diff --git a/docs/LOCAL_MODEL_SERVING.md b/docs/LOCAL_MODEL_SERVING.md
index d6b5eee..b8f56ad 100644
--- a/docs/LOCAL_MODEL_SERVING.md
+++ b/docs/LOCAL_MODEL_SERVING.md
@@ -276,3 +276,33 @@ biofoundry analyze-study \
 - Increase concurrency for throughput only before freezing a confirmatory protocol.
   Concurrency does not grant extra scheduled decisions, but it can change wall time and
   may expose backend-specific sampling behavior.
+
+## 8. Optional serving switches for other local servers
+
+Two `llm` options, both off by default, help with servers whose Responses endpoint is
+incomplete or whose models copy prompt placeholders into ID fields. With both off, request
+bodies, prompts and the recorded `action_schema_sha256` are exactly as before.
+
+```yaml
+llm:
+  transport: chat_completions            # default "responses"
+  extra_body:
+    chat_template_kwargs: {enable_thinking: false}
+  strict_id_patterns: true               # default false
+```
+
+- `transport: chat_completions` posts to `{base_url}/chat/completions` with
+  `response_format: {type: json_schema, json_schema: {name, schema, strict}}`, the same schema
+  object the Responses transport sends in `text.format`. Use it when a server's `/v1/responses`
+  drops `text.format` (llama-server does) or does not exist. `max_tokens`, `temperature` and
+  `reasoning_effort` map one-to-one; usage is recorded under the Responses names so traces and
+  the compute audit stay uniform; `finish_reason: length` raises the same incomplete error, now
+  with the length, head and tail of the truncated reply; a leading `<think>...</think>` block is
+  stripped if a server returns reasoning inline.
+- `extra_body` merges arbitrary fields into the request body of either transport, for
+  server-specific switches such as vLLM's `chat_template_kwargs`.
+- `strict_id_patterns` adds a regex `pattern` to the three free-string ID fields of the action
+  schema (`reply_to`, `target_agent_id`, `target_artifact_id`) so constrained decoding cannot
+  emit placeholders such as `NONE`. It changes `action_schema_sha256`, so traces made with it
+  are self-labelled. The backend must honour `pattern`: on vLLM use
+  `--structured-outputs-config '{"backend":"guidance"}'` (xgrammar rejects the schema).
diff --git a/src/biofoundry/cli.py b/src/biofoundry/cli.py
index b95f114..1b738ce 100644
--- a/src/biofoundry/cli.py
+++ b/src/biofoundry/cli.py
@@ -195,6 +195,7 @@ def _reproducibility_metadata(config: GameConfig) -> dict[str, Any]:
         allow_replies=replies_available(config),
         resource_names=scenario.resource_names if scenario is not None else None,
         operation_names=scenario.operation_names if scenario is not None else None,
+        strict_id_patterns=config.llm.strict_id_patterns,
     )["schema"]
     prompt_text = (
         SCENARIO_SYSTEM_PROMPT.format(
diff --git a/src/biofoundry/config.py b/src/biofoundry/config.py
index 66f940e..35753b6 100644
--- a/src/biofoundry/config.py
+++ b/src/biofoundry/config.py
@@ -127,6 +127,18 @@ class LLMConfig:
     # Work around constrained decoders that cannot terminate JSON number tokens.
     # Numbers travel as short decimal strings and are coerced before validation.
     grammar_safe_numbers: bool = False
+    # Replication-package addition (default off; see replication/patches/README.md). Small
+    # open-weight models copy the prompt's "use NONE when irrelevant" into the free-string ID
+    # fields, and the engine then rejects the whole action. With this flag the JSON schema
+    # constrains reply_to / target_agent_id / target_artifact_id to their ID grammar or "".
+    strict_id_patterns: bool = False
+    # Replication-package addition (default "responses"). "chat_completions" posts to
+    # {base_url}/chat/completions with response_format json_schema, for servers whose Responses
+    # route does not enforce text.format (llama-server converts /v1/responses internally and
+    # drops the schema). extra_body is merged into every request, e.g.
+    # {"chat_template_kwargs": {"enable_thinking": false}} for Qwen-style thinking models.
+    transport: str = "responses"
+    extra_body: dict[str, Any] | None = None
     # Some local tokenizers strongly prefer legal JSON whitespace forever under a
     # grammar.  A negative bias on tokenizer tokens that decode only to whitespace
     # preserves compact JSON while leaving content tokens unconstrained.
diff --git a/src/biofoundry/providers/openai_compatible.py b/src/biofoundry/providers/openai_compatible.py
index 9d238ac..27c96ac 100644
--- a/src/biofoundry/providers/openai_compatible.py
+++ b/src/biofoundry/providers/openai_compatible.py
@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import asyncio
+import copy
 import http.client
 import json
 import os
@@ -94,6 +95,52 @@ def _whitespace_token_ids(
     return tuple(token_ids)
 
 
+def _chat_usage_as_responses(usage: dict[str, Any]) -> dict[str, Any]:
+    """Map Chat Completions usage keys onto the Responses names the traces use."""
+    mapped: dict[str, Any] = {}
+    for src, dst in (
+        ("prompt_tokens", "input_tokens"),
+        ("completion_tokens", "output_tokens"),
+        ("total_tokens", "total_tokens"),
+    ):
+        if isinstance(usage.get(src), int):
+            mapped[dst] = usage[src]
+    return mapped
+
+
+def _extract_chat_text(data: dict[str, Any]) -> str:
+    """Extract assistant text from a non-streaming Chat Completions payload."""
+    error = data.get("error")
+    if error:
+        detail = str(error.get("message") or error) if isinstance(error, dict) else str(error)
+        raise ProviderError(f"model response failed: {detail[:1200]}")
+    choices = data.get("choices")
+    if not isinstance(choices, list) or not choices or not isinstance(choices[0], dict):
+        raise ProviderError(f"Chat Completions payload contains no choices: {data}")
+    choice = choices[0]
+    finish = choice.get("finish_reason")
+    message = choice.get("message") or {}
+    if finish == "length":
+        # Keep the head and tail of the truncated text in the error so the trace shows what
+        # the model was doing when it hit max_tokens (whitespace loop, repeated field, ...).
+        partial = message.get("content") if isinstance(message.get("content"), str) else ""
+        tail = ""
+        if partial:
+            tail = f" | {len(partial)} chars, head {partial[:120]!r}, tail {partial[-160:]!r}"
+        raise ProviderError(
+            "model response status is 'incomplete': {'reason': 'max_output_tokens'}" + tail
+        )
+    if message.get("refusal"):
+        raise ProviderError(f"model refused the request: {str(message['refusal'])[:1200]}")
+    content = message.get("content")
+    if not isinstance(content, str) or not content:
+        raise ProviderError(f"Chat Completions payload contains no content: {data}")
+    # Thinking models without a server-side reasoning parser prepend their reasoning block.
+    if content.lstrip().startswith("<think>") and "</think>" in content:
+        content = content.split("</think>", 1)[1]
+    return content.strip()
+
+
 def _extract_output_text(data: dict[str, Any]) -> str:
     """Extract assistant text from a non-streaming Responses API payload."""
     error = data.get("error")
@@ -159,7 +206,11 @@ def __init__(
         self.allow_replies = allow_replies
         self.resource_names = None if resource_names is None else tuple(resource_names)
         self.operation_names = None if operation_names is None else tuple(operation_names)
-        self.endpoint = config.base_url.rstrip("/") + "/responses"
+        if config.transport not in ("responses", "chat_completions"):
+            raise ValueError("llm.transport must be 'responses' or 'chat_completions'")
+        self.endpoint = config.base_url.rstrip("/") + (
+            "/chat/completions" if config.transport == "chat_completions" else "/responses"
+        )
         self.semaphore = asyncio.Semaphore(config.concurrency)
         self._json_whitespace_logit_bias: dict[str, float] = {}
         if config.json_whitespace_logit_bias is not None:
@@ -185,6 +236,8 @@ async def generate_record(
             return await asyncio.to_thread(self._generate_sync_record, messages)
 
     def _request_body(self, messages: Sequence[ChatMessage]) -> dict[str, Any]:
+        if self.config.transport == "chat_completions":
+            return self._chat_completions_body(messages)
         body = {
             "model": self.config.model,
             "input": [
@@ -208,12 +261,58 @@ def _request_body(self, messages: Sequence[ChatMessage]) -> dict[str, Any]:
                 allow_replies=self.allow_replies,
                 resource_names=self.resource_names,
                 operation_names=self.operation_names,
+                strict_id_patterns=self.config.strict_id_patterns,
             )
             body["text"] = {"format": output_format}
             if self._json_whitespace_logit_bias:
                 body["logit_bias"] = dict(self._json_whitespace_logit_bias)
         elif self.config.json_mode:
             body["text"] = {"format": {"type": "json_object"}}
+        if self.config.extra_body:
+            body.update(copy.deepcopy(self.config.extra_body))
+        return body
+
+    def _output_format(self) -> dict[str, Any]:
+        return bounded_action_plan_text_format(
+            self.config.max_plan_actions,
+            grammar_safe_numbers=self.config.grammar_safe_numbers,
+            allowed_action_types=self.allowed_action_types,
+            allow_addressing=self.allow_addressing,
+            allow_replies=self.allow_replies,
+            resource_names=self.resource_names,
+            operation_names=self.operation_names,
+            strict_id_patterns=self.config.strict_id_patterns,
+        )
+
+    def _chat_completions_body(self, messages: Sequence[ChatMessage]) -> dict[str, Any]:
+        """Replication-package addition: the same request as Chat Completions."""
+        body: dict[str, Any] = {
+            "model": self.config.model,
+            "messages": [
+                {"role": message.role, "content": message.content} for message in messages
+            ],
+            "max_tokens": self.config.max_tokens,
+        }
+        if self.config.reasoning_effort is not None:
+            body["reasoning_effort"] = self.config.reasoning_effort
+        else:
+            body["temperature"] = self.config.temperature
+        if self.config.structured_output:
+            output_format = self._output_format()
+            body["response_format"] = {
+                "type": "json_schema",
+                "json_schema": {
+                    "name": output_format["name"],
+                    "schema": output_format["schema"],
+                    "strict": bool(output_format.get("strict", True)),
+                },
+            }
+            if self._json_whitespace_logit_bias:
+                body["logit_bias"] = dict(self._json_whitespace_logit_bias)
+        elif self.config.json_mode:
+            body["response_format"] = {"type": "json_object"}
+        if self.config.extra_body:
+            body.update(copy.deepcopy(self.config.extra_body))
         return body
 
     def _generate_sync(self, messages: Sequence[ChatMessage]) -> str:
@@ -262,8 +361,13 @@ def _generate_sync_record(
         usage = data.get("usage", {})
         if not isinstance(usage, dict):
             usage = {}
+        if self.config.transport == "chat_completions":
+            usage = _chat_usage_as_responses(usage)
+            text = _extract_chat_text(data)
+        else:
+            text = _extract_output_text(data)
         return GenerationResult(
-            text=_extract_output_text(data),
+            text=text,
             usage={str(key): value for key, value in usage.items()},
             metadata={
                 "response_id": str(data.get("id", "")),
diff --git a/src/biofoundry/structured_output.py b/src/biofoundry/structured_output.py
index 6c94dcf..945a36f 100644
--- a/src/biofoundry/structured_output.py
+++ b/src/biofoundry/structured_output.py
@@ -621,6 +621,7 @@ def bounded_action_plan_text_format(
     allow_replies: bool = True,
     resource_names: Iterable[str] | None = None,
     operation_names: Iterable[str] | None = None,
+    strict_id_patterns: bool = False,
 ) -> dict[str, Any]:
     """Return the exact provider contract for one experimental capability set."""
 
@@ -676,9 +677,24 @@ def bounded_action_plan_text_format(
         if not retained:
             raise ValueError("allowed_action_types must retain at least one action")
         result["schema"]["$defs"]["action"]["anyOf"] = retained
+    if strict_id_patterns:
+        # Replication-package addition: pin free-string ID fields to their ID grammar or the
+        # empty string so constrained decoders cannot emit placeholders such as "NONE".
+        for branch in result["schema"]["$defs"]["action"]["anyOf"]:
+            for field, prefix in ID_FIELD_PREFIXES.items():
+                prop = branch["properties"].get(field)
+                if isinstance(prop, dict) and "enum" not in prop:
+                    prop["pattern"] = f"^({prefix}_[0-9]+)?$"
     return result
 
 
+ID_FIELD_PREFIXES = {
+    "reply_to": "message",
+    "target_agent_id": "agent",
+    "target_artifact_id": "artifact",
+}
+
+
 def coerce_transport_numbers(value: Any) -> Any:
     """Convert grammar-safe numeric strings to canonical plan numbers in-place."""
 
diff --git a/tests/test_local_model_serving.py b/tests/test_local_model_serving.py
new file mode 100644
index 0000000..da72aa8
--- /dev/null
+++ b/tests/test_local_model_serving.py
@@ -0,0 +1,103 @@
+"""Tests for the opt-in local-model serving options: chat-completions transport, extra_body,
+constrained ID fields, and the truncated-reply detail on finish_reason=length."""
+
+import hashlib
+import json
+
+import pytest
+
+from biofoundry.config import LLMConfig
+from biofoundry.providers.base import ChatMessage
+from biofoundry.providers.openai_compatible import (
+    OpenAICompatibleProvider,
+    ProviderError,
+    _chat_usage_as_responses,
+    _extract_chat_text,
+)
+from biofoundry.structured_output import ACTION_PLAN_SCHEMA, bounded_action_plan_text_format
+
+
+def _schema_hash(**kwargs: object) -> str:
+    schema = bounded_action_plan_text_format(12, **kwargs)["schema"]
+    return hashlib.sha256(json.dumps(schema, sort_keys=True).encode()).hexdigest()
+
+
+def test_strict_id_patterns_off_leaves_schema_unchanged() -> None:
+    assert _schema_hash() == _schema_hash(strict_id_patterns=False)
+
+
+def test_strict_id_patterns_constrains_id_fields_and_changes_hash() -> None:
+    assert _schema_hash(strict_id_patterns=True) != _schema_hash()
+    schema = bounded_action_plan_text_format(12, strict_id_patterns=True)["schema"]
+    text = json.dumps(schema)
+    assert "^(message_[0-9]+)?$" in text
+    assert "^(agent_[0-9]+)?$" in text
+    assert "^(artifact_[0-9]+)?$" in text
+    # the plain schema carries no such patterns
+    assert "message_[0-9]" not in json.dumps(bounded_action_plan_text_format(12)["schema"])
+
+
+def test_chat_completions_transport_builds_openai_chat_body() -> None:
+    provider = OpenAICompatibleProvider(
+        LLMConfig(
+            model="local-qwen",
+            max_tokens=4096,
+            structured_output=True,
+            transport="chat_completions",
+            extra_body={"chat_template_kwargs": {"enable_thinking": False}},
+        )
+    )
+    messages = [ChatMessage("system", "Act."), ChatMessage("user", "Return JSON.")]
+    body = provider._request_body(messages)
+    assert provider.endpoint == "http://127.0.0.1:8000/v1/chat/completions"
+    assert body["messages"] == [
+        {"role": "system", "content": "Act."},
+        {"role": "user", "content": "Return JSON."},
+    ]
+    assert body["max_tokens"] == 4096
+    assert body["temperature"] == 0.7
+    assert body["response_format"] == {
+        "type": "json_schema",
+        "json_schema": {
+            "name": "biofoundry_agent_plan",
+            "strict": True,
+            "schema": ACTION_PLAN_SCHEMA,
+        },
+    }
+    assert body["chat_template_kwargs"] == {"enable_thinking": False}
+    assert "input" not in body and "text" not in body and "store" not in body
+
+
+def test_default_transport_is_unchanged_responses_shape() -> None:
+    provider = OpenAICompatibleProvider(
+        LLMConfig(model="gpt-5.6-luna", max_tokens=64, structured_output=True)
+    )
+    body = provider._request_body([ChatMessage("user", "x")])
+    assert provider.endpoint.endswith("/responses")
+    assert "input" in body and "response_format" not in body
+
+
+def test_chat_usage_maps_to_responses_names() -> None:
+    usage = {"prompt_tokens": 12, "completion_tokens": 3, "total_tokens": 15}
+    assert _chat_usage_as_responses(usage) == {
+        "input_tokens": 12,
+        "output_tokens": 3,
+        "total_tokens": 15,
+    }
+
+
+def test_chat_text_extraction_and_think_block_strip() -> None:
+    content = '<think>\nplan\n</think>\n{"verb":3}'
+    data = {"choices": [{"finish_reason": "stop", "message": {"content": content}}]}
+    assert _extract_chat_text(data) == '{"verb":3}'
+
+
+def test_chat_length_finish_raises_incomplete_with_tail() -> None:
+    partial = "{" + "x" * 300
+    data = {"choices": [{"finish_reason": "length", "message": {"content": partial}}]}
+    with pytest.raises(ProviderError) as info:
+        _extract_chat_text(data)
+    message = str(info.value)
+    assert "max_output_tokens" in message
+    assert "301 chars" in message
+    assert partial[:20] in message and partial[-20:] in message