@sogni-ai/sogni-creative-agent-skill
Version:
Sogni Creative Agent Skill: agent skill and CLI for Sogni AI image, 3D, speech, video, and music generation.
1,401 lines • 107 kB
Markdown
## [3.66.1] - 2026-11-10
- Document the SAM 3.1 video selection options for SDK callers: a drawn region, 2 to 16 objects tracked together, the same selection on several frames, and how tracking carries between 32-frame chunks. They need the JavaScript or Python SDK 5.80.0 or later; the CLI has no flags for them.
- Use Sogni SDK 5.80.0 and intelligence-client 4.26.0.
## [3.66.0] - 2026-11-07
- Select an object across a video from a point, box or text description and download its animated mask with the source frame count and timing. Existing image selection stays available.
- Animate reference images with GPT Image 2.6 Sunburst or Flare: combine `-c` with `--output-format gif` and describe the motion. GIFs continue to exclude masks and annotations.
- Use Sogni SDK 5.79.0, intelligence-client 4.25.6 and the synchronized creative-agent 2.29.4 runtime.
## [3.65.1] - 2026-11-05
- Improve optional inline image previews while keeping full-resolution result links available.
- Refresh agent instructions and Claude plugin presentation, and clarify how network generation uses reference media.
## [3.65.0] - 2026-11-05
- Keep `test:all` offline and require an explicit environment opt-in for paid integration tests. Release checks never enable real GPU jobs automatically.
- Add explicit `--priority` for covered Fast renders on paid Unlimited, with 1.5× daily and monthly fair-use consumption. The default stays ordinary Fast.
- List account Personas with `--persona-list-account` and copy a selected one into the local library with `--persona-import-account`. Imports retain original photos, character sheets, audio voice recordings and descriptions; existing local Personas are preserved.
- Desktop tools now forward Fast/Relaxed choices, Priority and GPT Image 2.6 GIF, complexity and annotation controls. Model references and prompt guides match the current SDK chain.
## [3.64.0] - 2026-11-05
- Render on Relaxed from the CLI: `--network fast|relaxed` picks the Supernet for one run, and `defaultNetwork` now applies to renders. Before this, both only changed the connection and every job still ran on Fast. If you already set `defaultNetwork: "relaxed"`, renders now go to Relaxed: models with no Relaxed workers are refused with a hint rather than moved to Fast, and vendor models stay on Fast. Renders on Relaxed wait at least 30 minutes by default. Fast fair-use refusals (4087, 4089) now explain how to reach Relaxed, and hints for these refusals also show without `--json`.
- Add GPT Image 2.6 Sunburst and Flare as explicit choices, with animated GIF generation, planning complexity, and annotation-guided still-image edits. GIF settings and quoted costs stay together through submission.
* Seedance draft workflows require the user's choice after explaining their purpose and the draft, direct-final, and combined costs. After a failed or unusable 720p/1080p attempt, agents suggest testing the revision in a 480p draft, roughly 5.5 times cheaper than 1080p, and wait for consent before switching. Final-quality requests proceed to an unchanged approved draft's finalization or a revised final-resolution render without adding another unrequested draft.
## [3.63.0] - 2026-11-03
### Features
* `sogni-agent prompt-guide <model> [--mode <mode>] [--json]` prints the prompt rules Sogni's own prompt writers use for an image or video model and mode, so an agent can read them before it writes the prompt. It is free and offline (no credits, no API key), resolves the model the way `-m` does for a render, and with no model lists every family, mode and model id. The desktop extension adds a matching `prompt_guide` tool, and `generate_image` and `generate_video` ask for it before a prompt is written. The video and image prompting guides carry the same rules, generated from intelligence-client: each video family's main mode in full, with the command for every other mode.
* Seedance 2.5 draft mode: `--seedance-draft` renders a 480p draft at the 480p price, and `--seedance-finalize <job-id>` renders that same video at 1080p, with the draft's prompt, media, length, ratio, seed and audio. Iterate on drafts and pay for 1080p only on the take you keep. `--json` results carry the draft's `jobId` and `seedanceDraftExpiresAt` (about three hours), `--estimate-video-cost` quotes both, and a finalize refuses the inputs it inherits, in the CLI's own flag names.
* The video prompting guide now has sections for Seedance 2.x/2.5, WAN 2.2, Wan 3 and HappyHorse 1.1, written for an agent writing the prompt itself: the shot in one sentence first, one job per reference, timed beats, one primary camera move, pace shown through what rushes past, constraints stated as positive facts, cuts that follow the request, and the exact words for speech and on-screen text. SKILL.md, the plugin skills, `llm.txt` and the desktop extension point agents to the model's section before they write a prompt.
### Bug Fixes
* Camera wording keeps the request's pace: "zoom in" maps to a push-in or crash zoom, "pan" to a pan or whip pan, "orbit" to an orbit or fast arc, instead of always slowing the move.
* LTX text-to-video gives each shot one sentence and keeps every requested shot, each later sentence opening with its cut time; image-to-video and first/last-frame clips stay one continuous take.
* Every action the request names stays in the clip, on-screen text is kept exactly in quotes after "reads", and adult guides follow the requested intensity.
* A storyboard upload keeps your own words after the storyboard instructions instead of replacing them with a fixed prompt.
* Use intelligence-client 4.23.0, the creative-agent 2.27.9 runtime and Sogni SDK 5.74.0.
## [3.62.0] - 2026-11-03
### Features
* Image edits route by what you ask for: Krea 2 Identity Edit for people, FLUX 3 Image Klein for type and graphic design, Dark Beast Krea 2 Identity Edit for NSFW, and GPT Image 2.5 Sunburst as the Premium Spark best option. Qwen Image Edit runs only when you name it.
### Bug Fixes
* Seedance 2.5 and Seedance 2.5 Uncensored edits keep a 30-second source video and 30 seconds of reference audio. A local source was trimmed to 15 seconds before upload, so a 30-second edit came back half as long. Seedance 2.0 keeps its 15-second limit.
* Long hosted chat calls stay connected past two minutes and deliver their whole reply, including long `--json` output through a pipe.
* API key hints link straight to the API Keys page.
* Use intelligence-client 4.22.9 and Sogni SDK 5.73.0 across the CLI and standalone skill runtime.
### Updates
* New production guidance: perform flat cloned dialogue with a video model and keep only the sound, re-prompt Dark Beast instead of cleaning its stills with another editor, check every join of a stitched video at full size, check that every clip stays physically plausible, and keep a recurring character's voice reference and wording identical in every clip.
## [3.61.1] - 2026-11-01
### Updates
* Agents get better every session. They keep a short local lessons file (`~/.config/sogni/lessons.md`), read it before planning or rendering, and add a dated line when you correct them or a render fails for a reason the guides don't cover. With your yes, and showing you the exact text first, they can share a lesson as a GitHub issue so it reaches every install. Nothing is posted automatically, and lessons never include keys, private prompts, names or media.
* New guidance from real Seedance 2.5 productions: check the first frames for a jump, keep anchor images in agreement about who stands where, spell spoken words the way they sound, use an edit as a whole replacement shot, and expect a large Premium Spark hold on reference-video and edit requests. When both Seedance models refuse a shot, Wan 3.0 Enhanced is the closest-looking alternative.
* Cutting takes from different models together: join them with `--concat-fps 24` so one 30 fps Wan clip doesn't convert the whole sequence, and prefer plain cuts to added whooshes.
* GPT Image 2.5 masked edits: composite the result back onto the original through the same mask.
## [3.61.0] - 2026-10-30
### Features
* Show the recorded explanation for a failed project in status and result output.
### Bug Fixes
* Include Premium Spark in balance JSON and text output while preserving total Spark, account and subscription details. Distinguish zero Premium from unavailable information.
* Use intelligence-client 4.22.4 and Sogni SDK 5.71.0 across the CLI and standalone skill runtime.
## [3.60.6] - 2026-10-30
### Bug Fixes
* Correct speech option forwarding for Qwen preset voices, voice cloning and voice design while preserving existing defaults.
## [3.60.5] - 2026-10-29
### Bug Fixes
* Express a referenced individual's recognizable identity within the requested visual language. Artistic transformations can change proportions, skin treatment, accessories, expression, pose and framing; bounded edits and explicit preservation requests remain respected.
* Share the corrected image guidance across the creative runtime, tool descriptions, prompt-authoring reference, persona recipes and Hermes bundle. Preserve custom wording and existing model defaults.
* Align the bundled runtime with intelligence-client 4.22.2, creative-agent 2.26.5 and the published Sogni SDK 5.70.0.
* Clarify source-shot reference-video editing: apply requested changes while preserving the requested source continuity and untouched content.
## [3.60.4] - 2026-10-26
### Updates
* When you want an interactive, explorable or game-like world, agents offer the open-source [Sogni Worlds Kit](https://github.com/Sogni-AI/sogni-worlds-kit). With your go-ahead they clone it, you save your Sogni API key with `node world setup` in your own terminal, and they follow the kit's own skill through places, crossings, narration, music, build and export, submitting to Sogni's public community gallery only if you choose. The skill, the Claude Code and Codex plugins, the Hermes bundle and the MCP extension (as server instructions) all carry it.
* Agents serve stills as WebP (a JPEG social card, PNG only for icons) and films through the capped H.264 streaming profile with a half-size phone copy, the standard the kit's build and export enforce.
## [3.60.3] - 2026-10-26
### Bug Fixes
* Keep the scoped Seedance 2.5 reference synchronized across the main skill and Hermes package, including account requirements and supported output resolutions.
* Make the HTTPS first-frame routing test independent of external website responses.
* Align the development and ClawHub runtime pins with the published Sogni SDK 5.69.1 dependency.
### Documentation
* Preserve the proposed plugin-submission plan as a draft in the source documentation.
## [3.60.2] - 2026-10-23
### Bug Fixes
* Default MiniMax H3 reference-to-video to Balanced (`minimax-h3-r2v-balanced`, or `minimax-h3-r2v-balanced-2stage` for 1080p and 2K). Standard, LightX2V Turbo and FastH3 Reference remain available by name.
* List Mystic X v5 on MiniMax H3 Reference to Video.
### Updates
* Use intelligence client 4.21.0 and the creative agent 2.24.0 runtime.
## [3.60.1] - 2026-10-22
### Bug Fixes
* Prefer FastH3 Omni for MiniMax reference-to-video, including the two-stage option for 1080p and 2K. Explicit Standard, Balanced and LightX2V Turbo selections remain available.
## [3.60.0] - 2026-10-22
### Features
* Create images and edit from up to ten references with Seedream 5.0 Pro Uncensored. Choose it by name; it shares the one-time Seedance likeness agreement.
* Make MiniMax H3 reference videos on the FastH3 tier with FastH3 Reference (`minimax-h3-fasth3-r2v-turbo`, or `minimax-h3-fasth3-r2v-turbo-2stage` for 1080p and 2K). It takes the same pictures, clips, sound, catalog LoRAs and Personal LoRAs as the other Reference to Video choices.
### Updates
* Use the released image and video tool contracts in intelligence client 4.19.0 and Sogni SDK 5.69.0.
## [3.59.0] - 2026-10-21
### Features
* Agents know that Seedance 2.5, Seedance 2.0 and Seedance 2.0 Mini take photos and clips of you, or of people who have given you their consent, after the one-time likeness agreement. When one of them turns down a real person, agents offer the agreement and try again on the same model. Unlimited members hear about MiniMax H3 first.
## [3.58.4] - 2026-10-20
### Bug Fixes
* Agents keep your Sogni API key out of the chat. When you have a terminal, they ask you to save the key there yourself; they never read or print it, and check it with `sogni-agent doctor` instead.
* Agents give each new render its own filename. Unless you ask, they never write over your source files or an earlier render, or remove your media, personas or memories.
* In Claude Code and Codex, agents run `sogni-agent doctor` only after an install or upgrade, or when a command fails, not before every render.
## [3.58.3] - 2026-10-20
### Bug Fixes
* MiniMax H3 video tools now describe Mystic X v5 beside v4 and the Action Speed Slider, whose negative strengths give slow motion and positive strengths faster action. The H3 LoRA reference lists both and corrects Mystic X v4's catalog default to 0.5. Update the intelligence client to 4.16.3 for this guidance.
## [3.58.2] - 2026-10-19
### Bug Fixes
* FLUX 3 Image Klein edits now take up to ten references and GPT Image edits up to 16. With 7 or more `-c` images the CLI failed with "contextImages can have at most 6 images", because the intelligence client applied one limit of 6 to every model; it now checks each model's own limit. Update the intelligence client to 4.16.2 for this fix.
## [3.58.1] - 2026-10-19
### Bug Fixes
* When a model needs its one-time account agreement, show a direct review link, the Sogni account to use, and instructions for returning to the original request after accepting. Update the intelligence client to 4.16.1 for these recovery details.
## [3.58.0] - 2026-10-18
### Added
* Generate uncensored images from text with MOSS-Image and MOSS-Image 8-Step, including compatible Personal LoRAs.
### Changed
* Use the published MOSS contracts in intelligence client 4.15.0 and SDK 5.65.0.
## [3.57.1] - 2026-10-17
### Bug Fixes
* FLUX 3 Image Klein now renders at about 2 megapixels when no size is given, the 2K that Black Forest Labs recommends for its best results: a text-to-image renders at 1440x1440 (was 1024x1024), and an edit with several references sizes from the first one at up to 2 megapixels (was 1), never enlarged. An edit of one photo still keeps that photo's own shape up to 4 megapixels, and a size you give is kept. This matches Sogni Web, chat and the hosted agent.
* Update the intelligence client to 4.14.1 (creative agent 2.15.1) and the Sogni SDK (`@sogni-ai/sogni-client`) to 5.64.0 with protocol 1.0.0-alpha.51, the versions Sogni Web, chat and the hosted API run. 3.57.0 listed SDK 5.64.0 but still installed 5.63.3; this release installs 5.64.0, which routes Klein with up to ten references at up to 4096 px a side.
## [3.57.0] - 2026-10-16
### Features
* Add FLUX 3 Image Klein (`-m flux3_image_klein_sd_int8_convrot`; hosted selector `flux-3-klein`), Black Forest Labs' fast open image model: text-to-image, image-to-image, and edits from up to ten references, at up to 4096 px a side. A text-to-image with no size given renders at Klein's native 1024x1024. An edit of one photo now renders at that photo's own shape up to 4 megapixels when no size is given (larger sizes are capped, since the server refuses them), and several references size from the first at up to 1 megapixel, matching Sogni Web and the hosted agent. Personal LoRAs imported for Klein work with `--lora personal-<id> --no-filter`.
* Update the intelligence client to 4.13.0 (creative agent 2.13.0, Sogni SDK 5.64.0, protocol 1.0.0-alpha.51) for the Klein selector, its ten-reference limit and its prompt guides.
## [3.56.2] - 2026-10-14
### Bug Fixes
* Update the Sogni SDK (`@sogni-ai/sogni-client`) to 5.63.3 and the intelligence client to 4.12.4 (creative agent 2.12.8). A FLAC, Ogg or WebM voice sample given to `--reference-audio-identity` or `--voice-reference` is now uploaded with the type its bytes say; before, it went up with no declared type, so Sogni could hand it to a worker that cannot decode that format and the job failed there. A sign-in refresh that answers without a usable token pair now signs the client out and reports it, instead of repeating the refresh on every request (and, in a long-running process, retrying it without end). Whether a prompt was expanded (`promptExpanded`) is kept with the project, so a project read back later still says so.
* Update sharp to 0.35.5, which carries the librsvg fix for CVE-2026-96889 (a high-severity flaw in SVG decoding).
## [3.56.1] - 2026-10-10
### Bug Fixes
* `--upscale-video` rejected every video with "The source video dimensions, frame count, or frame rate could not be read" whenever it could not start ffprobe, even when ffprobe worked in the user's shell. It now says so, as `MISSING_FFPROBE`: `ffprobe is required for --upscale-video, but "<path>" could not be started (ENOENT)`, with a hint to set `FFPROBE_PATH` to the full path from `command -v ffprobe`. When ffprobe runs but cannot read the file, the hint now shows ffprobe's own error, such as "Invalid data found when processing input", and an ffprobe crash names its signal.
* `FFPROBE_PATH` and `FFMPEG_PATH` values that start with `~/`, such as a quoted `"~/.local/bin/ffprobe"`, now expand to the home directory. Before, the literal path was run and failed as if the tool were missing.
## [3.56.0] - 2026-10-08
### Features
* Seedance 2.0 Mini Uncensored: `-m seedance2-mini-uncensored` (and `-m seedance2-mini-uncensored-t2v`, or names such as "Seedance 2.0 Mini Uncensored", "Seedance Mini Uncensored" and "Seedance Mini Spicy" in any word order) runs Sogni model `seedance-2-0-mini-uncensored`. It has every Seedance 2.0 Mini mode and limit (480p or 720p, 4 to 15 s) and never becomes `seedance-2-0-mini`; a name with "Mini" never becomes Seedance 2.5 Uncensored. Like Seedance 2.5 Uncensored, it comes at a premium rate above its standard version, shares the same one-time likeness and consent agreement (error 4103 until it is accepted in the Sogni app) and is refused with error 4104 where Sogni has not released it yet.
* Update the Sogni SDK (`@sogni-ai/sogni-client`) to 5.63.0 and the intelligence client to 4.12.2 (creative agent 2.12.2), which carry the new model.
### Bug Fixes
* A Seedance content refusal now shows Sogni's own message, including its suggestion of which model to try instead, rather than a generic safety-filter line.
* A Seedance refusal of a photo that may show a real person now points to the uncensored version of that Seedance model or to MiniMax H3 (MiniMax H3 alone where the uncensored models are not released), instead of offering to stylize the people in the photo or to use LTX 2.3.
* `--json` output for a project that needs the consent agreement now includes `consentRequired` (its key and version), for both uncensored Seedance models; it was dropped before.
## [3.55.0] - 2026-10-06
### Features
* Seedance 2.5 Uncensored: `-m seedance2-5-uncensored` (with `-t2v`, `-ia2v` and `-v2v` forms, or the names "Seedance 2.5 Uncensored", "Seedance Uncensored", "Seedance 2.5 Spicy" and "Seedance Spicy") runs Sogni model `seedance-2-5-uncensored`, which has every Seedance 2.5 mode and limit and costs about twice as much. It never becomes `seedance-2-5`, and plain `seedance2-5` never becomes it. Each account accepts a one-time likeness and consent agreement in the Sogni app first; until then jobs fail with error 4103, reported as a non-retryable `PERMISSION_REQUIRED` error. Where Sogni has not released the model yet, jobs and estimates fail with error 4104, reported as a non-retryable `MODEL_UNAVAILABLE` error that names the models to use instead.
* Update the Sogni SDK (`@sogni-ai/sogni-client`) to 5.61.0 and the intelligence client to 4.11.0, which carry the new model and both errors. The intelligence client now reports a failed project's reason as the server's message instead of "[object Object]".
### Bug Fixes
* An inline image, video or audio input larger than about 3 MB could fail with "Maximum call stack size exceeded" in a long-running process: the SDK checked the whole data URI with a regular expression that overflows once Node stops optimizing new ones. Update the Sogni SDK (`@sogni-ai/sogni-client`) from 5.60.7 to 5.60.8, which checks it without one; the same inputs are accepted and refused as before.
* Update the intelligence client to 4.10.2 (creative agent 2.10.0). A ChatGPT, OpenAI, GPT or GPT Image request that names no version now runs on GPT Image 2.5; GPT Image 2 runs only when it is named.
## [3.54.5] - 2026-10-04
### Bug Fixes
* A reconnect to Sogni that took longer than 15 seconds could end the CLI with an error (exit 1) while it was still waiting for a render: the SDK's keep-alive pinged the connection before it had opened, which throws. Update the Sogni SDK (`@sogni-ai/sogni-client`) from 5.58.0 to 5.60.7, which pings only an open connection. It also keeps waiting on a running project when a status lookup fails, instead of counting the failed lookup toward ending the project as timed out, and no longer reconnects about once a second to a server that keeps closing the connection before signing it in.
* Update the intelligence client to 4.9.3 and the bundled runtime to creative agent 2.9.9. A batch image prompt (`-n` above 1 without `{a|b}` variations) is now sent as written: words such as "grid", "versions" or "4 different poses" are no longer deleted from it, so a caption or a count the user asked for reaches the model. The `generate_image` guidance for hosted agents says the same.
## [3.54.4] - 2026-10-02
### Features
* `--music` without `-m` now runs MiniMax Music 3 (60 s by default, 10 to 300 s), like every other Sogni surface; ACE-Step 1.5 XL Turbo stays available by name. Without lyrics it sends an `[Intro]`..`[Outro]` section outline, because a bare Music 3 instrumental ends early. `--bpm`, `--keyscale`, `--timesig` and `--language` are written into the prompt, since Music 3 has no such controls, and a setting only ACE-Step has (`--music-shift`, `--composer-mode`, `--creativity`, `--duration` over 300, `--steps` under 10, `--sampler euler_ancestral`) runs on ACE-Step with a note, so existing commands keep working. The OpenClaw `defaultMusicModel`, the desktop extension and the loop maker follow.
* `--status <projectId> --wait` and `--result <projectId> --wait` wait for a detached project over the live connection and fetch it when it finishes, making no requests while it renders (bounded by `-t`, default 1800 s; on timeout nothing is cancelled and the command to wait again is printed). A project this install submitted is taken back through the same app slot, never one a live process still holds. `--detach` prints this `waitCommand`; never re-run `--status` in a loop, and check without waiting at most once a minute.
### Bug Fixes
* `--detach` lost every detached project: it printed the project id and exited the moment the request was handed to the connection, before Sogni received it, so the project never ran (nothing was charged) while the CLI reported it as submitted and still running, and `--status` / `--result` answered "Project not found". It now waits up to 30 s for Sogni's answer for each project; a refusal is reported as the error it is, and no answer as `PROJECT_SUBMISSION_UNCONFIRMED` with advice to check once before submitting again.
* A 429 from Sogni is reported as a retryable `RATE_LIMITED` error that carries the server's Retry-After (`metadata.retryAfterSeconds`) and says not to retry sooner and not to poll. Sogni rate-limits per IP, so an agent that keeps asking can lock everyone behind that IP out.
* Personal LoRA imports have no live status event: the docs now say to check `--get-personal-lora` at most once a minute.
* Update the intelligence client to 4.7.1, which stops connect retry storms (capped backoff, stops on authentication and rate-limit refusals, disposes failed clients) and never resubmits a project Sogni already accepted.
* The speech reference and interactive-worlds guide say to design a recurring narrator voice once and clone that take for every later line, because voice design invents a new voice on every call.
* Local Seedance, Wan 3 and HappyHorse reference images, audio and video now upload with a presigned PUT (`/v1/image/uploadUrl` and `/v1/media/uploadUrl`), the way the SDK and `--api-chat` references already do, instead of the `/v2` form POST. Sogni's temporary generation media is moving to Cloudflare R2, which has no form POST, so the old path kept those uploads on the previous store. The results are the same: the reference reaches the model as a Sogni-hosted link that lasts 48 hours.
* The CLI now checks the limits the form policy used to enforce before it uploads anything: 100 MiB per file and a type Sogni accepts (PNG, JPEG, WebP or GIF images; M4A, MP3, FLAC or WAV audio; MP4, QuickTime or WebM video). A file outside them fails with `MEDIA_REFERENCE_TOO_LARGE` or `UNSUPPORTED_MEDIA_TYPE` instead of an error from the server.
## [3.54.3] - 2026-09-27
### Bug Fixes
* MiniMax H3 keyframes are now named in the prompt, the way MiniMax's own keyframe workflow does it. Each keyframe is `<Picture N>`, numbered after the first frame, last frame or references; it is listed with its time in the prompt's alignment line and tied to its shot with "the shot's keyframe corresponds to `<Picture N>`". On Reference to Video it also gets a keyframe entry, `keyframe completion` in the summary and a retention line. The guidance used to say keyframes are never named. The video prompting guide, the models reference and the hosted API notes follow.
* Update the intelligence client to 4.6.3 and the bundled runtime to creative agent 2.7.3, which write and check keyframe prompts in that format and state Sound to Video's real length (6 seconds renders 158 frames).
## [3.54.2] - 2026-09-27
### Bug Fixes
* Wan 3 and Wan 3.0 Enhanced can render 1080p from the CLI. Every Wan 3 size was cut to a 1536 px long side, so `--target-resolution 1080` (1920x1080) went out as 1536x864, which Sogni prices and renders as 720p. `-w 1080 -h 1920` did the same. Billing matched the 720p that was delivered, but there was no way to get 1080p. The CLI now sends Alibaba's exact size for the tier and ratio. At 1080 that is 1920x1080, 1440x1080, 1080x1080, 1080x1440 or 1080x1920; the ratio comes from `--wan3-ratio`, a ratio named in the prompt, or 16:9. Image-to-video keeps the first frame's shape with its short side at the tier (an 832x1216 frame at 1080 renders at 1080x1578), including a first frame given as a URL. A fixed `--wan3-ratio` that does not match the frame prints a notice instead of cropping it. Explicit `-w`/`-h` are sent as given, up to 1920 per side.
* The Wan 3 default stays 720p, the size and price CLI renders already had. 1080p is `--target-resolution 1080` and costs twice as much per second (52 instead of 26 Spark per second on Wan 3, 64 instead of 32 on Enhanced, before subscription discounts). `--estimate-video-cost` quotes the size that will be sent.
* Update the intelligence client to 4.6.2, which lets Wan 3 keep sizes up to 1920 instead of clamping them to 1536.
## [3.54.1] - 2026-09-27
### Bug Fixes
* `--upscale` now renders RTX VSR output above 2048 px, up to the 15,360 px long edge the help promises. Every image size above 2048 px was refused with "Width must be between 256 and 2048" before it reached Sogni, so GPT Image up to 3840 px and Krea 2 and Qwen up to 2560 px work again too.
* A result the Sensitive Content Filter withheld is reported as withheld instead of as "Result download failed: HTTP 404". When nothing was delivered the CLI stops with a `SAFETY_REJECTED` error that suggests rewording the prompt or `--no-filter`; a partly withheld batch saves the rest, marks each withheld result with `urlUnavailable: "sensitiveContent"` in `--json` (as `--result` does) and prints how many were withheld. Multi-angle edits, 360 segments and the looping return clip report it the same way.
* Update the intelligence client to 4.6.1 and the bundled runtime to creative agent 2.7.1, which carry the image size fix.
## [3.54.0] - 2026-09-27
### Features
* Pin MiniMax H3 keyframes: repeat `--keyframe <image>@<seconds>` (a local path or URL, up to 8) to make one H3 clip land on extra stills at chosen moments, in addition to `--ref` and `--ref-end`. It works on the H3 image-to-video and first/last-frame selectors (every tier, including FastH3 and the Two-Stage forms), the FastH3 Sound to Video selectors (`-ia2v-turbo`, `-flfa2v-turbo`, `-a2v-turbo` and their `-2stage` forms) and Ref2VA (`minimax-h3-r2v` and its Balanced, Turbo and Two-Stage forms). Text-to-video and every other model family refuse it and name the selectors that take it. Each time lands on frame `round(seconds × 24)`, which must be strictly inside the clip and on a frame of its own; anything else is refused with the range to use, and nothing is clamped or dropped. Each still is centre-cropped onto the canvas the job renders, like the last frame (at twice the canvas on Two-Stage models). Two keyframes are included in the price; each additional one adds 0.75 s of output time on FastH3 and 0.3 s on the other tiers, so an 8 s FastH3 clip costs 32 Spark with up to two keyframes and 50 with eight. `--estimate-video-cost` and the balance check include them, `--json` lists them, and a refusal because no online worker can pin keyframes yet (error 4100) is reported as a retryable `MODEL_UNAVAILABLE` error.
* `--keyframe` also works with `--api-chat` and `--durable-chat`: the images are uploaded after any frame images and the hosted agent is asked to pass them as `keyframes: [{ imageIndex, atSeconds }]`. `--api-workflow` refuses the flag and points at the step arguments instead.
* The desktop extension's `generate_video` MCP tool takes an optional `keyframes` array (up to 8 items of `{ image, at_seconds }`, where `image` is a path or URL, like `ref` and `ref_end`) and passes each item as `--keyframe`.
* Agent guidance for keyframes: `references/video-prompting.md` gains an Intermediate keyframes section (timing on the 24 fps grid, saying in the prompt what each keyframe shows at its time, starting a new shot at a keyframe that changes the angle, place or light, and a worked first/last-frame example); `references/models.md` lists which selectors take keyframes, the framing and the price; `references/hosted-api.md` covers the hosted `keyframes` argument. SKILL.md, README, llm.txt, the Claude plugin and Hermes skills and the video skill views point at the flag.
### Bug Fixes
* Keep Wan 2.2 sizes within 1,048,576 pixels per frame (1024x1024), with each side between 480 and 1536. Sizes the CLI chooses are fitted inside that limit (a 2314x1200 image-to-video reference now renders at 1408x736 instead of 1536x800), and source reels fit their Wan 2.2 sizes inside it too. An explicit Wan 2.2 `-w`/`-h` or exact pixel size in the prompt beyond the limit is refused with `INVALID_VIDEO_SIZE` and a fitted-size hint instead of being changed, and `--target-resolution` or `--reel-target-resolution` above 1024 is refused.
* Update the generation clients to SDK 5.58.0 and intelligence client 4.6.0, and the bundled runtime to creative agent 2.7.0, which carry MiniMax H3 keyframes. The CLI's keyframe checks and the keyframe framing now come from the intelligence client.
## [3.53.0] - 2026-09-25
### Features
* **A timeout no longer cancels the project.** When `-t` runs out, the project keeps rendering on Sogni and the run exits with `errorCode: "PROJECT_TIMEOUT_STILL_RUNNING"`. `details.resultCommands` holds the `sogni-agent --result <id>` commands that fetch it later, and `details.waitingReason` says why it was still waiting. `--cancel-on-timeout` restores the old behavior. The clock now starts at submission (uploads excluded). Time the account's own plan limit holds a project is not counted, because that wait ends by itself.
* Follow a project by its id. Every run prints each project id as soon as it is submitted. `--detach` (`--no-wait`) submits and returns with the id. `--status <id>` shows the project's state and, while it is queued, why. `--result <id>` fetches its finished media (`-o` saves it, with `-2`, `-3` suffixes for several renders). `--recent [hours]` lists this account's completed projects, newest first (default 24 hours, max 168). Projects that finished while no agent was connected are included; the server holds a result for a disconnected client for one hour only, but `--recent` and `--result` read the account's history. `--detach` refuses flows that submit a second project after the first (`--looping`, `--multi-angle`, source reels, hosted workflows and `--api-chat`).
* Say why a project is queued, in the same words as Sogni Web. A plan limit reads "you've reached your Unlimited plan limit for simultaneous MiniMax H3 videos": the account's own limit, which starts the project when one of its running jobs finishes. It is not a shortage of workers, and resubmitting only lengthens the line. Confirming payment and waiting for an available worker are reported separately.
* Before a video is submitted, note how many of the account's other video projects are already queued or rendering from other runs, apps or devices, since those may run first.
* Agent guidance in SKILL.md and the new `references/long-jobs-and-queues.md`: check `--recent` at the start of a session and offer unseen results; never resubmit or cancel a queued project; a batch of standard MiniMax H3 videos on Unlimited runs one after another, so estimate the wait for the whole batch; on hosts that cap a single tool call, submit long videos with `--detach` and collect them with `--result`.
### Bug Fixes
* Document the actual default video timeout, 1800 seconds (the skill said 300), and the 600-second music default.
* Update the generation clients to SDK 5.57.0 and intelligence client 4.4.2. The SDK now chooses a result's download endpoint from the model's media kind, the result's content type and the project's type instead of assuming an image, and recognizes Wan 2.2 sound-to-video and Qwen3-TTS results as media. `--status`, `--result` and `--recent` need SDK 5.57.0; an older installation says so and asks for a skill update.
## [3.52.4] - 2026-09-23
### Features
* Pass `--lora` on the MiniMax H3 FastH3 audio-to-video modes (`-ia2v-turbo`, `-flfa2v-turbo`, `-a2v-turbo` and their `-2stage` forms). Catalog and Personal H3 LoRAs resolve against the selected audio mode exactly as on the frame modes, on the released generation clients.
## [3.52.3] - 2026-09-23
### Bug Fixes
* Update the generation clients to preserve authoritative queue reasons across reconnects, partial batches, and recovered projects.
## [3.52.2] - 2026-09-23
### Bug Fixes
* Keep pending generation uploads, reconnect recovery, and chat requests attached to the account that started them. Routine credential refresh continues without interrupting work.
* Support WebP worker image exports and `--no-prompt-metadata`. Preserve non-blocking content labels on individual JSON results.
## [3.52.1] - 2026-09-19
### Bug Fixes
* Keep an aspect ratio stated in the prompt (for example "a 9:16 portrait shot") when `--target-resolution` sets the size. It used to render landscape at that size on every video model, including the MiniMax H3 Two-Stage selectors at 720 and 1080, and square on Seedance 2.5; it now renders portrait at the requested size (FastH3 Two-Stage 1080: 544x960).
## [3.52.0] - 2026-09-19
### Features
* Add MiniMax H3 Two-Stage Reference to Video: `-m minimax-h3-r2v-2stage` (Standard, 20 steps) and `-m minimax-h3-r2v-balanced-2stage` (Balanced, 8 steps) take the same loose `--ref`/`-c` images and repeatable `--ref-video`/`--ref-audio` references as `minimax-h3-r2v` and `minimax-h3-r2v-balanced`, and deliver twice the canvas with the same length and audio. `--target-resolution` picks the delivered size like the FastH3 Two-Stage selectors: 720 (672x384 canvas), 1080 (960x544) or 2K (1344x768 canvas delivering 2688x1536, the default). 1080p and 2K render at the 768p canvas and FastH3 refines the enlargement. `--estimate-video-cost` quotes the `_2stage` id on that canvas, and `--json` reports the delivered size.
### Bug Fixes
* Update the bundled runtime and exact dependencies to SDK 5.53.1, intelligence client 4.3.2 and creative agent 2.4.11. Generations whose request was sent on a connection that dropped before Sogni received it are now resent once after the reconnect instead of failing with "The server has no record of this generation", and hosted agents are told when to use the Reference Two-Stage selectors.
## [3.51.1] - 2026-09-16
Includes the integration prepared as 3.51.0 (unpublished) and the shared Ref2VA prompt repair.
### Features
* Manage the same Personal LoRA library as Sogni Web: list imports and supported models, import with confirmed rights, inspect asynchronous status, remove owned entries, and include ready personal adapters in model-filtered discovery. Generations preserve ordered strengths and use catalog defaults when no strength was selected.
* Expose hosted background removal, object segmentation, and single/multi-view Pixal3D reconstruction alongside the existing direct CLI modes. Document speech cloning, output controls, and FastH3 Two-Stage selectors.
* Use GPT Image 2.5 Sunburst and Seedance 2.5 at 1080p for new storyboard workflows; explicit existing model selections remain supported.
### Bug Fixes
* Keep GLB outputs typed as models through durable workflow and chat records. Update the bundled runtime and exact dependencies to SDK 5.51.0, intelligence client 4.2.1, and creative agent 2.3.2.
## [3.50.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.49.1...v3.50.0) (2026-09-14)
### Features
* Add MiniMax H3 FastH3 audio-to-video, driven by your own voice or song: `-m minimax-h3-fasth3-ia2v-turbo` takes a first frame (`--ref`) and `--ref-audio`, `-m minimax-h3-fasth3-flfa2v-turbo` a first and last frame (`--ref`, `--ref-end`) and `--ref-audio`, and `-m minimax-h3-fasth3-a2v-turbo` the audio alone. The upload drives the picture from frame 0 and is the delivered soundtrack, trimmed to the clip: `--duration` or `--frames` set the length on the usual 124-362 frame grid at 24 fps, and `--audio-start` picks where the window begins. `-m minimax-h3-fasth3-turbo --ref-audio` picks the mode from the frames you pass, and the `-2stage` selectors (and `minimax-h3-fastvideo-int8_{ia2v,flfa2v,a2v}_turbo[_2stage]` ids) deliver at twice the canvas with `--target-resolution`, printing a server refusal word for word like the other two-stage models. The CLI refuses a missing or extra upload, LoRAs, `--audio-duration` and `--no-generate-audio` before submitting, and every other H3 mode except r2v refuses `--ref-audio` and names the audio selector to use. Prices are FastH3's per-second price for the matching frame mode.
* Add Pixal3D multi-view: `--image-to-3d front.png` with any of `--left-view`, `--back-view` and `--right-view` builds the GLB with `pixal3d_multiview_int8_i23d` at the single-view options and prices. Views are named by the subject's own sides: the left view shows the subject turned so its own left side faces the camera (facing screen-left), the right view its own right side (facing screen-right), the back view the subject from behind. Templates that label the subject's right side "left" build a model turned 180 degrees. The view flags are refused without `--image-to-3d`, with other modes, and with `-m pixal3d_int8_i23d`; `--json` reports them under `orbitViews`.
### Bug Fixes
* Use Sogni Client 5.50.0 and intelligence-client 4.1.0, which carry the FastH3 audio-to-video ids, the hosted `sound_to_video` audio-guide selectors and Pixal3D multi-view uploads, in both the npm package and the installed skill runtime (`skill-package.json`); the wrapper now sizes audio-guide frames on the H3 canvas itself, so the CLI's local frame-fitting workaround is gone.
## [3.49.1](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.49.0...v3.49.1) (2026-09-14)
### Bug Fixes
* Include video inputs in Seedance estimates and all reference-video counts and measured durations in Wan 3 and MiniMax H3 estimates. Wan 3 quotes use the selected source window.
* Use intelligence-client 4.0.1 and SDK 5.48.1 consistently in the CLI and installed skill runtime.
## [3.49.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.47.0...v3.49.0) (2026-09-14)
### Features
* Add promptless Pixal3D image-to-GLB generation with `--image-to-3d`, original-image uploads, mesh/texture controls, 1024 or 1536 shape resolution, binary downloads, and model result metadata.
* Add BiRefNet background removal with `--remove-background`: transparent PNG cutouts by default, or soft foreground masks with `--matte`, preserving source dimensions.
* Add direct MiniMax Music 3 through `--music -m music3`, using the shared SDK defaults and model-specific duration and control validation.
* Add Qwen3-TTS studio speech, voice cloning, and voice design through `--speech`, with exact scripts, preset voices, written voice directions, original reference recordings, and WAV/MP3/FLAC output. Audio batches save every result to a separate file.
### Bug Fixes
* Include 3D artifacts in live model discovery and add `--model-media model` filtering.
* Refresh model guidance and Codex, Claude, and Hermes entrypoints, including Seedance 2.5 1080p edit/extend support, current utility contracts, and current installation commands. Keep the Hermes interactive-worlds reference synchronized.
* Check result download status and GLB headers before saving successful output metadata.
## [3.48.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.47.0...v3.48.0) (2026-09-13)
### Features
* Add MiniMax H3 FastH3 Two-Stage for 720p, 1080p and 2K video: `-m minimax-h3-fasth3-turbo-2stage` picks text-to-video, image-to-video or first/last-frame from the frames you pass (like `minimax-h3-fasth3-turbo`, no R2V), and `minimax-h3-fasth3-t2v-turbo-2stage`, `-i2v-turbo-2stage`, `-flf2v-turbo-2stage` and the `minimax-h3-fastvideo-int8_*_turbo_2stage` ids name a mode directly. FastH3 renders a half-size canvas and the clip is delivered at twice it with the same length and audio. `--target-resolution` names the delivered size: `2K` or `1440` (the default) renders 1344x768 and delivers 2688x1536, `1080` renders 960x544 and delivers 1920x1088, and `720` renders 672x384 and delivers 1344x768, in the prompt's or the `--ref`/`--ref-end` image's aspect; any other value is refused, and `2K` is refused on other models. 1080p costs 10 Spark per second (FastH3 plus 6), 2K costs 16 (plus 12), and 720p costs the FastH3 rate of 4. The summary line and `--json` output (`deliveredWidth`, `deliveredHeight`) report the delivered size, and a server refusal of two-stage (such as its "available soon" notice) is printed word for word. The `minimax-h3-fastvideo-int8_*_turbo_2stage_720p` ids the Socket records 720p work under are recognized when named directly.
* Seedance 2.5 renders 1080p as well as 480p and 720p (`--target-resolution 1080`), exports MOV with `--output-format mov`, and returns the clip's final frame as an image with `--return-last-frame`, for the first frame of a following clip. With `-o` the frame is saved next to the video as `<name>-last-frame.<ext>`; `--json` reports `lastFrameUrls` and `lastFramePaths`. Both flags are refused on every other model.
* `--2k` and `--output-scale`, added in 3.47.0, are removed. Sogni now refuses the `outputScale` request option, so neither flag can render; the CLI fails on either one and points at `-m minimax-h3-fasth3-turbo-2stage` (with `--target-resolution 1080` or `720`).
### Bug Fixes
* Align the public runtime with Sogni Client 5.48.0 and Intelligence Client 4.0.0 (Sogni Protocol 1.0.0-alpha.40), moving the exact pin, `skill-package.json` and the sogni-client override together. Intelligence Client 4.0.0 carries the two-stage sizing helpers the CLI uses.
* The bundled creative-agent runtime is regenerated from 2.0.0: `generate_video` and `animate_photo` drop `outputScale` and offer the FastH3 Two-Stage selectors, whose `targetResolution` names the delivered class, 720, 1080 or 2K, at the GPU-time prices above.
## [3.47.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.46.0...v3.47.0) (2026-09-12)
### Features
* Add `--output-scale 2` (alias `--2k`) for MiniMax H3 2K delivery: the clip renders on the requested canvas and comes back at twice its width and height (1344x768 becomes 2688x1536) with the same length and audio, for +10 Spark per second at 544/768p-class sizes and +6 at 480p. Other models refuse the flag by name. While Sogni holds 2K behind its "available soon" message, the CLI reports that refusal as `MODEL_UNAVAILABLE` with a hint to retry at the standard size.
### Bug Fixes
* Align the public runtime with Sogni Client 5.46.0 and Intelligence Client 3.33.0, moving the exact pin, `skill-package.json` and the sogni-client override together. 5.46.0 stops the SDK cancelling a project after its render moves to another worker, which killed the live retry.
* The bundled creative-agent runtime is regenerated from 1.65.0; its public contracts are unchanged.
## [3.46.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.43.0...v3.46.0) (2026-09-11)
3.44.0 and 3.45.0 were prepared but never published to npm; this release carries their changes too, so 3.43.0 is the last version npm actually shipped.
### Features
* Add `--segment` with `--segment-point`, `--segment-exclude`, `--segment-box` and `--segment-text`, returning a SAM 3 mask for an original still. Text and point prompts never combine, point prompts take at most one box, and the limits are 32 points, 16 boxes and 240 text characters.
### Bug Fixes
* Align the public runtime with Intelligence Client 3.31.1 and Sogni Client 5.44.1, moving the exact pin, `skill-package.json` and the sogni-client override together.
* The bundled creative-agent runtime (1.63.1) adds `detailPreference`, `processingSpeed` and `seed` to `upscale_video`; none of them changes the price.
## [3.45.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.43.0...v3.45.0) (2026-09-11)
3.44.0 was prepared but never published to npm; this release carries its changes too.
### Features
* Add GPT Image 2.5 Sunburst and Flare (`-m gpt-image-2.5-sunburst`, `-m gpt-image-2.5-flare`) for text-to-image and edits with up to 16 ordered `--context` references.
* Add `--image-quality low|medium|high` for every GPT Image model, plus `xhigh|max` on 2.5. Auto quality is rejected: Sogni never lets the provider choose the quality.
* Add `--image-background` (transparent on 2.5, PNG or WebP), `--image-output-compression 0-100` (JPEG or WebP) and `--image-mask`, a PNG alpha mask of the first `--context` image for edits.
* GPT Image defaults to 1024x1024 and a 10-minute timeout, and rejects sizes outside the provider's bounds before submitting.
* The MiniMax H3 reference keeps one continuous shot as the default, with beats and multi-shot sequences for when the context calls for them.
### Bug Fixes
* Align the public runtime with Intelligence Client 3.30.1 and Sogni Client 5.42.0.
## [3.44.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.43.0...v3.44.0) (2026-09-11)
### Features
* Add the beats-and-voices lessons to the MiniMax H3 reference: pace dramatic requests by beats with a final shot that may cut to the closing frame's composition, keep continuous motion for continuous action, write a fall or a bite as the physical act, describe each voice with accent, pitch, timbre and speaking rate, place dialogue after the action that motivates it, and say whose hands are in a first-person frame.
## [3.43.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.42.0...v3.43.0) (2026-09-10)
### Features
* Teach MiniMax H3 prompt writing from MiniMax's official Base guide: `[Shot 1]` opens with the picture's own look rather than a genre label, every sentence depicts something seen or heard, exclusions are positive facts, clock times live only on cut markers, first-to-last-frame prompts describe the path between the pictures in one continuous shot, and each vocal line binds to its speaker. Adds a first-to-last-frame worked example.
* The bundled creative-agent runtime (1.60.0) applies the same rules and flags in-shot clock times when it writes MiniMax H3 prompts.
### Bug Fixes
* `--upscale-video` relays the server's own refusal when a source is too long, with a hint to trim or split the clip, for direct and wrapped SDK errors.
* Align the public runtime with Intelligence Client 3.29.2 and Sogni Client 5.40.0.
## [3.42.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.41.0...v3.42.0) (2026-09-10)
### Features
* Add a Goose host launcher and document skill and MCP extension installation.
* Publish `sogni-agent-mcp` as a standalone command that uses the CLI bundled in the same npm package, with a global-install fallback for desktop extension bundles.
### Bug Fixes
* Include the shared attribution and version modules in the standalone Claude Desktop bundle, with a startup check that runs outside the repository.
* Align the public runtime dependencies with Intelligence Client 3.29.1 and Sogni Client 5.39.1.
* Let the server decide FlashVSR's maximum clip length, preserving the exact source timing and relaying any length rejection to the user.
## [3.41.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.40.1...v3.41.0) (2026-09-10)
### Features
* Add `--upscale-video <path|url>` with `--upscale-resolution 1080|1440` for promptless FlashVSR v1.1 video upscaling. The CLI reads the source's exact size, frame count and frame rate with local ffprobe, checks the public limits (short edge up to 768px and area up to about 1344x768, 1-362 frames and about 15 seconds, 1-60 fps, SDR with square pixels, 100 MB), and saves the MP4 with the original audio. Prompts, more than one output, timing or size overrides, and other media are rejected.
* Route pure video upscales to `--upscale-video` or the hosted `upscale_video` tool rather than `video_to_video`, and carry the `upscale_video` contracts in the public runtime.
### Bug Fixes
* Align the public runtime dependencies with Intelligence Client 3.29.0 and Sogni Client 5.38.0.
## [3.40.1](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.40.0...v3.40.1) (2026-09-10)
### Bug Fixes
* Align the public runtime dependencies with Intelligence Client 3.28.4 and Sogni Client 5.36.3, including the fix that keeps storyboard aspect ratios separate from panel counts.
* Forward explicit content-filter preferences when starting durable workflows through either HTTP or SDK transport; resumed runs retain their original preference.
* Preserve structured SDK failures across project creation and job/project events, retaining actionable error messages and codes.
* Give hosted chat, durable chat, and storyboard planning a 4,096-token default output budget so larger scene descriptions and tool arguments have room to complete. Explicit `--max-tokens` values still take precedence.
* Allow two minutes for HTTP chat and storyboard-planning responses, preventing larger tool calls from hitting the former 30-second timeout. Explicit HTTP timeout overrides remain supported.
* Reconcile durable-chat completion snapshots with streamed deltas so complete drafts appear once, while retaining completion-only messages and later rounds.
* Reject unsupported durable-chat tool restrictions before starting a run; direct callers to `--api-chat` for restricted tools or planning without execution.
### Documentation
* Align the root, Codex/Claude plugin, and Hermes skill guidance with tested Chat workflows: requested stage boundaries, exact prompts and output counts, complete variation prompts, ordered references, model-specific durations, and continuing from completed stages.
* Clarify user-confirmed recovery after generation failures, document the speech tool in the hosted surface, and distinguish local skill updates from server-side planner and repair behavior.
* Keep trial-access guidance current: cancelling a trial prevents the first charge while preserving trial access until its original end date.
# [3.40.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.39.1...v3.40.0) (2026-09-09)
### Features
* **skill:** publish current generation guidance ([ff83933](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/ff83933ae87074c6b175358fcd974adb785da0c0))
## [3.39.1](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.39.0...v3.39.1) (2026-09-02)
### Bug Fixes
* **video:** replace retired MiniMax H3 recipe wording ([6c9b312](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/6c9b3121a14f8f16d79a497ab90017fe2161a015))
# [3.39.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.38.0...v3.39.0) (2026-09-02)
### Bug Fixes
* **deps:** consume intelligence-client 3.27.0 ([7199baf](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/7199baffa673ed3ba5fc34d2891a9f42533c1665))
### Features
* **video:** add MiniMax H3 FastH3 Turbo selectors ([b17a468](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/b17a4685afa2f7503dcf2de5ee3316f5e4b0af41))
# [3.38.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.37.0...v3.38.0) (2026-08-29)
### Bug Fixes
* **deps:** consume released Wan 3 client stack ([d807c85](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/d807c85494ccd8f4bc1356ac07b089466de900ec))
* **release:** restore published 3.37 baseline ([a25d76d](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/a25d76ded2d4e111d7a7ead3c1d7d573b5ee7550))
* **video:** align enhanced Wan3 CLI ([f3bec1f](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/f3bec1f9955a414782aff3f993ce6841e2809fad))
### Features
* **video:** add Wan 3 Enhanced CLI ([477c116](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/477c11668c2ad2881f9578635a5e07b2ae6ce783))
# [3.37.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.36.0...v3.37.0) (2026-08-27)
### Bug Fixes
* **deps:** align the Sogni runtime on intelligence client 3.25.1 and Sogni client 5.22.0 ([5729a10](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/5729a107070c992b63186c542d6cd10e7bf2afc1))
### Features
* **video:** add MiniMax H3 Balanced, a fast, high-quality eight-step tier for text-to-video, image-to-video, first/last-frame, and reference-to-video generation ([bf2cfaa](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/bf2cfaa4cac80a9687adb28af331c2cfbdd85c20))
# [3.36.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.35.4...v3.36.0) (2026-08-27)
### Bug Fixes
* **deps:** consume intelligence client 3.25.0 ([b495645](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/b4956451ec39cc44406d7115f91dc8156fe1a850))
* forward H3 video durations ([986f373](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/986f373d085400add22aceda2d4d2fb5cde48ea6))
### Features
* **video:** harden local reference workflows ([eb3c19c](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/eb3c19c66f215a8322a864e0d7175939851f8bda))
### Documentation
* **hosted-api:** document workflow LoRA steps ([5bdf28f](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/5bdf28f3ae7a5d788b3f0fb95379c2b3c73ffb5c))
* **loras:** document the MiniMax H3 video adapters ([92d2524](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/92d2524f52346bebde33fc288421a490109d49a1))
## [3.35.4](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.35.3...v3.35.4) (2026-08-26)
### Bug Fixes
* **deps:** consume storyboard compiler correction ([ed60ec8](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/ed60ec85fdfa11e8c2a10dd7bbd283ae9a1087cf))
## [3.35.3](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.35.2...v3.35.3) (2026-08-26)
### Bug Fixes
* **deps:** consume final Wan 3 SDK releases ([389a4be](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/389a4be4640e61cd8374acb3d0ccb715df2ced53))
## [3.35.2](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.35.1...v3.35.2) (2026-08-26)
### Bug Fixes
* **video:** align exact-model skill contracts ([ba3a071](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/ba3a071406d2000ad65ee4c1a5d1ce1b71eeff53))
* **video:** align public model-native contracts ([9d44147](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/9d44147dd9aef334c60690e3ad9eda26f40e2ed7))
* **video:** correct Wan 3 provider semantics ([a54fb89](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/a54fb891b2606e096215058e2ce98674c7b2cb37))
* **video:** merge exact-model public contracts ([c8381b3](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/c8381b3c99d91912412ef0d2e6fd017822b4d403))
## [3.35.1](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.35.0...v3.35.1) (2026-08-26)
### Bug Fixes
* **deps:** consume corrected Wan 3 video quotes ([d805281](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/d805281fb85183696c26abfecb6ef87f49a4674c))
* **deps:** consume R2V canvas runtime ([82b4436](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/82b4436e36318614aaeeb295f0a8026ddf6f0a37))
* **video:** protect loose references from canvas resizing ([b996acc](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/b996accfeddb88a03c2d2b644bc7add751e852c5))
# [3.35.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.34.0...v3.35.0) (2026-08-25)
### Bug Fixes
* **deps:** consume Wan3 intelligence ([3aaa8d0](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/3aaa8d03354de24139fc0a07d923617ae1f8b88a))
* **video:** complete Wan3 CLI surface ([7b27d0f](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/7b27d0f7f662cd63a65d451dda1b9eb0c747605a))
### Features
* **video:** expose Alibaba Wan 3 ([f337a91](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/f337a9163bbab728822982bf7df0459b7c677b0b))
* **video:** merge Wan3 integration ([77f07a4](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/77f07a452a1afe535f3e8227e10d6815d6b7ec23))
# [3.34.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.33.1...v3.34.0) (2026-08-25)
### Bug Fixes
* **deps:** align Sogni runtime dependencies ([64fb49d](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/64fb49d300dae84165698125e1439dc0f65b5bae))
* **deps:** sync published intelligence runtime ([66ce6ab](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/66ce6abeb735405ea883072774b8830398660c6d))
* **deps:** update intelligence client to 3.23.2 ([ee62730](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/ee62730fff940fde2c7f6dccc8669b02d3bb542a))
### Features
* **prompts:** add exact model authoring guidance ([701d5a8](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/701d5a81698558f90ab183ca3fd913bae6970fab))
* **video:** add Seedance 2.5 R2V task modes ([ab094ef](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/ab094ef2c18f2364ac354cc4905061f21980873b))
## [3.33.1](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.33.0...v3.33.1) (2026-08-25)
### Bug Fixes
* **video:** distinguish H3 tokenizer prompt markup ([f931855](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/f93185556e06307a353a8d0acb8bb985b4f0a560))
# [3.33.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.32.0...v3.33.0) (2026-08-24)
### Bug Fixes
* **deps:** align model-native prompt runtime packages ([a013381](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/a013381c62e87158070d0a238d4908ceabc60f05))
* **video:** keep H3 references on the valid grid ([cd5c0f1](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/cd5c0f187fd175291de8e2a2bee7559fdf35dd94))
* **video:** pretrim H3 R2V audio windows ([#31](https://github.com/Sogni-AI/sogni-creative-agent-skill/issues/31)) ([1447ef5](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/1447ef5f7314c5f11d9f2ae8648dddccfef3a90d))
* **video:** quote H3 R2V reference images ([d9f20e9](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/d9f20e958f1147fe3f58ca4093ad4c84aa9844c3))
### Features
* **prompts:** document model-native authoring ([9cab39c](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/9cab39cc05f62e1788f807e656fed99d75913977))
# [3.32.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.31.0...v3.32.0) (2026-08-19)
### Bug Fixes
* **deps:** align shared media releases ([e1a3bed](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/e1a3bed4ae688c7497c0690126e78102cd24d758))
* **deps:** bump intelligence-client to 3.19.0 to enable LoRA discovery ([25cd43e](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/25cd43ee298cbee00097f156b29a9a038125da07))
* **hermes:** ship the LoRA reference in the Hub bundle ([d615a59](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/d615a59063899d9d609947f044d44eba8d0ccf36))
### Features
* add Hermes Skills Hub distribution ([#30](https://github.com/Sogni-AI/sogni-creative-agent-skill/issues/30)) ([52329e8](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/52329e81f5f89660fe71bb88e5a8a13f24ef789a))
* **loras:** add LoRA catalog discovery to the CLI ([8c50259](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/8c502597145ec9a261f2e1a8d03ea2688df35505))
* **upscale:** support NVIDIA RTX VSR through 16K ([d6ae213](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/d6ae2138833dccb88b5eac53a4878e7f1117e9eb))
# [3.31.1](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.31.0...v3.31.1) (2026-08-14)
### Bug Fixes
* **deps:** align shared media releases ([e1a3bed](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/e1a3bed4ae688c7497c0690126e78102cd24d758))
# [3.31.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.30.0...v3.31.0) (2026-08-14)
### Features
* **cli:** add RTX VSR upscale surface and route pro tier to live models ([d6ab96a](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/d6ab96a957999f223b85efd07964d6143815b6ee))
# [3.30.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.29.0...v3.30.0) (2026-08-14)
### Bug Fixes
* **deps:** adopt intelligence-client 3.16.2 ([0413dcc](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/0413dcc51549d71585708a2f49dd485c1a080eef))
* **video:** adopt the client's model-aware dimension ceilings ([bcb1c65](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/bcb1c654532b4162060456ddf3e33b53f2ed65a9))
### Features
* **video:** expose MiniMax H3 R2V Turbo ([813aa66](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/813aa66e3ad612ab579ea9e06eadec7df21e8476))
# [3.29.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.28.0...v3.29.0) (2026-08-13)
### Bug Fixes
* **deps:** align shared agent runtime dependencies ([a1ba39c](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/a1ba39cd6e41f0669a2ac057cf97105532f2b8ad))
### Features
* **video:** add LTX 2.5 skill routing ([b7a77ae](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/b7a77ae48212e3f4c9ae187ca1feb9cb22f4390a))
# [3.28.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.27.0...v3.28.0) (2026-08-10)
### Bug Fixes
* **deps:** align bundled skill client version ([8eb0119](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/8eb0119725bd6e5417ea3877d20b849337dd8598))
* **deps:** bump sogni-intelligence-client to 3.14.3 ([0b418ba](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/0b418bad2b930f2757f09e9ae29c66446b0054b9))
* **docs:** document Seedance 2.5 models and alias coverage ([9a4994b](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/9a4994b2486f6dd926911cf3433c3ea15b2939c8))
### Features
* **video:** support MiniMax H3 Turbo sampler overrides ([5cf7e53](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/5cf7e53c5f1f886ddd9e7dee13d8d9da187a4d1f))
## [3.27.0] - 2026-08-07
### Added
- **MiniMax H3 Turbo now has explicit public CLI selectors for every supported workflow.** Text-to-video,
image-to-video, and first/last-frame generation use the official structured prompt contracts and fixed
four-step Turbo path; Ref2VA remains available through the standard H3 model.
### Changed
- **MiniMax H3 Ref2VA now accepts a reference video as its only visual input.** Image and video references can
independently satisfy the visual-input requirement, while audio-only requests fail before dispatch with an
actionable validation error. Public guidance also reflects the worker capacity required by each H3 mode.
- **Updated the Sogni dependency stack.** The skill now uses
`@sogni-ai/sogni-intelligence-client` `3.14.2`, `@sogni-ai/sogni-client` `5.7.0`, and the producer-pinned
`@sogni-ai/sogni-protocol` `1.0.0-alpha.17` for the current public video contracts and LTX 2.3 model registry.
- **Subscription, queue, and app-ID recovery guidance now matches the current public product contract.** The
CLI and agent-facing documentation direct users to live plan limits and actionable retry choices.
### Fixed
- **Video project failures preserve the actionable server message.** Generic project-creation wrappers no
longer hide the underlying user-facing error returned by the generation service.
## [3.26.1](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.26.0...v3.26.1) (2026-08-07)
### Bug Fixes
* **video:** report the MiniMax H3 duration actually delivered instead of a 5-15s range no H3 render can produce ([#28](https://github.com/Sogni-AI/sogni-creative-agent-skill/issues/28)) ([33f1f8b](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/33f1f8bf4997dd369339ea5f41d5fd45c2572332))
* **skill:** point plain SKILL.md installs at their host launcher so Hermes renders stop reporting agentFramework "unknown" ([#28](https://github.com/Sogni-AI/sogni-creative-agent-skill/issues/28)) ([33f1f8b](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/33f1f8bf4997dd369339ea5f41d5fd45c2572332))
### Documentation
* document the host launchers in README.md and llm.txt, and note MiniMax H3 duration snapping ([#28](https://github.com/Sogni-AI/sogni-creative-agent-skill/issues/28)) ([33f1f8b](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/33f1f8bf4997dd369339ea5f41d5fd45c2572332))
# [3.26.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.25.0...v3.26.0) (2026-08-04)
### Features
* **video:** support MiniMax H3 multimodal workflows ([d9f1737](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/d9f17378db498d73bf231b286affe34f26b97c57))
# [3.25.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.24.0...v3.25.0) (2026-08-04)
### Features
* **video:** document MiniMax H3 reference workflows ([22bece3](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/22bece3815ca9fb684e416a7ad2958ddc815f045))
# [3.24.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.23.0...v3.24.0) (2026-08-04)
### Features
* **video:** document MiniMax H3 generation ([67b0c5a](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/67b0c5ae94c7ee784a260d0174008f42d2283038))
# [3.23.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.22.0...v3.23.0) (2026-08-03)
### Bug Fixes
* **video:** raise the default generation timeout to 30 minutes and explicitly cancel every known Sogni project before reporting a timeout
* **video:** keep desktop and OpenClaw process safety limits behind the CLI timeout so network cancellation can finish first
### Features
* **models:** add MiniMax H3 text-to-video, image-to-video, and first/last-frame aliases with native frame-grid alignment
# [3.22.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.21.0...v3.22.0) (2026-07-30)
### Features
* **images:** route identity-sensitive edits of referenced people and characters
through Krea 2 Identity Edit by default while preserving explicit model choices
* **images:** add concise Krea edit prompting, ordered base/detail references,
and worker-owned execution defaults for stronger likeness preservation while
preserving deliberate direct-CLI control overrides
* **models:** add the explicit `ltx23-eros` selector for LTX-2.3 10Eros
image-to-video, with required filter acknowledgement and fixed model settings
* **video:** add start-frame-aware 10Eros cinematic prompting guidance for
coherent performance, camera motion, atmosphere, dialogue, and sound
# [3.21.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.20.0...v3.21.0) (2026-07-31)
### Features
* **telemetry:** identify Codex, Claude Code, Claude Desktop, Hermes Agent, and OpenClaw through fixed host markers and allowlisted MCP client metadata; attach privacy-safe framework, surface, workload, and semantic operation lineage to Sogni-owned requests without sending prompts, user paths, credentials, or headers to presigned uploads
### Dependencies
* use `@sogni-ai/sogni-intelligence-client` 3.10.0 and `@sogni-ai/sogni-client` 5.3.0 for the published attribution contract
# [3.20.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.19.0...v3.20.0) (2026-07-29)
### Bug Fixes
* **deps:** use `@sogni-ai/sogni-intelligence-client` 3.9.0
### Features
* **images:** support up to eight ordered Krea 2 Turbo LoRAs with independently
validated bipolar strengths, including negative inverse-effect values
# [3.19.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.18.1...v3.19.0) (2026-07-28)
### Features
* **cli:** lease per-process app IDs from a persistent slot pool ([#26](https://github.com/Sogni-AI/sogni-creative-agent-skill/issues/26)) ([ef0ae8b](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/ef0ae8b973e65f88a78b775fb1b791daca719d42))
# [3.18.1](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.18.0...v3.18.1) (2026-07-28)
### Bug Fixes
* **cli:** persist one recognizable `sogni-agent-<uuid>` application ID across SDK sessions to prevent repeated CLI runs from exhausting the per-address app-ID allowance ([#25](https://github.com/Sogni-AI/sogni-creative-agent-skill/issues/25))
# [3.18.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.17.2...v3.18.0) (2026-07-27)
### Features
* **video:** default first/last-frame pairs to LTX-2.3 morph ([#24](https://github.com/Sogni-AI/sogni-creative-agent-skill/issues/24)) ([430edf5](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/430edf58e1d1be1f757824e586a5ba7ef18764fe))
## [3.17.2](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.17.1...v3.17.2) (2026-07-26)
### Bug Fixes
* **billing:** clarify that `tokenType: "spark"` is a denomination rather than proof of a Spark debit;
Unlimited coverage is determined by `paymentModel: "subscription"` or explicit subscription-billing
success ([d836b49](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/d836b490bccda5c9e8166b2bb3d928708d2fbae4)).
* **video:** use an already-compatible local LTX/10Eros reference as the implicit i2v canvas, avoiding
misleading intermediate dimension adjustments and no-op resize messages
([d836b49](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/d836b490bccda5c9e8166b2bb3d928708d2fbae4)).
## [3.17.1](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.17.0...v3.17.1) (2026-07-26)
### Bug Fixes
* **scripts:** refuse runtime sync on intelligence-client version mismatch ([247da8c](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/247da8c9cb7b6f0575a25543847ca00f44929c0a))
* **video:** stop discarding i2v resolution on sparse aspect ratios ([23125f7](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/23125f780d394ec988012aabd4db7a9e49c63341))
## [3.17.0] - 2026-07-26
### Added
* **video:** add LTX-2.3 10Eros image-to-video support for private mature-theme creativity, including
DR34ML4Y v3 video LoRA support through `--lora dr34ml4y-v3`.
* **models:** move live model discovery to the public Sogni model catalog API, including API-supplied
catalog tags, network and media filtering, and five-minute ETag-revalidated caching without requiring
an API key.
## [3.16.1] - 2026-07-25
### Fixed
* **release:** stamp and verify the Codex plugin manifest alongside every other package and plugin manifest.
## [3.16.0] - 2026-07-25
### Added
* **cli:** add live Supernet media model discovery through `sogni-client`, including separator-insensitive
ID/name search, image/video/audio and Fast/Relaxed filters, JSON output, and official catalog tag search
with repeatable `--model-tag` filters for labels such as `spicy` and `uncensored`.
### Changed
* **image edits:** consume `@sogni-ai/sogni-intelligence-client` `3.8.0` with Sogni Client `5.1.0-alpha.22`,
removing the duplicate Krea identity-edit CLI fallback while preserving the two-reference limit and 10-step
guidance defaults.
# [3.15.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.14.1...v3.15.0) (2026-07-21)
### Bug Fixes
* **packaging:** include the hosted SDK client module required by durable chat
### Features
* support Krea identity-edit models
* add the verified loop-maker workflow
# [3.14.1](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.14.0...v3.14.1) (2026-07-13)
### Bug Fixes
* **release:** prevent mixed-version npm tarballs after semantic-release bumps package.json
# [3.14.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.13.0...v3.14.0) (2026-07-05)
### Features
* **desktop:** add import_media tool bridging chat attachments to local files ([a259bcf](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/a259bcf36d4a747bb2cbe9d71ff8ec9da9e6c1c9))
# [3.13.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.12.0...v3.13.0) (2026-07-05)
### Features
* **cli:** port SourceReel folder-to-video reels from recovered branch ([0e61830](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/0e6183090a0e1af698bce5de112246bcd44d5eac))
# [3.12.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.11.1...v3.12.0) (2026-07-05)
### Features
* **unlimited:** surface plan entitlement and billing mode across the CLI ([373490f](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/373490fc30da1edea84d7928691ab419b44f5bcf))
## [3.11.1](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.11.0...v3.11.1) (2026-07-05)
### Bug Fixes
* **desktop:** fit inline image previews under the 1MB host result cap ([4666922](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/46669226a8c6fb9a70ff808ef58c387ae62a4778))
# [3.11.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.10.0...v3.11.0) (2026-07-05)
### Bug Fixes
* **desktop:** bound inline image fetches with a 20s abort timeout ([f20f42b](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/f20f42ba6138eeb1f1c07ed8f02a5feb79c9f1fd))
### Features
* **desktop:** add inline-image collection module for MCP results ([422043f](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/422043f0cdef1bcd36f168fb95ab3e3bf6450962))
* **desktop:** attach inline image blocks to successful tool results ([617bd5b](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/617bd5b63f3f0437cd81df3f6b59da8a587dcf16))
# [3.10.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.9.0...v3.10.0) (2026-07-04)
### Bug Fixes
* **cli:** recognize semantic-release h1 changelog entries in whats-new ([067bb65](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/067bb651e1ab924fb32d1b15585c80deeeb79d84))
* **desktop:** keep head of large success output, tail of errors ([394e1e1](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/394e1e131f5f75c4f01205a37286609ed7a59c99))
### Features
* **desktop:** add MCP tool registry mapping tools to sogni-agent argv ([cb951a7](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/cb951a79fcc5ef859ef3135f46b8aab81945305d))
* **desktop:** add MCPB manifest, npm packaging, and build script ([d464627](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/d4646273bcb4c2c8cfd7d08e7b6928e218f15e6f))
* **desktop:** add path resolution for the Claude Desktop MCP wrapper ([8cd5832](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/8cd58323982d63a756e75264441d0dd33feb1acc))
* **desktop:** implement dependency-free MCP stdio server for Claude Desktop ([de26cda](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/de26cda9f433a8da62dcc24c7b2b49e6c13e892a))
# [3.9.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.8.0...v3.9.0) (2026-06-30)
### Features
* wire ltx23 lora video controls ([968aa9a](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/968aa9aca2bf5cfc30d7f73b513f4ff0afb9cf94))
# [3.7.0](https://github.com/Sogni-AI/sogni-creative-agent-skill/compare/v3.6.4...v3.7.0) (2026-06-25)
### Features
* **video:** document Seedance 4K skill support ([0f3c150](https://github.com/Sogni-AI/sogni-creative-agent-skill/commit/0f3c1509801fcdb15741e6adb94514f38fead0e5))
# Changelog
All notable changes to this project are documented in this file.
The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
## [Unreleased]
### Added
- **LTX-2.3 LoRA controls are now wired through the public CLI.** Direct `sogni-agent` video generation now
supports `--control-type outpaint` with `--outpaint-position` / `--outpaint-aspect-ratio`,
`--control-type inpaint --mask <image>`, and automatic `transition` LoRA attachment for LTX-2.3 i2v jobs
that provide both `--ref` and `--ref-end`.
### Changed
- **Creative upscale routing is clearer in the bundled intelligence client.** Updated
`@sogni-ai/sogni-intelligence-client` to `3.7.0` so uploaded-video Seedance upscale, enhance, remaster,
restyle, and transform requests route through `video_to_video` with `seedance-v2v`, while quality-only upscale
and sharpening requests continue to use the LTX-2.3 detailer path without restyling.
## [3.8.0] - 2026-06-30
### Changed
- **Music generation now defaults to ACE-Step 1.5 XL in the CLI.** The `turbo` and `sft` music aliases resolve to
the XL model IDs, while the legacy ACE-Step 1.5 model IDs remain available for explicit selection.
- **Model guidance now reflects the current Premium Spark vendor catalog.** README, LLM-facing docs, skill
routing notes, and CLI billing guidance now list HappyHorse 1.1 alongside GPT Image 2 and Seedance as
Premium-Spark-only vendor paths.
- **Krea 2 Turbo discovery is clearer.** CLI help and model references now list the direct worker model ID
`krea2_turbo_fp8_scaled` and note the hosted/chat selector `krea-2-turbo`.
- **Updated the bundled Sogni intelligence client to `3.5.1`.** The skill now consumes the current Sogni Client
stack and model metadata used by the refreshed video, image, and music catalog.
## [3.7.0] - 2026-06-25
### Changed
- **Seedance 2.0 Mini is now available through the bundled runtime.** Mini is the default lower-cost
720p Seedance path, while `seedance2-fast` remains available as an explicit legacy selection.
- Bumped `@sogni-ai/sogni-intelligence-client` to `3.3.0` so the skill consumes the published
Seedance Mini runtime, tool metadata, and Sogni Client `5.1.0-alpha.14` dependency stack.
### Changed
- **Seedance 2.0 4K routing is now documented and available through the current Sogni dependency stack.**
The full `seedance2` model can be selected for native 4K/2160p output with `--target-resolution 2160`;
`seedance2-fast` remains the lower-resolution fast path. The bundled runtime and
`@sogni-ai/sogni-intelligence-client` dependency now consume the published Seedance 4K tool metadata.
- **ACE-Step 1.5 XL is now the default music model.** `generate_music` / `--music` now default to
ACE-Step 1.5 XL Turbo (`ace_step_1.5_xl_turbo`), with ACE-Step 1.5 XL SFT (`ace_step_1.5_xl_sft`) as the
quality variant. The `--music-model` keys are unchanged — `turbo` (default) and `sft` — but they now map to the
XL ids; the legacy `ace_step_1.5_turbo` / `ace_step_1.5_sft` ids are no longer the default. The 8-step / `euler`
/ `simple` turbo defaults are unchanged. Updated `references/models.md`, `README.md`,
`references/openclaw-config.md`, and the `defaultMusicModel` default in `openclaw.plugin.json`.
### Added
- **LTX-2.3 video-to-video outpaint and inpaint control modes** documented for `ltx23-22b-fp8_v2v_distilled`.
Outpaint extends/expands the video canvas (e.g. vertical → widescreen) and is positional and mask-free —
anchored with a position (`center|top|bottom|left|right`) and an optional target aspect ratio
(`16:9|9:16|1:1|4:3|3:4|21:9`), growing the canvas without cropping. Inpaint regenerates a masked region and
requires a mask image (white pixels = region to regenerate). Added a V2V outpaint/inpaint subsection to
`references/video-editing.md` and capability prose to `references/models.md` and `README.md`.
- **LTX-2.3 two-keyframe transition / morph LoRA** documented for image-to-video. When the LTX-2.3 i2v model
`ltx23-22b-fp8_i2v_distilled` receives both a start frame (`--ref`) and an end frame (`--ref-end`), it
auto-applies the ValiantCat transition/morph LoRA (lora id `transition`, trigger word `zhuanchang`,
strength ~1.0) and morphs the first image into the last in a single render — distinct from the manual
bridge-clip "transition between two videos" recipe. Added notes to `references/video-editing.md`, `SKILL.md`,
and `README.md`.
## [3.6.4] - 2026-06-22
### Fixed
- **Agents no longer run `sogni-agent doctor` as a routine preflight before every generation.** `SKILL.md` now
scopes `doctor` explicitly to install/upgrade verification and failure troubleshooting, and tells agents to go
straight to the generate command (which validates credentials, ffmpeg, and balance itself and returns a fix hint
on failure). The previous wording — featuring `doctor` prominently in Setup, Install Policy, and Troubleshooting
without a "not a preflight" caveat — led agents to run it before each render. Inside sandboxed runtimes like Codex
this was especially costly: `doctor` makes a live network/auth call, which the sandbox blocks on the first
attempt, forcing a network-approval prompt and a second `doctor` run before any work happened. The
memory/personality preflight bullet also now notes it is memory + personality only — no `doctor` call.
## [3.6.3] - 2026-06-22
### Added
- **Sogni Unlimited subscription guidance.** README, SKILL.md, and `llm.txt` now document the Unlimited plan,
and the CLI includes friendlier billing-error fallbacks for subscription-related failures.
### Fixed
- **`sogni-agent self-update` now gives clearer recovery guidance after failed global package updates.** When the
package manager exits nonzero, the CLI reports the failing exit code and prints platform-specific permission
guidance (for example, `sudo sogni-agent self-update` on macOS/Linux or an Administrator terminal on Windows)
instead of leaving users with only raw npm output.
- **Codex and Hermes setup docs now mention the target app must be started once before targeted installer runs.**
`README.md` and `llm.txt` now match the setup installer preflight: `--only=codex` and `--only=hermes` exit
before installing anything if `~/.codex/` or `~/.hermes/` has not been created yet.
- **Claude plugin setup guidance no longer tells agents to create a duplicate personal Claude skill.** The
plugin-bundled skill now installs only the global CLI, tells agents not to run the default setup installer from
inside the plugin, and clarifies that setup cleanup does not uninstall the Claude Code plugin itself.
## [3.6.1] - 2026-06-15
### Changed
- **Hosted-API guidance now recommends client-side planning over hosted re-planning.** The skill is driven by a
frontier LLM that out-plans Sogni's hosted planning model, so steering it to delegate planning through
`--api-chat` was a downgrade. `SKILL.md`, `references/hosted-api.md`, and `README.md` now tell the calling agent
to plan and select tools itself, use `--api-workflow` with an explicit `--workflow-input` step graph for durable
multi-step work (the server executes the authored plan without re-planning), and reserve `--api-chat` /
`--durable-chat` for deliberately offloading a long server-side loop or uploading several local files in one
turn. `--api-chat` and all hosted modes remain fully supported — only the recommended default changed.
### Fixed
- **Local Seedance reference images via `-c`/`--context` now auto-upload in direct CLI mode.** Local
loose-reference images were rejected with an HTTPS-only error that pushed users onto the unreliable
`--api-chat` / `--durable-chat` path; local `--ref-audio` and `--ref-video` already auto-uploaded through the
`/v2` presigned-POST flow, so images were the only modality missing it and one broken branch cascaded into
downstream failures (vision 1024px cap, HTTP timeout, no-content, missing durable SDK package). Local
`-c`/`--context` images now upload through the same `/v2/image` presigned flow and forward as Sogni-hosted URLs.
MIME type is resolved by magic-byte sniffing (falling back to extension), and the accepted set
(PNG/JPEG/WebP/GIF) mirrors the backend's `allowedContentTypes`. Adds local-PNG-upload and mislabeled-WebP
byte-sniff regression tests; verified end-to-end with a real Seedance 2.0 render from a local `-c` PNG.
## [3.6.0] - 2026-06-12
### Added
- **Agents now surface update notices (gstack-style).** Update notices were previously suppressed exactly where
agents live — non-TTY stderr, `--json` mode, and OpenClaw plugin invocations — so Claude Code / Codex / Hermes /
OpenClaw users never learned a newer skill existed. Any command may now print a single advisory stderr line,
`[sogni-agent] Update available: <current> -> <latest> ...`, throttled to at most once per 24 hours, telling
the agent to finish the current task, relay the update to the user, and offer `sogni-agent self-update`
(`--snooze-update` on decline). Interactive TTY users keep the existing banner. stdout is never touched, so
`--json` output stays machine-parseable; SKILL.md instructs agents how to handle the line. Background version
checks now also run in agent contexts (still skipped for CI, tests, `--no-update-check`,
`SOGNI_NO_UPDATE_CHECK`, and dev checkouts).
## [3.5.1] - 2026-06-12
### Fixed
- **OpenClaw no longer marks the skill "missing".** The SKILL.md frontmatter listed every optional override
variable (`SOGNI_CREDENTIALS_PATH`, `OPENCLAW_CONFIG_PATH`, `FFMPEG_PATH`, …) and optional config path
(including `~/Downloads/sogni`) under `metadata.openclaw.requires`, so OpenClaw treated them all as hard
requirements that could never be satisfied and flagged the skill `✗ missing` on every install. Requirements
are now just `bins: node` + `anyBins: ffmpeg`; the API key still comes from the credentials file or
`primaryEnv`, and the override variables remain documented in "Filesystem Paths and Overrides". Verified
live: `openclaw skills check` flips from missing to `✓ ready`.
## [3.5.0] - 2026-06-11
### Added
- **`sogni-agent doctor` (also `--doctor`).** One deterministic install health check: Node floor, credentials
presence and file permissions, config-dir writability, ffmpeg discovery, a live auth + balance probe with a
timeout, and version freshness. `--json` emits a structured `checks` array; exit 1 when a required check fails.
Every install path in the docs now ends with `sogni-agent doctor` as the verification gate.
- **Upgrade UX.** `--whats-new [since-version]` prints the bundled CHANGELOG entries (CHANGELOG.md now ships in
the npm tarball and `self-update` points at it after upgrading), and `--snooze-update` pauses the update
reminder with escalating backoff (1 day → 2 days → 1 week, reset by a newer release) instead of re-nagging
every 24 hours.
- **SSRF-guarded downloads.** New `fetchSafeUrl` in `ssrf-guard.mjs` fetches with manual redirects and re-validates
every hop, so a vetted public media URL can no longer redirect a download to a private/metadata address. Remote
`--ref`/`--ref-audio`/`--ref-video` fetches use it.
- **CI.** GitHub Actions workflow running the unit suite on Node 22.11.0 and 24, verifying npm tarball contents,
and validating the plugin manifests. `npm run sync:version` stamps the package.json version into every manifest
(`version.mjs`, `SKILL.md`, `.claude-plugin/plugin.json`, `openclaw.plugin.json`), enforced by a new
docs-consistency test suite that also fails on any documented flag missing from the CLI parser.
### Changed
- **SKILL.md restructured for progressive disclosure.** The always-loaded core shrank from 1,338 lines (~10k
words) to ~300 lines (~2.6k words) — every routing rule (photobooth-vs-context-edit, LTX prompt rewrite,
high-res routing, PWD output convention, insufficient-funds script, media/shell security rules) stays inline,
while deep guides moved to `references/` (video-prompting, video-editing, hosted-api, models, personas-memory,
openclaw-config) read on demand. `references/` and `skills/` now ship in the npm tarball, the Claude plugin,
and the OpenClaw link surface. Verified with a 7-scenario agent battery against the new layout.
- **Install docs are now accurate per platform.** Added the missing OpenAI Codex CLI section
(`~/.codex/skills/`), a real Hermes Agent section (`~/.hermes/skills/media/` + `/reset`), the ChatGPT
Custom-GPT path the installer prints, an explicit note that the npx installer does not configure OpenClaw, and
"pick one registration" guidance for Claude Code (plugin or personal skill, not both). The OpenClaw install
command is now `openclaw plugins install npm:@sogni-ai/sogni-creative-agent-skill` — the bare unscoped name
never resolved the scoped npm package.
- **OpenClaw branding modernized.** Frontmatter metadata key `clawdbot:` → `openclaw:`; `--list-media` now
defaults to `~/.openclaw/media/inbound` with automatic fallback to the legacy `~/.clawdbot/media/inbound`;
the ClawHub install hook no longer overwrites `package.json` in a git checkout (guarded copy).
- **`--json` stdout is now strictly machine-parseable.** Durable-workflow SSE progress frames stream to stderr in
JSON mode; `--last --json` wraps the record in a `{ "success": true, ... }` envelope and exits 1 with
`errorCode: "NO_LAST_RENDER"` when nothing has been rendered (previously raw record / exit 0). Human-mode
errors now print the same classified, friendly message JSON consumers get.
- **Paid-batch safety cap.** `-n/--count` is capped at 16 outputs per invocation (a typo like `-n 1000` no longer
launches a thousand paid renders); raise deliberately with `SOGNI_MAX_COUNT`. OpenClaw `defaultCount` is
clamped the same way.
- `npm test` now runs the offline unit suites only and works without the private `sogni-creative-agent` sibling
(the runtime freshness check skips with a warning; publishing still hard-requires it via `prepack`).
Integration tests are strictly opt-in: `SOGNI_INTEGRATION=1` / `npm run test:integration` — a `SOGNI_API_KEY`
in the environment no longer causes plain `npm test` to submit real paid GPU jobs.
- `engines.node` raised to `>=22.11.0` to match the runtime guard (Node 22.0–22.10 previously passed `npm
install` and then hard-exited at first run).
### Fixed
- **Ctrl-C and temp-file hygiene.** The CLI now handles SIGINT/SIGTERM/SIGHUP (conventional exit codes) and
removes every temporary directory it created on exit — interrupting a long video job no longer orphans
directories under the OS temp dir, and the multi-angle / loop flows no longer leak a temp dir on every run.
- Credentials file values containing an inline ` #` comment now trigger a clear warning instead of silently
corrupting the API key into a confusing 401; prompts that begin with `-` get a hint about the standalone `--`
separator.
## [3.4.0] - 2026-05-30
### Added
- **Video finishing without raw ffmpeg.** `--concat-videos` now uses the concat *filter* (not the demuxer): it
probes each clip and normalizes fps/size/SAR/pixel-format and synthesizes silent audio for clips with no audio
track, fixing frozen video with continuing audio when clips differ in frame rate or stream layout. Adds
`--concat-fps`, `--extract-first-frame` (mirror of `--extract-last-frame`), and `--remix-audio` for looping a
bed (`--audio-loop`), fades (`--audio-fade-in/out`), and mixing one extra track (`--mix-audio/--mix-at/--mix-gain`)
without re-encoding video. External `--concat-audio` is now padded/trimmed to the video length.
### Changed
- Bumped `@sogni-ai/sogni-intelligence-client` to `^3.0.13` (pins `sogni-client 5.0.0-alpha.17`), keeping the
bundled creative-agent runtime in sync with the current shared prompt contracts and repair recipes.
### Fixed
- **Idiotproof first run.** Added a zero-dependency Node.js version guard that prints a clear "requires Node >= 22"
message before native modules load, `fetchWithTimeout` on every REST/download call (a black-holing proxy now
fails with `NETWORK_TIMEOUT` instead of hanging; override via `SOGNI_HTTP_TIMEOUT_MS`), `OUTPUT_WRITE_FAILED`
mapping for filesystem errors so a paid render isn't lost to a raw `EACCES`/`ENOSPC`, a friendly
`MEDIA_REFERENCE_NOT_FOUND` for missing `--ref`/`-c` files, and leading `~` expansion in file arguments.
- **Invalid/rejected API key no longer crashes.** A bad or expired `SOGNI_API_KEY` previously threw from a
detached promise during connect and dumped a raw stack trace. Added invalid-key detection plus global
`uncaughtException`/`unhandledRejection` handlers that route fatals through the clean `Error:`/`Hint:` path
(JSON-aware, exit 1) with a dashboard.sogni.ai hint.
- Routed source-preserving image edits away from photobooth and preserved structured project-result errors so
insufficient-funds responses consistently surface the Spark Packs guidance.
## [3.3.5] - 2026-05-29
### Changed
- Added an explicit **Output Path Convention** section to `SKILL.md` instructing agents to save generated images,
videos, and music to the user's current working directory (PWD) rather than `/tmp`. The directive includes a
short ✓/✗ example block so agents prefer `./cat.png` or a bare filename over an absolute `/tmp/…` path. Final
user-visible renders belong in the user's working directory; `/tmp` is reserved for transient intermediate
files (audio re-encodes, frame extraction, concat staging) the CLI cleans up itself.
- Updated all 26 inline `-o /tmp/…` examples in `SKILL.md` and 2 in `README.md` to use relative paths so
agent transcripts model the recommended behavior.
## [3.3.4] - 2026-05-26
### Added
- Added a Sogni-aware default hosted-chat system prompt for `--api-chat` and `--durable-chat`, so hosted
chat describes and uses Sogni's real image, video, music, GPT Image 2, Seedance, workflow, and media-reference
capabilities instead of behaving like a generic text-only assistant.
- Added public guidance for batching prompt-only video variations with Dynamic Prompt syntax and `-n` when each
output shares the same source/end assets, duration, dimensions, and references.
- Added LTX-2.3 dialogue-duration guidance that budgets roughly three spoken words per second plus acting beats,
helping agents choose workable clip durations before submitting paid video jobs.
### Changed
- Bumped `@sogni-ai/sogni-intelligence-client` to `^3.0.8` so the skill consumes the current published Sogni Intelligence client stack, including `@sogni-ai/sogni-client@5.0.0-alpha.15` and `@sogni-ai/sogni-protocol@1.0.0-alpha.6`.
- Synced the bundled creative-agent runtime with the current shared prompt contracts and storyboard helpers, including
provider-neutral storyboard reference wording, field-tag sanitizer compatibility, and updated video Dynamic Prompt
instructions from the shared runtime.
- Tightened persona resolution to explicit saved persona names, ids, and tags/aliases. Relationship phrases such as
"my wife" or "my son" are no longer treated as persona identifiers by themselves.
- Refreshed README, root skill, Claude Code plugin skill, `llm.txt`, and per-skill docs to reflect the current image, video, music, hosted chat, durable workflow, replay, and Seedance/LTX/WAN feature surface.
### Fixed
- `--durable-chat` now recognizes v2 chat-run SSE event names such as `assistant_message_delta`, `run_completed`, `run_failed`, and `run_waiting_for_user` while preserving the legacy aliases.
- Vendor models such as Seedance and GPT Image 2 no longer fall back to SOGNI tokens under `--token-type auto`; they
require Premium Spark eligibility and fail clearly when that billing path is unavailable.
- Hosted API chat now keeps saved persona, memory, and personality injection while using the richer Sogni-specific
prompt, avoiding regressions where media requests could be framed as plain text-only chat.
- Replaced stale agent-facing quick examples that referenced removed `--image`, `--context-image`, and `--i2v --image` flags with current `sogni-agent`, `-c`, and `--video --ref` usage.
## [3.3.3] - 2026-05-22
### Changed
- Aligned skill, OpenClaw, and Claude Code plugin version metadata with the npm package version for the release.
## [3.3.2] - 2026-05-21
### Changed
- Synced the bundled creative-agent runtime with the same source SHA now deployed by `sogni-chat`, picking up the latest prompt contracts for exact Seedance prompts, Seedance V2V remaster routing, relative video segment windows, persona video gating, and non-empty text-only `finalize_response` answers.
- Bumped `@sogni-ai/sogni-intelligence-client` to `^2.4.1` so npm installs use the current shared Sogni runtime dependency set.
- Aligned skill, OpenClaw, and Claude Code plugin version metadata with the npm package version for the release.
### Fixed
- Active persona state now gates only explicit persona-video requests that lack a persona image, instead of broad video-generation requests.
- Text-only / no-action `finalize_response` guidance now requires a substantive final answer instead of an empty or placeholder summary.
## [3.3.1] - 2026-05-21
### Added
- **Seedance direct-gen media upload support.** Local Seedance `--ref-audio` and `--ref-video` references now upload through the Sogni Intelligence `/v2/media/*Url` presigned POST flow and are forwarded as Sogni-hosted URL references, matching the documented backend media-reference contract.
- Regression coverage for Seedance direct-gen local MP3 audio uploads, local V2V source uploads, and vendor policy-failure JSON shaping.
### Changed
- Seedance audio references are now treated as MP3-only. Local audio is trimmed to the requested Seedance clip window and converted to `audio/mpeg` before upload; HTTPS non-MP3 audio references are re-uploaded as prepared MP3 media instead of being forwarded directly.
- Seedance V2V local source clips are trimmed to the requested clip duration before upload so long local source videos are not submitted raw to the vendor.
### Fixed
- Seedance vendor content-policy cancellations now surface friendly, structured CLI errors (`SAFETY_REJECTED` / `content_refused`) without leaking raw vendor task IDs or terminal status payloads.
- Seedance invalid audio-format failures now classify as non-retryable parameter errors with actionable MP3 guidance instead of opaque vendor failures.
## [3.3.0] - 2026-05-20
### Added
- **Background npm update check.** `sogni-agent` now checks the npm registry at most once every 24 hours and surfaces a trailing "update available" notice when a newer version is published. The check times out at 1.5s, never blocks the foreground command, detects the package manager that installed the CLI (npm, pnpm, or yarn) so the suggested install command matches the user's environment, and persists throttle state at `~/.config/sogni/update-check.json`.
- **`--no-update-check`** flag to opt out of the update check for a single run.
- **Claude Code plugin install instructions** in README and `llm.txt`. Both now show the `npm install -g` prerequisite, the marketplace registration command, and the new `/plugin install sogni-creative-agent@sogni` step, with a brief explanation of what each command does.
## [3.2.0] - 2026-05-20
### Added
- **Seedance multi-modal references.** `--ref-audio` and `--ref-video` are now repeatable on Seedance models, and `-c`/`--context` image refs flow through to Seedance `referenceImageUrls` as loose `@ImageN` refs — matching the "up to 9 image / 3 video / 3 audio / 12 total" caps published by `@sogni-ai/sogni-intelligence-client@2.4.0` (sourced from `@sogni-ai/sogni-protocol`'s `SEEDANCE_REFERENCE_LIMITS` catalog).
- **Dedicated first-frame / last-frame mode parity** with sogni-socket's two-mode contract. Dedicated frame mode (`--ref` / `--ref-end`) and loose reference mode (`-c`/`--context`) are mutually exclusive on Seedance; the skill rejects mixed mode client-side with a message pointing to the right mode.
- **Per-job progress and ETA logging during durable chat runs.** `--durable-chat` now emits de-duplicated per-job progress, ETA, and result lines from hosted run events.
- **Hosted-intelligence guidance refresh** in `SKILL.md`, covering the recommended routing through `/v1/chat/completions`, `/v1/creative-agent/workflows`, and `/v1/chat/runs`.
### Changed
- Bumped `@sogni-ai/sogni-intelligence-client` to `^2.4.0` for the `SEEDANCE_REFERENCE_LIMITS` export and refreshed the rest of the Sogni runtime dependency surface.
- New `enforceSeedanceReferenceCaps()` helper translates `SeedanceReferenceLimitError` into a fatal CLI error with the canonical message. Non-Seedance video models reject repeated `--ref-audio` / `--ref-video` flags with a clear error.
### Tests
- 10 new `node:test` cases covering multi-ref forwarding (HTTPS extras → URL arrays), per-modality cap errors, the combined 12-asset cap, dedicated-vs-loose mutex, local-file extra rejection in CLI direct-gen, `seedance2-fast` parity, and non-seedance multi-ref rejection.
## [3.1.1] - 2026-05-20
### Changed
- Bumped `@sogni-ai/sogni-intelligence-client` to `^2.2.8` and refreshed the intelligence-client runtime bundle.
### Fixed
- The skill now invokes the globally installed `sogni-agent` command directly instead of assuming a specific install path, so agents that resolve the binary via `PATH` work in both global-install and `npm link` setups.
- Republished alongside the renamed `@sogni-ai/sogni-client` package so consumers pulling the latest skill no longer hit the unscoped/legacy client name.
## [3.1.0] - 2026-05-20
### Added
- **Claude Code plugin marketplace manifest.** Scaffolds the `sogni` marketplace and `sogni-creative-agent` plugin entry that ships a lean Claude-Code-focused `SKILL.md` from `plugin-skills/` while keeping the full skill spec at the repository root.
- **`setup-sogni-agent-skill` installer.** Adds the design spec and implementation plan for the upcoming bootstrap installer; the runtime work lands in a later release.
### Changed
- Bumped `@sogni-ai/sogni-intelligence-client` through `^2.2.4` to `^2.2.6` as the client stabilized.
### Fixed
- Root SDK is now loaded through a compatible module path so installs on Node module resolvers that disallow deep imports continue to work.
## [3.1.0-alpha.1] - 2026-05-20
### Changed
- Bumped `@sogni-ai/sogni-intelligence-client` to `^2.2.1`.
### Fixed
- Synced `version.mjs` with `package.json` so the runtime `--version` output matches the published npm version.
## [3.1.0-alpha.0] - 2026-05-18
### Added
- **Managed Agent parity with sogni-chat.** `buildSkillDynamicSystemPrompt()` injects the same persona / memory / personality framing as `buildChatDynamicSystemPrompt` (`User's people` / `PERSONA RULES` / `User preferences` / `USER PERSONALITY PREFERENCE`) so saved `--persona-*`, `--memory-set`, and `--personality-*` stores travel into `/v1/chat/completions`. Empty when no stores are populated, so fresh installs are unaffected.
- **`--no-filter` now propagates to `safeContentFilter: false`** on the hosted chat body, in addition to the existing per-tool `disableNSFWFilter` plumbing.
- **`--durable-chat` CLI flag** for `/v1/chat/runs` with SSE assistant deltas — the foundation for the per-job progress / ETA / result events added in 3.2.0.
- **`composition_planning` per-skill manifest** groups `enhance_prompt`, `compose_lyrics`, `compose_instrumental`, `compose_script`, `compose_workflow`, and `compose_workflow_template` into a single capability surface, matching the canonical `@sogni/creative-agent` manifest layout. SKILL.md and `skills/README.md` are updated to list it in the per-skill index, and the cross-surface parity test asserts `ALL_BUILT_IN_SKILLS` exposes it.
- **Conventional commits tooling.** Adopted commitlint + husky for the strict commit-message rules used across the Sogni ecosystem.
- **Semantic-release configuration** added; publish remains manual-gated for now (the CI auto-publish workflow was added and then removed in the same release pending an automation token).