core/agent+cli: defer DB write until turn success; auto/manual retry

Failed turns roll back memory and keep pending_retry_text; persist only
after success. CLI auto-retries 10s/20s/30s (cancellable) and /retry.

CLAUDE.md

@@ -55,7 +55,7 @@ Async SQLAlchemy + aiosqlite. `MemoryStore`: schema init, default local user, se
 - **`ChatClient`** Protocol for `chat_completions`.
 - **`@tool` / `ToolRegistry`**: decorator infers JSON Schema from type hints; execute tools by name.
 - **`run_chat_loop`**: multi-round tool loop, yields `AgentEvent` stream; optional `on_limit` to continue past max rounds.
-- **`ChatAgent`**: wrapper with optional `MemoryStore` bind (load/append messages).
+- **`ChatAgent`**: wrapper with optional `MemoryStore` bind (load/append messages on success only); `pending_retry_text` + `retry()` after failed turns.
 
 ### Tools (`tools/`)
 
@@ -73,7 +73,8 @@ Module-level `@tool` handlers. Call `set_workspace_root()` before use.
 Click app + readline REPL. Entry: `plyngent` / `python -m plyngent`.
 
 - **`plyngent chat`**: provider/model selection (flags or interactive), SQLite sessions via config `[database]` (file DB under user data if unset/`:memory:`), resumes latest session by default (`--new` / `--session`).
-- Slash: `/history`, `/sessions`, `/resume`, `/rounds`, …
+- Slash: `/history`, `/sessions`, `/resume`, `/rounds`, `/retry`, …
+- Failed turns: not written to DB; auto-retry 10s/20s/30s (Ctrl+C cancels wait); `/retry` manual.
 - **`plyngent providers`**: list config providers.
 - **`plyngent config path|edit`**: print or open config in `$EDITOR` (`shlex`-split, e.g. `codium --wait`).
 - If no providers and `$EDITOR` is set, chat/providers prompt to edit config then reload.

src/plyngent/agent/chat.py

@@ -31,6 +31,7 @@ class ChatAgent:
     temperature: float | None
     on_limit: LimitContinueHook | None
     messages: list[AnyChatMessage]
+    pending_retry_text: str | None
 
     def __init__(  # noqa: PLR0913
         self,
@@ -54,6 +55,7 @@ class ChatAgent:
         self.temperature = temperature
         self.on_limit = on_limit
         self.messages = list(messages) if messages is not None else []
+        self.pending_retry_text = None
 
     async def load_history(self) -> None:
         """Replace in-memory messages from the bound memory session."""
@@ -61,6 +63,7 @@ class ChatAgent:
             msg = "load_history requires memory and session_id"
             raise RuntimeError(msg)
         self.messages = await self.memory.list_messages(self.session_id)
+        self.pending_retry_text = None
 
     async def bind_session(self, session_id: int, *, load: bool = True) -> None:
         """Attach a memory session id; optionally load existing messages."""
@@ -75,24 +78,48 @@ class ChatAgent:
         if self.memory is not None and self.session_id is not None:
             _ = await self.memory.append_message(self.session_id, message)
 
+    async def _run_from_user_message(self, user_msg: UserChatMessage) -> AsyncIterator[AgentEvent]:
+        """Run the tool loop for an already-appended user message; persist only on success."""
+        pre_len = len(self.messages) - 1
+        if pre_len < 0 or self.messages[pre_len] is not user_msg:
+            pre_len = len(self.messages)
+            self.messages.append(user_msg)
+
+        try:
+            async for event in run_chat_loop(
+                self.client,
+                self.messages,
+                model=self.model,
+                tools=self.tools,
+                max_rounds=self.max_rounds,
+                temperature=self.temperature,
+                on_limit=self.on_limit,
+            ):
+                yield event
+        except Exception:
+            # Roll back the whole turn so a failed user message is not left half-applied.
+            del self.messages[pre_len:]
+            self.pending_retry_text = user_msg.content
+            raise
+
+        for message in self.messages[pre_len:]:
+            await self._persist(message)
+        self.pending_retry_text = None
+
     async def run(self, user_text: str) -> AsyncIterator[AgentEvent]:
-        """Append a user message, run the tool loop, yield events, persist new messages."""
+        """Append a user message, run the tool loop, yield events, persist only on success."""
         user_msg = UserChatMessage(content=user_text)
         self.messages.append(user_msg)
-        await self._persist(user_msg)
-
-        start_len = len(self.messages)
-        async for event in run_chat_loop(
-            self.client,
-            self.messages,
-            model=self.model,
-            tools=self.tools,
-            max_rounds=self.max_rounds,
-            temperature=self.temperature,
-            on_limit=self.on_limit,
-        ):
+        async for event in self._run_from_user_message(user_msg):
             yield event
 
-        # Persist messages appended by the loop after the user message.
-        for message in self.messages[start_len:]:
-            await self._persist(message)
+    async def retry(self) -> AsyncIterator[AgentEvent]:
+        """Re-run the last failed user turn (no duplicate user message in history/DB)."""
+        text = self.pending_retry_text
+        if text is None:
+            msg = "nothing to retry"
+            raise RuntimeError(msg)
+        user_msg = UserChatMessage(content=text)
+        self.messages.append(user_msg)
+        async for event in self._run_from_user_message(user_msg):
+            yield event

src/plyngent/cli/readline_setup.py

@@ -28,6 +28,7 @@ SLASH_COMMANDS: tuple[str, ...] = (
     "/model",
     "/tools",
     "/rounds",
+    "/retry",
 )
 
 _TOOLS_ARGS: tuple[str, ...] = ("on", "off")

src/plyngent/cli/repl.py

@@ -7,8 +7,8 @@ from typing import TYPE_CHECKING
 import click
 from msgspec import UNSET
 
-from plyngent.cli.display import render_events
 from plyngent.cli.readline_setup import setup_readline
+from plyngent.cli.retry import retry_pending_with_retries, run_user_text_with_retries
 from plyngent.cli.selection import select_model, select_provider
 from plyngent.lmproto.openai_compatible.model import (
     AssistantChatMessage,
@@ -35,6 +35,10 @@ Commands:
   /model [id]        Show or switch model
   /tools [on|off]    Show or toggle tools
   /rounds [n]        Show or set max tool-loop rounds
+  /retry             Retry the last failed user turn (after errors)
+
+On network/API errors, the user turn is not saved to the DB. Auto-retry
+waits 10s, 20s, then 30s (Ctrl+C cancels waits; use /retry later).
 
 Tab completes slash commands and some arguments (provider, model, tools).
 Use --session ID or /resume to continue a prior chat after restart.
@@ -141,6 +145,7 @@ def _cmd_rounds(state: ReplState, arg: str) -> None:
 
 def _cmd_clear(state: ReplState) -> None:
     state.agent.messages.clear()
+    state.agent.pending_retry_text = None
     click.echo("conversation cleared (in-memory only; DB history kept)")
 
 
@@ -197,6 +202,15 @@ def _cmd_history(state: ReplState, arg: str) -> None:
     click.echo(f"session={state.session_id}  messages={len(messages)}  showing={len(messages) - start}")
     for offset, message in enumerate(messages[start:]):
         click.echo(_format_history_message(start + offset, message))
+    if state.agent.pending_retry_text is not None:
+        click.secho(
+            f"(pending retry) user: {_preview_content(state.agent.pending_retry_text)}",
+            fg="yellow",
+        )
+
+
+async def _cmd_retry(state: ReplState) -> None:
+    _ = await retry_pending_with_retries(state.agent)
 
 
 async def _dispatch_slash(state: ReplState, command: str, arg: str) -> bool:
@@ -214,6 +228,7 @@ async def _dispatch_slash(state: ReplState, command: str, arg: str) -> bool:
         "model": lambda: _cmd_model(state, arg),
         "tools": lambda: _cmd_tools(state, arg),
         "rounds": lambda: _cmd_rounds(state, arg),
+        "retry": lambda: _cmd_retry(state),
     }
     handler = handlers.get(command)
     if handler is None:
@@ -269,8 +284,4 @@ async def run_repl(state: ReplState) -> None:
 
         click.secho("user: ", fg="green", nl=False)
         click.echo(line)
-        try:
-            await render_events(state.agent.run(line))
-        except Exception as exc:  # noqa: BLE001 — show API/runtime errors in REPL
-            click.secho(f"error: {exc}", fg="red")
-            click.echo()  # match render_events spacing before next prompt
+        _ = await run_user_text_with_retries(state.agent, line)

src/plyngent/cli/retry.py

@@ -0,0 +1,102 @@
+from __future__ import annotations
+
+import asyncio
+import time
+from typing import TYPE_CHECKING
+
+import click
+
+from plyngent.cli.display import render_events
+
+if TYPE_CHECKING:
+    from collections.abc import AsyncIterator, Callable
+
+    from plyngent.agent import AgentEvent
+    from plyngent.agent.chat import ChatAgent
+
+# Wait before retry attempt 1, 2, and 3 (after the first failure).
+DEFAULT_RETRY_DELAYS_SECONDS: tuple[float, ...] = (10.0, 20.0, 30.0)
+_PREVIEW_LEN = 80
+
+
+async def sleep_cancellable(seconds: float) -> bool:
+    """Sleep in short steps so Ctrl+C can cancel. Returns False if interrupted."""
+    deadline = time.monotonic() + seconds
+    try:
+        while True:
+            remaining = deadline - time.monotonic()
+            if remaining <= 0:
+                return True
+            await asyncio.sleep(min(0.5, remaining))
+    except asyncio.CancelledError:
+        return False
+    except KeyboardInterrupt:
+        return False
+
+
+async def run_turn_with_retries(
+    agent: ChatAgent,
+    *,
+    starter: Callable[[], AsyncIterator[AgentEvent]],
+    delays: tuple[float, ...] = DEFAULT_RETRY_DELAYS_SECONDS,
+) -> bool:
+    """Run a chat turn with automatic retries on failure.
+
+    ``starter`` produces the event stream for the current attempt (``agent.run``
+    or ``agent.retry``). Returns True if the turn completed successfully.
+    """
+    max_retries = len(delays)
+    attempt = 0
+    while True:
+        try:
+            await render_events(starter())
+        except KeyboardInterrupt:
+            click.echo()
+            click.secho("interrupted", fg="yellow")
+            return False
+        except Exception as exc:  # noqa: BLE001 — surface and optionally retry
+            click.secho(f"error: {exc}", fg="red")
+            if attempt >= max_retries:
+                if agent.pending_retry_text is not None:
+                    click.secho(
+                        "auto-retry exhausted; use /retry to try again, or send a new message",
+                        fg="yellow",
+                    )
+                click.echo()
+                return False
+
+            wait = delays[attempt]
+            attempt += 1
+            click.secho(
+                f"auto-retry {attempt}/{max_retries} in {wait:g}s "
+                f"(Ctrl+C to cancel; then /retry later)",
+                fg="yellow",
+            )
+            try:
+                ok = await sleep_cancellable(wait)
+            except KeyboardInterrupt:
+                ok = False
+            if not ok:
+                click.secho("auto-retry cancelled; use /retry to try again", fg="yellow")
+                click.echo()
+                return False
+            click.secho(f"retrying ({attempt}/{max_retries})…", fg="yellow")
+        else:
+            return True
+
+
+async def run_user_text_with_retries(agent: ChatAgent, text: str) -> bool:
+    """Send a new user message with auto-retry."""
+    return await run_turn_with_retries(agent, starter=lambda: agent.run(text))
+
+
+async def retry_pending_with_retries(agent: ChatAgent) -> bool:
+    """Retry the last failed user turn with auto-retry."""
+    if agent.pending_retry_text is None:
+        click.echo("nothing to retry")
+        return False
+    preview = agent.pending_retry_text
+    if len(preview) > _PREVIEW_LEN:
+        preview = preview[:_PREVIEW_LEN] + "…"
+    click.echo(f"retrying: {preview}")
+    return await run_turn_with_retries(agent, starter=agent.retry)

src/plyngent/cli/state.py

@@ -66,6 +66,7 @@ class ReplState:
         session = await self.memory.create_session(name=name)
         self.session_id = session.sid
         self.agent = self._make_agent()
+        self.agent.pending_retry_text = None
 
     async def resume_session(self, session_id: int) -> None:
         row = await self.memory.get_session(session_id)

tests/test_agent/test_loop.py

@@ -2,6 +2,8 @@ from __future__ import annotations
 
 from typing import TYPE_CHECKING, Literal, overload
 
+import pytest
+
 from plyngent.agent import (
     AssistantMessageEvent,
     ChatAgent,
@@ -225,3 +227,91 @@ async def test_chat_agent_memory_roundtrip() -> None:
     await agent2.load_history()
     assert len(agent2.messages) == 2  # noqa: PLR2004
     await store.close()
+
+
+async def test_chat_agent_failed_turn_not_persisted() -> None:
+    store = await MemoryStore.open(DatabaseConfig())
+    session = await store.create_session(name="t")
+
+    class BoomClient:
+        @overload
+        async def chat_completions(
+            self, param: ChatCompletionsParam, *, stream: Literal[False] = False
+        ) -> ChatCompletionResponse: ...
+
+        @overload
+        async def chat_completions(
+            self, param: ChatCompletionsParam, *, stream: Literal[True]
+        ) -> AsyncIterator[ChatCompletionChunk]: ...
+
+        async def chat_completions(
+            self, param: ChatCompletionsParam, *, stream: bool = False
+        ) -> ChatCompletionResponse | AsyncIterator[ChatCompletionChunk]:
+            del param, stream
+            msg = "network down"
+            raise RuntimeError(msg)
+
+    agent = ChatAgent(BoomClient(), model="m", memory=store, session_id=session.sid)
+    with pytest.raises(RuntimeError, match="network down"):
+        _ = [e async for e in agent.run("hello")]
+
+    assert agent.messages == []
+    assert agent.pending_retry_text == "hello"
+    loaded = await store.list_messages(session.sid)
+    assert loaded == []
+    await store.close()
+
+
+async def test_chat_agent_retry_after_failure() -> None:
+    store = await MemoryStore.open(DatabaseConfig())
+    session = await store.create_session(name="t")
+
+    class FlakyClient:
+        calls: int
+
+        def __init__(self) -> None:
+            self.calls = 0
+
+        @overload
+        async def chat_completions(
+            self, param: ChatCompletionsParam, *, stream: Literal[False] = False
+        ) -> ChatCompletionResponse: ...
+
+        @overload
+        async def chat_completions(
+            self, param: ChatCompletionsParam, *, stream: Literal[True]
+        ) -> AsyncIterator[ChatCompletionChunk]: ...
+
+        async def chat_completions(
+            self, param: ChatCompletionsParam, *, stream: bool = False
+        ) -> ChatCompletionResponse | AsyncIterator[ChatCompletionChunk]:
+            del param
+            self.calls += 1
+            if self.calls == 1:
+                msg = "temporary"
+                raise RuntimeError(msg)
+            if stream:
+
+                async def empty() -> AsyncIterator[ChatCompletionChunk]:
+                    empty_chunks: list[ChatCompletionChunk] = []
+                    for chunk in empty_chunks:
+                        yield chunk
+
+                return empty()
+            return _response(AssistantChatMessage(content="recovered"))
+
+    client = FlakyClient()
+    agent = ChatAgent(client, model="m", memory=store, session_id=session.sid)
+    with pytest.raises(RuntimeError, match="temporary"):
+        _ = [e async for e in agent.run("ping")]
+    assert agent.pending_retry_text == "ping"
+    assert await store.list_messages(session.sid) == []
+
+    events = [e async for e in agent.retry()]
+    assert any(isinstance(e, TextDeltaEvent) and e.content == "recovered" for e in events)
+    assert agent.pending_retry_text is None
+    loaded = await store.list_messages(session.sid)
+    assert len(loaded) == 2  # noqa: PLR2004
+    assert isinstance(loaded[0], UserChatMessage)
+    assert loaded[0].content == "ping"
+    await store.close()

tests/test_cli/test_retry.py

@@ -0,0 +1,138 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Literal, overload
+
+from plyngent.agent import ChatAgent
+from plyngent.cli.retry import retry_pending_with_retries, run_turn_with_retries, sleep_cancellable
+from plyngent.config.models import DatabaseConfig
+from plyngent.lmproto.openai_compatible.model import (
+    AssistantChatMessage,
+    ChatCompletionChoice,
+    ChatCompletionChunk,
+    ChatCompletionResponse,
+    ChatCompletionsParam,
+    UserChatMessage,
+)
+from plyngent.memory import MemoryStore
+
+if TYPE_CHECKING:
+    from collections.abc import AsyncIterator
+
+    import pytest
+
+
+def _response(message: AssistantChatMessage) -> ChatCompletionResponse:
+    return ChatCompletionResponse(
+        id="1",
+        object="chat.completion",
+        created=0,
+        model="test",
+        choices=[
+            ChatCompletionChoice(
+                index=0,
+                message=message,
+                logprobs={},
+                finish_reason="stop",
+            )
+        ],
+        system_fingerprint="",
+        usage={},
+    )
+
+
+class FlakyClient:
+    calls: int
+    fail_times: int
+
+    def __init__(self, fail_times: int) -> None:
+        self.calls = 0
+        self.fail_times = fail_times
+
+    @overload
+    async def chat_completions(
+        self, param: ChatCompletionsParam, *, stream: Literal[False] = False
+    ) -> ChatCompletionResponse: ...
+
+    @overload
+    async def chat_completions(
+        self, param: ChatCompletionsParam, *, stream: Literal[True]
+    ) -> AsyncIterator[ChatCompletionChunk]: ...
+
+    async def chat_completions(
+        self, param: ChatCompletionsParam, *, stream: bool = False
+    ) -> ChatCompletionResponse | AsyncIterator[ChatCompletionChunk]:
+        del param
+        self.calls += 1
+        if self.calls <= self.fail_times:
+            msg = f"fail-{self.calls}"
+            raise RuntimeError(msg)
+        if stream:
+
+            async def empty() -> AsyncIterator[ChatCompletionChunk]:
+                empty_chunks: list[ChatCompletionChunk] = []
+                for chunk in empty_chunks:
+                    yield chunk
+
+            return empty()
+        return _response(AssistantChatMessage(content="ok"))
+
+
+async def test_sleep_cancellable_completes() -> None:
+    assert await sleep_cancellable(0.01) is True
+
+
+async def test_auto_retry_eventually_succeeds(monkeypatch: pytest.MonkeyPatch) -> None:
+    async def no_wait(_seconds: float) -> bool:
+        return True
+
+    monkeypatch.setattr("plyngent.cli.retry.sleep_cancellable", no_wait)
+    store = await MemoryStore.open(DatabaseConfig())
+    session = await store.create_session(name="t")
+    client = FlakyClient(fail_times=2)
+    agent = ChatAgent(client, model="m", memory=store, session_id=session.sid)
+
+    ok = await run_turn_with_retries(
+        agent,
+        starter=lambda: agent.run("hello"),
+        delays=(0.01, 0.01, 0.01),
+    )
+    assert ok is True
+    assert client.calls == 3  # noqa: PLR2004
+    loaded = await store.list_messages(session.sid)
+    assert len(loaded) == 2  # noqa: PLR2004
+    assert isinstance(loaded[0], UserChatMessage)
+    assert loaded[0].content == "hello"
+    await store.close()
+
+
+async def test_manual_retry_after_exhausted(monkeypatch: pytest.MonkeyPatch) -> None:
+    async def no_wait(_seconds: float) -> bool:
+        return True
+
+    monkeypatch.setattr("plyngent.cli.retry.sleep_cancellable", no_wait)
+    store = await MemoryStore.open(DatabaseConfig())
+    session = await store.create_session(name="t")
+    client = FlakyClient(fail_times=5)
+    agent = ChatAgent(client, model="m", memory=store, session_id=session.sid)
+
+    ok = await run_turn_with_retries(
+        agent,
+        starter=lambda: agent.run("hold-me"),
+        delays=(0.01,),  # one auto-retry only → 2 attempts total, both fail
+    )
+    assert ok is False
+    assert agent.pending_retry_text == "hold-me"
+    assert await store.list_messages(session.sid) == []
+
+    # Now succeed on manual /retry path
+    client.fail_times = client.calls  # next call succeeds
+    ok2 = await retry_pending_with_retries(agent)
+    # retry_pending uses default delays; force zero-wait already patched
+    # But fail_times set so first attempt of manual retry succeeds immediately
+    assert ok2 is True
+    assert agent.pending_retry_text is None
+    loaded = await store.list_messages(session.sid)
+    assert len(loaded) == 2  # noqa: PLR2004
+    assert isinstance(loaded[0], UserChatMessage)
+    assert loaded[0].content == "hold-me"
+    await store.close()