diff --git a/decnet/orchestrator/drivers/email.py b/decnet/orchestrator/drivers/email.py
index 2ccd1c48..547fd44a 100644
--- a/decnet/orchestrator/drivers/email.py
+++ b/decnet/orchestrator/drivers/email.py
@@ -1,27 +1,27 @@
-"""Email driver — Ollama-backed EML generation + decky-side delivery.
+"""Email driver — pluggable-LLM EML generation + decky-side delivery.
 
 One :class:`EmailAction` becomes one EML written into the mail decky's
 configured emailgen spool directory (``/var/spool/decnet-emails/`` by
-default).  An integration follow-up wires the IMAP/POP3 service templates
-to read EMLs from that spool at request time so attackers see the
-generated mail in their MUA.
+default).  The IMAP/POP3 service templates read that spool at request
+time so attackers see the generated mail in their MUA.
 
-The Ollama call shells out via ``ollama run <model>`` — the prototype at
-``DECNET-EMAILs/main.py`` proved the round-trip works.  Output is
-parsed-and-repaired into a valid EML using :mod:`email.mime.*`; the
-worker then ``docker exec``\\s a ``tee`` to drop the file inside the
-target container.
+The LLM call goes through :mod:`decnet.orchestrator.emailgen.llm` —
+backend-agnostic by construction so swapping Ollama for the Anthropic
+API, vLLM, or llama.cpp is a config change, not a driver rewrite.
+Output is parsed-and-repaired into a valid EML using
+:mod:`email.mime.*`; the worker then ``docker exec``\\s a ``tee`` to
+drop the file inside the target container, followed by a
+``touch -d <Date>`` so the file's mtime matches the email's RFC 2822
+``Date:`` header.
 
 Per CLAUDE.md "no shell strings": every subprocess invocation uses an
-argv list, never ``shell=True``.  Ollama prompts and EML payloads are
-piped via ``stdin``, not interpolated into argv.
+argv list, never ``shell=True``.  EML payloads are piped via ``stdin``,
+not interpolated into argv.
 """
 from __future__ import annotations
 
 import asyncio
-import os
 import shlex
-import time
 from datetime import datetime, timezone
 from email.mime.text import MIMEText
 from email.utils import formatdate
@@ -29,6 +29,7 @@ from typing import Any, Optional
 
 from decnet.logging import get_logger
 from decnet.orchestrator.drivers.base import ActivityResult
+from decnet.orchestrator.emailgen.llm import LLMBackend, LLMTimeout, get_llm
 from decnet.orchestrator.emailgen.prompt import PromptInputs, build as build_prompt
 from decnet.orchestrator.emailgen.scheduler import EmailAction
 from decnet.orchestrator.emailgen.threads import new_message_id
@@ -36,12 +37,6 @@ from decnet.orchestrator.emailgen.threads import new_message_id
 log = get_logger("orchestrator.email")
 
 _DOCKER = "docker"
-_OLLAMA = "ollama"
-# Wall-clock cap for the LLM call.  Big enough for a 4070 running
-# llama3.1; small enough that a stuck Ollama server doesn't wedge the
-# emailgen tick.
-_DEFAULT_OLLAMA_TIMEOUT = float(os.environ.get("DECNET_EMAILGEN_TIMEOUT", "60"))
-_DEFAULT_MODEL = os.environ.get("DECNET_EMAILGEN_MODEL", "llama3.1")
 # docker-exec wall-clock cap for the per-EML write.
 _DOCKER_TIMEOUT = 8.0
 # Container suffix for the IMAP service on a mail decky.
@@ -156,31 +151,35 @@ def _build_eml(
 class EmailDriver:
     """Concrete driver for :class:`EmailAction`.
 
-    Stateless across calls — Ollama model + timeout are constructor
-    args, not per-call.  The driver does *not* know about the bus or
-    DB; it returns an :class:`ActivityResult` that the worker pipes
-    onward.
+    Stateless across calls — the LLM backend is constructed once at
+    init time (or injected for tests).  The driver itself does *not*
+    know about the bus or DB; it returns an :class:`ActivityResult`
+    that the worker pipes onward.
     """
 
     def __init__(
         self,
         *,
-        model: str = _DEFAULT_MODEL,
-        ollama_timeout: float = _DEFAULT_OLLAMA_TIMEOUT,
+        llm: Optional[LLMBackend] = None,
+        model: Optional[str] = None,
         spool_dir: str = _SPOOL_DIR,
     ) -> None:
-        self.model = model
-        self.ollama_timeout = ollama_timeout
+        # *llm* takes precedence so tests can inject a FakeBackend
+        # without env-var trickery.  *model* lets the worker honour
+        # ``--model`` from the CLI without each backend needing to know
+        # about CLI flags.
+        self._llm = llm if llm is not None else get_llm(model=model)
         self.spool_dir = spool_dir
 
+    @property
+    def model(self) -> str:
+        """Convenience accessor for telemetry / logging."""
+        return self._llm.model
+
     async def run(self, action: EmailAction) -> ActivityResult:
-        # Look up the mail-decky container name + services.  The driver
-        # receives a denormalised view via the action — the worker
-        # populates it from the same list the scheduler used.
         return await self._run_email(action)
 
     async def _run_email(self, action: EmailAction) -> ActivityResult:
-        t0 = time.monotonic()
         prompt, mannerisms_used = build_prompt(
             PromptInputs(
                 sender=action.sender,
@@ -190,30 +189,41 @@ class EmailDriver:
                 parent_excerpt=action.parent_excerpt,
             )
         )
-        rc, stdout, stderr = await _run_capture(
-            [_OLLAMA, "run", self.model],
-            stdin_data=prompt.encode("utf-8"),
-            timeout=self.ollama_timeout,
-        )
-        gen_ms = int((time.monotonic() - t0) * 1000)
-        if rc != 0 or not stdout.strip():
-            log.warning(
-                "emailgen ollama failed rc=%d stderr=%r model=%s",
-                rc, stderr[:200], self.model,
-            )
+        try:
+            llm_result = await self._llm.generate(prompt)
+        except LLMTimeout as exc:
+            log.warning("emailgen llm timeout model=%s: %s", self._llm.model, exc)
             return ActivityResult(
                 success=False,
                 payload={
-                    "stage": "ollama",
-                    "rc": rc,
-                    "stderr": stderr.strip()[:256],
-                    "generation_ms": gen_ms,
-                    "model": self.model,
+                    "stage": "llm",
+                    "error": "timeout",
+                    "model": self._llm.model,
                     "thread_id": action.thread_id,
                 },
             )
 
-        subject, body = _parse_subject_and_body(stdout)
+        gen_ms = llm_result.latency_ms
+        if not llm_result.success or not llm_result.text.strip():
+            log.warning(
+                "emailgen llm produced no usable output model=%s extra=%r",
+                self._llm.model, llm_result.extra,
+            )
+            return ActivityResult(
+                success=False,
+                payload={
+                    "stage": "llm",
+                    "model": self._llm.model,
+                    "generation_ms": gen_ms,
+                    "thread_id": action.thread_id,
+                    **{
+                        k: v for k, v in llm_result.extra.items()
+                        if k in ("rc", "stderr")
+                    },
+                },
+            )
+
+        subject, body = _parse_subject_and_body(llm_result.text)
         message_id = new_message_id(action.sender.email.split("@", 1)[1])
         ts = datetime.now(timezone.utc)
         eml_bytes = _build_eml(
diff --git a/decnet/orchestrator/emailgen/llm/__init__.py b/decnet/orchestrator/emailgen/llm/__init__.py
new file mode 100644
index 00000000..4ce31dcc
--- /dev/null
+++ b/decnet/orchestrator/emailgen/llm/__init__.py
@@ -0,0 +1,22 @@
+"""LLM backend for emailgen.
+
+Pluggable from day one (per the provider-subpackages convention used by
+:mod:`decnet.web.db` and :mod:`decnet.bus`): the worker only depends on
+:class:`LLMBackend` from :mod:`base`; concrete transports live under
+:mod:`impl` and are selected by :func:`get_llm`.
+
+This is the seam ANTI will pull on when swapping local Ollama for the
+Anthropic API, llama.cpp, vLLM, or any other inference server — change
+``DECNET_EMAILGEN_LLM`` (or pass ``llm=`` to the driver), no driver
+rewrite.
+"""
+from __future__ import annotations
+
+from decnet.orchestrator.emailgen.llm.base import (
+    LLMBackend,
+    LLMResult,
+    LLMTimeout,
+)
+from decnet.orchestrator.emailgen.llm.factory import get_llm
+
+__all__ = ["LLMBackend", "LLMResult", "LLMTimeout", "get_llm"]
diff --git a/decnet/orchestrator/emailgen/llm/base.py b/decnet/orchestrator/emailgen/llm/base.py
new file mode 100644
index 00000000..73eb7ca5
--- /dev/null
+++ b/decnet/orchestrator/emailgen/llm/base.py
@@ -0,0 +1,47 @@
+"""Backend protocol shared by every LLM transport.
+
+Deliberately narrow: emailgen needs one async ``generate`` call that
+takes a prompt string and returns the model's output text plus enough
+metadata for the worker to populate the orchestrator-email payload
+(model name, latency, success bit).  Streaming, embeddings, multi-turn
+chat — all out of scope here; emailgen only ever does one-shot
+single-prompt generations.
+"""
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any, Protocol
+
+
+class LLMTimeout(Exception):
+    """Raised when a generation exceeds the backend's wall-clock cap.
+
+    Backends MUST raise this rather than returning silently empty
+    output; the driver discriminates timeout from "model produced
+    nothing useful" so payloads carry the right ``stage`` value.
+    """
+
+
+@dataclass
+class LLMResult:
+    """Outcome of one ``generate`` call.
+
+    ``success`` is ``False`` when the backend ran cleanly but produced
+    no usable output (e.g. an empty stdout).  Hard failures (subprocess
+    crash, network error) raise; soft failures land here so the driver
+    can persist + log them as one event.
+    """
+    success: bool
+    text: str
+    model: str
+    latency_ms: int
+    extra: dict[str, Any] = field(default_factory=dict)
+
+
+class LLMBackend(Protocol):
+    """Minimal contract for an emailgen LLM provider."""
+
+    model: str
+    timeout: float
+
+    async def generate(self, prompt: str) -> LLMResult: ...
diff --git a/decnet/orchestrator/emailgen/llm/factory.py b/decnet/orchestrator/emailgen/llm/factory.py
new file mode 100644
index 00000000..71bf4edb
--- /dev/null
+++ b/decnet/orchestrator/emailgen/llm/factory.py
@@ -0,0 +1,46 @@
+"""Backend dispatch.
+
+Reads ``DECNET_EMAILGEN_LLM`` to pick a concrete :class:`LLMBackend`.
+Defaults to ``ollama`` because that's what the prototype proved out and
+what most dev boxes have on hand.
+
+Supported keys:
+
+* ``ollama`` — :class:`decnet.orchestrator.emailgen.llm.impl.ollama.OllamaBackend`
+* ``fake``   — :class:`decnet.orchestrator.emailgen.llm.impl.fake.FakeBackend`
+  (canned output, used by tests so they don't shell out)
+
+Anthropic / vLLM / llama.cpp slots in here as a third branch when the
+need shows up.  Per the provider-subpackages memory, do NOT collapse
+factory dispatch into the impl modules — keeps the ``__init__`` import
+graph cycle-free and the env contract auditable in one place.
+"""
+from __future__ import annotations
+
+import os
+from typing import Any
+
+from decnet.orchestrator.emailgen.llm.base import LLMBackend
+
+
+def get_llm(*, model: str | None = None, **kwargs: Any) -> LLMBackend:
+    """Instantiate the LLM backend selected by environment.
+
+    *model* (when provided) overrides whatever the backend's own default
+    is — e.g. for OllamaBackend that's ``llama3.1`` unless
+    ``DECNET_EMAILGEN_MODEL`` says otherwise.  Lets the worker honour
+    ``decnet emailgen run --model gpt-oss`` without each backend having
+    to know about CLI flags.
+    """
+    backend_key = os.environ.get("DECNET_EMAILGEN_LLM", "ollama").lower()
+
+    if backend_key == "ollama":
+        from decnet.orchestrator.emailgen.llm.impl.ollama import OllamaBackend
+        return OllamaBackend(model=model, **kwargs)
+    if backend_key == "fake":
+        from decnet.orchestrator.emailgen.llm.impl.fake import FakeBackend
+        return FakeBackend(model=model or "fake-model", **kwargs)
+    raise ValueError(
+        f"Unsupported DECNET_EMAILGEN_LLM={backend_key!r}; "
+        "expected one of: ollama, fake"
+    )
diff --git a/decnet/orchestrator/emailgen/llm/impl/__init__.py b/decnet/orchestrator/emailgen/llm/impl/__init__.py
new file mode 100644
index 00000000..33d12197
--- /dev/null
+++ b/decnet/orchestrator/emailgen/llm/impl/__init__.py
@@ -0,0 +1,6 @@
+"""Concrete LLM-backend implementations.
+
+Importers go through :func:`decnet.orchestrator.emailgen.llm.get_llm`,
+not these modules directly — same convention as
+:mod:`decnet.web.db.sqlite` and :mod:`decnet.bus.unix_client`.
+"""
diff --git a/decnet/orchestrator/emailgen/llm/impl/fake.py b/decnet/orchestrator/emailgen/llm/impl/fake.py
new file mode 100644
index 00000000..4b249032
--- /dev/null
+++ b/decnet/orchestrator/emailgen/llm/impl/fake.py
@@ -0,0 +1,50 @@
+"""In-process fake backend for tests.
+
+Returns a canned ``Subject:\\n\\nbody`` string so the driver path can be
+exercised without an Ollama install.  Configurable via ``DECNET_EMAILGEN_FAKE_OUTPUT``
+(env) or the ``output`` constructor arg — the env-var path lets
+integration tests run the worker end-to-end with deterministic output.
+"""
+from __future__ import annotations
+
+import os
+import time
+from typing import Optional
+
+from decnet.orchestrator.emailgen.llm.base import LLMBackend, LLMResult
+
+
+_DEFAULT_OUTPUT = (
+    "Subject: Quick update\n\n"
+    "Hi,\n\nFollowing up on the topic.\n\nBest regards,\nFake Persona\n"
+)
+
+
+class FakeBackend(LLMBackend):
+    def __init__(
+        self,
+        *,
+        model: str = "fake-model",
+        timeout: float = 1.0,
+        output: Optional[str] = None,
+        success: bool = True,
+    ) -> None:
+        self.model = model
+        self.timeout = timeout
+        self._output = (
+            output
+            if output is not None
+            else os.environ.get("DECNET_EMAILGEN_FAKE_OUTPUT", _DEFAULT_OUTPUT)
+        )
+        self._success = success
+
+    async def generate(self, prompt: str) -> LLMResult:    # noqa: ARG002
+        t0 = time.monotonic()
+        latency_ms = int((time.monotonic() - t0) * 1000)
+        return LLMResult(
+            success=self._success,
+            text=self._output if self._success else "",
+            model=self.model,
+            latency_ms=latency_ms,
+            extra={"rc": 0 if self._success else 1},
+        )
diff --git a/decnet/orchestrator/emailgen/llm/impl/ollama.py b/decnet/orchestrator/emailgen/llm/impl/ollama.py
new file mode 100644
index 00000000..85bdb9e5
--- /dev/null
+++ b/decnet/orchestrator/emailgen/llm/impl/ollama.py
@@ -0,0 +1,107 @@
+"""Ollama subprocess backend.
+
+Shells out to ``ollama run <model>`` with the prompt fed via stdin.
+Mirrors what the original prototype at ``DECNET-EMAILs/main.py`` did,
+but lifted out of the driver so the rest of emailgen never imports a
+specific transport.
+
+Why subprocess and not the Ollama HTTP API:
+* No new dependency (``ollama`` Python lib is optional).
+* Works on hosts where Ollama is bound to a unix socket, an unusual TCP
+  port, or behind a remote-mount layer — `ollama run` resolves all that.
+* Same path the operator uses by hand (``ollama run llama3.1``); easier
+  to debug discrepancies between worker output and a console session.
+
+Cost: per-call process spawn (~50ms on a warm box).  Acceptable for
+emailgen's tick rate (one email every 5 minutes by default).  When that
+cost matters, swap to an HTTP-API backend; the seam is in
+:mod:`decnet.orchestrator.emailgen.llm.factory`.
+"""
+from __future__ import annotations
+
+import asyncio
+import os
+import time
+from typing import Optional
+
+from decnet.logging import get_logger
+from decnet.orchestrator.emailgen.llm.base import (
+    LLMBackend,
+    LLMResult,
+    LLMTimeout,
+)
+
+log = get_logger("orchestrator.emailgen.llm")
+
+_OLLAMA = "ollama"
+_DEFAULT_MODEL = os.environ.get("DECNET_EMAILGEN_MODEL", "llama3.1")
+_DEFAULT_TIMEOUT = float(os.environ.get("DECNET_EMAILGEN_TIMEOUT", "60"))
+
+
+class OllamaBackend(LLMBackend):
+    """Concrete :class:`LLMBackend` that shells out to ``ollama run``."""
+
+    def __init__(
+        self,
+        *,
+        model: Optional[str] = None,
+        timeout: Optional[float] = None,
+    ) -> None:
+        self.model = model or _DEFAULT_MODEL
+        self.timeout = timeout if timeout is not None else _DEFAULT_TIMEOUT
+
+    async def generate(self, prompt: str) -> LLMResult:
+        t0 = time.monotonic()
+        try:
+            proc = await asyncio.create_subprocess_exec(
+                _OLLAMA, "run", self.model,
+                stdin=asyncio.subprocess.PIPE,
+                stdout=asyncio.subprocess.PIPE,
+                stderr=asyncio.subprocess.PIPE,
+            )
+        except FileNotFoundError as exc:
+            latency_ms = int((time.monotonic() - t0) * 1000)
+            return LLMResult(
+                success=False,
+                text="",
+                model=self.model,
+                latency_ms=latency_ms,
+                extra={"rc": 127, "stderr": f"argv[0] not found: {exc}"},
+            )
+        try:
+            stdout, stderr = await asyncio.wait_for(
+                proc.communicate(prompt.encode("utf-8")),
+                timeout=self.timeout,
+            )
+        except asyncio.TimeoutError as exc:
+            try:
+                proc.kill()
+            except ProcessLookupError:
+                pass
+            raise LLMTimeout(
+                f"ollama run {self.model} exceeded {self.timeout}s"
+            ) from exc
+
+        latency_ms = int((time.monotonic() - t0) * 1000)
+        rc = proc.returncode if proc.returncode is not None else -1
+        text = stdout.decode("utf-8", "replace")
+        stderr_s = stderr.decode("utf-8", "replace")
+        if rc != 0 or not text.strip():
+            log.warning(
+                "ollama backend non-zero / empty rc=%d model=%s stderr=%r",
+                rc, self.model, stderr_s[:200],
+            )
+            return LLMResult(
+                success=False,
+                text=text,
+                model=self.model,
+                latency_ms=latency_ms,
+                extra={"rc": rc, "stderr": stderr_s.strip()[:256]},
+            )
+        return LLMResult(
+            success=True,
+            text=text,
+            model=self.model,
+            latency_ms=latency_ms,
+            extra={"rc": rc},
+        )
diff --git a/tests/orchestrator/emailgen/test_driver.py b/tests/orchestrator/emailgen/test_driver.py
index 88b8d913..7737f1ac 100644
--- a/tests/orchestrator/emailgen/test_driver.py
+++ b/tests/orchestrator/emailgen/test_driver.py
@@ -1,14 +1,40 @@
-"""EmailDriver: stub the Ollama subprocess + docker exec; verify EML
-parse-and-repair and payload metadata."""
+"""EmailDriver: inject a fake LLM backend + stub docker exec; verify
+EML parse-and-repair and payload metadata."""
 from __future__ import annotations
 
 import pytest
 
 from decnet.orchestrator.drivers import email as email_driver
+from decnet.orchestrator.emailgen.llm.base import LLMResult, LLMTimeout
+from decnet.orchestrator.emailgen.llm.impl.fake import FakeBackend
 from decnet.orchestrator.emailgen.personas import EmailPersona
 from decnet.orchestrator.emailgen.scheduler import EmailAction
 
 
+class _RaisingBackend:
+    """Async stub that raises LLMTimeout on every call."""
+    model = "stuck-model"
+    timeout = 0.1
+
+    async def generate(self, prompt: str) -> LLMResult:    # noqa: ARG002
+        raise LLMTimeout("stuck")
+
+
+class _FailingBackend:
+    """Async stub that returns success=False."""
+    model = "broken-model"
+    timeout = 1.0
+
+    async def generate(self, prompt: str) -> LLMResult:    # noqa: ARG002
+        return LLMResult(
+            success=False,
+            text="",
+            model=self.model,
+            latency_ms=5,
+            extra={"rc": 1, "stderr": "model not found"},
+        )
+
+
 def _persona(name="John", email="john@corp.com"):
     return EmailPersona(
         name=name,
@@ -110,19 +136,20 @@ def test_container_for_pop3_only():
 
 @pytest.mark.asyncio
 async def test_driver_run_success_path(monkeypatch):
-    """Stub both subprocess calls (ollama + docker exec) as success."""
-    calls: list[list[str]] = []
+    """Inject a FakeBackend + stub docker exec; success end-to-end."""
+    docker_calls: list[list[str]] = []
 
     async def fake_run_capture(argv, *, stdin_data=None, timeout=8.0):
-        calls.append(list(argv))
-        if argv[0] == "ollama":
-            return 0, "Subject: Q3 budget\n\nHi Sarah,\nNumbers attached.\n", ""
-        # docker exec
+        docker_calls.append(list(argv))
         return 0, "", ""
 
     monkeypatch.setattr(email_driver, "_run_capture", fake_run_capture)
 
-    drv = email_driver.EmailDriver(model="llama3.1", ollama_timeout=1.0)
+    llm = FakeBackend(
+        model="llama3.1",
+        output="Subject: Q3 budget\n\nHi Sarah,\nNumbers attached.\n",
+    )
+    drv = email_driver.EmailDriver(llm=llm)
     result = await drv.run(_action())
     assert result.success is True
     assert result.payload["model"] == "llama3.1"
@@ -132,46 +159,56 @@ async def test_driver_run_success_path(monkeypatch):
     assert result.payload["message_id"].startswith("<")
     assert result.payload["eml_path"].endswith(".eml")
     assert result.payload["container"] == "mailhost-imap"
-    # Two subprocess calls: ollama, then docker exec.
-    assert calls[0][0] == "ollama"
-    assert calls[1][0] == "docker"
-    # docker exec shell command must include `touch -d` so the file's
-    # mtime matches the EML's Date: header — otherwise the spool's
-    # `ls -lt` clusters every email inside the worker tick window.
-    docker_sh = calls[1][-1]
+    # Only docker exec is shelled out now — the LLM call is in-process
+    # via the FakeBackend.
+    assert len(docker_calls) == 1
+    assert docker_calls[0][0] == "docker"
+    docker_sh = docker_calls[0][-1]
     assert "touch -d" in docker_sh
     assert "tee" in docker_sh
-    # And tee must come before touch so we don't touch a file that
-    # doesn't exist yet.
     assert docker_sh.index("tee") < docker_sh.index("touch -d")
 
 
 @pytest.mark.asyncio
-async def test_driver_run_ollama_failure_short_circuits(monkeypatch):
+async def test_driver_run_llm_failure_short_circuits(monkeypatch):
+    """When the backend reports success=False, no docker exec should fire."""
+    docker_called = False
+
     async def fake_run_capture(argv, *, stdin_data=None, timeout=8.0):
-        if argv[0] == "ollama":
-            return 1, "", "ollama: model not found"
+        nonlocal docker_called
+        docker_called = True
         return 0, "", ""
 
     monkeypatch.setattr(email_driver, "_run_capture", fake_run_capture)
 
-    drv = email_driver.EmailDriver()
+    drv = email_driver.EmailDriver(llm=_FailingBackend())
     result = await drv.run(_action())
     assert result.success is False
-    assert result.payload["stage"] == "ollama"
+    assert result.payload["stage"] == "llm"
+    assert "stderr" in result.payload
     assert "model not found" in result.payload["stderr"]
+    assert docker_called is False
+
+
+@pytest.mark.asyncio
+async def test_driver_run_llm_timeout_reported_distinctly(monkeypatch):
+    drv = email_driver.EmailDriver(llm=_RaisingBackend())
+    result = await drv.run(_action())
+    assert result.success is False
+    assert result.payload["stage"] == "llm"
+    assert result.payload["error"] == "timeout"
 
 
 @pytest.mark.asyncio
 async def test_driver_run_delivery_failure(monkeypatch):
     async def fake_run_capture(argv, *, stdin_data=None, timeout=8.0):
-        if argv[0] == "ollama":
-            return 0, "Subject: hi\n\nbody\n", ""
         return 1, "", "no such container"
 
     monkeypatch.setattr(email_driver, "_run_capture", fake_run_capture)
 
-    drv = email_driver.EmailDriver()
+    drv = email_driver.EmailDriver(
+        llm=FakeBackend(output="Subject: hi\n\nbody\n"),
+    )
     result = await drv.run(_action())
     assert result.success is False
     assert result.payload["stage"] == "delivery"
diff --git a/tests/orchestrator/emailgen/test_llm.py b/tests/orchestrator/emailgen/test_llm.py
new file mode 100644
index 00000000..6f600b50
--- /dev/null
+++ b/tests/orchestrator/emailgen/test_llm.py
@@ -0,0 +1,137 @@
+"""LLM backend factory + Ollama implementation."""
+from __future__ import annotations
+
+import asyncio
+
+import pytest
+
+from decnet.orchestrator.emailgen.llm import LLMTimeout, get_llm
+from decnet.orchestrator.emailgen.llm.impl.fake import FakeBackend
+from decnet.orchestrator.emailgen.llm.impl.ollama import OllamaBackend
+
+
+# ── factory dispatch ─────────────────────────────────────────────────────────
+
+
+def test_factory_default_is_ollama(monkeypatch):
+    monkeypatch.delenv("DECNET_EMAILGEN_LLM", raising=False)
+    backend = get_llm()
+    assert isinstance(backend, OllamaBackend)
+
+
+def test_factory_selects_fake(monkeypatch):
+    monkeypatch.setenv("DECNET_EMAILGEN_LLM", "fake")
+    backend = get_llm()
+    assert isinstance(backend, FakeBackend)
+
+
+def test_factory_unknown_raises(monkeypatch):
+    monkeypatch.setenv("DECNET_EMAILGEN_LLM", "vllm-someday")
+    with pytest.raises(ValueError, match="Unsupported"):
+        get_llm()
+
+
+def test_factory_passes_model_through(monkeypatch):
+    monkeypatch.setenv("DECNET_EMAILGEN_LLM", "ollama")
+    backend = get_llm(model="qwen2:7b")
+    assert backend.model == "qwen2:7b"
+
+
+# ── FakeBackend ──────────────────────────────────────────────────────────────
+
+
+@pytest.mark.asyncio
+async def test_fake_backend_returns_canned_output():
+    fb = FakeBackend(output="Subject: hi\n\nbody")
+    result = await fb.generate("any prompt")
+    assert result.success is True
+    assert result.text.startswith("Subject:")
+    assert result.model == "fake-model"
+
+
+@pytest.mark.asyncio
+async def test_fake_backend_can_simulate_failure():
+    fb = FakeBackend(success=False)
+    result = await fb.generate("prompt")
+    assert result.success is False
+    assert result.text == ""
+
+
+# ── OllamaBackend (subprocess stubbed) ───────────────────────────────────────
+
+
+@pytest.mark.asyncio
+async def test_ollama_backend_success(monkeypatch):
+    """Stub asyncio.create_subprocess_exec to return canned stdout."""
+
+    class _StubProc:
+        returncode = 0
+
+        async def communicate(self, _stdin):
+            return b"Subject: hi\n\nbody\n", b""
+
+    async def fake_create(*args, **kwargs):    # noqa: ARG001
+        return _StubProc()
+
+    monkeypatch.setattr(asyncio, "create_subprocess_exec", fake_create)
+
+    backend = OllamaBackend(model="m1", timeout=1.0)
+    result = await backend.generate("hello")
+    assert result.success is True
+    assert "Subject:" in result.text
+    assert result.model == "m1"
+
+
+@pytest.mark.asyncio
+async def test_ollama_backend_non_zero_rc_marks_failure(monkeypatch):
+    class _StubProc:
+        returncode = 1
+
+        async def communicate(self, _stdin):
+            return b"", b"model not found"
+
+    async def fake_create(*args, **kwargs):    # noqa: ARG001
+        return _StubProc()
+
+    monkeypatch.setattr(asyncio, "create_subprocess_exec", fake_create)
+
+    backend = OllamaBackend(model="m1", timeout=1.0)
+    result = await backend.generate("hello")
+    assert result.success is False
+    assert result.extra["rc"] == 1
+    assert "model not found" in result.extra["stderr"]
+
+
+@pytest.mark.asyncio
+async def test_ollama_backend_timeout_raises(monkeypatch):
+    class _StubProc:
+        returncode = None
+
+        async def communicate(self, _stdin):
+            await asyncio.sleep(10)    # well past the timeout
+            return b"", b""
+
+        def kill(self):
+            pass
+
+    async def fake_create(*args, **kwargs):    # noqa: ARG001
+        return _StubProc()
+
+    monkeypatch.setattr(asyncio, "create_subprocess_exec", fake_create)
+
+    backend = OllamaBackend(model="m1", timeout=0.05)
+    with pytest.raises(LLMTimeout):
+        await backend.generate("hello")
+
+
+@pytest.mark.asyncio
+async def test_ollama_backend_missing_binary_returns_failure(monkeypatch):
+    async def fake_create(*args, **kwargs):    # noqa: ARG001
+        raise FileNotFoundError("ollama: not found")
+
+    monkeypatch.setattr(asyncio, "create_subprocess_exec", fake_create)
+
+    backend = OllamaBackend(model="m1", timeout=1.0)
+    result = await backend.generate("hello")
+    assert result.success is False
+    assert result.extra["rc"] == 127
diff --git a/tests/orchestrator/emailgen/test_worker_integration.py b/tests/orchestrator/emailgen/test_worker_integration.py
index 4aef25ba..eb03f003 100644
--- a/tests/orchestrator/emailgen/test_worker_integration.py
+++ b/tests/orchestrator/emailgen/test_worker_integration.py
@@ -10,6 +10,7 @@ import pytest_asyncio
 from decnet.bus.fake import FakeBus
 from decnet.orchestrator.drivers import email as email_driver
 from decnet.orchestrator.emailgen import worker as eg_worker
+from decnet.orchestrator.emailgen.llm.impl.fake import FakeBackend
 from decnet.orchestrator.emailgen.scheduler import EmailAction  # noqa: F401
 from decnet.web.db.models import Topology, TopologyDecky
 from decnet.web.db.sqlite.repository import SQLiteRepository
@@ -82,9 +83,9 @@ async def _seed_mail_topology(repo: SQLiteRepository) -> str:
 async def test_one_tick_records_and_publishes(repo, fake_bus, monkeypatch):
     decky_uuid = await _seed_mail_topology(repo)
 
+    # Stub only the docker exec subprocess; the LLM call goes through
+    # an injected FakeBackend with deterministic output.
     async def fake_run_capture(argv, *, stdin_data=None, timeout=8.0):
-        if argv[0] == "ollama":
-            return 0, "Subject: Hi\n\nBody here.\n", ""
         return 0, "", ""
 
     monkeypatch.setattr(email_driver, "_run_capture", fake_run_capture)
@@ -101,7 +102,9 @@ async def test_one_tick_records_and_publishes(repo, fake_bus, monkeypatch):
     collector = asyncio.create_task(collect())
     await asyncio.sleep(0)
 
-    driver = email_driver.EmailDriver()
+    driver = email_driver.EmailDriver(
+        llm=FakeBackend(output="Subject: Hi\n\nBody here.\n"),
+    )
     await eg_worker._one_tick(repo, driver, fake_bus)
     await asyncio.wait_for(collector, timeout=2.0)
 
@@ -126,11 +129,13 @@ async def test_one_tick_noop_when_no_mail_decky(repo, fake_bus, monkeypatch):
     async def fake_run_capture(argv, *, stdin_data=None, timeout=8.0):
         nonlocal called
         called = True
-        return 0, "Subject: x\n\nb\n", ""
+        return 0, "", ""
 
     monkeypatch.setattr(email_driver, "_run_capture", fake_run_capture)
 
-    driver = email_driver.EmailDriver()
+    driver = email_driver.EmailDriver(
+        llm=FakeBackend(output="Subject: x\n\nb\n"),
+    )
     await eg_worker._one_tick(repo, driver, fake_bus)
     assert called is False
     assert await repo.list_orchestrator_emails() == []