coreinsight-cli 0.3.3__tar.gz → 0.3.4__tar.gz

{coreinsight_cli-0.3.3/coreinsight_cli.egg-info → coreinsight_cli-0.3.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: coreinsight-cli
-Version: 0.3.3
+Version: 0.3.4
 Summary: Local-first AI performance profiler that mathematically verifies optimizations for Python, C++, and CUDA
 Author: Varun Jani
 License: GPL-3.0-or-later

{coreinsight_cli-0.3.3 → coreinsight_cli-0.3.4}/coreinsight/analyzer.py

@@ -256,6 +256,13 @@ class AnalyzerAgent:
         hardware_target: str = "Generic CPU",
         stream_callback: Optional[Callable[[str], None]] = None,
     ) -> str:
+        from coreinsight.prompts import ModelTier
+        if language.lower() in ("cpp", "c++") and self.model_tier == ModelTier.SMALL:
+            raise ValueError(
+                f"C++ harness generation requires a MEDIUM or LARGE model "
+                f"(current tier: {self.model_tier}). "
+                "Switch to a 13B+ model (e.g. codellama:13b, mistral) for C++ support."
+            )
         try:
             context, original_code = _compress_for_small_model(
                 context, original_code, self.model_tier

{coreinsight_cli-0.3.3 → coreinsight_cli-0.3.4}/coreinsight/prompts/harness.py

@@ -76,6 +76,10 @@ CRITICAL TIMING:
 - Python: use `time.perf_counter()`. C++: use `std::chrono::high_resolution_clock`.
 - Clamp: `orig_time = max(end - start, 1e-9)` to prevent zero-division.
 - Speedup: `speedup = orig_time / opt_time`.
+- TIME BUDGET: each individual timing call MUST complete within 0.5 seconds.
+  If the original function at N=1000 takes more than 0.5s, cap N at 500 or 100.
+  If a measurement exceeds 0.5s, print a row of zeros and continue — do NOT hang.
+- HARD CAP: maximum N is 10,000. Never exceed this regardless of RAM available.
 
 ISOLATION RULES (CRITICAL):
 - This runs in an empty Docker container. NO local files exist.
@@ -151,6 +155,8 @@ REQUIREMENTS:
 7. Use high-resolution timers: Python → time.perf_counter(), C++ → std::chrono::high_resolution_clock.
 8. Clamp: orig_time = max(end - start, 1e-9) to prevent zero-division.
 9. Speedup = orig_time / opt_time.
+10. TIME BUDGET: each timing call MUST complete within 0.5 seconds. If N=1000 is too slow, use N=500 or N=100 instead. Print zeros and continue if a run hangs — do NOT exceed the budget.
+11. HARD CAP: maximum N is 10,000. Never exceed this.
 
 ISOLATION RULES (CRITICAL — runs in empty Docker container):
 - NO local imports. Define everything inline.
@@ -215,6 +221,7 @@ CRITICAL RULES:
 2. Generate realistic dummy args matching the function signature scaled to N.
 3. No internet. No pip installs.
 4. Print CSV header first. If any N crashes, print zeros and continue.
+5. TIME LIMIT: each timing call must finish in under 0.5s. Use smaller N if needed. Max N = 10,000.
 
 PYTHON STRUCTURE EXAMPLE (only if language=python):
     import time

{coreinsight_cli-0.3.3 → coreinsight_cli-0.3.4}/coreinsight/prompts/test_cases.py

@@ -37,8 +37,19 @@ Example (do NOT copy this — generate cases specific to {func_name}):
 
 CONSTRAINTS:
 - All values must be plain JSON types — no numpy arrays, no custom objects.
+- Do NOT instantiate any classes (no ClassName(...), no dataclasses, no enums). Use plain dicts instead.
 - If the function operates on a matrix, represent it as a list-of-lists.
 - If the function takes a size integer N, generate concrete data of that size inline.
 - Do NOT include function calls or expressions — only literal values.
 - Produce exactly {num_cases} test cases.
+- SANDBOX RULE: these test cases run in an isolated environment where project-specific
+  imports (custom classes, dataclasses, enums) have been stripped. If the function
+  accepts custom objects, pass plain dicts with the same field names instead of class
+  instances — e.g. {{"id": 1, "amount": 100.0}} not Transaction(id=1, amount=100.0).
+  The function code may use attribute access (obj.field) — dict args are fine for that
+  because the sandbox converts them via SimpleNamespace automatically.
+- IN-PLACE MUTATION: if the function mutates its input in-place (e.g. sorts a list,
+  fills an array) and returns None, add an optional "expected_mutation" key to the
+  test case object — a copy of the input argument in its expected post-call state.
+  The harness will compare the mutated arg against expected_mutation if present.
 """

{coreinsight_cli-0.3.3 → coreinsight_cli-0.3.4}/coreinsight/sandbox.py

@@ -39,6 +39,7 @@ import json, sys, math, importlib.util, traceback
 
 spec = importlib.util.spec_from_file_location("user_module", "/workspace/source.py")
 mod  = importlib.util.module_from_spec(spec)
+sys.modules["user_module"] = mod   # ← register BEFORE exec so @dataclass can resolve cls.__module__
 spec.loader.exec_module(mod)
 
 original_fn  = getattr(mod, "{original_func_name}",  None)

{coreinsight_cli-0.3.3 → coreinsight_cli-0.3.4/coreinsight_cli.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: coreinsight-cli
-Version: 0.3.3
+Version: 0.3.4
 Summary: Local-first AI performance profiler that mathematically verifies optimizations for Python, C++, and CUDA
 Author: Varun Jani
 License: GPL-3.0-or-later

{coreinsight_cli-0.3.3 → coreinsight_cli-0.3.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "coreinsight-cli"
-version = "0.3.3"
+version = "0.3.4"
 description = "Local-first AI performance profiler that mathematically verifies optimizations for Python, C++, and CUDA"
 license = {text = "GPL-3.0-or-later"}
 authors = [