• Non-apparel physical products (cosmetics, electronics, home goods, food, etc.) → Use

  linkfox-aigc-imagegen-product

  series.
• Pure image editing operations (background replacement, watermark removal, angle change, etc. with clear creative directions) → Use

  linkfox-aigc-imagegen

  .
• Pose/action variation of existing model images (model variation) → Not within this scope (business side constructs posePrompt upstream, no dictionary reading).
• Pure text generation →

  linkfox-aigc-textgen

  .
• Video generation →

  linkfox-aigc-videogen

  .

• Input:

  imageUrls

• Empty Array Guard (Mandatory): If

  imageUrls

  is empty / missing / all invalid, stop immediately, do not call downstream with empty images. Prioritize filling from context (original image paths uploaded by users in current/historical rounds); if still unable to fill, truthfully report error to user and request re-upload. FORBID continuing with empty

  imageUrls

  .
• Operation: Check each value in

  imageUrls

  in order: if it is a public URL (starts with http/https) → pass through directly; if it is a local file path (not http/https) → call

  linkfox-file-upload

  to upload and get a public URL for replacement.
• Output:

  imageUrls

  (all are publicly accessible HTTPS URLs)

references/types/<type>.md

Constraints for ①Static Direct Generation Branch (WHITE_BG): Use

scripts/build_imagegen_prompt.py --type WHITE_BG

to build the imagegen parameter file. The script will automatically extract the static content of white background images from the

text

code block in

references/types/white-bg.md

as the final prompt. The content is long, always let the script read it, do not manually copy or pass via shell variables.

Constraints for ②Placeholder Direct Generation Branch (MODEL_IMAGE, SIZE Phase 2): Must use the

scripts/build_imagegen_prompt.py

auxiliary script of this skill to build the imagegen prompt parameter file (the script automatically completes template reading, placeholder filling, null parameter replacement, and sensitive word avoidance instruction appending for SIZE Phase 2). FORBID manually splicing prompts with Write tool or capturing large text with shell variables. Inject the

analysis_result

of SIZE Phase 2 using

--analysis-file

to avoid passing large text via shell.

Constraints for ③textgen Rewriting Branch (SCENE, SELLING_POINT, *_APLUS, SIZE Phase 1): Must use the

build_textgen_params.py

auxiliary script of this skill to build the textgen parameter file (the script automatically completes template reading, placeholder filling, model/thinkingLevel setting, and sensitive word avoidance instruction appending for Selling Point & A+), then execute in chain call mode as specified in textgen SKILL.md.

Image-set Must Follow Three-phase Pipeline: For image-set path, run plan / dispatch / summary three phases with

scripts/run_collection_pipeline.py

, agent sends 3 Bash commands in sequence; concurrency is handled by

ThreadPoolExecutor

at skill layer during dispatch, agent itself does not handle concurrency:

1. --phase plan

   writes

   image-plan.json

   /

   collection-state.json

   , stdout directly outputs markdown table + 1 line of status JSON (including

   specs[]

   /

   run_one_task_script

   /

   state_file

   /

   total

   )
2. After user confirmation, agent sends only 1 dispatch Bash:

   python <root of this skill>/scripts/run_collection_pipeline.py --phase dispatch --state <state_file>

   ; dispatch internally uses

   ThreadPoolExecutor

   to run

   run_one_task.py --index 1..N

   concurrently, each subtask outputs

   task-result-<id>.json

   fragment; dispatch itself does not emit

   Saved full response

   in stdout, only outputs 1 line

   {"status":"dispatch_complete",...}

   at the end for internal agent parsing (this Bash does not render images, it is normal)

3. --phase summary --state <state_file>

   reads fragments → stdout outputs final markdown details (each successful image comes with

   ![label](abs_path)

   inline reference, frontend markdown renderer renders the entire set of images based on this) + 1 line of status JSON; product variant also writes

   collection-asset-manifest.json

FORBID Agent manually completing S1 reasoning or reading full text of S0~S3 in conversation; FORBID splitting into N Bash commands to run

run_one_task.py

after user confirmation (old protocol; concurrency is now handled at skill layer, self-concurrency will only result in duplicate images); FORBID stripping

![]()

lines when forwarding summary markdown (the entire set of images relies on these lines for rendering); FORBID attaching

Saved full response: [...]

outside summary (causes double rendering with markdown

![]()

). The following bash single-step link only applies to single-image direct generation.

Path Rule: This link spans two skill directories (parameter construction scripts belong to this skill, text generation and parameter injection scripts belong to

linkfox-aigc-textgen

), the current working directory is not fixed. All scripts must be called with absolute paths, do not use bare

scripts/...

relative paths (relative paths will point to wrong directories and cause execution failure). Where

<root of this skill>

= absolute path of the directory where this SKILL.md is located;

<textgen root directory>

is obtained by parsing the absolute path of the directory where SKILL.md of

skill:linkfox-aigc-textgen

is located.

1. Parameter Construction (this skill):

   python <root of this skill>/scripts/build_textgen_params.py --type <TYPE> ... --out <textgen_params.json>

2. Text Generation (textgen skill):

   PROMPT=$(python <textgen root directory>/scripts/aigc_textgen.py --stdin --content-only < <textgen_params.json>)

   ——

   --content-only

   outputs single-line content captured into variable (line breaks are flattened to

   ⏎

   , the mechanism for downstream automatic restoration is in textgen SKILL.md).
3. Parameter Injection (Step 3, inline with

   jq

   ): Use

   jq

   to inline

   $PROMPT

   into downstream imagegen parameter JSON, then pass to

   linkfox-aigc-imagegen

   for image generation.
   bash

   PROMPT=$(python <textgen root directory>/scripts/aigc_textgen.py --stdin --content-only < <textgen_params.json>)
   PARAMS=$(jq -nc --arg p "$PROMPT" --argjson imgs '<URL JSON array from Step 1>' \
     '{prompt:$p, imageUrls:$imgs, provider:"<provider>", outputNum:1, aspectRatio:"<ratio>", resolution:"<resolution>"}')
   # Then pass $PARAMS to image generation according to the calling method in linkfox-aigc-imagegen SKILL.md

FORBID manually constructing textgen parameter files with Write tool (must use

build_textgen_params.py

), extracting original content with

jq -r

(will reintroduce real line breaks and break splicing), or directly writing Python to call API to bypass this link.

• Input:

  prompt

  from Step 2,

  imageUrls

  from Step 1,

  ratio

  ,

  resolution

  ,

  provider

• Operation: Read SKILL.md of

  skill:linkfox-aigc-imagegen

  and execute image generation script according to its calling method. Pass parameters:

  • prompt

    : Final prompt produced in Step 2

  • imageUrls

    : All image URLs from Step 1

  • provider

    : Model specified by input (default

    BANANA_PRO

    )

  • aspectRatio

    : Pass

    ratio

    as is

  • resolution

    : Pass as is

  • outputNum

    :

    1

• Output: Image (local path after

  Saved full response:

  )
• Single-image Direct Generation Delivery (By Type): For single-image direct generation scenario of this skill, follow §3.1 of

  references/runtime/03-deliver.md

  . After imagegen Bash completes, agent must:
  1. Parse

     Saved full response: ["..."]

     in stdout to get local path (1 image on success)
  2. Append markdown inline block to conversation reply content:
     markdown

     - Image 1 · <Chinese name of type>
       ![<Chinese name of type>](<abs_path>)

  3. Chinese Name Mapping: SCENE→Lifestyle Marketing Image / SELLING_POINT→Selling Point Image / MODEL_IMAGE→Model Image / SIZE→Size Chart Image / WHITE_BG→White Background Image / PREMIUM_APLUS→Premium A+ Image / STANDARD_APLUS→Standard A+ Image / PHONE_APLUS→Mobile A+ Image
  4. FORBID repeating the protocol text

     Saved full response: [...]

     /

     Saved full response: <json path>

     in conversation reply —— it is an internal protocol line between skill ↔ bridge, repeating it will expose raw JSON array path to users (i.e., "protocol line exposure" fault); only read it to get the path, do not copy the literal text
  5. FORBID manually echoing

     Saved full response:

     literal text to let bridge render again —— the "task output" resource card of bridge path A is automatically triggered by imagegen Bash stdout, agent should not copy it manually
  6. Double rendering (markdown inline image + bridge "task output" resource card displayed simultaneously) is expected behavior under current protocol, not a fault
  7. For failure cases (stdout is

     Saved full response: <data path>.json

     single file), truthfully inform user according to "Error Handling and Failure Guards" in this SKILL.md, do not append inline image

linkfox-aigc-imagegen

skill returns exit 0 for business failures (non-zero exit is not used), so cannot judge success/failure by command exit code, must parse stdout.

• Success: stdout contains

  Saved full response: ["xxx.png", ...]

  (JSON array path) → Image has been stored in media/, get path for display.
• Failure: stdout contains

  Saved full response: xxx.json

  (single JSON file path, not array) → No image product, need to read this JSON to view error details (

  errcode

  /

  errmsg

  /

  error

  ).

errmsg

errorMsg

provider

aspectRatio

resolution

quality

AskUserQuestion

errcode==401

LINKFOX_AGENT_API_KEY

error

Connection failed

• ❌ FORBID switching to other skills due to image generation failure —— The only image generation path for this skill is

  linkfox-aigc-imagegen

  , handle or report according to the above table on failure, never "try another skill".
• ❌ FORBID unlimited retries: Except for "transient errors with at most 1 retry", do not retry the same type of failure repeatedly.
• ❌ FORBID treating failed return body as success and continuing (display empty image, tell user "generated successfully").

1. If agent layer passes

   resolution

   (value is 2K or 4K): use directly
2. If agent layer does not pass: use

   AskUserQuestion

   to let user select from 2K / 4K

1. No Upload Link: Apparel images are directly passed in the form of

   imageUrls

   , this skill does not involve local file upload operations (except URL verification in Step 1). FORBID probing or calling any upload scripts.
2. Planning Tool Usage Specification: Do not use planning tool for single-image direct generation. Must create task plan for N images before concurrent image generation for image sets (see

   references/runtime/02-confirm.md

   ).
3. Image-set Storage: plan / summary are written to disk by

   run_collection_pipeline.py

   ; FORBID Agent writing the entire image-plan from memory with Write.
4. Listing Report Asset Manifest Write-back (Mandatory): If image-set results are to be merged into Listing HTML, caller must retain

   $DATADIR/collection-asset-manifest.json

   , and replace local

   media/

   paths in manifest with public OSS URLs via

   linkfox-file-upload

   before merging. This skill does not upload directly, but must not imply that local paths can be directly written into HTML.

• White background image template uses static instructions, strongly depends on model's judgment of perspective (front/back) and version (men's/women's/children's clothing); wrong judgment will affect silhouette and volume reconstruction.
• Model image link does not call textgen or check sensitive words, all decoding and composition judgment rely on single image generation by the model; template only consumes

  {image_desc}

  , strong control of market-oriented race / brand tone needs to be passed in via

  imageDesc

  text description.
• Lifestyle marketing image strongly depends on

  linkfox-aigc-textgen

  (

  GEM_3_FLASH

  ) judgment of target audience / age (children's vs adult) / gender; wrong judgment will directly affect model portrait and scene composition.
• Selling point image/A+ share dictionary template, distinguished by

  {layoutInfo}

  ; wrong intention filling will lead to layout mismatch. Selling point image/A+ include sensitive word avoidance in textgen phase.
• Size chart image uses two-phase link: quality of

  analysis_result

  in Phase 1 directly determines category judgment and prompt copy of size chart in Phase 2; sensitive word avoidance in Phase 2 is downgraded to "append avoidance instructions at the end of final prompt sent to imagegen, self-reviewed by image generation model", no deterministic word list fallback.
• When multiple reference images are provided, the model judges the role of main apparel automatically, which may misjudge the main image in complex combinations.
• Branches that do not rewrite sensitive words (lifestyle marketing image/model image/white background image): if selling points/scenes contain prohibited content, this link does not intercept, controlled by upstream or manual review.

• API Key Not Configured: Environment variable

  LINKFOX_AGENT_API_KEY

  is not configured, nor is

  LINKFOXAGENT_API_KEY

  .
• Response with 401 or 402 Status Code
• Response Indicates Insufficient Quota or Balance: Message contains "insufficient quota balance/insufficient billing/insufficient balance/quota exceeded/insufficient balance/package expired/need to recharge/please recharge" or similar meanings.