@wji.ai/cli 0.7.1 → 0.8.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@wji.ai/cli",
-  "version": "0.7.1",
+  "version": "0.8.0",
   "description": "Command line interface for the WJI AI platform",
   "license": "UNLICENSED",
   "author": "WJI AI",

package/skill/short-drama/SKILL.md

@@ -211,7 +211,7 @@ Read `{baseDir}/reference/stage-characters.md`.
    sheet prompt prefixed with `@image1 is the face to match`), and a **voice** audio node
    (`seed-audio-1.0`, `--from` the headshot — the model reads the face and invents a voice for it;
    the composed prompt asks for a dry voice, because the same model also makes rooms, scores and
-   sound effects — that is what its stage-5 soundtrack nodes are for;
+   sound effects — that is what the episode score over the edit is for;
    `prompt voice C01` composes the text: a parenthetical carrying age, identity, personality,
    timbre, pitch, pace, accent and emotion, then the acted line).
    Recipes in `reference/canvas-recipes.md`.
@@ -345,15 +345,27 @@ must be connected: pictures, then clips, then voices.
    segment's keyframes only, inspect, then continue. Keyframes for one segment are a batch (quote
    the total). Do not switch a project to keyframes on your own — it doubles the images to pay for
    and approve.
-5. **Soundtrack node** per segment (`params.audioTrack`, default `bed`): `prompt audio E0N-01`
-   derives a Seed Audio prompt from `segment.sound` — timecoded `[start s:end s]` ambience, sfx on
-   their cuts, music that enters / swells / ducks / falls silent at the cues you wrote on
-   `sound.music.cues` — and `canvas add audio -m seed-audio-1.0 --prompt-file …` makes it (~0.15
-   credits a minute, no `--from`; `--full` adds the lines in the connected `@Audio1..3` voices).
-   Seed Audio is the voice-sample model, but it is an audio-scene generator: one prompt is a whole
-   track. Attach the asset to the shot dataset; bind `wji.nodes.track` / `wji.assets.track`.
-6. After each clip: `fs set <shot dataset> --add-images <track> [<keyframes…>] --set primaryVideo=<clip>`;
-   bind asset ids. `manifest --ep N` must come back clean before you call the episode done.
+5. **No BGM in the clips** (`params.audioTrack: post`, the default). Each segment is a separate
+   generation, so music or sound design made inside a clip stops at its last frame and cannot carry
+   across a cut. The video prompt keeps the lines, the ambience and the sfx the world makes, says
+   there is no music, and leaves out every sfx marked `post: true` (stingers, risers, whooshes,
+   impact hits). Still write `sound.music` (intent + cues) and the `post: true` hits — they are the
+   spotting notes for the score.
+6. After each clip: `fs set <shot dataset> [--add-images <keyframes…>] --set primaryVideo=<clip>`;
+   bind asset ids.
+7. **Score the edited episode.** When every clip is in, edit them into a locked cut (picture lock),
+   `wji asset upload` it and bind `wji.assets.edit`. `prompt score E0N [--fps 24]` prints the planned
+   spotting sheet — SMPTE rows for the scene beds, every line (where the music ducks), each
+   segment's music cues at episode time, the `post` sound design, the sfx already in the clips, the
+   hook / turns / 爽点 / cliff (起承转合) and the ≤120 s Seed Audio sections. Conform it to the edit's
+   real cuts and write the cue sheet and the section prompts (`reference/mix-over-video.md`; craft in
+   `reference/bgm-sfx.md`, loudness and terms in `reference/film-mixing-guideline.md`): BGM, BGE /
+   ambience and sound design only — never what the clips already carry. One `canvas add audio -n
+   "E0N score K" -m seed-audio-1.0 --prompt-file …` per section (~0.15 credits a minute — quote
+   first), mix the sections under the edit at their offsets, upload and bind `wji.assets.mix`.
+   `manifest --ep N` must come back clean before you call the episode done. (`audioTrack: bed` /
+   `full` instead make a Seed Audio node per segment from `prompt audio E0N-01` and put the music in
+   the clip.)
 
 - **A failed generation is not always a bad prompt.** `wan-3.0-ref` rejections from DashScope's
   content filter (`DataInspectionFailed`) charge nothing and are partly random — retry the same node

package/skill/short-drama/reference/bgm-sfx.md

@@ -0,0 +1,124 @@
+# BGM timed to video + short-drama SFX
+
+## Segment `sound` SoT
+
+```json
+"sound": {
+  "ambience": "neon KTV booth hum, leather sofa creak",
+  "sfx": [
+    { "cut": 2, "event": "wine glass tip and splash on table" },
+    { "cut": 7, "event": "SPECIAL: cliff freeze sting — single high piano hit then silence", "post": true }
+  ],
+  "music": {
+    "intent": "romantic neon slow pulse, soft keys, restrained satin bass",
+    "cues": [
+      { "at": 1, "do": "enter", "note": "under dialogue" },
+      { "at": "10.5s", "do": "swell", "note": "rise into kiss handoff" },
+      { "at": 4, "do": "duck", "note": "under roast lines" },
+      { "at": 6, "do": "drop", "note": "after climax into plea" },
+      { "at": 7, "do": "silence", "note": "before P02 reveal" },
+      { "at": "end", "do": "out" }
+    ]
+  }
+}
+```
+
+Where it goes depends on `params.audioTrack`:
+
+- **`post` (default)** — the clip gets only the ambience, the in-picture sfx and the lines; `music`
+  and every `post: true` sfx stay out of it. They are the spotting notes for the episode score:
+  `$SD prompt <project> score E01` lays them on the episode timeline (SMPTE) for scoring over the
+  edit (`mix-over-video.md`). A cue `at` is still a cut of *this* segment — the sheet adds the
+  segment's offset.
+- **`bed` / `full`** — `prompt audio … --bed` / `--full` expands this into `[start s:end s]` lines
+  for a Seed Audio node per segment, and the music intent also goes into the clip prompt.
+
+## BGM cue verbs
+
+`enter | swell | duck | drop | silence | out`
+
+Put mood vocabulary in `intent` + `note`:
+
+| Want | How to write |
+| --- | --- |
+| rise / build | `swell` + note "intensity climb / rise into …" |
+| drone / dread | intent "low dread drone, barely moving" + `enter`/`drop` |
+| romantic | intent "romantic …" + `swell` on romantic beats |
+| comedy | intent "light comedy pizzicato" + duck under punchlines |
+| cliff | `drop` or `silence` then one special sting SFX |
+| triumphant | intent "bright brass sting" — use sparingly |
+
+Rules: one music intent per segment (or continuous across a scene); instrumental unless vocals are a story beat; **duck under dialogue**; hit 爽点 / cliff on purpose. Continuous texture belongs in **ambience**, not twenty music mood words.
+
+### With a `params.cinematography` preset (sound is half the look)
+
+Craft in `cinematography-<preset>.md`; not gated.
+
+| preset | music / sfx |
+| --- | --- |
+| `low-key` | nothing extra — the usual dread / tension vocabulary |
+| `sweet-romance` | the OST is the segment's `sound.music`, cues on the cut: `enter` on the eye-lock's last cut, `swell` on the look back / the carry / the gap closing, `silence` as the hand lands or just before a confession line (so it lands in a hush), `drop` on the 壁咚; rain on the umbrella, a heartbeat if wanted |
+| `period-nostalgia` | diegetic music of the era — the radio, a period song on the cassette player — in `sound.sfx` or as the segment's `music` intent; cicadas in the courtyard, the bicycle bell, a street hawker, a distant train |
+| `costume-idol` | follows `sweet-romance` for the 心动 set pieces (执手, 簪花, 共乘一马, 月下对饮, 灯会) — `enter` / `swell` / `silence`; wind, chimes, the lantern festival crowd as ambience |
+| `costume-dark` | fire crackle, wind, rain on tiles, armour creaking, distant drums; **`silence`** in the court before a verdict; violence carried by sound and consequence rather than choreography |
+| `scifi-hard` | hums, fans, relays, alarms, radio chatter with squelch, breathing in a helmet; **silence in vacuum** (only the suit's breathing and radio); music sparse — a low drone, a single piano line |
+| `scifi-soft` | rain, crowds in many languages, synth hums, holographic chimes, drones passing; music synthwave / ambient pads |
+
+### By segment grammar (`segment.grammar`, else `params.grammar`)
+
+Craft in `grammar-faceslap.md` §3, `grammar-comedy.md` and
+`action.md` / `action-<style>.md` ("Sound"); gate `grammar-advice` checks the 打脸 music hit.
+
+| grammar | music / sfx |
+| --- | --- |
+| `faceslap` | `silence` or `drop` under the insult, **`enter` / `swell` on the entrance or the proof**, full through the reaction cascade; sfx: the document slapped on the table, the gasp, the cup falling, the murmur rising |
+| `comedy` | **silence before the punchline** (a `silence` cue or a pause in the ambience); `slapstick` — cartoon sfx on every beat (boing, whoosh, record scratch) and a comic sting after the punchline; `deadpan` — no comic sfx, often no music, the ambience (clock, fly, fridge hum) fills the hold; `romcom` — playful pops and plinks, a heartbeat, a cute sting, the OST entering on the blush |
+| `action` | **every impact gets an sfx on its cut**; the music `drop`s out before the decisive blow; per style: `bullet-time` stretched / pitched-down sound then snap back, `gun-fu` driving low synth, `hk-kungfu` exaggerated whooshes and snaps with a comic sting, `historical-realist` metal on metal and often no score, `street-brawl` none or a low drone, `car-chase` engines then the silence after the crash |
+
+## SFX + special short-drama FX
+
+Every story-visible hit needs `sfx[]` tied to a **cut** (or absolute time in a hand-written prompt). Prefer concrete / onomatopoeic events.
+
+Mark hero moments with `SPECIAL:` in `event` so QC can hear them.
+
+**In the clip or in post?** An sfx the world makes — a glass, a door, a slap, footsteps, rain — is
+in-clip: the video model makes it in sync with the picture. Sound design the world does not make —
+stinger, riser, whoosh, impact hit, reverse cymbal, drone hit, the silence drop — gets `"post": true`:
+it stays out of the clip prompt and is made over the edit, where it can land across a cut. A hit that
+is both (the slap *and* a stinger on it) is two entries.
+
+| Beat | Typical special FX |
+| --- | --- |
+| 打脸 / reveal | sharp whoosh + muted impact + brief silence |
+| 身份揭露 | rising drone → glass-ring sting |
+| 强宠 / romantic turn | soft shimmer / heartbeat under VO |
+| Comedy mishap | spill / choke / bonk — diegetic if live-action |
+| Cliffhanger | near-silence + one iconic prop sound |
+| Transition | whip-pan whoosh (sparingly) |
+
+Few high-value FX per segment. If the beat is only atmosphere, put it in ambience.
+
+## Intent vocabulary bank (examples — not a closed list)
+
+Write these into `music.intent` and cue `note`. Pair with `enter | swell | duck | drop | silence | out`. Mix freely.
+
+**Tension / suspense:** suspense pad · ticking urgency · dread drone · cold empty room · stalking pulse · pressure build · quiet danger · held breath
+
+**Rise / climax:** slow rise · intensity climb · heartbeat acceleration · approaching climax · swelling strings · peak hit
+
+**Romantic:** romantic neon · soft satin · afterglow haze · tender piano · yearning guitar · candlelit hush · close whisper hush
+
+**Comedy / farce:** light comedy pizzicato · clumsy slapstick · awkward silence bed · wink-and-nudge · drunken wobble · karaoke parody · punchline duck-and-sting
+
+**Face-slap / twist / 爽点:** smug turnaround · reveal sting · power-flip brass · gotcha hit · vindication swell · identity-drop silence then sting
+
+**Sad / bruised:** lonely night · soft regret · bruised tenderness · rain-on-window · hollow after
+
+**Anger / control:** cold command · sharp authority · icy control · simmering rage · light disciplinary march
+
+
+**Place texture:** neon booth haze · muffled corridor bass · office night hum · leather sofa creak bed · fluorescent buzz · city window night
+
+**Transition / end:** whip cut · hard reset · freeze-frame sting · cliff hang · unresolved chord · fade-to-black
+
+Example: `intent: "awkward comedy pizzicato under neon booth haze"` · note: `swell — intensity climb into climax` · note: `drop — freeze-frame sting then silence`

package/skill/short-drama/reference/canvas-recipes.md

@@ -162,8 +162,11 @@ that character's lines to the label. Caps: 3 audio references on `minimax-h3-ref
 pictures then voices, in the order to connect them; drop them only if the model shows no audio
 reference support (`params.voiceSamples: false`).
 
-The segment's **soundtrack node** (Seed Audio; `params.audioTrack`, default `bed`) sits beside the
-video node — no `--from` for a bed, the voice display nodes in speaking order for `--full`:
+**Soundtrack.** By default (`params.audioTrack: post`) there is **no per-segment soundtrack node**:
+the clips carry no music, the episode is edited first, and the score + sound design are made over the
+edit, one Seed Audio node per ≤ 120 s section (see the post recipe below). Only in `bed` / `full`
+does each segment get a **soundtrack node** beside its video node — no `--from` for a bed, the voice
+display nodes in speaking order for `--full`:
 
 ```bash
 SD prompt <project> audio E01-01 > /tmp/E01-01-audio.txt                    # timecoded ambience / sfx / music
@@ -173,6 +176,20 @@ wji canvas gen A1 --yes --wait                                                #
 # --full: wji canvas add audio … --from D_C02_VOICE D_C01_VOICE  (→ @Audio1 @Audio2, the order the prompt names them)
 ```
 
+**Post soundtrack (`audioTrack: post`)** — after every clip is bound and the episode is edited:
+
+```bash
+wji asset upload ./E01-edit.mp4 --json | jq -r .id                            # the locked edit, no music → EDIT
+SD bind <project> --ep 1 wji.assets.edit EDIT
+SD prompt <project> score E01 > /tmp/E01-score.md                            # planned spotting sheet + ≤120 s sections
+# Sound Designer conforms it to the edit's real cuts and writes one bed prompt per section
+wji canvas add audio -c E1 -n "E01 score 1" -m seed-audio-1.0 --prompt-file /tmp/E01-score-1.txt --json | jq -r .id
+wji canvas gen <node> --yes --wait                                            # ask first; ~0.15 credits a minute
+# mix the sections under the edit at their offsets (ffmpeg), then:
+wji asset upload ./E01-mix.mp4 --json | jq -r .id                             # → MIX
+SD bind <project> --ep 1 wji.assets.mix MIX
+```
+
 `wan` mode, one clip per segment (the model speaks the dialogue itself, so no voice nodes):
 
 ```bash