@nodaro/shared@1.10.0

Minor Changes

• 1686b80: New LLM_TEXT_INPUT_MAX (100,000) — the input ceiling for LLM text-generation nodes (Generate Text / llm-chat, AI Writer, Generate Script). Their systemPrompt / userInput / prompt fields were capped at a flat 10,000 chars, which falsely blocked pasting a long document to summarize or rewrite; LLM contexts are far larger (Claude 200K / GPT 128K+ / Gemini 1M tokens), so the routes now accept up to 100K input chars (output stays bounded by each route's maxTokens).

• f23dd98: New structuredOutputMode: "responses-json-schema" — KIE's codex/v1/responses endpoint natively enforces text.format JSON schemas (live-verified 2026-07-14, text and vision inputs). Applied to gpt-5.4, gpt-5.5, and the GPT-5.6 family (gpt-5.6-luna / gpt-5.6-terra / gpt-5.6-sol), which therefore now appear in STRUCTURED_VISION_MODELS (guaranteed-structured vision models, e.g. the describe-to-picker model gate).

• c971db1: Video-analysis robustness: quality tiers + layered audio.

  • Quality tiers — new VIDEO_ANALYSIS_TIERS (fast/pro), resolveVideoAnalysisModel, VIDEO_ANALYSIS_TIER_LABELS/ORDER, and DEFAULT_VIDEO_ANALYSIS_* (default pro). Users select a tier; the underlying model is never surfaced. resolveVideoAnalysisModel accepts a tier or a raw model id (back-compat) and falls back to the default.
  • Layered audio (breaking shape change to WindowAnalysis/VideoAnalysisResult) — a scene's audio is now an ARRAY of concurrent layers (AudioLayer[]) instead of a single {mode,content,voice} object, so simultaneous music + speech + sfx are all captured; an empty array means silence and the silence mode is removed. All in-repo consumers are updated; external consumers of VideoAnalysisResult.audio must read it as an array.

Patch Changes

• 44cc24d: Seedance 2 continuation references are now 2 seconds — KIE rejects r2v reference videos shorter than 1.8s for the seedance-2 family ("video duration … must be greater than or equal to 1.8 … in r2v"), which made every 1-second continuation tail fail deterministically.

  • New exports: SEEDANCE_2_CONTINUATION_REF_SEC (2 — the reference length every Seedance-2 chaining feature cuts and bills) and SEEDANCE_2_R2V_MIN_REF_VIDEO_SEC (1.8 — the verified provider floor).
  • SEEDANCE_2_EXTEND_STITCH.referenceTailSeconds is now SEEDANCE_2_CONTINUATION_REF_SEC (was 1).

cli-v1.3.0

Full Changelog: https://github.com/nodaroai/app.nodaro.ai/compare/@nodaro/cli@1.3.0...cli-v1.3.0

@nodaro/shared@1.9.0

Minor Changes

• 774aa2d: Add reasoning-effort control and 6 new KIE LLM models (gpt-5.6-luna/terra/sol, gpt-5.5, claude-sonnet-5, claude-opus-4.8) end-to-end.

  • @nodaro/shared: new LLM_REASONING_EFFORTS (none/low/medium/high/xhigh/max) + LlmReasoningEffort type, EFFORT_TIER_BUMP set, and effectiveReasoningEffort() helper (clamps a requested effort down to the highest level the target model actually supports). LLM_MODELS gains 6 new entries plus per-model reasoningEfforts, supportsTemperature, and preferKie capability fields. buildLlmCreditIdentifier() / resolveLlmCreditId() take an optional reasoningEffort third argument — xhigh/max (after clamping) bill one credit tier up (economy→standard, standard→premium, premium stays premium); high is the Claude-family server default and never bumps.
  • @nodaro/sdk: prompt-helper wizard resources' CommonInput gains an optional reasoningEffort field, forwarded automatically by the existing request-builder spread.
  • @nodaro/cli: nodaro prompt wizard subcommands gain a --reasoning-effort <level> flag (model-dependent; accepts none|low|medium|high|xhigh|max).

  grok-4.5 was evaluated but deferred — its KIE chat endpoint is not yet live, so no registry entry, rate row, or docs were added for it in this release.

@nodaro/sdk@1.3.0

Minor Changes

• 774aa2d: Add reasoning-effort control and 6 new KIE LLM models (gpt-5.6-luna/terra/sol, gpt-5.5, claude-sonnet-5, claude-opus-4.8) end-to-end.

  • @nodaro/shared: new LLM_REASONING_EFFORTS (none/low/medium/high/xhigh/max) + LlmReasoningEffort type, EFFORT_TIER_BUMP set, and effectiveReasoningEffort() helper (clamps a requested effort down to the highest level the target model actually supports). LLM_MODELS gains 6 new entries plus per-model reasoningEfforts, supportsTemperature, and preferKie capability fields. buildLlmCreditIdentifier() / resolveLlmCreditId() take an optional reasoningEffort third argument — xhigh/max (after clamping) bill one credit tier up (economy→standard, standard→premium, premium stays premium); high is the Claude-family server default and never bumps.
  • @nodaro/sdk: prompt-helper wizard resources' CommonInput gains an optional reasoningEffort field, forwarded automatically by the existing request-builder spread.
  • @nodaro/cli: nodaro prompt wizard subcommands gain a --reasoning-effort <level> flag (model-dependent; accepts none|low|medium|high|xhigh|max).

  grok-4.5 was evaluated but deferred — its KIE chat endpoint is not yet live, so no registry entry, rate row, or docs were added for it in this release.

Patch Changes

• Updated dependencies [774aa2d]
  • @nodaro/shared@1.9.0

@nodaro/cli@1.3.0

Minor Changes

• 774aa2d: Add reasoning-effort control and 6 new KIE LLM models (gpt-5.6-luna/terra/sol, gpt-5.5, claude-sonnet-5, claude-opus-4.8) end-to-end.

  • @nodaro/shared: new LLM_REASONING_EFFORTS (none/low/medium/high/xhigh/max) + LlmReasoningEffort type, EFFORT_TIER_BUMP set, and effectiveReasoningEffort() helper (clamps a requested effort down to the highest level the target model actually supports). LLM_MODELS gains 6 new entries plus per-model reasoningEfforts, supportsTemperature, and preferKie capability fields. buildLlmCreditIdentifier() / resolveLlmCreditId() take an optional reasoningEffort third argument — xhigh/max (after clamping) bill one credit tier up (economy→standard, standard→premium, premium stays premium); high is the Claude-family server default and never bumps.
  • @nodaro/sdk: prompt-helper wizard resources' CommonInput gains an optional reasoningEffort field, forwarded automatically by the existing request-builder spread.
  • @nodaro/cli: nodaro prompt wizard subcommands gain a --reasoning-effort <level> flag (model-dependent; accepts none|low|medium|high|xhigh|max).

  grok-4.5 was evaluated but deferred — its KIE chat endpoint is not yet live, so no registry entry, rate row, or docs were added for it in this release.

Patch Changes

• Updated dependencies [774aa2d]
  • @nodaro/shared@1.9.0
  • @nodaro/sdk@1.3.0

@nodaro/shared@1.7.0

Minor Changes

• aac8660: HappyHorse 1.1: the happyhorse / happyhorse-i2v / happyhorse-ref2v ids now target KIE's happyhorse-1-1/* endpoints (1.0 was delisted; identical parameter surface, so existing workflows keep working). Catalog gains the model's full 9-ratio aspect set (adds 4:5, 5:4, 21:9, 9:21 for T2V/Ref2V) and per-second pricing tiers (<id>:<N>s:<720p|1080p>, N = 3–15) in DURATION_PRICED_PROVIDERS / VIDEO_DURATION_TIERS / RESOLUTION_DURATION_PRICING. Prompt-wizard capability blurbs updated accordingly.

@nodaro/shared@1.6.0

Minor Changes

• 4e9f1b2: Add boards to CHARACTER_ATTACH_COLUMNS (worker auto-attach of identity boards), the CHARACTER_PICKER_DISPLAY_ORDER display constant + characterBucketDisplayRank / sortCharacterEntriesForDisplay helpers (boards-first picker menus), and an optional display-only bucket field on ConnectedReference.

• 254e7ef: Add a ref-only reference role that injects only the bare reference pointer — reference image A on image nodes, @image_1 / @video_1 / @audio_1 on video nodes — with no the {label} from … phrase.

  • roleToPhrase("ref-only", binding) returns the bare binding; ref-only is now the first curated preset for wired-character / wired-location.
  • Plain image / video / audio references now default to ref-only (DEFAULT_LABEL_BY_SOURCE manual/wired-image → empty label). Character / location / object / animal asset defaults are unchanged.
  • Video/audio label-less body tokens resolve to the bare @kind_N (was the subject in @kind_N).

• 7ea3412: Variant + Role Separation for mention tokens: a non-mode 4th segment now parses as a per-mention role coexisting with the variant — @kira:1:walking:clothes attaches the walking image and injects "the clothes from …" (image and video resolvers; @lib:1:weather/rain:lighting for locations). Any role works, curated, custom, or ref-only. Every pre-existing token shape parses byte-identically; CharacterMentionTokenInfo gains an optional role field (mirroring the location parser's).

• da90853: Migrate to zod 4 (4.4.x). No API changes — schema exports and their parse
  behavior are unchanged. @nodaro/shared now declares zod: ^4.4.0;
  @nodaro/prompts's bundled zod moves 3.25 → 4.4 and its schema-builder
  types use zod-4 generics (z.ZodType<Output, Input>).

Patch Changes

• 269d1b6: Fix reference chips when an entity node feeds one generate node via both its identity handle and its plain image handle. New sourceRefKey() scopes an entity's image-handle ref to ${nodeId}::image so the identity ref and the plain-image ref no longer collide on the node-id-keyed assembly maps (which previously dropped one non-deterministically — a literal @name:N token + lost character, or the image missing from the picker).

@nodaro/sdk@1.2.0

Minor Changes

• de3f995: Voice Changer Pro: VoiceChangerProInput.orderedVoices now accepts null entries (keep-slots). A null at position i means "keep speaker i's original voice" instead of recasting it. Additive — existing non-null callers are unaffected.

@nodaro/prompts@1.1.1

Patch Changes

• aac8660: HappyHorse 1.1: the happyhorse / happyhorse-i2v / happyhorse-ref2v ids now target KIE's happyhorse-1-1/* endpoints (1.0 was delisted; identical parameter surface, so existing workflows keep working). Catalog gains the model's full 9-ratio aspect set (adds 4:5, 5:4, 21:9, 9:21 for T2V/Ref2V) and per-second pricing tiers (<id>:<N>s:<720p|1080p>, N = 3–15) in DURATION_PRICED_PROVIDERS / VIDEO_DURATION_TIERS / RESOLUTION_DURATION_PRICING. Prompt-wizard capability blurbs updated accordingly.
• Updated dependencies [aac8660]
  • @nodaro/shared@1.7.0

@nodaro/prompts@1.1.0

Minor Changes

• 254e7ef: Add a ref-only reference role that injects only the bare reference pointer — reference image A on image nodes, @image_1 / @video_1 / @audio_1 on video nodes — with no the {label} from … phrase.

  • roleToPhrase("ref-only", binding) returns the bare binding; ref-only is now the first curated preset for wired-character / wired-location.
  • Plain image / video / audio references now default to ref-only (DEFAULT_LABEL_BY_SOURCE manual/wired-image → empty label). Character / location / object / animal asset defaults are unchanged.
  • Video/audio label-less body tokens resolve to the bare @kind_N (was the subject in @kind_N).

• 7ea3412: Variant + Role Separation for mention tokens: a non-mode 4th segment now parses as a per-mention role coexisting with the variant — @kira:1:walking:clothes attaches the walking image and injects "the clothes from …" (image and video resolvers; @lib:1:weather/rain:lighting for locations). Any role works, curated, custom, or ref-only. Every pre-existing token shape parses byte-identically; CharacterMentionTokenInfo gains an optional role field (mirroring the location parser's).

• da90853: Migrate to zod 4 (4.4.x). No API changes — schema exports and their parse
  behavior are unchanged. @nodaro/shared now declares zod: ^4.4.0;
  @nodaro/prompts's bundled zod moves 3.25 → 4.4 and its schema-builder
  types use zod-4 generics (z.ZodType<Output, Input>).

Patch Changes

• Updated dependencies [4e9f1b2]
• Updated dependencies [254e7ef]
• Updated dependencies [269d1b6]
• Updated dependencies [7ea3412]
• Updated dependencies [da90853]
  • @nodaro/shared@1.6.0