agentforge-chat 0.2.1__py3-none-any.whl

agentforge_chat/__init__.py

@@ -0,0 +1,40 @@
+"""`agentforge-chat` — Chat-agent runtime for AgentForge (feat-020).
+
+Public surface: `ChatSession` (chunk 3) + history drivers
+(`InMemoryChatHistory`, `SqliteChatHistory`) + four truncation
+strategies (`SlidingWindow`, `TokenBudget`, `SummariseOldest`,
+`Hybrid`).
+"""
+
+from __future__ import annotations
+
+from agentforge_chat.build import build_chat_session_from_config
+from agentforge_chat.history import InMemoryChatHistory
+from agentforge_chat.session import ChatSession, SafetyMode
+from agentforge_chat.sqlite import SqliteChatHistory
+from agentforge_chat.tokenisers import (
+    Tokeniser,
+    anthropic_tokeniser,
+    tiktoken_tokeniser,
+)
+from agentforge_chat.truncation import (
+    Hybrid,
+    SlidingWindow,
+    SummariseOldest,
+    TokenBudget,
+)
+
+__all__ = [
+    "ChatSession",
+    "Hybrid",
+    "InMemoryChatHistory",
+    "SafetyMode",
+    "SlidingWindow",
+    "SqliteChatHistory",
+    "SummariseOldest",
+    "TokenBudget",
+    "Tokeniser",
+    "anthropic_tokeniser",
+    "build_chat_session_from_config",
+    "tiktoken_tokeniser",
+]

agentforge_chat/_idempotency.py

@@ -0,0 +1,38 @@
+"""Tiny LRU+TTL cache for per-session idempotency keys (feat-020).
+
+Keyed by ``(session_id, key)``; values are the previous
+`ChatResponse`. Entries past TTL are evicted on lookup; entries
+past `max_entries` are evicted oldest-first.
+"""
+
+from __future__ import annotations
+
+import time
+from collections import OrderedDict
+
+
+class IdempotencyCache[V]:
+    def __init__(self, *, ttl_s: float, max_entries: int = 256) -> None:
+        self._ttl = ttl_s
+        self._max = max_entries
+        self._store: OrderedDict[tuple[str, str], tuple[float, V]] = OrderedDict()
+
+    def get(self, session_id: str, key: str) -> V | None:
+        k = (session_id, key)
+        entry = self._store.get(k)
+        if entry is None:
+            return None
+        ts, value = entry
+        if (time.monotonic() - ts) > self._ttl:
+            self._store.pop(k, None)
+            return None
+        # Mark as recently used.
+        self._store.move_to_end(k)
+        return value
+
+    def put(self, session_id: str, key: str, value: V) -> None:
+        k = (session_id, key)
+        self._store[k] = (time.monotonic(), value)
+        self._store.move_to_end(k)
+        while len(self._store) > self._max:
+            self._store.popitem(last=False)

agentforge_chat/_locks.py

@@ -0,0 +1,115 @@
+"""Per-session lock registry (feat-020).
+
+`ChatSession.send` / `stream` acquires a session-scoped lock so
+concurrent calls against the same `session_id` queue. v0.1 shipped
+an in-process `asyncio.Lock` via `WeakValueDictionary`. v0.2 extends
+the surface to support cross-process locks (Redis-backed) via a
+`SessionLock` Protocol that both shapes satisfy.
+
+Default factory keeps the in-process behaviour. Multi-worker
+deployments inject `redis_session_lock_factory(...)` from
+`agentforge-chat-history-redis`.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import weakref
+from collections.abc import Callable
+from types import TracebackType
+from typing import Protocol
+
+
+class SessionLock(Protocol):  # pragma: no cover — Protocol method stubs
+    """Async-context-manager lock keyed by `session_id`.
+
+    `ChatSession` calls `async with lock:` once per turn.
+    Implementations:
+
+    - :class:`InMemorySessionLock` — wraps a per-session
+      ``asyncio.Lock``. Default; single-process only.
+    - ``RedisSessionLock`` (in `agentforge-chat-history-redis`) —
+      cross-process; uses Redis ``SET NX PX`` + UUID fencing.
+    """
+
+    async def __aenter__(self) -> SessionLock: ...
+
+    async def __aexit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc: BaseException | None,
+        tb: TracebackType | None,
+    ) -> None: ...
+
+
+SessionLockFactory = Callable[[str], SessionLock]
+"""Build a `SessionLock` for one ``session_id``. v0.2 lets callers
+inject this on `ChatSession` / `ChatServer` construction."""
+
+
+class InMemorySessionLock:
+    """Wraps a per-session `asyncio.Lock` so multiple chat turns on
+    the same session_id queue inside one process.
+
+    Conforms structurally to `SessionLock`.
+    """
+
+    def __init__(self, lock: asyncio.Lock) -> None:
+        self._lock = lock
+
+    async def __aenter__(self) -> InMemorySessionLock:
+        await self._lock.acquire()
+        return self
+
+    async def __aexit__(
+        self,
+        exc_type: type[BaseException] | None,
+        exc: BaseException | None,
+        tb: TracebackType | None,
+    ) -> None:
+        self._lock.release()
+
+
+class _LockRegistry:
+    def __init__(self) -> None:
+        self._locks: weakref.WeakValueDictionary[str, asyncio.Lock] = weakref.WeakValueDictionary()
+
+    def get(self, session_id: str) -> asyncio.Lock:
+        lock = self._locks.get(session_id)
+        if lock is None:
+            lock = asyncio.Lock()
+            self._locks[session_id] = lock
+        return lock
+
+
+_REGISTRY = _LockRegistry()
+
+
+def lock_for(session_id: str) -> asyncio.Lock:
+    """Return the (shared, weak-referenced) raw `asyncio.Lock`.
+
+    Retained for backward-compatibility with v0.1 callers that read
+    `ChatSession._lock` directly. New code should use
+    :func:`default_session_lock_factory` or inject a custom
+    `SessionLockFactory`.
+    """
+    return _REGISTRY.get(session_id)
+
+
+def default_session_lock_factory(session_id: str) -> SessionLock:
+    """Build the default in-process `SessionLock` for ``session_id``.
+
+    Wraps the shared `asyncio.Lock` from the weak-ref registry so
+    multiple `ChatSession` instances bound to the same session_id
+    still queue correctly.
+    """
+    return InMemorySessionLock(_REGISTRY.get(session_id))
+
+
+__all__ = [
+    "InMemorySessionLock",
+    "SessionLock",
+    "SessionLockFactory",
+    "default_session_lock_factory",
+    "lock_for",
+]

agentforge_chat/_segment.py

@@ -0,0 +1,45 @@
+"""Sentence segmenter for the buffer-then-stream path (feat-020).
+
+v0.2 ships `ChatSession.stream()` in the spec's
+`safety_mode: "buffer-then-stream"` semantics: the agent runs to
+completion, then the assistant turn is sliced into sentence-ish
+chunks for the wire format. Real per-token streaming follows in a
+later release without changing this surface.
+"""
+
+from __future__ import annotations
+
+import re
+
+_SENTENCE_BOUNDARY = re.compile(r"(?<=[.!?])\s+")
+_MAX_CHUNK_CHARS = 200
+"""Soft cap so a single uninterrupted paragraph still emits as
+multiple chunks."""
+
+
+def segment_for_stream(text: str) -> list[str]:
+    """Split ``text`` into wire-format-friendly chunks.
+
+    Prefers sentence boundaries (``.!?`` followed by whitespace);
+    falls back to paragraph boundaries; falls back to a hard
+    `_MAX_CHUNK_CHARS` cap.
+    """
+    if not text:
+        return []
+    parts = [p for p in _SENTENCE_BOUNDARY.split(text) if p]
+    out: list[str] = []
+    for part in parts:
+        out.extend(_split_long(part))
+    return out
+
+
+def _split_long(text: str) -> list[str]:
+    if len(text) <= _MAX_CHUNK_CHARS:
+        return [text]
+    pieces: list[str] = []
+    cursor = 0
+    while cursor < len(text):
+        end = min(cursor + _MAX_CHUNK_CHARS, len(text))
+        pieces.append(text[cursor:end])
+        cursor = end
+    return pieces

agentforge_chat/_window.py

@@ -0,0 +1,86 @@
+"""Sentence-window buffer for the streaming output-guardrail path
+(feat-020 v0.3 polish).
+
+When `ChatSession.safety_mode == "sentence-window"`, streamed text
+tokens accumulate in a `_SentenceWindowBuffer`. Each `push(text)` call
+returns the completed sentences ready to validate (terminator
+followed by whitespace, OR newline, OR 200-char hard cap so a
+paragraph without punctuation still flushes). The buffer's
+`flush()` returns whatever residual remains so callers can pipe
+the partial through the guardrail one last time at end-of-stream.
+
+Boundary heuristic mirrors :mod:`agentforge_chat._segment` so the
+streaming and buffer-then-stream paths produce comparable chunk
+shapes. Multi-language sentence segmentation is out of scope for
+v0.3; the regex is English-centric.
+"""
+
+from __future__ import annotations
+
+import re
+
+_SOFT_MAX_CHARS = 200
+"""Soft cap so an unpunctuated paragraph still emits as chunks."""
+
+_BOUNDARY_RE = re.compile(r"[.!?]\s+|\n+")
+
+
+class _SentenceWindowBuffer:
+    """Accumulates streamed tokens; releases completed sentences.
+
+    Not thread-safe — `ChatSession._stream_per_token` already
+    serialises per-session via a lock, so each buffer instance is
+    single-writer / single-reader.
+    """
+
+    def __init__(self) -> None:
+        self._buf = ""
+
+    def push(self, text: str) -> list[str]:
+        """Append `text` to the buffer; return any completed sentences.
+
+        A sentence is "completed" when:
+        - a `.!?` is followed by whitespace, OR
+        - a newline appears, OR
+        - the buffer length exceeds `_SOFT_MAX_CHARS`.
+
+        Remaining partial text stays buffered until the next push
+        or a `flush()`.
+        """
+        if not text:
+            return []
+        self._buf += text
+        completed: list[str] = []
+        while True:
+            cut = self._find_cut()
+            if cut is None:
+                break
+            sentence = self._buf[:cut].rstrip()
+            if sentence:
+                completed.append(sentence)
+            self._buf = self._buf[cut:].lstrip()
+        return completed
+
+    def flush(self) -> str:
+        """Return the residual buffer contents + reset internal state.
+
+        Callers run this through their per-sentence pipeline one
+        last time so end-of-stream text isn't dropped on the floor.
+        Returns an empty string when the buffer is already empty.
+        """
+        residual, self._buf = self._buf, ""
+        return residual
+
+    def _find_cut(self) -> int | None:
+        """Return the byte index at which to slice off a completed
+        sentence, or `None` if nothing is ready yet.
+
+        Priority: punctuation/newline boundary first; hard-cap
+        fallback at `_SOFT_MAX_CHARS`.
+        """
+        match = _BOUNDARY_RE.search(self._buf)
+        if match is not None:
+            return match.end()
+        if len(self._buf) >= _SOFT_MAX_CHARS:
+            return _SOFT_MAX_CHARS
+        return None

agentforge_chat/build.py

@@ -0,0 +1,112 @@
+"""Config-driven `ChatSession` construction (feat-020).
+
+`build_chat_session_from_config(config, agent)` reads
+`modules.chat:` and assembles:
+
+  - the history-store driver (resolver category `chat.history`),
+  - the truncation strategy (resolver category `chat.truncation`),
+  - per-turn / per-session budget + idempotency knobs.
+
+Drivers that expose `from_config(cfg)` are preferred; otherwise
+the class is constructed with `**cfg`. Async-factory drivers
+(e.g. `SqliteChatHistory.from_path`) are recognised by the
+`from_path` classmethod returning an awaitable.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from agentforge.agent import Agent
+from agentforge_core.config.schema import AgentForgeConfig
+from agentforge_core.contracts.chat import ChatHistoryStore, HistoryTruncationStrategy
+from agentforge_core.production.exceptions import ModuleError
+from agentforge_core.resolver import Resolver
+
+from agentforge_chat.history import InMemoryChatHistory
+from agentforge_chat.session import ChatSession, SafetyMode
+from agentforge_chat.truncation import SlidingWindow
+
+
+async def build_chat_session_from_config(
+    config: AgentForgeConfig,
+    agent: Agent,
+    *,
+    session_id: str | None = None,
+    owner: str | None = None,
+    system_prompt: str | None = None,
+) -> ChatSession:
+    """Instantiate a `ChatSession` driven by `modules.chat:`.
+
+    Caller still owns the `Agent`; the chat session merely wraps it.
+    `session_id` defaults to a fresh ULID-ish hex when omitted.
+    """
+    chat_cfg = config.modules.chat
+    history: ChatHistoryStore = InMemoryChatHistory()
+    truncation: HistoryTruncationStrategy = SlidingWindow(50)
+    per_turn = None
+    per_session = None
+    idem_window = 60.0
+    safety_mode: SafetyMode = "buffer-then-stream"
+    if chat_cfg is not None:
+        if chat_cfg.history is not None:
+            history = await _build_history(chat_cfg.history.driver, chat_cfg.history.config)
+        if chat_cfg.truncation is not None:
+            truncation = _build_truncation(chat_cfg.truncation.strategy, chat_cfg.truncation.config)
+        per_turn = chat_cfg.session.per_turn_budget_usd
+        per_session = chat_cfg.session.per_session_budget_usd
+        idem_window = chat_cfg.session.idempotency_window_s
+        safety_mode = chat_cfg.session.safety_mode
+    return ChatSession(
+        agent=agent,
+        session_id=session_id,
+        history_store=history,
+        system_prompt=system_prompt,
+        truncation=truncation,
+        owner=owner,
+        per_turn_budget_usd=per_turn,
+        per_session_budget_usd=per_session,
+        idempotency_window_s=idem_window,
+        safety_mode=safety_mode,
+    )
+
+
+async def _build_history(driver: str, cfg: dict[str, Any]) -> ChatHistoryStore:
+    cls = Resolver.global_().resolve("chat.history", driver)
+    instance = await _maybe_async(_instantiate(cls, cfg))
+    if not isinstance(instance, ChatHistoryStore):
+        raise ModuleError(
+            f"Resolved chat.history driver {driver!r} ({cls.__name__}) does not "
+            f"implement ChatHistoryStore."
+        )
+    return instance
+
+
+def _build_truncation(name: str, cfg: dict[str, Any]) -> HistoryTruncationStrategy:
+    cls = Resolver.global_().resolve("chat.truncation", name)
+    instance = _instantiate(cls, cfg)
+    if not isinstance(instance, HistoryTruncationStrategy):
+        raise ModuleError(
+            f"Resolved chat.truncation {name!r} ({cls.__name__}) does not "
+            f"implement HistoryTruncationStrategy."
+        )
+    return instance
+
+
+def _instantiate(cls: type, cfg: dict[str, Any]) -> Any:
+    from_config = getattr(cls, "from_config", None)
+    if callable(from_config):
+        return from_config(cfg)
+    from_path = getattr(cls, "from_path", None)
+    if callable(from_path) and "path" in cfg:
+        return from_path(cfg["path"])
+    return cls(**cfg)
+
+
+async def _maybe_async(value: Any) -> Any:
+    if hasattr(value, "__await__"):
+        return await value
+    return value
+
+
+__all__ = ["build_chat_session_from_config"]

agentforge_chat/history.py

@@ -0,0 +1,126 @@
+"""`InMemoryChatHistory` — process-local default `ChatHistoryStore`.
+
+Backs `ChatSession` when no driver is configured. Useful for tests
+and tiny demos; not persistent across process restarts.
+"""
+
+from __future__ import annotations
+
+import asyncio
+from collections.abc import Mapping
+from datetime import UTC, datetime
+from typing import Any
+
+from agentforge_core.contracts.chat import ChatHistoryStore
+from agentforge_core.values.chat import ChatTurn, SessionInfo
+
+
+class InMemoryChatHistory(ChatHistoryStore):
+    """Thread-safe in-memory implementation of `ChatHistoryStore`."""
+
+    def __init__(self) -> None:
+        self._turns: dict[str, list[ChatTurn]] = {}
+        self._meta: dict[str, dict[str, Any]] = {}
+        self._owners: dict[str, str | None] = {}
+        self._created_at: dict[str, datetime] = {}
+        self._last_active: dict[str, datetime] = {}
+        self._lock = asyncio.Lock()
+
+    async def append(self, turn: ChatTurn) -> None:
+        async with self._lock:
+            self._turns.setdefault(turn.session_id, []).append(turn)
+            now = datetime.now(UTC)
+            self._created_at.setdefault(turn.session_id, now)
+            self._last_active[turn.session_id] = now
+
+    async def load(
+        self,
+        session_id: str,
+        *,
+        limit: int | None = None,
+        before: datetime | None = None,
+        after: datetime | None = None,
+        roles: list[str] | None = None,
+    ) -> list[ChatTurn]:
+        async with self._lock:
+            turns = list(self._turns.get(session_id, []))
+        if before is not None:
+            turns = [t for t in turns if t.timestamp < before]
+        if after is not None:
+            turns = [t for t in turns if t.timestamp > after]
+        if roles is not None:
+            allowed = set(roles)
+            turns = [t for t in turns if t.role in allowed]
+        turns.sort(key=lambda t: t.timestamp)
+        if limit is not None:
+            turns = turns[:limit]
+        return turns
+
+    async def count(self, session_id: str) -> int:
+        async with self._lock:
+            return len(self._turns.get(session_id, []))
+
+    async def delete_session(self, session_id: str) -> int:
+        async with self._lock:
+            removed = len(self._turns.pop(session_id, []))
+            self._meta.pop(session_id, None)
+            self._owners.pop(session_id, None)
+            self._created_at.pop(session_id, None)
+            self._last_active.pop(session_id, None)
+            return removed
+
+    async def list_sessions(
+        self,
+        *,
+        owner: str | None = None,
+        limit: int = 100,
+        before: datetime | None = None,
+    ) -> list[SessionInfo]:
+        async with self._lock:
+            out = [self._build_info(sid) for sid in self._turns]
+        if owner is not None:
+            out = [s for s in out if s.owner == owner]
+        if before is not None:
+            out = [s for s in out if s.last_active_at < before]
+        out.sort(key=lambda s: s.last_active_at, reverse=True)
+        return out[:limit]
+
+    async def update_session_metadata(self, session_id: str, metadata: Mapping[str, Any]) -> None:
+        async with self._lock:
+            bag = self._meta.setdefault(session_id, {})
+            for k, v in metadata.items():
+                bag[k] = v
+            if "owner" in metadata:
+                self._owners[session_id] = metadata["owner"]
+
+    async def expire_before(self, cutoff: datetime) -> int:
+        async with self._lock:
+            doomed = [sid for sid, last in self._last_active.items() if last < cutoff]
+            for sid in doomed:
+                self._turns.pop(sid, None)
+                self._meta.pop(sid, None)
+                self._owners.pop(sid, None)
+                self._created_at.pop(sid, None)
+                self._last_active.pop(sid, None)
+            return len(doomed)
+
+    async def close(self) -> None:
+        return None
+
+    def capabilities(self) -> set[str]:
+        return {"ttl"}
+
+    def _build_info(self, sid: str) -> SessionInfo:
+        turns = self._turns.get(sid, [])
+        return SessionInfo(
+            id=sid,
+            owner=self._owners.get(sid),
+            created_at=self._created_at.get(sid, datetime.now(UTC)),
+            last_active_at=self._last_active.get(sid, datetime.now(UTC)),
+            turn_count=len(turns),
+            total_cost_usd=sum(t.cost_usd for t in turns),
+            metadata=dict(self._meta.get(sid, {})),
+        )
+
+
+__all__ = ["InMemoryChatHistory"]

agentforge_chat/manifest.yaml

@@ -0,0 +1,32 @@
+# Module manifest for `agentforge add module chat` (feat-010).
+name: chat
+description: |
+  Chat-agent runtime (feat-020). Adds the `agentforge_chat` package
+  with `ChatSession`, in-memory + SQLite history drivers, and four
+  truncation strategies.
+
+distribution:
+  pip_name: agentforge-chat
+
+config_block:
+  modules.chat:
+    history:
+      driver: memory   # memory | sqlite (sqlite needs `agentforge-chat[sqlite]`)
+      config: {}
+    truncation:
+      strategy: sliding_window
+      max_turns: 50
+    session:
+      per_turn_budget_usd: null
+      per_session_budget_usd: null
+      idempotency_window_s: 60
+
+entry_points:
+  agentforge.chat.history:
+    memory: agentforge_chat.history:InMemoryChatHistory
+    sqlite: agentforge_chat.sqlite:SqliteChatHistory
+  agentforge.chat.truncation:
+    sliding_window: agentforge_chat.truncation:SlidingWindow
+    token_budget: agentforge_chat.truncation:TokenBudget
+    summarise_oldest: agentforge_chat.truncation:SummariseOldest
+    hybrid: agentforge_chat.truncation:Hybrid