0.3.2

A voice on the subject, not on the timeline (issue #10). Every subject slot has an
Add voice button (Refs ON), or drop an audio file on the slot. That clip is sent to the model
as an <Audio N> reference tied to that subject:

<Audio 1> is the voice-timbre reference for <Subject 1> (S1).
<Audio 2> is the voice-timbre reference for <Subject 2> (S2).

It is not on the timeline, so it is not bounded by the render window and takes no output time —
two voice references and a background track no longer compete for one lane, and the render stays
as short as you made it. It does not depend on the audio-track switch either: a voice on a subject
is as deliberate as the subject's image. Voices go first in <Audio N> order (slot order), then
whatever sits on the timeline, inside H3's cap of three audio clips; a clip that does not fit is
named. The timeline's audio clips and their Voice of menu work as before. A slot with no
reference image cannot be a <Subject N>, so its voice goes out as a plain voice reference, which
the tooltip says. Override Audio together with a reference video still wins (a reference video's
soundtrack is numbered first), and it names the voice it set aside. The log warns when a voice file
is under the 2 s the model card asks for.

0.3.1

@Brioch's pull request #18, carried over by hand: it had stopped
merging after 0.2.3–0.3.0 touched the same planner and docs, so its behaviour was re-implemented
on the current code and its tests re-written against it. Every idea below is theirs.

• A reference no longer spends output time. A reference is an input to the model, not content
  in the video — <Video k> and <Audio j> are never composited — but a clip had to overlap the
  render window to be sent at all, and adding one stretched the window to cover it. The model card
  wants each reference audio clip 10–15 s long, so three of them grew the output to 45 s: three
  times the longest video H3 can make. Neither reference track touches the output length now,
  and a clip is sent from wherever it sits; it lands after the last one, which for anything
  longer than the window is out in the shaded area, where a reference belongs. A 5 s render
  sends all three. Position still decides one thing, for audio only: whether the clip is also
  part of the muxed soundtrack — inside the window it is both, past it a reference only, which
  the clip's info panel says. A parked reference video is sent whole on its own trim instead of
  losing its head to a window start it has nothing to do with. A retake keeps the old rule.
  (lock clips are not references and still have to be in the window: a lock is a position in time.)
• Every audio clip on the track plays its own file again. With two clips loaded both showed
  up in the prompt and both drew their waveform, but only one could be heard: the editor stamped
  the decoded buffer onto every clip that matched it, and a standalone clip has no blob URL to
  match on, so the test was undefined === undefined for the whole track. Only the preview was
  affected; what was sent to the model was always right.
• A reference clip that cannot be read says which one. A missing file or a trim past the end of
  the audio was dropped in silence — and because <Audio N> is numbered before any file is opened,
  that label and every one after it then named a reference the model never got. The log now names
  the clip and its label.
• Both per-type caps (3 audio, 3 video) name the clip they drop, like every other limit.
• Override Audio wins over the audio track outside the editor too: sent together, every
  <Audio N> label would be one off (a reference video's soundtrack is numbered first). The clips
  set aside are named, and Override Audio warns when the video it wants the soundtrack from is
  parked past the window.

0.3.0

Your own audio in the video, with a way to choose how — and the lines the model needs to believe it.

• A clip set to fully_copy / partially_copy now has a real chance of being copied. Setting
  the marker was always possible (right-click a clip) and it wrote the guide's
  retention_analysis line — but with non_diegetic_music: N/A and no soundscape line the
  model sang something of its own. Measured on an 8 s singing clip, 20 steps: without the two
  sound lines the audio was unrelated to the clip in all three runs (envelope correlation
  0.08–0.17); with them the clip itself came back in four runs of six (0.96–0.98) and not in
  the other two (0.10, 0.21), the same prompt on a different seed. So it is the guide's
  behaviour and the model follows it most of the time, not always. The Director now writes
  those two lines whenever a clip is set to a copy marker and the boxes are empty, in the
  guide's own wording ("<Audio 1> is directly reused as the complete audience-only score");
  text you write there still wins. The default stays reference, which follows a voice or
  style and never copies. For audio that must come out as given, use lock.
• New: lock (exact audio) on an audio clip. Not a prompt wording but a different
  mechanism, the one the LTX Director used: the clip is encoded with the audio VAE into the
  model's audio stream and held there with a zero noise mask, and the video is generated around
  it. The gaps between locked clips, and anything after the last, stay the model's. It works with
  Refs OFF too, where the guide offers no audio references at all. Measured at 0.99 in every
  run (five of five, both checkpoints), against four of six for the copy markers. A
  locked clip is not an <Audio N> reference, gets no label and does not count against the
  limit of three. It needs the audio VAE and says so if it is missing.
• Only one fully_copy. The marker means "the complete final audio track"; a second clip
  with it is written as partially_copy and the warnings area says so.
• The right-click menu now calls the audio setting Audio use and lists lock beside the
  guide's four markers; with Refs OFF it offers mixdown only / lock.
• README: a table of what each setting does to the sound, and what was measured.

0.2.5

• The node now says when a timeline audio clip is not being sent to the model. With
  References off (fl2va) the model has no audio input at all, and with the audio track
  switched off in ref2va nothing is sent either; in both cases the clip is carried only by the
  combined_audio output. That is how it was designed, but it looked exactly like a clip
  that was ignored: the voice in the result was the model's own, spoken from the text, and
  nothing said why. The warning names which case it is and what to do — wire combined_audio
  into CreateVideo to hear the clip itself, or use Refs ON with the audio track enabled so
  the model gets it as an <Audio> reference. Even then it is a reference the voice follows,
  not a copy of the waveform.

0.2.4

• prepend_text on the Director ([#29]). A connection-only text input; whatever is wired
  into it goes in front of the global prompt, joined by a comma. It was asked for so LoRA
  trigger words can come straight from a LoRA loader. It is deliberately not put ahead of
  the whole compiled prompt: the minimax format opens with the task statement and labelled
  sections, and text in front of them would break the shape both guides ask for. The
  Director's existing prompt output carries it, so a second pass can reuse the composed
  text (the other half of the request, which already worked). The panel's live preview cannot
  show it, because the text does not exist until the graph runs. The new socket is declared
  last, so workflows saved before it keep every value where it was.

0.2.3

Three bugs reported by users, and one control that never did anything.

• Save / Save As kept only part of the timeline (#23, reported and diagnosed by
  @SveSop). The Director stores its timeline from an allowlist,
  and the .json written by Save carried its own, shorter copy of it. overall_soundscape,
  non_diegetic_music and the hand-written prompt override (prompt_override,
  prompt_override_on) were on the first list and not the second, so they came back empty
  after a Load — no error anywhere. The two lists now match, and test_node.py compares them
  by key name so the next field added to one cannot drift from the other.
• Retake Stitch could not find the base video's audio when the input folder is not the
  default one (#30, reported with the fix by @Jaminblack).
  The frames were fetched through the media library, which resolves the timeline's relative
  filename against the input folder; the soundtrack was handed to PyAV as-is and looked for
  whatdreamscost/… in the working directory. The head and tail of the result came back
  silent and the log said could not read audio. The audio slice resolves the name the same
  way now, and says plainly when the file is not in the input folder.
• A base URL ending in /v1 no longer produces a 404 (#31). The Analyze request
  appended /v1/chat/completions to whatever was typed, so the address Anthropic's and LM
  Studio's docs give — https://api.anthropic.com/v1/ — asked for /v1/v1/chat/completions.
  Both spellings work now, and the tooltip and README no longer say "no path".
• Video Strength and Video Attn are gone from reference-video segments (#26). They are
  LTX conditioning weights inherited from the editor this one was forked from; nothing in the
  H3 planner reads them, so moving them changed no output at all. The stored values stay in
  old timelines and are ignored, as before.

0.2.2

Two contributions from @Brioch — #16 and #17, closing
#20 and #19 — and a new node.

• New node: MiniMax H3 Save Last Frame. It goes straight after VAEDecode, writes the
  last frame of the batch as a PNG the way Save Image would — same counter, same
  %date:…% tokens, same embedded workflow metadata — and passes the batch on. Every H3
  render ends on a frame worth keeping, since it is the one you feed back in as the next
  shot's opening keyframe, and fishing it out otherwise meant a second graph with
  ImageFromBatch wired to a SaveImage, rebuilt every time the length changed.

  Two things it deliberately does not do. It never touches the pixels: the IMAGE output is
  the same tensor object that came in, so it cannot change what anything downstream decodes
  or muxes. And it does not need to be bypassed — save is a widget, because a node that
  has to be disabled between runs is a node that will be left enabled by accident.

• The resolution panel covers every aspect ratio the model card lists, in both
  orientations. Six presets reached 16:9, 9:16 and 1:1; 21:9, 4:3 and 3:4 are in MiniMax's
  own output envelope and had to be worked out by hand — from an area cap, in multiples of
  32 — and typed into Width and Height. There are 26, at two sizes:

  Ratio	Native	Fast
  21:9	1344×576	1120×480
  2:1	1344×672	960×480
  16:9	1344×768	864×480
  3:2	1152×768	736×480
  4:3	1024×768	640×480
  5:4	960×768	608×480
  1:1	992×992	640×640
  4:5	768×960	480×608
  3:4	768×1024	480×640
  2:3	768×1152	480×736
  9:16	768×1344	480×864
  1:2	672×1344	480×960
  9:21	576×1344	480×1120

  Native keeps H3's 768 px short edge, except at the two widest ratios where 1344 is the
  long-edge cap and the short edge gives way instead. Fast is the same list at a 480 px
  short edge, 1:1 aside, which stays area-matched to its tier rather than dropping to
  480×480. Every edge is a multiple of 32 — H3's own step, and what divisible_by defaults
  to — so a preset is never quietly floored to something else on the way in.

  0.2.1's 1920×1088 is still there, under a Past native heading of its own, since it
  is the one canvas here that leaves the trained envelope and the heading is what says so.
  Everything above it is inside the envelope, so the list no longer mixes the two.

• Or name a shape and a pixel budget instead of a canvas. A new Aspect / MP row
  takes an aspect ratio and a figure in megapixels and fills Width and Height with the best
  pair of /32 edges that holds the ratio. Holding the ratio outweighs hitting the budget
  exactly, and overshooting the budget is penalised twice as hard as undershooting it, since
  memory is what a budget is protecting: 16:9 at 1.03 MP lands on H3's own 1344×768 rather
  than the 1376×768 that is closer to true 16:9 and 2.6% more canvas. Either box works on
  its own — a budget with no ratio picked rescales the shape already in the boxes — and
  typing Width and Height by hand still works, with both menus following along and reading
  — for a shape the ratio list does not name.

• A timeline image no longer collapses the canvas to one pixel. resize_image's
  cover-crop sized the scaled image with int(W * ratio), and a ratio that is exact in
  arithmetic comes back a hair under its integer in floating point — 704 * (480 / 704) is
  479.99999999999994. The cover then landed one pixel short of the canvas, the centre slice
  started at −1, and a 1px-wide image came out: a 704×1408 timeline image fitted to 480×640
  hit the Director's canvas guard as "the canvas came out 1x640", naming a width nothing had
  asked for. Roughly one source width in nine does this at a 480 px canvas, and the crop is
  the default fit, so it was reachable from any preset. Rounding rather than truncating — and
  requiring a cover to be at least the size of what it covers — makes it unreachable.

  The same truncation was quietly costing a whole 32-block in the aspect-preserving methods,
  where nothing crashed and so nobody looked: maintain aspect ratio fitted a 1920×1080
  image into a 1024×1024 box at 992×544 rather than the 1024×576 the code's own docstring
  promises.

• A voice reference's declaration ends on the speaker's global ID, which is what the
  guide means by "reuse that speaker's global ID in the definition":

  subject_definitions:  <Audio 1> is the voice-timbre reference for <Subject 2> (S2).
  

  The ID is the speaking order rather than the subject number, so it cannot be known when
  that sentence is written — it is filled in once the numbering pass has walked the finished
  body. A subject who never speaks has no ID to reuse and the sentence ends on the label,
  which already says whose voice the clip is. A hand-written declaration gets the same
  treatment when it names the subject, and is left alone when it writes an ID of its own.

• A spoken line stays where it was written. Dialogue is lifted out of a shot prompt to be
  rendered and was then appended to the end of the shot, which only looked right when the
  line already came last. A line with prose after it moved:

  @ref1: before mid segment          [Shot 1] mid segment. a woman (S1) says,
  mid segment                    →     <d>[English] before mid segment</d> a woman (S1)
  @ref1: after mid segment             says, <d>[English] after mid segment</d>
  

  It is now put back where it sat, which is how the guide writes a shot — its own Shot 1 goes
  action, the line that action motivates, then more action. The same pass decides both
  orderings the guide cares about, so they can no longer disagree: which mention of a subject
  is its first — and so is named in full rather than by its called name — and the order
  of vocal events that hands out (Sx). A frame anchor still rides with the prose rather than
  trailing the shot, since a spoken line is not something a shot "begins from". </d> no
  longer picks up a full stop it never needed.

• The live preview reports a line that reads as dialogue and stayed prose. The colon is
  what makes a tagged line dialogue, and @ref1 says "hello sir" — the natural thing to
  type — has none, so it reached the model as narration: no speaker ID, no <d>[Language] …</d>, and nothing for an <Audio N> voice reference to reuse. The line is not
  reinterpreted, because prose quotes things nobody says out loud, but it is no longer
  silent:

  [Shot 2] `@ref1 says "hello sir"` reads as dialogue but has no colon, so it stayed
  narration — no speaker ID, no <d>[Language] …</d>. Dialogue is `@ref1 <how it is said>:
  the words`.
  

  @char1 says: hello is reported the same way: the alias resolves to a subject label
  everywhere, but only @refN speaks.

• Two required fields are reported when they are empty. detailed_description joins
  overall_soundscape, which was already checked — and unlike the soundscape it has no
  N/A, so an empty one leaves the section out of the prompt entirely rather than visibly
  blank. A voice reference bound to a subject who never speaks is reported too, naming the
  tag that would give them a line: the declaration has no (Sx) to reuse in that case, which
  is correct and reads like a bug.

0.2.1

Four reports, and one of them was right about something this pack had been saying wrongly
since 0.1.0.

• width and height can be wired in (#14). The settings panel owns
  custom_width / custom_height and hides them, and a hidden widget's input slot is
  never laid out — LiteGraph left it at the node's own origin, so five invisible sockets
  sat stacked under the title bar. There was no reachable way to drive the canvas from a
  resolution node. There are now two connection-only inputs beside start / end /
  duration, and the five orphaned slots are dropped when nothing is wired to them (a
  link saved in an older workflow keeps its socket).

  A wire carries no minimum where a widget does, so 0 — what an upstream node hands over
  when its own value was never set — is refused by name rather than passed into the VAE,
  the same guard duration grew in 0.1.2.

• A subject description reaches the ComfyUI prompt format (#14). That format resolves
  @ref1 to the bare <Picture 1> and has no subject_definitions section, so a
  description typed into a slot had nowhere to go and was simply dropped: the prompt said
  "<Picture 1> steps out of the doorway" and never once said who that is. It now goes
  into the flat Reference notes: line, subjects before pictures, so the thing is
  introduced before the frame that shows it. The MiniMax format is unchanged — it has had
  the section all along, and started using the description in 0.2.0.

• An API key for a cloud vision model (#15). Nothing here ever sent an
  Authorization header, so the Custom (OpenAI-compatible) provider could only reach a
  server that did not ask for one. Both the Analyze button and the Enhance node send a
  bearer token now, and a 401 says where to put the key instead of echoing the endpoint's
  own body.

  Where the key is kept mattered more than sending it. The Analyze settings live in
  timeline_data, which is serialised into the workflow JSON — a key typed there would
  travel with every workflow you share. So the gear menu's field writes to ComfyUI's user
  settings instead, and the Enhance node's widget takes the name of an environment
  variable rather than a key, because widget values are saved with the workflow.
  Failing that: MINIMAX_DIRECTOR_VLM_API_KEY, then OPENAI_API_KEY.

• 4–15 seconds is the trained range, not a limit (#12). Nothing in this pack ever
  capped the length, but the README said "H3's trained range is the limit" and the console
  warned of "a VRAM wall" — which nobody here had measured, and which a 45-second render on
  a 5090 contradicts. Both now say what is actually true: it renders, quality leaves the
  envelope the model card describes, and the clock climbs faster than the video does
  because attention cost goes with the square of the sequence while memory goes with its
  length.

• The 2K resolution preset is labelled past native (#14). The model card's 2K is
  not these weights at 1920×1088 — it is H3-Regenerate-2K, a separate in-context
  regeneration pass, and MiniMax says "this module is not yet open-sourced. We will release
  it once it is ready." The preset still exists; it no longer claims to be something it is
  not.

• New offline test file, test_node.py — 35 checks for the parts that need ComfyUI
  imported, which test_plan.py deliberately cannot reach: the automation-socket guards,
  API-key resolution, and a check that every schema input has a matching execute()
  parameter. test_plan.py is up to 270.

0.2.0

Full-reference mode, from
references/ref-en.txt.
Until now a reference image was always a character, always followed as tightly as
possible, and a timeline image was always a frame anchor. The guide describes far more
than that, and two of its sections were structurally missing from the output.

• A reference no longer has to be a character. The guide defines <Subject N> as
  "people, animals, or objects; scenes, backgrounds, or environments; clothing, props,
  interfaces, or visual effects; styles, actions, expressions, or poses". Each slot now
  carries a kind that supplies the noun when no description is written, so a slot can
  hold a location or a look instead of a face. The panel goes from three slots to nine,
  and @ref1 … @ref9 address them. @char1 … @char3 still resolve.

• A subject no longer has to have an image, which is what made subjects unreachable on
  the Refs OFF (fl2va) path: the slot's text boxes only appeared once something had
  been dropped on it, and fl2va discards the drop. The one thing that path can carry was
  behind the one thing it throws away.

  references/base-en.txt
  has no subject_definitions section at all — with no reference files there is nothing to
  declare — so a subject there lives in the prose, "established when a speaker first
  appears" and referred to consistently afterwards. The slot now says both halves of that:

  Box	Becomes
  describes	the full identity, written where the subject first appears
  called	what to call it at every mention after that

  @ref1 places a fresh loaf on the counter.
  @ref1 says: First batch of the morning.
  
  → [Shot 1] a middle-aged baker with a calm, slightly raspy voice places a fresh loaf on
    the counter. the baker (S1) says, <d>[English] First batch of the morning.</d>
  

  Tags now resolve left to right through the finished video — the global block, then each
  shot in turn, its prose before its dialogue — so "first" means first on screen and not
  first in whichever box is being scanned. A subject introduced by its own spoken line is
  named in full there. An empty called repeats the description at every mention, which
  is what every timeline written before this field did. The same treatment applies with
  references on to a slot that has a description but no image: no picture, no
  <Subject 1> to lean on, so the prose has to carry the identity.

  Images left in a slot with references off are kept — switching the toolbar back must not
  cost the upload — but they are dimmed, and the prompt panel now says outright that
  fl2va sends none of them.

• The fl2va prompt now stops at the fields the base guide defines. Its structure is
  closed — the alignment instruction, then "Three Core Fields (Required, in this order)" —
  and summary is not among them; it belongs to the reference guide. A filled summary box
  was emitting a section H3 was never trained to read on that path, task-type prefix and
  all. The box is now hidden with references off and keeps its text for when the toolbar
  goes back.

  The other half of "required" also applies: non_diegetic_music is written as N/A when
  the box is empty, which is the value both guides use for having none.
  overall_soundscape deliberately does not get filled in the same way — there N/A
  means the video was asked to be silent, which is a claim only you can make — so an empty
  one is called out in the prompt panel instead. H3 generates the audio; leaving that field
  out hands the whole soundtrack to a guess.

  The 350-500 word figure is the reference guide's, for its own detailed_description.
  The base guide gives no word count, so the fl2va path no longer cites one. The word
  figure is still reported either way.

• A reference no longer has to be followed exactly. Every reference carries one of the
  guide's four retention markers, written into retention_analysis verbatim because
  the guide calls them "fixed English values in the output format":

  Marker	Meaning
  fully_preserved	The defined role of the referenced content is fully preserved
  partially_preserved	Still used, some defined characteristics changed
  attribute_transfer	Its characteristics move to a different target subject
  weak_reference	Broad similarity in style, category, composition or atmosphere only

  Audio uses its own set — fully_copy, partially_copy, reference, weak_reference —
  because copying a signal and imitating one are different jobs. An off-spec value coming
  from an edited timeline is clamped rather than passed through into the prompt.

• Both sentences about a reference are yours to write. The guide's lines name actual
  features rather than stock phrases — fully_preserved - the Samoyed's thick white fur, pointed ears, dark nose, and curved tail are retained — and name relationships the
  timeline cannot work out, like <Video 1> is the source video for the target video edit.
  Every reference therefore has up to two boxes, one per section it feeds:

  Box	Becomes
  describes	the reference's line in subject_definitions — what it is
  retained	the sentence after the marker in retention_analysis — what survives

  Subject slots carry both in the panel; timeline images, reference videos and audio clips
  carry theirs in the properties panel. Either left empty falls back to the generated
  sentence, so the boxes override and never oblige. Frame and storyboard anchors are the
  one exception with no describes box: their declaration states where the image sits in
  the video, which the timeline already knows and should not be contradicted.

• An image only gets a <Picture N> entry when it really is one. The guide: "If an
  image is used only to define a character, scene, costume, or style, do not create a
  standalone picture entry. Instead, cite the image source inside the corresponding
  <Subject N> definition." A timeline image can now be a frame anchor (unchanged),
  a storyboard reference, or subject-defining — the last getting no picture entry
  and, because it is no longer a keyframe, no longer cropped to the output canvas either.

• subject_definitions declares every label; retention_analysis scores every label.
  Both sections were previously incomplete: pictures were never declared, and retention
  was prose. The two sections now follow the guide's shapes, and a subject's
  (appears in [Shot 1], [Shot 3]) is read back off the shot text rather than assumed.

  Both are written one entry per line, the way the guide lays them out. They are lists of
  records rather than prose, and running a dozen of them together into a paragraph made it
  genuinely hard to see where one label ended and the next began.

  0.1.5   subject_definitions: <Subject 1> is the character shown in <Picture 1>.
          retention_analysis: Keep the identity, face and clothing of <Subject 1>
                              consistent across every shot. [Shot 1] begins from <Picture 2>.
  
  0.2.0   subject_definitions:
          <Subject 1> is a woman in a red coat, shown in <Picture 1>.
          <Picture 2> is the first frame of [Shot 1].
  
          retention_analysis:
          <Subject 1> (appears in [Shot 1]): fully_preserved - the collar shape and the
            silver necklace are retained.
          <Picture 2> ([Shot 1] first frame): fully_preserved - the framing and
            composition of <Picture 2> are retained.
  

• New summary section with a derived [task type] prefix — keyframe completion,
  reference generation, audio reuse, audio reference, combined with + in the
  guide's own order. It is derived from what the references are used for, not from what
  is connected: the guide warns that "the mere presence of video or audio does not
  automatically create a corresponding task type", so a reference video supplying only
  camera movement stays reference generation. The gear menu's Task Type field
  overrides it, which is the only way to reach video editing and video continuation —
  neither of which this node can produce on its own.

• Frame anchors are also named inside their shot, as the guide's section 5.3 asks:
  [Shot 1] she enters. The shot begins from <Picture 2>. The phrase goes after the
  shot's own text, because a later shot opens At 00:05.000, and a capitalised clause
  cannot continue out of that comma.

• The reference panel resizes, and the images grow with it. The previews were locked to
  52px inside a fixed 148px slot — too small to judge a reference by, and with no room for
  a second text box. The panel now has the same drag strip the prompt and global-prompt
  panels have, its height is remembered per node, and the previews row is the only part
  that flexes, so every pixel gained goes to the image rather than to the text.

• overall_soundscape, non_diegetic_music and the new summary box sit beside the
  global prompt, not on top of it. The strip was absolutely positioned over the textarea,
  which was shortened by a hard-coded calc() to compensate — so the two fought over the
  same space and the strip sat on the box's own border. As a flex sibling the panel divides
  itself, with no second copy of the height to keep in step.

• Analyze no longer assumes every slot is a character. Its prompt said "describe the
  character's physical appearance", which produced what you would expect when the image was
  a coffee shop. It now asks about the slot's actual kind — a place, a garment, a pose —
  and returns the description and the retained sentence togethe...

0.1.6

• The compiled prompt can be written by hand. EDIT on the prompt panel turns it into
  a textarea, prefilled with what the timeline just compiled; REVERT throws the edit away
  and compiles again. Only the text is replaced — which images, videos and audio clips get
  loaded still comes from the timeline, because the tokenizer emits <Picture i> in the
  order the plan decided and a rewritten sentence cannot renumber that. The edit is stored
  rather than merged: discarding it when a segment moves would lose work without asking,
  and keeping it silently would leave a prompt that no longer matches the screen
  (#4).
• A timeline picture opens its shot rather than ending the video
  (#4). On a hard
  cut the second image starts shot 2; 0.1.5 read the fl2va anchor roles and announced it as
  what the shot "ends on". ends on is now reserved for a segment explicitly flagged as an
  end frame, which is the only place that intention is stated rather than inferred from
  geometry. Also from #4: Keep the identity, face and clothing of … is now appearance,
  since a subject slot holds whatever was put in it and only some of those wear anything.
• The preview's playback rate is a choice, not a decision made for you. true speed
  (the default, unchanged) spreads the sampled frames across the shot's real duration, so
  the preview lasts as long as the finished clip — with latent2rgb that caps at
  preview_fps / 3.35, because there is one image per latent frame and H3 compresses time
  by that much. source fps plays them at the shot's own rate the way ComfyUI's preview
  does: motion reads normally, the clip ends early. Measured on a 124-frame shot: 4.65 fps
  against 24.0.
• Subject slots are a number you choose, 1 to 9, still 3 on a fresh node
  (#8). Three was
  hardcoded in four places and came from nowhere in particular; the model card caps
  reference images at nine and each slot holds two. @char4 and beyond resolve now.
• Images on the ref_images wire can be described. One line of ref_image_notes per
  image, emitted in the guide's own shape (<Picture 3> is a storyboard reference for …).
  Blank lines are kept as placeholders so line 3 always belongs to the third image. Before
  this they were numbered and the prompt said nothing about what was in them
  (#8).
• An audio reference can say whose voice it is
  (#10). Pick a
  subject on the clip and the prompt uses the reference guide's own sentence, <Audio 1> is the voice-timbre reference for <Subject 1> (S1). With more than one subject there was
  previously no way to say which voice belonged to whom. Unassigned clips keep the old
  wording.
• unload_after reaches llama.cpp, not just Ollama
  (#9).
  llama-server in router mode has POST /models/unload; a plain one does not, and the log
  now names --sleep-idle-seconds instead of leaving a checkbox that quietly does nothing.
• A canvas too small to encode is refused by name. Under 32px per side the video latent
  loses an edge and ComfyUI dies six frames deep with a bare float division by zero.
  Reachable from the widgets: custom_width=4 with divisible_by=1 was enough. Not the
  cause of the crash reported in #4, which is still open.
• Fixed: adding a widget shifted every saved value after it, so playback first landed on
  webp_quality and blocked the node with "The value 80 is not available". New widgets go
  last now, a combo whose saved value is not among its options falls back to the default
  with a warning instead of blocking, and both nodes grow to their minimum height on load so
  the panels below a new widget cannot end up hanging outside the node body.
• CONTRIBUTING.md, and test_plan.py is up to 123 offline checks.