fusiontest 0.2.2__tar.gz → 0.2.3__tar.gz

{fusiontest-0.2.2 → fusiontest-0.2.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: fusiontest
-Version: 0.2.2
+Version: 0.2.3
 Summary: AI-powered UI testing for mobile and desktop — by FusionLeap.io
 Author-email: FusionLeap <hello@fusionleap.io>
 License: MIT

{fusiontest-0.2.2 → fusiontest-0.2.3}/fusiontest/core/action_model.py

@@ -210,6 +210,52 @@ GEMINI_DEFAULT_MODEL = "gemini-2.5-flash"
 LLM_MAX_OUTPUT_TOKENS = 1024
 
 
+# Page text the action model sees per step (ADR-007 cost work). It needs every
+# clickable element but only the gist of the page — the verifier, which judges
+# content, gets the full page. Measured on fusionleap.io, page text was ~65% of
+# each ~2.6k-token step prompt.
+_ACTION_CONTENT_CHARS = 1500
+_ACTION_LABEL_CHARS = 80
+_ACTION_ELEMENT_CHARS = 120
+_CONTENT_LINE_RE = re.compile(r'^\s*(\w+): "(.*)"$')
+_ELEMENT_LINE_RE = re.compile(r'^\s*\[\d+\] ')
+
+
+def compact_screen_for_action(screen_text: str) -> str:
+    """Trim a screen's page-content section for the per-step action prompt.
+
+    Every interactive element is kept (the model can only act on what it sees),
+    but very long labels — e.g. list items holding whole paragraphs — are cut:
+    the model acts by index. In the page content, headings are kept whole,
+    other text is cut to 80 characters, and the section stops at ~1,500 characters.
+    """
+    lines: list[str] = []
+    in_content = False
+    used = 0
+    for line in screen_text.splitlines():
+        if line.startswith("-- "):
+            in_content = line.startswith("-- Page content")
+            lines.append(line)
+            continue
+        if not in_content:
+            if len(line) > _ACTION_ELEMENT_CHARS and _ELEMENT_LINE_RE.match(line):
+                line = line[:_ACTION_ELEMENT_CHARS] + '…"'
+            lines.append(line)
+            continue
+        match = _CONTENT_LINE_RE.match(line)
+        if match:
+            kind, label = match.groups()
+            if kind != "heading" and len(label) > _ACTION_LABEL_CHARS:
+                label = label[:_ACTION_LABEL_CHARS] + "…"
+            line = f'  {kind}: "{label}"'
+        if used + len(line) > _ACTION_CONTENT_CHARS:
+            lines.append("  … (more page text omitted for brevity)")
+            break
+        used += len(line)
+        lines.append(line)
+    return "\n".join(lines)
+
+
 def openai_compat_extra_args(model: str) -> dict:
     """Extra chat.completions args for OpenAI-compatible models (Groq/OpenAI/Ollama)."""
     if model.startswith("openai/gpt-oss"):
@@ -321,6 +367,10 @@ class LLMRequest:
     json_object: bool = False   # OpenAI-compatible response_format={"type": "json_object"}
     empty: str = ""             # returned when the model produces no text
     max_output_tokens: int = LLM_MAX_OUTPUT_TOKENS  # includes hidden reasoning/thinking
+    # Let reasoning models think before answering. Off for per-step action and
+    # verifier calls (short answers; thinking tokens are billed as output and
+    # add latency); on for open-ended generation like goal discovery.
+    thinking: bool = False
 
 
 class LLMCaller:
@@ -609,6 +659,11 @@ class LLMCaller:
                     system_instruction=request.system,
                     temperature=self.temperature,
                     max_output_tokens=request.max_output_tokens,
+                    # Flash models can skip thinking entirely; Pro models can't.
+                    thinking_config=(
+                        genai_types.ThinkingConfig(thinking_budget=0)
+                        if not request.thinking and "flash" in resolved_model else None
+                    ),
                 ),
             )
             meta = getattr(resp, "usage_metadata", None)
@@ -820,7 +875,7 @@ class ActionModel(LLMCaller):
             pruned_invalid = invalid_actions[-self.max_invalid_actions_in_prompt:]
             sections += ["", "ACTIONS THAT FAILED (do NOT repeat):", *[f"  - {a}" for a in pruned_invalid]]
 
-        trimmed_screen = screen_text
+        trimmed_screen = compact_screen_for_action(screen_text)
         if self.max_screen_chars and len(trimmed_screen) > self.max_screen_chars:
             # Truncate at the last newline before the limit to avoid cutting an
             # element in half. Interactive elements are listed first, so only

{fusiontest-0.2.2 → fusiontest-0.2.3}/fusiontest/core/runner.py

@@ -60,7 +60,8 @@ class StepResult:
     screenshots: list[bytes] = field(default_factory=list)
     error: str = ""
     duration_seconds: float = 0.0
-    token_usage: TokenUsage = field(default_factory=TokenUsage)
+    token_usage: TokenUsage = field(default_factory=TokenUsage)       # all LLM calls
+    verifier_tokens: TokenUsage = field(default_factory=TokenUsage)   # the verifier's share
     # The goal could not be evaluated (LLM capacity / infrastructure). Not a
     # test result: excluded from stability and reported as an error (ADR-007).
     infra_error: bool = False
@@ -80,6 +81,10 @@ class StepResult:
             "error":            self.error,
             "duration_seconds": self.duration_seconds,
             "token_usage":      self.token_usage.to_dict(),
+            "token_breakdown":  {
+                "action":   (self.token_usage - self.verifier_tokens).to_dict(),
+                "verifier": self.verifier_tokens.to_dict(),
+            },
             "mode":             self.mode,
         }
 
@@ -506,7 +511,7 @@ class FusionTestRunner:
                     action_str = _redact(f"done: {action.reasoning}")
                     step.actions.append(action_str)
                     verified = self.verifier.is_complete(goal, screen_text, history)
-                    step.token_usage += getattr(self.verifier, "last_token_usage", TokenUsage())
+                    self._add_verifier_tokens(step)
                     step.success = verified
                     if verified:
                         self._last_recording = self._make_recording(recorded, screen_text)
@@ -607,7 +612,7 @@ class FusionTestRunner:
                     f"  Goal verified complete at max_steps "
                     f"({self.config.max_steps_per_goal}) — last-chance check passed"
                 )
-            step.token_usage += getattr(self.verifier, "last_token_usage", TokenUsage())
+            self._add_verifier_tokens(step)
 
         if not step.success and self.config.screenshot_on_failure:
             try:
@@ -744,6 +749,12 @@ class FusionTestRunner:
                 )
         return replace(action, value=resolver.resolve(action.value)), None
 
+    def _add_verifier_tokens(self, step: StepResult) -> None:
+        usage = getattr(self.verifier, "last_token_usage", None)
+        if isinstance(usage, TokenUsage):
+            step.token_usage += usage
+            step.verifier_tokens += usage
+
     def _write_step_log(
         self, goal: str, screen_text: str, action: str, outcome: str
     ) -> None:

{fusiontest-0.2.2 → fusiontest-0.2.3}/fusiontest/core/tokens.py

@@ -25,6 +25,12 @@ class TokenUsage:
             output_tokens=self.output_tokens + other.output_tokens,
         )
 
+    def __sub__(self, other: TokenUsage) -> TokenUsage:
+        return TokenUsage(
+            input_tokens=self.input_tokens - other.input_tokens,
+            output_tokens=self.output_tokens - other.output_tokens,
+        )
+
     def __iadd__(self, other: TokenUsage) -> TokenUsage:
         self.input_tokens += other.input_tokens
         self.output_tokens += other.output_tokens

{fusiontest-0.2.2 → fusiontest-0.2.3}/fusiontest/discovery/goal_generator.py

@@ -147,6 +147,7 @@ _REQUEST = LLMRequest(
     empty="{}",
     # 20 goals of JSON plus the model's hidden reasoning/thinking.
     max_output_tokens=8192,
+    thinking=True,   # open-ended generation benefits from thinking
 )
 
 

{fusiontest-0.2.2 → fusiontest-0.2.3}/fusiontest.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: fusiontest
-Version: 0.2.2
+Version: 0.2.3
 Summary: AI-powered UI testing for mobile and desktop — by FusionLeap.io
 Author-email: FusionLeap <hello@fusionleap.io>
 License: MIT

{fusiontest-0.2.2 → fusiontest-0.2.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "fusiontest"
-version = "0.2.2"
+version = "0.2.3"
 description = "AI-powered UI testing for mobile and desktop — by FusionLeap.io"
 readme = "README.md"
 requires-python = ">=3.10"