[songji-tian]

Problem

When the server-side DSML -> tool_calls parser fails to convert the whole tool call, it can leave an untranslated tail in message.content and return no tool_calls. The turn then ends as a plain text reply: no tool runs, no error is raised, and the agent silently stalls.

The current guard only matches one exact form:

if content.rstrip().endswith('</｜DSML｜ calls>'):

Real leaks (14 stalled rolls out of 18 sampled agent transcripts, DeepSeek official endpoint) end with other forms, so the guard never fires:

leaked tail	current check
</｜DSML｜ calls>	fires
</｜｜DSML｜｜ parameter>	misses
</｜DSML｜ parameter>	misses
</｜DSML｜ invoke>	misses
`...parameter>
</code_run>`	misses

Measured on those 14 rolls: current check 0/14, this patch 14/14.

Change (1 line, net zero)

-        if content.rstrip().endswith('</｜DSML｜ calls>'):
+        if 'DSML' in content[-200:] and re.search(r'</[^>]*>\s*$', content.rstrip()):

The 'DSML' in content[-200:] guard keeps it conservative: the leak marker must still be present in the tail, so a normal answer that happens to end with </code_run> is not retried.

Verification

• 14 real stalled rolls: 0/14 before, 14/14 after.
• 54 normal text-only rolls / 1960 historical responses: 0 false positives.
• Retry is effective: 19/19 leaked turns returned a proper tool_calls on the next attempt.

Related: #804 (same leak, closed as an upstream serving-side issue).

Co-Authored-By: GenericAgent bot@gaagent.ai

[ZhulongNT]

Related: I ran into the same leak and opened #837 with a fix in #838 - it attacks the marker matching itself (variant-agnostic regex + a stateful filter that re-joins markers split across stream chunks) instead of only checking the content tail, so a leak anywhere in the turn and closing tags such as </||DSML|| parameter> are handled too. Your #836 covers the retry side; if you want, feel free to fold the two together.