minu 0.1.0__tar.gz

minu-0.1.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 mr-jones123
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

minu-0.1.0/PKG-INFO

@@ -0,0 +1,134 @@
+Metadata-Version: 2.4
+Name: minu
+Version: 0.1.0
+Summary: Tiny agent harness for local models (llama.cpp, MLX, Ollama)
+Keywords: agent,llm,local-llm,llama.cpp,mlx,ollama,tui,coding-agent
+Author: mr-jones123
+Author-email: mr-jones123 <lacapxyniljhed@gmail.com>
+License-Expression: MIT
+License-File: LICENSE
+Classifier: Development Status :: 3 - Alpha
+Classifier: Environment :: Console
+Classifier: Intended Audience :: Developers
+Classifier: Operating System :: MacOS
+Classifier: Operating System :: POSIX :: Linux
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Topic :: Software Development
+Requires-Dist: textual>=1.0
+Requires-Dist: textual-image>=0.8
+Requires-Dist: pillow>=10.3
+Requires-Dist: mlx-lm ; platform_machine == 'arm64' and sys_platform == 'darwin' and extra == 'mlx'
+Requires-Python: >=3.13
+Project-URL: Homepage, https://github.com/mr-jones123/minu
+Project-URL: Repository, https://github.com/mr-jones123/minu
+Project-URL: Issues, https://github.com/mr-jones123/minu/issues
+Provides-Extra: mlx
+Description-Content-Type: text/markdown
+
+# minu
+
+A very small agent harness for local models. One tool (`sh`) and a system prompt of about 86 tokens.
+
+```
+uv tool install minu              # or: pipx install minu   (needs Python 3.13; uv fetches it for you)
+minu                              # run it in any project folder
+```
+Upgrade with `uv tool upgrade minu`, remove with `uv tool uninstall minu`. Sessions are saved in `~/.minu/`.
+
+```
+minu                              # opens the model picker
+minu --hf LiquidAI/LFM3-1.7B-GGUF
+minu --mlx mlx-community/Qwen3-1.7B-4bit
+minu --ollama gemma3:4b
+minu --url http://localhost:8080           # server that's already running
+```
+
+`-y` runs commands without asking. `--native` uses the server's tool-call API instead of ```sh blocks.
+
+## Models
+`ctrl+o` (or `/model`) opens the picker. It has:
+- the GGUF and MLX models in your Hugging Face cache, plus your Ollama models
+- **⬇ download from Hugging Face**: search, or paste `org/repo`, then pick a size. For vision repos the `mmproj` file is downloaded too.
+- **⬇ pull from Ollama**: any name from ollama.com/library, with progress shown in the status bar.
+
+Switching models stops the old server and starts the new one. The chat stays.
+
+**✕ delete a model** (in the picker, or `/delete`) lists your downloaded models with their size on disk. It asks before
+deleting (the default is "keep it"). For GGUF it removes that quant, plus the whole repo (mmproj, cache entry) once
+no weights are left in it; MLX repos are removed whole; Ollama models go through `ollama`'s API. A loaded model is stopped first.
+`esc` cancels a download.
+
+If something is missing, minu offers to install it (it always asks first):
+| needed for | tool | install |
+|---|---|---|
+| GGUF | llama-server | `brew install llama.cpp` |
+| MLX (Apple Silicon only) | mlx_lm | `uv tool install mlx-lm` |
+| HF downloads | hf | `uv tool install huggingface_hub` |
+| Ollama | ollama | `brew install ollama` / ollama.com install script |
+
+If Ollama is installed but not running, minu starts `ollama serve` itself and stops it when you quit.
+
+## Images
+Put an image path in your message (dragging a file into the terminal pastes its path):
+`what's in ~/Desktop/shot.png?`
+The image is shown in the chat and sent to the model. That needs a vision model: GGUF repos with an
+mmproj (marked `[vision]` in the picker), or an Ollama vision model such as `gemma3:4b`. MLX models are text only.
+Image files the model mentions in its answer are shown too, and `/view path.png` shows any image.
+Images render as real pixels in Kitty, Ghostty, WezTerm and iTerm2, and as colored blocks in other terminals.
+
+## Sessions
+Each chat is saved to `~/.minu/sessions/<project>/<timestamp>.jsonl` (one folder per project path, e.g.
+`-Users-me-code-app`): a `{"meta": …}` line, then one message per line. Set `MINU_HOME` to move `~/.minu`.
+Sessions from older versions in `./.minu/sessions` still show up in `/resume`.
+`/resume` opens a session picker · `minu -c` continues the latest one · `minu -r FILE` opens a specific file · `/new` (or `ctrl+n`) starts a new session.
+
+## Keys
+`enter` send · `ctrl+j` newline · `ctrl+o` model · `ctrl+y` auto-run · `esc` stop · `ctrl+n` new session · `ctrl+c` quit
+Type `/` for a menu of commands (↑/↓ to move, tab to complete, enter to run, esc to close):
+`/new` `/model` `/delete` `/resume` `/search <query>` `/fetch <url>` `/view <img>` `/auto` `/clear` `/exit`
+
+## Web
+The model can search on its own (`minu-search "q"`, `minu-fetch URL` in its shell). You can also do it yourself:
+`/search <query>` gets results and the model answers from them; `/fetch <url>` reads a page and the model summarizes it.
+Search tries DuckDuckGo first, then Wikipedia's API. DuckDuckGo sometimes rate-limits; for search that never gets blocked,
+point minu at your own SearXNG instance: `MINU_SEARXNG=http://localhost:8888`.
+Env: `MINU_CTX` (llama.cpp context, default 16384), `MINU_HOME` (default `~/.minu`), `OLLAMA_HOST`, `HF_HOME`. Server log: `~/.cache/minu/server.log`.
+
+## Development
+```
+git clone https://github.com/mr-jones123/minu && cd minu
+uv tool install -e .              # global `minu` that runs your checkout (changes to src/ apply right away)
+uv run minu                       # or run it from the repo
+uv build                          # sdist + wheel into dist/
+```
+
+## Releasing to PyPI
+Publishing runs in GitHub Actions (`.github/workflows/publish.yml`) with PyPI Trusted Publishing, so no token is stored.
+One-time setup: on pypi.org → *Your account → Publishing → Add a pending publisher*: project `minu`,
+owner `mr-jones123`, repo `minu`, workflow `publish.yml`, environment `pypi`. Then for each release:
+```
+uv version --bump patch           # or minor / major; edits pyproject.toml
+git commit -am "release v$(uv version --short)" && git tag v$(uv version --short) && git push --follow-tags
+```
+CI builds, smoke-tests the wheel, checks the tag matches the version, and uploads.
+
+## Code layout (`src/minu/`, the `minu` package)
+| file | what it does |
+|---|---|
+| `cli.py` | command-line flags, plus `minu search` / `minu fetch` for the agent's shell |
+| `config.py` | paths, limits, system prompts, banner |
+| `app.py` | the TUI: layout, status bar, input handling |
+| `agent.py` | agent loop: stream a reply, run its command, send the output back |
+| `model_flows.py` | model picker, HF/Ollama downloads, install offers, loading |
+| `sessions.py` | saving and resuming `~/.minu/sessions/<project>/*.jsonl` |
+| `server.py` | starting llama-server / mlx_lm / Ollama and streaming chat |
+| `models.py` | finding local models and searching Hugging Face |
+| `tools.py` | the `sh` tool, web search/fetch, parsing commands from replies |
+| `images.py` | attaching images and showing them |
+| `installs.py` | finding tools and their install commands |
+| `widgets.py` | input box, list picker, text prompt |
+
+## License
+MIT, see [LICENSE](LICENSE).

minu-0.1.0/README.md

@@ -0,0 +1,105 @@
+# minu
+
+A very small agent harness for local models. One tool (`sh`) and a system prompt of about 86 tokens.
+
+```
+uv tool install minu              # or: pipx install minu   (needs Python 3.13; uv fetches it for you)
+minu                              # run it in any project folder
+```
+Upgrade with `uv tool upgrade minu`, remove with `uv tool uninstall minu`. Sessions are saved in `~/.minu/`.
+
+```
+minu                              # opens the model picker
+minu --hf LiquidAI/LFM3-1.7B-GGUF
+minu --mlx mlx-community/Qwen3-1.7B-4bit
+minu --ollama gemma3:4b
+minu --url http://localhost:8080           # server that's already running
+```
+
+`-y` runs commands without asking. `--native` uses the server's tool-call API instead of ```sh blocks.
+
+## Models
+`ctrl+o` (or `/model`) opens the picker. It has:
+- the GGUF and MLX models in your Hugging Face cache, plus your Ollama models
+- **⬇ download from Hugging Face**: search, or paste `org/repo`, then pick a size. For vision repos the `mmproj` file is downloaded too.
+- **⬇ pull from Ollama**: any name from ollama.com/library, with progress shown in the status bar.
+
+Switching models stops the old server and starts the new one. The chat stays.
+
+**✕ delete a model** (in the picker, or `/delete`) lists your downloaded models with their size on disk. It asks before
+deleting (the default is "keep it"). For GGUF it removes that quant, plus the whole repo (mmproj, cache entry) once
+no weights are left in it; MLX repos are removed whole; Ollama models go through `ollama`'s API. A loaded model is stopped first.
+`esc` cancels a download.
+
+If something is missing, minu offers to install it (it always asks first):
+| needed for | tool | install |
+|---|---|---|
+| GGUF | llama-server | `brew install llama.cpp` |
+| MLX (Apple Silicon only) | mlx_lm | `uv tool install mlx-lm` |
+| HF downloads | hf | `uv tool install huggingface_hub` |
+| Ollama | ollama | `brew install ollama` / ollama.com install script |
+
+If Ollama is installed but not running, minu starts `ollama serve` itself and stops it when you quit.
+
+## Images
+Put an image path in your message (dragging a file into the terminal pastes its path):
+`what's in ~/Desktop/shot.png?`
+The image is shown in the chat and sent to the model. That needs a vision model: GGUF repos with an
+mmproj (marked `[vision]` in the picker), or an Ollama vision model such as `gemma3:4b`. MLX models are text only.
+Image files the model mentions in its answer are shown too, and `/view path.png` shows any image.
+Images render as real pixels in Kitty, Ghostty, WezTerm and iTerm2, and as colored blocks in other terminals.
+
+## Sessions
+Each chat is saved to `~/.minu/sessions/<project>/<timestamp>.jsonl` (one folder per project path, e.g.
+`-Users-me-code-app`): a `{"meta": …}` line, then one message per line. Set `MINU_HOME` to move `~/.minu`.
+Sessions from older versions in `./.minu/sessions` still show up in `/resume`.
+`/resume` opens a session picker · `minu -c` continues the latest one · `minu -r FILE` opens a specific file · `/new` (or `ctrl+n`) starts a new session.
+
+## Keys
+`enter` send · `ctrl+j` newline · `ctrl+o` model · `ctrl+y` auto-run · `esc` stop · `ctrl+n` new session · `ctrl+c` quit
+Type `/` for a menu of commands (↑/↓ to move, tab to complete, enter to run, esc to close):
+`/new` `/model` `/delete` `/resume` `/search <query>` `/fetch <url>` `/view <img>` `/auto` `/clear` `/exit`
+
+## Web
+The model can search on its own (`minu-search "q"`, `minu-fetch URL` in its shell). You can also do it yourself:
+`/search <query>` gets results and the model answers from them; `/fetch <url>` reads a page and the model summarizes it.
+Search tries DuckDuckGo first, then Wikipedia's API. DuckDuckGo sometimes rate-limits; for search that never gets blocked,
+point minu at your own SearXNG instance: `MINU_SEARXNG=http://localhost:8888`.
+Env: `MINU_CTX` (llama.cpp context, default 16384), `MINU_HOME` (default `~/.minu`), `OLLAMA_HOST`, `HF_HOME`. Server log: `~/.cache/minu/server.log`.
+
+## Development
+```
+git clone https://github.com/mr-jones123/minu && cd minu
+uv tool install -e .              # global `minu` that runs your checkout (changes to src/ apply right away)
+uv run minu                       # or run it from the repo
+uv build                          # sdist + wheel into dist/
+```
+
+## Releasing to PyPI
+Publishing runs in GitHub Actions (`.github/workflows/publish.yml`) with PyPI Trusted Publishing, so no token is stored.
+One-time setup: on pypi.org → *Your account → Publishing → Add a pending publisher*: project `minu`,
+owner `mr-jones123`, repo `minu`, workflow `publish.yml`, environment `pypi`. Then for each release:
+```
+uv version --bump patch           # or minor / major; edits pyproject.toml
+git commit -am "release v$(uv version --short)" && git tag v$(uv version --short) && git push --follow-tags
+```
+CI builds, smoke-tests the wheel, checks the tag matches the version, and uploads.
+
+## Code layout (`src/minu/`, the `minu` package)
+| file | what it does |
+|---|---|
+| `cli.py` | command-line flags, plus `minu search` / `minu fetch` for the agent's shell |
+| `config.py` | paths, limits, system prompts, banner |
+| `app.py` | the TUI: layout, status bar, input handling |
+| `agent.py` | agent loop: stream a reply, run its command, send the output back |
+| `model_flows.py` | model picker, HF/Ollama downloads, install offers, loading |
+| `sessions.py` | saving and resuming `~/.minu/sessions/<project>/*.jsonl` |
+| `server.py` | starting llama-server / mlx_lm / Ollama and streaming chat |
+| `models.py` | finding local models and searching Hugging Face |
+| `tools.py` | the `sh` tool, web search/fetch, parsing commands from replies |
+| `images.py` | attaching images and showing them |
+| `installs.py` | finding tools and their install commands |
+| `widgets.py` | input box, list picker, text prompt |
+
+## License
+MIT, see [LICENSE](LICENSE).

minu-0.1.0/pyproject.toml

@@ -0,0 +1,53 @@
+[project]
+name = "minu"
+version = "0.1.0"
+description = "Tiny agent harness for local models (llama.cpp, MLX, Ollama)"
+readme = "README.md"
+requires-python = ">=3.13"
+license = "MIT"
+license-files = ["LICENSE"]
+keywords = [
+    "agent",
+    "llm",
+    "local-llm",
+    "llama.cpp",
+    "mlx",
+    "ollama",
+    "tui",
+    "coding-agent",
+]
+classifiers = [
+    "Development Status :: 3 - Alpha",
+    "Environment :: Console",
+    "Intended Audience :: Developers",
+    "Operating System :: MacOS",
+    "Operating System :: POSIX :: Linux",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.13",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+    "Topic :: Software Development",
+]
+dependencies = [
+    "textual>=1.0",
+    "textual-image>=0.8",
+    "pillow>=10.3",
+]
+
+[[project.authors]]
+name = "mr-jones123"
+email = "lacapxyniljhed@gmail.com"
+
+[project.optional-dependencies]
+mlx = ["mlx-lm; sys_platform == 'darwin' and platform_machine == 'arm64'"]
+
+[project.urls]
+Homepage = "https://github.com/mr-jones123/minu"
+Repository = "https://github.com/mr-jones123/minu"
+Issues = "https://github.com/mr-jones123/minu/issues"
+
+[project.scripts]
+minu = "minu.cli:main"
+
+[build-system]
+requires = ["uv_build>=0.9,<0.10"]
+build-backend = "uv_build"

minu-0.1.0/pyproject.toml.orig

@@ -0,0 +1,37 @@
+[project]
+name = "minu"
+version = "0.1.0"
+description = "Tiny agent harness for local models (llama.cpp, MLX, Ollama)"
+readme = "README.md"
+requires-python = ">=3.13"
+license = "MIT"
+license-files = ["LICENSE"]
+authors = [{ name = "mr-jones123", email = "lacapxyniljhed@gmail.com" }]
+keywords = ["agent", "llm", "local-llm", "llama.cpp", "mlx", "ollama", "tui", "coding-agent"]
+classifiers = [
+    "Development Status :: 3 - Alpha",
+    "Environment :: Console",
+    "Intended Audience :: Developers",
+    "Operating System :: MacOS",
+    "Operating System :: POSIX :: Linux",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.13",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+    "Topic :: Software Development",
+]
+dependencies = ["textual>=1.0", "textual-image>=0.8", "pillow>=10.3"]
+
+[project.optional-dependencies]
+mlx = ["mlx-lm; sys_platform == 'darwin' and platform_machine == 'arm64'"]
+
+[project.urls]
+Homepage = "https://github.com/mr-jones123/minu"
+Repository = "https://github.com/mr-jones123/minu"
+Issues = "https://github.com/mr-jones123/minu/issues"
+
+[project.scripts]
+minu = "minu.cli:main"
+
+[build-system]
+requires = ["uv_build>=0.9,<0.10"]
+build-backend = "uv_build"

minu-0.1.0/src/minu/__init__.py

@@ -0,0 +1,5 @@
+"""minu: a tiny agent harness for local models (llama.cpp / MLX / Ollama)."""
+from importlib.metadata import PackageNotFoundError, version
+
+try: __version__ = version("minu")
+except PackageNotFoundError: __version__ = "0.0.0"

minu-0.1.0/src/minu/__main__.py

@@ -0,0 +1,3 @@
+from .cli import main
+
+main()

minu-0.1.0/src/minu/agent.py

@@ -0,0 +1,119 @@
+"""The agent loop: stream a reply, run its command, feed the output back."""
+import json, re, time
+
+from textual import work
+from textual.widgets import Markdown
+
+from .config import BLOCK, STEPS
+from .images import mentioned_images
+from .tools import fetch, search, sh, split_think, to_cmd
+
+
+class AgentMixin:
+    def approve(s, cmd):
+        if s.auto: s.ui(s.note, f"$ {cmd}", "cmd"); return True
+        s.ui(s.note, f"$ {cmd}", "cmd")
+        return s.ask("run it?")
+
+    def exec_cmd(s, cmd):
+        if not s.approve(cmd): return "[user declined to run this]"
+        s.set_state("running…")
+        out = sh(cmd); s.state = ""
+        lines = out.splitlines()
+        s.ui(s.note, "\n".join(lines[:15]) + (f"\n… ({len(lines) - 15} more lines)" if len(lines) > 15 else ""), "out")
+        return out
+
+    def web(s, kind, arg):  # /search and /fetch: look it up, show it, hand it to the model
+        s.set_state(f"{'searching' if kind == 'search' else 'fetching'}…")
+        try: res = search(arg) if kind == "search" else fetch(arg)
+        except Exception as e: res = f"{kind} failed: {e}"
+        lines = res.splitlines()
+        s.ui(s.note, "\n".join(lines[:20]) + (f"\n… ({len(lines) - 20} more lines)" if len(lines) > 20 else ""), "out")
+        if not s.srv.url: return s.say("▸ no model loaded, so just showing results (ctrl+o to pick one)")
+        ask = (f'Web search results for "{arg}":\n\n{res}\n\nAnswer "{arg}" now from these results, citing URLs. Do not run any commands.' if kind == "search"
+               else f"Contents of {arg}:\n\n{res}\n\nSummarize this page.")
+        s.push({"role": "user", "content": ask}); return True
+
+    @work(thread=True, exclusive=True, group="turn")
+    def turn(s, pre=None):
+        s.busy = True; s.stop_ev.clear()
+        try:
+            if not pre or s.web(*pre):  # web() returns None when there's no model to answer
+                for _ in range(STEPS):
+                    if s.step(): break
+        except Exception as e:
+            msg = str(e)
+            if re.search(r"image|mmproj|multimodal|vision", msg, re.I) and any(isinstance(m["content"], list) for m in s.msgs):
+                for m in s.msgs:  # drop images so the chat can continue
+                    if isinstance(m["content"], list):
+                        m["content"] = m["content"][0]["text"] + "\n[image removed: model can't see images]"
+                msg = "this model can't read images. Pick one marked [vision] (ctrl+o), or e.g. gemma3:4b on Ollama"
+            s.say(f"▸ error: {msg}")
+        s.busy = False; s.set_state("")
+
+    def step(s):
+        md = think_box = think_w = None
+        raw, reasoning, calls, block = "", "", {}, None
+        last, first, think_t, n, final_tps = 0.0, None, None, 0, None
+        s.set_state("waiting…")
+        for ch in s.srv.stream(s.msgs, s.native, s.stop_ev):
+            if u := ch.get("usage"):
+                s.used = u.get("total_tokens") or u.get("prompt_tokens", 0) + u.get("completion_tokens", 0)
+                if u.get("completion_tokens") and first: final_tps = u["completion_tokens"] / max(time.time() - first, 1e-3)
+            if (tm := ch.get("timings")) and tm.get("predicted_per_second"): final_tps = tm["predicted_per_second"]  # llama.cpp
+            if not ch.get("choices"): continue
+            d = ch["choices"][0].get("delta") or {}
+            piece, rpiece = d.get("content") or "", d.get("reasoning_content") or d.get("reasoning") or ""
+            raw += piece; reasoning += rpiece
+            for tc in d.get("tool_calls") or []:
+                c = calls.setdefault(tc.get("index", 0), {"id": "", "name": "", "args": ""})
+                f = tc.get("function") or {}
+                c["id"] = tc.get("id") or c["id"]; c["name"] += f.get("name") or ""; c["args"] += f.get("arguments") or ""
+            if piece or rpiece or d.get("tool_calls"):
+                n += 1; first = first or time.time()
+            think, text = s.split(reasoning, raw)
+            if not s.native and (block := BLOCK.search(text)):
+                break
+            if time.time() - last > 0.08:
+                last = time.time()
+                if n > 1: s.tps = (n - 1) / max(last - first, 1e-3)  # ~1 token per streamed chunk
+                if think:
+                    think_t = think_t or time.time()
+                    if not think_box: think_box, think_w = s.ui(s.add_think)
+                    s.ui(s.render, think_w, think)
+                if text:
+                    if not md: md = s.ui(s.add, Markdown(""))
+                    if think_box and not think_box.collapsed: s.ui(s.end_think, think_box, time.time() - think_t)
+                    s.ui(s.render, md, text)
+                s.set_state("writing…" if text else "thinking…" if think else "waiting…")
+        think, text = s.split(reasoning, raw)
+        if block: text = text[:block.end()]
+        if final_tps: s.tps = final_tps
+        if think:
+            if not think_box: think_box, think_w = s.ui(s.add_think)
+            s.ui(s.render, think_w, think); s.ui(s.end_think, think_box, time.time() - (think_t or first or time.time()))
+        if text or not calls:
+            if not md: md = s.ui(s.add, Markdown(""))
+            s.ui(s.render, md, text or "_(empty)_")
+        if s.stop_ev.is_set():
+            s.push({"role": "assistant", "content": text}); s.say("▸ stopped"); return True
+        if calls:
+            s.push({"role": "assistant", "content": text, "tool_calls": [
+                {"id": c["id"] or f"call{i}", "type": "function", "function": {"name": c["name"], "arguments": c["args"]}}
+                for i, c in calls.items()]})
+            for i, c in calls.items():
+                try: cmd = json.loads(c["args"] or "{}").get("command", "")
+                except json.JSONDecodeError: cmd = c["args"]
+                s.push({"role": "tool", "tool_call_id": c["id"] or f"call{i}", "content": s.exec_cmd(cmd)})
+            return s.stop_ev.is_set()
+        s.push({"role": "assistant", "content": text})
+        if block:
+            s.push({"role": "user", "content": "Output:\n" + s.exec_cmd(to_cmd(block))})
+            return s.stop_ev.is_set()
+        for p in mentioned_images(text): s.ui(s.show_image, p)
+        return True
+
+    @staticmethod
+    def split(reasoning, raw):  # reasoning field (llama.cpp/Ollama) + inline <think> tags -> (thinking, answer)
+        think, text = split_think(raw)
+        return "\n".join(x for x in (reasoning.strip(), think) if x), text