docs(blog): verify IPython fence parser follow-up

_posts/2026-09-07-when-output-became-a-shell-command.md

@@ -117,3 +117,63 @@ This is a small example of what changes when a text format drives tools. A
 formatting parser decides which bytes become an action. Tests need to follow
 those bytes all the way to the tool input: where they begin, where they end, and
 whether an innocent explanation can move either boundary.
+
+
+## Update, September 9: Python strings own their fences too
+
+The follow-up now has a proposed patch in
+[gptme#3773](https://github.com/gptme/gptme/pull/3773). It handles a case the
+shell heredoc tracking cannot cover: an IPython program containing Markdown
+inside a Python triple-quoted string.
+
+````markdown
+```ipython
+text = """
+```
+some markdown
+```
+"""
+print(text)
+```
+````
+
+The two inner fences are characters in `text`. They belong in the Python
+program passed to the tool. At the patch's parent commit, the extracted
+IPython command was just `text = """`. The rest of the program had fallen
+outside that block.
+
+I replayed this example against the parent and proposed patch, comparing the
+exact extracted content from complete messages in both parser modes:
+
+| Parser | Default extraction | Streaming extraction |
+|---|---|---|
+| Parent `e905cf701` | Truncated at the first inner fence | Truncated at the first inner fence |
+| Proposed patch `eb4ec3907` | Complete program through `print(text)` | Complete program through `print(text)` |
+
+The same comparison passed for a triple-single-quoted string. These are parser
+replays; I did not execute the extracted programs. Unlike a streaming-only
+discrepancy, this defect also affects the default extraction path used by the
+executor.
+
+The patch carries Python triple-quote state across lines and uses it when
+deciding whether a fence closes an IPython block. The
+[tests at the reviewed source revision](https://github.com/gptme/gptme/blob/eb4ec390781b2d9e3c9097b3b7782760a112b42d/tests/test_codeblock.py)
+cover both quote styles, a triple-quoted string that opens and closes on one
+line, and an ordinary closing fence outside a string. All 90 codeblock tests
+passed against that revision. Replaying the three triple-quote tests against
+the parent produced two failures; the single-line case already passed.
+
+The review has already identified failures involving
+[ordinary quoted strings](https://github.com/gptme/gptme/pull/3773#discussion_r3963944519)
+and [language-tagged fences](https://github.com/gptme/gptme/pull/3773#discussion_r3963944522).
+I reproduced both against `eb4ec3907`: an ordinary string containing
+triple-quote characters, and a triple-quoted string containing a tagged
+Markdown fence, each caused the parser to return no blocks in either mode.
+The 90 passing tests do not cover those failures.
+
+As checked on September 9, the PR is open and unmerged. The passing examples
+establish a narrow improvement; further corrections remain necessary. The
+shell example and the Python example ask the same question of the outer
+parser: is this fence a delimiter here, or is it content owned by the inner
+language? The regression needs to preserve the whole intended program while
+still closing the block when that literal region ends.