diff --git a/src/openchronicle/config.py b/src/openchronicle/config.py
index 21794c89..2fcf7ef8 100644
--- a/src/openchronicle/config.py
+++ b/src/openchronicle/config.py
@@ -18,6 +18,14 @@ class ModelConfig:
     api_key: str = ""
     api_key_env: str = "OPENAI_API_KEY"
     max_tokens: int | None = None
+    # Per-call timeout in seconds. ``None`` falls back to the writer
+    # default (writer.llm.DEFAULT_TIMEOUT_SECONDS) so a hung connection
+    # can never block a reducer / classifier thread indefinitely.
+    timeout_seconds: int | None = None
+    # Number of automatic retries on transient failures (429s, network
+    # blips). ``None`` falls back to writer.llm.DEFAULT_NUM_RETRIES.
+    # litellm honors Retry-After headers between attempts.
+    num_retries: int | None = None
 
 
 @dataclass
@@ -207,6 +215,8 @@ def load(path: Path | None = None) -> Config:
 api_key_env = "OPENAI_API_KEY"
 # base_url = ""
 # api_key = ""          # overrides api_key_env if set
+# timeout_seconds = 120 # per-call timeout; raise this if you run a slow local model
+# num_retries = 2       # automatic retries on transient errors (429, network blips)
 
 [models.compact]
 # Accuracy-sensitive — match or exceed the default.
diff --git a/src/openchronicle/writer/llm.py b/src/openchronicle/writer/llm.py
index 6e4317bd..ed87885e 100644
--- a/src/openchronicle/writer/llm.py
+++ b/src/openchronicle/writer/llm.py
@@ -11,6 +11,20 @@
 
 logger = get("openchronicle.writer")
 
+# litellm.completion() defaults to no client-side timeout and zero
+# retries. Without these, a stuck connection (slow provider, partial
+# response, dead TCP socket) blocks the reducer / classifier daemon
+# thread *forever* — the daemon stays "alive" but no durable facts
+# get written.  Two minutes is generous for a long session reduce
+# without being absurdly long; a slow local model can override per
+# stage via ``[models.<stage>] timeout_seconds = N``.
+DEFAULT_TIMEOUT_SECONDS = 120
+# Two retries gives 3 total attempts. litellm uses tenacity-style
+# backoff between attempts and respects Retry-After headers from
+# providers, so this is enough to absorb a brief 429 / 502 blip
+# without piling on extra latency.
+DEFAULT_NUM_RETRIES = 2
+
 
 def call_llm(
     cfg: Config,
@@ -30,9 +44,21 @@ def call_llm(
     import litellm  # imported lazily to keep CLI startup fast
 
     model_cfg = cfg.model_for(stage)
+    timeout = (
+        model_cfg.timeout_seconds
+        if model_cfg.timeout_seconds is not None
+        else DEFAULT_TIMEOUT_SECONDS
+    )
+    num_retries = (
+        model_cfg.num_retries
+        if model_cfg.num_retries is not None
+        else DEFAULT_NUM_RETRIES
+    )
     kwargs: dict[str, Any] = {
         "model": model_cfg.model,
         "messages": messages,
+        "timeout": timeout,
+        "num_retries": num_retries,
     }
     if model_cfg.base_url:
         kwargs["api_base"] = model_cfg.base_url
@@ -47,7 +73,10 @@ def call_llm(
     if model_cfg.max_tokens:
         kwargs["max_tokens"] = model_cfg.max_tokens
 
-    logger.debug("llm call stage=%s model=%s", stage, model_cfg.model)
+    logger.debug(
+        "llm call stage=%s model=%s timeout=%ds retries=%d",
+        stage, model_cfg.model, timeout, num_retries,
+    )
     return litellm.completion(**kwargs)
 
 
diff --git a/tests/test_config.py b/tests/test_config.py
index f5e742db..418e017e 100644
--- a/tests/test_config.py
+++ b/tests/test_config.py
@@ -49,3 +49,54 @@ def test_api_key_precedence(tmp_path: Path, monkeypatch) -> None:
     assert config.resolve_api_key(cfg) == "direct"
     cfg2 = config.ModelConfig(api_key="", api_key_env="ENV_KEY")
     assert config.resolve_api_key(cfg2) == "from-env"
+
+
+def test_timeout_and_retries_inherit_from_default(tmp_path: Path) -> None:
+    """[models.<stage>] without explicit timeout_seconds inherits the default."""
+    path = tmp_path / "config.toml"
+    path.write_text(
+        """
+[models.default]
+model = "gpt-5.4-nano"
+timeout_seconds = 200
+num_retries = 4
+
+[models.reducer]
+model = "claude-sonnet-4"
+"""
+    )
+    cfg = config.load(path)
+    reducer = cfg.model_for("reducer")
+    assert reducer.model == "claude-sonnet-4"
+    # timeout_seconds / num_retries come down from [models.default]
+    assert reducer.timeout_seconds == 200
+    assert reducer.num_retries == 4
+
+
+def test_timeout_per_stage_overrides_default(tmp_path: Path) -> None:
+    """A stage that explicitly sets timeout_seconds wins over the default."""
+    path = tmp_path / "config.toml"
+    path.write_text(
+        """
+[models.default]
+timeout_seconds = 120
+
+[models.reducer]
+timeout_seconds = 600
+"""
+    )
+    cfg = config.load(path)
+    assert cfg.model_for("default").timeout_seconds == 120
+    assert cfg.model_for("reducer").timeout_seconds == 600
+
+
+def test_missing_timeout_defaults_to_none(tmp_path: Path) -> None:
+    """An old config that doesn't mention timeout_seconds parses to None.
+
+    The llm.py wrapper interprets None as "use module default" so a
+    user upgrading from a config that pre-dates this field gets the
+    DEFAULT_TIMEOUT_SECONDS automatically — no breakage.
+    """
+    cfg = config.load(tmp_path / "missing.toml")
+    assert cfg.model_for("reducer").timeout_seconds is None
+    assert cfg.model_for("reducer").num_retries is None
diff --git a/tests/test_writer_llm.py b/tests/test_writer_llm.py
new file mode 100644
index 00000000..a4a46431
--- /dev/null
+++ b/tests/test_writer_llm.py
@@ -0,0 +1,107 @@
+"""Tests for the litellm wrapper: timeout / retries / per-stage override."""
+
+from __future__ import annotations
+
+from typing import Any
+from unittest.mock import MagicMock, patch
+
+from openchronicle.config import Config, ModelConfig
+from openchronicle.writer import llm as llm_mod
+
+
+def _cfg_with_models(models: dict[str, ModelConfig]) -> Config:
+    """Build a Config whose ``model_for(stage)`` returns the given mapping.
+
+    The real ``Config.model_for`` falls back from stage → 'default' →
+    bare ``ModelConfig()``, so we make sure 'default' exists.
+    """
+    if "default" not in models:
+        models["default"] = ModelConfig()
+    cfg = Config()
+    cfg.models = models
+    return cfg
+
+
+def _stub_completion_response() -> Any:
+    """Return-shape doesn't matter for these tests — they only inspect kwargs."""
+    msg = MagicMock()
+    msg.content = ""
+    msg.tool_calls = None
+    choice = MagicMock()
+    choice.message = msg
+    choice.finish_reason = "stop"
+    resp = MagicMock()
+    resp.choices = [choice]
+    return resp
+
+
+def test_call_llm_passes_default_timeout_and_retries(monkeypatch) -> None:
+    """Stage with no explicit override → module defaults are forwarded to litellm."""
+    monkeypatch.delenv("OPENCHRONICLE_LLM_MOCK", raising=False)
+    cfg = _cfg_with_models({})
+
+    with patch("litellm.completion", return_value=_stub_completion_response()) as mock:
+        llm_mod.call_llm(
+            cfg, "reducer",
+            messages=[{"role": "user", "content": "hi"}],
+        )
+        kwargs = mock.call_args.kwargs
+
+    assert kwargs["timeout"] == llm_mod.DEFAULT_TIMEOUT_SECONDS
+    assert kwargs["num_retries"] == llm_mod.DEFAULT_NUM_RETRIES
+
+
+def test_call_llm_per_stage_timeout_override(monkeypatch) -> None:
+    """A stage-level timeout_seconds in the config overrides the default."""
+    monkeypatch.delenv("OPENCHRONICLE_LLM_MOCK", raising=False)
+    cfg = _cfg_with_models(
+        {
+            "reducer": ModelConfig(timeout_seconds=300, num_retries=5),
+        }
+    )
+
+    with patch("litellm.completion", return_value=_stub_completion_response()) as mock:
+        llm_mod.call_llm(
+            cfg, "reducer",
+            messages=[{"role": "user", "content": "hi"}],
+        )
+        kwargs = mock.call_args.kwargs
+
+    assert kwargs["timeout"] == 300
+    assert kwargs["num_retries"] == 5
+
+
+def test_call_llm_other_stages_keep_defaults_when_one_is_overridden(monkeypatch) -> None:
+    """Setting timeout for ``reducer`` must not leak into ``classifier``."""
+    monkeypatch.delenv("OPENCHRONICLE_LLM_MOCK", raising=False)
+    cfg = _cfg_with_models(
+        {
+            "reducer": ModelConfig(timeout_seconds=300),
+            "classifier": ModelConfig(),  # no override
+        }
+    )
+
+    with patch("litellm.completion", return_value=_stub_completion_response()) as mock:
+        llm_mod.call_llm(
+            cfg, "classifier",
+            messages=[{"role": "user", "content": "hi"}],
+        )
+        kwargs = mock.call_args.kwargs
+
+    assert kwargs["timeout"] == llm_mod.DEFAULT_TIMEOUT_SECONDS
+
+
+def test_call_llm_mock_path_bypasses_litellm(monkeypatch) -> None:
+    """OPENCHRONICLE_LLM_MOCK=1 still short-circuits before importing litellm."""
+    monkeypatch.setenv("OPENCHRONICLE_LLM_MOCK", "1")
+    cfg = _cfg_with_models({})
+
+    with patch("litellm.completion") as mock:
+        resp = llm_mod.call_llm(
+            cfg, "reducer",
+            messages=[{"role": "user", "content": "hi"}],
+        )
+
+    assert mock.call_count == 0
+    # The mock response shape: choices[0].message.content is a string.
+    assert resp.choices[0].message.content