feat: extract shared provider substrate

README.md

@@ -157,8 +157,8 @@ edit both products' core copies, that is evidence it belongs here.
    bus/fan-in orchestration, resource management, task budgets, and the
    backend-injected DAG scheduler. Product tool assembly, persistence backends,
    and workflow policy remain downstream.
-9. **Provider substrate** (next): converge the shared provider transports,
-   fallback engine, prompt cache, usage metering, and non-blocking streaming
+9. **Provider substrate** (complete in AgentCore): shared provider transports,
+   fallback engine, prompt cache, usage metadata normalization, and non-blocking streaming
    behind product configuration and session-affinity adapters.
 10. Remove product compatibility facades after downstream imports have moved to
    `agent_core`.

agent_core/llm.py

@@ -46,6 +46,21 @@ class StreamDelta:
     # consumer read an empty vendor, which split one model's usage across
     # a ``provider=""`` bucket and a named bucket.
     provider: str = ""
+    # VERBATIM provider content blocks for the assembled turn, carried only by
+    # providers whose replay contract needs the structured block list rather
+    # than flattened text — currently Anthropic extended thinking, where each
+    # ``thinking`` block's cryptographic ``signature`` (and any opaque
+    # ``redacted_thinking`` payload) MUST be echoed back unmodified on the next
+    # turn or the model cannot continue the signed reasoning state.
+    #
+    # ``content``/``reasoning_content`` above are the flattened text channels;
+    # they cannot represent a signature, so without this field the streaming
+    # path silently produced reasoning that ``thinking_format="content_block"``
+    # could not replay. Providers that need no block list leave it empty and
+    # the stream assembler keeps using the flattened string.
+    reasoning_blocks: list[dict[str, Any]] = field(
+        default_factory=list[dict[str, Any]],
+    )
 
 
 @runtime_checkable

agent_core/llm_utils.py

@@ -0,0 +1,27 @@
+"""Small product-neutral LLM helpers."""
+
+from __future__ import annotations
+
+
+def extract_boxed_content(text: str) -> str:
+    r"""Return the content of the last balanced ``\boxed{...}``."""
+    if not text:
+        return ""
+    matches: list[str] = []
+    offset = 0
+    while (start := text.find(r"\boxed{", offset)) >= 0:
+        content_start = start + 7
+        depth = 1
+        pos = content_start
+        while pos < len(text) and depth:
+            depth += (text[pos] == "{") - (text[pos] == "}")
+            pos += 1
+        if depth == 0:
+            matches.append(text[content_start : pos - 1])
+            offset = pos
+        else:
+            offset = content_start
+    return matches[-1] if matches else ""
+
+
+__all__ = ["extract_boxed_content"]

agent_core/providers/__init__.py

@@ -0,0 +1,43 @@
+"""Provider transports and provider-neutral client wrappers."""
+
+from agent_core.providers.anthropic import AnthropicClient
+from agent_core.providers.fallback import FallbackEntry, LLMFallbackChain
+from agent_core.providers.nonblocking_stream import NonBlockingStream, nonblocking_stderr
+from agent_core.providers.openai_chat import (
+    OpenAIClient,
+    SessionQueryResolver,
+    SessionScopeResolver,
+    configure_session_query_resolver,
+    configure_session_scope_resolver,
+)
+from agent_core.providers.openai_responses import OpenAIResponsesClient
+from agent_core.providers.prompt_cache import (
+    AnthropicPromptCacheAdapter,
+    maybe_wrap_for_prompt_cache,
+)
+from agent_core.providers.protocol_client import (
+    build_protocol_client,
+    protocol_of,
+    provider_label,
+    thinking_format_for_protocol,
+)
+
+__all__ = [
+    "AnthropicClient",
+    "AnthropicPromptCacheAdapter",
+    "FallbackEntry",
+    "LLMFallbackChain",
+    "NonBlockingStream",
+    "OpenAIClient",
+    "OpenAIResponsesClient",
+    "SessionQueryResolver",
+    "SessionScopeResolver",
+    "build_protocol_client",
+    "configure_session_query_resolver",
+    "configure_session_scope_resolver",
+    "maybe_wrap_for_prompt_cache",
+    "nonblocking_stderr",
+    "protocol_of",
+    "provider_label",
+    "thinking_format_for_protocol",
+]