ThinkingParameterExtraction

ThinkingParameterExtraction

This skill defines extraction and evidence rules only. It does not orchestrate the full websocket workflow.

Scope

• Goal: produce normalized evidence that can be compiled into a thinking recipe preview.
• Non-goal: trigger save automatically.
• Save boundary: human confirms in frontend, then backend save API persists.

Inputs

Accept only:

• Official documentation URL (doc_url)
• Official code sample or parameter snippet (snippet)

Required runtime context:

• model_name
• base_url

Routing Rules

• Only snippet: run code_extractor_subagent only.
• Only doc_url: run url_extractor_subagent only.
• doc_url + snippet: run both extractors, then check consistency.
• Run compare_subagent only when dual-source consistency check reports conflict.

Subagent Contracts

code_extractor_subagent

Responsibilities:

• Clean snippet (remove unrelated prose/noise).
• Extract candidate thinking control fields and allowed values.
• Keep source evidence and exact quoted fragments.

Output shape:

{
  "source_type": "snippet",
  "evidence": [{"kind": "snippet", "source": "inline", "quote": "..."}],
  "candidates": {
    "enabled_param": "thinking.enabled",
    "effort_param": "reasoning_effort",
    "allowed_efforts": ["low", "medium", "high"]
  },
  "confidence": 0.0
}

url_extractor_subagent

Responsibilities:

• Fetch page content.
• Clean HTML and extract main body text + API example blocks.
• Keep source URL and evidence anchors.

Output shape:

{
  "source_type": "url",
  "evidence": [{"kind": "url", "source": "https://...", "quote": "..."}],
  "candidates": {
    "enabled_param": "enable_thinking",
    "effort_param": "thinking.effort",
    "allowed_efforts": ["high", "max"]
  },
  "confidence": 0.0
}

compare_subagent

Run only when both sources exist and conflict.

Responsibilities:

• Compare candidate parameters from URL and snippet evidence.
• Score and rank by:
  • credibility
  • completeness
  • executability
  • recency
  • model match
• Produce one merged final candidate and decision reason.

Output shape:

{
  "source_type": "merged",
  "decision": "url|snippet|hybrid",
  "scoring": {
    "url": {"credibility": 0.0, "completeness": 0.0, "executability": 0.0, "recency": 0.0, "model_match": 0.0},
    "snippet": {"credibility": 0.0, "completeness": 0.0, "executability": 0.0, "recency": 0.0, "model_match": 0.0}
  },
  "evidence": [{"kind": "url|snippet", "source": "...", "quote": "..."}],
  "candidates": {
    "enabled_param": "thinking.enabled",
    "effort_param": "reasoning_effort",
    "allowed_efforts": ["high", "max"]
  }
}

Tool Boundaries

• thinking_recipe_lookup: can be called to check existing global recipe.
• thinking_recipe_compile: can be called to turn evidence into preview recipe.
• thinking_recipe_save: do not call automatically from this skill.

Quality Gates

• Do not invent undocumented parameters.
• Mark uncertain fields explicitly.
• Every candidate field must have evidence.
• Preserve original source URL/snippet fragments for UI confirmation.