kon-coding-agent 0.2.0__tar.gz → 0.2.2__tar.gz

kon_coding_agent-0.2.2/.kon/skills/kon-release-publish/SKILL.md

@@ -0,0 +1,85 @@
+---
+name: kon-release-publish
+description: Tag, publish to PyPI, and create GitHub release for Kon with validation and rollback-safe steps
+---
+
+# Kon Release + PyPI Publish
+
+Use this skill when the user asks to cut a new Kon version, tag it, publish to PyPI, and/or create a GitHub release.
+
+## Inputs to confirm
+
+- Target version (example: `0.2.1`)
+- Base range for notes (usually previous tag, example: `v0.2.0..HEAD`)
+- Whether to push `main`
+- Whether to publish to PyPI now
+- Whether to create GitHub release now
+
+## Files to bump
+
+- `pyproject.toml` → `[project].version`
+- `src/kon/ui/app.py` → fallback `VERSION = "..."`
+- `uv.lock` → local package version block
+
+## Release workflow
+
+1. **Preflight**
+   - `git status --short --branch` must be clean (or confirm with user)
+   - `git tag --list` and `git log --oneline <prev_tag>..HEAD` to summarize changes
+
+2. **Version bump**
+   - Update version in all 3 files above
+
+3. **Quality gates**
+   - `uv run ruff format .`
+   - `uv run ruff check .`
+   - `uv run pyright .`
+   - `uv run pytest`
+
+4. **Commit**
+   - Commit message: `build: bump version to <version>`
+
+5. **Tag**
+   - Annotated tag: `git tag -a v<version> -m "v<version> ..."`
+   - Include concise “changes since previous tag” bullets
+
+6. **Push**
+   - `git push origin main`
+   - `git push origin v<version>`
+
+7. **Build + verify artifacts**
+   - `rm -rf dist && uv build`
+   - `uv run python -m twine check dist/*`
+
+8. **Publish to PyPI**
+   - Prefer token file if present (example `~/.pypi-token`):
+   - `TWINE_USERNAME=__token__ TWINE_PASSWORD="$(< ~/.pypi-token)" uv run python -m twine upload dist/*`
+   - Verify:
+     - `https://pypi.org/project/kon-coding-agent/<version>/`
+     - `https://pypi.org/pypi/kon-coding-agent/json` reports latest version
+
+9. **Create GitHub release**
+   - If token exists at `~/.github-token`, call Releases API:
+   - `POST /repos/<owner>/<repo>/releases` with:
+     - `tag_name: v<version>`
+     - `target_commitish: main`
+     - `name: v<version>`
+     - `generate_release_notes: true`
+   - If 403 occurs, report missing token scopes/permissions (`contents:write` required)
+
+## Important notes
+
+- **Tagging and GitHub release are separate**:
+  - Tag = git ref in repository
+  - Release = GitHub object attached to a tag (notes/assets)
+- You can do either independently, but most projects do both together for user-facing releases.
+- If PyPI publish succeeds but GitHub release fails, do **not** retag/re-publish. Just fix auth and create the release for the existing tag.
+
+## Output checklist to report
+
+- Version bumped in all files
+- Checks passed
+- Commit hash
+- Tag created and pushed
+- PyPI upload URL
+- GitHub release URL (or exact error + remediation)

kon_coding_agent-0.2.2/LOCAL.md

@@ -0,0 +1,25 @@
+# Local Models
+
+This document provides detailed information about running and configuring local models with Kon.
+
+## Tested Models
+
+| Model | Quantization | Context Length | TPS | System Specs |
+| ----- | -------------- | -------------- | --- | ------------ |
+| `qwen/qwen3-coder-next` | Q4_K_M | 64,000 | N/A | i7-14700F × 28, 64GB RAM, 24GB VRAM (RTX 3090) |
+| `zai-org/glm-4.7-flash` | Q4_K_M | 64,000 | ~80-90 | i7-14700F × 28, 64GB RAM, 24GB VRAM (RTX 3090) |
+
+Run a local model using llama-server with the following command:
+
+```bash
+./llama-server -m <models-dir>/GLM-4.7-Flash-GGUF/GLM-4.7-Flash-Q4_K_M.gguf -n 8192 -c 64000
+```
+
+Then start kon:
+
+```bash
+kon --model zai-org/glm-4.7-flash --provider openai --base-url http://localhost:8080/v1 --api-key ""
+```
+
+> [!NOTE]
+> I was not able to run qwen-coder-next reliably on my system. Either the provider config had some issues or it's too big for my system (i'm not sure)

{kon_coding_agent-0.2.0 → kon_coding_agent-0.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: kon-coding-agent
-Version: 0.2.0
+Version: 0.2.2
 Summary: Minimal coding agent
 License-File: LICENSE
 Requires-Python: >=3.12
@@ -215,34 +215,32 @@ UI (app.py)
 
 ## Supported Models
 
-Kon works well with local models exposed through an OpenAI-compatible `/v1` API (for example LM Studio).
+Kon works well with local models exposed through an OpenAI-compatible `/v1` API.
 
-### Example on LM Studio
+### Example using llama-server
 
-To run a local model from LM Studio:
+To run a local model using llama-server:
 
 ```bash
-# GLM-4.7-flash
-kon --provider openai-responses \
-  --base-url http://127.0.0.1:1234/v1 \
-  --model zai-org/glm-4.7-flash \
-  --api-key ""
-
-# Qwen3-coder-next
-kon --provider openai-responses \
-  --base-url http://127.0.0.1:1234/v1 \
-  --model qwen/qwen3-coder-next \
-  --api-key ""
+./llama-server -m <models-dir>/GLM-4.7-Flash-GGUF/GLM-4.7-Flash-Q4_K_M.gguf \
+    -n 8192 \
+    -c 64000
+
+# Then use Kon with:
+kon --model zai-org/glm-4.7-flash \
+    --provider openai \
+    --base-url http://localhost:8080/v1 \
+    --api-key ""
 ```
 
-For detailed configuration and performance benchmarks, see [LOCAL.md](LOCAL.md).
+`GLM-4.7-Flash-Q4` ran at 80-90 tps on my i7-14700F × 28, 64GB RAM, 24GB VRAM (RTX 3090)
 
 ### All Supported Providers
 
 | Model (local=*) | Provider | Thinking | Vision |
 | ----- | -------- | -------- | ------ |
-| `*zai-org/glm-4.7-flash` | OpenAI Responses | Yes | No |
-| `*qwen/qwen3-coder-next` | OpenAI Responses | Yes | No |
+| `*zai-org/glm-4.7-flash` | OpenAI Completions | Yes | No |
+| `*qwen/qwen3-coder-next` | OpenAI Completions | Yes | No |
 | `glm-4.7` | ZhiPu (OpenAI Completions) | Yes | No |
 | `glm-5` | ZhiPu (OpenAI Completions) | Yes | No |
 | `claude-sonnet-4.5` | GitHub Copilot | Yes | Yes |
@@ -264,7 +262,8 @@ Most important knobs:
 - `llm.default_thinking_level`
 - `llm.system_prompt` (**you can fully override Kon’s system prompt here**)
 - `llm.tool_call_idle_timeout_seconds` (fallback timeout for stalled tool-call streaming)
-- `compaction.on_overflow`, `compaction.buffer_tokens`, `compaction.default_context_window`
+- `compaction.on_overflow`, `compaction.buffer_tokens`
+- `agent.max_turns`, `agent.default_context_window`
 
 You can also theme the UI via `[ui.colors]` values.
 
@@ -275,7 +274,7 @@ Example:
 default_provider = "openai-codex"
 default_model = "gpt-5.3-codex"
 default_thinking_level = "high"
-tool_call_idle_timeout_seconds = 10
+tool_call_idle_timeout_seconds = 60
 system_prompt = """Your custom system prompt here"""
 
 [compaction]

{kon_coding_agent-0.2.0 → kon_coding_agent-0.2.2}/README.md

@@ -199,34 +199,32 @@ UI (app.py)
 
 ## Supported Models
 
-Kon works well with local models exposed through an OpenAI-compatible `/v1` API (for example LM Studio).
+Kon works well with local models exposed through an OpenAI-compatible `/v1` API.
 
-### Example on LM Studio
+### Example using llama-server
 
-To run a local model from LM Studio:
+To run a local model using llama-server:
 
 ```bash
-# GLM-4.7-flash
-kon --provider openai-responses \
-  --base-url http://127.0.0.1:1234/v1 \
-  --model zai-org/glm-4.7-flash \
-  --api-key ""
-
-# Qwen3-coder-next
-kon --provider openai-responses \
-  --base-url http://127.0.0.1:1234/v1 \
-  --model qwen/qwen3-coder-next \
-  --api-key ""
+./llama-server -m <models-dir>/GLM-4.7-Flash-GGUF/GLM-4.7-Flash-Q4_K_M.gguf \
+    -n 8192 \
+    -c 64000
+
+# Then use Kon with:
+kon --model zai-org/glm-4.7-flash \
+    --provider openai \
+    --base-url http://localhost:8080/v1 \
+    --api-key ""
 ```
 
-For detailed configuration and performance benchmarks, see [LOCAL.md](LOCAL.md).
+`GLM-4.7-Flash-Q4` ran at 80-90 tps on my i7-14700F × 28, 64GB RAM, 24GB VRAM (RTX 3090)
 
 ### All Supported Providers
 
 | Model (local=*) | Provider | Thinking | Vision |
 | ----- | -------- | -------- | ------ |
-| `*zai-org/glm-4.7-flash` | OpenAI Responses | Yes | No |
-| `*qwen/qwen3-coder-next` | OpenAI Responses | Yes | No |
+| `*zai-org/glm-4.7-flash` | OpenAI Completions | Yes | No |
+| `*qwen/qwen3-coder-next` | OpenAI Completions | Yes | No |
 | `glm-4.7` | ZhiPu (OpenAI Completions) | Yes | No |
 | `glm-5` | ZhiPu (OpenAI Completions) | Yes | No |
 | `claude-sonnet-4.5` | GitHub Copilot | Yes | Yes |
@@ -248,7 +246,8 @@ Most important knobs:
 - `llm.default_thinking_level`
 - `llm.system_prompt` (**you can fully override Kon’s system prompt here**)
 - `llm.tool_call_idle_timeout_seconds` (fallback timeout for stalled tool-call streaming)
-- `compaction.on_overflow`, `compaction.buffer_tokens`, `compaction.default_context_window`
+- `compaction.on_overflow`, `compaction.buffer_tokens`
+- `agent.max_turns`, `agent.default_context_window`
 
 You can also theme the UI via `[ui.colors]` values.
 
@@ -259,7 +258,7 @@ Example:
 default_provider = "openai-codex"
 default_model = "gpt-5.3-codex"
 default_thinking_level = "high"
-tool_call_idle_timeout_seconds = 10
+tool_call_idle_timeout_seconds = 60
 system_prompt = """Your custom system prompt here"""
 
 [compaction]

{kon_coding_agent-0.2.0 → kon_coding_agent-0.2.2}/TODO.md

@@ -1,5 +1,3 @@
-- show new release update in ui to prompt the user to upgrade
-- show tokens streamed in for bash, edit and write tools (which can large at times)
 - if @ or / menu is open pressing esc closes it but interrupts stream as well
 - bug in how we report context size, tokens in and out and cached tokens for codex?
 - steer (immediate) and normal queues

{kon_coding_agent-0.2.0 → kon_coding_agent-0.2.2}/pyproject.toml

@@ -14,7 +14,7 @@ default = true
 
 [project]
 name = "kon-coding-agent"
-version = "0.2.0"
+version = "0.2.2"
 description = "Minimal coding agent"
 readme = "README.md"
 requires-python = ">=3.12"

{kon_coding_agent-0.2.0 → kon_coding_agent-0.2.2}/src/kon/config.py

@@ -59,12 +59,16 @@ class LLMConfig(BaseModel):
     default_model: str
     default_thinking_level: str
     system_prompt: str
-    tool_call_idle_timeout_seconds: float = 10
+    tool_call_idle_timeout_seconds: float = 60
 
 
 class CompactionConfig(BaseModel):
     on_overflow: OnOverflowMode = "continue"
     buffer_tokens: int = 20000
+
+
+class AgentConfig(BaseModel):
+    max_turns: int = 500
     default_context_window: int = 200000
 
 
@@ -72,6 +76,7 @@ class ConfigSchema(BaseModel):
     llm: LLMConfig
     ui: UIConfig
     compaction: CompactionConfig
+    agent: AgentConfig
 
 
 class _BinariesConfig:
@@ -126,6 +131,10 @@ class Config:
     def compaction(self) -> CompactionConfig:
         return self._parsed.compaction
 
+    @property
+    def agent(self) -> AgentConfig:
+        return self._parsed.agent
+
     @property
     def binaries(self) -> _BinariesConfig:
         return _BinariesConfig(AVAILABLE_BINARIES)

{kon_coding_agent-0.2.0 → kon_coding_agent-0.2.2}/src/kon/defaults/config.toml

@@ -2,7 +2,7 @@
 default_provider = "openai-codex" # "zhipu", "github-copilot", "openai-codex"
 default_model = "gpt-5.3-codex"
 default_thinking_level = "high"
-tool_call_idle_timeout_seconds = 10
+tool_call_idle_timeout_seconds = 120
 system_prompt = """You are an expert coding assistant called `Kon`.
 You help users by reading, searching, executing commands, editing code, and writing new files.
 
@@ -21,6 +21,9 @@ You help users by reading, searching, executing commands, editing code, and writ
 [compaction]
 on_overflow = "continue" # "continue" or "pause"
 buffer_tokens = 20000
+
+[agent]
+max_turns = 500
 default_context_window = 200000
 
 [ui.colors]

{kon_coding_agent-0.2.0 → kon_coding_agent-0.2.2}/src/kon/events.py

@@ -100,6 +100,14 @@ class ToolArgsDeltaEvent:
     delta: str = ""
 
 
+@dataclass
+class ToolArgsTokenUpdateEvent:
+    type: Literal["tool_args_token_update"] = "tool_args_token_update"
+    tool_call_id: str = ""
+    tool_name: str = ""
+    token_count: int = 0
+
+
 @dataclass
 class ToolEndEvent:
     type: Literal["tool_end"] = "tool_end"
@@ -180,6 +188,7 @@ StreamEvent = (
     | TextEndEvent
     | ToolStartEvent
     | ToolArgsDeltaEvent
+    | ToolArgsTokenUpdateEvent
     | ToolEndEvent
     | ToolResultEvent
     | RetryEvent

{kon_coding_agent-0.2.0 → kon_coding_agent-0.2.2}/src/kon/llm/providers/mock.py

@@ -15,6 +15,7 @@ Scenarios (set via scenario parameter):
 - "unknown_tool": call unknown tool
 - "long_text": multiple text chunks
 - "tool_hang": emits a tool call and then never sends StreamDone
+- "tool_with_many_chunks": tool call with many argument chunks for token counting tests
 """
 
 import asyncio
@@ -138,6 +139,45 @@ class MockProvider(BaseProvider):
 
                 return tool_hang_iter()
 
+            case "tool_with_many_chunks":
+
+                async def tool_with_many_chunks_iter():
+                    # Tool call with many chunks to test token counting
+                    # 24 chunks of 8 chars each = 192 chars = 48 tokens
+                    # Should trigger token update events at chunks 12, 16, 20, 24
+                    yield ToolCallStart(id="call-1", name="bash", index=0)
+                    chunks = [
+                        "aaaaaaa",
+                        "bbbbbbb",
+                        "ccccccc",
+                        "ddddddd",
+                        "eeeeeee",
+                        "fffffff",
+                        "ggggggg",
+                        "hhhhhhh",
+                        "iiiiiii",
+                        "jjjjjjj",
+                        "kkkkkkk",
+                        "lllllll",
+                        "mmmmmmm",
+                        "nnnnnnn",
+                        "ooooooo",
+                        "ppppppp",
+                        "qqqqqqq",
+                        "rrrrrrr",
+                        "sssssss",
+                        "ttttttt",
+                        "uuuuuuu",
+                        "vvvvvvv",
+                        "wwwwwww",
+                        "xxxxxxxx",
+                    ]
+                    for chunk in chunks:
+                        yield ToolCallDelta(index=0, arguments_delta=chunk)
+                    yield StreamDone(stop_reason=StopReason.TOOL_USE)
+
+                return tool_with_many_chunks_iter()
+
             case _:
                 # Fallback to default
                 async def default_iter():

{kon_coding_agent-0.2.0 → kon_coding_agent-0.2.2}/src/kon/loop.py

@@ -76,7 +76,7 @@ def build_system_prompt(cwd: str, context: Context | None = None) -> str:
 
 @dataclass
 class AgentConfig:
-    max_turns: int = 200
+    max_turns: int | None = None
     system_prompt: str | None = None
     cwd: str | None = None
     context: Context | None = None
@@ -137,7 +137,12 @@ class Agent:
         system_prompt = self.config.system_prompt or build_system_prompt(cwd, self.config.context)
 
         try:
-            while turn < self.config.max_turns:
+            max_turns = (
+                self.config.max_turns
+                if self.config.max_turns is not None
+                else kon_config.agent.max_turns
+            )
+            while turn < max_turns:
                 if cancel_event and cancel_event.is_set():
                     was_interrupted = True
                     stop_reason = StopReason.INTERRUPTED
@@ -194,7 +199,7 @@ class Agent:
                 if stop_reason != StopReason.TOOL_USE:
                     break
 
-            if turn >= self.config.max_turns and not was_interrupted:
+            if turn >= max_turns and not was_interrupted:
                 stop_reason = StopReason.LENGTH
 
         except Exception as e:  # intentionally broad — top-level boundary; crash = broken TUI
@@ -219,7 +224,7 @@ class Agent:
         if last_usage is None:
             return
 
-        context_window = self.config.context_window or kon_config.compaction.default_context_window
+        context_window = self.config.context_window or kon_config.agent.default_context_window
         max_output = self.config.max_output_tokens or self.provider.config.max_tokens
         buffer_tokens = kon_config.compaction.buffer_tokens
 

{kon_coding_agent-0.2.0 → kon_coding_agent-0.2.2}/src/kon/session.py

@@ -58,6 +58,7 @@ class ModelChangeEntry(EntryBase):
     type: Literal["model_change"] = "model_change"
     provider: str
     model_id: str
+    base_url: str | None = None
 
 
 class CompactionEntry(EntryBase):
@@ -241,13 +242,16 @@ class Session:
         self._append_entry(entry)
         return entry.id
 
-    def append_model_change(self, provider: str, model_id: str) -> str:
+    def append_model_change(
+        self, provider: str, model_id: str, base_url: str | None = None
+    ) -> str:
         entry = ModelChangeEntry(
             id=self._generate_entry_id(),
             parent_id=self._leaf_id,
             timestamp=_now_iso(),
             provider=provider,
             model_id=model_id,
+            base_url=base_url,
         )
         self._append_entry(entry)
         return entry.id
@@ -372,20 +376,25 @@ class Session:
         return self._initial_thinking_level
 
     @property
-    def model(self) -> tuple[str, str] | None:
+    def model(self) -> tuple[str, str, str | None] | None:
         for entry in reversed(self._entries):
             if isinstance(entry, ModelChangeEntry):
-                return (entry.provider, entry.model_id)
+                return (entry.provider, entry.model_id, entry.base_url)
 
         if self._initial_provider and self._initial_model_id:
-            return (self._initial_provider, self._initial_model_id)
+            return (self._initial_provider, self._initial_model_id, None)
         return None
 
-    def set_model(self, provider: str, model_id: str) -> None:
+    def set_model(self, provider: str, model_id: str, base_url: str | None = None) -> None:
         current = self.model
-        if current and current[0] == provider and current[1] == model_id:
+        if (
+            current
+            and current[0] == provider
+            and current[1] == model_id
+            and current[2] == base_url
+        ):
             return
-        self.append_model_change(provider, model_id)
+        self.append_model_change(provider, model_id, base_url)
 
     def set_thinking_level(self, thinking_level: str) -> None:
         if self.thinking_level == thinking_level:

{kon_coding_agent-0.2.0 → kon_coding_agent-0.2.2}/src/kon/turn.py

@@ -59,6 +59,7 @@ from .events import (
     ThinkingEndEvent,
     ThinkingStartEvent,
     ToolArgsDeltaEvent,
+    ToolArgsTokenUpdateEvent,
     ToolEndEvent,
     ToolResultEvent,
     ToolStartEvent,
@@ -70,7 +71,14 @@ from .llm.base import LLMStream
 from .tools import BaseTool, get_tool, get_tool_definitions
 
 _STREAM_EXHAUSTED = object()
-_DEFAULT_TOOL_CALL_IDLE_TIMEOUT_SECONDS = 10.0
+_DEFAULT_TOOL_CALL_IDLE_TIMEOUT_SECONDS = 60.0
+_TOOL_ARGS_TOKEN_DISPLAY_THRESHOLD = 20
+_TOOL_ARGS_TOKEN_CHUNK_UPDATE_INTERVAL = 4
+
+
+def _count_tokens(text: str) -> int:
+    """Estimate token count from text (approx 4 chars per token)."""
+    return len(text) // 4
 
 
 class StreamState(StrEnum):
@@ -228,6 +236,10 @@ async def run_single_turn(
     pending_tool_calls: list[dict] = []
     current_tool_call: dict | None = None
 
+    # Token counting for tool argument streaming
+    _tool_arg_chunk_counter = 0
+    _tool_arg_token_count = 0
+
     current_state: StreamState | None = None
     stop_reason: StopReason = StopReason.STOP
     interrupted = False
@@ -381,6 +393,10 @@ async def run_single_turn(
                     pending_tool_calls.append(current_tool_call)
                     current_tool_call = None
 
+                # Reset token counters when starting a new tool call
+                _tool_arg_chunk_counter = 0
+                _tool_arg_token_count = 0
+
                 current_state = StreamState.TOOL_CALL
                 current_tool_call = {"id": id, "name": name, "arguments": ""}
 
@@ -391,6 +407,20 @@ async def run_single_turn(
                     current_tool_call["arguments"] += delta
                     yield ToolArgsDeltaEvent(tool_call_id=current_tool_call["id"], delta=delta)
 
+                    # Count tokens and fire update event every Nth chunk after threshold tokens
+                    _tool_arg_chunk_counter += 1
+                    _tool_arg_token_count += _count_tokens(delta)
+
+                    if (
+                        _tool_arg_token_count > _TOOL_ARGS_TOKEN_DISPLAY_THRESHOLD
+                        and _tool_arg_chunk_counter % _TOOL_ARGS_TOKEN_CHUNK_UPDATE_INTERVAL == 0
+                    ):
+                        yield ToolArgsTokenUpdateEvent(
+                            tool_call_id=current_tool_call["id"],
+                            tool_name=current_tool_call["name"],
+                            token_count=_tool_arg_token_count,
+                        )
+
             case StreamDone(stop_reason=reason):
                 stop_reason = reason
 

{kon_coding_agent-0.2.0 → kon_coding_agent-0.2.2}/src/kon/ui/app.py

@@ -4,8 +4,10 @@ import os
 import shutil
 import sys
 import time
+import tomllib
 from collections import deque
 from importlib.metadata import PackageNotFoundError, version
+from pathlib import Path
 from typing import ClassVar
 
 from rich.console import Console
@@ -32,6 +34,7 @@ from ..events import (
     ThinkingDeltaEvent,
     ThinkingEndEvent,
     ThinkingStartEvent,
+    ToolArgsTokenUpdateEvent,
     ToolEndEvent,
     ToolResultEvent,
     ToolStartEvent,
@@ -65,12 +68,24 @@ from .session_ui import SessionUIMixin
 from .styles import STYLES
 from .widgets import InfoBar, QueueDisplay, StatusLine, format_path
 
-_PYPI_PACKAGE_NAME = "kon-coding-agent"
+
+def _get_package_name() -> str:
+    pyproject_path = Path(__file__).parent.parent.parent.parent / "pyproject.toml"
+    if pyproject_path.exists():
+        try:
+            data = tomllib.loads(pyproject_path.read_text())
+            return data["project"]["name"]
+        except Exception:
+            pass
+    return "kon-coding-agent"
+
+
+_PYPI_PACKAGE_NAME = _get_package_name()
 
 try:
     VERSION = version(_PYPI_PACKAGE_NAME)
 except PackageNotFoundError:
-    VERSION = "0.2.0"
+    VERSION = "0.2.2"
 
 _COPILOT_API_TYPES: frozenset[ApiType] = frozenset(
     {ApiType.GITHUB_COPILOT, ApiType.GITHUB_COPILOT_RESPONSES, ApiType.ANTHROPIC_COPILOT}
@@ -202,16 +217,20 @@ class Kon(CommandsMixin, SessionUIMixin, App[None]):
             if self._session.entries:
                 model_info = self._session.model
                 if model_info:
-                    provider, self._model = model_info
+                    provider, self._model, session_base_url = model_info
                     self._model_provider = provider
+                    if self._base_url is None and session_base_url:
+                        self._base_url = session_base_url
                 self._thinking_level = self._session.thinking_level
         elif self._continue_recent:
             self._session = Session.continue_recent(self._cwd)
             if self._session.entries:
                 model_info = self._session.model
                 if model_info:
-                    provider, self._model = model_info
+                    provider, self._model, session_base_url = model_info
                     self._model_provider = provider
+                    if self._base_url is None and session_base_url:
+                        self._base_url = session_base_url
                 self._thinking_level = self._session.thinking_level
 
         model_info = get_model(self._model, self._model_provider)
@@ -257,7 +276,7 @@ class Kon(CommandsMixin, SessionUIMixin, App[None]):
                 model_id=self._model,
                 thinking_level=self._thinking_level,
             )
-            self._session.append_model_change(model_provider, self._model)
+            self._session.append_model_change(model_provider, self._model, base_url)
 
         self._project_context = Context.load(self._cwd)
         # TODO: Surface self._project_context.skill_warnings in UI (e.g. chat info/error messages)
@@ -567,7 +586,6 @@ class Kon(CommandsMixin, SessionUIMixin, App[None]):
             tools = get_tools(DEFAULT_TOOLS)
             model_info = get_model(self._model, self._model_provider)
             agent_config = AgentConfig(
-                max_turns=50,
                 system_prompt=self._get_system_prompt(),
                 context_window=model_info.context_window if model_info else None,
                 max_output_tokens=model_info.max_tokens if model_info else None,
@@ -621,6 +639,10 @@ class Kon(CommandsMixin, SessionUIMixin, App[None]):
                             chat.start_tool(name, id, "")
                             self._current_block_type = "tool_call"
                             status.increment_tool_calls()
+                            status.set_streaming_tokens(0)  # Reset token count for new tool
+
+                        case ToolArgsTokenUpdateEvent(token_count=tc):
+                            status.set_streaming_tokens(tc)
 
                         case ToolEndEvent(tool_call_id=id, display=display):
                             chat.update_tool_call_msg(id, display)
@@ -743,6 +765,7 @@ def main():
         dest="resume_session",
         help="Resume a specific session by ID (full or unique prefix)",
     )
+    parser.add_argument("--version", action="version", version=f"kon {VERSION}")
     args = parser.parse_args()
 
     app = Kon(