Parser fails when using step-3.7-flash via Chat Completions API

[li-xiu-qi]

Problem

When using cloud reasoning models (e.g. step-3.7-flash) via Chat Completions API, the parser fails on the first step:

ValueError: ui_action must contain 'action' or 'action_type'. Got keys: ['cot']

Root Cause

ask_llm_v2.py detects reasoning_content in the API response and wraps it as a <think> block, prepending it to the model's content:

result = "<think>" + reasoning + "</think>" + "\n" + llm_content

But the model's content already contains <THINK> tags (required by the parser prompt), so the final output has two layers:

<think>reasoning_content from API</think>       ← prepended by ask_llm_v2.py
<THINK>model's own thinking</THINK>             ← from model content
explain:...  action:CLICK  point:500,300  ...   ← parser needs this

The parser's str2action() uses split("</THINK>")[1] to extract key-value pairs, which grabs content after the first </THINK> — i.e., the second THINK block text, not the action parameters.

Three API Comparison

API	Reasoning Location	Two-layer THINK?
Chat Completions	reasoning_content field → prepended by ask_llm_v2.py	Yes
Messages (Anthropic)	thinking type block (separate)	No
Responses	reasoning type output item (separate)	No

The model's content always has exactly one <THINK> layer. The problem is purely in ask_llm_v2.py's handling.

Reproduction

1. Configure model_config.yaml with StepFun API:

stepfun:
    api_base: "https://api.stepfun.com/v1"
    api_key: "YOUR_API_KEY"

2. Set model_provider to stepfun and model_name to step-3.7-flash.

3. Run any task. First step fails with ValueError.

Fix

PR #64 — stop prepending reasoning_content in ask_llm_v2.py, return model's content as-is. Verified with real Android device.