gemx 0.2.4__tar.gz → 0.2.5__tar.gz

{gemx-0.2.4 → gemx-0.2.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: gemx
-Version: 0.2.4
+Version: 0.2.5
 Summary: Drive the Gemini web UI from Python via Playwright
 Project-URL: Homepage, https://github.com/ra0x3/gemx
 Project-URL: Repository, https://github.com/ra0x3/gemx
@@ -26,13 +26,12 @@ Description-Content-Type: text/markdown
 `gemini.google.com` as if it were an API, using Playwright to enter a prompt,
 submit it, and capture the structured reply.
 
-It exists because the obvious approaches don't work: Gemini's editor is
-[Quill](https://quilljs.com/), which keeps its own document model and ignores
-DOM surgery, Playwright `fill()`, and synthetic input events — those leave the
-model empty and the turn errors with *"I encountered an error."* Gemx injects
-text via `execCommand('insertText')` (the same trusted input pipeline a manual
-paste uses), reads the reply from `message-content .markdown`, and retains the
-*peak* streamed text because Gemini re-mounts the response node mid-stream.
+Gemx enters prompts using the input routine verified by the browser-console
+recovery script: it replaces the editor contents with a paragraph and dispatches
+a bubbling `input` event before clicking Send. It reads replies from
+`message-content .markdown` and retains the *peak* streamed text because Gemini
+can re-mount the response node mid-stream. Responses must settle for six seconds
+by default; JSON and XML must also parse before Gemx returns them.
 
 ## Install
 

{gemx-0.2.4 → gemx-0.2.5}/README.md

@@ -4,13 +4,12 @@
 `gemini.google.com` as if it were an API, using Playwright to enter a prompt,
 submit it, and capture the structured reply.
 
-It exists because the obvious approaches don't work: Gemini's editor is
-[Quill](https://quilljs.com/), which keeps its own document model and ignores
-DOM surgery, Playwright `fill()`, and synthetic input events — those leave the
-model empty and the turn errors with *"I encountered an error."* Gemx injects
-text via `execCommand('insertText')` (the same trusted input pipeline a manual
-paste uses), reads the reply from `message-content .markdown`, and retains the
-*peak* streamed text because Gemini re-mounts the response node mid-stream.
+Gemx enters prompts using the input routine verified by the browser-console
+recovery script: it replaces the editor contents with a paragraph and dispatches
+a bubbling `input` event before clicking Send. It reads replies from
+`message-content .markdown` and retains the *peak* streamed text because Gemini
+can re-mount the response node mid-stream. Responses must settle for six seconds
+by default; JSON and XML must also parse before Gemx returns them.
 
 ## Install
 

{gemx-0.2.4 → gemx-0.2.5}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "gemx"
-version = "0.2.4"
+version = "0.2.5"
 description = "Drive the Gemini web UI from Python via Playwright"
 readme = "README.md"
 requires-python = ">=3.12"

{gemx-0.2.4 → gemx-0.2.5}/src/gemx/__init__.py

@@ -28,4 +28,4 @@ __all__ = [
     "parse_output",
 ]
 
-__version__ = "0.2.4"
+__version__ = "0.2.5"

{gemx-0.2.4 → gemx-0.2.5}/src/gemx/client.py

@@ -3,11 +3,9 @@
 This is a formalized version of a hack that drives ``gemini.google.com/app`` as
 if it were an API. The non-obvious parts (and why this exists):
 
-* **Input** must go through ``document.execCommand('insertText')``. Gemini's
-  editor is Quill, which keeps its own document model and ignores DOM surgery,
-  Playwright ``fill()``, and synthetic ``InputEvent``s — those leave the model
-  empty and the turn errors with "I encountered an error". ``execCommand`` fires
-  the trusted ``beforeinput``/``input`` pair Quill honors, exactly like a paste.
+* **Input** follows the browser-console recovery script: replace the editor's
+  contents with a paragraph and dispatch a bubbling ``input`` event before
+  clicking Send.
 * **The response** lives in ``message-content .markdown`` (not a bare
   ``.markdown``), and Gemini re-mounts / empties that node after streaming, so we
   retain the *peak* text seen rather than reading the DOM once at the end.
@@ -109,13 +107,13 @@ _INSERT_TEXT_JS = """(text) => {
     const editor = document.querySelector('.ql-editor[contenteditable="true"]');
     if (!editor) return false;
     editor.focus();
-    const sel = window.getSelection();
-    sel.removeAllRanges();
-    const range = document.createRange();
-    range.selectNodeContents(editor);
-    range.collapse(false);
-    sel.addRange(range);
-    return document.execCommand('insertText', false, text);
+    editor.classList.remove('ql-blank');
+    editor.textContent = '';
+    const p = document.createElement('p');
+    p.textContent = text;
+    editor.appendChild(p);
+    editor.dispatchEvent(new Event('input', { bubbles: true }));
+    return true;
 }"""
 
 _INSTALL_RESPONSE_OBSERVER_JS = """(selectors) => {
@@ -297,9 +295,8 @@ class GemxConfig:  # pylint: disable=too-many-instance-attributes
     scroll_input_into_view: bool = True
     # The peak response must exceed this length and stay non-growing (or be
     # emptied) for this many poll ticks before it is accepted as complete.
-    # Matches the original bot/gemini.py driver (stable_count >= 2).
     min_response_chars: int = 50
-    stable_ticks: int = 2
+    stable_ticks: int = 3
     # How often (seconds) to emit page diagnostics while waiting for the response
     # node to appear.
     diagnostics_interval_s: int = 10

{gemx-0.2.4 → gemx-0.2.5}/src/gemx/formats.py

@@ -40,13 +40,13 @@ def format_instruction(fmt: OutputFormat) -> str:
     """Return the prompt instruction that asks Gemini for ``fmt`` output."""
     if fmt is OutputFormat.JSON:
         return (
-            "Respond with a single valid JSON object only. Do not include prose, "
-            "markdown fences, or commentary outside the JSON."
+            "Respond with exactly one ```json fenced code block holding a single "
+            "valid JSON object, and nothing outside the block."
         )
     if fmt is OutputFormat.XML:
         return (
-            "Respond with a single well-formed XML document only. Do not include "
-            "prose, markdown fences, or commentary outside the XML."
+            "Respond with exactly one ```xml fenced code block holding a single "
+            "well-formed XML document, and nothing outside the block."
         )
     return (
         "Respond with plain text only. Do not include markdown fences, JSON, or "
@@ -140,8 +140,17 @@ def parse_output(text: str, fmt: OutputFormat) -> Any:
             end = repaired.rfind("}") + 1
             candidate = repaired[start:end] if start != -1 and end != 0 else repaired
         try:
-            return json.loads(candidate)
+            return json.loads(candidate, strict=False)
         except json.JSONDecodeError as exc:
+            restarts = list(
+                re.finditer(r"```json[ \t]*\r?\n", text, flags=re.IGNORECASE)
+            )
+            if restarts:
+                candidate = text[restarts[-1].end() :].strip().removesuffix("```").rstrip()
+                try:
+                    return json.loads(_clean_json_text(candidate), strict=False)
+                except json.JSONDecodeError:
+                    pass
             raise ValueError(f"Could not parse JSON from response: {exc}") from exc
 
     start = cleaned.find("<")