load, unload, browsed_song color, sampler and broadcast verbs; executing them would rewrite the library. The channel's own return value proves nothing either (rule 4) — the observable is whether the parse errors, not what it says.
  • E_FAIL is silence, not denial. A snippet returning E_FAIL is not evidence of a parse failure; only E_INVALIDARG-style structural errors are, and even those need the nonsense control treatment before being called a contradiction.
  • Snippets naming effects, skins or files this install lacks will fail for environmental reasons. Classify those separately or the signal drowns.

DONE 2026-09-03. tools/check_corpus_parses.py (just corpus-parses, gated in just check) sends all 1,427 snippets through /query: 1,099 parse and none contradicts a grammar claim. Outcomes are classified rather than counted — 152 not-implemented, 97 other-error, 28 E_FAIL no-value, 28 surface-gated (action-position verbs sent to a query surface), 15 pipeline (& param_* needs an inbound value HTTP cannot supply), 1 placeholder, 7 structural.

The seven residual cases turned out to be execute-position semantics — loop 50% multiplies rather than sets, pitch_range +1 and sampler_loop ±1 are relative — and each has an absolute equivalent that answers normally. Written up in VDJScript Grammar; a relative or multiplying argument is execute-only, and asking a query for one errors rather than going silent.

It also falsified something on its first run — my own extractor. Eight snippets read color &apos;red&apos;: extract_script_corpus.py never XML-unescaped attribute values, so the corpus had been storing escaped entities since it was built an hour earlier. Fixed, and it is exactly what a regression set is for.

15. Two Loose Ends

Status: Done

Note: Both loose ends closed; details below. Added 2026-09-03.

  • timecode_cd_mode — CLOSED 2026-09-03. VirtualDJ restarted at 09:42 and the verb reads no on all four decks, back to its pre-probe value. So it is runtime-only state: settable from script, not clearable from script, and not persisted to settings.xml. The 2026-09-03 execute-probe incident is closed with no lasting change to the user's configuration.

    Worth keeping as a documented verb property rather than only an incident: a verb can latch for the life of a session. The execute prober's pre-flight round-trip now catches that class before probing (it flips once and requires the value to return), and timecode is back inside the category deny-list that was inert at the time.

  • What the sampler all means — DONE 2026-09-03. It is the selected slot, not an aggregate: with slot 1 selected get_sample_name all reads slot 1's name, after sampler_select 5 it reads slot 5's, and sampler_loaded all stays yes while sampler_loaded 8 is no. Recorded in VDJScript Verbs, along with the sampler_select quirk found while restoring state — 1-based slot on execute, 0-1 slider value on query.

Progress log for open tasks

No open task carries a progress log here at present. The H4 log moved under its closed block on 2026-09-19.

S4. Establish a build-scoped button XML ownership pilot

Status: Done

Note: 2026-09-22. The initial structural pilot on build 18.0.9644 (arm64) preserves outer/shared/child XML read paths and unresolved alternatives, with verified memory-capture image identity and reproducible disk analysis. just skin-schema button is the compact entry point. No skin behavior was promoted and no complete-schema claim was made. Skin Schema Recovery records the method, remaining conditional-node/helper gaps and the next discriminating fixture work. General class extraction now supports fresh output without requiring a historical app or overwriting older captures.

S5. Resolve conditional child ownership and test it live

Status: Done

Note: 2026-09-22, build 18.0.9644 arm64. Guarded current-build models now retain named child and same-name sibling paths through the conditional selectors. just skin-schema button uses the new capture and separately joins the scoped live result. The synthetic pos/size/up fixture separates false conditions from true/nonsense controls through click readbacks and saved screenshots, repeated after reload in reversed order. Original skin and empty/stopped decks were restored; installed fixture removed; initially unset test variables remain zero. Skin Schema Recovery and the linked fixture retain the exact limits and execution anomalies. Full schema completeness, outer/child precedence and additional XML getter families remain open.

S7. Read the Atomix add-on skins into the skin XML inventory

Status: Done

Note: Completed 2026-09-24. The inventory and literal nesting graph now read Atomix add-on skins with separate addon provenance; the inventory lists source files per element. New names were triaged in docs/Skin SDK.md, and the deliberate zzscratchwave negative control was excluded. The reader vocabulary was re-anchored to installed build 18.0.9644; just check passes.

Original scope: opened 2026-09-24. The add-on trees (examples/Skins/Official-Addons/, 32 skins in 40 XML files) reached the repo on 2026-09-21/22 and joined the script corpus in bff50cd, but the XML vocabulary side never read them: tools/extract_xml_inventory.py and tools/skin_relations.py both carry fixed glob lists naming only examples/Skins/Built-In/, the SDK browser example, this repo's own skeletons and tests/Skins/. So just element, just list-skin-elements --undocumented and the observed-nesting queries still answer from the bundle skins alone. This task closes that gap. It is desk work: no live instance needed.

Why it is worth a task rather than a one-line glob edit. A first-pass comparison against the committed inventory's skins family found roughly 12 element names and about 226 element/attribute pairs that no bundle file writes — led by tooltip_localized (214 uses), filelist (63), pannel (29), panel@deckn, textselected@format, panel@brheight, down@nb/down@smooth, textzone@geterrormessage. That figure is a rough scan, not a finding: some names may already be documented under another family, and pannel is probably the misspelling the repo has already recorded. Verify each before writing any of it down.

The blast radius is the reason for care. Regenerating the inventory also moves extract_skin_readers.py (which diffs the inventory against the binary's reader vocabulary), extract_skin_classes.py, skin-xml-relations.json, just topic, and the undocumented-element and attribute lists the docs cite. A batch of genuinely new vendor vocabulary lands as new documentation gaps in docs/Skin SDK.md, which is the point: the reason to read these files is that the controller screen skins and the 2018 default skin exercise browser, sideview and playlist surfaces no bundle skin does.

Work, in order:

  1. Decide the grading first, before any regeneration. The add-ons are Tier 2 vendor examples, same as Built-in skin, but they are catalog downloads rather than bundle members, so tools/check_bundle_copies.py cannot hold them against a file the installed app still ships. The script corpus answered this by giving them their own addon source; decide whether the inventory and the relations report need the same separation, or whether a shared Published skin label is enough. Record the decision where the tool declares its sources.
  2. Add the roots and regenerate, then read the diff rather than trusting the totals: python3 tools/extract_xml_inventory.py, just skin-relations, and whatever just check then reports stale — expect the reader and class artifacts to follow, the way attested-tails and the action catalog followed the corpus change.
  3. Triage the new names, one at a time. For each: is it genuinely absent from the docs, or present under another family or another spelling? just element <tag> and just grep-verb-docs answer most of it. A name that only the add-ons write is the interesting half — say so per name, with the file that writes it.
  4. Route the real gaps into the docs. Element-level gaps belong in docs/Skin SDK.md; attribute-level gaps that no doc explains are what just element --undocumented is for, so leaving them listed is a legitimate outcome. Do not infer semantics from a name: an attribute appearing in shipped XML is attested vocabulary, never a tested behavior (Evidence Standards).
  5. Keep the counts honest. Inventory totals are derived artifacts, so name the command that answers rather than writing the number into prose; a per-build observation gets a stamp.

Start here:

  • tools/extract_xml_inventory.py — the FAMILIES glob list that excludes the add-ons
  • tools/skin_relations.py — the SOURCES tuple, same exclusion
  • examples/Skins/Official-Addons/README.md — provenance, add-on ids, zip hashes
  • tools/extract_script_corpus.py — how the same provenance question was settled for the script corpus (addon source)
  • docs/Skin Element Discovery.md — how the element queries are meant to be used