drydock-cli 3.0.87__tar.gz → 3.0.88__tar.gz

{drydock_cli-3.0.87/drydock_cli.egg-info → drydock_cli-3.0.88}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 3.0.87
+Version: 3.0.88
 Summary: Drydock — a local, provider-agnostic terminal coding agent for local LLMs
 Author: Frank Bobe III
 License-Expression: Apache-2.0

{drydock_cli-3.0.87 → drydock_cli-3.0.88}/drydock/tools/__init__.py

@@ -85,7 +85,7 @@ SCHEMAS = [
     },
     {
         "name": "ViewImage",
-        "description": "Look at an image file with your vision — use this when you need to SEE a screenshot, mockup, diagram, photo, or rendered output (.png/.jpg/.jpeg/.gif/.webp/.bmp). The image is shown to you so you can describe it, read text from it, or debug it. (Reading an image with the Read tool gives binary garbage; use ViewImage instead.)",
+        "description": "Look at an image file with your vision — use this when you need to SEE a screenshot, mockup, diagram, photo, or rendered output (.png/.jpg/.jpeg/.gif/.webp/.bmp). The image is shown to you so you can describe it, read text from it, or debug it. To READ TEXT OR DATA out of an image — a scanned document, invoice, receipt, form, or a screenshot with text — use ViewImage FIRST (you can read it directly); prefer it over OCR tools like tesseract/pdftotext, which are often unreliable or not installed. (Reading an image with the Read tool gives binary garbage; use ViewImage instead.)",
         "input_schema": {
             "type": "object",
             "properties": {

{drydock_cli-3.0.87 → drydock_cli-3.0.88}/drydock/tuning.py

@@ -95,7 +95,9 @@ _DEFAULT_SYSTEM_PROMPT = (
     ".jpeg/.gif/.webp/.bmp on disk — a screenshot, mockup, diagram, or photo), it "
     "is attached and you can SEE it; you can also call the `ViewImage` tool to "
     "look at any image you find yourself (Read gives binary garbage for images — "
-    "use ViewImage). Describe it, read text from it, or debug with it — never "
+    "use ViewImage). To read text/data out of a scanned document, invoice, "
+    "receipt, or screenshot, use ViewImage directly rather than OCR tools like "
+    "tesseract. Describe it, read text from it, or debug with it — never "
     "claim you cannot view images."
 )
 
@@ -123,8 +125,10 @@ _GEMMA_SYSTEM_PROMPT = (
     "(.png/.jpg/.jpeg/.gif/.webp/.bmp that exists on disk — a screenshot, mockup, "
     "diagram, or photo), that image is attached and you can SEE it. You can also "
     "call the `ViewImage` tool to look at any image you find yourself (Read gives "
-    "binary garbage for images — use ViewImage). Describe it, read text from it, "
-    "or use it to debug — never claim you cannot view images."
+    "binary garbage for images — use ViewImage). To read TEXT or DATA out of a "
+    "scanned document, invoice, receipt, form, or a screenshot, use ViewImage "
+    "directly — don't reach for OCR tools like tesseract first. Describe it, read "
+    "text from it, or use it to debug — never claim you cannot view images."
 )
 
 

{drydock_cli-3.0.87 → drydock_cli-3.0.88/drydock_cli.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 3.0.87
+Version: 3.0.88
 Summary: Drydock — a local, provider-agnostic terminal coding agent for local LLMs
 Author: Frank Bobe III
 License-Expression: Apache-2.0

{drydock_cli-3.0.87 → drydock_cli-3.0.88}/pyproject.toml

@@ -7,7 +7,7 @@ build-backend = "setuptools.build_meta"
 # PyPI distribution name is drydock-cli (the established install name, continued
 # from the retired v2 fork); the import package + CLI command stay `drydock`.
 name = "drydock-cli"
-version = "3.0.87"
+version = "3.0.88"
 description = "Drydock — a local, provider-agnostic terminal coding agent for local LLMs"
 readme = "README.md"
 requires-python = ">=3.11"

{drydock_cli-3.0.87 → drydock_cli-3.0.88}/tests/test_viewimage.py

@@ -74,3 +74,15 @@ def test_image_load_error_classifier():
     assert is_image_load_error("invalid_request_error: could not decode image")
     assert not is_image_load_error("rate limit exceeded")
     assert not is_image_load_error("context_length_exceeded")
+
+
+def test_viewimage_nudges_over_ocr_for_documents():
+    """The tool description + both system prompts steer the model to ViewImage
+    (its own vision) for reading text off invoices/receipts/scans, instead of
+    reaching for OCR tools first (observed on the financial-doc tbench task)."""
+    from drydock.tuning import system_prompt_for_model
+    desc = reg.get("ViewImage").schema["description"]
+    assert "OCR" in desc and ("invoice" in desc or "receipt" in desc)
+    for model in ("gemma4", "some-other-model"):
+        p = system_prompt_for_model(model)
+        assert "OCR" in p and "ViewImage" in p