@sogni-ai/sogni-creative-agent-skill 3.1.1 → 3.2.0

package/README.md

@@ -42,6 +42,7 @@ With this skill, an agent can:
 - [Requirements](#requirements)
 - [Installation](#installation)
   - [Node CLI (default)](#node-cli-default)
+  - [Claude Code plugin](#claude-code-plugin)
   - [OpenClaw plugin](#openclaw-plugin)
   - [Hermes Agent / Manus / other frameworks](#hermes-agent--manus--other-frameworks)
   - [Manual install from source](#manual-install-from-source)
@@ -111,6 +112,17 @@ sogni-agent --version
 
 Then point your agent/runtime at this repository's [`SKILL.md`](./SKILL.md). When an install request is ambiguous, install the CLI and skill source together — that's the supported default.
 
+### Claude Code plugin
+
+The Claude Code plugin shells out to the `sogni-agent` CLI installed above, so both steps are required. From inside Claude Code, register the marketplace and install the plugin:
+
+```text
+/plugin marketplace add Sogni-AI/sogni-creative-agent-skill
+/plugin install sogni-creative-agent@sogni
+```
+
+The first command registers a `sogni` marketplace with one plugin entry (`sogni-creative-agent`) backed by a lean Claude-Code-focused [`plugin-skills/sogni-creative-agent/SKILL.md`](./plugin-skills/sogni-creative-agent/SKILL.md); the second installs the plugin into Claude Code. The full skill spec still lives at the repository root [`SKILL.md`](./SKILL.md).
+
 ### OpenClaw plugin
 
 For the published plugin:
@@ -275,6 +287,10 @@ sogni-agent --api-chat --task-profile reasoning --no-thinking \
   "Plan a concise multi-step product launch workflow"
 sogni-agent --list-api-models
 
+# Durable hosted chat run with SSE progress events
+SOGNI_SKILL_USE_SDK_TRANSPORT=1 sogni-agent --durable-chat \
+  "Create a product launch storyboard and render the first hero image"
+
 # Durable hosted workflow (/v1/creative-agent/workflows)
 sogni-agent --api-workflow \
   --video-prompt "The camera slowly pushes in as the sketch comes alive" \
@@ -299,7 +315,7 @@ sogni-agent --get-replay run_abc123 --json
 
 # Opt in to SDK transport for hosted operations (durable workflows + chat).
 # Validates restEndpoint/socketEndpoint via the skill's SSRF guard, then
-# calls sogni.workflows.* / .chat.completions.* directly.
+# calls the SDK workflow/chat methods directly.
 # Falls back to the legacy SSRF-validated fetch path when the env is unset.
 export SOGNI_SKILL_USE_SDK_TRANSPORT=1
 sogni-agent --api-workflow storyboard-video "10s neon city flyover"
@@ -508,6 +524,7 @@ Hosted API modes require `SOGNI_API_KEY`.
 
 - **`--api-chat`** targets `/v1/chat/completions` with Sogni creative-agent tools — best for text-first natural-language workflows. The CLI sanitizes prompt-injection markers before forwarding messages and can use the current server-side creative-agent media tools, including video extension, segment replacement, overlays, subtitles, stitch/orbit/dance composition, and generated artifact indexing. Tune with `--api-tools creative-agent|creative-tools|none`, `--no-api-tool-execution`, `--llm-model`, and `--system`.
 - **Sogni Intelligence controls** include `--task-profile general|coding|reasoning`, `--max-tokens`, and `--thinking` / `--no-thinking`, which forward to `/v1/chat/completions` as `task_profile`, `max_tokens`, and `chat_template_kwargs.enable_thinking`. Use `--list-api-models` or `--get-api-model <id>` to inspect `/v1/models`.
+- **`--durable-chat`** starts a hosted `/v1/chat/runs` record through the SDK transport. Set `SOGNI_SKILL_USE_SDK_TRANSPORT=1` before using it. The CLI streams assistant deltas and de-duplicated per-job progress / ETA / result lines from hosted run events.
 - **`--api-workflow`** targets `/v1/creative-agent/workflows` for durable, async workflow records with event streaming and cancellation. Requests carry `input.steps` plus snake_case controls such as `token_type`, `media_references`, `max_estimated_capacity_units`, and `confirm_cost`.
 - **`--workflow-input`** forwards exact durable workflow JSON (`{ title?, steps: [...] }`). Use this when you need exact multi-step behavior such as repeated `replace_video_segment` steps with `replacementStartSeconds` / `replacementEndSeconds` for interleaved video slices.
 - **`--api-workflow storyboard-video`** generates a storyline, creates a single GPT Image 2 storyboard sheet, then passes that artifact into Seedance as the video reference. The `-Q fast|hq|pro` preset maps to GPT Image 2 low/medium/high quality for that storyboard sheet.
@@ -601,7 +618,7 @@ With both repos checked out as siblings, refresh the generated runtime before pu
 npm run sync:creative-agent-runtime
 ```
 
-Reusable workflow rules should be added to `../sogni-creative-agent` first, then synced here. Keep storyboard planning, tool argument validation, prompt linting, typed media turn intent, and typed repair/control semantics aligned with `sogni-chat`, `sogni-client`, and `sogni-api` hosted chat/workflow endpoints rather than recreating skill-only regex guards. Prefer generated or copied shared helpers for hosted workflow compilation, schema argument validation, `CreativeTurnPlannerFields` / `classifyMediaTurnIntent()` media-routing contracts, repair-control decisions, and guard telemetry summaries over skill-local guard code — this keeps public-agent behavior close to `/v1/chat/completions` and `/v1/creative-agent/workflows`.
+Reusable workflow rules should come from the shared Sogni runtime before they are synced into this public package. Keep storyboard planning, tool argument validation, prompt linting, media-routing decisions, chat-run progress extraction, and repair/control behavior aligned with the hosted `/v1/chat/completions` and `/v1/creative-agent/workflows` APIs. Prefer typed helpers exported by `@sogni-ai/sogni-intelligence-client` or the generated runtime over new skill-local regex guards.
 
 Public-skill regex should stay limited to CLI argument/fact extraction such as file paths, URLs, extensions, dimensions, durations, and explicit positions. Hosted-style decisions such as latest-video continuation, uploaded-video modification, image-selection waits, stitch-after-batch state, and repair/control routing belong upstream in typed planner/runtime fields before they are synced here.
 

package/SKILL.md

@@ -128,12 +128,16 @@ Path override environment variables:
 
 ## Recommended path: route through the hosted Sogni Intelligence endpoints
 
-For any natural-language creative request — anything that should be planned, multi-step, or that benefits from tool selection, repair, or durable workflows — prefer the hosted endpoints over the direct-to-SDK flags. The hosted endpoints are the canonical home for tool dispatch, Structured Contracts v1 (gating policies, repair recipes, prompt contracts), durable workflows, replay, and asset-manifest mapping. They stay aligned with `sogni-chat` and the rest of the `@sogni/creative-agent` consumers automatically.
+For any natural-language creative request — anything that should be planned, multi-step, resumable, or that benefits from tool selection, repair, or durable workflows — prefer the hosted Sogni Intelligence endpoints over the direct-to-SDK media flags. The hosted surfaces are the canonical home for OpenAI-compatible chat, server-side creative tool dispatch, Structured Contracts v1 (gating policies, repair recipes, prompt contracts), durable chat runs, durable workflows, workflow templates, replay, and asset-manifest mapping. They stay aligned with `sogni-chat`, `sogni-api`, and the rest of the `@sogni/creative-agent` consumers.
 
 ```bash
 # Natural-language creative request (LLM picks the tool, dispatches, repairs)
 sogni-agent --api-chat "Turn the attached product photo into a launch poster" --ref product.jpg
 
+# Durable hosted chat run (persisted event log + SSE stream)
+SOGNI_SKILL_USE_SDK_TRANSPORT=1 sogni-agent --durable-chat \
+  "Create a four-shot launch campaign, generate the key art, and animate the hero clip"
+
 # Multi-step durable workflow (resumable, replay-friendly, server-orchestrated)
 sogni-agent --api-workflow \
   --video-prompt "The camera slowly pushes in" \
@@ -196,11 +200,15 @@ sogni-agent --api-chat --ref product.jpg \
 
 # Sogni Intelligence model/replay utilities
 sogni-agent --list-api-models
-sogni-agent --api-chat --task-profile reasoning --no-thinking \
+sogni-agent --api-chat --task-profile reasoning --max-tokens 2000 \
   "Plan a concise multi-step product launch workflow"
 sogni-agent --list-replays 20
 sogni-agent --get-replay run_abc123 --json
 
+# Draft a savable workflow template through the hosted creative-agent tool loop
+sogni-agent --api-chat \
+  "Design a reusable workflow for a 9:16 product teaser from one product photo"
+
 # Durable API workflow: generated keyframe to video with resumable workflow record
 sogni-agent --api-workflow \
   --video-prompt "The camera slowly pushes in as the sketch comes alive" \
@@ -214,43 +222,128 @@ sogni-agent --api-workflow \
   "Animate the referenced sketch"
 
 # Exact durable workflow input with explicit steps
-sogni-agent --api-workflow --workflow-input @workflow.json
+sogni-agent --api-workflow --workflow-input @workflow-input.json \
+  --workflow-idempotency-key product-teaser-v1
 
 # Durable storyboard-video workflow: storyline -> GPT Image 2 storyboard -> Seedance
 sogni-agent --api-workflow storyboard-video --storyboard-frames 6 --duration 12 -Q hq \
   "Create a 9:16 bakery launch video with a neon street-window reveal"
+
+# Workflow management
+sogni-agent --list-workflows
+sogni-agent --resume-workflow wf_durable_workflow_123
 ```
 
 Use `--api-chat` for text-first natural-language workflows that should go through
-Sogni API's OpenAI-compatible `/v1/chat/completions` tool loop. This path
-sanitizes prompt-injection markers before forwarding messages and uses the
-current hosted creative-agent tool surface. Use `--api-workflow` when the caller
-already knows it wants an async durable workflow record under
-`/v1/creative-agent/workflows`. Use `--workflow-input @workflow.json` when the
-caller already has exact durable workflow input with `steps`; the skill forwards
-that body to the API as-is. This is the preferred hosted path for
-exact multi-step plans, including repeated `replace_video_segment` operations
-with `replacementStartSeconds` / `replacementEndSeconds` when interleaving
-existing video slices. Use `--api-workflow storyboard-video`
-when the caller wants the hosted sequence to generate a storyline, create one GPT
-Image 2 storyboard sheet, and feed that image artifact into Seedance as the video
-reference. The `-Q fast|hq|pro` preset maps to GPT Image 2 low|medium|high
-quality for the storyboard sheet. Hosted API requests forward media references
-from `-c`, `--ref`, `--ref-end`, `--ref-audio`,
-`--reference-audio-identity`, and `--ref-video` as `media_references`
-metadata; workflow JSON can bind them into step arguments with
-`sourceStepId: "$input_media"`, and API chat also attaches image refs as vision
-inputs. Local file references are uploaded to Sogni media storage first, then
-forwarded as retrievable URLs for hosted chat and durable workflows. Use the
-direct CLI path for private media that must not leave the local machine.
-Use `--workflow-max-cost <n>` plus `--confirm-cost` / `--no-confirm-cost` to
-forward explicit workflow cost policy.
-Sogni Intelligence utilities are exposed through the same API key path:
-`--list-api-models` / `--get-api-model <id>` read `/v1/models`,
-`--task-profile`, `--max-tokens`, and `--thinking` / `--no-thinking` tune
-`/v1/chat/completions`, and `--list-replays`, `--get-replay`, and
-`--ingest-replay` manage `/v1/replay/records` RunRecords for replay/debug
-viewers.
+Sogni API's OpenAI-compatible `POST /v1/chat/completions` loop. The public
+REST body uses snake_case controls such as `tool_choice`, `response_format`,
+`task_profile`, `token_type`, `app_source`, `media_references`,
+`chat_template_kwargs`, `sogni_tools`, and `sogni_tool_execution`. The endpoint
+normalizes OpenAI `developer` messages to `system`; when a developer message is
+present and no explicit `task_profile` is supplied, the server treats the task
+as `coding`. The CLI sanitizes prompt-injection markers before forwarding
+messages and sends API-key auth so hosted Sogni tools can execute server-side.
+
+Hosted tool surfaces are split by `sogni_tools`:
+
+- `creative-tools` is the public API default when `sogni_tools` is omitted or
+  true. It exposes generation/editing tools (`generate_image`,
+  `generate_video`, `generate_music`, `edit_image`, `apply_style`,
+  `restore_photo`, `refine_result`, `animate_photo`, `change_angle`,
+  `video_to_video`, `stitch_video`, `orbit_video`, `dance_montage`,
+  `sound_to_video`, `extend_video`, `replace_video_segment`, `overlay_video`,
+  `add_subtitles`), media-analysis tools (`analyze_image`, `analyze_video`,
+  `extract_metadata`), and lightweight composition tools (`enhance_prompt`,
+  `compose_lyrics`, `compose_instrumental`, `compose_script`).
+- `creative-agent` is this CLI's default for `--api-chat`. It includes the
+  `creative-tools` surface plus session-control tools
+  (`ask_clarifying_question`, `finalize_response`), asset-manifest tools
+  (`create_asset_manifest`, `inspect_asset`, `label_asset`,
+  `map_assets_for_model`, `validate_asset_references`), and durable planning
+  tools (`compose_workflow`, `compose_workflow_template`). Use this surface
+  when the model should design one-shot workflow plans, draft savable workflow
+  templates, or maintain stable asset references across a multi-step turn.
+- `none` disables Sogni tool injection and leaves only caller-supplied OpenAI
+  tools on raw API/SDK requests. In the CLI, use it with
+  `--no-api-tool-execution` when you want text-only planning without hosted
+  Sogni tool dispatch.
+
+Use `--durable-chat` for long-running, LLM-in-the-loop turns that should be
+persisted as `POST /v1/chat/runs` records instead of a single
+`/v1/chat/completions` request. Chat runs keep an event log, stream via
+`/v1/chat/runs/:id/events/stream`, support cancellation, and can pause for
+persisted cost approval (`/v1/chat/runs/:id/confirm-cost`) in first-party
+clients. The CLI can start and stream durable chat runs through the SDK
+transport when `SOGNI_SKILL_USE_SDK_TRANSPORT=1` is set.
+
+Use `--api-workflow` when the caller already knows it wants an async durable
+workflow under `POST /v1/creative-agent/workflows`. The API now accepts either
+an inline durable plan (`input.steps`) or a saved workflow template invocation
+(`workflow_id` plus `inputs`) and rejects requests that provide both. The CLI's
+generated-keyframe and `storyboard-video` presets submit inline `input.steps`;
+`--workflow-input @workflow-input.json` supplies that `input` object directly.
+Saved template CRUD lives at `/v1/creative-agent/workflows/templates`, and a
+saved template can later be run by API/SDK callers with `workflow_id + inputs`.
+Use `compose_workflow_template` through `--api-chat` to draft a savable template;
+the caller is still responsible for persisting the returned `template_draft`.
+
+Exact multi-step workflow plans should use explicit step dependencies, including
+`replace_video_segment` steps with bounded `replacementStartSeconds` /
+`replacementEndSeconds` when interleaving existing video slices. Workflow JSON
+can bind request media into step arguments with `sourceStepId: "$input_media"`.
+Use `--api-workflow storyboard-video` when the hosted sequence should generate a
+storyline, create one GPT Image 2 storyboard sheet, and feed that image artifact
+into Seedance as the video reference. The `-Q fast|hq|pro` preset maps to GPT
+Image 2 low|medium|high quality for the storyboard sheet.
+
+Hosted API requests forward media references from `-c`, `--ref`, `--ref-end`,
+`--ref-audio`, `--reference-audio-identity`, and `--ref-video` as
+`media_references` metadata. `--ref-audio` and `--ref-video` are repeatable in
+api-chat / durable-chat mode — each entry uploads independently and is exposed
+to the hosted LLM at `@Audio1` / `@Audio2` / `@Video1` etc. API chat also
+attaches image refs as vision inputs. Local file references are uploaded to
+Sogni media storage first, then forwarded as retrievable URLs for hosted chat
+and durable workflows. Use the direct CLI path for private media that must not
+leave the local machine.
+
+### Seedance reference modes (mutually exclusive)
+
+When `--video -m seedance2` or `-m seedance2-fast` is selected, the skill
+exposes the same two-mode pattern that the hosted chat surfaces. Pick one
+mode per video request:
+
+- **Dedicated frame mode — `--ref` and/or `--ref-end`.** First-class
+  first-frame / last-frame anchoring; the Seedance worker pins them as
+  parameter-mode firstFrame / lastFrame. Max 2 images.
+- **Loose reference mode — `-c/--context` plus optional `--ref-audio`
+  extras and `--ref-video` extras.** Anchor frame intent in the prompt with
+  `@Image1` / `@Image2` / `@Video1` / `@Audio1` etc. (e.g. *"Use @Image1 as
+  the opening shot reference"*). Supports up to 9 image refs, 3 video refs,
+  3 audio refs, and 12 total reference assets per video request. The
+  numeric caps come from the canonical
+  `@sogni-ai/sogni-protocol/catalogs/seedance-reference-limits.json` catalog,
+  surfaced through `@sogni-ai/sogni-intelligence-client/tools` as
+  `SEEDANCE_REFERENCE_LIMITS` and `validateSeedanceReferenceCounts()`.
+
+Combining `--ref` / `--ref-end` with `-c/--context` on Seedance is rejected
+client-side with a clear error pointing to the correct mode. In CLI direct-gen
+mode, additional `--ref-audio` / `--ref-video` entries beyond the first must
+be HTTPS URLs (the primary entry can still be a local file path); for local
+multi-file Seedance uploads, use `--api-chat` / `--durable-chat` instead. Use
+`--workflow-max-cost <n>` plus `--confirm-cost` / `--no-confirm-cost` to forward
+explicit workflow cost policy, and `--workflow-idempotency-key` when retrying a
+workflow start request.
+
+Sogni Intelligence utilities are exposed through the same API-key path:
+`--list-api-models` / `--get-api-model <id>` read `/v1/models`, `--task-profile`
+and `--max-tokens` tune `/v1/chat/completions`, and `--list-replays`,
+`--get-replay`, and `--ingest-replay` manage `/v1/replay/records` RunRecords for
+replay/debug viewers. The public chat endpoint also accepts OpenAI-standard
+`reasoning_effort` / `reasoning.effort` in raw API requests. The CLI's
+`--thinking` / `--no-thinking` flags are forwarded as
+`chat_template_kwargs.enable_thinking`; current hosted Qwen requests may
+normalize thinking on server-side, so do not rely on `--no-thinking` as a hard
+suppression switch for `/v1/chat/completions`.
 Hosted API modes require `SOGNI_API_KEY`; this skill's CLI uses API-key
 authentication.
 
@@ -263,15 +356,15 @@ SSRF-validated fetch path. The skill's `sogni-hosted-client.mjs`
 factory still validates `restEndpoint` / `socketEndpoint` against the
 SSRF guard before constructing the SDK client, so the safety contract
 holds.
+For `--durable-chat`, stream output as the run advances; the CLI reports
+assistant deltas plus de-duplicated per-job progress / ETA / result lines from
+hosted run events.
 
 When changing hosted API chat/workflow behavior, keep reusable validation,
-workflow compilation, repair-control, and guard telemetry logic in
-`../sogni-creative-agent` first. The public skill should consume generated or
-copied shared contracts instead of adding skill-local regex guards. Media-routing
-decisions should come from typed planner/runtime contracts such as
-`CreativeTurnPlannerFields`, `classifyMediaTurnIntent()`, `videoContinuation`,
-`videoModification`, `outputGrouping`, `imageSelectionPolicy`, and
-`pendingStitchAfterBatch`; regex is appropriate only for bounded CLI/fact
+workflow compilation, repair-control, and guard telemetry logic in the shared
+Sogni runtime first, then sync it into this public skill. The public skill
+should consume generated or shared typed contracts instead of adding
+skill-local regex guards. Keep local regex limited to bounded CLI/fact
 extraction such as paths, URLs, extensions, dimensions, durations, and explicit
 positions.
 
@@ -358,13 +451,15 @@ positions.
 | `--concat-audio <path>` | Optional audio track to mux over `--concat-videos` output | - |
 | `--concat-audio-start <sec>` | Start offset into `--concat-audio` | - |
 | `--list-media [type]` | List recent inbound media (images\|audio\|all) | images |
-| `--api-chat` | Call `/v1/chat/completions` with Sogni creative-agent tool injection | - |
-| `--api-tools <mode>` | API tool mode: creative-agent\|creative-tools\|none | creative-agent |
+| `--api-chat` | Call OpenAI-compatible `/v1/chat/completions`; CLI default sends the hosted `creative-agent` tool surface | - |
+| `--durable-chat` | Start and stream a durable `/v1/chat/runs` record through SDK transport; requires `SOGNI_SKILL_USE_SDK_TRANSPORT=1` | - |
+| `--api-tools <mode>` | API tool mode: creative-agent\|creative-tools\|none. CLI default is creative-agent; raw API default is creative-tools. | creative-agent |
 | `--no-api-tool-execution` | Plan/tool-call via API chat without executing Sogni tools | - |
 | `--llm-model <id>` | LLM model for `--api-chat` | qwen3.6-35b-a3b-gguf-iq4xs |
 | `--task-profile <profile>` | Sogni Intelligence task profile: general\|coding\|reasoning | - |
 | `--max-tokens <n>` | Max hosted chat completion tokens | 1600 |
-| `--thinking`, `--no-thinking` | Toggle `chat_template_kwargs.enable_thinking` for hosted chat | server default |
+| `--thinking`, `--no-thinking` | Forward `chat_template_kwargs.enable_thinking` for hosted chat; current public Qwen requests may normalize thinking on server-side | server default |
+| `--system <text>` | Override the base system prompt for hosted chat | built-in creative assistant prompt |
 | `--list-api-models`, `--get-api-model <id>` | Inspect Sogni Intelligence LLM model metadata | - |
 | `--list-replays [n]`, `--get-replay <id>`, `--ingest-replay <json\|@path>` | Manage Sogni Intelligence replay RunRecords. List/get output is run through `redactRunRecord` from `@sogni/creative-agent/replay` before printing, so signed URLs, bearer tokens, JWTs, and PEM blocks cannot leak via the CLI. Use `@path` to load JSON from a file. | - |
 | `--skip-redact`, `--no-redact` | Bypass the replay redactor on `--list-replays` / `--get-replay`. Debug-only — emits unredacted RunRecord payloads. | redacted |
@@ -376,9 +471,10 @@ positions.
 | `--storyboard-plan-frames <n>` | Frame count for `--storyboard-plan`. | inferred |
 | `--storyboard-plan-model <id>` | Adapter target for `--storyboard-plan` (seedance, seedance2, gpt-image-2, ltx23, wan). | inferred |
 | `--storyboard-plan-stage <stage>` | Compilation stage for `--storyboard-plan` (storyboard_image, scene_clip). | storyboard_image |
-| `--api-workflow` | Start a durable workflow with explicit `input.steps`; optional `storyboard-video` preset | - |
-| `--workflow-input <json\|@path>` | Durable workflow input JSON. Use `@path` to load from a file. | - |
+| `--api-workflow` | Start `/v1/creative-agent/workflows` with generated inline `input.steps`; optional `storyboard-video` preset | - |
+| `--workflow-input <json\|@path>` | Durable workflow `input` JSON for the start request. Use `@path` to load from a file. | - |
 | `--workflow-title <text>` | Title for generated or storyboard durable workflow input | - |
+| `--workflow-idempotency-key <key>`, `--idempotency-key <key>` | Reuse safely when retrying a durable workflow start request | - |
 | `--workflow-max-cost <n>` | Reject hosted workflow starts above this estimated capacity-unit ceiling | - |
 | `--confirm-cost`, `--no-confirm-cost` | Forward explicit hosted workflow cost confirmation | - |
 | `--storyboard-frames <n>` | Beat count for storyboard-video workflow | - |
@@ -387,7 +483,7 @@ positions.
 | `--generate-audio`, `--no-generate-audio` | Toggle audio generation for generated video steps | - |
 | `--expand-prompt`, `--no-expand-prompt` | Toggle prompt expansion for generated video steps | - |
 | `--watch-workflow` | Stream durable workflow events after start | - |
-| `--list-workflows`, `--get-workflow <id>`, `--workflow-events <id>`, `--stream-workflow <id>`, `--cancel-workflow <id>` | Durable workflow management helpers | - |
+| `--list-workflows`, `--get-workflow <id>`, `--workflow-events <id>`, `--stream-workflow <id>`, `--cancel-workflow <id>`, `--resume-workflow <id>` | Durable workflow management helpers | - |
 | `--api-base-url <url>` | Sogni API base for hosted API modes. Credentials are only sent to `https://api.sogni.ai` by default; use `SOGNI_API_ALLOWED_HOSTS` for trusted custom hosts or `SOGNI_ALLOW_UNSAFE_API_BASE_URL=1` for isolated local testing. | https://api.sogni.ai |
 | `--no-filter` | Disable NSFW content filter | - |
 | `--memory-set <key> <value>` | Save a user preference | - |

package/generated/creative-agent-runtime.mjs

@@ -2285,7 +2285,7 @@ const PROMPT_CONTRACTS = [
         "contractId": "sound_to_video_v1",
         "version": "1.0.0",
         "toolName": "sound_to_video",
-        "baseDescription": "sound_to_video creates audio-synced video from an audio source. Works with uploaded audio\nfiles (mp3, m4a, wav) OR previously generated music from generate_music (auto-detected).\n\nWhen the user asks to \"turn that song/music into a video\" after generate_music, use\nsound_to_video — it will automatically find the generated audio.\n\nFor music visualization (syncing video to a specific song or audio track), use the\ngenerate_music → sound_to_video pipeline. Do NOT use animate_photo or generate_video for\naudio-driven visualization.\n\nanimate_photo and generate_video produce audio natively via LTX 2.3 — never pre-generate\naudio for those tools. sound_to_video is only for when the audio IS the primary creative\ninput driving the video output.",
+        "baseDescription": "sound_to_video creates audio-synced video from an audio source. Works with uploaded audio\nfiles (mp3, m4a, wav) OR previously generated music from generate_music (auto-detected).\n\nWhen the user asks to \"turn that song/music into a video\" after generate_music, use\nsound_to_video — it will automatically find the generated audio.\n\nFor music visualization (syncing video to a specific song or audio track), use the\ngenerate_music → sound_to_video pipeline. Do NOT use animate_photo or generate_video for\naudio-driven visualization.\n\nanimate_photo and generate_video produce audio natively via LTX 2.3 — never pre-generate\naudio for those tools. sound_to_video is only for when the audio IS the primary creative\ninput driving the video output.\n\nPERSONA VOICE CLIPS: Persona voice clips returned by resolve_personas are voice identity\nreferences for LTX Audio ID, not audio tracks to synchronize. If the user explicitly asks\nfor a registered/persona/reference voice or voice clone, call resolve_personas first, then\nuse generate_video with ltx23 and voicePersonaName when there is no image source, or\nanimate_photo with ltx23 and voicePersonaName when an image/source frame should be\nanimated. Do not call sound_to_video for a persona voice clip unless a separate uploaded\naudio track is the primary sync driver.",
         "parameterDocs": {
             "audioSource": "Uploaded audio file or reference to a prior generate_music result. Auto-detected when omitted after generate_music."
         }
@@ -2351,7 +2351,7 @@ const PROMPT_CONTRACTS = [
         "contractId": "resolve_personas_v1",
         "version": "1.0.0",
         "toolName": "resolve_personas",
-        "baseDescription": "resolve_personas is the required first step when the user explicitly names a saved Persona\nor says to use a Persona Image, Persona reference photo, Persona Voice, registered voice,\nor voice clone. Do not answer in prose, ask a follow-up, or finalize before calling this\ntool when a listed Persona name is present.\n\nDIRECT PERSONA IMAGE / VOICE VIDEO: If the user says to use the Persona image/reference\ndirectly/originally, call resolve_personas first, then call animate_photo using the injected\npersona photo as an uploaded image index. For one named Persona, use sourceImageIndex=-1\nor sourceImageIndices=[-1,...] for a multi-clip batch. If Persona Voice was explicitly\nrequested, set voicePersonaName to the exact resolved Persona name and use an LTX model.\nDo not call generate_video for Persona image/voice videos. Do not generate a new image first\nwhen the user explicitly requested the existing Persona image directly.\n\nMULTI-CLIP PERSONA BATCHES: If the user asks for several separate clips from the same\nPersona Image, make one animate_photo call after resolve_personas with repeated persona\nsource indices, one prompt per clip, and the requested per-clip duration. If the user asks\nto stitch the clips, call stitch_video with the returned video indices after animate_photo.",
+        "baseDescription": "resolve_personas is the required first step when the user explicitly names a saved Persona\nor says to use a Persona Image, Persona reference photo, Persona Voice, registered voice,\nor voice clone. Do not answer in prose, ask a follow-up, or finalize before calling this\ntool when a listed Persona name is present.\n\nDIRECT PERSONA IMAGE / VOICE VIDEO: If the user says to use the Persona image/reference\ndirectly/originally, call resolve_personas first, then call animate_photo using the injected\npersona photo as an uploaded image index. For one named Persona, use sourceImageIndex=-1\nor sourceImageIndices=[-1,...] for a multi-clip batch. If Persona Voice was explicitly\nrequested, set voicePersonaName to the exact resolved Persona name and use ltx23. Use\nanimate_photo when an image/source frame should be animated. Use generate_video with\nltx23 when the user requested a persona voice video but no image source exists yet.\nDo not call sound_to_video for Persona voice clips; they are voice identity references,\nnot the audio track to synchronize. Do not generate a new image first when the user\nexplicitly requested the existing Persona image directly.\n\nMULTI-CLIP PERSONA BATCHES: If the user asks for several separate clips from the same\nPersona Image, make one animate_photo call after resolve_personas with repeated persona\nsource indices, one prompt per clip, and the requested per-clip duration. If the user asks\nto stitch the clips, call stitch_video with the returned video indices after animate_photo.",
         "parameterDocs": {
             "names": "Persona names to load. Use the exact listed Persona name; call this before any Persona image/voice video or image generation."
         }

package/llm.txt

@@ -11,6 +11,11 @@ runtimes.
 npm install -g @sogni-ai/sogni-creative-agent-skill@latest
 sogni-agent --version
 
+# Claude Code plugin (requires the CLI install above; the plugin shells out to sogni-agent).
+# Run both slash commands from inside Claude Code:
+#   /plugin marketplace add Sogni-AI/sogni-creative-agent-skill
+#   /plugin install sogni-creative-agent@sogni
+
 # OpenClaw plugin
 openclaw plugins install sogni-creative-agent-skill
 
@@ -67,6 +72,9 @@ for `/v1/chat/completions` with rich creative-agent tools and sanitized
 message forwarding, or
 `sogni-agent --api-workflow --video-prompt "motion" "image prompt"`
 for durable `/v1/creative-agent/workflows` execution.
+Use `SOGNI_SKILL_USE_SDK_TRANSPORT=1 sogni-agent --durable-chat "prompt"` for
+durable `/v1/chat/runs` execution with SSE assistant deltas and per-job
+progress / ETA / result events.
 Sogni Intelligence utilities are available with `--list-api-models`,
 `--get-api-model <id>`, `--task-profile general|coding|reasoning`,
 `--max-tokens <n>`, `--thinking` / `--no-thinking`, `--list-replays [n]`,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sogni-ai/sogni-creative-agent-skill",
-  "version": "3.1.1",
+  "version": "3.2.0",
   "description": "Sogni Creative Agent Skill: agent skill and CLI for Sogni AI image, video, and music generation.",
   "type": "module",
   "main": "sogni-agent.mjs",
@@ -62,11 +62,12 @@
     "openclaw.plugin.json",
     "env.mjs",
     "ssrf-guard.mjs",
+    "update-check.mjs",
     "generated/creative-agent-runtime.mjs",
     "sogni-agent.mjs"
   ],
   "dependencies": {
-    "@sogni-ai/sogni-intelligence-client": "^2.2.6",
+    "@sogni-ai/sogni-intelligence-client": "^2.4.0",
     "execa": "^9.6.1",
     "json5": "^2.2.3",
     "sharp": "^0.34.5"
@@ -77,11 +78,11 @@
     ]
   },
   "devDependencies": {
-    "@commitlint/cli": "^20.5.3",
-    "@commitlint/config-conventional": "^20.5.3",
+    "@commitlint/cli": "^21.0.1",
+    "@commitlint/config-conventional": "^21.0.1",
     "@semantic-release/changelog": "^6.0.3",
     "@semantic-release/git": "^10.0.1",
     "husky": "^9.1.7",
-    "semantic-release": "^24.2.9"
+    "semantic-release": "^25.0.3"
   }
 }

package/skill-package.json

@@ -3,7 +3,7 @@
   "private": true,
   "type": "module",
   "dependencies": {
-    "@sogni-ai/sogni-intelligence-client": "^2.2.6",
+    "@sogni-ai/sogni-intelligence-client": "^2.4.0",
     "execa": "^9.6.1",
     "json5": "^2.2.3",
     "sharp": "^0.34.5"

package/sogni-agent.mjs

@@ -14,6 +14,13 @@ import sharp from 'sharp';
 import { getEnv, hasEnv } from './env.mjs';
 import { PACKAGE_VERSION } from './version.mjs';
 import { assertSafeUrl } from './ssrf-guard.mjs';
+import {
+  INTERNAL_FLAG as UPDATE_CHECK_INTERNAL_FLAG,
+  runForegroundCheck as runUpdateCheckForeground,
+  maybeSpawnBackgroundCheck as maybeSpawnUpdateCheck,
+  getQueuedNotice as getUpdateCheckNotice,
+  runSelfUpdate as runSogniSelfUpdate,
+} from './update-check.mjs';
 import {
   LTX23_WORKFLOW_MODELS,
   PUBLIC_SKILL_DEFAULT_TOOL_DEFINITIONS,
@@ -56,6 +63,14 @@ import {
   redactPayload,
   redactRunRecord
 } from '@sogni-ai/sogni-intelligence-client/replay';
+import {
+  extractToolCallProgressUpdate
+} from '@sogni-ai/sogni-intelligence-client/chatRun';
+import {
+  SEEDANCE_REFERENCE_LIMITS,
+  SeedanceReferenceLimitError,
+  validateSeedanceReferenceCounts
+} from '@sogni-ai/sogni-intelligence-client/tools';
 
 const require = createRequire(import.meta.url);
 const rootClientModule = process.env.SOGNI_AGENT_TEST_STATE_PATH
@@ -121,6 +136,26 @@ const IS_OPENCLAW_INVOCATION = Boolean(getEnv('OPENCLAW_PLUGIN_CONFIG'));
 const RAW_ARGS = process.argv.slice(2);
 const CLI_WANTS_JSON = RAW_ARGS.includes('--json');
 const JSON_ERROR_MODE = CLI_WANTS_JSON || IS_OPENCLAW_INVOCATION;
+
+// --- Update-check entry points --------------------------------------------
+// Internal mode: the detached background child that fetches the npm registry.
+if (RAW_ARGS[0] === UPDATE_CHECK_INTERNAL_FLAG) {
+  await runUpdateCheckForeground({ currentVersion: PACKAGE_VERSION });
+  process.exit(0);
+}
+// User-facing subcommand: `sogni-agent self-update`
+if (RAW_ARGS[0] === 'self-update') {
+  process.exit(runSogniSelfUpdate({}));
+}
+// Fire-and-forget background check (no-op when throttled or skipped)
+try { maybeSpawnUpdateCheck({ cliPath: process.argv[1] }); } catch { /* never break the CLI */ }
+// Trailing notice on exit, if a newer version is on file
+process.on('exit', () => {
+  try {
+    const notice = getUpdateCheckNotice({ currentVersion: PACKAGE_VERSION });
+    if (notice) process.stderr.write(notice + '\n');
+  } catch { /* never break exit */ }
+});
 const SOCKET_EVENT_SUBSCRIPTIONS = Object.freeze({
   modelAvailability: false
 });
@@ -1108,12 +1143,14 @@ const options = {
   angles360Video: null,
   refImage: null, // Reference image for video (start frame)
   refImageEnd: null, // End frame for video interpolation
-  refAudio: null, // Uploaded/generated audio for ia2v/a2v, or s2v lip-sync
+  refAudio: null, // Uploaded/generated audio for ia2v/a2v, or s2v lip-sync (primary)
+  refAudios: [], // Additional Seedance loose audio refs; first --ref-audio fills refAudio, subsequent calls append here
   audioStart: null, // Optional start offset into reference audio
   audioDuration: null, // Optional duration slice for reference audio
   referenceAudioIdentity: null, // Voice identity reference for LTX native audio
   voicePersonaName: null,
-  refVideo: null, // Reference video for animate workflows
+  refVideo: null, // Reference video for animate workflows (primary)
+  refVideos: [], // Additional Seedance loose video refs; first --ref-video fills refVideo, subsequent calls append here
   videoStart: null, // Optional start offset into reference video
   contextImages: [], // Context images for image editing
   looping: false, // Create looping video (i2v only): generate A→B then B→A and concatenate
@@ -1238,11 +1275,13 @@ const cliSet = {
   refImage: false,
   refImageEnd: false,
   refAudio: false,
+  refAudios: false,
   audioStart: false,
   audioDuration: false,
   referenceAudioIdentity: false,
   voicePersonaName: false,
   refVideo: false,
+  refVideos: false,
   videoStart: false,
   context: false,
   looping: false,
@@ -1529,8 +1568,13 @@ for (let i = 0; i < args.length; i++) {
   } else if (arg === '--ref-audio' || arg === '--audio') {
     const raw = requireFlagValue(args, i, arg);
     i++;
-    options.refAudio = raw;
-    cliSet.refAudio = true;
+    if (!options.refAudio) {
+      options.refAudio = raw;
+      cliSet.refAudio = true;
+    } else {
+      options.refAudios.push(raw);
+      cliSet.refAudios = true;
+    }
   } else if (arg === '--audio-start') {
     const raw = requireFlagValue(args, i, arg);
     i++;
@@ -1554,8 +1598,13 @@ for (let i = 0; i < args.length; i++) {
   } else if (arg === '--ref-video') {
     const raw = requireFlagValue(args, i, arg);
     i++;
-    options.refVideo = raw;
-    cliSet.refVideo = true;
+    if (!options.refVideo) {
+      options.refVideo = raw;
+      cliSet.refVideo = true;
+    } else {
+      options.refVideos.push(raw);
+      cliSet.refVideos = true;
+    }
   } else if (arg === '--video-start' || arg === '--video-start-offset') {
     const raw = requireFlagValue(args, i, arg);
     i++;
@@ -1947,6 +1996,9 @@ for (let i = 0; i < args.length; i++) {
     options.showBalance = true;
   } else if (arg === '--version' || arg === '-V') {
     options.showVersion = true;
+  } else if (arg === '--no-update-check') {
+    // Update-check opt-out handled at module load; no-op here so the parser
+    // doesn't reject it as an unknown option.
   } else if (arg === '--help') {
     console.log(`
 sogni-agent - Generate images, videos, and music using Sogni AI
@@ -2015,14 +2067,30 @@ Video Options:
   --auto-resize-assets  Auto-resize video reference assets (default)
   --no-auto-resize-assets  Disable auto-resize for video assets
   --estimate-video-cost Estimate video cost and exit
-  --ref <path|url>      Reference image for video (start frame)
-  --ref-end <path|url>  End frame for interpolation/morphing
-  --ref-audio <path|url> Uploaded/generated audio for ia2v/a2v, or s2v lip-sync
+  --ref <path|url>      Reference image for video (start/first frame on Seedance)
+  --ref-end <path|url>  End frame for interpolation/morphing (last frame on Seedance)
+  --ref-audio <path|url> Audio reference. Repeatable on Seedance models (up to 3 total);
+                         first entry is the primary, extras must be HTTPS URLs in CLI
+                         direct-gen (use --api-chat for multi local-file uploads).
+                         On LTX/WAN: single primary only (for ia2v/a2v/s2v lip-sync).
   --audio-start <sec>   Start offset into --ref-audio for audio-driven clips
   --audio-duration <sec> Duration slice from --ref-audio
   --reference-audio-identity <path>  Voice identity clip for LTX native audio
   --voice-persona <name>  Use saved persona voice clip as LTX voice identity
-  --ref-video <path|url> Reference video for animate/v2v workflows
+  --ref-video <path|url> Video reference. Repeatable on Seedance models (up to 3 total);
+                         first entry is the primary, extras must be HTTPS URLs in CLI
+                         direct-gen. On LTX/WAN: single primary for animate/v2v workflows.
+
+Seedance Reference Modes (mutually exclusive on seedance2 / seedance2-fast):
+  - DEDICATED FRAME MODE: --ref (first frame) and/or --ref-end (last frame).
+    Best when you want canonical first/last frame anchoring; max 2 images.
+  - LOOSE REFERENCE MODE: -c/--context image refs plus optional --ref-audio /
+    --ref-video extras. Anchor frame intent in the prompt with @Image1, @Image2,
+    @Video1, @Audio1 etc. (e.g. "Use @Image1 as the opening shot reference").
+    Up to 9 image / 3 video / 3 audio / 12 total references per video request.
+  Combining --ref/--ref-end with -c/--context on Seedance is rejected client-side.
+  All three modalities pull caps from the canonical
+  @sogni-ai/sogni-protocol seedance-reference-limits catalog.
   --video-start <sec>   Start offset into --ref-video for segmented V2V/animate
   --controlnet-name <n> ControlNet type for v2v: canny|pose|depth|detailer
   --controlnet-strength <n>  ControlNet strength for v2v (0.0-1.0, default: 0.8)
@@ -2076,6 +2144,8 @@ General:
   --token-type <type>   Token type: spark|sogni|auto (default: spark, auto retries with alternate)
   --balance, --balances Show SPARK/SOGNI balances and exit
   --version, -V         Show sogni-agent version and exit
+  --no-update-check     Skip the once-daily npm update check for this run
+  self-update           Upgrade sogni-agent in place (npm/pnpm/yarn/bun auto-detected)
   --extract-last-frame <video> <image>  Extract last frame from a video (safe ffmpeg wrapper)
   --concat-videos <out> <clips...>      Concatenate video clips (safe ffmpeg wrapper, min 2 clips)
   --concat-audio <path> Optional audio track to mux over --concat-videos output
@@ -2916,8 +2986,47 @@ if (options.video) {
   if (options.videoStart !== null && !options.refVideo) {
     fatalCliError('--video-start requires --ref-video.', { code: 'INVALID_ARGUMENT' });
   }
-  if (isSeedanceVideo && options.refAudio && !options.refImage && !options.refImageEnd && !options.refVideo) {
-    fatalCliError('Seedance audio references require --ref or --ref-video.', { code: 'INVALID_ARGUMENT' });
+  if (isSeedanceVideo && options.refAudio && !options.refImage && !options.refImageEnd && !options.refVideo
+      && (!Array.isArray(options.contextImages) || options.contextImages.length === 0)) {
+    fatalCliError('Seedance audio references require --ref, --ref-video, or -c/--context image refs.', { code: 'INVALID_ARGUMENT' });
+  }
+
+  // Seedance reference modes are mutually exclusive:
+  //   - DEDICATED FRAME MODE: --ref (first frame) and/or --ref-end (last frame).
+  //     Up to 2 images; the platform pins them as parameter-mode firstFrame/lastFrame.
+  //   - LOOSE REFERENCE MODE: -c/--context (repeatable image refs), --ref-audio extras,
+  //     --ref-video extras. Up to 9 images / 3 videos / 3 audios / 12 total.
+  //     Anchor frame intent in the prompt with @Image1 / @Video1 / @Audio1 etc.
+  // Mixing dedicated frames with loose image refs is rejected at sogni-socket
+  // (jobsController.js) so we catch it client-side with a clearer message.
+  if (isSeedanceVideo
+      && (options.refImage || options.refImageEnd)
+      && Array.isArray(options.contextImages) && options.contextImages.length > 0) {
+    fatalCliError(
+      'Seedance reference modes are mutually exclusive: --ref/--ref-end (dedicated first/last frame) cannot be combined with -c/--context (loose image references). '
+      + 'Pick one: use --ref/--ref-end for first-class first-frame/last-frame anchoring (max 2 images), '
+      + 'or use -c/--context (plus optional @Image1/@Image2 prompt language) for up to 9 loose image references.',
+      { code: 'INVALID_ARGUMENT', details: {
+          dedicatedFrames: [options.refImage, options.refImageEnd].filter(Boolean),
+          looseImageRefs: options.contextImages,
+        } },
+    );
+  }
+  // Non-Seedance video models do not understand multi-ref audio/video extras —
+  // they only support a single primary --ref-audio / --ref-video each.
+  if (!isSeedanceVideo) {
+    if (Array.isArray(options.refAudios) && options.refAudios.length > 0) {
+      fatalCliError('Multiple --ref-audio entries are only supported for Seedance models (seedance2, seedance2-fast).', {
+        code: 'INVALID_ARGUMENT',
+        details: { model: options.model, extras: options.refAudios },
+      });
+    }
+    if (Array.isArray(options.refVideos) && options.refVideos.length > 0) {
+      fatalCliError('Multiple --ref-video entries are only supported for Seedance models (seedance2, seedance2-fast).', {
+        code: 'INVALID_ARGUMENT',
+        details: { model: options.model, extras: options.refVideos },
+      });
+    }
   }
 
   if (options.referenceAudioIdentity && !['t2v', 'i2v'].includes(options.videoWorkflow)) {
@@ -3246,8 +3355,9 @@ function apiRequestHeaders(apiKey, extra = {}) {
  * Phase 6 P0 — SDK transport dispatch for hosted workflow operations.
  *
  * When `SOGNI_SKILL_USE_SDK_TRANSPORT=1` is set, route hosted workflow
- * start / get / list / events / cancel through `@sogni-ai/sogni-client`
- * via the SSRF-validated `SogniHostedClientFactory` in
+ * start / get / list / events / cancel through
+ * `@sogni-ai/sogni-intelligence-client`'s SDK-backed client via the
+ * SSRF-validated `SogniHostedClientFactory` in
  * `sogni-hosted-client.mjs`. Otherwise fall back to the legacy
  * `fetchApiJson` path so existing users on older SDK versions are
  * unaffected.
@@ -3325,8 +3435,9 @@ async function dispatchWorkflowActionViaSdk(action, apiKey, params) {
  * Phase 6 P0 — SDK transport dispatch for hosted chat completions.
  *
  * When `SOGNI_SKILL_USE_SDK_TRANSPORT=1` is set, route synchronous
- * hosted chat through `@sogni-ai/sogni-client` via the SSRF-validated
- * factory. The SDK's `chat.hosted.create` accepts the same field
+ * hosted chat through `@sogni-ai/sogni-intelligence-client`'s SDK-backed
+ * client via the SSRF-validated factory. The SDK's `chat.hosted.create`
+ * accepts the same field
  * names the legacy fetch sends (`model`, `messages`, `temperature`,
  * `max_tokens`, `token_type`, `app_source`, `sogni_tools`,
  * `sogni_tool_execution`, `task_profile`, `chat_template_kwargs`,
@@ -3464,8 +3575,14 @@ function getApiModeMediaReferences() {
   if (options.refImage) refs.push({ flag: '--ref', value: options.refImage, kind: 'image' });
   if (options.refImageEnd) refs.push({ flag: '--ref-end', value: options.refImageEnd, kind: 'image' });
   if (options.refAudio) refs.push({ flag: '--ref-audio', value: options.refAudio, kind: 'audio' });
+  for (const value of options.refAudios || []) {
+    if (value) refs.push({ flag: '--ref-audio', value, kind: 'audio' });
+  }
   if (options.referenceAudioIdentity) refs.push({ flag: '--reference-audio-identity', value: options.referenceAudioIdentity, kind: 'audio' });
   if (options.refVideo) refs.push({ flag: '--ref-video', value: options.refVideo, kind: 'video' });
+  for (const value of options.refVideos || []) {
+    if (value) refs.push({ flag: '--ref-video', value, kind: 'video' });
+  }
   return refs;
 }
 
@@ -4014,6 +4131,17 @@ async function runApiChatDurable(log, { apiKey, body }) {
       }
       if (!options.json) log(`Durable chat run started: ${runId}`);
 
+      // Per-job tool_call_progress dedupe state. The sogni-api throttled
+      // emitter sends 1 Hz `jobETA` countdowns + per-step progress
+      // ticks per job; we log only when the value actually changes
+      // (and only in non-JSON CLI mode) so a 16-image batch doesn't
+      // pour ~16 lines/sec into the log file.
+      const perJobLogState = new Map();
+      const logJobUpdate = (line) => {
+        if (options.json) return;
+        log(line);
+      };
+
       for await (const event of helpers.sdkChatRunsStreamEvents(client, runId, {})) {
         const type = event?.type || event?.event || '';
         const payload = event?.data || event;
@@ -4028,6 +4156,50 @@ async function runApiChatDurable(log, { apiKey, body }) {
             process.stdout.write(delta);
           }
         }
+        // Per-job progress / ETA / completion / error log lines for
+        // CLI watchers. The sogni-api `tool_call_progress` SSE event
+        // packs `jobIndex` + per-job fields (`jobProgress`,
+        // `jobEtaSeconds`, `resultUrl`, `jobError`) for vendor-emulated
+        // jobs (GPT, Seedance — 1 Hz `jobETA` heartbeat from
+        // sogni-socket) and real workers (per-step progress).
+        // Untouched payloads from older sogni-api builds simply lack
+        // `jobIndex` and skip this block — forward-compatible.
+        if (type === 'tool_call_progress' && payload && typeof payload === 'object') {
+          const {
+            jobIndex,
+            jobProgress,
+            jobEtaSeconds,
+            resultUrl,
+            jobError,
+          } = extractToolCallProgressUpdate(payload);
+          if (jobIndex !== undefined) {
+            const state = perJobLogState.get(jobIndex) ?? {};
+            if (jobError && state.error !== jobError) {
+              logJobUpdate(`[job ${jobIndex}] error: ${jobError}`);
+              state.error = jobError;
+            } else if (resultUrl && state.resultUrl !== resultUrl) {
+              logJobUpdate(`[job ${jobIndex}] done${jobProgress !== undefined ? ` (${Math.round(jobProgress * 100)}%)` : ''} → ${resultUrl}`);
+              state.resultUrl = resultUrl;
+              state.progress = jobProgress ?? state.progress;
+            } else if (jobProgress !== undefined || jobEtaSeconds !== undefined) {
+              // Dedupe: only emit when progress moved >=5% or ETA changed.
+              const pctBefore = state.progress !== undefined ? Math.round(state.progress * 100) : -1;
+              const pctNow = jobProgress !== undefined ? Math.round(jobProgress * 100) : pctBefore;
+              const progressChanged = jobProgress !== undefined && Math.abs(pctNow - pctBefore) >= 5;
+              const etaChanged = jobEtaSeconds !== undefined && jobEtaSeconds !== state.eta;
+              if (progressChanged || etaChanged) {
+                const parts = [`[job ${jobIndex}]`];
+                if (jobProgress !== undefined) parts.push(`${pctNow}%`);
+                else if (state.progress !== undefined) parts.push(`${pctBefore}%`);
+                if (jobEtaSeconds !== undefined) parts.push(`(${jobEtaSeconds}s)`);
+                logJobUpdate(parts.join(' '));
+                if (jobProgress !== undefined) state.progress = jobProgress;
+                if (jobEtaSeconds !== undefined) state.eta = jobEtaSeconds;
+              }
+            }
+            perJobLogState.set(jobIndex, state);
+          }
+        }
         const eventToolCalls =
           payload?.toolCalls
           || payload?.tool_calls
@@ -5306,6 +5478,51 @@ async function appendSafeSeedanceReferenceUrl(target, pathOrUrl, label) {
   return true;
 }
 
+// Effective Seedance reference counts for the current `options` snapshot.
+// Mirrors the per-modality bookkeeping sogni-chat does in
+// uploadedModalityReferenceIndices(...) (chatService.ts ~6149), translated to
+// the skill's primary + extras CLI shape:
+//   images = refImage + refImageEnd + contextImages (loose Seedance @ImageN refs)
+//   audios = refAudio + refAudios (extras)
+//   videos = refVideo + refVideos (extras)
+function effectiveSeedanceReferenceCounts() {
+  const images =
+    (options.refImage ? 1 : 0)
+    + (options.refImageEnd ? 1 : 0)
+    + (Array.isArray(options.contextImages) ? options.contextImages.length : 0);
+  const audios =
+    (options.refAudio ? 1 : 0)
+    + (Array.isArray(options.refAudios) ? options.refAudios.length : 0);
+  const videos =
+    (options.refVideo ? 1 : 0)
+    + (Array.isArray(options.refVideos) ? options.refVideos.length : 0);
+  return { images, audios, videos };
+}
+
+// Wraps the shared validateSeedanceReferenceCounts() so a thrown
+// SeedanceReferenceLimitError is re-raised as a CLI fatal error with the same
+// human message the hosted chat surfaces. Source of truth for the numeric caps
+// (9 / 3 / 3 / 12) is @sogni-ai/sogni-protocol's seedance-reference-limits
+// catalog, surfaced through @sogni-ai/sogni-intelligence-client/tools.
+function enforceSeedanceReferenceCaps() {
+  try {
+    validateSeedanceReferenceCounts(effectiveSeedanceReferenceCounts());
+  } catch (err) {
+    if (err instanceof SeedanceReferenceLimitError) {
+      fatalCliError(err.message, {
+        code: err.code,
+        details: {
+          limitKind: err.limitKind,
+          requestedCount: err.requestedCount,
+          maxCount: err.maxCount,
+          limits: SEEDANCE_REFERENCE_LIMITS,
+        },
+      });
+    }
+    throw err;
+  }
+}
+
 function resolveMultiAngleOutputConfig(outputPath, outputFormat) {
   if (!outputPath) return null;
   const ext = extname(outputPath);
@@ -6499,6 +6716,11 @@ async function main() {
       if (options.refVideo) log(`Reference video: ${options.refVideo}`);
 
       const isSeedanceVideo = isSeedanceModel(options.model);
+      if (isSeedanceVideo) {
+        // Source of truth: @sogni-ai/sogni-protocol catalogs/seedance-reference-limits.json
+        // surfaced through @sogni-ai/sogni-intelligence-client/tools.
+        enforceSeedanceReferenceCaps();
+      }
       const seedanceReferenceImageUrls = [];
       const seedanceReferenceVideoUrls = [];
       const seedanceReferenceAudioUrls = [];
@@ -6518,6 +6740,45 @@ async function main() {
         && options.videoStart === null
         && await appendSafeSeedanceReferenceUrl(seedanceReferenceVideoUrls, options.refVideo, 'Reference video');
 
+      // Seedance loose-reference extras: -c/--context images beyond start/end,
+      // plus repeated --ref-audio / --ref-video entries past the first. The
+      // Sogni Client SDK accepts only URL arrays for these (createJobRequestMessage),
+      // so extras MUST be HTTPS URLs. For multi-file local uploads, use --api-chat /
+      // --durable-chat where the LLM upload pipeline handles per-file uploads.
+      if (isSeedanceVideo) {
+        for (const ctxImage of (Array.isArray(options.contextImages) ? options.contextImages : [])) {
+          if (!ctxImage) continue;
+          if (!isHttpsUrl(ctxImage)) {
+            fatalCliError(
+              `Seedance extra image reference "${ctxImage}" must be an HTTPS URL. ` +
+              'Local file uploads beyond --ref / --ref-end are only supported in --api-chat / --durable-chat mode.',
+              { code: 'INVALID_ARGUMENT', details: { flag: '-c/--context', value: ctxImage } },
+            );
+          }
+          await appendSafeSeedanceReferenceUrl(seedanceReferenceImageUrls, ctxImage, 'Seedance image reference');
+        }
+        for (const extraAudio of options.refAudios) {
+          if (!isHttpsUrl(extraAudio)) {
+            fatalCliError(
+              `Additional --ref-audio "${extraAudio}" must be an HTTPS URL. ` +
+              'Local file uploads beyond the primary --ref-audio are only supported in --api-chat / --durable-chat mode.',
+              { code: 'INVALID_ARGUMENT', details: { flag: '--ref-audio', value: extraAudio } },
+            );
+          }
+          await appendSafeSeedanceReferenceUrl(seedanceReferenceAudioUrls, extraAudio, 'Seedance audio reference');
+        }
+        for (const extraVideo of options.refVideos) {
+          if (!isHttpsUrl(extraVideo)) {
+            fatalCliError(
+              `Additional --ref-video "${extraVideo}" must be an HTTPS URL. ` +
+              'Local file uploads beyond the primary --ref-video are only supported in --api-chat / --durable-chat mode.',
+              { code: 'INVALID_ARGUMENT', details: { flag: '--ref-video', value: extraVideo } },
+            );
+          }
+          await appendSafeSeedanceReferenceUrl(seedanceReferenceVideoUrls, extraVideo, 'Seedance video reference');
+        }
+      }
+
       let imageBuffer = options.refImage && !useRefImageUrl ? await fetchMediaBuffer(options.refImage) : undefined;
       let endImageBuffer = options.refImageEnd && !useRefImageEndUrl ? await fetchMediaBuffer(options.refImageEnd) : undefined;
       let audioBuffer = options.refAudio && !useRefAudioUrl ? await fetchMediaBuffer(options.refAudio) : undefined;

package/update-check.mjs

@@ -0,0 +1,303 @@
+/**
+ * sogni-agent update check — trailing-notification style.
+ *
+ * Public API:
+ *   shouldSkipForEnvironment(opts)  → boolean   (pure)
+ *   compareSemver(a, b)             → -1|0|1    (pure)
+ *   detectPackageManager(env)       → { manager, installCmd }
+ *   formatUpdateNotice(opts)        → string    (pure)
+ *   readState(path)                 → state | null
+ *   writeState(path, state)         → void
+ *   runForegroundCheck(opts)        → Promise<void>   (used by --__update-check)
+ *   maybeSpawnBackgroundCheck(opts) → 'spawned' | 'skipped' | 'fresh'
+ *   getQueuedNotice(opts)           → string | null
+ *   runSelfUpdate(opts)             → number (exit code)
+ */
+
+import { spawn, spawnSync } from 'child_process';
+import { existsSync, mkdirSync, readFileSync, writeFileSync, unlinkSync } from 'fs';
+import { dirname, join } from 'path';
+import { homedir } from 'os';
+import https from 'https';
+
+export const PACKAGE_NAME = '@sogni-ai/sogni-creative-agent-skill';
+export const DEFAULT_STATE_PATH = join(homedir(), '.config', 'sogni', 'update-check.json');
+export const DEFAULT_THROTTLE_MS = 24 * 60 * 60 * 1000; // 24h
+const REGISTRY_URL = `https://registry.npmjs.org/${encodeURIComponent(PACKAGE_NAME)}/latest`;
+const REGISTRY_TIMEOUT_MS = 1500;
+const MAX_RESPONSE_BYTES = 1024 * 1024;
+const INTERNAL_FLAG = '--__update-check';
+
+export { INTERNAL_FLAG };
+
+// ---------- pure helpers ----------
+
+function parseSemverPart(value) {
+  const [main, prerelease] = String(value).split('-', 2);
+  const nums = main.split('.').map((n) => Number.parseInt(n, 10));
+  if (nums.length !== 3 || nums.some((n) => !Number.isFinite(n) || n < 0)) return null;
+  return { nums, prerelease: prerelease || '' };
+}
+
+export function compareSemver(a, b) {
+  const pa = parseSemverPart(a);
+  const pb = parseSemverPart(b);
+  if (!pa || !pb) return 0;
+  for (let i = 0; i < 3; i++) {
+    if (pa.nums[i] !== pb.nums[i]) return pa.nums[i] < pb.nums[i] ? -1 : 1;
+  }
+  if (pa.prerelease === pb.prerelease) return 0;
+  if (!pa.prerelease) return 1;
+  if (!pb.prerelease) return -1;
+  return pa.prerelease < pb.prerelease ? -1 : 1;
+}
+
+export function detectPackageManager(env = process.env) {
+  const ua = env.npm_config_user_agent || '';
+  if (ua.startsWith('pnpm/')) {
+    return { manager: 'pnpm', installCmd: `pnpm add -g ${PACKAGE_NAME}` };
+  }
+  if (ua.startsWith('yarn/')) {
+    return { manager: 'yarn', installCmd: `yarn global add ${PACKAGE_NAME}` };
+  }
+  if (ua.startsWith('bun/')) {
+    return { manager: 'bun', installCmd: `bun add -g ${PACKAGE_NAME}` };
+  }
+  return { manager: 'npm', installCmd: `npm install -g ${PACKAGE_NAME}` };
+}
+
+export function shouldSkipForEnvironment({
+  argv = process.argv,
+  env = process.env,
+  stderr = process.stderr,
+  cliPath = process.argv[1] || '',
+} = {}) {
+  if (Array.isArray(argv) && argv.includes('--no-update-check')) return true;
+  if (env.SOGNI_NO_UPDATE_CHECK === '1' || env.SOGNI_NO_UPDATE_CHECK === 'true') return true;
+  if (env.NO_UPDATE_NOTIFIER === '1' || env.NO_UPDATE_NOTIFIER === 'true') return true;
+  if (env.CI) return true;
+  if (env.SOGNI_AGENT_TEST_STATE_PATH) return true;
+  if (env.OPENCLAW_PLUGIN_CONFIG) return true;
+  if (env.NODE_ENV === 'test') return true;
+  if (env.npm_lifecycle_event) return true; // running under `npm <script>`
+  if (Array.isArray(argv) && argv.includes('--json')) return true;
+  if (stderr && stderr.isTTY === false) return true;
+  // Dev / source checkout: CLI directory contains .git
+  if (cliPath) {
+    try {
+      const cliDir = dirname(cliPath);
+      if (existsSync(join(cliDir, '.git'))) return true;
+    } catch {
+      // ignore
+    }
+  }
+  return false;
+}
+
+export function formatUpdateNotice({
+  currentVersion,
+  latestVersion,
+  installCmd,
+  useColor,
+} = {}) {
+  const color = useColor !== false && !process.env.NO_COLOR && process.stderr.isTTY;
+  const c = {
+    dim: color ? '\x1b[2m' : '',
+    bold: color ? '\x1b[1m' : '',
+    yellow: color ? '\x1b[33m' : '',
+    cyan: color ? '\x1b[36m' : '',
+    reset: color ? '\x1b[0m' : '',
+  };
+  const headline = `Update available ${c.dim}${currentVersion}${c.reset} → ${c.bold}${c.yellow}${latestVersion}${c.reset}`;
+  const cta = `Run ${c.cyan}${installCmd}${c.reset} to update`;
+  const tip = `${c.dim}(or run ${c.reset}${c.cyan}sogni-agent self-update${c.reset}${c.dim}, disable with --no-update-check)${c.reset}`;
+  return ['', headline, cta, tip, ''].join('\n');
+}
+
+// ---------- state file ----------
+
+export function readState(path = DEFAULT_STATE_PATH) {
+  try {
+    if (!existsSync(path)) return null;
+    const raw = readFileSync(path, 'utf8');
+    const parsed = JSON.parse(raw);
+    if (!parsed || typeof parsed !== 'object') return null;
+    return parsed;
+  } catch {
+    return null;
+  }
+}
+
+export function writeState(path, state) {
+  try {
+    const dir = dirname(path);
+    if (!existsSync(dir)) mkdirSync(dir, { recursive: true });
+    writeFileSync(path, JSON.stringify(state, null, 2));
+  } catch {
+    // best-effort; never throw
+  }
+}
+
+export function clearState(path = DEFAULT_STATE_PATH) {
+  try {
+    if (existsSync(path)) unlinkSync(path);
+  } catch {
+    // ignore
+  }
+}
+
+// ---------- network ----------
+
+function fetchLatestVersion({ url = REGISTRY_URL, timeoutMs = REGISTRY_TIMEOUT_MS } = {}) {
+  return new Promise((resolve, reject) => {
+    let settled = false;
+    const finish = (fn, value) => {
+      if (settled) return;
+      settled = true;
+      fn(value);
+    };
+    let req;
+    try {
+      req = https.get(url, { headers: { accept: 'application/json' } }, (res) => {
+        if (res.statusCode !== 200) {
+          res.resume();
+          finish(reject, new Error(`registry status ${res.statusCode}`));
+          return;
+        }
+        let received = 0;
+        const chunks = [];
+        res.on('data', (chunk) => {
+          received += chunk.length;
+          if (received > MAX_RESPONSE_BYTES) {
+            res.destroy();
+            finish(reject, new Error('registry response too large'));
+            return;
+          }
+          chunks.push(chunk);
+        });
+        res.on('end', () => {
+          try {
+            const body = Buffer.concat(chunks).toString('utf8');
+            const parsed = JSON.parse(body);
+            if (parsed && typeof parsed.version === 'string') {
+              finish(resolve, parsed.version);
+            } else {
+              finish(reject, new Error('registry response missing version'));
+            }
+          } catch (err) {
+            finish(reject, err);
+          }
+        });
+        res.on('error', (err) => finish(reject, err));
+      });
+    } catch (err) {
+      finish(reject, err);
+      return;
+    }
+    req.setTimeout(timeoutMs, () => {
+      req.destroy(new Error('registry timeout'));
+    });
+    req.on('error', (err) => finish(reject, err));
+  });
+}
+
+// ---------- foreground (child) check ----------
+
+export async function runForegroundCheck({
+  currentVersion,
+  statePath = DEFAULT_STATE_PATH,
+  url = REGISTRY_URL,
+  timeoutMs = REGISTRY_TIMEOUT_MS,
+  fetcher = fetchLatestVersion,
+  now = Date.now,
+} = {}) {
+  try {
+    const latest = await fetcher({ url, timeoutMs });
+    writeState(statePath, {
+      lastCheckedAt: now(),
+      lastKnownLatest: latest,
+      currentVersion: currentVersion || null,
+    });
+  } catch {
+    // Still record the attempt timestamp so we don't hammer the registry
+    // when offline. Keep any previously-known latest version so the user
+    // still sees the notice for an older known update.
+    const prev = readState(statePath) || {};
+    writeState(statePath, {
+      lastCheckedAt: now(),
+      lastKnownLatest: prev.lastKnownLatest || null,
+      currentVersion: currentVersion || null,
+    });
+  }
+}
+
+// ---------- parent helpers ----------
+
+export function maybeSpawnBackgroundCheck({
+  cliPath = process.argv[1],
+  statePath = DEFAULT_STATE_PATH,
+  throttleMs = DEFAULT_THROTTLE_MS,
+  now = Date.now,
+  spawnFn = spawn,
+  execPath = process.execPath,
+  env = process.env,
+} = {}) {
+  if (shouldSkipForEnvironment({ env })) return 'skipped';
+  const state = readState(statePath);
+  if (state && typeof state.lastCheckedAt === 'number' && now() - state.lastCheckedAt < throttleMs) {
+    return 'fresh';
+  }
+  try {
+    const child = spawnFn(execPath, [cliPath, INTERNAL_FLAG], {
+      detached: true,
+      stdio: 'ignore',
+      env,
+    });
+    child.on('error', () => {});
+    if (typeof child.unref === 'function') child.unref();
+    return 'spawned';
+  } catch {
+    return 'skipped';
+  }
+}
+
+export function getQueuedNotice({
+  currentVersion,
+  statePath = DEFAULT_STATE_PATH,
+  env = process.env,
+} = {}) {
+  if (shouldSkipForEnvironment({ env })) return null;
+  const state = readState(statePath);
+  if (!state || typeof state.lastKnownLatest !== 'string') return null;
+  if (compareSemver(state.lastKnownLatest, currentVersion) <= 0) return null;
+  const { installCmd } = detectPackageManager(env);
+  return formatUpdateNotice({
+    currentVersion,
+    latestVersion: state.lastKnownLatest,
+    installCmd,
+  });
+}
+
+export function runSelfUpdate({
+  env = process.env,
+  statePath = DEFAULT_STATE_PATH,
+  spawnSyncFn = spawnSync,
+  stdio = 'inherit',
+} = {}) {
+  const { manager, installCmd } = detectPackageManager(env);
+  const [command, ...args] = installCmd.split(' ');
+  console.error(`Running: ${installCmd}`);
+  const result = spawnSyncFn(command, args, { stdio, env });
+  if (result.error) {
+    console.error(`self-update failed: ${result.error.message}`);
+    if (manager === 'npm' && /EACCES|EPERM/i.test(result.error.message)) {
+      console.error('Hint: re-run with sudo, or install with a Node version manager (nvm/fnm/volta).');
+    }
+    return 1;
+  }
+  if (typeof result.status === 'number' && result.status !== 0) {
+    return result.status;
+  }
+  clearState(statePath);
+  return 0;
+}

package/version.mjs

@@ -1 +1 @@
-export const PACKAGE_VERSION = '3.1.1';
+export const PACKAGE_VERSION = '3.2.0';