refactor(runtime): extract shared LLM call layer

README.md

@@ -21,6 +21,8 @@ Version `0.1.x` contains the converged foundation layer:
 - loop configuration, lifecycle contexts, observer protocol, intervention
   merging, and observer dispatch helpers.
 - streamed tool-call recovery checks for missing required arguments.
+- LLM binding, response normalization, streaming assembly/watchdogs, retry
+  classification, runaway recovery, and physical-call orchestration.
 
 The initial extraction is based on the already-merged integration branches:
 
@@ -30,10 +32,10 @@ The initial extraction is based on the already-merged integration branches:
 Those revisions are provenance, not runtime dependencies. AgentCore tests and
 builds without either product checkout.
 
-The LLM call runtime and agent loop remain in the products until tool
-execution, model profiles, retry classification, execution-context storage,
-and runtime hooks have product-neutral boundaries. Moving those files before
-that boundary exists would only hide product coupling inside this package.
+The agent loop remains in the products until tool execution, model profiles,
+tool-call parsing, and execution-context storage have converged boundaries.
+The LLM runtime accepts the three remaining product decisions through explicit
+hooks: wall-deadline lookup, provider-chain state, and sticky-session policy.
 
 ## Repository boundary
 
@@ -109,11 +111,10 @@ edit both products' core copies, that is evidence it belongs here.
 
 1. **Foundation** (this version): messages, token estimation, compaction,
    context budget, trimming.
-2. **Runtime contracts** (in progress): LLM protocols and loop types are now
-   shared; errors, execution-context storage, retry classification, and
-   explicit product hooks remain.
-3. **LLM runtime:** binding, calls, streaming, response normalization, runaway
-   recovery, and the public `llm_client` facade.
+2. **Runtime contracts** (complete): LLM protocols, loop types, errors, retry
+   classification, and explicit product hooks are shared.
+3. **LLM runtime** (complete): binding, calls, streaming, response
+   normalization, runaway recovery, and the public `llm_client` facade.
 4. **Agent loop:** model/tool parsing, tool execution, and `agent_loop`.
 5. Remove product compatibility facades after downstream imports have moved to
    `agent_core`.

agent_core/__init__.py

@@ -1,5 +1,6 @@
 """Product-neutral building blocks for Apodex agent runtimes."""
 
+from agent_core.errors import AgentCoreError
 from agent_core.llm import LLMClient, LLMResponse, StreamDelta
 from agent_core.messages import (
     Message,
@@ -11,6 +12,7 @@
 )
 
 __all__ = [
+    "AgentCoreError",
     "LLMClient",
     "LLMResponse",
     "Message",

agent_core/errors.py

@@ -0,0 +1,164 @@
+"""Exception hierarchy for AgentCore."""
+
+from __future__ import annotations
+
+from typing import Any
+
+
+class AgentCoreError(Exception):
+    """Base exception for all AgentCore errors."""
+
+
+# ── Kernel errors ───────────────────────────────────────────────────────────
+
+
+class KernelError(AgentCoreError):
+    """Errors originating from the OS kernel layer."""
+
+
+class TaskNotFoundError(KernelError):
+    def __init__(self, task_id: str) -> None:
+        super().__init__(f"Task not found: {task_id}")
+        self.task_id = task_id
+
+
+class InvalidStateTransition(KernelError):
+    def __init__(self, task_id: str, current: str, target: str) -> None:
+        super().__init__(f"Invalid transition for {task_id}: {current} → {target}")
+
+
+class ServiceNotRegistered(KernelError):
+    def __init__(self, service_type: type) -> None:
+        super().__init__(f"Service not registered: {service_type.__name__}")
+
+
+class PermissionDenied(KernelError):
+    def __init__(self, role: str, tool: str) -> None:
+        super().__init__(f"Role '{role}' has no permission for tool '{tool}'")
+
+# LLM request errors
+
+class LLMError(AgentCoreError):
+    """Errors from the LLM/provider layer."""
+
+
+class LLMReasoningRunaway(LLMError):
+    """A live stream spent its semantic budget on reasoning-only output.
+
+    Unlike :class:`LLMStreamStalled`, the provider is healthy and actively
+    emitting chunks. The failure is semantic: no non-whitespace visible text
+    or tool-call delta appeared before the configured time/token guard fired.
+
+    ``partial_response`` is intentionally carried separately from provider
+    usage. Early stream cancellation often happens before the terminal usage
+    chunk arrives, so its estimated reasoning tokens must never be presented
+    as authoritative billing data.
+    """
+
+    def __init__(
+        self,
+        *,
+        elapsed_s: float,
+        estimated_tokens: int,
+        trigger: str,
+        partial_response: Any,
+    ) -> None:
+        self.elapsed_s = float(elapsed_s)
+        self.estimated_tokens = int(estimated_tokens)
+        self.trigger = trigger
+        self.partial_response = partial_response
+        super().__init__(
+            "reasoning-only stream exceeded "
+            f"{trigger} guard (elapsed={self.elapsed_s:.1f}s, "
+            f"estimated_tokens={self.estimated_tokens})",
+        )
+
+
+class LLMStreamStalled(LLMError, TimeoutError):
+    """A streaming LLM call went silent mid-flight.
+
+    Subclasses ``asyncio.TimeoutError`` so every existing transient-
+    timeout handler (retry/backoff in ``call_llm``, chain wrappers,
+    classification) treats it identically without changes; carried
+    fields make the distinct failure mode visible in logs and traces.
+    """
+
+    def __init__(
+        self, stall_s: float, chunks_seen: int, elapsed_s: float,
+    ) -> None:
+        self.stall_s = stall_s
+        self.chunks_seen = chunks_seen
+        self.elapsed_s = elapsed_s
+        super().__init__(
+            f"stream stalled: no chunks for {stall_s:.0f}s "
+            f"(chunks_seen={chunks_seen}, elapsed={elapsed_s:.0f}s)",
+        )
+
+
+class LLMDeadlineExceeded(LLMError, TimeoutError):
+    """An LLM attempt was stopped by an enclosing runtime deadline.
+
+    ``reason`` is deliberately carried on the underlying exception as well as
+    on :class:`LLMCallExhausted`. Some callers unwrap ``last_exc`` before
+    handing it to a provider-chain policy; a dedicated type prevents that
+    policy from mistaking an exhausted run budget for an ordinary transient
+    provider timeout.
+    """
+
+    def __init__(self, reason: str, detail: str) -> None:
+        self.reason = reason
+        super().__init__(f"{reason}: {detail}")
+
+
+class LLMCallExhausted(LLMError, RuntimeError):
+    """Raised by ``call_llm`` when retries are exhausted or the error is
+    structurally unrecoverable (4xx without proxy-wrap, or a chain-aware
+    fallback signal like ``model_not_found``).
+
+    Wraps the last exception encountered so the caller (typically the
+    product's agent loop) can surface it to a provider-chain wrapper for
+    L1→L2→L3 rotation. Carries ``last_exc`` separately because
+    ``raise from`` is too opaque for chain-aware classification — a chain
+    wrapper calls ``classify_error(last_exc)`` directly.
+
+    ``last_exc`` must always agree with ``reason``: it is the exception that
+    *caused this raise*, not merely the most recent failure seen. A deadline
+    refusal therefore carries :class:`LLMDeadlineExceeded` even when earlier
+    attempts failed for unrelated reasons. The wrapper's ``reason`` remains
+    authoritative, while the underlying exception preserves the same reason
+    if a caller unwraps it before classification.
+
+    ``prior_exc`` is where that earlier, superseded failure goes: diagnostic
+    context for logs and post-mortems, deliberately outside the field
+    classification reads.
+    """
+
+    def __init__(
+        self,
+        last_exc: BaseException,
+        reason: str,
+        *,
+        prior_exc: BaseException | None = None,
+    ) -> None:
+        self.last_exc = last_exc
+        self.reason = reason
+        self.prior_exc = prior_exc
+        detail = f"call_llm {reason}: {last_exc!r}"
+        if prior_exc is not None and prior_exc is not last_exc:
+            detail += f" (after {prior_exc!r})"
+        super().__init__(detail)
+
+
+__all__ = [
+    "AgentCoreError",
+    "InvalidStateTransition",
+    "KernelError",
+    "LLMCallExhausted",
+    "LLMDeadlineExceeded",
+    "LLMError",
+    "LLMReasoningRunaway",
+    "LLMStreamStalled",
+    "PermissionDenied",
+    "ServiceNotRegistered",
+    "TaskNotFoundError",
+]

agent_core/llm.py

@@ -37,9 +37,9 @@ class StreamDelta:
     usage: dict[str, int] = field(default_factory=dict[str, int])
     finish_reason: str = ""
     model: str = ""
-    # Vendor label of the leg serving this stream, stamped by
-    # ``LLMFallbackChain.stream`` (constant once the chain commits to an
-    # entry — failover only fires before the first yield). The stream
+    # Vendor label of the leg serving this stream, stamped by a product's
+    # provider-chain wrapper (constant once the chain commits to an entry —
+    # failover only fires before the first yield). The stream
     # assembler folds it into ``LLMResponse.response_metadata`` so per-call
     # billing attribution works for streamed calls too — without this the
     # streaming path had no channel for the provider and every billing

agent_core/messages.py

@@ -20,8 +20,8 @@
 is preserved as the ``content`` value when the underlying client returns it;
 callers that need flat text use :func:`text_of`.
 
-This module replaces ``langchain_core.messages`` (BaseMessage / SystemMessage /
-HumanMessage / AIMessage / ToolMessage). It is intentionally dependency-free.
+This module replaces the former framework-specific message classes with a
+small, dependency-free wire contract.
 """
 
 from __future__ import annotations
@@ -34,8 +34,8 @@
 class ToolCall(TypedDict):
     """OpenAI-style tool_call payload — ``function.arguments`` is JSON-encoded.
 
-    Wire key order is fixed ``{type, id, function}`` to match the LangChain
-    serializer the served checkpoints were aligned against; do not reorder.
+    Wire key order is fixed ``{type, id, function}`` to match the serializer
+    byte shape the served checkpoints were aligned against; do not reorder.
     """
 
     id: str
@@ -141,7 +141,7 @@ def for_wire(messages: list[Message]) -> list[Message]:
 
 # Key insertion order: ``content`` first, then ``role``. Some served
 # checkpoints are sensitive to this byte shape (wire byte-equality with
-# LangChain's ``_convert_message_to_dict`` — see migration gotcha #2). Do
+# the legacy message serializer — see migration gotcha #2). Do
 # not reorder these dict literals.
 
 

agent_core/runtime/env.py

@@ -0,0 +1,35 @@
+"""Shared env-variable prefix cascade.
+
+Several modules independently re-walked the same
+``AGENT_CORE_ / MIROHARNESS_ / FRONTIER_AGENT_`` prefix order looking for a
+configured value — one copy per module, easy to drift if a prefix is ever
+added or reordered in only one of them. This is the single implementation
+they converge on: import :func:`first_configured` and let it own the order
+rather than passing a locally-spelled tuple back in.
+"""
+
+from __future__ import annotations
+
+import os
+
+# The portable ``AGENT_CORE_`` spelling wins when multiple aliases are
+# configured, followed by the MiroHarness and FrontierAgent compatibility
+# names. Order matters:
+# callers rely on the first configured prefix winning.
+ENV_PREFIXES = ("AGENT_CORE_", "MIROHARNESS_", "FRONTIER_AGENT_")
+
+
+def first_configured(suffix: str, prefixes: tuple[str, ...] = ENV_PREFIXES) -> tuple[str, str] | None:
+    """Return the ``(name, value)`` of the first non-empty ``{prefix}{suffix}`` env var.
+
+    ``None`` when none of the prefixed names are set (or all are blank).
+    """
+    for prefix in prefixes:
+        name = f"{prefix}{suffix}"
+        raw = os.environ.get(name, "").strip()
+        if raw:
+            return name, raw
+    return None
+
+
+__all__ = ["ENV_PREFIXES", "first_configured"]

agent_core/runtime/llm_request_overrides.py

@@ -0,0 +1,68 @@
+"""Task-local overrides for one physical LLM request.
+
+The runtime occasionally needs to change generation behaviour for one retry
+without mutating a cached/shared client.  ``ContextVar`` keeps that override
+isolated across concurrent tasks and automatically restores the client's
+normal profile on exit.
+
+Provider adapters opt in to the semantic override they understand.  Today the
+OpenAI-compatible adapter maps it onto SGLang/Qwen
+``chat_template_kwargs`` and an explicitly configured ``reasoning_effort``.
+Unsupported adapters simply keep their normal request shape; the runtime's
+retry prompt and output cap remain the portable fallback.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Generator
+from contextlib import contextmanager
+from contextvars import ContextVar
+from dataclasses import dataclass
+
+
+@dataclass(frozen=True)
+class ThinkingRetryOverride:
+    """Semantic thinking controls for a single retry attempt."""
+
+    mode: str = "reduced"
+    thinking_budget: int | None = None
+    reasoning_effort: str | None = None
+
+    @property
+    def disabled(self) -> bool:
+        return self.mode == "disabled"
+
+
+_THINKING_RETRY_OVERRIDE: ContextVar[ThinkingRetryOverride | None] = ContextVar(
+    "agent_core_thinking_retry_override",
+    default=None,
+)
+
+
+def current_thinking_retry_override() -> ThinkingRetryOverride | None:
+    """Return the override active for the current async task, if any."""
+
+    return _THINKING_RETRY_OVERRIDE.get()
+
+
+@contextmanager
+def thinking_retry_override(
+    override: ThinkingRetryOverride | None,
+) -> Generator[None, None, None]:
+    """Apply ``override`` only inside this context and async task."""
+
+    if override is None:
+        yield
+        return
+    token = _THINKING_RETRY_OVERRIDE.set(override)
+    try:
+        yield
+    finally:
+        _THINKING_RETRY_OVERRIDE.reset(token)
+
+
+__all__ = [
+    "ThinkingRetryOverride",
+    "current_thinking_retry_override",
+    "thinking_retry_override",
+]