venice-video-harness 2.11.2 → 2.13.0

package/.claude/agents/prompt-engineer.md

@@ -11,6 +11,8 @@ If the **Seedance 2.0 Skill OS** is installed (see this repo's README "Directing
 
 **Division of labor with the harness:** identity is locked by R2V references + the Seedance → Wan keyframe pass, and durations/model routing are decided at generation time. So direct **intention/camera/light/blocking/performance/sound**; let the pipeline own identity, duration, and routing.
 
+**Blocking is geometry, not vibes (rule 49).** Identity references lock what things look like, not where they are — placement is re-inferred on every generation unless the prompt states it. Every prompt with characters or key objects must say, in concrete spatial terms: WHERE each subject is relative to the location's named landmarks (`Location.spatialAnchors`), WHERE they sit in the frame (screen left/center/right; foreground/background), and WHICH WAY they face (eyeline target). Across a scene's shots, keep screen sides, relative positions, and eyelines constant (180-degree rule) unless a movement is written into the action. The shot's authored `blocking` field is injected verbatim by the pipeline — when writing prompts by hand, restate it, never paraphrase it differently per shot.
+
 ## Role
 Build optimized Venice AI image generation prompts that maintain character consistency across all storyboard panels.
 
@@ -28,13 +30,15 @@ Build optimized Venice AI image generation prompts that maintain character consi
 
 [SHOT] {shot type}, {camera angle}, {lens mm}, {camera movement}
 
-[SETTING] {location} - {time of day}, {atmosphere details from scene action}
+[SETTING] {location} - {time of day}, {atmosphere details from scene action}, fixed layout: {location spatialAnchors — named landmarks and their relative positions}
 
 [CHARACTERS]
-- {NAME} ({position}, {facing}): {FULL description - age, ethnicity, hair, eyes, face, build, height}, wearing {wardrobe}, expression: {emotion from context}
+- {NAME} ({position relative to a named landmark}, {screen side + depth}, {facing/eyeline}): {FULL description - age, ethnicity, hair, eyes, face, build, height}, wearing {wardrobe}, expression: {emotion from context}
 
 [ACTION] {what's happening in this specific shot}
 
+[BLOCKING] {the shot's authored blocking, verbatim — who is where relative to which anchor, frame side, depth, facing; consistent with adjacent shots in the scene}
+
 [MOOD] {scene mood}
 
 [LIGHTING] {lighting setup based on location + time + mood}
@@ -62,6 +66,7 @@ In addition to image prompts, build video-generation prompts for each shot. Vide
 
 1. **Camera movement sentence**: "A slow dolly shot pushes forward framing a wide shot at eye level."
 2. **Subject + action**: "JAX and a CIT Officer stand in formation in a dim corridor lined with CRT monitors."
+   - **Include the blocking geometry**: "JAX screen left at the security desk, facing right; the Officer screen right in the doorway, facing him." State each subject's position vs a named landmark, frame side, depth, and eyeline — and keep it identical to the adjacent shots' geometry unless the action moves someone.
 3. **Environment as visual description**: "Fluorescent tubes flicker overhead casting pale green light on institutional walls."
 4. **Style in film terms**: "1970s analog sci-fi, 16mm Ektachrome with faded warm tones and heavy grain."
 5. **Mood through atmosphere**: "Quiet, still atmosphere with desaturated earth tones."

package/.claude/agents/storyboard-qa.md

@@ -18,6 +18,13 @@ Analyze generated storyboard panels for character consistency and setting contin
 - **Location**: architectural style, neon signs, background elements should be consistent
 - **Color palette**: should match the locked aesthetic profile
 
+### Spatial Continuity (against the shot's blocking + prior same-location panels)
+- **Stated blocking honored**: each character/object is on the stated frame side, at the stated depth, facing the stated direction, positioned correctly relative to the location's named landmarks (`Location.spatialAnchors`)
+- **Screen sides held across coverage**: characters keep the side of frame they had in the previous same-location panel unless the script moves them
+- **Eyelines / screen direction (180-degree rule)**: if A looks right at B in one panel, A keeps looking right and B keeps looking left in the coverage that follows
+- **Landmark stability**: doors, windows, counters, and key props have not moved, mirrored, vanished, or rearranged between panels of the same location
+- A side-swap or mirrored geography that breaks the scene is **CRITICAL**; a single character on the wrong frame side vs stated blocking, or a relocated landmark, is **MODERATE**
+
 ### Aesthetic Adherence
 - **Style**: panels should match the series aesthetic (webtoon, anime noir, etc.)
 - **Palette**: color grading should be consistent with the locked palette
@@ -40,10 +47,11 @@ Summarize with:
 ## How To Invoke
 
 This agent is called automatically after `storyboard-episode` generates panels. It reads:
-1. `series.json` for character descriptions, wardrobe, and aesthetic
-2. `script.json` for which characters appear in which shots
+1. `series.json` for character descriptions, wardrobe, aesthetic, and each location's `spatialAnchors`
+2. `script.json` for which characters appear in which shots and each shot's `blocking`
 3. The generated panel PNGs in `scene-001/`
 4. Character reference images in `characters/<name>/front.png` for comparison
+5. The nearest prior panel from the same location (attached automatically by `qa-storyboard`) for spatial-continuity comparison
 
 ## Integration
 - Runs as a sub-agent via the Task tool after panel generation

package/.claude/commands/qa-storyboard.md

@@ -32,11 +32,18 @@ Run a visual QA analysis on generated storyboard panels for character and settin
    - Art style matches locked aesthetic?
    - Color grading consistent?
 
+   **Spatial Continuity** (compare against the shot's `blocking`, the location's `spatialAnchors`, and the nearest prior panel in the same location):
+   - Each character/object on the stated frame side, at the stated depth, facing the stated direction?
+   - Positions correct relative to the location's named landmarks?
+   - Screen sides held from the previous same-location panel (no side-swaps)?
+   - Eyelines / screen direction preserved (180-degree rule)?
+   - Landmarks (doors, windows, counters, key props) unmoved — no mirroring, vanishing, or rearranging?
+
 5. Rate each panel:
-   - **PASS**: Character and setting match descriptions
-   - **FLAG-CRITICAL**: Character appearance is wrong (wrong hair, wrong outfit, wrong body type)
-   - **FLAG-MODERATE**: Minor drift but recognizable (slightly different shade, small detail off)
-   - **FLAG-LOW**: Stylistic variance within acceptable range
+   - **PASS**: Character, setting, and spatial geometry match descriptions
+   - **FLAG-CRITICAL**: Character appearance is wrong (wrong hair, wrong outfit, wrong body type) OR a spatial flip that breaks the scene (characters swapped sides, geography mirrored/rearranged vs the previous panel)
+   - **FLAG-MODERATE**: Minor drift but recognizable (slightly different shade, small detail off); character on the wrong frame side vs stated blocking; a relocated landmark
+   - **FLAG-LOW**: Stylistic variance or small placement deviation within acceptable range
 
 6. Present the QA report to the user:
    - Show each flagged panel inline with the specific issues
@@ -52,5 +59,6 @@ Run a visual QA analysis on generated storyboard panels for character and settin
 ## Important
 - Always load and show the character reference images alongside flagged panels for visual comparison
 - Be specific about what's wrong: "Sera's hair is shoulder-length bob in shot 4 but should be long flowing dark hair per her description"
-- Focus most on CRITICAL issues (wrong appearance) -- these break immersion across episodes
+- Be equally specific about spatial issues: "JAX was screen left at the counter in shot 3 but is screen right in shot 4 with no scripted movement — side-swap, regenerate"
+- Focus most on CRITICAL issues (wrong appearance, broken scene geography) -- these break immersion across episodes
 - Run this QA automatically after every storyboard-episode generation

package/.claude/commands/workshop-episode.md

@@ -2,6 +2,16 @@ Workshop an episode script collaboratively with the user.
 
 **Direct before you draft.** If the idea is vague, run the Seedance 2.0 Skill OS `seedance-interview` (short form for a fast brief); for a longer or multi-clip story run `seedance-sequence` to set the story spine and one directorial voice. Then, for every shot, name **one intention** and derive camera/light/blocking/performance/sound from it — do NOT stack "cinematic" adjectives. Read each drafted `description` back through `directing-engine`, and strip empty quality-boosters with `seedance-antislop` + `vocab/*`. If Seedance OS isn't installed, apply the same principle from memory (see this repo's README "Directing layer"). The in-code workshop system prompt (`src/mini-drama/cli.ts`) already carries this instruction; a directed concept from the user is what makes it land.
 
+**Author the geometry, not just the action (rule 49).** Spatial consistency is
+built into the script, not patched at generation time. Give every location a
+`spatialAnchors` field (3-5 named landmarks and their fixed relative positions)
+and every character shot a `blocking` field: each character/object's position
+relative to those named anchors, their frame side (screen left/center/right),
+depth (foreground/background), and facing/eyeline. Across consecutive shots in
+a scene, characters keep their screen sides and relative positions unless a
+movement is written into the action; preserve screen direction and eyelines
+(180-degree rule); close-ups still name what's behind/beside the subject.
+
 1. Ask the user for the episode concept/outline (what happens in this episode?)
 2. Draft a shot-by-shot script in the EpisodeScript JSON format
 3. Each shot should specify:
@@ -9,6 +19,8 @@ Workshop an episode script collaboratively with the user.
    - duration (3s/5s/8s/10s/13s/15s)
    - videoModel ("action" for movement/dialogue, "atmosphere" for establishing/static)
    - description (visual scene description)
+   - location (slug of a defined location)
+   - blocking (concrete geometry: positions vs named location anchors, frame side, depth, facing/eyeline — consistent with adjacent shots)
    - characters present
    - dialogue (if any)
    - sfx (optional sound effects hint)

package/.claude/skills/character-consistency/SKILL.md

@@ -95,8 +95,8 @@ The most critical consistency layer for video generation. **R2V models are the o
 - **Prompt integration:** Character names replaced with `@Element1`, `@Element2` tokens
 
 ### Critical Rules
-1. **Only R2V models support elements/reference_image_urls.** The Kling multi-shot model (`kling-o3-pro-image-to-video`), Kling V3 Pro, Veo 3.1, and all other non-R2V models have ZERO reference support.
-2. **Never group shots with different characters into multi-shot units.** Multi-shot units use non-R2V models. Characters lose all identity anchoring.
+1. **Only R2V models support elements/reference_image_urls.** Kling V3 Pro, Veo 3.1, `kling-o3-pro-image-to-video`, and all other non-R2V models have ZERO reference support.
+2. **Never group shots with different characters into multi-shot units.** Multi-shot units render on Seedance R2V Enhanced by default (2026-08-05) with ONE slot plan per generation — disjoint characters can't all anchor correctly. (Legacy `videoDefaults.multiShotModel: kling-o3-pro-image-to-video` overrides have no references at all.)
 3. **For talk shows, interviews, and panel formats:** Set `mustStaySingle: true` on all shots to force R2V singles.
 
 ### Format-Specific Guidance
@@ -107,6 +107,31 @@ The most critical consistency layer for video generation. **R2V models are the o
 | Continuous action scene | Multi-shot OK if same characters | Temporal continuity benefits outweigh |
 | Drama with cuts | Singles for close-ups/reactions, multi-shot for continuous action | Balance both |
 
+## Layer 5: Spatial Anchoring (placement, not just appearance)
+
+Identity references lock what a character LOOKS like; they say nothing about
+WHERE the character is. Placement is re-inferred on every generation unless
+each prompt states the geometry the same way (rule 49):
+
+- **`Location.spatialAnchors`** locks a place's geography: 3-5 named landmarks
+  and their fixed relative positions. Injected as "Fixed layout (never
+  rearrange): …" into every panel/plate/video prompt for shots in that
+  location, and baked into the location's reference angles.
+- **`ShotScript.blocking`** locks the shot's geometry: each character/object's
+  position relative to the named anchors, frame side (screen left/center/
+  right), depth, and facing/eyeline. Injected verbatim (with @ImageN name
+  substitution) into panel, plate, and video prompts.
+- **Storyboard blocking plates** encode the same geometry visually and carry a
+  role clause restricting them to composition/blocking authority; the video
+  prompt additionally forbids mirroring/swapping who stands where.
+- **Continuity rules:** characters keep their screen sides and relative
+  positions across a scene's shots unless a movement is scripted; eyelines
+  and screen direction obey the 180-degree rule; close-ups still name what's
+  behind/beside the subject.
+- **QA:** `qa-storyboard` checks a SPATIAL CONTINUITY dimension against the
+  stated blocking and the previous same-location panel; a side-swap or
+  mirrored geography is FLAG-CRITICAL.
+
 ## Usage
 ```typescript
 import { buildCharacterReferencePrompt } from './mini-drama/prompt-builder.js';

package/.claude/skills/shot-composition/SKILL.md

@@ -47,6 +47,31 @@ When the **Seedance 2.0 Skill OS** is installed (`.claude/skills/seedance-20/`),
 5. Transitions between scenes match the screenplay's indicated transition
 6. Insert shots for objects mentioned prominently in action lines
 
+## Spatial Blocking Rules (rule 49 — placement is authored, not inferred)
+
+AI video models re-infer placement on every generation, so shot plans must
+state the geometry explicitly and keep it stable across a scene's coverage:
+
+1. **Anchor to the location's named landmarks.** Each `Location` carries
+   `spatialAnchors` (3-5 landmarks with fixed relative positions). Blocking
+   language must reference those names ("at the bar counter", "in the
+   doorway") so "by the window" means one specific window in every shot.
+2. **Every character shot gets a `blocking` field**: each character/key
+   object's position relative to the named anchors, their frame side (screen
+   left / center / right), depth (foreground / midground / background), and
+   facing/eyeline direction.
+3. **Hold screen sides across coverage.** Characters keep the side of frame
+   they had in the master unless a movement is scripted — and if they move,
+   the movement is the shot's action, written into the description.
+4. **Respect the 180-degree rule.** Pick the scene's axis at the master and
+   keep eyelines/screen direction consistent in every single and reaction:
+   if A looks right at B, A keeps looking right and B keeps looking left.
+5. **Close-ups still name the geography.** State what is behind/beside the
+   subject so backgrounds match the wide ("the neon window behind her
+   shoulder"), or backgrounds will drift per generation.
+6. **Establishing shots restate the full landmark layout** so the scene's
+   geography is re-locked at every scene boundary.
+
 ## Transition Types
 - **CUT**: Standard, default between shots
 - **DISSOLVE**: Time passage, dream, memory

package/.claude/skills/venice-agent-guide/SKILL.md

@@ -35,10 +35,11 @@ in the other.
 - Front-load style at the START of every prompt; keep Seedance prompts under ~60 words with the 5-part structure (Subject, Action, Camera, Style, Constraints).
 - Direct the beat, do not decorate it: name one intention and derive camera/light/blocking/performance from it. Do not stack "cinematic / epic / 4k" adjectives.
 - Re-anchor every separately-rendered shot to the SAME locked references and restate the character invariant traits (including relative size) in every prompt.
+- State placement explicitly in every prompt: lock each location's landmark geography (`spatialAnchors`) and give every character shot a `blocking` field (position vs named anchors, screen side, depth, facing/eyeline). Keep screen sides and eyelines constant across a scene unless a movement is scripted.
 - Prefer native model dialogue (Seedance 2.0, HappyHorse 1.1 with voice-donor references) over exact TTS lip-sync.
 
 ## Where the full knowledge lives
-- `AGENTS.md` — 47 rules and 27 production anti-patterns, shipped in the package.
+- `AGENTS.md` — 49 rules and 28 production anti-patterns, shipped in the package.
 - `.claude/skills/` — `venice-api`, `venice-video-model-routing`, `character-consistency`, `shot-composition`, `burn-in-subtitles`, `video-editing`, and more.
 - `.claude/commands/` — 20 step-by-step playbooks; `.claude/agents/` — 10 sub-agent roles.
 - Read the relevant playbook before running a workflow. Validate model capabilities against `src/venice/models.ts` before an API call.

package/.claude/skills/venice-video-model-routing/README.md

@@ -119,7 +119,8 @@ venice-video-model-routing/
 - `veo3.1-fast-image-to-video` -- atmosphere tier (establishing shots, inserts)
 - `kling-o3-standard-reference-to-video` -- character consistency tier (identity-critical shots, default)
 - `kling-o3-pro-reference-to-video` -- character consistency tier (higher quality, slower)
-- `kling-o3-pro-image-to-video` -- multi-shot units
+- `seedance-2-0-enhanced-reference-to-video` -- multi-shot units (default since 2026-08-05; reference-first with `Lens switch.` separators)
+- `kling-o3-pro-image-to-video` -- legacy multi-shot lane (explicit `videoDefaults.multiShotModel` override only; no reference support)
 - `vidu-q3-image-to-video` -- 1080p output with reference image support
 - `wan-2.6-image-to-video` -- configurable audio, various durations
 - `sora-2-image-to-video` -- high quality, requires aspect ratio

package/.claude/skills/venice-video-model-routing/SKILL.md

@@ -162,7 +162,9 @@ ACTION_MODEL              = 'seedance-2-0-image-to-video'
 ATMOSPHERE_MODEL          = 'seedance-2-0-image-to-video'
 CHARACTER_CONSISTENCY_MODEL = 'seedance-2-0-reference-to-video'
 KLING_R2V_MODEL           = 'kling-o3-standard-reference-to-video'  (fallback for 3+ chars)
-MULTISHOT_MODEL           = 'kling-o3-pro-image-to-video'
+MULTISHOT_MODEL           = 'seedance-2-0-enhanced-reference-to-video'  (default since 2026-08-05; resolveMultiShotModel())
+# Legacy: 'kling-o3-pro-image-to-video' — explicit videoDefaults.multiShotModel override only.
+# It has NO elements and NO reference_image_urls, so it drops all identity anchoring.
 
 # Single image default for ALL panels (Venice removed the Seedance face restriction, 2026-07)
 IMAGE_GENERATION          = 'nano-banana-2'          # character + faceless panels alike
@@ -318,6 +320,8 @@ Construct prompts differently depending on the resolved model's capabilities:
 - Dialogue speaker uses image ref: `[@Image1, voice description, delivery]: "dialogue line"`
 - Attach `reference_image_urls` at the API layer in the slot-planner order (up to 9: character primaries → storyboard blocking plate → location angles → second character angles)
 - Emit a role clause per non-character slot: location angles ("@Image4 is a second angle of the same location"), blocking plates ("use ONLY for composition, blocking, and spatial relationships")
+- Carry the shot's authored `blocking` verbatim (`Blocking: @Image1 at the bar counter, screen left, facing right; …`) and the location's `spatialAnchors` ("Fixed layout (never rearrange): …") — placement is stated per shot, never left to the model (rule 49)
+- With a blocking plate bound, the prompt also pins geometry to it: characters keep their side of the scene and position relative to the plate's landmarks — no mirroring, swapping, or rearranging; plateless location shots pin geography to the first location `@ImageN` slot instead
 - **Pure reference mode:** no `image_url` start frame when the slot plan is populated — references carry all consistency
 - Keep prompts concise. Use the 5-part structure: Subject, Action, Camera, Style, Constraints.
 - Use physics-aware language: describe forces and materials, not just actions
@@ -345,9 +349,22 @@ Construct prompts differently depending on the resolved model's capabilities:
 - Structure: camera term first, then description, then dialogue with delivery cues
 - Veo requires `resolution: '720p'`; Kling does NOT accept `resolution`/`aspect_ratio` (derived from input image)
 
-### Native Multi-Shot (Kling 3.0, up to 6 shots, 15s max)
+### Native Multi-Shot (default: Seedance R2V Enhanced with `Lens switch.`)
 
-Kling 3.0 supports native multi-shot generation — a single prompt produces a single video with multiple distinct shots and automatic cuts between them. This works with the R2V model (`kling-o3-standard-reference-to-video`) which supports `elements` for identity anchoring across all shots.
+**Multi-shot units render on `seedance-2-0-enhanced-reference-to-video` by default (2026-08-05)** — the same reference-first lane as singles. `buildMultiShotPrompt()` produces ONE Seedance generation covering the window's beats:
+
+1. **Identity declarations up front** from the unit's @Image slot plan: `@Image1 is Bob — wearing …`
+2. **Role clauses** for the non-character slots (blocking plate, location angles) — identical clause text to singles
+3. **Per-beat blocks** labeled `Shot N (Xs):` with camera, description (names → `@ImageN`), authored `Blocking:` restated per beat (rule 49), and `[@ImageN, voice, delivery]: "line"` dialogue
+4. **Literal `Lens switch.` lines** between beats (rule 21)
+5. **Geometry hold** pinned to the plate (or first location angle): no mirroring/swapping across the internal cuts
+6. Compact aesthetic + audio-exclusion suffix; ≤2500 chars
+
+The unit renders in **pure reference mode** — `reference_image_urls` pushed in slot-plan order, no `image_url` start frame — and takes voice-donor `reference_audio_urls` (@AudioN) for its dialogue speakers. Grouping windows never span locations (one slot plan per generation).
+
+### Legacy Multi-Shot Format (Kling 3.0 — explicit override only)
+
+Used only when `videoDefaults.multiShotModel` is explicitly set to a non-@Image-tag model (e.g. `kling-o3-pro-image-to-video`, which has NO reference support — identity is prompt-text only).
 
 **Prompt structure (per [Kling 3.0 Prompting Guide](https://blog.fal.ai/kling-3-0-prompting-guide/)):**
 1. **Define subjects up front** with `@Element` refs and traits — Kling locks these across all shots
@@ -400,7 +417,7 @@ For formats with **frequent speaker cuts** (talk shows, interviews, debate panel
 
 ### All Character Shots Must Be R2V Singles
 
-- **Never group shots with different speakers into multi-shot units.** The Kling multi-shot model (`kling-o3-pro-image-to-video`) does NOT support `elements` or `reference_image_urls` — characters lose all identity anchoring.
+- **Never group shots with different speakers into multi-shot units.** Even on the reference-first Seedance multi-shot default, a unit builds ONE slot plan for the window — disjoint speakers mean the plan can't anchor everyone correctly. (On the legacy Kling override the failure is total: `kling-o3-pro-image-to-video` has NO `elements` or `reference_image_urls`, so characters lose all identity anchoring.)
 - **Set `mustStaySingle: true`** on all shots in talk show scripts, or ensure the generation planner only groups shots that share the same characters.
 - **Every character shot uses R2V** — `seedance-2-0-reference-to-video` for 1-2 character shots (flat `reference_image_urls` with `@Image` tags), auto-fallback to `kling-o3-standard-reference-to-video` for 3+ characters (structured `elements`).
 
@@ -443,7 +460,7 @@ Seedance 2.0 (now the default for both atmosphere and character shots) accepts *
 
 ### Identity Failures
 
-- **Grouping different-character shots into multi-shot units:** The Kling multi-shot model has no reference image support. Characters rendered in a multi-shot unit with no `elements` lose identity completely. Always verify pairwise character overlap before grouping.
+- **Grouping different-character shots into multi-shot units:** A multi-shot unit builds ONE reference slot plan for the window — disjoint characters can't all anchor correctly (and on the legacy Kling override there are no references at all). Always verify pairwise character overlap before grouping.
 - **Frame chaining when new character enters:** The video model invents the new character's appearance from nothing. Always use the panel image (which was refined against character references) as the start frame when a new character appears.
 - **Multi-edit with more than 2 character references:** The multi-edit endpoint accepts max 3 images total (base + 2 refs). Exceeding this drops references silently.
 - **Sequential action in image descriptions:** Causes comic-panel layouts instead of single frames. Separate the single-frame panel description from the full video action description.

package/.claude/skills/venice-video-model-routing/references/decision-trees.md

@@ -202,7 +202,12 @@ Legend: Y = supported, X = not supported, (400) = returns HTTP 400 error
 
 ## 7. Multi-Shot Grouping Decision Tree
 
-Determines whether consecutive shots should be grouped into a single Kling multi-shot unit or rendered as individual R2V singles.
+Determines whether consecutive shots should be grouped into a single native
+multi-shot unit or rendered as individual R2V singles. Multi-shot units render
+on `seedance-2-0-enhanced-reference-to-video` by default (2026-08-05) — one
+reference-first generation with `Lens switch.` separators, carrying the full
+@Image slot plan. The Kling i2v lane is an explicit
+`videoDefaults.multiShotModel` override only.
 
 ```
 INPUT: window of consecutive shots
@@ -216,13 +221,16 @@ START
   |   -> SINGLES (these should always render independently)
   |
   +-- total duration exceeds 15s?
-  |   -> SINGLES (exceeds Kling multi-shot limit)
+  |   -> SINGLES (exceeds the single-generation limit on both lanes)
+  |
+  +-- shots span more than one location?
+  |   -> SINGLES (one slot plan per generation — rule 21b)
   |
   +-- do consecutive pairs share at least one character?
   |   |
   |   NO -> SINGLES (different characters lose R2V anchoring)
   |   |
-  |   YES -> GROUP (temporal continuity benefits outweigh)
+  |   YES -> GROUP (identity/environment/lighting hold inside one generation)
   |
   +-- CRITICAL: for talk shows / interviews / panels:
       -> PREFER SINGLES (identity anchoring > temporal continuity)

package/AGENTS.md

@@ -383,7 +383,7 @@ Use `POST /video/quote` (via `quoteVideo()`) to estimate costs before committing
 15. **Match lighting across consecutive shots in the same location.** When generating panels for sequential shots in the same environment, style-match later shots against earlier ones. Explicitly describe the established lighting in each subsequent prompt.
 16. **Use `silhouetteCharacters` for distant/silhouetted figures.** Characters visible only as silhouettes (e.g., figure in doorway) go in `silhouetteCharacters`, not `characters`. This ensures they appear in panels without triggering R2V routing or "no people" negative prompts.
 17. **Describe the Venice AI logo as crossed-keys, never as "triple-V" or "VVV".** The actual logo is two ornate skeleton keys crossed in an X with a chevron/book at top. Use the full geometric description in prompts, or multi-edit with the logo PNG as reference.
-18. **Use Kling 3.0 native multi-shot for sequences within a single generation.** Structure: define subjects with `@Element` refs up front, label shots as `Shot N (Xs):`, use `[Character, voice description]: "dialogue"` format, and separate shots with `Immediately, cut to:`. This produces a single video with multiple shots — no concatenation needed. Max 6 shots, 15s total. See [Kling 3.0 Prompting Guide](https://blog.fal.ai/kling-3-0-prompting-guide/).
+18. **Multi-shot units render on Seedance 2.0 R2V Enhanced by default (2026-08-05) — the Kling multi-shot lane is an explicit override only.** `DEFAULT_MULTISHOT_MODEL` / `resolveMultiShotModel()` in `src/series/types.ts` route every multi-shot generation unit to the SAME reference-first lane as singles: one Seedance native multi-shot generation with `Lens switch.` separators (rule 21), anchored to the full @Image slot plan (character sheets, blocking plate, location angles) in pure reference mode — no start frame. `buildMultiShotPrompt()` dispatches on the resolved model. The old default, `kling-o3-pro-image-to-video`, has NO `elements` and NO `reference_image_urls` support, so every multi-shot unit silently dropped ALL identity anchoring (anti-pattern 1's trap, and the reason grouping rules had to police characters so hard). To deliberately use the Kling format anyway, set `videoDefaults.multiShotModel: "kling-o3-pro-image-to-video"` — the legacy Kling 3.0 structure (subjects up front, `Shot N (Xs):` labels, `[Character, voice]: "dialogue"`, `Immediately, cut to:` separators, max 6 shots / 15s; see the [Kling 3.0 Prompting Guide](https://blog.fal.ai/kling-3-0-prompting-guide/)) is kept behind that override.
 19. **Seedance 2.0 R2V uses `@Image` tags, not `@Element` tags.** When the resolved model is Seedance R2V, replace character names with `@Image1`, `@Image2` in prompts. Do NOT use `@Element` tags — Seedance does not support structured elements. The prompt builder handles this automatically via `useImageTags`.
 20. **Keep Seedance prompts under 60 words for best results.** Seedance responds to precision, not volume. Use the 5-part structure: Subject, Action (present tense, one verb), Camera (shot size + movement), Style (lighting, color), Constraints (what to exclude). See [Seedance prompting guide](https://venice.ai/blog/seedance-sota-video-generation-live-on-venice).
 21. **Default to Seedance native multi-shot for any 2–3 beat scene.** Before planning a bundle of separate Seedance renders, first ask whether the beats can fit into ONE generation up to 15s with `Lens switch.` separators between them. The native multi-shot path is the default; bundled separate renders are the fallback. Identity, environment, and lighting hold across the lens switches inside a single generation, and the result costs and takes ~3× less than three separate renders. Reach for a bundle only when (a) a beat needs Wan 2.7 lip-sync to a specific dialogue MP3, (b) beats span different locations or non-overlapping characters, or (c) total runtime > 15s. Prompt structure: one front-loaded STYLE + character anchor at the top, then per-beat `Shot N (Xs): ...` blocks with the 5-part structure (Subject, Action, Camera, Style, Constraints) kept under ~50 words each, separated by literal `Lens switch.` lines. Pass character refs once via `reference_image_urls` and reference them inline as `@Image1`, `@Image2`, etc. in every beat.
@@ -422,14 +422,16 @@ Use `POST /video/quote` (via `quoteVideo()`) to estimate costs before committing
 
 48. **The CLI is self-describing — discover state and order from it, do not guess.** `venice-video agent-guide [--json]` prints the core operating rules (this section in miniature) and ships inside the binary, so it is available even on a bare global install with no `AGENTS.md`. `venice-video pipeline [--json]` prints the ordered stages, their gates, and the command that advances each; it mirrors `classifyEpisode` in `src/session/status.ts` (rule 45 — change both together). `venice-video status -p <project> [--json]` reports where a project stands and the next command. `--json` is supported on `status`, `pipeline`, `agent-guide`, `doctor`, and `queue` (and globally as `venice-video --json <command>`); it prints exactly one JSON object on stdout, and the human text rendering is unchanged. Exit codes are honest — `status` with no project exits non-zero — and ordinary errors are a clean `error:` line, not a stack trace (`VENICE_VIDEO_DEBUG=1` for the stack). The condensed guide is kept in `src/agent/guide.ts` and duplicated as an installable skill at `.claude/skills/venice-agent-guide/`; when a non-negotiable here changes, change those two as well.
 
+49. **Spatial consistency is authored, not inferred — every prompt states placement relative to locked anchors (2026-08-05).** Visual consistency (identity, wardrobe, palette) is handled by the reference stack, but *spatial* consistency — who stands where, which side of frame, facing which way, relative to which landmark — drifts unless it is stated the same way in every generation. The harness now carries geometry as first-class data: (a) **`Location.spatialAnchors`** is the locked geography of a place — 3-5 named landmarks and their fixed relative positions ("bar counter along the back wall; entrance door opposite it; neon window left of the door as seen from the counter"). It is baked into the location reference angles at generation time, injected as "Fixed layout (never rearrange): …" into every panel and video prompt for shots tagged with the location, and is sticky on merge — an existing anchor set is never overwritten by a later script part. (b) **`ShotScript.blocking`** is the shot's authored geometry — 1-2 sentences placing each character/object relative to the named anchors, the frame (screen left/right, foreground/background), and their facing/eyeline. The workshop and script LLM prompts require it for every character shot, with continuity rules: characters keep their screen side and relative positions across consecutive shots unless a movement is written into the action; screen direction and eyelines are preserved (180-degree rule); close-ups still name what is behind/beside the subject. It is injected verbatim (with @ImageN/@ElementN name substitution) into the panel prompt (`BLOCKING: …`), the video prompt (`Blocking: …`), the multi-shot per-beat blocks (both the Seedance native lane and the legacy Kling format), and it seeds the beat's storyboard blocking-plate description. (c) The video prompt's plate clause now forbids mirroring/swapping ("each character stays on the same side of the scene… do not mirror, swap, or rearrange who stands where"), and plateless location shots get a geography-hold clause anchored to the location's first `@ImageN` slot. (d) **QA reads geometry**: `qa-storyboard` evaluates a fourth SPATIAL CONTINUITY dimension against the shot's stated blocking and the location's landmarks, and attaches the nearest prior panel from the same location so side-swaps, mirrored geography, and moved landmarks are caught against real coverage — a spatial flip that breaks the scene is FLAG-CRITICAL. (e) `workshop-episode` warns when character shots are missing `blocking` or locations are missing `spatialAnchors`; `insert-shot` inherits the anchor shot's location and blocking (same-scene splices keep the established geography) and takes `--location` / `--blocking` overrides; `add-location` takes `--spatial-anchors`. The failure mode this kills: characters teleporting across the room, swapping frame sides between coverage, and set geography silently mirroring between shots that read as the same scene (see anti-patterns 20 and 26).
+
 ## Learned Anti-Patterns (Production Issues Log)
 
 Issues discovered during production and their fixes. The agent should internalize these to avoid repeating them.
 
 ### 1. Multi-Shot Grouping Bug: Wrong Character Overlap Check
-**Symptom:** Shots cutting between different characters (e.g., Chad-only → Vivienne-only) were grouped into Kling multi-shot units, which use `kling-o3-pro-image-to-video` — a model with NO `elements` or `reference_image_urls` support. Characters lost all identity anchoring.
+**Symptom:** Shots cutting between different characters (e.g., Chad-only → Vivienne-only) were grouped into multi-shot units, which at the time used `kling-o3-pro-image-to-video` — a model with NO `elements` or `reference_image_urls` support. Characters lost all identity anchoring.
 **Root cause:** `hasOverlappingCharacters()` checked each shot's characters against the union pool instead of requiring pairwise overlap between consecutive shots.
-**Fix:** Rewrote to require every consecutive pair of shots to share at least one character. Shots with disjoint characters now always render as singles with R2V.
+**Fix:** Rewrote to require every consecutive pair of shots to share at least one character. Shots with disjoint characters now always render as singles with R2V. (2026-08-05: the referenceless-model half of this trap was removed at the root — multi-shot units now default to Seedance R2V Enhanced with the full reference slot plan, see rule 18. The pairwise-overlap check remains because grouping disjoint characters is still wrong on any lane.)
 **File:** `src/mini-drama/generation-planner.ts`
 
 ### 2. Character Reference Style Inconsistency Across Angles
@@ -594,6 +596,11 @@ Issues discovered during production and their fixes. The agent should internaliz
 **Symptom:** (a) R2V/i2v requests with human-face references returned 409 `needs_consent`. (b) After ~20 failed queue attempts, the account tripped a 30s 429.
 **Fix:** (a) Attach `consents.seedance: { confirmed_terms_and_privacy: true, confirmed_legal_right: true, confirmed_screening_acknowledged: true }` to `/video/queue` for face-bearing requests — surface the policy text once per session for the user to ack, then auto-attach. (b) Back off on repeated 4xx; don't hammer the queue endpoint.
 
+### 28. Spatial Drift Across Shots: Side-Swaps, Teleporting Props, Mirrored Geography
+**Symptom:** Two shots that read as the same scene disagree spatially — a character who was screen-left in the master is screen-right in coverage; a prop on the table has moved or vanished; the room's layout is mirrored (door now on the other side); eyelines cross the 180-degree line so speakers appear to look the same way.
+**Root cause:** Prompts described the *action* but left placement implicit, so every generation re-inferred the geometry from scratch. Identity references lock what things look like, not where they are; the blocking plate helps, but a plate alone (without matching placement language in the prompt) is a weak anchor the model can reinterpret or mirror.
+**Fix:** Author the geometry once and restate it identically everywhere (rule 49): lock each location's landmark layout in `Location.spatialAnchors`, write per-shot `blocking` (position vs named anchors + frame side + depth + facing/eyeline), and let the harness inject both into panel, plate, and video prompts verbatim. Keep screen sides and eyelines constant across a scene's shots unless a movement is scripted. QA panels for spatial continuity against the previous same-location panel (the fourth `qa-storyboard` dimension); treat a side-swap or mirrored geography as FLAG-CRITICAL and re-roll before rendering video.
+
 ## Output
 
 Generated project output belongs in:

package/CHANGELOG.md

@@ -1,5 +1,111 @@
 # Changelog
 
+## 2.13.0 — 2026-08-05
+
+### Changed
+
+- **Multi-shot units now default to Seedance 2.0 R2V Enhanced — the Kling i2v
+  lane is an explicit override only.** `DEFAULT_MULTISHOT_MODEL` /
+  `resolveMultiShotModel()` (new, `src/series/types.ts`) route every
+  multi-shot generation unit to `seedance-2-0-enhanced-reference-to-video`,
+  the same reference-first lane as singles. The old default,
+  `kling-o3-pro-image-to-video`, has NO `elements` and NO
+  `reference_image_urls` support, so every multi-shot unit silently dropped
+  all identity anchoring (anti-pattern 1's underlying trap). Specifics:
+  - **New `buildMultiShotPrompt()`** dispatches on the resolved model. The
+    Seedance path builds ONE native multi-shot generation per rule 21:
+    identity declarations from the unit's @Image slot plan up front, role
+    clauses for the blocking plate and location angles, per-beat
+    `Shot N (Xs):` blocks (names → `@ImageN`, authored `Blocking:` restated
+    per beat, `[@ImageN, voice, delivery]: "line"` dialogue), literal
+    `Lens switch.` separators, and a geometry-hold clause pinned to the plate
+    (rule 49). ≤2500-char video prompt cap with aesthetic-first trimming.
+  - **Pure reference mode for the whole unit:** `reference_image_urls` pushed
+    in slot-plan order (union of the window's characters + the beat's plate +
+    location angles), no `image_url` start frame, no `end_image_url`. The
+    unit also carries voice-donor `reference_audio_urls` (@AudioN) for its
+    dialogue speakers, deduped across beats within Venice's 3-clip budget.
+  - **Planner:** unit type renamed `kling-multishot` → `multishot` (the old
+    name still parses from existing generation-plan.json files); units carry
+    the resolved model; multi-shot windows can no longer span locations
+    (rule 21b — one slot plan per generation); the 15s window limit message
+    no longer names Kling (both lanes cap at 15s).
+  - **Override:** `videoDefaults.multiShotModel` (new) selects another lane
+    explicitly — setting it to `kling-o3-pro-image-to-video` restores the
+    legacy Kling 3.0 format (`buildKlingMultiShotPrompt` is retained and
+    still exported). `KLING_MULTISHOT_MODEL` is deprecated but resolvable.
+  - **Docs:** AGENTS.md rule 18 rewritten (Seedance default, Kling as
+    override), anti-pattern 1 annotated, `venice-video-model-routing`
+    (SKILL/README/decision trees) and `character-consistency` updated.
+  - **Tests:** `test-spatial-consistency.mjs` now covers the dispatcher —
+    default resolution, Lens-switch structure, slot plan (plate + location),
+    per-beat blocking with @ImageN substitution, geometry hold, and the
+    explicit Kling override path; `audio-routing.test.mjs` asserts the
+    grouped unit renders on Seedance R2V.
+
+## 2.12.0 — 2026-08-05
+
+### Added
+
+- **Spatial consistency is now authored data, not per-generation inference
+  (rule 49).** Visual consistency was already reference-anchored, but *where*
+  characters and objects sit — screen side, depth, facing, position relative to
+  the set — was re-inferred by the model on every generation, which is where
+  side-swaps, teleporting props, and mirrored geography came from
+  (anti-pattern 28). Two new fields carry the geometry through the whole
+  pipeline:
+  - **`Location.spatialAnchors`** — the locked geography of a place: 3-5 named
+    landmarks and their fixed relative positions. Baked into the location's
+    reference angles at generation time, injected as
+    `Fixed layout (never rearrange): …` into every panel and video prompt for
+    shots tagged with the location, and sticky on merge (an existing anchor
+    set is never overwritten by a later script part). `add-location` takes
+    `--spatial-anchors`.
+  - **`ShotScript.blocking`** — the shot's authored geometry: 1-2 sentences
+    placing each character/object relative to the named anchors, the frame
+    (screen left/right, foreground/background), and their facing/eyeline.
+    Injected verbatim (with `@ImageN`/`@ElementN` name substitution) into the
+    panel prompt (`BLOCKING: …`), the video prompt (`Blocking: …`), and the
+    Kling multi-shot per-shot blocks; it also seeds the beat's storyboard
+    blocking-plate description. `insert-shot` inherits the anchor shot's
+    location and blocking for same-scene splices and takes `--location` /
+    `--blocking` overrides.
+- **The script LLM is now required to author the geometry.** Both workshop
+  system prompts (`workshop` in `workshop.ts` and `workshop-script` in
+  `cli.ts`) demand `spatialAnchors` per location and `blocking` per character
+  shot, with continuity rules: characters keep their screen sides and relative
+  positions across consecutive shots unless a movement is written into the
+  action, screen direction and eyelines obey the 180-degree rule, and blocking
+  always references the location's named anchors. `workshop-script` warns when
+  character shots are missing `blocking` or locations are missing
+  `spatialAnchors` (post-condition advisory, same pattern as the duration and
+  no-music checks).
+- **`qa-storyboard` reads geometry.** A fourth SPATIAL CONTINUITY dimension
+  checks each panel against the shot's stated blocking and the location's
+  landmarks, and the command now attaches the nearest prior panel from the
+  same location so side-swaps, mirrored geography, and moved landmarks are
+  caught against real coverage instead of prose alone. A spatial flip that
+  breaks the scene is FLAG-CRITICAL; a wrong frame side or relocated landmark
+  is FLAG-MODERATE.
+- **Stronger geometry clauses in video prompts.** The blocking-plate clause now
+  explicitly forbids mirroring/swapping ("each character stays on the same
+  side of the scene… do not mirror, swap, or rearrange who stands where"), and
+  plateless location shots get a geography-hold clause pinned to the location's
+  first `@ImageN` slot. Blocking plates themselves are prompted for legible
+  placement (screen side, distance, facing, landmark relations readable in one
+  look) and carry the location's fixed layout.
+- **Docs and knowledge pack updated together:** AGENTS.md rule 49 +
+  anti-pattern 28, README "carry these" rule 8, `agent-guide` (binary +
+  `venice-agent-guide` skill), `shot-composition` (spatial blocking rules),
+  `character-consistency` (Layer 5: spatial anchoring),
+  `venice-video-model-routing` (blocking/fixed-layout prompt lines),
+  `prompt-engineer` ([BLOCKING] template section), `storyboard-qa`, and the
+  `qa-storyboard` / `workshop-episode` playbooks.
+- **Tests:** `tests/test-spatial-consistency.mjs` covers blocking injection in
+  panel/video/multi-shot prompts, `@ImageN` substitution inside blocking,
+  fixed-layout injection, the no-mirroring and geography-hold clauses, plate
+  descriptions inheriting blocking, and the workshop prompt contract.
+
 ## 2.11.2 — 2026-08-05
 
 ### Changed

package/README.md

@@ -40,7 +40,7 @@ knowledge reaches you, which is the single largest predictor of output quality.
 
 | Surface | How it runs | What you get | Use when |
 |---|---|---|---|
-| **Repo-resident agent** | Agent's cwd is a clone of this repo | Everything: `AGENTS.md` (47 rules, 20 anti-patterns), `.claude/commands/`, `.claude/agents/`, `.claude/skills/`, `.cursor/rules/` | Authoring and iteration — the best results by a wide margin |
+| **Repo-resident agent** | Agent's cwd is a clone of this repo | Everything: `AGENTS.md` (49 rules, 28 anti-patterns), `.claude/commands/`, `.claude/agents/`, `.claude/skills/`, `.cursor/rules/` | Authoring and iteration — the best results by a wide margin |
 | **MCP** | `venice-video-mcp` (on npm) shells out to this CLI | 7 action-discriminated tools, structured JSON responses, progress notifications, plus 4 companion skills carrying the pipeline order | Any agent that supports MCP — Hermes, OpenClaw, Cursor, Claude — with no clone required |
 | **Bare global CLI** | `npm install -g`, shell tool, `--help` | The compiled CLI, this README, `AGENTS.md`, `.claude/skills/`, and the self-describing commands below (`agent-guide`, `pipeline`) | When your runner has a shell but no MCP — start with `venice-video agent-guide` |
 
@@ -420,8 +420,8 @@ everything else falls back to a default rather than prompting.
 
 ### The rules that most affect output quality
 
-Full text lives in `AGENTS.md` > "Agent Rules" (47 rules) and "Learned
-Anti-Patterns" (20 entries). If you can only carry a few, carry these:
+Full text lives in `AGENTS.md` > "Agent Rules" (49 rules) and "Learned
+Anti-Patterns" (28 entries). If you can only carry a few, carry these:
 
 1. **Direct the scene, don't decorate it.** Name one intention for the beat and
    derive camera, light, blocking, performance, and sound from it. Stacking
@@ -443,16 +443,23 @@ Anti-Patterns" (20 entries). If you can only carry a few, carry these:
 7. **Re-anchor every separately-rendered shot to the same locked references** and
    restate the character's invariant traits — including relative size — in every
    prompt.
-8. **Pass `aspectRatio` explicitly** on reference-to-video generation.
-9. **Never multi-edit close-up face shots on 16:9 panels.** The square-to-16:9
-   crop removes roughly 25% top and bottom, losing foreheads and chins.
-10. **Archive prior renders; never delete generated shot assets.**
-11. **Validate model capabilities before sending** `elements`,
+8. **State placement explicitly — spatial consistency is authored, not
+   inferred.** Lock each location's landmark geography in
+   `Location.spatialAnchors` and give every character shot a `blocking` field:
+   each subject's position relative to the named anchors, screen side, depth,
+   and facing/eyeline. Keep screen sides and eyelines constant across a scene's
+   shots unless a movement is scripted (180-degree rule). The harness injects
+   both verbatim into panel, blocking-plate, and video prompts.
+9. **Pass `aspectRatio` explicitly** on reference-to-video generation.
+10. **Never multi-edit close-up face shots on 16:9 panels.** The square-to-16:9
+    crop removes roughly 25% top and bottom, losing foreheads and chins.
+11. **Archive prior renders; never delete generated shot assets.**
+12. **Validate model capabilities before sending** `elements`,
     `reference_image_urls`, `scene_image_urls`, `end_image_url`, or `audio_url`.
     The registry is `src/venice/models.ts` in a clone; from a global install use
     `.claude/skills/venice-video-model-routing/SKILL.md` or the model tables
     below.
-12. **Ask before burning in subtitles**, and derive caption timings from
+13. **Ask before burning in subtitles**, and derive caption timings from
     `ffmpeg silencedetect` on the rendered voiceover rather than estimating them.
 
 ### Checkpoints where you should stop and ask
@@ -463,11 +470,11 @@ with TTS, burning in subtitles, upscaling to a 4K master, and any run whose
 quote you have not shown the operator. Post a short summary of what you are
 about to do and wait.
 
-## Supported Venice Models (April 2026)
+## Supported Venice Models
 
 ### Video Models
 
-Live catalog as of **2026-05-20** (synced against `GET /api/v1/models?type=video` — 103 entries). Families the harness routes to today; private / `-video-to-video` / `-extend-video` variants exist in the live catalog but aren't surfaced here.
+Live catalog (synced against `GET /api/v1/models?type=video` — 103 entries). Families the harness routes to today; private / `-video-to-video` / `-extend-video` variants exist in the live catalog but aren't surfaced here.
 
 | Family | Image-to-Video | Text-to-Video | Max Duration | Audio | Special Features |
 |--------|---------------|---------------|-------------|-------|-----------------|
@@ -501,7 +508,7 @@ Live catalog as of **2026-05-20** (synced against `GET /api/v1/models?type=video
 
 > **Seedance face rule (removed 2026-07):** Seedance 2.0 used to reject face-bearing input images that weren't produced by `seedream-v5-lite` / `seedream-v5-lite-edit`. Venice removed that restriction — any image family now works for face-bearing inputs, so the harness uses `nano-banana-2` for all panels. See [Image / Video Family Pairing](#image--video-family-pairing) below.
 
-### Image Models (28 entries, 2026-05-20 sync)
+### Image Models (28 entries)
 
 `nano-banana-pro`, `nano-banana-2`, `gpt-image-2` (high-quality alternative to `nano-banana-pro`), `gpt-image-1-5`, `flux-2-pro`, `flux-2-max`, `grok-imagine-image`, `grok-imagine-image-quality`, `hunyuan-image-v3`, `imagineart-1.5-pro`, `qwen-image-2`, `qwen-image-2-pro`, `recraft-v4`, `recraft-v4-pro`, `seedream-v4`, `seedream-v5-lite`, `chroma`, `hidream`, `venice-sd35`, `lustify-sdxl`, `lustify-v7`, `lustify-v8`, `wai-Illustrious`, `z-image-turbo`, `ernie-image`, `ernie-image-turbo`, `wan-2-7-text-to-image`, `wan-2-7-pro-text-to-image`, `bria-bg-remover`
 
@@ -519,7 +526,7 @@ New since the last sync: `grok-imagine-image`, `grok-imagine-image-quality`, `lu
 - **SFX**: `elevenlabs-sound-effects-v2`, `mmaudio-v2-text-to-audio`
 - **TTS (ElevenLabs)**: `elevenlabs-tts-v3`, `elevenlabs-tts-multilingual-v2`
 
-### The intelligence model (2026-08-05 sync)
+### The intelligence model
 
 Three steps in the pipeline reason rather than render: the **workshop** develops
 the project, **workshop-script** writes the shot script, and **qa-storyboard**

package/dist/agent/guide.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"guide.d.ts","sourceRoot":"","sources":["../../src/agent/guide.ts"],"names":[],"mappings":"AAkBA,MAAM,WAAW,YAAY;IAC3B,KAAK,EAAE,MAAM,CAAC;IACd,MAAM,EAAE,MAAM,EAAE,CAAC;CAClB;AAED,eAAO,MAAM,WAAW,EAAE,SAAS,YAAY,EAoD9C,CAAC;AAEF,wBAAgB,WAAW,IAAI;IAAE,OAAO,EAAE,CAAC,CAAC;IAAC,QAAQ,EAAE,SAAS,YAAY,EAAE,CAAA;CAAE,CAE/E;AAED,wBAAgB,WAAW,IAAI,MAAM,CAWpC"}
+{"version":3,"file":"guide.d.ts","sourceRoot":"","sources":["../../src/agent/guide.ts"],"names":[],"mappings":"AAkBA,MAAM,WAAW,YAAY;IAC3B,KAAK,EAAE,MAAM,CAAC;IACd,MAAM,EAAE,MAAM,EAAE,CAAC;CAClB;AAED,eAAO,MAAM,WAAW,EAAE,SAAS,YAAY,EAqD9C,CAAC;AAEF,wBAAgB,WAAW,IAAI;IAAE,OAAO,EAAE,CAAC,CAAC;IAAC,QAAQ,EAAE,SAAS,YAAY,EAAE,CAAA;CAAE,CAE/E;AAED,wBAAgB,WAAW,IAAI,MAAM,CAWpC"}