Compare commits

..

2 Commits

Author SHA1 Message Date
Jörn Fischer b37d004cb7 Readme auf multimodales jcode angepasst 2026-09-23 12:00:08 +02:00
Jörn Fischer b0c3d32e3f editor replaced (jprompt) multimodal tools added 2026-09-23 11:44:17 +02:00
3 changed files with 914 additions and 38 deletions

View File

@ -1,8 +1,8 @@
# jcode # jcode
A minimal, single-file coding agent (< 1000 lines of Python) for local LLMs — with optional cloud fallback. A minimal, single-file coding agent (~1,200 lines of Python) for local LLMs — with optional cloud fallback.
jcode connects to a local **LM Studio** or **Ollama** server (or, via `settings.json`, to **OpenAI**, **Anthropic**, or **Gemini**), takes a task in plain language, and works autonomously inside a sandboxed `./code` directory: it reads, writes, and edits files, runs commands, writes its own smoke tests, and keeps fixing the code until the tests pass. jcode connects to a local **LM Studio** or **Ollama** server (or, via `settings.json`, to **OpenAI**, **Anthropic**, or **Gemini**), takes a task in plain language, and works autonomously inside a sandboxed `./code` directory: it reads, writes, and edits files, runs commands, searches the web, looks at images/screenshots/PDF pages, writes its own smoke tests, and keeps fixing the code until the tests pass.
``` ```
_ _ _ _
@ -15,15 +15,22 @@ jcode connects to a local **LM Studio** or **Ollama** server (or, via `settings.
## Features ## Features
- **Autonomous test-fix loop** — the agent writes a test for every change, runs it, analyzes failures, fixes the code, and repeats until green (up to `MAX_STEPS`, default 40). - **Autonomous test-fix loop** — the agent writes a test for every change, runs it, analyzes failures, fixes the code, and repeats until green (up to `MAX_STEPS`, default 200).
- **Sandboxed** — all file operations are confined to `./code`; path-escape attempts (`../`, absolute paths) are blocked. Destructive shell commands (`rm -rf /`, `sudo`, `dd`, `mkfs`, shutdown, fork bombs, `curl | sh`, …) are refused. - **Sandboxed** — all file operations are confined to `./code`; path-escape attempts (`../`, absolute paths) are blocked. Destructive shell commands (`rm -rf /`, `sudo`, `dd`, `mkfs`, shutdown, fork bombs, `curl | sh`, …) are refused.
- **Zero-config local backends** — auto-detects LM Studio (port 1234) first, then Ollama (port 11434). - **Zero-config local backends** — auto-detects LM Studio (port 1234) first, then Ollama (port 11434).
- **Cloud providers via `settings.json`** — switch to OpenAI, Anthropic, or Gemini at runtime with `/backend <name>`; all are used through their OpenAI-compatible endpoints, so the single `openai` client dependency covers everything. - **Cloud providers via `settings.json`** — switch to OpenAI, Anthropic, or Gemini at runtime with `/backend <name>`; all are used through their OpenAI-compatible endpoints, so the single `openai` client dependency covers everything.
- **Project guidelines via `AGENTS.md`** — project-specific rules (coding style, per-language test conventions) are loaded from `code/AGENTS.md` and injected as binding instructions. - **Project guidelines via `AGENTS.md`** — project-specific rules (coding style, per-language test conventions) are loaded from `code/AGENTS.md` and injected as binding instructions.
- **Vision** — the agent can look at images, screenshots and PDF pages (see the `view_image`/`take_screenshot`/`view_pdf_page` tools below) to visually check its own work, e.g. a generated plot or a running GUI.
- **Session-wide context management** — token usage is tracked (server-reported or estimated); when the context nears its limit, older history is compacted into an LLM-written summary so long sessions keep working. - **Session-wide context management** — token usage is tracked (server-reported or estimated); when the context nears its limit, older history is compacted into an LLM-written summary so long sessions keep working.
- **Adjustable reasoning effort** — `/reasoning high|medium|low|max|none` controls the model's thinking effort at runtime (backend-dependent; `none` turns it off).
- **Robust against flaky servers** — visible retries with backoff; if the backend goes down, the session is preserved and the task can simply be resubmitted. - **Robust against flaky servers** — visible retries with backoff; if the backend goes down, the session is preserved and the task can simply be resubmitted.
- **Transparent** — every request, tool call, and result is logged to `logFile.md`; file changes are shown as colored unified diffs in the terminal. - **Transparent** — every request, tool call, and result is logged to `logFile.md`; file changes are shown as colored unified diffs in the terminal.
- **Tools available to the model** — `list_files`, `read_file` (optionally with line numbers), `search_files` (regex grep over the sandbox, hits as `file:line: content`), `write_file`, `edit_file` (search/replace), `insert_lines`, `delete_lines`, `run_command`, `web_search` (DuckDuckGo), `fetch_url`. - **Tools available to the model:**
- `list_files`, `read_file` (optionally with line numbers), `search_files` (regex grep over the sandbox, hits as `file:line: content`)
- `write_file`, `edit_file` (search/replace), `insert_lines`, `delete_lines`
- `run_command` (blocks known-destructive patterns; `pip`/`pip3` calls are redirected to jcode's own Python interpreter)
- `web_search` (DuckDuckGo) and `fetch_url` (fetches a page and extracts its text)
- `view_image` (png/jpg/gif/webp/bmp/svg), `take_screenshot` (whole screen, e.g. of a GUI app under test), `view_pdf_page` (renders one PDF page) — images are delivered to the model as an inline image right after the tool result
## Requirements ## Requirements
@ -33,6 +40,11 @@ jcode connects to a local **LM Studio** or **Ollama** server (or, via `settings.
- [LM Studio](https://lmstudio.ai/) with the local server running (default `http://localhost:1234`), or - [LM Studio](https://lmstudio.ai/) with the local server running (default `http://localhost:1234`), or
- [Ollama](https://ollama.com/) (`http://localhost:11434`), or - [Ollama](https://ollama.com/) (`http://localhost:11434`), or
- an API key for OpenAI / Anthropic / Gemini (see `settings.json` below) - an API key for OpenAI / Anthropic / Gemini (see `settings.json` below)
- Optional, only needed for the vision tools (the agent will tell you what's missing and how to install it when it tries to use one):
- `Pillow` — `view_image`, `take_screenshot`, `view_pdf_page`
- `cairosvg` — `view_image` on `.svg` files
- `pymupdf` — `view_pdf_page`
- `certifi` — CA bundle for `fetch_url`; falls back to the system default if not installed. Set `JCODE_INSECURE_SSL=1` to disable TLS verification (e.g. behind a re-signing corporate proxy) — use with care.
## Usage ## Usage
@ -55,8 +67,10 @@ All generated files land in `./code`, which is created automatically.
| `/models` | List models available on the current backend | | `/models` | List models available on the current backend |
| `/model <name>` | Switch model | | `/model <name>` | Switch model |
| `/backend <name>` | Switch to `lm-studio`, `ollama`, or a provider from `settings.json` (e.g. `openai`, `anthropic`, `gemini`) | | `/backend <name>` | Switch to `lm-studio`, `ollama`, or a provider from `settings.json` (e.g. `openai`, `anthropic`, `gemini`) |
| `/reasoning <level>` | Set reasoning effort: `high`, `medium`, `low`, `max`, or `none` (off); no argument resets to the backend default |
| `/max-turns N` | Set the step limit per task | | `/max-turns N` | Set the step limit per task |
| `Ctrl+C` | Quit | | `Ctrl+C` | Abort the current input or running task (session is preserved) |
| `Ctrl+D`, or `Ctrl+C` on an empty prompt | Quit |
## Configuration ## Configuration
@ -105,12 +119,16 @@ If `code/AGENTS.md` exists, it is loaded at session start and treated as binding
| Constant | Default | Meaning | | Constant | Default | Meaning |
|---|---|---| |---|---|---|
| `MODEL` | (auto) | Model name; auto-corrected to the first available model | | `MODEL` | (auto) | Model name; auto-corrected to the closest available model |
| `MAX_STEPS` | 40 | Tool-use steps per task | | `MAX_STEPS` | 200 | Tool-use steps per task |
| `CMD_TIMEOUT` | 60 s | Timeout per shell command | | `CMD_TIMEOUT` | 60 s | Default timeout per shell command (overridable per call via the `timeout` argument) |
| `CONTEXT_LIMIT` | 256 000 | Model context window (tokens) | | `CONTEXT_LIMIT` | 256 000 | Model context window (tokens) |
| `COMPACT_THRESHOLD` | 250 000 | History compaction kicks in above this | | `COMPACT_THRESHOLD` | 250 000 | History compaction kicks in above this |
| `REQUEST_TIMEOUT` | 600 s | Timeout per LLM request (local models are slow) | | `KEEP_RECENT` | 8 | Most recent messages kept uncompressed when compacting |
| `REQUEST_TIMEOUT` | 11 600 s | Timeout per LLM request (local models are slow) |
| `REASONING_EFFORT` | (backend default) | Reasoning effort sent to the backend; set at runtime with `/reasoning` |
| `MAX_IMAGE_SIDE` | 1280 px | Longest edge images are scaled to before being sent to the model |
| `MAX_IMAGES_IN_CONTEXT` | 4 | Older images beyond this are replaced by a text stub to save context |
## How it works ## How it works

273
jcode.py
View File

@ -11,12 +11,13 @@ import re
import subprocess import subprocess
import sys import sys
import time import time
import ssl
import urllib.request import urllib.request
try: try:
from prompt_toolkit import prompt from jprompt import prompt # own multiline editor (jprompt.py next to jcode.py)
except ImportError: except ImportError:
print("ERROR: 'prompt_toolkit' not installed. Run: pip install prompt_toolkit", print("ERROR: 'jprompt.py' not found - it must be in the same directory as jcode.py.",
file=sys.stderr) file=sys.stderr)
sys.exit(1) sys.exit(1)
from html.parser import HTMLParser from html.parser import HTMLParser
@ -26,8 +27,9 @@ from openai import OpenAI
# --------------------------------------------------------------- Configuration # --------------------------------------------------------------- Configuration
APP_NAME = "jcode" APP_NAME = "jcode"
APP_VERSION = "0.9.3" APP_VERSION = "0.10.0"
MODEL = "qwen3.6-35b-a3b-uncensored-hauhaucs-aggressive" DEFAULT_MODEL = "qwen3.6-35b-a3b-uncensored-hauhaucs-aggressive"
MODEL = DEFAULT_MODEL
_LM_STUDIO_URL = "http://localhost:1234/v1" _LM_STUDIO_URL = "http://localhost:1234/v1"
_OLLAMA_URL = "http://localhost:11434/v1" _OLLAMA_URL = "http://localhost:11434/v1"
@ -40,6 +42,19 @@ def _fetch_models(url: str) -> list[str] | None:
except Exception: except Exception:
return None return None
def _pick_model(models: list[str]) -> str:
"""Prefers DEFAULT_MODEL (exact, then case-insensitive, then partial match, e.g.
'publisher/<name>' or '<name>@q4_k_m'); falls back to the first available model."""
if not models: return DEFAULT_MODEL
want = DEFAULT_MODEL.lower()
for match in (lambda m: m == DEFAULT_MODEL,
lambda m: m.lower() == want,
lambda m: m.lower().split("/")[-1].split("@")[0] == want,
lambda m: want in m.lower()):
for m in models:
if match(m): return m
return models[0]
SETTINGS_FILE = Path("settings.json") SETTINGS_FILE = Path("settings.json")
def _load_settings() -> dict: def _load_settings() -> dict:
"""Reads settings.json: providers (OpenAI/Anthropic/Gemini via OpenAI-compatible endpoints) for '/backend <name>'.""" """Reads settings.json: providers (OpenAI/Anthropic/Gemini via OpenAI-compatible endpoints) for '/backend <name>'."""
@ -56,8 +71,7 @@ def _detect_backend() -> tuple[str, str, str]:
(_OLLAMA_URL, "ollama", "Ollama")]: (_OLLAMA_URL, "ollama", "Ollama")]:
models = _fetch_models(url) models = _fetch_models(url)
if models is not None: if models is not None:
if models and MODEL not in models: MODEL = _pick_model(models)
MODEL = models[0]
return url, key, name return url, key, name
return _LM_STUDIO_URL, "lm-studio", "LM Studio" return _LM_STUDIO_URL, "lm-studio", "LM Studio"
@ -76,19 +90,20 @@ def _switch_backend(name: str) -> None:
key, bname = ("ollama", "Ollama") if url == _OLLAMA_URL else ("lm-studio", "LM Studio") key, bname = ("ollama", "Ollama") if url == _OLLAMA_URL else ("lm-studio", "LM Studio")
models = _fetch_models(url) models = _fetch_models(url)
if models is None: print(f"{RED}\u2717 {bname} not reachable.{RESET}"); return if models is None: print(f"{RED}\u2717 {bname} not reachable.{RESET}"); return
if models and MODEL not in models: MODEL = models[0] MODEL = _pick_model(models)
BASE_URL, _API_KEY, BACKEND = url, key, bname BASE_URL, _API_KEY, BACKEND = url, key, bname
client = OpenAI(base_url=BASE_URL, api_key=_API_KEY, timeout=REQUEST_TIMEOUT, max_retries=0) client = OpenAI(base_url=BASE_URL, api_key=_API_KEY, timeout=REQUEST_TIMEOUT, max_retries=0)
print(f"{DIM}Backend: {BACKEND} ({BASE_URL}) | Model: {MODEL}{RESET}") print(f"{DIM}Backend: {BACKEND} ({BASE_URL}) | Model: {MODEL}{RESET}")
SANDBOX = Path("code").resolve() SANDBOX = Path("code").resolve()
MAX_STEPS = 40 MAX_STEPS = 200
CMD_TIMEOUT = 60 # seconds per command (compiling C/C++/Java takes time) CMD_TIMEOUT = 60 # seconds per command (compiling C/C++/Java takes time)
CONTEXT_LIMIT = 256_000 # context window of the model (tokens) CONTEXT_LIMIT = 256_000 # context window of the model (tokens)
COMPACT_THRESHOLD = 250_000 # history is compacted from here on COMPACT_THRESHOLD = 250_000 # history is compacted from here on
KEEP_RECENT = 8 # most recent messages kept uncompressed KEEP_RECENT = 8 # most recent messages kept uncompressed
REQUEST_TIMEOUT = 600 # seconds per LLM request (local models are slow) REQUEST_TIMEOUT = 11600 # seconds per LLM request (local models are slow)
RETRY_WAITS = (5, 15, 30, 60) # wait times between retry attempts RETRY_WAITS = (5, 15, 30, 60) # wait times between retry attempts
REASONING_EFFORT = None # None = backend default; else "high"|"medium"|"low"|"max"|"none" (Ollama /v1)
LOG_FILE = Path("logFile.md") LOG_FILE = Path("logFile.md")
_REQ_NR = 0 # running number of LM Studio requests _REQ_NR = 0 # running number of LM Studio requests
@ -127,7 +142,10 @@ SYSTEM_PROMPT = (
"You are a precise coding agent. Your working directory is already the " "You are a precise coding agent. Your working directory is already the "
"project directory. ALWAYS give paths relative to it, e.g. 'fib.py' or " "project directory. ALWAYS give paths relative to it, e.g. 'fib.py' or "
"'tests/test_fib.py' - NEVER with a leading 'code/' and NEVER absolute. " "'tests/test_fib.py' - NEVER with a leading 'code/' and NEVER absolute. "
"Use the tools to read, write, list, and edit files and to run commands.\n\n" "Use the tools to read, write, list, and edit files and to run commands. "
"Images (view_image), screenshots (take_screenshot) and PDF pages "
"(view_pdf_page) are delivered as an image in the message right after the "
"tool result - describe only what you actually see, never guess.\n\n"
"Language: reply in the same language as the task (German task -> German answers, English task -> English).\n\n" "Language: reply in the same language as the task (German task -> German answers, English task -> English).\n\n"
"How to work:\n" "How to work:\n"
"1. Locate code with search_files and read a file before you edit it. " "1. Locate code with search_files and read a file before you edit it. "
@ -443,6 +461,21 @@ class _TextExtractor(HTMLParser):
blank = True blank = True
return "\n".join(out).strip() return "\n".join(out).strip()
def _make_ssl_context() -> ssl.SSLContext:
"""CA bundle for https fetches: certifi (pip install certifi) if available,
otherwise the OS/Python default. JCODE_INSECURE_SSL=1 disables verification
(e.g. behind a corporate proxy that re-signs TLS) - use with care."""
if os.environ.get("JCODE_INSECURE_SSL") == "1":
ctx = ssl.create_default_context()
ctx.check_hostname = False; ctx.verify_mode = ssl.CERT_NONE
return ctx
try:
import certifi # noqa: PLC0415
return ssl.create_default_context(cafile=certifi.where())
except Exception: # noqa: BLE001
return ssl.create_default_context()
_SSL_CONTEXT = _make_ssl_context()
def fetch_url(url: str, max_chars: int = 6000) -> str: def fetch_url(url: str, max_chars: int = 6000) -> str:
"""Loads a web page and returns the extracted text content.""" """Loads a web page and returns the extracted text content."""
if not url.lower().startswith(("http://", "https://")): if not url.lower().startswith(("http://", "https://")):
@ -452,7 +485,7 @@ def fetch_url(url: str, max_chars: int = 6000) -> str:
url, headers={"User-Agent": f"Mozilla/5.0 ({APP_NAME}/{APP_VERSION})"} url, headers={"User-Agent": f"Mozilla/5.0 ({APP_NAME}/{APP_VERSION})"}
) )
try: try:
with urllib.request.urlopen(req, timeout=15) as resp: with urllib.request.urlopen(req, timeout=15, context=_SSL_CONTEXT) as resp:
ctype = resp.headers.get("Content-Type", "") ctype = resp.headers.get("Content-Type", "")
raw = resp.read(2_000_000) # load at most 2 MB raw = resp.read(2_000_000) # load at most 2 MB
charset = resp.headers.get_content_charset() or "utf-8" charset = resp.headers.get_content_charset() or "utf-8"
@ -474,6 +507,128 @@ def fetch_url(url: str, max_chars: int = 6000) -> str:
text = text[:max_chars] + f"\n...[truncated, {len(text)} characters total]..." text = text[:max_chars] + f"\n...[truncated, {len(text)} characters total]..."
return f"Content of {url}:\n{text}" return f"Content of {url}:\n{text}"
# ----------------------------------------------------------------- Image tools
# Tool results in the OpenAI chat format are text-only. Images are therefore
# queued here and delivered to the model as a separate user message directly
# after the tool results of the current step (works with LM Studio/Ollama for
# vision models such as Qwen3-VL and Gemma).
MAX_IMAGE_SIDE = 1280 # longest edge in pixels sent to the model
MAX_IMAGES_IN_CONTEXT = 4 # older images are replaced by a text stub
IMAGE_TOKENS = 1000 # rough token estimate per image
_PENDING_IMAGES: list[tuple[str, str]] = [] # (label, data URI)
def _pil():
try:
from PIL import Image
return Image
except ImportError:
raise RuntimeError("package 'Pillow' not installed - run 'pip install Pillow' with run_command")
def _attach_image(img, label: str) -> str:
"""Scales a PIL image, encodes it as data URI and queues it for the next message."""
import base64, io
w, h = img.size
if (scale := MAX_IMAGE_SIDE / max(w, h)) < 1:
img = img.resize((round(w * scale), round(h * scale)), _pil().LANCZOS)
buf = io.BytesIO()
if img.mode in ("RGBA", "LA") or "transparency" in img.info:
img.convert("RGBA").save(buf, "PNG"); mime = "image/png"
else:
img.convert("RGB").save(buf, "JPEG", quality=90); mime = "image/jpeg"
uri = f"data:{mime};base64,{base64.b64encode(buf.getvalue()).decode()}"
_PENDING_IMAGES.append((label, uri))
return f"{label} ({w}x{h}px) attached - it follows in the next message."
def view_image(path: str) -> str:
"""Loads an image file from the sandbox (png/jpg/gif/webp/bmp/svg) for visual analysis."""
p = _resolve(path)
if not p.is_file():
return f"File '{path}' not found. Use list_files() to see available files."
Image = _pil()
if p.suffix.lower() == ".svg":
try:
import cairosvg # pip install cairosvg
except ImportError:
return ("ERROR: package 'cairosvg' not installed - run 'pip install cairosvg' "
"with run_command (or read the SVG source with read_file).")
import io
img = Image.open(io.BytesIO(cairosvg.svg2png(url=str(p), output_width=MAX_IMAGE_SIDE)))
else:
img = Image.open(p)
img.load()
return _attach_image(img, f"Image '{path}'")
def take_screenshot(save_as: str = "", delay: int = 0) -> str:
"""Captures the whole screen (e.g. of a running GUI program) for visual analysis."""
Image = _pil()
time.sleep(max(0, min(int(delay), 30)))
try:
from PIL import ImageGrab
img = ImageGrab.grab()
except Exception as e: # noqa: BLE001 - e.g. Wayland: use a system screenshot tool
import shutil, tempfile
tool = next((c for c in (["grim"], ["gnome-screenshot", "-f"], ["spectacle", "-bno"])
if shutil.which(c[0])), None)
if tool is None:
return f"ERROR: screenshot failed ({e}) and no grim/gnome-screenshot/spectacle found."
tmp = Path(tempfile.mkdtemp()) / "shot.png"
subprocess.run(tool + [str(tmp)], capture_output=True, timeout=20)
if not tmp.is_file():
return f"ERROR: screenshot via {tool[0]} failed."
img = Image.open(tmp); img.load()
label = "Screenshot"
if save_as:
p = _resolve(save_as if Path(save_as).suffix else save_as + ".png")
p.parent.mkdir(parents=True, exist_ok=True)
img.save(p); label += f" (saved as '{p.relative_to(SANDBOX)}')"
return _attach_image(img, label)
def view_pdf_page(path: str, page: int = 1, dpi: int = 110) -> str:
"""Renders one page of a PDF from the sandbox as an image for visual analysis."""
p = _resolve(path)
if not p.is_file():
return f"File '{path}' not found. Use list_files() to see available files."
try:
try:
import pymupdf as fitz
except ImportError:
import fitz # older pymupdf versions
except ImportError:
return "ERROR: package 'pymupdf' not installed - run 'pip install pymupdf' with run_command."
Image = _pil()
with fitz.open(p) as doc:
n = doc.page_count
if not 1 <= page <= n:
return f"page={page} invalid - '{path}' has {n} page(s)."
pix = doc[page - 1].get_pixmap(dpi=max(50, min(int(dpi), 200)))
img = Image.frombytes("RGB", (pix.width, pix.height), pix.samples)
return _attach_image(img, f"Page {page}/{n} of '{path}'")
def _image_message() -> dict:
"""Builds the user message carrying the queued images and clears the queue."""
labels = ", ".join(l for l, _ in _PENDING_IMAGES)
content = [{"type": "text", "text": f"[Attached from the tool result(s) above: {labels}]"}]
content += [{"type": "image_url", "image_url": {"url": uri}} for _, uri in _PENDING_IMAGES]
_PENDING_IMAGES.clear()
_log("### Images attached", labels)
return {"role": "user", "content": content}
def _prune_images(messages: list) -> None:
"""Keeps only the MAX_IMAGES_IN_CONTEXT newest images; older ones become text stubs."""
seen = 0
for msg in reversed(messages):
for part in reversed(msg["content"]) if isinstance(msg.get("content"), list) else ():
if part.get("type") == "image_url":
seen += 1
if seen > MAX_IMAGES_IN_CONTEXT:
part.clear(); part.update({"type": "text", "text": "[older image removed from context]"})
def _content_text(content) -> str:
"""Text view of a message content (image parts become '[image]')."""
if isinstance(content, list):
return " ".join("[image]" if p.get("type") == "image_url" else p.get("text", "") for p in content)
return str(content or "")
TOOL_FUNCS = { TOOL_FUNCS = {
"list_files": list_files, "list_files": list_files,
"read_file": read_file, "read_file": read_file,
@ -485,6 +640,9 @@ TOOL_FUNCS = {
"run_command": run_command, "run_command": run_command,
"web_search": web_search, "web_search": web_search,
"fetch_url": fetch_url, "fetch_url": fetch_url,
"view_image": view_image,
"take_screenshot": take_screenshot,
"view_pdf_page": view_pdf_page,
} }
# Per-tool descriptions and parameter descriptions. The builder below derives # Per-tool descriptions and parameter descriptions. The builder below derives
@ -575,6 +733,32 @@ _TOOL_META = {
"max_chars": "Max. text length (500-20000), default 6000", "max_chars": "Max. text length (500-20000), default 6000",
}, },
}, },
"view_image": {
"description": (
"Shows an image file from the project (png, jpg, gif, webp, bmp, svg) "
"to you for visual analysis, e.g. a generated plot, an icon or a "
"mockup. The image arrives in the message after the tool result."
),
"params": {"path": "Image file, e.g. 'plot.png' or 'assets/logo.svg'"},
},
"take_screenshot": {
"description": (
"Captures the whole screen and shows it to you, e.g. to check a GUI "
"program you started. Start GUI programs in the background first "
"(Linux/macOS: 'python app.py > app.log 2>&1 &', Windows: "
"'start python app.py'), then take the screenshot with a delay."
),
"params": {"save_as": "Optional file name to keep the screenshot, e.g. 'shot.png'",
"delay": "Seconds to wait before capturing (0-30), default 0"},
},
"view_pdf_page": {
"description": (
"Renders ONE page of a PDF file from the project as an image and "
"shows it to you. The result states the total number of pages."
),
"params": {"page": "Page number (1-based), default 1",
"dpi": "Render resolution (50-200), default 110"},
},
} }
def _build_tool_schema(fn) -> dict: def _build_tool_schema(fn) -> dict:
@ -627,11 +811,12 @@ def splash_screen() -> None:
print(f" Sandbox : {SANDBOX}") print(f" Sandbox : {SANDBOX}")
print(f" Max steps : {MAX_STEPS}") print(f" Max steps : {MAX_STEPS}")
print(" " + "\u2500" * 50) print(" " + "\u2500" * 50)
_submit = "Esc, Enter" if sys.platform == "darwin" else "Alt+Enter" print(" Describe a task: Enter submits, Ctrl+J starts a new line, Ctrl+C aborts")
print(f" Describe a task (multiline, submit with {_submit}). History is kept") print(" the input or a running task, Ctrl+D or Ctrl+C on an empty prompt exits.")
print(" across tasks. '/status' shows usage, '/new' clears the session,") print(" History is kept across tasks. '/status' shows usage, '/new' clears the session,")
print(" '/model' sets model, '/models' lists models, '/backend <name>' switches to") print(" '/model' sets model, '/models' lists models, '/backend <name>' switches to")
print(" lm-studio, ollama or a settings.json provider, '/max-turns N' sets step limit.\n") print(" lm-studio, ollama or a settings.json provider, '/max-turns N' sets step limit,")
print(" '/reasoning high|medium|low|max|none' sets thinking effort (none = off).\n")
# Descriptive action label per tool call - WITHOUT code/file content. # Descriptive action label per tool call - WITHOUT code/file content.
def _action_label(name: str, args: dict) -> str: def _action_label(name: str, args: dict) -> str:
@ -664,6 +849,12 @@ def _action_label(name: str, args: dict) -> str:
return f"\U0001f50d Researching \"{args.get('query', '')}\"" return f"\U0001f50d Researching \"{args.get('query', '')}\""
if name == "fetch_url": if name == "fetch_url":
return f"\U0001f310 Loading page {args.get('url', '')}" return f"\U0001f310 Loading page {args.get('url', '')}"
if name == "view_image":
return f"\U0001f5bc\ufe0f Viewing {target}"
if name == "take_screenshot":
return f"\U0001f4f8 Screenshot {args.get('save_as') or ''}".rstrip()
if name == "view_pdf_page":
return f"\U0001f4c4 Rendering {target} (page {args.get('page', 1)})"
return f"\u2699\ufe0f {name}" return f"\u2699\ufe0f {name}"
# Compact result status - never output file content or code. # Compact result status - never output file content or code.
@ -724,12 +915,15 @@ LAST_USAGE_TOKENS = 0 # last real usage.total_tokens value from the server
def _estimate_tokens(messages: list) -> int: def _estimate_tokens(messages: list) -> int:
"""Rough token estimate (~4 chars/token) as a fallback without usage data.""" """Rough token estimate (~4 chars/token) as a fallback without usage data."""
chars = 0 chars, images = 0, 0
for msg in messages: for msg in messages:
chars += len(str(msg.get("content") or "")) content = msg.get("content")
chars += len(_content_text(content))
if isinstance(content, list):
images += sum(1 for p in content if p.get("type") == "image_url")
for tc in msg.get("tool_calls") or []: for tc in msg.get("tool_calls") or []:
chars += len(json.dumps(tc, ensure_ascii=False)) chars += len(json.dumps(tc, ensure_ascii=False))
return chars // 4 return chars // 4 + images * IMAGE_TOKENS
def _compact_cut_index(messages: list) -> int: def _compact_cut_index(messages: list) -> int:
"""Determines the index from which messages are kept. """Determines the index from which messages are kept.
@ -758,7 +952,7 @@ def _compact_messages(messages: list) -> list:
parts = [] parts = []
for msg in old: for msg in old:
role = msg.get("role", "?") role = msg.get("role", "?")
content = str(msg.get("content") or "") content = _content_text(msg.get("content"))
if role == "tool" and len(content) > 400: if role == "tool" and len(content) > 400:
content = content[:400] + "...[truncated]" content = content[:400] + "...[truncated]"
calls = ", ".join( calls = ", ".join(
@ -877,7 +1071,7 @@ def run(task: str, messages: list) -> list:
The history is preserved across tasks; the compaction therefore works The history is preserved across tasks; the compaction therefore works
session-wide automatically. Returns the updated history (compaction session-wide automatically. Returns the updated history (compaction
creates a new list). is done in place, so the caller's list stays the live history).
""" """
global _REQ_NR, LAST_USAGE_TOKENS global _REQ_NR, LAST_USAGE_TOKENS
messages.append({"role": "user", "content": task}) messages.append({"role": "user", "content": task})
@ -889,18 +1083,19 @@ def run(task: str, messages: list) -> list:
# response, otherwise the character heuristic. Compact on overflow. # response, otherwise the character heuristic. Compact on overflow.
used = max(last_total_tokens, _estimate_tokens(messages)) used = max(last_total_tokens, _estimate_tokens(messages))
if used > COMPACT_THRESHOLD: if used > COMPACT_THRESHOLD:
messages = _compact_messages(messages) messages[:] = _compact_messages(messages)
last_total_tokens = 0 last_total_tokens = 0
LAST_USAGE_TOKENS = 0 LAST_USAGE_TOKENS = 0
_prune_images(messages)
try: try:
_REQ_NR += 1 _REQ_NR += 1
_log(f"### Request {_REQ_NR}", _log(f"### Request {_REQ_NR}",
f"{len(messages)} messages in context, last: " f"{len(messages)} messages in context, last: "
f"{str(messages[-1].get('content') or '')[:300]}") f"{_content_text(messages[-1].get('content'))[:300]}")
resp = _chat( resp = _chat(
model=MODEL, messages=messages, tools=TOOLS, temperature=0.2 model=MODEL, messages=messages, tools=TOOLS, temperature=0.2,
) **({"reasoning_effort": REASONING_EFFORT} if REASONING_EFFORT else {}))
except LLMUnavailable as e: except LLMUnavailable as e:
_log("### Error", str(e)) _log("### Error", str(e))
print(f"\n{RED}\u2717 {e}{RESET}") print(f"\n{RED}\u2717 {e}{RESET}")
@ -949,13 +1144,24 @@ def run(task: str, messages: list) -> list:
messages.append( messages.append(
{"role": "tool", "tool_call_id": tc.id, "content": str(result)} {"role": "tool", "tool_call_id": tc.id, "content": str(result)}
) )
if _PENDING_IMAGES:
messages.append(_image_message())
print("\u2514\u2500") print("\u2514\u2500")
print("Maximum number of steps reached.") print("Maximum number of steps reached.")
return messages return messages
def _discard_incomplete(messages: list) -> None:
"""After Ctrl+C: drops a trailing tool-call round whose tool results are incomplete."""
i = len(messages)
while i > 0 and messages[i - 1].get("role") == "tool":
i -= 1
if i > 0 and messages[i - 1].get("role") == "assistant" and messages[i - 1].get("tool_calls"):
if len(messages) - i < len(messages[i - 1]["tool_calls"]):
del messages[i - 1:]
def main() -> None: def main() -> None:
global LAST_USAGE_TOKENS, MODEL, MAX_STEPS global LAST_USAGE_TOKENS, MODEL, MAX_STEPS, REASONING_EFFORT
SANDBOX.mkdir(exist_ok=True) SANDBOX.mkdir(exist_ok=True)
if (d := SETTINGS.get("default_provider")): _switch_backend(d) if (d := SETTINGS.get("default_provider")): _switch_backend(d)
if len(sys.argv) > 1: if len(sys.argv) > 1:
@ -967,9 +1173,12 @@ def main() -> None:
session = new_session() session = new_session()
while True: while True:
try: try:
task = prompt("\nTask> ", multiline=True).strip() task = prompt("\nTask>").strip()
except (EOFError, KeyboardInterrupt): except (EOFError, KeyboardInterrupt) as e:
print("\nSee you soon.") if getattr(e, "text", "").strip(): # Ctrl+C with text: discard it only
print(f"{DIM}(input discarded){RESET}")
continue
print("See you soon.") # Ctrl+D, or Ctrl+C on an empty prompt
break break
if not task: if not task:
continue continue
@ -987,9 +1196,15 @@ def main() -> None:
_switch_backend(task[8:].strip()); continue _switch_backend(task[8:].strip()); continue
if task.lower().startswith("/model"): if task.lower().startswith("/model"):
MODEL = task[6:].strip() or MODEL; print(f"{DIM}Current model: {MODEL}{RESET}"); continue MODEL = task[6:].strip() or MODEL; print(f"{DIM}Current model: {MODEL}{RESET}"); continue
if task.lower().startswith("/reasoning"):
r = task[10:].strip().lower(); REASONING_EFFORT = r if r in ("high", "medium", "low", "max", "none") else (None if r in ("", "default") else REASONING_EFFORT); print(f"{DIM}Reasoning effort: {REASONING_EFFORT or 'default'}{RESET}" if r in ("high", "medium", "low", "max", "none", "", "default") else f"{RED}\u2717 Use: /reasoning high|medium|low|max|none{RESET}"); continue
if task.lower().startswith("/max-turns"): if task.lower().startswith("/max-turns"):
MAX_STEPS = int(n) if (n := task[10:].strip()).isdigit() else MAX_STEPS; print(f"{DIM}Max steps: {MAX_STEPS}{RESET}"); continue MAX_STEPS = int(n) if (n := task[10:].strip()).isdigit() else MAX_STEPS; print(f"{DIM}Max steps: {MAX_STEPS}{RESET}"); continue
session = run(task, session) try:
session = run(task, session)
except KeyboardInterrupt:
_discard_incomplete(session)
print(f"\n{RED}\u2717 Task aborted (Ctrl+C) - the session is preserved.{RESET}")
if __name__ == "__main__": if __name__ == "__main__":
main() main()

643
jprompt.py 100644
View File

@ -0,0 +1,643 @@
#!/usr/bin/env python3
# MIT License - Copyright (c) 2026 Prof. Dr. Joern Fischer (full text: see README.md)
"""jprompt - a small, dependency-free multiline prompt editor for the terminal.
The editor mimics the input box of Claude Code: the text starts on its own
line (below an optional header such as "Task>"), wraps at the terminal
width, scrolls vertically when it is taller than the terminal and keeps the
cursor visible at all times. The terminal is read character by character
(cbreak-like raw mode); nothing is echoed by the terminal itself.
Keys
Enter submit the text
Ctrl+J insert a new line (Shift+Enter / Alt+Enter work too
where the terminal reports them)
Ctrl+C abort the input -> raises Abort (a KeyboardInterrupt
whose .text holds the discarded input, "" if it was empty)
Ctrl+D exit the application -> raises EOFError
Ctrl+V / Cmd+V paste. Terminals normally deliver a paste themselves
(bracketed paste is enabled so newlines in the pasted
text do NOT submit). If the terminal passes Ctrl+V
through, the system clipboard is read as a fallback.
Arrows move the cursor. Up on the first row and Down on the
last row do nothing.
Home/End, Ctrl+A/Ctrl+E, Backspace, Delete, Tab (= 4 spaces)
Usage
from jprompt import prompt
text = prompt("Task>") # header is printed, editing starts below
Architecture (every part can be tested without a terminal)
Buffer lines + cursor and all editing operations
Layout wraps the Buffer lines to a width, maps cursor <-> screen cell
Editor Buffer + Layout + rendering (ANSI) into any file-like object
KeyReader turns raw characters into KeyEvents (escape sequences, paste)
prompt() wires everything to the real terminal (raw mode, SIGWINCH)
Known limitation: terminals that *reflow* old text on resize (iTerm2,
kitty, Windows Terminal) may leave a stale copy of the editor above the
freshly drawn one after a resize; the editor itself and its cursor stay
correct because every redraw is relative to the cursor position.
"""
from __future__ import annotations
import codecs
import os
import select
import shutil
import signal
import subprocess
import sys
import unicodedata
from dataclasses import dataclass
try:
import termios
except ImportError: # Windows
termios = None
try:
import msvcrt
except ImportError:
msvcrt = None
__all__ = ["prompt", "Abort", "Buffer", "Layout", "Editor", "KeyReader", "KeyEvent", "RESIZE"]
class Abort(KeyboardInterrupt):
"""Raised on Ctrl+C; `.text` is the input that was being edited."""
def __init__(self, text: str = "") -> None:
super().__init__("input aborted")
self.text = text
# ------------------------------------------------------------- Text helpers
def char_width(ch: str) -> int:
"""Display width of one character (0 for combining marks, 2 for CJK)."""
if unicodedata.combining(ch) or unicodedata.category(ch) in ("Mn", "Me", "Cf"):
return 0
return 2 if unicodedata.east_asian_width(ch) in ("W", "F") else 1
def str_width(s: str) -> int:
return sum(char_width(ch) for ch in s)
def _clean(text: str) -> str:
"""Normalises newlines, expands tabs and drops control characters."""
text = text.replace("\r\n", "\n").replace("\r", "\n").replace("\t", " ")
return "".join(ch for ch in text if ch == "\n" or unicodedata.category(ch) != "Cc")
# ------------------------------------------------------------------- Buffer
class Buffer:
"""Editable text: a list of logical lines plus a cursor (row, col)."""
def __init__(self, text: str = "") -> None:
self.lines: list[str] = _clean(text).split("\n")
self.row = len(self.lines) - 1
self.col = len(self.lines[-1])
# -- state
@property
def text(self) -> str:
return "\n".join(self.lines)
@property
def cursor(self) -> tuple[int, int]:
return self.row, self.col
# -- editing
def insert(self, text: str) -> None:
parts = _clean(text).split("\n")
line = self.lines[self.row]
head, tail = line[: self.col], line[self.col :]
if len(parts) == 1:
self.lines[self.row] = head + parts[0] + tail
self.col += len(parts[0])
return
self.lines[self.row : self.row + 1] = [head + parts[0], *parts[1:-1], parts[-1] + tail]
self.row += len(parts) - 1
self.col = len(parts[-1])
def backspace(self) -> None:
if self.col > 0:
line = self.lines[self.row]
self.lines[self.row] = line[: self.col - 1] + line[self.col :]
self.col -= 1
elif self.row > 0:
self.col = len(self.lines[self.row - 1])
self.lines[self.row - 1] += self.lines.pop(self.row)
self.row -= 1
def delete(self) -> None:
line = self.lines[self.row]
if self.col < len(line):
self.lines[self.row] = line[: self.col] + line[self.col + 1 :]
elif self.row + 1 < len(self.lines):
self.lines[self.row] += self.lines.pop(self.row + 1)
# -- horizontal movement
def left(self) -> None:
if self.col > 0:
self.col -= 1
elif self.row > 0:
self.row -= 1
self.col = len(self.lines[self.row])
def right(self) -> None:
if self.col < len(self.lines[self.row]):
self.col += 1
elif self.row + 1 < len(self.lines):
self.row += 1
self.col = 0
def home(self) -> None:
self.col = 0
def end(self) -> None:
self.col = len(self.lines[self.row])
# ------------------------------------------------------------------- Layout
@dataclass(frozen=True)
class VisualRow:
line: int # index of the logical line
start: int # first character (inclusive)
end: int # last character (exclusive)
last: bool # last visual row of its logical line?
class Layout:
"""Wraps logical lines to `width` display columns.
A logical line of exactly `width` columns is NOT split: the cursor may sit
at display column `width` (which is why Editor uses terminal width - 1).
A cursor at the end of a non-last visual row is shown at the start of the
next row - the same convention prompt_toolkit uses.
"""
def __init__(self, lines: list[str], width: int) -> None:
self.lines = lines
self.width = max(1, width)
self.rows: list[VisualRow] = []
for li, line in enumerate(lines):
start = acc = 0
for i, ch in enumerate(line):
w = char_width(ch)
if acc + w > self.width and i > start:
self.rows.append(VisualRow(li, start, i, False))
start, acc = i, 0
acc += w
self.rows.append(VisualRow(li, start, len(line), True))
def row_text(self, r: VisualRow) -> str:
return self.lines[r.line][r.start : r.end]
def cursor_to_visual(self, row: int, col: int) -> tuple[int, int]:
"""(logical row, col) -> (visual row index, display column)."""
for vr, r in enumerate(self.rows):
if r.line == row and (r.start <= col < r.end or (r.last and col == r.end)):
return vr, str_width(self.lines[row][r.start : col])
return len(self.rows) - 1, 0 # unreachable for a consistent buffer
def visual_to_cursor(self, vr: int, want_col: int) -> tuple[int, int]:
"""(visual row index, wanted display column) -> nearest (row, col)."""
r = self.rows[vr]
line = self.lines[r.line]
limit = r.end if r.last else r.end - 1 # end of a wrapped row belongs to the next row
col, acc = r.start, 0
while col < limit:
w = char_width(line[col])
if acc + w > want_col:
break
acc, col = acc + w, col + 1
return r.line, col
# ---------------------------------------------------------------- KeyEvents
@dataclass(frozen=True)
class KeyEvent:
kind: str # text, newline, enter, backspace, delete, left, right, up,
# down, home, end, ctrl_c, ctrl_d, resize, ignore
text: str = "" # only for kind == "text" (typed char or pasted block)
RESIZE = object() # sentinel a key source returns when the window was resized
# ------------------------------------------------------------------- Editor
class Editor:
"""Buffer + rendering. `out` needs write()/flush(); `size()` -> (cols, rows).
The editor assumes the terminal cursor is at column 0 of the line where
the text should start when render() is first called. All later drawing is
relative to the cursor, so scrolling of the terminal does not matter.
"""
def __init__(self, out, size, text: str = "") -> None:
self.buf = Buffer(text)
self.out = out
self._size = size
self._want_col: int | None = None # sticky column for Up/Down
self._top = 0 # first visible visual row (vertical scrolling)
self._cur_row = 0 # screen row of the cursor relative to the first drawn row
self._drawn = 0 # number of rows drawn by the last render()
self.hidden_above = 0 # visual rows scrolled out of view (shown as markers)
self.hidden_below = 0
# -- geometry
@property
def width(self) -> int:
"""Wrap width: one less than the terminal so the cursor never touches the auto-wrap column."""
return max(1, self._size()[0] - 1)
@property
def height(self) -> int:
"""Max. visible rows: one less than the terminal so a header line above survives."""
return max(1, self._size()[1] - 1)
@property
def text(self) -> str:
return self.buf.text
def layout(self) -> Layout:
return Layout(self.buf.lines, self.width)
# -- keys
def handle(self, key: KeyEvent) -> bool:
"""Applies one key; returns True when the text was submitted."""
k = key.kind
if k == "ctrl_c":
raise Abort(self.text)
if k == "ctrl_d":
raise EOFError
if k == "enter":
return True
if k == "text":
self.buf.insert(key.text)
elif k == "newline":
self.buf.insert("\n")
elif k == "backspace":
self.buf.backspace()
elif k == "delete":
self.buf.delete()
elif k == "left":
self.buf.left()
elif k == "right":
self.buf.right()
elif k == "home":
self.buf.home()
elif k == "end":
self.buf.end()
elif k == "up":
self.move_up()
elif k == "down":
self.move_down()
if k not in ("up", "down", "ignore"):
self._want_col = None
self.render()
return False
def move_up(self) -> bool:
return self._vertical(-1)
def move_down(self) -> bool:
return self._vertical(+1)
def _vertical(self, delta: int) -> bool:
"""Moves one *visual* row; returns False at the top/bottom edge."""
lay = self.layout()
vr, vc = lay.cursor_to_visual(*self.buf.cursor)
target = vr + delta
if not 0 <= target < len(lay.rows):
return False
if self._want_col is None:
self._want_col = vc
self.buf.row, self.buf.col = lay.visual_to_cursor(target, self._want_col)
return True
# -- drawing
def _viewport(self, vr: int, n: int) -> tuple[int, int, bool, bool]:
"""Chooses the first visible row so the cursor row is on screen.
Returns (top, rows shown, marker above?, marker below?). Each marker
line ("... N more lines ...") costs one row of the available height,
so the choice is iterated until it is stable.
"""
h, top = self.height, self._top
above = below = False
for _ in range(3):
avail = max(1, h - above - below)
if vr < top:
top = vr
elif vr >= top + avail:
top = vr - avail + 1
top = max(0, min(top, n - avail))
if h < 3: # no room for markers
return top, avail, False, False
new = (top > 0, top + avail < n)
if new == (above, below):
return top, avail, above, below
above, below = new
return top, avail, above, below
def _marker(self, count: int, where: str) -> str:
text = f"\u2026 {count} more line{'s' if count != 1 else ''} {where} \u2026"
if str_width(text) > self.width:
arrow = "\u2191" if where == "above" else "\u2193"
text = f"\u2026{count}{arrow}"
return f"\x1b[2m{text}\x1b[0m"
def render(self) -> None:
lay = self.layout()
vr, vc = lay.cursor_to_visual(*self.buf.cursor)
n = len(lay.rows)
top, avail, above, below = self._viewport(vr, n)
self._top = top
self.hidden_above, self.hidden_below = top, max(0, n - top - avail)
lines = [lay.row_text(r) for r in lay.rows[top : top + avail]]
if above:
lines.insert(0, self._marker(self.hidden_above, "above"))
if below:
lines.append(self._marker(self.hidden_below, "below"))
crow = vr - top + int(above)
parts = []
if self._cur_row > 0: # back to the first drawn row
parts.append(f"\x1b[{self._cur_row}A")
parts.append("\r\x1b[J") # column 0, clear to end of screen
parts.append("\r\n".join(lines))
back = len(lines) - 1 - crow
if back > 0:
parts.append(f"\x1b[{back}A")
parts.append("\r")
if vc > 0:
parts.append(f"\x1b[{vc}C")
self.out.write("".join(parts))
self.out.flush()
self._cur_row, self._drawn = crow, len(lines)
def finish(self) -> None:
"""Moves the cursor below the text and starts a fresh line."""
down = self._drawn - 1 - self._cur_row
self.out.write((f"\x1b[{down}B" if down > 0 else "") + "\r\n")
self.out.flush()
# ---------------------------------------------------------------- KeyReader
class KeyReader:
"""Decodes characters from a `source` into KeyEvents.
source.read() blocks; returns one character or RESIZE
source.pending(timeout) True if a character arrives within `timeout` s
"""
ESC_TIMEOUT = 0.05
_CTRL = {"\r": "enter", "\n": "newline", "\x7f": "backspace", "\x08": "backspace",
"\x03": "ctrl_c", "\x04": "ctrl_d", "\x01": "home", "\x05": "end"}
_CSI_FINAL = {"A": "up", "B": "down", "C": "right", "D": "left", "H": "home", "F": "end"}
_CSI_TILDE = {"1": "home", "7": "home", "4": "end", "8": "end", "3": "delete"}
def __init__(self, source) -> None:
self.source = source
self._resized = False
def read_key(self) -> KeyEvent:
if self._resized:
self._resized = False
return KeyEvent("resize")
ch = self.source.read()
if ch is RESIZE:
return KeyEvent("resize")
if ch == "\x1b":
return self._escape()
if ch == "\x16": # Ctrl+V passed through by the terminal
text = _clipboard()
return KeyEvent("text", text) if text else KeyEvent("ignore")
if ch == "\t":
return KeyEvent("text", " ")
if ch in self._CTRL:
return KeyEvent(self._CTRL[ch])
if unicodedata.category(ch) != "Cc":
return KeyEvent("text", ch)
return KeyEvent("ignore")
def _next(self) -> str:
"""Next character inside an escape sequence; remembers resizes for later."""
while True:
ch = self.source.read()
if ch is RESIZE:
self._resized = True
continue
return ch
def _escape(self) -> KeyEvent:
if not self.source.pending(self.ESC_TIMEOUT):
return KeyEvent("ignore") # a lone Esc
c = self._next()
if c in ("\r", "\n"):
return KeyEvent("newline") # Alt+Enter
if c == "O": # SS3: cursor keys in application mode
return KeyEvent(self._CSI_FINAL.get(self._next(), "ignore"))
if c != "[":
return KeyEvent("ignore") # Alt+<key>
seq = ""
while True:
d = self._next()
seq += d
if "\x40" <= d <= "\x7e":
break
if len(seq) > 32:
return KeyEvent("ignore")
return self._csi(seq)
def _csi(self, seq: str) -> KeyEvent:
final, params = seq[-1], seq[:-1]
if seq == "200~": # bracketed paste
return self._paste()
if final in self._CSI_FINAL:
return KeyEvent(self._CSI_FINAL[final])
if final == "~":
return KeyEvent(self._CSI_TILDE.get(params.split(";")[0], "ignore"))
if final == "u" and params.split(";")[0] == "13": # kitty protocol: Shift/Ctrl+Enter
return KeyEvent("newline")
return KeyEvent("ignore")
def _paste(self) -> KeyEvent:
end, buf = "\x1b[201~", []
while True:
buf.append(self._next())
if len(buf) >= len(end) and "".join(buf[-len(end):]) == end:
return KeyEvent("text", "".join(buf[: -len(end)]))
def _clipboard() -> str:
"""Best-effort read of the system clipboard (fallback for a raw Ctrl+V)."""
for cmd in (["pbpaste"], ["wl-paste", "-n"], ["xclip", "-selection", "clipboard", "-o"],
["xsel", "-b", "-o"], ["powershell", "-NoProfile", "-Command", "Get-Clipboard"]):
if shutil.which(cmd[0]):
try:
return subprocess.run(cmd, capture_output=True, text=True, timeout=2).stdout
except Exception: # noqa: BLE001
continue
return ""
# -------------------------------------------------------- Terminal plumbing
class _TtySource:
"""Character source for a POSIX tty; `wake_fd` (optional) signals resizes."""
def __init__(self, fd: int, wake_fd: int | None = None) -> None:
self.fd, self.wake = fd, wake_fd
self._dec = codecs.getincrementaldecoder("utf-8")(errors="replace")
self._buf: list[str] = [] # decoded but not yet consumed characters
def read(self):
fds = [self.fd] if self.wake is None else [self.fd, self.wake]
while not self._buf:
ready, _, _ = select.select(fds, [], [])
if self.wake in ready:
try:
os.read(self.wake, 64)
except OSError:
pass
return RESIZE
# Read whole chunks: a large paste must be drained quickly, the
# kernel's tty input queue is small (1 KB on macOS).
b = os.read(self.fd, 65536)
if not b:
raise EOFError
self._buf.extend(self._dec.decode(b))
return self._buf.pop(0)
def pending(self, timeout: float) -> bool:
if self._buf:
return True
ready, _, _ = select.select([self.fd], [], [], timeout)
return bool(ready)
class _WinSource:
"""Character source for the Windows console (msvcrt); special keys become CSI sequences."""
_SPECIAL = {"H": "A", "P": "B", "K": "D", "M": "C", "G": "H", "O": "F", "S": "3~"}
def __init__(self) -> None:
self._queue: list[str] = []
def read(self):
if self._queue:
return self._queue.pop(0)
ch = msvcrt.getwch()
if ch in ("\x00", "\xe0"):
code = self._SPECIAL.get(msvcrt.getwch())
if code is None:
return self.read()
self._queue.extend("[" + code)
return "\x1b"
return ch
def pending(self, timeout: float) -> bool:
return bool(self._queue) or msvcrt.kbhit()
class _RawMode:
"""Character-at-a-time input without echo; Enter stays '\\r', Ctrl+J stays '\\n'."""
def __init__(self, fd: int) -> None:
self.fd = fd
def __enter__(self):
if termios is None:
return self
self.old = termios.tcgetattr(self.fd)
new = termios.tcgetattr(self.fd)
new[0] &= ~(termios.ICRNL | termios.INLCR | termios.IXON) # iflag
new[3] &= ~(termios.ECHO | termios.ICANON | termios.ISIG | termios.IEXTEN) # lflag
new[6][termios.VMIN], new[6][termios.VTIME] = 1, 0
termios.tcsetattr(self.fd, termios.TCSANOW, new)
return self
def __exit__(self, *exc):
if termios is not None:
termios.tcsetattr(self.fd, termios.TCSADRAIN, self.old)
class _ResizeWatcher:
"""Turns SIGWINCH into a readable byte on `self.fd` (None when unsupported)."""
def __enter__(self):
self.fd = None
if not hasattr(signal, "SIGWINCH"):
return self
r, w = os.pipe()
try:
os.set_blocking(w, False)
self._old_wake = signal.set_wakeup_fd(w)
self._old_handler = signal.signal(signal.SIGWINCH, lambda *_: None)
self.fd, self._pipe = r, (r, w)
except (ValueError, OSError): # e.g. not in the main thread
os.close(r), os.close(w)
return self
def __exit__(self, *exc):
if self.fd is None:
return
signal.signal(signal.SIGWINCH, self._old_handler)
signal.set_wakeup_fd(self._old_wake)
for fd in self._pipe:
os.close(fd)
def _terminal_size() -> tuple[int, int]:
size = shutil.get_terminal_size((80, 24))
return size.columns, size.lines
def _enable_windows_vt() -> None:
try:
import ctypes
k32 = ctypes.windll.kernel32
handle, mode = k32.GetStdHandle(-11), ctypes.c_uint32()
if k32.GetConsoleMode(handle, ctypes.byref(mode)):
k32.SetConsoleMode(handle, mode.value | 0x0004) # ENABLE_VIRTUAL_TERMINAL_PROCESSING
except Exception: # noqa: BLE001
pass
def _interactive() -> bool:
try:
return sys.stdin.isatty() and sys.stdout.isatty() and (termios or msvcrt) is not None
except (AttributeError, ValueError):
return False
# ------------------------------------------------------------------- prompt
def prompt(header: str = "", text: str = "") -> str:
"""Prints `header` (if any) and edits `text` on the following line(s).
Returns the entered text on Enter. Raises Abort (a KeyboardInterrupt with
the current input in `.text`) on Ctrl+C and EOFError on Ctrl+D. Without a
terminal (pipe), one line is read from stdin instead.
"""
if header:
print(header, flush=True)
if not _interactive():
line = sys.stdin.readline()
if not line:
raise EOFError
return line.rstrip("\r\n")
out = sys.stdout
editor = Editor(out, _terminal_size, text)
if termios is None:
_enable_windows_vt()
source, fd = _WinSource(), None
else:
fd = sys.stdin.fileno()
with _RawMode(fd), _ResizeWatcher() as watcher:
if termios is not None:
source = _TtySource(fd, watcher.fd)
reader = KeyReader(source)
out.write("\x1b[?25h\x1b[?2004h") # cursor visible, bracketed paste on
try:
editor.render()
while not editor.handle(reader.read_key()):
pass
finally:
editor.finish()
out.write("\x1b[?2004l")
out.flush()
return editor.text
if __name__ == "__main__":
try:
result = prompt("jprompt demo - Enter submits, Ctrl+J = new line, Ctrl+C aborts, Ctrl+D exits\nText>")
print(f"You entered: {result!r}")
except KeyboardInterrupt:
print("(aborted)")
except EOFError:
print("(exit)")