fix: strip <answer> XML tags before do/finish markers in _parse_response

[nankingjing]

Summary

Fixes an ordering bug in ModelClient._parse_response (phone_agent/model/client.py) that corrupts every action when the model answers in the tagged format requested by the system prompt.

The bug

The system prompt (phone_agent/config/prompts_zh.py / prompts_en.py) instructs the model to output:

<think>{think}</think>
<answer>{action}</answer>

where {action} is a do(...) or finish(...) call. _parse_response checked its rules in this order:

1. finish(message= marker
2. do(action= marker
3. <answer> XML tags (documented as the fallback)

Because a tagged answer always contains a do(action= or finish(message= call inside the <answer> block, Rule 1/2 match before the tags are stripped. For input:

<think>I should tap the search box</think><answer>do(action="Tap", element=[500, 300])</answer>

the old code returned:

• thinking = '<think>I should tap the search box</think><answer>' (tags leaked in)
• action = 'do(action="Tap", element=[500, 300])</answer>' (trailing </answer>)

The action string is not valid Python, so the downstream parse_action (phone_agent/actions/handler.py) raises ValueError. In PhoneAgent._execute_step that exception is swallowed and turned into finish(message=<raw string>), so the agent ends the task immediately instead of tapping. As a consequence, Rule 3 (the XML-tag path) was effectively dead code — it could never be reached for a real action.

The fix

Check for the <answer> tag first. When present, strip <think>/</think>/<answer>/</answer> and return the clean thinking/action pair. Tag-less responses (the AutoGLM-Phone-9B happy path) are unchanged: they fall through to the unchanged finish(message=/do(action= marker rules.

This is a pure reordering — no tag-less behaviour changes.

Tests

Adds tests/test_parse_response_order.py:

• Tagged do/finish responses now yield a clean action with no </answer> remnant (regression cases).
• Tag-less do, finish, finish-precedes-do, and plain-text fallback still behave exactly as before.

Verification

Not executed in CI here. Verified by reading the code path end-to-end and by simulating the reordered _parse_response against all test inputs (all produce the asserted (thinking, action) tuples). Both edited files pass ast.parse.

[nankingjing]

Reviewed — moving the tag check before the marker checks correctly prevents trailing from leaking into the action string. Good test coverage. Ready for review.

[nankingjing]

#378 and this PR fix the same bug: the <answer> tag check sits after the do(action= / finish(message= marker checks, so a tagged response leaves a trailing </answer> inside the action string, which then fails parse_action and aborts the step. Only one should merge, so here are the concrete differences.

This PR

• Reorders the three rules only. ModelClient._parse_response keeps its name and signature, so no caller breaks. +24/-14, no unrelated edits.
• Strips <think>/</think> from everything before <answer>, so the thinking comes back tag-free whether or not the closing tag is present.
• 104 lines of tests, 7 cases, covering the tagged path and the tag-less paths it must not regress.

#378

• Adds a debug print to the request path at phone_agent/model/client.py:147:

  print(f"{thinking=}, {action=}")

  That writes every parsed model response to stdout on every step.

• Moves _parse_response off the class to a module-level parse_response. Any caller using client._parse_response(...) breaks.

• Its rule 1 requires both tags (if "<answer>" in content and "</answer>" in content). With only the opening tag present it falls through to the do(action= marker, which leaves the raw tags in the thinking string — and its own last test pins that as the expected result:

  content = """<think>先思考</think>
  <answer>
  do(action="Launch", app="知乎")"""
  thinking, action = parse_response(content)
  assert thinking == "<think>先思考</think>\n<answer>"   # tags leak in here

  A truncated model response is exactly this shape, so that assertion locks the leak in rather than fixing it. This PR returns 先思考 for the same input.

One thing #378 does better: it truncates the action at the first closing tag (parts[1].split("</answer>", 1)[0]) instead of deleting all of them, which drops any trailing text after </answer>. I'll fold that into rule 1 here if there's interest.

I've left a note on #378 rather than closing it, since it isn't mine.