mycode-cli 0.4.2__tar.gz → 0.5.0__tar.gz

{mycode_cli-0.4.2 → mycode_cli-0.5.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: mycode-cli
-Version: 0.4.2
+Version: 0.5.0
 Summary: Minimal coding agent CLI and web UI built on mycode-sdk.
 Project-URL: Homepage, https://github.com/legibet/mycode
 Project-URL: Repository, https://github.com/legibet/mycode
@@ -22,7 +22,7 @@ Classifier: Programming Language :: Python :: 3.13
 Classifier: Topic :: Software Development
 Requires-Python: >=3.12
 Requires-Dist: fastapi>=0.115.0
-Requires-Dist: mycode-sdk==0.4.2
+Requires-Dist: mycode-sdk==0.5.0
 Requires-Dist: prompt-toolkit>=3.0.52
 Requires-Dist: pyyaml>=6.0.0
 Requires-Dist: rich>=14.2.0
@@ -36,13 +36,11 @@ Description-Content-Type: text/markdown
 
 A minimal coding agent. Inspired by [pi](https://github.com/badlogic/pi-mono).
 
-- Minimal core (under 5k lines of code).
-- Unified message format and robust cross-provider replay.
-- 4 built-in tools: `read`, `write`, `edit`, `bash`, expanded via skills.
-- Inspectable runtime, append-only JSONL sessions.
-- Native image and pdf input support.
+- Minimal agent core (< 5k lines) with multiple provider support and robust message replay.
+- 4 built-in tools only (`read`, `write`, `edit`, `bash`), expanded via skills.
 - Mobile-friendly web UI.
-- Lightweight `mycode-sdk` for embedding the agent in other Python apps.
+- Native image and pdf input support.
+- Lightweight Python SDK for building custom agents.
 
 ## Quick Start
 
@@ -52,7 +50,7 @@ Requires Python 3.12+. Install via [uv](https://docs.astral.sh/uv/):
 uv tool install mycode-cli
 ```
 
-Interactive terminal session:
+Interactive terminal UI:
 
 ```bash
 mycode
@@ -61,7 +59,7 @@ mycode
 Web UI (default at `http://localhost:8000`):
 
 ```bash
-mycode web (--port <port> --hostname <hostname>)
+mycode web [--port <port>] [--hostname <hostname>]
 ```
 
 Single message, non-interactive:
@@ -70,32 +68,33 @@ Single message, non-interactive:
 mycode run "explain how the session store works"
 ```
 
-API keys are discovered automatically from environment variables (see Providers & Models).
+API keys are discovered automatically from environment variables (see Providers).
+
+## Providers
 
-## Providers & Models
+| Provider          | id            | Env var              |
+| ----------------- | ------------- | -------------------- |
+| Anthropic         | `anthropic`   | `ANTHROPIC_API_KEY`  |
+| OpenAI            | `openai`      | `OPENAI_API_KEY`     |
+| Google Gemini     | `google`      | `GEMINI_API_KEY`     |
+| Moonshot          | `moonshotai`  | `MOONSHOT_API_KEY`   |
+| MiniMax           | `minimax`     | `MINIMAX_API_KEY`    |
+| DeepSeek          | `deepseek`    | `DEEPSEEK_API_KEY`   |
+| Z.AI              | `zai`         | `ZAI_API_KEY`        |
+| OpenRouter        | `openrouter`  | `OPENROUTER_API_KEY` |
+| OpenAI-compatible | `openai_chat` | —                    |
 
-| Provider          | id            | Env var              | Default models                                     |
-| ----------------- | ------------- | -------------------- | -------------------------------------------------- |
-| Anthropic         | `anthropic`   | `ANTHROPIC_API_KEY`  | `claude-sonnet-4-6`, `claude-opus-4-7`             |
-| OpenAI            | `openai`      | `OPENAI_API_KEY`     | `gpt-5.4`, `gpt-5.4-mini`                          |
-| Google Gemini     | `google`      | `GEMINI_API_KEY`     | `gemini-3.1-pro-preview`, `gemini-3-flash-preview` |
-| Moonshot          | `moonshotai`  | `MOONSHOT_API_KEY`   | `kimi-k2.5`                                        |
-| MiniMax           | `minimax`     | `MINIMAX_API_KEY`    | `MiniMax-M2.7`, `MiniMax-M2.7-highspeed`           |
-| DeepSeek          | `deepseek`    | `DEEPSEEK_API_KEY`   | `deepseek-chat`, `deepseek-reasoner`               |
-| Z.AI              | `zai`         | `ZAI_API_KEY`        | `glm-5.1`, `glm-5-turbo`                           |
-| OpenRouter        | `openrouter`  | `OPENROUTER_API_KEY` | `openrouter/auto`                                  |
-| OpenAI-compatible | `openai_chat` | —                    | (configured per provider)                          |
+Run `/model` in tui to see the available models.
 
 ## Configuration
 
-No config file is required. It is only used for:
+A config file is optional — API keys from the environment are usually sufficient.
 
-1. Setting default provider, model, and other options
-2. Overriding built-in provider settings (e.g. changing the available model list)
-3. Adding custom providers with any built-in provider type.
-4. Customize model metadata for built-in and custom models.
+Create `~/.mycode/config.json` (global) or `<workspace>/.mycode/config.json` (project, takes precedence) to:
 
-Config is loaded from `~/.mycode/config.json` (global) and `<workspace>/.mycode/config.json` (project-specific, takes precedence).
+- set a default provider, model, and reasoning effort
+- expose additional models on an existing provider (e.g. OpenRouter's catalog)
+- register a custom endpoint, such as a private or regional deployment
 
 ```json
 {
@@ -133,12 +132,12 @@ Config is loaded from `~/.mycode/config.json` (global) and `<workspace>/.mycode/
 }
 ```
 
-- Built-in provider ids can be overridden by key without specifying `type`. Custom providers must set `type`.
+- To override a built-in provider, reuse its id as the key — no `type` needed. Custom providers must declare a `type` — one of the built-in protocols.
 - `reasoning_effort` controls extended thinking for supported models: `auto` (default) · `none` · `low` · `medium` · `high` · `xhigh`.
 - API keys in config accept `${ENV_VAR}` references.
-- Model metadata is sourced from models.dev and bundled — no manual config needed for built-in models.
+- Model metadata is bundled from [models.dev](https://models.dev) — `{}` is enough for most models. Provide explicit fields only for models not listed there.
 
-> Built-in Moonshot, MiniMax, and Z.AI defaults use international endpoints. Override `base_url` in config for China endpoints.
+> Built-in Moonshot, MiniMax, and Z.AI providers default to international endpoints. Override `base_url` for China endpoints.
 
 ## CLI Reference
 
@@ -152,12 +151,12 @@ mycode web --dev                  API only, no static files
 mycode session list               list saved sessions
 ```
 
-Interactive slash commands: `/new` `/resume` `/provider` `/model` `/effort` `/clear` `/q`
+Interactive slash commands: `/new` `/resume` `/rewind` `/provider` `/model` `/effort` `/clear` `/q`
 
 ## Development
 
 ```bash
-git clone <repo> && cd mycode
+git clone https://github.com/legibet/mycode.git && cd mycode
 uv sync --dev
 uv run mycode
 ```
@@ -169,6 +168,14 @@ uv run mycode web --dev
 pnpm --dir web install && pnpm --dir web dev
 ```
 
+Or start both together:
+
+```bash
+just dev
+```
+
+Other useful shortcuts: `just check` · `just test` · `just fmt`
+
 Build distributable artifacts:
 
 ```bash
@@ -178,6 +185,8 @@ uv build --package mycode-cli
 
 ## mycode-sdk
 
+Agent core as a lightweight Python SDK for building agents. Install via: `uv add mycode-sdk`
+
 ```python
 import asyncio
 from mycode import Agent, bash_tool, read_tool
@@ -197,7 +206,7 @@ async def main() -> None:
 asyncio.run(main())
 ```
 
-See `mycode/README.md` for details.
+See [mycode/README.md](mycode/README.md) for details.
 
 ## License
 

{mycode_cli-0.4.2 → mycode_cli-0.5.0}/README.md

@@ -4,13 +4,11 @@
 
 A minimal coding agent. Inspired by [pi](https://github.com/badlogic/pi-mono).
 
-- Minimal core (under 5k lines of code).
-- Unified message format and robust cross-provider replay.
-- 4 built-in tools: `read`, `write`, `edit`, `bash`, expanded via skills.
-- Inspectable runtime, append-only JSONL sessions.
-- Native image and pdf input support.
+- Minimal agent core (< 5k lines) with multiple provider support and robust message replay.
+- 4 built-in tools only (`read`, `write`, `edit`, `bash`), expanded via skills.
 - Mobile-friendly web UI.
-- Lightweight `mycode-sdk` for embedding the agent in other Python apps.
+- Native image and pdf input support.
+- Lightweight Python SDK for building custom agents.
 
 ## Quick Start
 
@@ -20,7 +18,7 @@ Requires Python 3.12+. Install via [uv](https://docs.astral.sh/uv/):
 uv tool install mycode-cli
 ```
 
-Interactive terminal session:
+Interactive terminal UI:
 
 ```bash
 mycode
@@ -29,7 +27,7 @@ mycode
 Web UI (default at `http://localhost:8000`):
 
 ```bash
-mycode web (--port <port> --hostname <hostname>)
+mycode web [--port <port>] [--hostname <hostname>]
 ```
 
 Single message, non-interactive:
@@ -38,32 +36,33 @@ Single message, non-interactive:
 mycode run "explain how the session store works"
 ```
 
-API keys are discovered automatically from environment variables (see Providers & Models).
+API keys are discovered automatically from environment variables (see Providers).
+
+## Providers
 
-## Providers & Models
+| Provider          | id            | Env var              |
+| ----------------- | ------------- | -------------------- |
+| Anthropic         | `anthropic`   | `ANTHROPIC_API_KEY`  |
+| OpenAI            | `openai`      | `OPENAI_API_KEY`     |
+| Google Gemini     | `google`      | `GEMINI_API_KEY`     |
+| Moonshot          | `moonshotai`  | `MOONSHOT_API_KEY`   |
+| MiniMax           | `minimax`     | `MINIMAX_API_KEY`    |
+| DeepSeek          | `deepseek`    | `DEEPSEEK_API_KEY`   |
+| Z.AI              | `zai`         | `ZAI_API_KEY`        |
+| OpenRouter        | `openrouter`  | `OPENROUTER_API_KEY` |
+| OpenAI-compatible | `openai_chat` | —                    |
 
-| Provider          | id            | Env var              | Default models                                     |
-| ----------------- | ------------- | -------------------- | -------------------------------------------------- |
-| Anthropic         | `anthropic`   | `ANTHROPIC_API_KEY`  | `claude-sonnet-4-6`, `claude-opus-4-7`             |
-| OpenAI            | `openai`      | `OPENAI_API_KEY`     | `gpt-5.4`, `gpt-5.4-mini`                          |
-| Google Gemini     | `google`      | `GEMINI_API_KEY`     | `gemini-3.1-pro-preview`, `gemini-3-flash-preview` |
-| Moonshot          | `moonshotai`  | `MOONSHOT_API_KEY`   | `kimi-k2.5`                                        |
-| MiniMax           | `minimax`     | `MINIMAX_API_KEY`    | `MiniMax-M2.7`, `MiniMax-M2.7-highspeed`           |
-| DeepSeek          | `deepseek`    | `DEEPSEEK_API_KEY`   | `deepseek-chat`, `deepseek-reasoner`               |
-| Z.AI              | `zai`         | `ZAI_API_KEY`        | `glm-5.1`, `glm-5-turbo`                           |
-| OpenRouter        | `openrouter`  | `OPENROUTER_API_KEY` | `openrouter/auto`                                  |
-| OpenAI-compatible | `openai_chat` | —                    | (configured per provider)                          |
+Run `/model` in tui to see the available models.
 
 ## Configuration
 
-No config file is required. It is only used for:
+A config file is optional — API keys from the environment are usually sufficient.
 
-1. Setting default provider, model, and other options
-2. Overriding built-in provider settings (e.g. changing the available model list)
-3. Adding custom providers with any built-in provider type.
-4. Customize model metadata for built-in and custom models.
+Create `~/.mycode/config.json` (global) or `<workspace>/.mycode/config.json` (project, takes precedence) to:
 
-Config is loaded from `~/.mycode/config.json` (global) and `<workspace>/.mycode/config.json` (project-specific, takes precedence).
+- set a default provider, model, and reasoning effort
+- expose additional models on an existing provider (e.g. OpenRouter's catalog)
+- register a custom endpoint, such as a private or regional deployment
 
 ```json
 {
@@ -101,12 +100,12 @@ Config is loaded from `~/.mycode/config.json` (global) and `<workspace>/.mycode/
 }
 ```
 
-- Built-in provider ids can be overridden by key without specifying `type`. Custom providers must set `type`.
+- To override a built-in provider, reuse its id as the key — no `type` needed. Custom providers must declare a `type` — one of the built-in protocols.
 - `reasoning_effort` controls extended thinking for supported models: `auto` (default) · `none` · `low` · `medium` · `high` · `xhigh`.
 - API keys in config accept `${ENV_VAR}` references.
-- Model metadata is sourced from models.dev and bundled — no manual config needed for built-in models.
+- Model metadata is bundled from [models.dev](https://models.dev) — `{}` is enough for most models. Provide explicit fields only for models not listed there.
 
-> Built-in Moonshot, MiniMax, and Z.AI defaults use international endpoints. Override `base_url` in config for China endpoints.
+> Built-in Moonshot, MiniMax, and Z.AI providers default to international endpoints. Override `base_url` for China endpoints.
 
 ## CLI Reference
 
@@ -120,12 +119,12 @@ mycode web --dev                  API only, no static files
 mycode session list               list saved sessions
 ```
 
-Interactive slash commands: `/new` `/resume` `/provider` `/model` `/effort` `/clear` `/q`
+Interactive slash commands: `/new` `/resume` `/rewind` `/provider` `/model` `/effort` `/clear` `/q`
 
 ## Development
 
 ```bash
-git clone <repo> && cd mycode
+git clone https://github.com/legibet/mycode.git && cd mycode
 uv sync --dev
 uv run mycode
 ```
@@ -137,6 +136,14 @@ uv run mycode web --dev
 pnpm --dir web install && pnpm --dir web dev
 ```
 
+Or start both together:
+
+```bash
+just dev
+```
+
+Other useful shortcuts: `just check` · `just test` · `just fmt`
+
 Build distributable artifacts:
 
 ```bash
@@ -146,6 +153,8 @@ uv build --package mycode-cli
 
 ## mycode-sdk
 
+Agent core as a lightweight Python SDK for building agents. Install via: `uv add mycode-sdk`
+
 ```python
 import asyncio
 from mycode import Agent, bash_tool, read_tool
@@ -165,7 +174,7 @@ async def main() -> None:
 asyncio.run(main())
 ```
 
-See `mycode/README.md` for details.
+See [mycode/README.md](mycode/README.md) for details.
 
 ## License
 

{mycode_cli-0.4.2 → mycode_cli-0.5.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "mycode-cli"
-version = "0.4.2"
+version = "0.5.0"
 description = "Minimal coding agent CLI and web UI built on mycode-sdk."
 requires-python = ">=3.12"
 license = "MIT"
@@ -26,7 +26,7 @@ classifiers = [
 ]
 keywords = ["agent", "cli", "llm", "coding-agent", "fastapi"]
 dependencies = [
-    "mycode-sdk==0.4.2",
+    "mycode-sdk==0.5.0",
     "fastapi>=0.115.0",
     "uvicorn[standard]>=0.32.0",
     "rich>=14.2.0",

{mycode_cli-0.4.2 → mycode_cli-0.5.0}/src/mycode_cli/config.py

@@ -1,11 +1,4 @@
-"""Application configuration and provider resolution.
-
-This module keeps runtime config loading intentionally simple:
-
-- read the global config, then the workspace config
-- merge provider entries by name
-- resolve one runnable provider from explicit args, config defaults, or env
-"""
+"""Config loading, provider resolution, and CLI-owned filesystem paths."""
 
 from __future__ import annotations
 
@@ -27,9 +20,24 @@ from mycode.providers import (
     provider_default_models,
     provider_env_api_key_names,
 )
-from mycode.session import resolve_mycode_home
 from mycode.utils import as_bool, as_int
 
+_DEFAULT_MYCODE_HOME = "~/.mycode"
+
+
+def resolve_mycode_home() -> Path:
+    """Resolve the mycode home directory (``$MYCODE_HOME`` or ``~/.mycode``)."""
+
+    raw = os.environ.get("MYCODE_HOME", _DEFAULT_MYCODE_HOME)
+    return Path(raw).expanduser().resolve(strict=False)
+
+
+def resolve_sessions_dir() -> Path:
+    """Resolve the directory used for persisted sessions."""
+
+    return resolve_mycode_home() / "sessions"
+
+
 _API_KEY_ENV_REF_RE = re.compile(r"^\$\{([A-Za-z_][A-Za-z0-9_]*)\}$")
 # Full set of user-facing choices. "auto" means "do not send an explicit effort
 # to the provider" and is represented internally as `None` after normalization.

{mycode_cli-0.4.2 → mycode_cli-0.5.0}/src/mycode_cli/main.py

@@ -13,7 +13,7 @@ import typer
 from mycode.agent import Agent
 from mycode.messages import ConversationMessage
 from mycode.session import SessionStore
-from mycode_cli.config import ResolvedProvider, Settings, get_settings, resolve_provider
+from mycode_cli.config import ResolvedProvider, Settings, get_settings, resolve_provider, resolve_sessions_dir
 
 from .runtime import ResolvedSession, build_agent, resolve_session
 from .tui.chat import TerminalChat
@@ -82,7 +82,7 @@ def _bootstrap(
         raise typer.BadParameter("--session and --continue are mutually exclusive")
 
     cwd = os.path.abspath(os.getcwd())
-    store = SessionStore()
+    store = SessionStore(data_dir=resolve_sessions_dir())
     view = TerminalView()
     settings = get_settings(cwd)
 
@@ -91,10 +91,7 @@ def _bootstrap(
         resolved_session = asyncio.run(
             resolve_session(
                 store=store,
-                provider=resolved_provider.provider,
                 cwd=cwd,
-                model=resolved_provider.model,
-                api_base=resolved_provider.api_base,
                 requested_session_id=session,
                 continue_last=continue_last,
             )
@@ -109,7 +106,6 @@ def _bootstrap(
         settings=settings,
         resolved_provider=resolved_provider,
         session_id=resolved_session.session_id,
-        messages=resolved_session.messages,
         max_turns=max_turns,
     )
 
@@ -222,7 +218,7 @@ def session_list(
     """List saved sessions."""
 
     cwd = os.path.abspath(os.getcwd())
-    store = SessionStore()
+    store = SessionStore(data_dir=resolve_sessions_dir())
     view = TerminalView()
 
     sessions = asyncio.run(store.list_sessions(cwd=None if all_workspaces else cwd))

{mycode_cli-0.4.2 → mycode_cli-0.5.0}/src/mycode_cli/runtime.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 
 from dataclasses import dataclass
 from typing import Any
+from uuid import uuid4
 
 from mycode.agent import Agent
 from mycode.providers import (
@@ -48,22 +49,24 @@ def build_agent(
     settings: Settings,
     resolved_provider: ResolvedProvider,
     session_id: str,
-    messages: list[dict[str, Any]] | None = None,
     max_turns: int | None = None,
     reasoning_effort: str | None = None,
 ) -> Agent:
-    """Build an agent from the resolved provider, honoring model config overrides."""
+    """Build an agent from the resolved provider, honoring model config overrides.
 
-    model_config = _model_config_for(settings, resolved_provider)
+    History is auto-loaded from disk when ``session_id`` already exists under
+    the store; callers never pass messages explicitly.
+    """
+
+    model_config = model_config_for(settings, resolved_provider)
     return Agent(
         model=resolved_provider.model,
         provider=resolved_provider.provider,
         cwd=cwd,
-        session_dir=store.session_dir(session_id),
+        session_dir=store.data_dir,
         session_id=session_id,
         api_key=resolved_provider.api_key,
         api_base=resolved_provider.api_base,
-        messages=messages,
         reasoning_effort=reasoning_effort if reasoning_effort is not None else resolved_provider.reasoning_effort,
         max_tokens=model_config.max_output_tokens if model_config else None,
         context_window=model_config.context_window if model_config else None,
@@ -77,7 +80,7 @@ def build_agent(
     )
 
 
-def _model_config_for(settings: Settings, resolved: ResolvedProvider) -> ModelConfig | None:
+def model_config_for(settings: Settings, resolved: ResolvedProvider) -> ModelConfig | None:
     """Return the user-configured overrides for the resolved provider+model, if any."""
 
     provider_config = settings.providers.get(resolved.provider_name or "")
@@ -86,18 +89,20 @@ def _model_config_for(settings: Settings, resolved: ResolvedProvider) -> ModelCo
     return provider_config.models.get(resolved.model)
 
 
-def clone_agent(agent: Agent, *, store: SessionStore, session_id: str, messages: list[dict[str, Any]]) -> Agent:
-    """Keep the current runtime config while swapping session state."""
+def clone_agent(agent: Agent, *, store: SessionStore, session_id: str) -> Agent:
+    """Keep the current runtime config while swapping session state.
+
+    History auto-loads from disk when ``session_id`` exists under the store.
+    """
 
     return Agent(
         model=agent.model,
         provider=agent.provider,
         cwd=agent.cwd,
-        session_dir=store.session_dir(session_id),
+        session_dir=store.data_dir,
         session_id=session_id,
         api_key=agent.api_key,
         api_base=agent.api_base,
-        messages=messages,
         max_turns=agent.max_turns,
         max_tokens=agent.max_tokens,
         context_window=agent.context_window,
@@ -171,10 +176,7 @@ def supports_reasoning_effort(agent: Agent) -> bool:
 async def resolve_session(
     *,
     store: SessionStore,
-    provider: str,
     cwd: str,
-    model: str,
-    api_base: str | None,
     requested_session_id: str | None,
     continue_last: bool,
 ) -> ResolvedSession:
@@ -205,9 +207,9 @@ async def resolve_session(
                 "resumed",
             )
 
-    data = store.draft_session(None, provider=provider, model=model, cwd=cwd, api_base=api_base)
-    session = data.get("session") or {}
-    return ResolvedSession(str(session.get("id") or ""), session, [], "new")
+    # New sessions: the id is allocated here; the on-disk session is created
+    # lazily by Agent.achat on the first persist.
+    return ResolvedSession(uuid4().hex, {}, [], "new")
 
 
 def apply_resolved_provider(agent: Agent, resolved: ResolvedProvider, settings: Settings) -> bool:
@@ -233,7 +235,7 @@ def apply_resolved_provider(agent: Agent, resolved: ResolvedProvider, settings:
     agent.reasoning_effort = resolved.reasoning_effort
 
     if runtime_changed:
-        model_config = _model_config_for(settings, resolved)
+        model_config = model_config_for(settings, resolved)
         agent.refresh_capabilities(
             max_tokens=model_config.max_output_tokens if model_config else None,
             context_window=model_config.context_window if model_config else None,

{mycode_cli-0.4.2 → mycode_cli-0.5.0}/src/mycode_cli/server/deps.py

@@ -8,6 +8,7 @@ from typing import Annotated
 from fastapi import Depends
 
 from mycode.session import SessionStore
+from mycode_cli.config import resolve_sessions_dir
 from mycode_cli.server.run_manager import RunManager
 
 
@@ -15,7 +16,7 @@ from mycode_cli.server.run_manager import RunManager
 def get_store() -> SessionStore:
     """Return the shared session store for server requests."""
 
-    return SessionStore()
+    return SessionStore(data_dir=resolve_sessions_dir())
 
 
 @lru_cache

{mycode_cli-0.4.2 → mycode_cli-0.5.0}/src/mycode_cli/server/routers/chat.py

@@ -17,7 +17,6 @@ from fastapi.responses import StreamingResponse
 from mycode.messages import (
     build_message,
     document_block,
-    flatten_message_text,
     image_block,
     text_block,
 )
@@ -31,7 +30,7 @@ from mycode_cli.config import (
     resolve_provider,
     resolve_provider_choices,
 )
-from mycode_cli.runtime import build_agent
+from mycode_cli.runtime import build_agent, model_config_for
 from mycode_cli.server.deps import RunManagerDep, StoreDep
 from mycode_cli.server.run_manager import ActiveRunError, RunState
 from mycode_cli.server.schemas import ChatRequest, StreamEvent
@@ -165,19 +164,19 @@ async def chat(chat: ChatRequest, store: StoreDep, runs: RunManagerDep):
 
     data = await store.load_session(session_id)
     session = (data or {}).get("session")
-    messages = (data or {}).get("messages") or []
+    existing_messages = (data or {}).get("messages") or []
 
     if not session and chat.rewind_to is not None:
         raise HTTPException(status_code=400, detail="rewind_to requires an existing session")
 
     if chat.rewind_to is not None:
-        if not (0 <= chat.rewind_to < len(messages)):
+        if not (0 <= chat.rewind_to < len(existing_messages)):
             raise HTTPException(
                 status_code=400,
-                detail=f"rewind_to must reference a visible message index between 0 and {len(messages) - 1}",
+                detail=f"rewind_to must reference a visible message index between 0 and {len(existing_messages) - 1}",
             )
 
-        target = messages[chat.rewind_to]
+        target = existing_messages[chat.rewind_to]
         raw_blocks = target.get("content")
         blocks = raw_blocks if isinstance(raw_blocks, list) else []
         has_user_content = any(
@@ -194,7 +193,29 @@ async def chat(chat: ChatRequest, store: StoreDep, runs: RunManagerDep):
                 detail="rewind_to must reference a real user message",
             )
 
-        messages = messages[: chat.rewind_to]
+    # Capability check before any disk mutation — a failed check must not
+    # leave an empty session on disk or land a premature rewind marker.
+    model_config = model_config_for(settings, resolved)
+    caps = resolve_model_metadata(
+        provider=resolved.provider,
+        model=resolved.model,
+        supports_image_input=model_config.supports_image_input if model_config else None,
+        supports_pdf_input=model_config.supports_pdf_input if model_config else None,
+    )
+    content_types = {b.get("type") for b in (user_message.get("content") or []) if isinstance(b, dict)}
+    if "image" in content_types and not caps.supports_image_input:
+        raise HTTPException(status_code=400, detail="current model does not support image input")
+    if "document" in content_types and not caps.supports_pdf_input:
+        raise HTTPException(status_code=400, detail="current model does not support PDF input")
+
+    # All validation passed. Land the rewind marker (if any) and create the
+    # session so the response payload has a meta dict and the agent's
+    # auto-resume sees the correct on-disk state.
+    if chat.rewind_to is not None:
+        await store.append_rewind(session_id, chat.rewind_to)
+    if not session:
+        created = await store.create_session(session_id, cwd=cwd)
+        session = created["session"]
 
     agent = build_agent(
         store=store,
@@ -202,37 +223,14 @@ async def chat(chat: ChatRequest, store: StoreDep, runs: RunManagerDep):
         settings=settings,
         resolved_provider=resolved,
         session_id=session_id,
-        messages=messages,
         reasoning_effort=reasoning_effort,
     )
 
-    content_types = {b.get("type") for b in (user_message.get("content") or []) if isinstance(b, dict)}
-    if "image" in content_types and not agent.supports_image_input:
-        raise HTTPException(status_code=400, detail="current model does not support image input")
-    if "document" in content_types and not agent.supports_pdf_input:
-        raise HTTPException(status_code=400, detail="current model does not support PDF input")
-
-    if not session:
-        title = flatten_message_text(user_message).replace("\n", " ").strip()[:48] or "New chat"
-        created = await store.create_session(
-            title,
-            session_id=session_id,
-            provider=resolved.provider,
-            model=resolved.model,
-            cwd=cwd,
-            api_base=resolved.api_base,
-        )
-        session = created["session"]
-
-    if chat.rewind_to is not None:
-        # Land the rewind marker before the agent appends the next user turn.
-        await store.append_rewind(session_id, chat.rewind_to)
-
     try:
         run = await runs.start_run(
             session_id=session_id,
             user_message=user_message,
-            base_messages=messages,
+            base_messages=agent.messages,
             agent=agent,
         )
     except ActiveRunError as exc: