Sn Infographic

sn-infographic

Info graphic generation scene skill (tier 1), relying on the sn-image-generate, sn-image-recognize, and sn-text-optimize tools provided by sn-image-base (tier 0).

Features:

• Evaluation of prompt quality (auto mode)
• Prompt expansion (force/auto mode)
• Multiple rounds of image generation and VLM review
• Output the best result based on quality ranking

Input Specification

Who extracts what: Main Agent parameter extraction resolves max_rounds, output_mode, prompts_expand_mode, and aspect_ratio / image_size (each only when the user gives an explicit value). aspect_ratio and image_size without an explicit value are inferred by the Worker in Step 0.

API Configuration

All API calls in this skill are executed through the sn_agent_runner.py of the sn-image-base skill, with authentication parameters using default values (CLI > environment variables > built-in defaults),无需显式传入。

When encountering MissingApiKeyError or needing to specify a model: pass explicitly via CLI parameters, parameter reference $SN_IMAGE_BASE/references/api_spec.md.

$SN_IMAGE_BASE path explanation: $SN_IMAGE_BASE is the installation directory of the sn-image-base skill (SKILL.md exists). The agent can locate this path by skill name sn-image-base in the list of installed skills.

Architecture: Main Agent + Worker Agent

This skill uses a two-tier agent architecture:

Responsibility Boundaries:

• Worker Agent does not send any messages to the user directly, only returns structured JSON
• Main Agent is responsible for sending all user-visible messages
• Worker Agent's last message must be and only be the JSON string defined in the Return Contract
• Worker Agent's internal VLM calls always execute directly, without spawning subagents

Workflow

Main Agent Workflow

1. Parameter extraction from the user request, in three passes:

   1. Inline KV directives — parse tokens of the form key=value where key ∈ {max_rounds, output_mode, prompts_expand_mode, aspect_ratio, image_size}; strip recognized tokens from the user message, and the remainder becomes user_prompt. Example: "生成一张信息图 max_rounds=3 output_mode=verbose" → user_prompt="生成一张信息图", max_rounds=3, output_mode=verbose.

   2. Keyword recognition (case-insensitive, applied to the stripped text) — fill any parameter not yet set by inline KV using the table below:

      Parameter	Trigger keywords	Resolved value
      output_mode	verbose, 详细, 详尽, 完整统计	verbose
      	friendly, 简洁, 精简	friendly
      max_rounds	N 轮, N rounds, 重试 N 次 (parse N)	N, clamped to [1, 8]
      prompts_expand_mode	强制扩写, force expand, force expansion	force
      	不扩写, 跳过扩写, disable expansion, no expand	disable
      aspect_ratio	an explicit supported ratio (16:9 9:16 4:3 3:4 1:1 2:3 3:2 4:5 5:4 21:9 9:21), with or without a 宽高比 / 画面比例 / 比例 / aspect ratio lead-in	that ratio (validated against the supported set in runtime-parameters.md; unsupported value → leave unset for Worker inference)
      image_size	an explicit supported size (2k 4k), with or without an image_size / 分辨率 / 清晰度 / image size lead-in	that size (vague quality words like 高清 / 超清 do not count); unsupported value → leave unset for Worker inference

   3. Defaults — any parameter still unset falls back to max_rounds=1, output_mode=friendly, prompts_expand_mode=auto. aspect_ratio and image_size have no Main-Agent default: when no explicit value is detected they are left unset for the Worker to infer in Step 0.

   Precedence: inline KV > keyword recognition > default. Values from inline KV are validated against the Input Specification (out-of-range max_rounds clamped to [1, 8]; unrecognized enum values fall back to default and Main Agent should log the mismatch).

2. Send uniform preflight message: "Using sn-infographic skill to generate infographic, please wait..."

3. Start Worker Agent (Sub-Agent), passing in complete parameters and working directory

4. When Worker Agent returns status=ok and need_main_agent_send=true:

   • max_rounds = 1: Generate the Text Summary (see Output Format → friendly mode for length/language rules) from expanded_prompt in the returned JSON (always present for status=ok, see Return Contract), send it, then send the rank=1 single image
   • max_rounds > 1, friendly mode: Generate the Text Summary based on the rank=1 round's result and violations, send it, then send the rank=1 single image
   • max_rounds > 1, verbose mode: Render the verbose template (see Output Format → verbose mode for substitution rules) and send it, then send all images in rank order

5. If Worker Agent returns status=error, report the real error field content to the user

Worker Agent Workflow

Worker Agent receives user_prompt, max_rounds, prompts_expand_mode, an optional aspect_ratio and image_size (each set only when the user gave an explicit value), and the working directory of this skill (SKILL_DIR). (output_mode stays on the Main Agent side — Worker has no branch that depends on it.)

Worker Environment

Variables referenced as $NAME in the bash snippets below. Worker must bind each before the step that consumes it.

Naming: $SKILL_DIR for own files; $SN_<SKILL_NAME> (e.g. $SN_IMAGE_BASE) for cross-skill references.

JSON parsing: every sn_agent_runner.py ... -o json call prints a JSON envelope on stdout ({"status", "result", "model", ...}; diagnostics go to stderr). A failed call sets status to a non-ok value (e.g. failed) and omits result, so always confirm .status == ok on the envelope before reading .result — otherwise a missing .result surfaces as the literal null, which is itself valid JSON and slips past both jq -r and extract_json.py (no error raised). The LLM's own JSON (evaluation / analysis steps) lives inside the result string and may carry stray prose or ```json fences. Before any jq, pipe the runner output through $SN_IMAGE_BASE/scripts/extract_json.py (reads stdin, prints the recovered JSON, exits non-zero when none is found); for steps that parse the inner LLM/VLM JSON, pipe .result through it as well. A non-ok status or a non-zero extract_json.py exit means the response is unusable → return the Error Flow JSON.

Step 0 — Initialization

1. Generate task_id (timestamp, format YYYYMMDD_HHMMSS) and create the uniform temporary directory /tmp/openclaw/sn-infographic/<task_id>/ as TEMP_DIR. TEMP_DIR must exist before any subsequent step writes to it:

   TASK_ID=$(date +%Y%m%d_%H%M%S)
   TEMP_DIR="/tmp/openclaw/sn-infographic/${TASK_ID}"
   mkdir -p "$TEMP_DIR"

2. Initialize an empty rounds list

3. Resolve aspect_ratio and image_size (bind to ASPECT_RATIO / IMAGE_SIZE): use the explicit Main Agent value when present, else infer from user_prompt per $SKILL_DIR/references/runtime-parameters.md. Defaults: aspect_ratio → 16:9; image_size inference currently has a single option, 2k.

Step 1 — Decide whether to rewrite the image prompt (always runs)

This step decides whether to rewrite/expand the user's image-generation prompt text before Step 3 generates the image, and produces the boolean should_expand. When Step 2 is skipped it also sets EXPANDED_PROMPT and records prompts_expand_skipped = true.

Scope (do not over-read the step name). "Expand" here means rewriting the text prompt for image generation, nothing more. This is not task decomposition, plan generation, or a plan-review gate, and Step 1 starts no agents of its own — it is a single sn-text-optimize call made by the Worker itself. Do not map it onto any subagent-driven-development / delegate_task-style workflow, and write no plan files. (The Worker is the only sub-agent this skill uses, started once by the Main Agent; the Worker spawns none of its own — see Responsibility Boundaries.)

PROMPTS_EXPAND_MODE is the already-resolved input handed over by the Main Agent — Step 1 acts on it, it does not re-parse the user request. Resolving the mode value (Main Agent parameter extraction) and running this decision are two different jobs: completing the former does not complete Step 1. Do not skip this step; in auto mode the evaluation call below is mandatory — never infer should_expand from the prompt's apparent quality.

Only Step 2 (prompt expansion) is ever skipped, and only when this step decides so. The branch depends on PROMPTS_EXPAND_MODE:

auto mode (default):

1. Run the evaluation call (mandatory). The inner evaluation JSON lives inside .result; its schema is $SKILL_DIR/references/evaluation-standard.md (required_results, optional_results).

   EVAL_ENVELOPE=$(python "$SN_IMAGE_BASE/scripts/sn_agent_runner.py" sn-text-optimize \
     --system-prompt-path "$SKILL_DIR/references/evaluation-standard.md" \
     --user-prompt "$USER_PROMPT" \
     --output-format json | python "$SN_IMAGE_BASE/scripts/extract_json.py")
   
   # A failed runner envelope (status != ok) carries no .result.
   EVAL_STATUS=$(printf '%s' "$EVAL_ENVELOPE" | jq -r '.status')
   
   EVAL=$(printf '%s' "$EVAL_ENVELOPE" | jq -r '.result' \
     | python "$SN_IMAGE_BASE/scripts/extract_json.py")

2. Decide should_expand:

   • required_pass: all answer in required_results are "yes"
   • optional_pass: count of answer="yes" in optional_results / total ≥ 0.6
   • should_expand = not (required_pass and optional_pass)

3. Conservative fallback: if EVAL_STATUS is not ok (the evaluation call itself failed) or extract_json.py exits non-zero, default should_expand = true. Unlike Steps 2.0 / 2.3, a failed evaluation does not abort the Worker — auto falls back to expanding.

4. If should_expand = true: execute Step 2.

5. If should_expand = false: skip Step 2, set EXPANDED_PROMPT to the original prompt, record prompts_expand_skipped = true:

   EXPANDED_PROMPT="$USER_PROMPT"
   echo "$EXPANDED_PROMPT" > "$TEMP_DIR/expanded-prompt.txt"

force mode:

• Skip the evaluation, always execute Step 2 (expansion is mandatory).
• prompts_expand_skipped is not recorded — the field appears in the Return JSON only when Step 2 is skipped (see Return Contract rules).

disable mode:

• Skip both the evaluation and Step 2; use user_prompt directly as expanded_prompt:

  EXPANDED_PROMPT="$USER_PROMPT"
  echo "$EXPANDED_PROMPT" > "$TEMP_DIR/expanded-prompt.txt"

• Record prompts_expand_skipped = true.