localcode 0.3.17__tar.gz → 0.3.19__tar.gz

{localcode-0.3.17/src/localcode.egg-info → localcode-0.3.19}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: localcode
-Version: 0.3.17
+Version: 0.3.19
 Summary: High-performance AI coding on consumer hardware.
 Author: LocalCode contributors
 License-Expression: Apache-2.0

{localcode-0.3.17 → localcode-0.3.19}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "localcode"
-version = "0.3.17"
+version = "0.3.19"
 description = "High-performance AI coding on consumer hardware."
 readme = "README.md"
 requires-python = ">=3.10"

{localcode-0.3.17 → localcode-0.3.19}/src/localcode/__init__.py

@@ -1,3 +1,3 @@
 __all__ = ["__version__"]
 
-__version__ = "0.3.17"
+__version__ = "0.3.19"

{localcode-0.3.17 → localcode-0.3.19}/src/localcode/agent/loop.py

@@ -361,6 +361,31 @@ def run_agent_loop(
     if not messages or messages[-1].get("content") != user_text:
         messages.append({"role": "user", "content": user_text})
 
+    # Attach any images pasted in the TUI to THIS turn's user message.
+    # app.ask()/compose_messages is text-only, so the base64 images ride
+    # on the app object (stashed by the chat screen). We rebuild the
+    # matching user message here — where the loop owns the message list —
+    # via the composer's OpenAI-compatible image_url parts. Cleared after
+    # so images send exactly once, and only when the model can see them.
+    _pending_imgs = getattr(app, "_pending_images", None)
+    if _pending_imgs:
+        app._pending_images = []
+        _profile = getattr(app, "profile", None)
+        if _profile is not None and getattr(_profile, "supports_vision", False):
+            try:
+                from ..composer import _build_user_message
+                _provider = getattr(getattr(app, "config", None).runtime, "provider", "llama_cpp")
+                for _i in range(len(messages) - 1, -1, -1):
+                    if (messages[_i].get("role") == "user"
+                            and messages[_i].get("content") == user_text):
+                        messages[_i] = _build_user_message(
+                            user_text, list(_pending_imgs), _profile, _provider
+                        )
+                        break
+            except Exception:
+                # Never let image plumbing break a normal text turn.
+                pass
+
     full_response: list[str] = []
     start_time = time.time()
     tools_called: list[str] = []
@@ -1301,7 +1326,7 @@ def run_agent_loop(
                 try:
                     from ..tools.project_check import run_project_check
                     out.print_info("Verifying — running the project's typecheck…")
-                    _proj_errors = run_project_check(str(app.repo_root))
+                    _proj_errors = run_project_check(str(app.repo_root), ctx_tokens=_ctx_tokens_turn)
                 except Exception:
                     _proj_errors = None
                 try:

{localcode-0.3.17 → localcode-0.3.19}/src/localcode/agent/prompts.py

@@ -54,7 +54,10 @@ Plan & tools:
 - For files and dirs use list_files/read_file/write_file/edit_file — NOT bash (`ls`, `cat`, `cat >`, `>`). bash is only for running commands (npm, git, build, test).
 - edit_file for existing files: anchor `old_string` on 2–4 adjacent lines (matching is whitespace-tolerant; the leading `<n>\\t` from read_file is stripped for you). write_file to create or fully rewrite.
 - On a tool error, read it, fix the specific cause, and retry — don't give up after one failure, and don't repeat the same failing call.
-- New project → prefer a small multi-file layout with a thin entrypoint. Write code valid for the file's real language and match the project's existing conventions.
+- New project → prefer a small multi-file layout with a thin entrypoint. Write code valid for the file's real language and match the project's existing conventions. Put source under `src/`, config at the repo root.
+- Define the shared types / data model FIRST, then write everything against them. Don't write pages/UI before their types exist — naming mismatches force a full rewrite later.
+- Don't invent dependencies: before importing a library confirm it's in package.json (or requirements.txt/Cargo.toml) or a neighboring file; when installing, don't pin a guessed version — `npm install <pkg>` lets the resolver pick a real one.
+- Never assume a library's API. Before using a package's exports, types, or signatures, verify them — read its files in node_modules, check how neighboring files use it, or web_fetch its docs. Don't guess names. Port known algorithms from the real reference (web_fetch), not from memory.
 
 Runtime facts:
 - bash returns an exit code; non-zero = failure. Background long-running commands (`cmd &`) so bash returns.

{localcode-0.3.17 → localcode-0.3.19}/src/localcode/config.py

@@ -45,6 +45,7 @@ llama_cpp_lookup_cache = false
 kv_cache_type_k = "q8_0"
 kv_cache_type_v = "turbo4"
 llama_cpp_binary = ""
+vision_enabled = false
 temperature = 1.0
 max_context_chars = 200000
 request_timeout_seconds = 600
@@ -116,6 +117,11 @@ class RuntimeConfig:
     diffusion_cli_binary: str = ""         # llama-diffusion-cli path (DiffusionGemma runner; built on demand)
     cohere_server_binary: str = ""         # llama-server with cohere2moe (North-Mini-Code; built on demand from PR #24260)
     model_dir: str = ""                    # directory where GGUFs download to (blank → ~/.local/share/localcode/models)
+    # Vision toggle. Tracks whether the multimodal projector should be
+    # loaded (--mmproj) alongside the text decoder. Persisted so turning
+    # vision OFF just relaunches WITHOUT --mmproj (frees RAM) while KEEPING
+    # the projector file on disk — re-enabling is instant, no re-download.
+    vision_enabled: bool = False
     temperature: float = 1.0  # Unsloth's official Gemma 4 recommendation — prevents IQ3_S mode-collapse loops
     max_context_chars: int = 200000
     request_timeout_seconds: int = 600
@@ -241,6 +247,7 @@ def save_config(config: AppConfig) -> Path:
         f'diffusion_cli_binary = "{config.runtime.diffusion_cli_binary}"\n'
         f'cohere_server_binary = "{config.runtime.cohere_server_binary}"\n'
         f'model_dir = "{config.runtime.model_dir}"\n'
+        f"vision_enabled = {'true' if config.runtime.vision_enabled else 'false'}\n"
         f"temperature = {config.runtime.temperature}\n"
         f"max_context_chars = {config.runtime.max_context_chars}\n"
         f"request_timeout_seconds = {config.runtime.request_timeout_seconds}\n"
@@ -321,6 +328,7 @@ def load_config() -> AppConfig:
         diffusion_cli_binary=os.environ.get("LOCALCODE_DIFFUSION_CLI_BINARY", runtime_data.get("diffusion_cli_binary", "")),
         cohere_server_binary=os.environ.get("LOCALCODE_COHERE_SERVER_BINARY", runtime_data.get("cohere_server_binary", "")),
         model_dir=os.environ.get("LOCALCODE_MODEL_DIR", runtime_data.get("model_dir", "")),
+        vision_enabled=str(os.environ.get("LOCALCODE_VISION_ENABLED", runtime_data.get("vision_enabled", False))).lower() in {"1", "true", "yes", "on"},
         # `llama_cpp_spec_type` migration (2026-04-26): old configs
         # auto-saved "ngram-mod" or "ngram-simple" from a benchmark
         # preset. N-gram speculative decoding looks up matching

{localcode-0.3.17 → localcode-0.3.19}/src/localcode/runtime.py

@@ -410,17 +410,21 @@ class LocalCodeRuntimeGateway(_DiffusionMixin):
             "--threads", str(threads),
             "--flash-attn", "on" if flash_attn_on else "off",
         ]
-        # Multimodal projector — if the currently-selected catalog entry
-        # has an mmproj sidecar AND the sidecar is on disk, pass --mmproj
-        # so llama-server loads the vision encoder alongside the text
-        # decoder. We don't auto-download here; the chat layer prompts
-        # the user on first vision use (see voice.py / vision flow).
+        # Multimodal projector — pass --mmproj only when vision is
+        # ENABLED (config flag) AND the currently-selected catalog entry
+        # has an mmproj sidecar present on disk. Gating on the flag (not
+        # file presence alone) is what lets /vision OFF free the encoder
+        # RAM by relaunching without --mmproj while KEEPING the projector
+        # file on disk — so re-enabling never re-downloads. We don't
+        # auto-download here; the chat layer prompts the user on first
+        # vision use (see voice.py / vision flow).
         try:
-            from .models_catalog import by_filename
-            from pathlib import Path as _P
-            choice = by_filename(_P(model_path).name)
-            if choice is not None and choice.mmproj_path and choice.mmproj_path.is_file():
-                cmd.extend(["--mmproj", str(choice.mmproj_path)])
+            if getattr(self.config, "vision_enabled", False):
+                from .models_catalog import by_filename
+                from pathlib import Path as _P
+                choice = by_filename(_P(model_path).name)
+                if choice is not None and choice.mmproj_path and choice.mmproj_path.is_file():
+                    cmd.extend(["--mmproj", str(choice.mmproj_path)])
         except Exception:
             # Never block server launch on mmproj lookup
             pass

{localcode-0.3.17 → localcode-0.3.19}/src/localcode/telemetry.py

@@ -122,18 +122,20 @@ class TurnTrace:
         self._active_tool = ToolEvent(name=name, args=args or "",
                                       start_ts=time.monotonic())
 
-    def tool_finished(self, result_chars: int, error: bool) -> None:
-        """Close out the most recent tool call. If no tool is active (e.g.
-        tool_start event was lost), we silently drop — the loss is more
-        useful than a noisy assertion."""
+    def tool_finished(self, result_chars: int, error: bool) -> int:
+        """Close out the most recent tool call and return its duration in ms
+        (0 if no tool is active — e.g. a lost tool_start event). If no tool is
+        active we silently drop; the loss is more useful than a noisy assertion.
+        The returned duration lets the UI render a per-tool time on the done row."""
         if self._active_tool is None:
-            return
+            return 0
         t = self._active_tool
         t.duration_ms = int((time.monotonic() - t.start_ts) * 1000)
         t.result_chars = result_chars
         t.error = bool(error)
         self.tools.append(t)
         self._active_tool = None
+        return t.duration_ms
 
     # ── Finalization ──
 

{localcode-0.3.17 → localcode-0.3.19}/src/localcode/tools/bash.py

@@ -762,6 +762,38 @@ def _quoting_error_hint(output: str) -> str:
     )
 
 
+# Package/version-not-found signatures across npm / pnpm / yarn / pip. A small
+# local model routinely invents package names or versions (e.g. `@types/dexie`,
+# a `^9.9.9` that doesn't exist) → the installer 404s and, with only raw stderr,
+# the model retries the SAME bad name in a loop. None of the reference harnesses
+# handle this deterministically (they rely on a "verify the library exists"
+# prompt line, which isn't enough for a small model), so we catch it here.
+_DEP_NOT_FOUND_MARKERS = (
+    "e404", "etarget", "no matching version found", "notarget",
+    "404 not found - get", "is not in this registry",
+    "could not find a version that satisfies",  # pip
+    "no matching distribution found",           # pip
+    "err_pnpm_no_matching_version",             # pnpm
+)
+
+
+def _dependency_error_hint(output: str) -> str:
+    """Actionable guidance when an install fails because a package/version does
+    not exist — so the model corrects the name instead of retrying it."""
+    low = output.lower()
+    if not any(m in low for m in _DEP_NOT_FOUND_MARKERS):
+        return ""
+    return (
+        "HINT: that package or version does not exist in the registry. Do NOT "
+        "retry the same name/version — it will 404 again. Instead: (1) read "
+        "package.json (or requirements.txt / Cargo.toml) and use a name that is "
+        "actually there; (2) install WITHOUT a pinned version so the resolver "
+        "picks a real latest (e.g. `npm install <pkg>` not `<pkg>@^9.9.9`); or "
+        "(3) if you invented the package, drop it. Never guess package names or "
+        "version numbers."
+    )
+
+
 def _redirect_shell_dir_listing(cmd: str) -> str:
     """Route `ls <path>` to list_files, which can't break on spaces/parens.
 
@@ -1052,6 +1084,11 @@ def execute(ctx: ToolContext, args: dict) -> str:
             _qh = _quoting_error_hint(output)
             if _qh:
                 output = _qh + "\n\n" + output
+            # Package/version-not-found (npm E404 / pip / etc.) → tell the model
+            # the name is wrong so it fixes it instead of retrying the same 404.
+            _dh = _dependency_error_hint(output)
+            if _dh:
+                output = _dh + "\n\n" + output
         # AirPlay-collision detector: if the model curls localhost:5000
         # or :7000, those are squatted by macOS AirPlay Receiver /
         # AirTunes — bare `curl -s` sees a 200 with empty body and

{localcode-0.3.17 → localcode-0.3.19}/src/localcode/tools/project_check.py

@@ -23,13 +23,22 @@ import shutil
 import subprocess
 
 _TIMEOUT = 60.0
-_MAX_ERR_CHARS = 2500
+_MAX_ERR_CHARS = 2500  # floor; scaled up to the real context window at call time
 _ANSI = re.compile(r"\x1b\[[0-9;]*m")
 
 
-def run_project_check(repo_root: str) -> str | None:
+def run_project_check(repo_root: str, ctx_tokens: int = 0) -> str | None:
     """Run the first available typecheck/lint for the project; return a bounded
-    error summary, or None if clean / nothing to run."""
+    error summary, or None if clean / nothing to run.
+
+    `ctx_tokens` (the model's real context window) scales the error-text cap so a
+    big-window model can see more of a wall of tsc/build errors it has room for,
+    instead of a fixed 2500-char slice; floored at 2500 so small machines are
+    unchanged.
+    """
+    max_err_chars = _MAX_ERR_CHARS
+    if ctx_tokens:
+        max_err_chars = max(_MAX_ERR_CHARS, int(ctx_tokens * 3.5 * 0.04))
     env = {**os.environ, "NO_COLOR": "1", "FORCE_COLOR": "0"}
     for label, argv, cwd in _detect_commands(repo_root):
         try:
@@ -42,10 +51,12 @@ def run_project_check(repo_root: str) -> str | None:
         if r.returncode != 0:
             out = _ANSI.sub("", ((r.stdout or "") + "\n" + (r.stderr or "")).strip())
             # Keep only the most useful lines, capped, so we don't flood a
-            # small model's context with a wall of output.
-            lines = [l for l in out.splitlines() if l.strip()][:40]
+            # small model's context with a wall of output. The line cap also
+            # scales with the window (40 on a small machine → more on a big one).
+            max_lines = 40 if not ctx_tokens else max(40, int(ctx_tokens / 4000))
+            lines = [l for l in out.splitlines() if l.strip()][:max_lines]
             if lines:
-                return f"[{label}] reported errors:\n" + "\n".join(lines)[:_MAX_ERR_CHARS]
+                return f"[{label}] reported errors:\n" + "\n".join(lines)[:max_err_chars]
     return None
 
 

localcode-0.3.19/src/localcode/tui/clipboard_image.py

@@ -0,0 +1,131 @@
+"""Read an image off the OS clipboard as PNG bytes.
+
+Terminals never hand an application image data through a Cmd+V / Ctrl+V
+paste — the OS only pastes text, so an image paste arrives as an EMPTY
+paste event. When that happens the TUI calls :func:`read_clipboard_png`
+to pull the image directly out of the OS clipboard and attach it.
+
+macOS-only. On every other platform this returns ``None`` (no
+dependency, no crash) so callers can attempt it unconditionally.
+
+Implementation notes
+--------------------
+We shell out to ``osascript`` (always present on macOS, no new Python
+dependency). Two cases are handled:
+
+1. Raw image data on the clipboard (e.g. a screenshot from Cmd+Shift+4
+   or "Copy Image" in a browser). We coerce the clipboard to
+   ``«class PNGf»`` and write the bytes to a temp file, then read them
+   back in Python.
+2. A copied image FILE (e.g. Finder "Copy" on a .png). The clipboard
+   exposes a file URL (``«class furl»``); if it points at a common
+   image extension we read that file directly.
+"""
+
+from __future__ import annotations
+
+import os
+import subprocess
+import sys
+import tempfile
+from pathlib import Path
+
+_IMAGE_EXTS = {".png", ".jpg", ".jpeg", ".gif", ".webp"}
+_PNG_MAGIC = b"\x89PNG\r\n\x1a\n"
+
+
+def read_clipboard_png() -> bytes | None:
+    """Return image bytes from the OS clipboard, or ``None``.
+
+    Prefers a PNG coercion of raw clipboard image data; falls back to a
+    copied image file. Returns ``None`` on non-macOS platforms and
+    whenever the clipboard holds no image.
+    """
+    if sys.platform != "darwin":
+        return None
+    data = _read_png_via_osascript()
+    if data:
+        return data
+    return _read_image_from_file_url()
+
+
+def _read_png_via_osascript() -> bytes | None:
+    """Coerce raw clipboard image data to PNG and read the bytes."""
+    fd, tmp_path = tempfile.mkstemp(suffix=".png")
+    os.close(fd)
+    try:
+        # AppleScript: try to grab the clipboard as PNG; bail with a
+        # sentinel if there's no image. Write to a temp file we can read
+        # back. `«class PNGf»` is the PNG pasteboard type; `«class furl»`
+        # (used below) is a file URL.
+        script = (
+            "try\n"
+            "\tset pngData to (the clipboard as «class PNGf»)\n"
+            "on error\n"
+            "\treturn \"NO_IMAGE\"\n"
+            "end try\n"
+            f"set outFile to (POSIX file \"{tmp_path}\")\n"
+            "try\n"
+            "\tset fh to open for access outFile with write permission\n"
+            "\tset eof fh to 0\n"
+            "\twrite pngData to fh\n"
+            "\tclose access fh\n"
+            "on error\n"
+            "\ttry\n"
+            "\t\tclose access fh\n"
+            "\tend try\n"
+            "\treturn \"WRITE_ERROR\"\n"
+            "end try\n"
+            "return \"OK\"\n"
+        )
+        result = subprocess.run(
+            ["osascript", "-e", script],
+            capture_output=True,
+            text=True,
+            timeout=15,
+        )
+        if (result.stdout or "").strip() != "OK":
+            return None
+        data = Path(tmp_path).read_bytes()
+        return data if _looks_like_png(data) else None
+    except Exception:
+        return None
+    finally:
+        try:
+            os.unlink(tmp_path)
+        except Exception:
+            pass
+
+
+def _read_image_from_file_url() -> bytes | None:
+    """Read a copied image FILE (Finder copy) from its clipboard URL."""
+    script = (
+        "try\n"
+        "\treturn POSIX path of (the clipboard as «class furl»)\n"
+        "on error\n"
+        "\treturn \"\"\n"
+        "end try\n"
+    )
+    try:
+        result = subprocess.run(
+            ["osascript", "-e", script],
+            capture_output=True,
+            text=True,
+            timeout=10,
+        )
+    except Exception:
+        return None
+    path_str = (result.stdout or "").strip()
+    if not path_str:
+        return None
+    p = Path(path_str)
+    if p.suffix.lower() not in _IMAGE_EXTS or not p.is_file():
+        return None
+    try:
+        return p.read_bytes()
+    except Exception:
+        return None
+
+
+def _looks_like_png(data: bytes) -> bool:
+    return bool(data) and data[:8] == _PNG_MAGIC