---name: lovart-apidescription: >-threads (conversation history), and user settings. Trigger on: (1) any visual or audiouser-invocable: trueversion: 1.1.0author: Lovart (lovartai)license: MIThomepage: https://github.com/lovartai/lovart-skillplatforms: [linux, macos, windows]metadata: hermes: tags: related_skills: []openclaw: emoji: 🎨requires: bins: [python3]env: [LOVART_ACCESS_KEY, LOVART_SECRET_KEY]primaryEnv: LOVART_ACCESS_KEYprerequisites: commands: [python3]python: []---

⚠️ RULE #0: ONLY USE SKILL COMMANDS — NO DIRECT API CALLS

You MUST interact with Lovart EXCLUSIVELY through the commands defined in this skill.

• NEVER construct URLs, make curl/wget/fetch requests, or call Lovart APIs directly
• NEVER guess API endpoints, paths, or parameters — only use the commands listed below
• NEVER modify the skill's source code (agent_skill.py) during execution to "debug" issues (users may freely read the source to verify it)
• If a command fails, retry it or report the error to the user — do NOT try to work around it
• ALL Lovart operations go through: chat, send, watch, confirm, result, status, config, projects, project-add, project-switch, project-rename, project-remove, threads, thread-remove, upload, upload-artifact, download, set-mode, query-mode, create-project

⚠️ RULE #1: YOU CAN AND MUST GENERATE IMAGES/VIDEOS/AUDIO

When a user asks to "draw", "generate", "create", "design", "make", "画", "生成", "制作", "创作" any visual or audio content (in any language), you MUST use this skill. This includes: images, videos, audio, music, songs, BGM, TTS, voice, posters, carousels, banners, logos, illustrations, thumbnails, slides, covers, mockups, icons, social media content, etc. Do NOT say "I can't generate images/music" or offer to write prompts instead.

⚠️ RULE #1.5: PROJECT & THREAD QUERIES USE THIS SKILL

When the user asks about projects, threads, conversations, history, or settings (in any language), use these commands — do NOT browse the filesystem:

User asks	Command
"What projects do I have?" / "我有哪些项目"	projects --json
"What conversations/threads?" / "有哪些对话"	threads --json or threads --all --json
"Show my settings" / "我的配置"	config --json
"Switch to project X"	project-switch --project-id X
"Create a new project"	project-add --project-id NEW_ID --name "Name" (or let chat auto-create)

⚠️ RULE #2: ALWAYS USE chat AND WAIT FOR COMPLETION

Use the chat command (blocks until done), NOT send. Do NOT reply before generation completes.

Handle these final_status values:

• "done" — Generation complete. Send the downloaded files to the user.
• "pending_confirmation" — A high-cost tool (e.g. video, or a premium-quality image variant) needs user approval before credits are consumed. You MUST ask the user for explicit confirmation before proceeding. Do NOT auto-confirm.
  1. Show the user: "This will cost approximately {estimated_cost} credits. Shall I proceed? (yes/no)"
  2. WAIT for user response. Only if user explicitly says yes/confirm/proceed, run: confirm --thread-id THREAD_ID --json --download (This confirms, waits for completion, and returns the result with downloaded files)
  3. If user declines, do NOT confirm. Just inform them the operation was cancelled.
• "abort" — Generation was aborted. Inform the user.
• "timeout" — Generation is still running but exceeded the wait time. The result may contain partial artifacts.
  1. Send any downloaded files that are already available
  2. Tell the user: "Generation is still in progress. Checking again..."
  3. Run: result --thread-id THREAD_ID --json --download to get the latest results
  4. If status is still "running", wait and retry. If "done", send remaining files.

Handle errors:

If chat throws an error (AgentSkillError), handle it by HTTP status and structured code. The message field already contains a user-ready explanation — surface it to the user as-is.

HTTP status	code	What it means	What to tell the user
402	2012	Quota / billing / risk-control rejection	Show AgentSkillError.message directly — the server already returns a specific message (insufficient credits, free-tier reached, concurrent limit, risk control, phone verification, team plan required, etc.) and a suggested next step.
409	2011	Another task is still running on this thread	"A task is still running on this conversation. Wait for it to finish (status) before sending a new prompt, or start a new thread."
429	1429	API rate limit hit	"Slowing down; rate limit hit. Retry in ~60s."
401	—	AK/SK misconfigured	"API key authentication failed. Please check your LOVART_ACCESS_KEY and LOVART_SECRET_KEY."
—	—	Project.*does not exist in message	"Project not found. Please check the project ID or create a new one."

Rule of thumb: prefer AgentSkillError.message for user-facing copy. Do not try to parse internal codes out of the response — the server already maps them to human-readable messages before returning.

Detect silent generation failures (done with no artifact):

Some prompts end with final_status: "done" but produce no artifacts / empty downloaded. This usually means the upstream image model refused the prompt (content moderation), timed out, or the LLM chose to reply with text instead of calling a tool. The skill flags this automatically — when chat() returns, check:

• result["generation_succeeded"] — boolean. False means no artifact was produced.
• result["warning"] — explanation string (present only when generation_succeeded is False).
• result["agent_message"] — the agent's plain-text reply that hints at why (present when available).

Typical triggers:

• GPT Image 2 with very long/complex prompts involving weapons, specific bodies, or policy-sensitive wording — retry with a different model (--include-tools generate_image_midjourney or generate_image_nano_banana_pro) or simplify the prompt.
• Prompt that describes a task the agent can't fulfill — show agent_message to the user.

⚠️ RULE #3: ALWAYS DELIVER RESULTS + PROJECT LINK

After EVERY generation, you MUST:

1. Use --download flag with chat (or result)
2. Send each downloaded file to the user as a file attachment (images, videos, audio/mp3 — ALL file types):
   • ALWAYS send downloaded[].local_path as file attachments, regardless of file type (.png, .jpg, .mp4, .mp3, etc.)
   • NEVER just paste the URL when a local file has been downloaded — send the actual file
   • Only fall back to displaying URLs if no files were downloaded
3. Append the project canvas link: https://www.lovart.ai/canvas?projectId={project_id}
4. Check failures in the result. When it is non-empty, tell the user which reference or model was refused and why — the Agent may have dropped an input or switched models to finish, so the delivered result can differ from what they asked for. Never report a clean success while failures is non-empty.

⚠️ RULE #4: CHECK LOCAL STATE ON FIRST USE (MANDATORY — DO NOT SKIP)

Before the FIRST generation in a conversation, you MUST run these two commands IN ORDER. This is NOT optional. Do NOT call chat until you have done both.

Step 1: config --json

• Check local state (~/.lovart/state.json) for active_project
• If active_project is set → proceed to Step 2. Do NOT create a new project. Do NOT ask the user.
• If active_project is missing → ask the user: "Do you have an existing Lovart project ID, or should I create a new one?" WAIT for their answer.
• Save with: project-add --project-id PID --name "name"

Step 2: threads --json

• Check if there's a recent thread to continue
• If recent thread exists and topic is related → REUSE it (pass --thread-id THREAD_ID to chat)
• If no threads or completely different topic → omit --thread-id (creates new thread)

CRITICAL RULES:

• NEVER create a new project if config --json already shows an active_project. Reuse it.
• NEVER omit --thread-id when a relevant recent thread exists. Always reuse threads by default.
• NEVER call chat without first running config --json and threads --json in the same conversation.
• The chat command auto-reads active_project from local state — you do NOT need to pass --project-id every time.
• Only create a new project if the user explicitly asks for one.
• Only create a new thread if the topic is completely unrelated to the most recent thread.
• When in doubt, REUSE both the existing project and the existing thread.

---

Lovart Agent OpenAPI Skill

Interact with Lovart AI Agent to generate images, videos, and visual assets via natural language.

Lovart is an AI design platform. The Agent understands user requests and automatically selects the best model and workflow.

Terminology

• Thread — A conversation flow (chat session) with the Lovart AI Agent, NOT a programming thread. Each thread has a unique thread_id and preserves multi-turn context. Reusing a thread means continuing the same conversation so the Agent remembers previous images/videos and can iterate on them.
• Project — A workspace/canvas that groups threads and generated artifacts together. One project can contain multiple threads.

Prerequisites

export LOVART_ACCESS_KEY="ak_xxx"
export LOVART_SECRET_KEY="sk_xxx"

No third-party dependencies. Python standard library only.

Features

1. Chat - Send a message to the AI Agent, get text replies and generated images/videos
2. Confirm - Confirm and wait for high-cost operations (e.g. video generation)
3. Create Project - Create a new project
4. Upload File - Upload a local image/video file, get back a CDN URL
5. Upload Artifact - Upload a link artifact to a project
6. Status/Result - Check thread status and retrieve results
7. Set/Query Mode - Switch between fast (credits) and unlimited (queue) mode

Usage

0. First-time setup (saves to ~/.lovart/state.json)

python3 {baseDir}/scripts/agent_skill.py project-add --project-id PROJECT_ID --name "My Project"

1. Send a message (reads project_id from local state)

python3 {baseDir}/scripts/agent_skill.py chat --prompt "USER_PROMPT" --json --download

To override project: add --project-id PROJECT_ID To continue a conversation: add --thread-id THREAD_ID To list saved threads: python3 {baseDir}/scripts/agent_skill.py threads

2. Create a project

python3 {baseDir}/scripts/agent_skill.py create-project

3. Upload a file (local image/video → CDN URL)

python3 {baseDir}/scripts/agent_skill.py upload --file /path/to/image.png
# Returns: {"url": "https://assets-persist.lovart.ai/img/{user_uuid}/xxx.png"}

Use this when the user sends an image/video file that needs to be passed as an attachment to chat.

4. Upload an artifact

python3 {baseDir}/scripts/agent_skill.py upload-artifact --project-id PROJECT_ID --url "ARTIFACT_URL" --type image

5. Check status / get result

# Status
python3 {baseDir}/scripts/agent_skill.py status --thread-id THREAD_ID

# Result (auto-syncs to gallery/canvas, idempotent)
python3 {baseDir}/scripts/agent_skill.py result --thread-id THREAD_ID --json --download

6. Download artifacts

# Download during chat
python3 {baseDir}/scripts/agent_skill.py chat --prompt "draw a cat" --json --download --output-dir /tmp/lovart

# Download from existing result
python3 {baseDir}/scripts/agent_skill.py result --thread-id THREAD_ID --download --output-dir /tmp/lovart

# Download specific URLs
python3 {baseDir}/scripts/agent_skill.py download --urls URL1 URL2 --output-dir /tmp/lovart --prefix myimg

Typical Workflows

Scenario 1: Generate images/videos/audio (most common)

First, run config --json to check if project_id is set. If not, ask the user and save with project-add.

1. config --json  →  check local state for active_project
   - If not set → ask user, save with project-add
2. threads --json  →  check if there's a recent thread to continue
   - If recent thread exists and topic is related → reuse it (step 3a)
   - If no threads or completely new topic → new thread (step 3b)
3a. chat --thread-id THREAD_ID --prompt "user's request" --json --download
3b. chat --prompt "user's request" --json --download
4. Send each downloaded[].local_path file as an IM attachment to the user
5. The chat command auto-syncs artifacts to canvas and gallery

IDs are auto-persisted locally (~/.lovart/state.json):

• project_id is saved after first chat, reused automatically
• thread_id + topic are saved after each chat for thread switching
• Only create a new project if the user explicitly asks for one
• Only create a new thread (omit --thread-id) when starting a completely new topic
• Run threads to list saved threads for the user to pick from

Scenario 2: Edit with attachments

1. User sends a reference image/video via IM → save to local file
2. upload --file /path/to/image.png  →  get CDN URL
3. chat --prompt "edit this image to..." --project-id PID --attachments "CDN_URL" --json --download
4. Continue as Scenario 1

Scenario 3: Follow-up on same topic (continue context)

1. chat --prompt "change the background to a beach" --project-id PROJECT_ID --thread-id THREAD_ID --json --download

The Agent remembers the previous conversation and can continue editing based on context.

Scenario 4: New topic (new thread)

1. chat --prompt "completely new request" --project-id PROJECT_ID --json --download

Omitting --thread-id creates a new conversation without previous memory.

Scenario 5: Streaming / incremental delivery (multiple artifacts)

Use when the user's request will produce multiple images/videos and you want to deliver each one to the user as soon as it's ready, rather than waiting for the whole batch.

python3 {baseDir}/scripts/agent_skill.py watch --prompt "generate 4 variations of a cyberpunk cat" --json

watch emits NDJSON to stdout (one event per line). Parse line-by-line and deliver each artifact event's local_path to the user immediately:

{"event": "started", "thread_id": "xxx", "project_id": "yyy"}
{"event": "artifact", "type": "image", "url": "https://...", "local_path": "/tmp/lovart/lovart_ab12cd.png"}
{"event": "artifact", "type": "image", "url": "https://...", "local_path": "/tmp/lovart/lovart_ef34gh.png"}
{"event": "pending_confirmation", "thread_id": "xxx", "pending_confirmation": {...}}
{"event": "finished", "thread_id": "xxx", "final_status": "done", "artifact_count": 4}

Files are saved with URL-hash filenames so re-running watch on the same thread won't re-download.

You can also attach to an already-running thread: watch --thread-id THREAD_ID.

When NOT to use watch: single-image requests — use chat (simpler, one-shot response).

Output Format

chat --json returns:

{
  "thread_id": "xxx",
  "status": "done",
  "project_id": "xxx",
  "final_status": "done",
  "items": [
    {"type": "assistant", "text": "Agent's reply"},
    {"type": "generator", "name": "artifacts", "artifacts": [
      {"type": "image", "content": "https://assets-persist.lovart.ai/artifacts/agent/xxx.png"},
      {"type": "video", "content": "https://assets-persist.lovart.ai/artifacts/agent/xxx.mp4"}
    ]}
  ],
  "downloaded": [
    {"type": "image", "url": "https://...", "local_path": "/tmp/lovart/lovart_01.png"}
  ],
  "generation_succeeded": true,
  "failures": []
}

failures lists tool calls that were rejected during the run, even when artifacts were still produced. warning is set alongside it with a one-line summary. See "Checking What Was Rejected" below.

Core Principle

You are a messenger, not a creator. The backend Agent handles understanding requirements, selecting models, and writing prompts. Your job:

1. Relay: Pass the user's original description verbatim to chat
2. Wait: Poll until generation completes
3. Deliver: Send result files to the user

Do NOT rewrite/expand prompts, break down tasks, or add your own style descriptions.

Lovart Generation Mode (MUST use API, not prompt)

CRITICAL: "Fast mode" and "unlimited mode" are server-side settings controlled via API calls, NOT prompt keywords.

Do NOT put "快速模式" or "fast mode" in the prompt text. Instead, call the set-mode command:

# User says "fast mode" / "快速模式" / "skip queue" / "use credits" → RUN THIS:
python3 {baseDir}/scripts/agent_skill.py set-mode --fast

# User says "unlimited mode" / "无限模式" / "free mode" / "save credits" → RUN THIS:
python3 {baseDir}/scripts/agent_skill.py set-mode --unlimited

# Check which mode is active:
python3 {baseDir}/scripts/agent_skill.py query-mode

How it works:

• set-mode --fast calls the Lovart backend API to switch the user's account to fast generation (costs credits, no queue)
• set-mode --unlimited switches to unlimited generation (free, may queue)
• This is a persistent server-side setting — it stays until changed again
• It affects ALL subsequent image/video generations, not just one request
• It has nothing to do with your (the assistant's) response style or behavior

Specifying Models

Option 1: In the prompt (simple, the Agent routes automatically):

python3 {baseDir}/scripts/agent_skill.py chat --prompt "generate ocean waves video using kling" --json --download

Option 2: Via --prefer-models (precise, same as frontend's model selector):

# Prefer a specific image model
python3 {baseDir}/scripts/agent_skill.py chat --prompt "draw a cat" --prefer-models '{"IMAGE":["generate_image_midjourney"]}' --json --download

# Prefer a specific video model
python3 {baseDir}/scripts/agent_skill.py chat --prompt "generate ocean waves" --prefer-models '{"VIDEO":["generate_video_kling_3_0"]}' --json --download

# Combine image and video preferences
python3 {baseDir}/scripts/agent_skill.py chat --prompt "create content" --prefer-models '{"IMAGE":["generate_image_seedream_3_0"],"VIDEO":["generate_video_kling_3_0"]}' --json --download

Available models for --prefer-models:

IMAGE:

Tool name	Display name
generate_image_gpt_image_2_5_flare	GPT Image 2.5 Flare Auto
generate_image_gpt_image_2_5_flare_low	GPT Image 2.5 Flare Low
generate_image_gpt_image_2_5_flare_medium	GPT Image 2.5 Flare Medium
generate_image_gpt_image_2_5_flare_high	GPT Image 2.5 Flare High
generate_image_gpt_image_2_5_flare_xhigh	GPT Image 2.5 Flare xhigh
generate_image_gpt_image_2_5_flare_max	GPT Image 2.5 Flare Max
generate_image_gpt_image_2_5_sunburst	GPT Image 2.5 Sunburst Auto
generate_image_gpt_image_2_5_sunburst_low	GPT Image 2.5 Sunburst Low
generate_image_gpt_image_2_5_sunburst_medium	GPT Image 2.5 Sunburst Medium
generate_image_gpt_image_2_5_sunburst_high	GPT Image 2.5 Sunburst High
generate_image_gpt_image_2_5_sunburst_xhigh	GPT Image 2.5 Sunburst xhigh
generate_image_gpt_image_2_5_sunburst_max	GPT Image 2.5 Sunburst Max
generate_image_gpt_image_2	GPT Image 2 Auto
generate_image_gpt_image_2_low	GPT Image 2 Low
generate_image_gpt_image_2_medium	GPT Image 2 Medium
generate_image_gpt_image_2_high	GPT Image 2 High
generate_image_nano_banana_pro	Nano Banana Pro
generate_image_nano_banana_2	Nano Banana 2
generate_image_seedream_v5_pro	Seedream 5.0 Pro
generate_image_gpt_image_1_5	GPT Image 1.5
generate_image_seedream_v5	Seedream 5.0 Lite
generate_image_luma_uni_1	Luma uni-1
generate_image_luma_uni_1_max	Luma uni-1-max
generate_image_flux_2_max	Flux.2 Max
generate_image_flux_2_pro	Flux.2 Pro
generate_image_seedream_v4_5	Seedream 4.5
generate_image_nano_banana	Nano Banana
generate_image_seedream_v4	Seedream 4
generate_image_midjourney	Midjourney
generate_image_ideogram_v4	Ideogram 4
generate_image_qwen_image3	Qwen Image3
generate_image_qwen_image3_pro	Qwen Image3 Pro
generate_image_nano_banana_2_lite	Nano Banana 2 Lite
generate_image_p_image_ideogram	Ideogram P-Image

VIDEO:

Tool name	Display name
generate_video_seedance_v2_5	Seedance 2.5
generate_video_seedance_v2_0	Seedance 2.0
generate_video_seedance_v2_0_fast	Seedance 2.0 Fast
generate_video_seedance_v2_0_mini	Seedance 2.0 Mini
generate_video_kling_v3	Kling 3.0
generate_video_kling_v3_omni	Kling 3.0 Omni
generate_video_minimax_h3	MiniMax H3
generate_video_kling_v2_6	Kling 2.6
generate_video_wan_v2_6	Wan 2.6
generate_video_veo3_1	Veo 3.1
generate_video_veo3_1_fast	Veo 3.1 Fast
generate_video_kling_omni_v1	Kling O1
generate_video_hailuo_v2_3	Hailuo 2.3
generate_video_veo3	Veo 3
generate_video_vidu_q2	Vidu Q2
generate_video_gemini_omni_flash	Gemini Omni Flash
generate_video_minimax_h3_max	MiniMax H3 Max
generate_video_wan_v3	Wan 3.0
generate_video_wan_v3_prime	Wan 3.0 Prime

3D:

Tool name	Display name
generate_3d_tripo	Tripo

When the user requests a specific model, prefer --prefer-models over putting model names in the prompt.

Option 3: Via --include-tools (strongest steer toward specific tools):

# Steer to upscale
python3 {baseDir}/scripts/agent_skill.py chat --prompt "upscale this image to 4K" --include-tools upscale_image --attachments "IMAGE_URL" --json --download

# Steer to a specific video model
python3 {baseDir}/scripts/agent_skill.py chat --prompt "generate a video" --include-tools generate_video_kling_3_0 --json --download

--include-tools strongly instructs the Agent to prioritize the listed tools. Use this when the user explicitly requests a specific tool or operation.

Two limits worth knowing:

• It is a strong instruction, not an enforced whitelist. The Agent normally follows it, but may pick another tool — for example after the requested one rejects the input. Check failures in the result to see when that happened.
• --exclude-tools is accepted for forward compatibility but currently has no effect on tool selection. To steer away from a tool, name the one you do want with --include-tools.

Reference Subjects from the Asset Library — --subjects

--attachments takes any image URL, and every new URL is reviewed again before a model that requires reviewed inputs will accept it. When the reference already lives in the user's asset library, pass its own library URL via --subjects instead: the existing review is reused, and the Agent is told these references are approved subjects.

python3 {baseDir}/scripts/agent_skill.py chat \
  --prompt "put these two characters in a hallway conversation" \
  --subjects '[{"url":"LIBRARY_URL_A","asset_id":"asset_xxx","display_name":"Bune","channel":"ark_sd2"},
               {"url":"LIBRARY_URL_B","asset_id":"asset_yyy","display_name":"Leo","channel":"ark_sd2"}]' \
  --json --download

Each entry takes url (required) plus optional type (subject_image by default, or subject_audio / subject_video), asset_id, display_name and channel. Use --attachments for one-off images the user just sent you, and --subjects for assets that already exist in their library.

--kits references a brand kit by ID. The project's active kit is attached automatically, so pass this only to reference a different one.

Checking What Was Rejected — failures

A thread can finish with final_status: "done" and still have had tool calls rejected along the way. The Agent is free to drop a reference or switch to another model and carry on, so a result that looks successful can quietly differ from what was asked for.

The result carries a failures array whenever that happens:

{
  "final_status": "done",
  "generation_succeeded": true,
  "warning": "2 tool calls were rejected. generate_video_seedance_v2_0_fast was rejected: ...",
  "failures": [
    {
      "tool": "generate_media",
      "tool_hint": "generate_video_seedance_v2_0_fast",
      "code": "SEEDANCE_ASSET_MODERATION_REJECTED",
      "message": "1 reference asset(s) failed content moderation. Do not retry with the same asset(s); replace them with compliant assets."
    },
    {
      "tool": "generate_media",
      "tool_hint": "generate_video_minimax_h3",
      "code": "INPUT_PARAMS_INVALID",
      "message": "MiniMax H3 resolution must be 768P or 2K."
    }
  ]
}

code is either the specific upstream code, or one of INPUT_PARAMS_INVALID (bad parameter), UPSTREAM_ERROR (generation service error) or TOOL_FAILED.

Always read failures before telling the user the run succeeded. When it is non-empty, tell them what was refused and why. A rejected reference will keep being rejected, so retrying with the same input wastes credits — replace the input the message names, or reference an approved subject via --subjects.

Reasoning Mode — --mode thinking / --mode fast

Lovart has two reasoning modes you can select per thread:

• fast (default) — lightweight single-pass response. Use for simple, one-shot generations where speed matters.
• thinking — deep structured reasoning with planning and multi-step analysis. Use for complex brand systems, multi-asset campaigns, anything that benefits from deliberate planning. Slower but higher quality.

Omitting --mode is equivalent to --mode fast, matching the web UI's default.

# Thinking mode — strategic, multi-step
python3 {baseDir}/scripts/agent_skill.py chat --prompt "design a brand identity system for a sustainable coffee startup" --mode thinking --json --download

# Fast mode — quick one-shot
python3 {baseDir}/scripts/agent_skill.py chat --prompt "draw a cat" --mode fast --json --download

Mode is locked to the thread on its first message. Once you start a thread with --mode thinking, subsequent messages on the same --thread-id stay in thinking mode regardless of later --mode flags. To switch modes, start a new thread (omit --thread-id).

Task-Specific Tool Selection (IMPORTANT)

When the user's request matches a specific operation, use --include-tools to ensure the correct tool:

User says	Use --include-tools
"upscale", "放大", "enlarge", "enhance resolution", "超分"	upscale_image
"edit image", "modify", "change style"	(let Agent decide)
"generate image", "draw", "画"	(let Agent decide, or use --prefer-models)

CRITICAL: When the user asks to "upscale", "enlarge", or increase resolution of an existing image, you MUST use --include-tools upscale_image. Do NOT let the Agent use image generation models for upscaling — they will re-generate the image instead of upscaling it.

Notes

• All APIs use AK/SK HMAC-SHA256 signature authentication
• Video generation takes several minutes; the chat command auto-polls until complete
• Gallery and canvas sync is idempotent — safe to call result multiple times without duplicates
• Connection failures auto-retry 3 times with SSL fallback
• After status becomes "done", waits 5 seconds to re-confirm (guards against sub-agent startup race)