Orange Star Dream Factory is an AI-driven comic drama production platform. Create projects, submit scripts, extract characters, generate storyboards, configure the final cut, and ultimately synthesize videos via API.
Beta Notice
You are currently using Funai-skill Beta. When interacting with users, the Agent should clearly inform them that it is currently in the beta stage.
Important Conclusion
The real backend is not a "linear process that only looks at
workflow.currentStep", but rather:
- One layer is project-level steps:
workflow.currentStep- One layer is chapter-level hidden tasks:
comicPreset.data.chapters[]insceneCaptionsTaskStatus、dialogTaskStatus、sceneTaskStatusIf the Agent only looks at
currentStepand continues to call the next step, it is very easy to get errors.
Skill Update Constraint (must be executed before creation-type operations)
Local machine-readable version file:VERSION
Note: External copy no longer repeatedly hardcodes the version number; the actual version is based on the local
VERSIONand thelatest_version / minimum_required_versionin the remote setup metadata.
Whenever the Agent is about to call this Skill to perform any operation that creates, advances, or modifies a platform task,it must first access the fixed update entry point:
https://neirong.funshion.net/skills/setup-skill.md
At minimum check the following fields:
latest_versionminimum_required_versionforce_updateskill_package_url
Execution rules:
- If this Skill is not yet installed locally, it must be automatically installed first.
- If the remote requires a forced update, or the local version is lower than
minimum_required_version, it must be automatically updated first. - Only after the version check is complete can operations such as creating projects, submitting scripts, advancing steps, generating images, generating videos, rolling back, and synthesizing continue.
Quick Configuration
1. Get Token
Guide the user to visit https://ai.fun.tv/#/openclaw and log in. After obtaining the Token, the Agent writes it to the configuration file on the user's behalf.
2. Configure Token
cp config/.env.example config/.env
# 编辑 config/.env
# AIFUN_TOKEN=...
3. Verify Configuration
./scripts/api-client.sh setup-check
./scripts/api-client.sh check
4. Run the full self-check example
./examples/create-comic.sh
The example script is aself-check script, which will use the real-time default values returned by
select-optionsto automatically create a test project.In Agent interactions with real users, the real-time available options must still be read first; among them,
aspect / stylemust be shown to the user for the user to choose,generateType / scriptTypemay be silently used according to the current default value,model / videoModelmay continue to use the real-time default value when the user has not specified it.In addition, to verify the entire chain, the example script will submit the script first and then show the "confirmation point" in the self-check scenario; this does not mean that real user interactions can skip the pacing requirement of "confirm first, then continue."
Execution Rules the Agent Must Follow
0. Use the default mode silently; do not proactively explain to the user
Unless the user explicitly requests to switch modes, the Agent should directly use the platform default mode:
generateType = 1(text-to-image)scriptType = 解说漫模式(based on the default script mode returned byselect-options)
However, this "silent default"only applies to the mode level, and does not apply to the project visual configuration.
When creating a project, the following two fieldsmust not be silently defaulted, the user must be made to explicitly choose:
aspect(aspect ratio)style(style)
Important:
- This is the Agent's internal default decision; do not proactively treat "currently defaulting to text-to-image / commentary comic mode" as a configuration item that needs to be told to the user.
- Only when the user explicitly requests switching to another mode does it need to be explained and adjusted.
- Even if the default mode is used silently, the underlying layer must still first call
select-options, and use the default values returned by that call's interface, rather than hardcoding constants. aspect / stylemust not, just because the interface returnedisDefault, directly decide on behalf of the user; the user must be shown the options and made to confirm.
1. Before creating a project, you must first call select-options
POST /service/workflow/project/story/select-options
{"appCode":"ai-story","generateType":"1"}
Rules:
- Must use the interface-returned
ratios、generateTypes、models、videoModels、styles、scriptTypes - Forbiddenhardcoding style, model, ratio, script type
isDefaultisrequest-related, do not hardcode historical default values in code or documentation- If the user has not explicitly requested a mode switch, then directly and silently use the default returned by that call's interface
generateType / scriptType - But
aspect / stylemust be shown to the user and chosen by the user, and must not be silently defaulted model / videoModelIf the user has no additional requirements, you may continue to useselect-optionsthe current default value returned- When interacting with the user, there is no need to proactively bring out "text-to-image / commentary comic mode" as a default item to explain; but the user must be clearly made to choose
aspect / style
2. In commentary drama / commentary comic mode, AI ghostwritten scripts must be written in narration-style novel form
If the user asks the AI to help write a script and the current mode iscommentary drama / commentary comic mode, the Agent must write the script asnarration-style novel / story body text, not as a storyboard script.
Must conform to:
- Continuous narrative
- Mainly uses narration to continuously tell the story's progression
- A small amount of natural dialogue may be interspersed, but overall it should still be narrative body text
- Suitable for later handing off to the platform for automatic splitting of characters, storyboard, and narration
The following kinds of writing are forbidden:
镜头1/镜头2旁白:...This kind of explicit script label林舟说:...【音效:脚步声】/砰!/咚——This kind of sound effect description- Stage-direction-style entries
- Storyboard table / script format
Correct style example:
清晨,城市公园的林荫道上,年轻人林舟正在慢跑。阳光穿过树叶,在青石板路上投下斑驳的金色光影。
他刚想加快脚步,一只金毛犬忽然从灌木旁窜了出来,差点撞上他的膝盖。林舟猛地停住脚步,呼吸一乱,循声抬头望去。
不远处,一个扎着马尾的女孩正快步跑来。她一边抓紧手里的牵引绳,一边有些不好意思地朝林舟笑了笑。
Incorrect style example:
镜头1:公园远景。
旁白:清晨的城市公园里,林舟正在跑步。
林舟说:今天一定要把状态跑出来。
镜头2:金毛犬冲出。
3. After a project is created successfully, first show the project link
After successful creation, you must immediately tell the user the project link:
https://ai.fun.tv/#/comic/multiple?projectId={projectId}
4. Do not only look at workflow.currentStep
The following three types of readiness must all be checked:
| Scenario | Check location | Continue condition |
|---|---|---|
| Which major step the project is in | GET /project/{projectId} → workflow.currentStep |
Current step correct |
| Whether the character image is complete | GET /comic/roles/{presetResourceId} → roles[].taskStatus / imgUrl |
All SUCCESS and imgUrl non-empty |
| Whether the chapter-level hidden task is complete | GET /resource/comicPreset/{presetResourceId} → chapters[] |
The corresponding status field is SUCCESS |
5. Do not hardcode tts / voice-clone-select input
novel_extract_roles It is not enough to just pass chapterNum . The actually runnable approach is:
- First call
GET /comic/roles/{presetResourceId} - From
data.voiceInputs[0]ordata.roles[0].inputs[0]extract the livettsinput object - Submit it together with
chapterNumtogether
6. Do not treat sceneUuid as sceneId
scene-list in sceneUuid is only a short ID and cannot be used directly for storyboard/video APIs.
Correct flow:
comicPreset/{presetResourceId}
→ chapters[].sceneTaskId
→ GET /task/{sceneTaskId}
→ data.resourceId (storyboardId)
→ GET /resource/{storyboardId}
→ data.data.scenes[].id (完整 sceneId)
7. Prefer returning to the user videoPlayUrl
If the backend returns a new playable field:
- scene video resource:
videoPlayUrl - final rendered resource:
videoPlayUrl
then the correct approach is:
- Externally return only
videoPlayUrl - Do not expose the raw video in high-level helpers, example scripts, or model-visible output
url - The project link can still be returned as a supplementary entry
8. scene/ai/video must use the scene's current base storyboard image
The correct image source for image-to-video is the current scene'sbase storyboard image:
- Prefer using
GET /storyboard/scene/{sceneId}/prompt/detail?sceneId={sceneId}invideo.imageId - then go to
GET /storyboard/scene/{sceneId}/resourcesfind the one matching thisimageIdcorresponding image resource - If there is still currently
selected == trueimage, it is usually the same image
Do not directly use:
GET /comic/roles/{presetResourceId}returned byroles[].imgUrl
Important: This is not a constraint that "the backend will definitely report an error," but a constraint that "the Agent must proactively ensure semantic correctness."
In live tests, the backend accepted roles[].imgUrl + roles[].imageResourceId, and successfully generated a video resource. That is, when the wrong image is passed, the API may still return 200 success, but what is generated isa video with incorrect semantics。
Therefore:
- Correct image source = the scene's current base storyboard image (prefer
video.imageIdcorresponding image) - Incorrect image source = character image / other images not from the current scene
- Do not rely on the backend to cover for you
Authoritative execution order for Agents
Step 0: Do the setup check first
First read https://neirong.funshion.net/skills/setup-skill.md, and confirm that the local skill version is available.
Step 1: Get project options
POST /service/workflow/project/story/select-options
{"appCode":"ai-story","generateType":"1"}
Step 2: Create project
Use the real-time values returned by the API in the previous step to create the project.
The live backend usually lets
novel_configcomplete automatically, so after project creation the commoncurrentStepisnovel_input, and this is normal.
If the user has not specifically stated otherwise, do not additionally tell the user "currently defaults to text-to-image / narrated comic mode."
But before actually creating the project, the user must first explicitly choose:
- Aspect ratio
aspect- Style
styleOnly after these two fields are confirmed can you combine them with the real-time default
generateType / scriptType / model / videoModelto create the project.
Step 3: Submit script
POST /service/workflow/project/step/next
{
"userProjectId": "项目ID",
"step": "novel_input",
"inputs": [
{"name": "content", "type": "text", "value": "剧本内容"},
{"name": "scriptType", "type": "text", "value": "0"},
{"name": "isSplit", "type": "text", "value": "0"}
]
}
Return resultId then poll GET /task/{resultId}。
If the script is AI-written and the current mode is narrated drama / narrated comic, the script body must remainnarrative novel style, do not write it as a storyboard script, and do not write sound effect descriptions.
Step 4: Intelligent episode splitting
POST /service/workflow/project/step/next
{"userProjectId":"项目ID","step":"novel_opt","inputs":[]}
Return resultId then poll the task.
Step 5: Wait for the character images to complete, then perform the character step
Call first:
GET /service/workflow/comic/roles/{presetResourceId}
Must satisfy:
roles[].taskStatus == SUCCESSroles[].imgUrlnon-empty
Then call:
POST /service/workflow/project/step/next
{
"userProjectId": "项目ID",
"step": "novel_extract_roles",
"inputs": [
{"type": "number", "name": "chapterNum", "value": 1},
{ ...来自 roles/voiceInputs 的 tts 输入对象... }
]
}
If this step is skipped, common errors are:
角色 [xxx] 图片未完成
If the user requests modifying the character appearance at this stage, it must satisfy:
- Current
workflow.currentStep == novel_extract_roles - First initiate new image generation for the target character
- Wait for the new image to become
SUCCESSin the character candidate image list andimgUrlnon-empty - Then explicitly apply this new image
- Finally re-read the character list, confirm the character main image has switched, and return the new image URL to the user
Generating the image without applying it does not count as completing the modification.
Step 6: Wait for sceneCaptionsTaskStatus=SUCCESS, then execute intelligent storyboard
Call first:
GET /service/workflow/resource/comicPreset/{presetResourceId}
Wait:
chapters[chapterNum].sceneCaptionsTaskStatus == SUCCESS
Additional note: in live data, this type of chapter field may first appear as a null value within a short time, then enter RUNNING / SUCCESS. A null value does not necessarily indicate an anomaly, and you must not advance to the next step early based on it.
Then call:
POST /service/workflow/project/step/next
{
"userProjectId": "项目ID",
"step": "novel_scene_captions",
"inputs": [
{"type": "number", "name": "chapterNum", "value": 1},
{"name": "imgGenTypeRef", "value": 1}
]
}
This step returns dialogTaskId, not resultId。
If called too early, common errors are:
该章节提取分镜场景任务未完成,请稍后再试
Step 7: After polling dialogTaskId , you still need to wait for sceneTaskStatus=SUCCESS
After polling dialogTaskId successfully, you still cannot immediately advance to the final video.
You also need to check:
chapters[chapterNum].sceneTaskStatus == SUCCESS
Additional note: in live data,sceneTaskStatus may be empty for a period of time, then become SUCCESS; when judging readiness, you should continue polling rather than directly treating the empty value as a failure.
If called too early, common errors are:
该章节分镜图任务未完成,请稍后再试
Step 8: Advance to the final video step
Wait for sceneTaskStatus=SUCCESS , then call:
POST /service/workflow/project/step/next
{
"userProjectId": "项目ID",
"step": "novel_chapter_scenes",
"inputs": [{"type": "number", "name": "chapterNum", "value": 1}]
}
Step 9: Parse storyboard and sceneId
Use the sceneTaskId -> task -> resourceId -> resource chain to obtain:
storyboardIdscenes[].idscenes[].image
Step 9.5: Single-scene image-to-video (if the user explicitly requests it)
First confirm:
- The current step is already at
novel_chapter_scenes sceneTaskStatus == SUCCESS- Use the complete
sceneId - First get
prompt/detail.video.imageId, then fromresourcesfind the corresponding image resource
Recommended payload form:
{
"sceneId": "完整 sceneId",
"prompt": "来自 prompt/detail 的视频提示词",
"cameraPrompt": "来自 prompt/detail 的镜头提示词",
"model": "doubao-pro",
"duration": "-1",
"firstImage": "scene 当前基准分镜图 URL",
"tailImage": "",
"clarity": "720p",
"imageId": "scene 当前基准 image 资源ID"
}
Do not write it as:
firstImage = roles[].imgUrl
imageId = roles[].imageResourceId
Because this may not report an error, but it will generate a video with incorrect semantics.
Step 10: Final video configuration and composition
Here it is necessary to clearly distinguishthe final video settings stageandthe video composition stage。
Step 10A: Enter novel_chapter_video = enter the final video settings stage
When the chapter advances from novel_chapter_scenes to novel_chapter_video , the user cannot immediately see the final video.
In live testing, common characteristics of this stage are:
workflow.currentStep = novel_chapter_videochapters[].currentStep = novel_chapter_videochapters[].currentSubStep = setting
This indicates that it is currently inthe final video settings stage, and at this point the final video parameters need to be prepared and confirmed.
Step 10B: Save final video settings
The order is as follows:
POST /storyboard/prepareVideoCompositeGET /storyboard/getVideoSetting?userProjectId=...&type=2&chapterNum=1POST /storyboard/saveVideoComposeConfig
This stage mainly corresponds to:
- Cover settings
- Subtitle settings
- Background music
- and composition-related parameters that need to be confirmed before subsequently generating the video
In live testing, aftersaveVideoComposeConfig succeeds, the chapter may still remain:
currentSubStep = setting
That is,successfully saving settings ≠ video composition has started。
Step 10C: Click "Generate Video" = actually initiate video composition
Only when calling:
POST /project/step/actionwithstep=novel_chapter_video
does video composition actually begin.
In the live test, this step will:
- Return a new
resultId(video composition task ID) - Update the chapter's
videoTaskIdto the new task ID - Move the chapter
currentSubStepfromsettingintovideo
Only then is it necessary to:
- Poll the returned
resultId
In the live self-check, this final composition task may remain in the QUEUED state noticeably longer than ordinary steps; do not directly judge a timeout by reusing the short polling budget for the earlier ordinary tasks.
Step 10D: Only after the composition task succeeds can the user see the final finished video
Only after the video composition task SUCCESS :
- will a new final video
resourceId - Only then can the user see the final finished video result in the platform
In other words:
- Entering
novel_chapter_videois not the completion of the final finished video - Saving the cover / subtitles / background music is also not the completion of the final finished video
- The final finished video only truly becomes visible after the video composition task succeeds
saveVideoComposeConfig 's runnable payload must contain at least:
selectCovercoverListselectCoverUrlselectMusicIdselectTitleFontselectSubtitleFonttitlebgmVolumetypechapterNumuserProjectId
After rollback, composition must be redone
The live test confirms:
- After
novel_chapter_videocomposition succeeds, the chapter will record the currentvideoTaskId - Once rolled back to
novel_chapter_scenesand adjusted again, after re-enteringnovel_chapter_video, the old composition result will not automatically become the new final finished video - When
project/step/actionis called again, it will producenewvideoTaskId - and ultimately will also generatenewvideo
resourceId
This shows:
After each rollback and adjustment from the finished video step, the video must be recomposed.
Task briefing and step-by-step confirmation mechanism
The Agent's role should be like a "continuously accompanying intelligent comic drama assistant", not a head-down executor.
Each time an important user-perceivable task is completed, a briefing should immediately be given to the user, containing at least:
- What was just completed
- What key result this step produced
- What is planned next
- If user decision is needed, clearly tell the user what to choose
- Clearly ask the user: whether to continue; if the current step supports modification, then also ask whether to modify the current result
If the project has been created, the briefing should preferentially include the project link:
https://ai.fun.tv/#/comic/multiple?projectId={projectId}
Mandatory pacing requirements
Except for purely technical intermediate actions, the Agent must notautomatically complete the entire chain all the way through without user confirmation.
The correct pacing is:
完成一个大步骤
→ 详细汇报当前结果
→ 询问用户是否满意、是否继续下一步;如果当前步骤支持修改,再询问是否需要修改
→ 等用户确认后再继续
The "major steps" here include at least:
- Project creation completed
- Script writing / submission completed
- Character organization completed
- Intelligent storyboard completed
- Storyboard image generation completed
- Image-to-video completed
- Finished video configuration completed / composition completed
Three mandatory stop points (cannot be skipped)
After the following three nodes are completed, the Agent mustimmediately stop and wait for user confirmation, cannot automatically proceed to the next step:
-
After the novel / script generation is complete
- the complete or core script content must be returned to the user
- explicitly ask: whether the plot, character settings, or tone style need to be modified
- only after the user confirms can the script be submitted or the subsequent steps be entered
-
After character generation is complete
- thecharacter image URLof each character must be returned to the user
- If the current project is still at the
novel_extract_roles(characters and voiceover) step, you should proactively ask the user whether they need to modify the appearance of a certain character - If the user requests a modification, you must first generate a new image, wait for the new image to complete, and then explicitly apply the new image; only after the main image is actually switched is the modification considered complete
- After the modification is complete, the new character image URL must be returned to the user
- Only after the user confirms it is correct can you continue to the scene / storyboard-related steps
-
After scene / storyboard generation is complete
- thescene descriptionof the scene must be returned to the user
- If there are storyboard images that can be displayed, they should also be returned or the viewing method in the project link should be explained
- If the user is dissatisfied with one or more scenes, they should be allowed to continue adjusting the image prompt or video prompt of a single scene at the
novel_chapter_scenesstep and then regenerate - You must proactively ask the user:
- whether to continue converting the storyboard images into videos
- or to directly enter the "final cut" interface and use the current storyboard images for direct compositing
- Only after the user makes an explicit choice can you continue to the image-to-video or final cut step
Content that must be included during character confirmation
The report-back at the character confirmation stage must include at least:
- Character name
- Character image
You cannot only say "character generation is complete" without returning the character image to the user.
Modification rules during character confirmation
Character appearance modifications are only allowed to be performed at the characters and voiceover step.
Character confirmation stage:
- You may ask "whether to modify the character appearance"
- but only on the premise that it is still at
novel_extract_roles - If the project has already advanced to subsequent steps, it can no longer be modified through the character image modification flow
- As long as a modification is performed, you must wait for the new image to complete, explicitly apply it, and return the new image URL to the user
Content that must be included during scene confirmation
The report-back at the scene confirmation stage must include at least:
- Scene name or scene number
- Scene description
- If a visual result has been generated, a storyboard image or the viewing method via the project link may be attached
You cannot only say "scene generation is complete" without returning the scene description to the user for confirmation.
Adjustment rules when dissatisfied with a scene
If the user is dissatisfied with the result of one or more scenes, the Agent should first determine which category the problem belongs to:
- The image itself is wrong: adjust
sceneDescription + promptand then regenerate the image - The image is basically correct, but the video performance is wrong: adjust
prompt(动作) + cameraPrompt(运镜)and then regenerate the video
Storyboard image regeneration image
When regenerating an image, the real prompt is composed of two parts:
sceneDescriptionprompt
Among them, prompt may contain:
@[人物](id:xxxx)
This type of live character reference is used to maintain character consistency and should not be deleted arbitrarily.
Image to video
When converting an image to video, distinguish between:
prompt= actioncameraPrompt= camera movement
Here, do not use the image-side @[人物](id:xxxx) structure again; instead, write plain-text action and camera descriptions directly.
Optional enhancement for multimodal models
If the current Agent / model has multimodal image-viewing capability, after each image or video regeneration, first view the current result, then decide whether to continue fine-tuning:
- On the image side, first check the current
selected=truescene image - On the video side, first check the current video resource's
videoCover
Then make targeted prompt modifications for the problem, rather than retrying blindly.
After the storyboard image is completed, the next step must let the user choose one of two options
When the storyboard image has already been generated, the Agent cannot directly default to entering the final video.
It must proactively ask the user:
- Whether to convert the storyboard image to video(this will increase Dream Value consumption)
- Or directly enter the final video interface for compositing(default is more economical)
Only after receiving this choice can execution continue.
Which actions do not require separate confirmation
The following still belong to technical actions within the same step and can be completed together in one step:
- Poll task status
- Read
comicPreset/storyboard/task - Parse
sceneId/storyboardId - Fall back to the storyboard image step
- Read
select-options
However, once a new user-perceivable result is formed, you must stop and wait for confirmation.
Image generation / video generation constraints
Image generation or video generation must be executed under the storyboard image step , and cannot be executed directly in the final video step.
Check order:
GET /project/{projectId}Checkworkflow.currentStep- If it is currently
novel_chapter_video, firstPOST /project/step/prev - Return to
novel_chapter_scenesbefore continuing
Live test confirmation:
- After a successful fallback,
workflow.currentStepwill becomenovel_chapter_scenes - The chapter's
currentSubStepmay still remain asvideo
This does not indicate an anomaly. Before actually continuing operations, continue to confirm:
sceneTaskStatus == SUCCESS- The scene's current baseline image resource is available
Dream Value rules
Recommended process that does not consume Dream Value by default
The default process is:
分镜图生成完成 → 直接配置成片 → 合成最终视频
But this does not mean the Agent can automatically execute all the way to the end.
The correct approach is:
分镜图生成完成
→ 先向用户展示分镜结果并询问是否修改 / 是否继续
→ 用户确认后再进入成片配置与合成
High-consumption operations that require user confirmation
The following operations consume Dream Value and must first obtain explicit user consent:
- Batch convert to video
check/batchAiVideoThe corresponding real video conversion action - Single image to video
scene/ai/video - Redraw storyboard image
scene/ai/image
check/batchAiVideo is only a query, not an execution
What it returns is:
noVideoCount- How many frames currently have not been converted to motion
Do not mistake this interface for "batch video conversion has already started".
Common issues and solutions
1. Token expired
- Error:
401/未登录或登陆状态已失效 - Solution: re-fetch the token and update
config/.env
2. Character step failed
- Error:
角色 [xxx] 图片未完成 - Cause: the character image hasn't finished generating, or live tts wasn't passed as input
ttsinput - Solution: first check
GET /comic/roles/{presetResourceId}, and once the character is ready, submitvoiceInputs[0]together
3. Smart storyboard step failed
- Error:
该章节提取分镜场景任务未完成,请稍后再试 - Solution: wait for
sceneCaptionsTaskStatus=SUCCESS
4. Advance to final video step failed
- Error:
该章节分镜图任务未完成,请稍后再试 - Solution: wait for
sceneTaskStatus=SUCCESS
5. sceneUuid misuse
- Error:
分镜场景资源不存在 - Cause: treating the short ID as the complete
sceneId - Solution: resolve the complete sceneId according to
sceneTaskId -> storyboardId -> scenes[].idresolve the complete sceneId
6. Image-to-video picks the wrong image source
- Symptom: the API may return
200 success, but the generated video content doesn't match the current storyboard - Cause: treating
roles[].imgUrl/roles[].imageResourceIdas thescene/ai/videooffirstImage/imageId - Solution: must switch to the scene's current baseline storyboard image resource (prefer the image resource corresponding to
prompt/detail.video.imageId)
7. Video configuration save failed
- Error:
请选择封面 - Cause:
saveVideoComposeConfigis missingcoverList/selectCover/selectCoverUrl - Solution: first get
getVideoSetting, then build the complete payload according to the returned content
Code and reference documentation
scripts/api-client.sh: reusable API utility functions and wait logicexamples/create-comic.sh: full self-check flow executed according to real backend behaviorreferences/01-project-management.md: project creation and configuration optionsreferences/02-script-management.md: request bodies and hidden chapter states for each stepreferences/03-character-management.md: character image status andttsinputreferences/05-storyboard-management.md: storyboard / sceneId / storyboard resourcesreferences/06-video-generation.md: final video configuration, batch video checking, and compositionreferences/07-task-status.md: task polling and chapter-level hidden statesreferences/09-error-codes.md: common errors and handling methodsreferences/10-models-and-config.md: real-time configuration and workflow overviewreferences/11-best-practices.md: execution recommendations for Agents