@jokerized/decksmith 0.1.4 → 0.3.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (40) hide show
  1. package/README.md +161 -26
  2. package/dist/cli.js +3864 -1816
  3. package/dist/ds-morph.js +1 -0
  4. package/dist/index.js +2701 -780
  5. package/dist/mcp.js +2509 -633
  6. package/dist/types/emit/archetypes/annotated-figure.d.ts +3 -2
  7. package/dist/types/emit/archetypes/data-table.d.ts +0 -19
  8. package/dist/types/emit/archetypes/equation-morph.d.ts +2 -0
  9. package/dist/types/emit/archetypes/equation-walk.d.ts +63 -1
  10. package/dist/types/emit/archetypes/index.d.ts +1 -1
  11. package/dist/types/emit/archetypes/pipeline.d.ts +2 -2
  12. package/dist/types/emit/archetypes/stack.d.ts +20 -1
  13. package/dist/types/emit/archetypes/title.d.ts +31 -14
  14. package/dist/types/emit/camera.d.ts +25 -0
  15. package/dist/types/emit/composition.d.ts +20 -2
  16. package/dist/types/emit/depth.d.ts +95 -0
  17. package/dist/types/emit/kit.d.ts +153 -0
  18. package/dist/types/emit/morph-runtime.d.ts +178 -0
  19. package/dist/types/emit/svg.d.ts +77 -13
  20. package/dist/types/emit/theme.d.ts +1 -0
  21. package/dist/types/images/illustrate.d.ts +34 -0
  22. package/dist/types/images/providers.d.ts +108 -0
  23. package/dist/types/index.d.ts +27 -8
  24. package/dist/types/mcp/prereqs.d.ts +18 -0
  25. package/dist/types/mcp/tools.d.ts +23 -0
  26. package/dist/types/plan/codex.d.ts +25 -3
  27. package/dist/types/plan/duration.d.ts +190 -15
  28. package/dist/types/plan/prompt.d.ts +7 -1
  29. package/dist/types/plan/refs.d.ts +41 -3
  30. package/dist/types/prefs.d.ts +21 -6
  31. package/dist/types/render/capture.d.ts +99 -0
  32. package/dist/types/render/render.d.ts +18 -0
  33. package/dist/types/server/options.d.ts +7 -0
  34. package/dist/types/server/pipeline.d.ts +14 -0
  35. package/dist/types/server/queue.d.ts +1 -1
  36. package/dist/types/types.d.ts +294 -4
  37. package/dist/types/verify/apparent.d.ts +123 -0
  38. package/dist/types/verify/index.d.ts +114 -6
  39. package/dist/types/verify/typefloor.d.ts +3 -1
  40. package/package.json +2 -2
@@ -23,10 +23,27 @@
23
23
  * else, so hitting 60s over 12 slides needs a shorter sentence AND a faster
24
24
  * animation. Neither alone gets there; that is the whole content of this file.
25
25
  *
26
+ * WHICH SLIDE COUNT THE BUDGET IS STRUCK AT. `prefs.slides` is what was ASKED
27
+ * FOR and `beats` is what the planner came back with, and they are routinely not
28
+ * the same number — four of the last five real plans came back short. Every
29
+ * derived number here is `duration / count`, so striking the budget at the
30
+ * request while the deck has fewer beats paces each scene for a slide shorter
31
+ * than the one it is: eight beats against a sixty-second target were paced for
32
+ * five-second beats that are really seven and a half, and the video came in
33
+ * under its target with every gate green and nothing anywhere saying so. So
34
+ * every caller holding a storyboard passes its `beats.length`. The two that
35
+ * cannot are `systemPrompt`, which builds the prompt before a plan exists, and
36
+ * the MCP `estimate`, which is a pre-flight with no plan at all; they get the
37
+ * request, which is the only number they have. The SHORTFALL itself is not this
38
+ * file's to report — `scanBeatCount` in src/verify/index.ts reports it, because
39
+ * a beat count is arithmetic rather than a judgement and a prompt rule alone can
40
+ * always be met cosmetically.
41
+ *
26
42
  * Everything here is pure arithmetic over preferences, so it is testable with no
27
43
  * planner, no TTS and no ffmpeg.
28
44
  */
29
45
  import type { Prefs } from "../prefs.js";
46
+ import type { Source } from "../types.js";
30
47
  /**
31
48
  * Characters of text per second of speech.
32
49
  *
@@ -174,6 +191,17 @@ export declare const FF_BEAT_SECONDS = 8;
174
191
  * for both put 60 seconds at eight slides when the reference is twelve.
175
192
  */
176
193
  export declare const MIN_BEAT_SECONDS = 4;
194
+ /**
195
+ * The longest a slide may sit and still be one thought.
196
+ *
197
+ * Past twenty seconds a beat stops being a slide and becomes a section of a
198
+ * lecture: the demo sits at 20.5s and is already the slowest thing this project
199
+ * ships. It was the bare `20` inside `slidesFor`'s tempo clamp; it is named
200
+ * because that function now uses it twice — once to stop a long target from
201
+ * flattening into a handful of enormous beats, and once as the floor a thin
202
+ * DOCUMENT may shorten the deck to. Both are the same rule.
203
+ */
204
+ export declare const MAX_BEAT_SECONDS = 20;
177
205
  /**
178
206
  * Subtitle rate a short-form viewer will take, as against broadcast television.
179
207
  *
@@ -247,10 +275,27 @@ export declare const SHORT_FORM_CPS = 22;
247
275
  * IT COSTS A STEP, KNOWINGLY. The demo's real p95 cue at `+20%` is 21.321,
248
276
  * inside the ceiling of 22 — but this predicts 22.13 for that step and refuses
249
277
  * it, so the deck takes `+10%` and speaks slower than the artifact proves it
250
- * could. A central estimate of ~1.26 would admit `+20%` and track the artifact
251
- * to within 0.6%; it would also under-predict on about half of future runs. The
252
- * trade was made deliberately in favour of never being under, on a statistic
253
- * measured once over 39 cues.
278
+ * could. It also costs the 60-second case its full sentence: twelve slides at
279
+ * low density plan 66 characters against the 72 of `EXPLAINING_CHARS`, where
280
+ * ~1.26 would plan exactly 72.
281
+ *
282
+ * 1.26 IS NOT AVAILABLE, THOUGH, and it is worth writing down why because the
283
+ * arithmetic above makes it look free. **The floor is 1.2784**, which is this
284
+ * deck's own `DEMO_P95_CUE_CPS / SPEECH_CPS.latin` — 18.409 / 14.4 — and
285
+ * `test/duration.test.ts` fails anything under it, one-sided and on purpose.
286
+ * A constant below the anchor's measured ratio admits a step whose captions are
287
+ * over the ceiling, which is exactly the 1.17 bug this replaced.
288
+ *
289
+ * The five ratios in the table above centre on 1.2604 and tempt you under that
290
+ * floor. They are a DIFFERENT synthesis of the same text: edge-tts does not
291
+ * repeat itself, and the `+0%` re-run came back at 1.2467 where the artifact in
292
+ * `demo/audio` sits at 1.2784. The shipped artifact wins — it is the deck that
293
+ * actually exists. So the honest range is [1.2784, 1.2945], the second number
294
+ * being the highest ratio the re-run saw, and 1.30 clears both.
295
+ *
296
+ * The 60-second sentence is therefore not bought back by tuning this. It is
297
+ * bought with seconds or with slides, which is what the rest of this file has
298
+ * said all along, and 90s plans 102.
254
299
  *
255
300
  * `test/duration.test.ts` pins the `+0%` end against the artifact.
256
301
  */
@@ -344,7 +389,16 @@ export interface DurationPlan {
344
389
  chars?: number;
345
390
  /** Seconds of speech one beat can afford. Absent without a target. */
346
391
  speechSeconds?: number;
347
- /** The per-beat length the target implies. Absent without a target. */
392
+ /**
393
+ * The per-beat length the target implies, at the count the plan was struck at.
394
+ * Absent without a target.
395
+ *
396
+ * A BUDGET, NOT THE ANSWER, and the name collides with the thing that IS the
397
+ * answer. `beatSeconds` in src/emit/composition.ts — `max(authored·speed,
398
+ * lastHold + SETTLE, speechEnd + SETTLE)` — is what actually sizes a scene,
399
+ * and it never sees this number. `speed` is the only value this file sends
400
+ * into the emitted timeline; everything else here is advice to the planner.
401
+ */
348
402
  beatSeconds?: number;
349
403
  /** Said, never fatal: what the target costs, or cannot buy. */
350
404
  warnings: string[];
@@ -359,8 +413,14 @@ export interface DurationPlan {
359
413
  * `duration` set OVERRIDES `animationSpeed`: the target owns the pace, because
360
414
  * the two cannot both be honoured and the target is the one the user stated a
361
415
  * number for. Said in a warning rather than silently.
416
+ *
417
+ * `beats` is the count the budget is struck at, and it defaults to the count
418
+ * that was asked for. Pass the storyboard's own `beats.length` wherever one is
419
+ * in hand — see the header for why, and for the two callers that structurally
420
+ * cannot. The default is not a convenience: it is the right answer for those
421
+ * two, and it is what keeps a run where the planner hit its number identical.
362
422
  */
363
- export declare function durationPlan(prefs: Prefs): DurationPlan;
423
+ export declare function durationPlan(prefs: Prefs, beats?: number): DurationPlan;
364
424
  /**
365
425
  * How many slides a target should have, by TEMPO.
366
426
  *
@@ -387,11 +447,90 @@ export declare function durationPlan(prefs: Prefs): DurationPlan;
387
447
  * 30s -> 8 120s -> 12 600s -> 30
388
448
  * 60s -> 12 300s -> 15
389
449
  *
450
+ * AND HOW MUCH THE DOCUMENT HAS TO SAY, which is the half that was missing. The
451
+ * table above is flat at twelve from 48s to 240s no matter what it is pointed
452
+ * at: a four-page workshop note and a forty-page survey both got twelve beats,
453
+ * because the only input was a clock. Tempo says how long a beat may LAST; it
454
+ * cannot say how many points exist to spend beats on. So `source`, when the
455
+ * caller has one, scales the tempo count by what the document actually contains
456
+ * (`sourcePoints`), bounded by `SUPPLY_RANGE` because the measure is coarse and
457
+ * its authority should be too.
458
+ *
459
+ * 60s + a thin note -> 6 240s + a thin note -> 12
460
+ * 60s + a full paper -> 15 240s + a full paper -> 17
461
+ *
462
+ * The clock still wins at both ends. Sixty seconds cannot hold seventeen beats
463
+ * at `MIN_BEAT_SECONDS` apiece, so a rich document against a short target spends
464
+ * its extra points on nothing; and past four minutes `MAX_BEAT_SECONDS` holds
465
+ * the count up, so a thin note asked for a long video does not become six
466
+ * forty-second slides. Only where the clock leaves room does the document move
467
+ * the number — which is most of the range people actually ask for.
468
+ *
469
+ * WHO CAN PASS A SOURCE. The CLI's `plan` verb reads `source.json` before it
470
+ * resolves preferences, so it can and does. `parseOptions` (src/server/options.ts)
471
+ * derives the count while the upload is still a form, before anything is
472
+ * ingested, and the MCP `estimate` is a pre-flight with no document at all;
473
+ * both get the tempo number, which is what they got before. Omitting the
474
+ * argument is therefore not a degraded path, it is the old behaviour, byte for
475
+ * byte.
476
+ *
390
477
  * This is the DEFAULT, never an override. `slides` is one of the three knobs the
391
478
  * owner asked to hold — "user can give you the number of slides they want in the
392
- * video with duration of their choice" — so an explicit count is obeyed.
479
+ * video with duration of their choice" — so an explicit count is obeyed, which
480
+ * is `loadPrefs`'s business: it calls this only when nobody named a number.
481
+ */
482
+ export declare function slidesFor(prefs: Prefs, source?: Source): number;
483
+ /**
484
+ * Characters of section prose one beat is worth.
485
+ *
486
+ * A GUESS, and marked one the way `SPEECH_CPS.cjk` is: no full-length document
487
+ * is stored in this repository to measure against — `demo/source.json` is a
488
+ * 231-character stub whose twelve beats were written by hand, and the fixture
489
+ * papers are 1.9 KB each. What matters is this number's RATIO to
490
+ * `REFERENCE_POINTS`, and the two are set together so an eight-section
491
+ * conference paper — around 30k characters, five figures, two tables, four
492
+ * equations — comes out above the reference rather than at it. Replace both with
493
+ * a measurement the first time a corpus of real sources exists.
494
+ */
495
+ export declare const PROSE_CHARS_PER_BEAT = 1800;
496
+ /**
497
+ * The supply a document needs to earn the tempo's own beat count.
498
+ *
499
+ * Twenty points is roughly a solid conference paper minus its trimmings. Below
500
+ * it the deck shortens, above it the deck lengthens, and `SUPPLY_RANGE` decides
501
+ * how far either can go.
502
+ */
503
+ export declare const REFERENCE_POINTS = 20;
504
+ /**
505
+ * How far the document may move the tempo's count, as a multiplier.
506
+ *
507
+ * Bounded rather than open because `sourcePoints` is a proxy and a proxy should
508
+ * not be trusted past the range where it is obviously right. Half is the
509
+ * shortest a deck of a real document should get before the answer is "this is
510
+ * not enough source"; one and a half keeps a rich paper under the twenty-second
511
+ * beat that makes a deck a lecture. A survey with a hundred points is not worth
512
+ * a hundred beats, and the clamp is what says so.
513
+ */
514
+ export declare const SUPPLY_RANGE: {
515
+ readonly min: 0.5;
516
+ readonly max: 1.5;
517
+ };
518
+ /**
519
+ * How many distinct points a document offers a deck.
520
+ *
521
+ * Two supplies, added, because they are genuinely different things to make a
522
+ * beat out of: PROSE, which carries the argument, and EXHIBITS — the figures,
523
+ * tables and equations the authors made on purpose because a sentence was not
524
+ * enough. RULE 2 in the prompt tells the planner every figure should earn a
525
+ * beat, so a document with eight of them is asking for a longer deck than one
526
+ * with none, and this is where that is counted rather than hoped for.
527
+ *
528
+ * Headings are deliberately NOT counted. A section boundary is a formatting
529
+ * decision — one author writes six headings over 20k characters and another
530
+ * writes twenty-four over the same prose — so counting them would measure the
531
+ * template rather than the document.
393
532
  */
394
- export declare function slidesFor(prefs: Prefs): number;
533
+ export declare function sourcePoints(source: Source): number;
395
534
  /**
396
535
  * How fast this deck speaks.
397
536
  *
@@ -442,12 +581,48 @@ export declare function p95CueRate(cues: readonly {
442
581
  text: string;
443
582
  }[]): number;
444
583
  /**
445
- * Whether this deck's own captions survive being sped up this much.
446
- *
447
- * Computed from the deck's MEASURED p95 cue rate against broadcast practice, not
448
- * from a hardcoded ceiling: a deck whose captions already run at 18 cps has no
449
- * headroom at all, and one written in short lines has plenty. Warns, never
450
- * refuses how readable a caption has to be is the user's call, and they can
451
- * only make it if they are told the number.
584
+ * What is wrong with speeding this deck up this much — everything that is.
585
+ *
586
+ * TWO CEILINGS, AND THEY ARE NOT THE SAME KIND OF THING. The caption one is
587
+ * computed from the deck's MEASURED p95 cue rate against broadcast practice
588
+ * rather than from a constant: a deck whose captions already run at 18 cps has
589
+ * no headroom at all, one written in short lines has plenty. That one stays
590
+ * pure advice how readable a caption has to be is the user's call, and they
591
+ * can only make it if they are told the number. `MAX_PLAYBACK` is the other,
592
+ * and `playbackRefusal` below turns it into a refusal, because time-stretched
593
+ * audio is not a matter of taste.
594
+ *
595
+ * BOTH CLAUSES, NOT THE FIRST ONE. This used to `return` out of the caption
596
+ * branch, which meant a dense deck — every deck this tool plans is dense, by
597
+ * construction — could only ever be told about the ceiling that is advisory,
598
+ * and never about the one that is enforced. The measured shape: 2.14× on a
599
+ * 19.6 cps deck printed "readable captions cap out near 0.87×" and suppressed
600
+ * "past the 1.25× that reads comfortably", so the only line naming an
601
+ * enforceable limit was the one the user did not get.
452
602
  */
453
603
  export declare function playbackWarning(factor: number, p95: number): string | undefined;
604
+ /**
605
+ * Why this target cannot be reached from this deck, when it cannot.
606
+ *
607
+ * `MAX_PLAYBACK` IS THE ONE THAT CAN BE ENFORCED. The caption ceiling cannot
608
+ * be, and the arithmetic is not close: `COMFORTABLE_CPS / DEMO_P95_CUE_CPS` is
609
+ * 17/18.409 = 0.92 on this project's own anchor deck, so a caption-derived
610
+ * limit is already breached at 1× and would refuse every speed-up there is —
611
+ * `--duration` would be dead on arrival. Every deck DeckSmith plans is denser
612
+ * than 17 cps by construction; see `CUE_OVERHEAD`. So the captions warn and
613
+ * the playback factor refuses, and `test/duration.test.ts` pins the reason.
614
+ *
615
+ * REFUSING IS NOT NEW POLICY, it is the policy the planner already keeps.
616
+ * `durationPlan` sizes a deck so the residual gap fits inside `MAX_PLAYBACK`,
617
+ * and `test/duration.test.ts` asserts exactly that at 60s, 120s, 180s and
618
+ * 240s. A 2.14× request is not a tight fit that slipped — it is 71% past a
619
+ * bound the rest of the file treats as arithmetic. What was shipping instead
620
+ * was a video nobody could follow, announced by a mid-render log line that 40
621
+ * lines of capture progress scrolled away.
622
+ *
623
+ * Both remedies are named because the honest one is not the flag: the length
624
+ * of a deck is decided at plan time by how much it says, so the fix is fewer
625
+ * words or more seconds, and `--allow-fast-playback` is for the person who has
626
+ * looked at the alternative and wants the fast video anyway.
627
+ */
628
+ export declare function playbackRefusal(actualSeconds: number, targetSeconds: number, p95: number): string | undefined;
@@ -43,5 +43,11 @@ export declare function systemPrompt(prefs: Prefs): string;
43
43
  * planner's other entry points keep working unchanged.
44
44
  */
45
45
  export declare const SYSTEM: string;
46
- /** The document plus every id the model is allowed to cite, and nothing else. */
46
+ /**
47
+ * The document plus every id the model is allowed to cite, and nothing else.
48
+ *
49
+ * The figure block carries two facts beyond the caption — the section the image
50
+ * sat under and the prose that refers to it — because a planner that cannot see
51
+ * the picture has nothing else to decide what the picture is FOR.
52
+ */
47
53
  export declare function renderSource(source: Source): string;
@@ -1,7 +1,39 @@
1
1
  import type { Source, Storyboard } from "../types.js";
2
- export declare function assertRefsResolve(storyboard: Storyboard, source: Source): void;
2
+ /** A slot that asks for a picture nobody has drawn yet: a brief with no `figureId` beside it. */
3
+ export interface PendingIllustration {
4
+ beatId: string;
5
+ /** The field the figure id will land in, e.g. `params.left.figureId`. */
6
+ where: string;
7
+ }
3
8
  /**
4
- * Every `inside` names a part the previous beat actually draws.
9
+ * Every pending slot, in beat order. A claim-figure without a `figureId` is
10
+ * pending by construction — the schema insists on one of the two — and a
11
+ * split-compare side is pending only when it carries a brief, because a side
12
+ * with neither is a list.
13
+ */
14
+ export declare function pendingIllustrations(storyboard: Storyboard): PendingIllustration[];
15
+ /**
16
+ * Whether this deck was illustrated — the fact a pack records as
17
+ * `images.enabled`, the way `narration.enabled` records whether it was spoken.
18
+ * A brief with a figure the source really has is the trace `illustrate` leaves;
19
+ * a brief alone is a picture still owed, which `assertRefsResolve` refuses
20
+ * before anyone asks this; a figure alone is an ordinary figure.
21
+ */
22
+ export declare function hasIllustrations(storyboard: Storyboard, source: Source): boolean;
23
+ export interface RefsOptions {
24
+ /**
25
+ * What a pending illustration means here. `refuse`, the default, is right for
26
+ * every reader that needs the figure — `build` would otherwise emit a slide
27
+ * around a picture that does not exist. Only the planner and the `plan` verb
28
+ * say `allow`, and only when images are on: a brief is what they were asked
29
+ * to produce, and `illustrate` is the step that makes it resolve.
30
+ */
31
+ pending?: "allow" | "refuse";
32
+ }
33
+ export declare function assertRefsResolve(storyboard: Storyboard, source: Source, opts?: RefsOptions): void;
34
+ /**
35
+ * Every `inside` names a part the previous beat actually draws — and, where the
36
+ * plan said which part it meant, the RIGHT one.
5
37
  *
6
38
  * THE SAME SHAPE AS THE CHECK ABOVE, one level in: a schema proves `inside` has
7
39
  * a beat and an element, and cannot prove the element exists. RULE 11 says only a
@@ -9,6 +41,11 @@ export declare function assertRefsResolve(storyboard: Storyboard, source: Source
9
41
  * entering, and a real plan asked to fly into `stage1` of an ANNOTATED-FIGURE —
10
42
  * which has notes and leader lines and no stages at all.
11
43
  *
44
+ * AND ONE LEVEL IN AGAIN, because existing is not enough. `element` is an index,
45
+ * so a plan that means the second stage and writes the third names a part that
46
+ * does exist; `partLabelProblem` compares `inside.label` — what the plan says is
47
+ * there — against what the archetype reports drawing.
48
+ *
12
49
  * WHY IT MOVED HERE. The emitter already refuses this, so nothing shipped
13
50
  * broken. But it refuses at BUILD, which is after `narrate` has spent a minute
14
51
  * and a dozen network round trips synthesising speech for a storyboard that was
@@ -17,6 +54,7 @@ export declare function assertRefsResolve(storyboard: Storyboard, source: Source
17
54
  *
18
55
  * Cheap enough to be exact rather than a table: the previous beat is emitted and
19
56
  * `enterableIds` is asked what it drew, so this cannot drift from the emitter the
20
- * way a hardcoded list of archetype interiors would.
57
+ * way a hardcoded list of archetype interiors would. The same `Scene` carries
58
+ * `parts`, so the label comparison rides along for nothing.
21
59
  */
22
60
  export declare function assertInsideResolves(storyboard: Storyboard, source: Source): void;
@@ -1,13 +1,15 @@
1
1
  import type { z } from "zod";
2
- import { prefsSchema } from "./types.js";
2
+ import { prefsSchema, type Source } from "./types.js";
3
3
  export type Prefs = z.infer<typeof prefsSchema>;
4
4
  /**
5
- * A partial Prefs. `narration` is the one nested object, so it is the one place
6
- * a patch merges rather than replaces: `--voice` must not wipe a `rate` the
7
- * config file set.
5
+ * A partial Prefs. `narration` and `images` are the two nested objects, so they
6
+ * are the two places a patch merges rather than replaces: `--voice` must not
7
+ * wipe a `rate` the config file set, and `--image-style` must not wipe its
8
+ * `provider`.
8
9
  */
9
- export type PrefsPatch = Partial<Omit<Prefs, "narration">> & {
10
+ export type PrefsPatch = Partial<Omit<Prefs, "narration" | "images">> & {
10
11
  narration?: Partial<Prefs["narration"]>;
12
+ images?: Partial<Prefs["images"]>;
11
13
  };
12
14
  export declare const CONFIG_FILE = "decksmith.config.json";
13
15
  /**
@@ -16,8 +18,15 @@ export declare const CONFIG_FILE = "decksmith.config.json";
16
18
  * `cwd` is where the search for a config file starts; the search stops at the
17
19
  * filesystem root. Returns a fully-populated Prefs — every consumer downstream
18
20
  * reads fields, never optionals.
21
+ *
22
+ * `source` is the document the deck is being planned FROM, when the caller has
23
+ * already read one. It is optional because most callers have not: `build`,
24
+ * `narrate` and `pack` work from a storyboard that was planned long ago, and
25
+ * `illustrate` never needed a beat count. Only the planning path passes it, and
26
+ * only so `slidesFor` can size the deck to what the document contains rather
27
+ * than to a clock — see its header for what changes when it is absent.
19
28
  */
20
- export declare function loadPrefs(overrides?: PrefsPatch, cwd?: string): Promise<Prefs>;
29
+ export declare function loadPrefs(overrides?: PrefsPatch, cwd?: string, source?: Source): Promise<Prefs>;
21
30
  /**
22
31
  * Turn CLI flags into a patch. Values arrive as strings from commander and are
23
32
  * not validated here beyond the numbers: `loadPrefs` runs them through the
@@ -40,4 +49,10 @@ export interface PrefFlags {
40
49
  subtitles?: boolean;
41
50
  /** `--narration-density`. Spelled apart from `density`, which is a slide's. */
42
51
  narrationDensity?: string;
52
+ /** `--images`. A boolean like `narrate`, so `flags()` in cli.ts reads it beside `--no-subtitles`. */
53
+ images?: boolean;
54
+ imageProvider?: string;
55
+ imageModel?: string;
56
+ imageStyle?: string;
57
+ imageMax?: string | number;
43
58
  }
@@ -0,0 +1,99 @@
1
+ /**
2
+ * The capture path, as a thing other code can drive.
3
+ *
4
+ * "Look at the artifact" is this project's strongest gate — six of its bugs
5
+ * shipped past green gates and every one was caught by a human looking at a
6
+ * frame. So the frame a human looks at has to be the frame the renderer would
7
+ * produce, and `hyperframes snapshot` is NOT that frame: it calls
8
+ * `player.renderSeek(t)` with no options, so `suppressEvents` is falsy and a
9
+ * GSAP `onUpdate` FIRES. Invariant 11 says the capture path suppresses events,
10
+ * so callback-driven motion plays under snapshot and renders frozen. Snapshot is
11
+ * permissive in exactly the direction that hides the project's most dangerous
12
+ * failure shape, and it was measured lying about CSS 3D as well — at t=3.9s it
13
+ * produced a flat, un-rotated frame where the render produced correct
14
+ * perspective (see `.planning/2026-09-04-css-3d-spike.md`).
15
+ *
16
+ * `fidelity` already drove the right path privately, and was measured doing so:
17
+ * against `experiments/015-decision/out/vocab-18.mp4`, a real render, at all
18
+ * twelve of its holds, it agreed to a worst case of 0.11 and a mean of 0.03
19
+ * percentage points — the residual being H.264 quantisation. This module is that
20
+ * code lifted out from under the ink arithmetic so the gate and a person asking
21
+ * for a PNG go through the same three calls: inject the pinned runtime,
22
+ * `renderSeek(t, { suppressEvents: true })`, then `Page.captureScreenshot` with
23
+ * the renderer's own clip.
24
+ */
25
+ /**
26
+ * The browser `render` already uses.
27
+ *
28
+ * `@puppeteer/browsers` and `puppeteer-core` are DIRECT dependencies of this
29
+ * package. They arrive with hyperframes too — a hard dependency of any verb that
30
+ * opens a browser — but relying on that made module imports resolve through
31
+ * somebody else's dependency tree, and the day hyperframes swaps its automation
32
+ * library those verbs break for a reason nothing in this repo mentions. They are
33
+ * still imported DYNAMICALLY, which is a different concern: a machine that
34
+ * cannot supply a browser must degrade rather than throw after the capture has
35
+ * already run.
36
+ *
37
+ * `need` completes the error message, because the three callers want different
38
+ * advice from it — captions can fall back to a sidecar, `frames` cannot. This
39
+ * was two copies until `frames` became the third caller the older of them said
40
+ * to promote it on.
41
+ */
42
+ export declare function chromePath(need?: string): Promise<string>;
43
+ type Browser = Awaited<ReturnType<typeof import("puppeteer-core").launch>>;
44
+ type Page = Awaited<ReturnType<Browser["newPage"]>>;
45
+ /** An open deck, seekable and photographable. Close it. */
46
+ export interface DeckPage {
47
+ /** The canvas the deck declares, which is also the screenshot's clip. */
48
+ width: number;
49
+ height: number;
50
+ /** Commit the frame at an absolute time, events suppressed. */
51
+ seek(t: number): Promise<void>;
52
+ /** PNG bytes of the frame currently committed. */
53
+ shoot(): Promise<Buffer>;
54
+ /** For callers that need to read the DOM at a stop, as `fidelity` does. */
55
+ page: Page;
56
+ close(): Promise<void>;
57
+ }
58
+ export interface OpenOptions {
59
+ timeoutMs?: number;
60
+ }
61
+ /**
62
+ * Open `dir`'s deck in the renderer's own browser, ready to seek.
63
+ *
64
+ * Throws for every environmental reason — no Chrome, no `index.html`, no canvas
65
+ * size. Callers that must not fail a build over the environment (the `fidelity`
66
+ * gate) catch it; callers a person invoked directly (`frames`) let it surface,
67
+ * because a person who asked for a PNG and got nothing needs to hear why.
68
+ */
69
+ export declare function openDeck(dir: string, opts?: OpenOptions): Promise<DeckPage>;
70
+ /** One written frame. */
71
+ export interface CapturedFrame {
72
+ t: number;
73
+ path: string;
74
+ bytes: number;
75
+ }
76
+ /**
77
+ * What one frame is called.
78
+ *
79
+ * The index leads, zero-padded to the width of the largest, because a directory
80
+ * listing is how these actually get looked at and lexical order is what a file
81
+ * browser gives you: `t10.000s` sorts before `t3.900s`, so the reader opens the
82
+ * wrong frame and believes it. Padding the INDEX rather than the time keeps the
83
+ * two orders — asked-for and on-disk — identical without deciding how many
84
+ * digits a deck's duration might need.
85
+ *
86
+ * The time is still in the name, to 3 decimals, because that is the thing a
87
+ * reader is checking against and invariant 10 says 3 decimals is what the
88
+ * renderer rounds to. A frame whose filename disagreed with the timeline would
89
+ * be worse than one with no time in it at all.
90
+ */
91
+ export declare function frameName(index: number, t: number, total: number): string;
92
+ /**
93
+ * Write a PNG per requested time into `outDir`.
94
+ *
95
+ * Times are rounded to 3 decimals to match invariant 10, so a filename names
96
+ * exactly the instant the renderer would.
97
+ */
98
+ export declare function captureFrames(dir: string, times: readonly number[], outDir: string, opts?: OpenOptions): Promise<CapturedFrame[]>;
99
+ export {};
@@ -103,8 +103,20 @@ export interface RenderOptions {
103
103
  * length is decided at plan time by how much it says, and this closes whatever
104
104
  * gap survived that. A video already inside the target is left alone rather
105
105
  * than padded — dead air is worse than eight seconds short.
106
+ *
107
+ * A gap wider than `MAX_PLAYBACK` is refused, not closed; see the check after
108
+ * `readTiming` and `playbackRefusal`.
106
109
  */
107
110
  targetSeconds?: number;
111
+ /**
112
+ * Speed past `MAX_PLAYBACK` anyway.
113
+ *
114
+ * For the person who has read the refusal, cannot re-plan, and wants the fast
115
+ * video regardless. It does not silence anything: both clauses of
116
+ * `playbackWarning` still print, and the summary still carries the factor and
117
+ * the caption rate it bought.
118
+ */
119
+ allowFastPlayback?: boolean;
108
120
  /** Leave the per-piece intermediates on disk. */
109
121
  keep?: boolean;
110
122
  /** Progress, one line at a time. */
@@ -119,5 +131,11 @@ export interface RenderResult {
119
131
  burned: boolean;
120
132
  /** What `targetSeconds` cost, if anything. 1 means the file was not respeeded. */
121
133
  playback: number;
134
+ /**
135
+ * p95 characters per second of the subtitles AS SHIPPED — measured off the
136
+ * cues that were written, so at `playback > 1` it is already the sped-up rate
137
+ * a viewer actually reads at rather than the deck's rate at rest.
138
+ */
139
+ captionCps: number;
122
140
  }
123
141
  export declare function render(opts: RenderOptions): Promise<RenderResult>;
@@ -61,6 +61,13 @@ export interface JobOptions {
61
61
  prefs: Prefs;
62
62
  narrate: boolean;
63
63
  video: boolean;
64
+ /**
65
+ * Run the `illustrate` stage between `plan` and `build`. The same bit lands in
66
+ * `prefs.images.enabled`, which is what lets the planner write a brief; this
67
+ * copy is what `stagesFor` reads, the way `narrate` is read beside
68
+ * `prefs.narration.enabled`.
69
+ */
70
+ images: boolean;
64
71
  /**
65
72
  * Whether the REQUEST said so, as opposed to the schema defaulting.
66
73
  *
@@ -1,3 +1,4 @@
1
+ import { type ImageProvider, type Runner } from "../index.js";
1
2
  import type { JobOptions } from "./options.js";
2
3
  import type { JobHandle, JobResult, Stage } from "./queue.js";
3
4
  import { type Upload } from "./upload.js";
@@ -15,6 +16,19 @@ export interface PipelineInput {
15
16
  * owner's own box, own papers), it is fetched with a count and a timeout.
16
17
  */
17
18
  fetchRemoteFigures: boolean;
19
+ /**
20
+ * The rungs `illustrate` draws through. A test injects the tool's own SVG and
21
+ * nothing else; absent, the stage resolves its providers from the environment
22
+ * at call time, like `narrate` does — so there is no server option to plumb
23
+ * and nothing to configure but the env vars.
24
+ */
25
+ imageChain?: ImageProvider[];
26
+ /**
27
+ * The planner's `Runner`, handed straight to `codexPlanner`. The seam
28
+ * test/plan.test.ts already drives the parse path through; here it is what
29
+ * lets a test carry a job PAST `plan` — into `illustrate` — without a Codex.
30
+ */
31
+ run?: Runner;
18
32
  }
19
33
  /** Which rows the step list should have, decided before anything runs. */
20
34
  export declare function stagesFor(options: JobOptions): Stage[];
@@ -15,7 +15,7 @@
15
15
  */
16
16
  import { type JobError } from "./errors.js";
17
17
  export type JobState = "queued" | "running" | "done" | "error";
18
- export type Stage = "ingest" | "plan" | "narrate" | "build" | "render";
18
+ export type Stage = "ingest" | "plan" | "illustrate" | "narrate" | "build" | "render";
19
19
  export type StepState = "pending" | "running" | "done" | "skipped" | "error";
20
20
  export interface StepView {
21
21
  name: Stage;