drydock-cli 3.0.76__tar.gz → 3.0.78__tar.gz

{drydock_cli-3.0.76/drydock_cli.egg-info → drydock_cli-3.0.78}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 3.0.76
+Version: 3.0.78
 Summary: Drydock — a local, provider-agnostic terminal coding agent for local LLMs
 Author: Frank Bobe III
 License-Expression: Apache-2.0
@@ -59,6 +59,11 @@ A full agentic CLI harness — every tool below is clean-room and dependency-fre
 
 - **Files & shell** — `Read` (with a structure index for huge files), `Write`,
   `Edit`, `Bash`, `Glob`, `Grep`.
+- **Vision (multimodal)** — reference an image path in your message (a `.png`/
+  `.jpg` screenshot, mockup, or diagram) and it's attached for a vision-capable
+  model to *see*; the agent can also call the `ViewImage` tool to look at an
+  image it discovers on its own — describe a UI, read text off a screenshot,
+  debug a diagram. Works with any `--mmproj` server (e.g. llama.cpp + a vision model).
 - **Version control** — `GitStatus`, `GitDiff`, `GitLog`, `GitCommit`
   (structured + truncated; commit is local and reversible).
 - **Internet** — `WebSearch` + `WebFetch` (DuckDuckGo; offline-safe).

{drydock_cli-3.0.76 → drydock_cli-3.0.78}/README.md

@@ -35,6 +35,11 @@ A full agentic CLI harness — every tool below is clean-room and dependency-fre
 
 - **Files & shell** — `Read` (with a structure index for huge files), `Write`,
   `Edit`, `Bash`, `Glob`, `Grep`.
+- **Vision (multimodal)** — reference an image path in your message (a `.png`/
+  `.jpg` screenshot, mockup, or diagram) and it's attached for a vision-capable
+  model to *see*; the agent can also call the `ViewImage` tool to look at an
+  image it discovers on its own — describe a UI, read text off a screenshot,
+  debug a diagram. Works with any `--mmproj` server (e.g. llama.cpp + a vision model).
 - **Version control** — `GitStatus`, `GitDiff`, `GitLog`, `GitCommit`
   (structured + truncated; commit is local and reversible).
 - **Internet** — `WebSearch` + `WebFetch` (DuckDuckGo; offline-safe).

{drydock_cli-3.0.76 → drydock_cli-3.0.78}/drydock/providers.py

@@ -250,16 +250,11 @@ _IMAGE_MIME = {
 }
 
 
-def _user_content_with_images(content):
-    """Vision support: if the user's text references image file paths that exist
-    on disk, attach them as OpenAI multimodal image_url blocks (works with any
-    --mmproj-enabled server). Text-only prompts pass through unchanged as a plain
-    string, so display / loop-detection / compaction / token-counting (which all
-    assume string content) are untouched — the multimodal list is built ONLY here,
-    at the API boundary."""
+def detect_image_paths(content) -> list[str]:
+    """Image file paths referenced in a user message that exist on disk. Shared
+    by the multimodal attach (below) and the TUI (to confirm '📎 attached')."""
     if not isinstance(content, str) or "." not in content:
-        return content
-    import base64
+        return []
     import os
     import re
 
@@ -281,21 +276,50 @@ def _user_content_with_images(content):
         p = os.path.expanduser(raw)
         if os.path.isfile(p) and p not in seen:
             seen.append(p)
+    return seen
+
+
+_MAX_IMAGE_BYTES = 20_000_000  # don't base64 a >20MB file into a request
+
+
+def _image_url_block(path: str) -> dict | None:
+    """A base64 data-URL image_url block for an on-disk image, or None if it
+    can't be read / is too large."""
+    import base64
+    import os
+
+    try:
+        if os.path.getsize(path) > _MAX_IMAGE_BYTES:
+            return None
+        with open(path, "rb") as f:
+            b64 = base64.b64encode(f.read()).decode()
+    except OSError:
+        return None
+    mime = _IMAGE_MIME.get(os.path.splitext(path)[1].lower(), "image/png")
+    return {"type": "image_url", "image_url": {"url": f"data:{mime};base64,{b64}"}}
+
+
+def _content_with_images(content):
+    """Turn text that references on-disk image paths into a multimodal content
+    list ([text, image_url…]); text without resolvable images passes through
+    unchanged. Used for BOTH user messages (the user names an image) and
+    ViewImage tool results (the AGENT chose to view one) — built ONLY here at the
+    API boundary so display/compaction/token-counting still see plain strings."""
+    seen = detect_image_paths(content)
     if not seen:
         return content
     blocks: list[dict] = [{"type": "text", "text": content}]
     for p in seen:
-        mime = _IMAGE_MIME.get(os.path.splitext(p)[1].lower(), "image/png")
-        try:
-            with open(p, "rb") as f:
-                b64 = base64.b64encode(f.read()).decode()
-        except OSError:
-            continue
-        blocks.append({"type": "image_url",
-                       "image_url": {"url": f"data:{mime};base64,{b64}"}})
+        block = _image_url_block(p)
+        if block:
+            blocks.append(block)
     return blocks if len(blocks) > 1 else content
 
 
+# Back-compat alias (user-message vision).
+_user_content_with_images = _content_with_images
+
+
 def messages_to_openai(messages: list, system: str) -> list:
     """Convert neutral messages to OpenAI API format."""
     # value type is mixed (str content, multimodal list content, tool_calls list)
@@ -322,10 +346,17 @@ def messages_to_openai(messages: list, system: str) -> list:
                 ]
             result.append(msg)
         elif role == "tool":
+            content = m["content"]
+            # Agent-side vision: a ViewImage result names an image the agent chose
+            # to look at — attach it so the model SEES it (servers read images
+            # from tool messages too, verified). Only ViewImage, so an unrelated
+            # tool mentioning a .png path never balloons the request.
+            if m.get("name") == "ViewImage" and isinstance(content, str):
+                content = _content_with_images(content)
             result.append({
                 "role": "tool",
                 "tool_call_id": m["tool_call_id"],
-                "content": m["content"],
+                "content": content,
             })
     return result
 

{drydock_cli-3.0.76 → drydock_cli-3.0.78}/drydock/tools/__init__.py

@@ -83,6 +83,17 @@ SCHEMAS = [
             "required": ["file_path"],
         },
     },
+    {
+        "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.)",
+        "input_schema": {
+            "type": "object",
+            "properties": {
+                "path": {"type": "string", "description": "Path to the image file"},
+            },
+            "required": ["path"],
+        },
+    },
     {
         "name": "Write",
         "description": "Write content to a file, creating parent directories as needed.",
@@ -474,6 +485,33 @@ def _file_index(lines: list[str], fp: str) -> str:
     return f"{head}\nKey locations ({len(anchors)}):\n{body}"
 
 
+_IMAGE_EXTS = (".png", ".jpg", ".jpeg", ".gif", ".webp", ".bmp")
+_MAX_VIEW_IMAGE_BYTES = 20_000_000
+
+
+def tool_viewimage(params: dict, config: dict) -> str:
+    """Validate an image path and return a result that names it; the API boundary
+    (providers.messages_to_openai) attaches the actual image to THIS tool result
+    so the vision model sees it. Returns a plain error string on any problem
+    (never an attachable path) so a bad call can't try to attach nothing."""
+    raw = (params.get("path") or params.get("file_path") or "").strip()
+    if not raw:
+        return "Error: ViewImage needs a `path` to an image file."
+    fp = _resolve_path(raw, config)
+    if os.path.splitext(fp)[1].lower() not in _IMAGE_EXTS:
+        return (f"Error: {raw} is not a supported image. ViewImage handles "
+                f"{', '.join(_IMAGE_EXTS)}.")
+    if not os.path.isfile(fp):
+        return f"Error: no image file at {raw} (resolved to {fp})."
+    size = os.path.getsize(fp)
+    if size > _MAX_VIEW_IMAGE_BYTES:
+        return f"Error: {raw} is {size // 1_000_000}MB — too large to view (limit 20MB)."
+    kind = os.path.splitext(fp)[1].lstrip(".").upper()
+    # The absolute path in this text is what the API boundary attaches.
+    return (f"Loaded image {fp} ({kind}, {size // 1024 or 1}KB) — it is now visible "
+            "to you. Describe it, read any text in it, or use it to answer the task.")
+
+
 def tool_read(params: dict, config: dict) -> str:
     fp = _resolve_path(params["file_path"], config)
     limit = params.get("limit")  # None = caller didn't specify a window
@@ -1010,7 +1048,7 @@ def tool_todo(params: dict, config: dict) -> str:
 # Tools a sub-agent may use. Excludes Write/Edit/todo (no silent mutations the
 # parent can't see) and `task` itself (no recursion). Bash is included for real
 # exploration (ls/cat/grep/find/git log) and is bounded by the same bash_safety.
-SUBAGENT_TOOLS = ("Read", "Glob", "Grep", "Bash")
+SUBAGENT_TOOLS = ("Read", "ViewImage", "Glob", "Grep", "Bash")
 
 _SUBAGENT_SYSTEM = (
     "You are a focused exploration sub-agent inside Drydock. You have read-only "
@@ -1428,7 +1466,8 @@ def register_all():
     for schema in SCHEMAS:
         name = schema["name"]
         func = {
-            "Read": tool_read, "Write": tool_write, "Edit": tool_edit,
+            "Read": tool_read, "ViewImage": tool_viewimage,
+            "Write": tool_write, "Edit": tool_edit,
             "Bash": tool_bash, "Glob": tool_glob, "Grep": tool_grep,
             "todo": tool_todo, "task": tool_task, "Dispatch": tool_dispatch,
             "Knowledge": tool_knowledge,
@@ -1441,8 +1480,8 @@ def register_all():
         # Read-only w.r.t. the parent's files (GitStatus/Diff/Log inspect only;
         # GitCommit + GraphAdd write).
         read_only = name in (
-            "Read", "Glob", "Grep", "task", "Dispatch", "Knowledge", "GraphQuery",
-            "StigRules", "StigRule",
+            "Read", "ViewImage", "Glob", "Grep", "task", "Dispatch", "Knowledge",
+            "GraphQuery", "StigRules", "StigRule",
             "WebSearch", "WebFetch", "GitStatus", "GitDiff", "GitLog",
         )
         register(ToolDef(name=name, schema=schema, func=func, read_only=read_only))

{drydock_cli-3.0.76 → drydock_cli-3.0.78}/drydock/tui/app.py

@@ -409,6 +409,14 @@ class DrydockApp(App):
         # running, queue this one and drain it when the current turn finishes
         # instead of dropping it on the floor.
         self._mount(UserMessage(text))
+        # Confirm any image attachments so the user SEES vision is active (the
+        # actual attach happens at the API boundary in providers).
+        from drydock import providers
+        imgs = providers.detect_image_paths(text)
+        if imgs:
+            import os as _os
+            names = ", ".join(_os.path.basename(p) for p in imgs)
+            self._info(f"📎 attached {len(imgs)} image(s) for the model to see: {names}")
         if self._busy:
             self._queue.append(text)
             self._refresh_status()

{drydock_cli-3.0.76 → drydock_cli-3.0.78}/drydock/tuning.py

@@ -90,7 +90,13 @@ _DEFAULT_SYSTEM_PROMPT = (
     "start to lay out the plan, then START WORKING THROUGH IT IMMEDIATELY — do "
     "NOT stop to ask whether to proceed. Call `todo` again to flip a step to "
     "done and the next to in-progress as you finish each, and only stop when "
-    "every step is done. Don't write a TODO.md file for this."
+    "every step is done. Don't write a TODO.md file for this.\n"
+    "You have VISION: when the user references an image file path (.png/.jpg/"
+    ".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 "
+    "claim you cannot view images."
 )
 
 # Short, imperative prompt. Small local models do better with "act now" than
@@ -112,7 +118,13 @@ _GEMMA_SYSTEM_PROMPT = (
     "(one task per line, '[ ]'/'[~]'/'[x]'), then IMMEDIATELY start doing the "
     "steps — do NOT stop to ask whether to proceed. Call `todo` again to update "
     "it as you finish each step, and only stop when every step is done. Do not "
-    "write a TODO.md file for this."
+    "write a TODO.md file for this.\n"
+    "You have VISION: when the user's message references an image file path "
+    "(.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."
 )
 
 
@@ -221,11 +233,15 @@ _DRYDOCK_COMMANDS_HELP = (
     "- Custom skills (reusable `/<name>` prompts): `/skills new <name> <prompt "
     "text>` creates one (use $ARGS in the prompt for trailing input); `/skills` "
     "lists them; then they run it as `/<name>`.\n"
+    "- Images / vision: the user just references an image file path in their "
+    "message (a .png/.jpg screenshot, mockup, or diagram on disk) and it is "
+    "attached for you to SEE — no command needed.\n"
     "- Other: `/model` (model/endpoint), `/cwd`, `/undo` (revert last write), "
-    "`/back` (rewind a turn), `/compact` (shrink context), `/loop <n> <prompt>` "
-    "(repeat a prompt), `/mcp` (connected MCP servers), `/status`, `/clear`, "
-    "`/help`, `/quit`. Internet + git are tools you call yourself "
-    "(WebSearch/WebFetch, GitStatus/GitDiff/GitLog/GitCommit)."
+    "`/back` (rewind a turn), `/compact` (shrink context), `/context` (view/set "
+    "the context-window budget), `/loop <n> <prompt>` (repeat a prompt), `/mcp` "
+    "(connected MCP servers), `/rmf` & `/stig` (NIST 800-53 / DISA STIG), "
+    "`/status`, `/clear`, `/help`, `/quit`. Internet + git are tools you call "
+    "yourself (WebSearch/WebFetch, GitStatus/GitDiff/GitLog/GitCommit)."
 )
 
 

{drydock_cli-3.0.76 → drydock_cli-3.0.78/drydock_cli.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: drydock-cli
-Version: 3.0.76
+Version: 3.0.78
 Summary: Drydock — a local, provider-agnostic terminal coding agent for local LLMs
 Author: Frank Bobe III
 License-Expression: Apache-2.0
@@ -59,6 +59,11 @@ A full agentic CLI harness — every tool below is clean-room and dependency-fre
 
 - **Files & shell** — `Read` (with a structure index for huge files), `Write`,
   `Edit`, `Bash`, `Glob`, `Grep`.
+- **Vision (multimodal)** — reference an image path in your message (a `.png`/
+  `.jpg` screenshot, mockup, or diagram) and it's attached for a vision-capable
+  model to *see*; the agent can also call the `ViewImage` tool to look at an
+  image it discovers on its own — describe a UI, read text off a screenshot,
+  debug a diagram. Works with any `--mmproj` server (e.g. llama.cpp + a vision model).
 - **Version control** — `GitStatus`, `GitDiff`, `GitLog`, `GitCommit`
   (structured + truncated; commit is local and reversible).
 - **Internet** — `WebSearch` + `WebFetch` (DuckDuckGo; offline-safe).

{drydock_cli-3.0.76 → drydock_cli-3.0.78}/drydock_cli.egg-info/SOURCES.txt

@@ -91,6 +91,7 @@ tests/test_tool_arg_parsing.py
 tests/test_tools_undo.py
 tests/test_tui.py
 tests/test_tuning.py
+tests/test_viewimage.py
 tests/test_vision_input.py
 tests/test_web_tools.py
 tests/test_xccdf.py

{drydock_cli-3.0.76 → drydock_cli-3.0.78}/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 = "3.0.76"
+version = "3.0.78"
 description = "Drydock — a local, provider-agnostic terminal coding agent for local LLMs"
 readme = "README.md"
 requires-python = ">=3.11"

drydock_cli-3.0.78/tests/test_viewimage.py

@@ -0,0 +1,58 @@
+"""Agent-side vision: the ViewImage tool + its API-boundary attachment.
+The agent can choose to SEE an image it discovers (not just user-referenced
+ones); the image rides back on the ViewImage tool result."""
+from __future__ import annotations
+
+import struct
+import zlib
+
+import drydock.tools  # noqa: F401 — triggers tool registration
+from drydock import tool_registry as reg
+from drydock.providers import messages_to_openai
+
+
+def _png(path):
+    w = h = 8
+    raw = b"".join(b"\x00" + bytes([10, 20, 200]) * w for _ in range(h))
+    def ch(t, d):
+        c = t + d
+        return struct.pack(">I", len(d)) + c + struct.pack(">I", zlib.crc32(c) & 0xFFFFFFFF)
+    path.write_bytes(b"\x89PNG\r\n\x1a\n"
+                     + ch(b"IHDR", struct.pack(">IIBBBBB", w, h, 8, 2, 0, 0, 0))
+                     + ch(b"IDAT", zlib.compress(raw)) + ch(b"IEND", b""))
+    return str(path)
+
+
+def test_viewimage_validates(tmp_path):
+    _png(tmp_path / "a.png")
+    cfg = {"cwd": str(tmp_path)}
+    assert "now visible to you" in reg.execute("ViewImage", {"path": "a.png"}, cfg)
+    assert "no image file" in reg.execute("ViewImage", {"path": "missing.png"}, cfg)
+    assert "not a supported image" in reg.execute("ViewImage", {"path": str(tmp_path / "x.txt")}, cfg)
+    assert "needs a `path`" in reg.execute("ViewImage", {}, cfg)
+
+
+def test_viewimage_is_read_only_and_subagent_visible():
+    td = reg.get("ViewImage")
+    assert td.read_only is True
+    from drydock.tools import SUBAGENT_TOOLS
+    assert "ViewImage" in SUBAGENT_TOOLS
+
+
+def test_viewimage_result_becomes_multimodal(tmp_path):
+    _png(tmp_path / "shot.png")
+    cfg = {"cwd": str(tmp_path)}
+    res = reg.execute("ViewImage", {"path": "shot.png"}, cfg)
+    out = messages_to_openai([{"role": "tool", "tool_call_id": "c1",
+                               "name": "ViewImage", "content": res}], "sys")
+    content = out[-1]["content"]
+    assert isinstance(content, list)
+    assert any(b.get("type") == "image_url" for b in content)
+    assert content[0]["type"] == "text"
+
+
+def test_non_viewimage_tool_with_png_path_stays_text(tmp_path):
+    img = _png(tmp_path / "shot.png")
+    out = messages_to_openai([{"role": "tool", "tool_call_id": "c2",
+                               "name": "Grep", "content": f"found {img} in code"}], "sys")
+    assert isinstance(out[-1]["content"], str)   # no accidental image ballooning

{drydock_cli-3.0.76 → drydock_cli-3.0.78}/tests/test_vision_input.py

@@ -82,3 +82,23 @@ def test_image_path_in_backticks_and_punctuation(tmp_path):
         out = _user_content_with_images(prompt)
         assert isinstance(out, list), f"not attached: {prompt!r}"
         assert any(b.get("type") == "image_url" for b in out), prompt
+
+
+def test_detect_image_paths_shared_helper(tmp_path):
+    from drydock.providers import detect_image_paths
+    img = _write_png(tmp_path / "shot.png")
+    # plain, quoted, backticked, and trailing-punctuation forms all resolve
+    assert detect_image_paths(f"look at {img}") == [img]
+    assert detect_image_paths(f'see "{img}".') == [img]
+    assert detect_image_paths(f"debug `{img}`,") == [img]
+    # nonexistent path or text-only → nothing attached
+    assert detect_image_paths(f"{tmp_path}/nope.png") == []
+    assert detect_image_paths("no images here, just prose") == []
+    # de-dupes repeats
+    assert detect_image_paths(f"{img} and again {img}") == [img]
+
+
+def test_vision_in_system_prompt():
+    from drydock.tuning import system_prompt_for_model
+    assert "VISION" in system_prompt_for_model("gemma4")
+    assert "VISION" in system_prompt_for_model("some-other-model")