diff --git a/pepper_wizard/cli.py b/pepper_wizard/cli.py
index 2c765cf..f005afc 100644
--- a/pepper_wizard/cli.py
+++ b/pepper_wizard/cli.py
@@ -782,6 +782,7 @@ def llm_talk_session(robot_client, config, verbose=False):
     from .stt_client import STTClient
     from .llm.client import LLMClient, LLMUnavailable
     from .llm.config_watcher import LLMConfigWatcher
+    from .llm.identity import build_config_snapshot
 
     logger = get_logger("LLMTalk")
     stt_config = config.stt_config
@@ -799,6 +800,7 @@ def llm_talk_session(robot_client, config, verbose=False):
         )
         return
 
+    logger.info("LLMConfigSnapshot", build_config_snapshot(watcher.current(), "session_start"))
     stt_client = STTClient(stt_config.get("zmq_address", "tcp://localhost:5562"))
     if not stt_client.ping():
         print_formatted_text(HTML(
@@ -963,20 +965,21 @@ def _announce_llm_reload(old_config: dict, new_config: dict, logger):
         return
     summary = "config reloaded: " + ", ".join(changed)
     print_formatted_text(HTML(f"<ansigray>{escape(summary)}</ansigray>"))
-    logger.info("LLMConfigReload", {
-        "changed": changed,
-        "old": old_config,
-        "new": new_config,
-    })
+    from .llm.identity import build_config_snapshot
+    logger.info(
+        "LLMConfigSnapshot",
+        build_config_snapshot(new_config, "reload", changed=changed),
+    )
 
 
 def _dispatch_to_llm(user_text, *, source, llm, stt, robot_client, logger):
     # Mute while Pepper is speaking so stt-service ignores self-hearing.
-    # TODO - for the person interrupting pepper this needs some neance. Outside the scope of current PR. 
+    # TODO - for the person interrupting pepper this needs some neance.
     stt.mute()
     logger.info("MuteStart", {})
     try:
-        reply = llm.reply(user_text)
+        result = llm.reply(user_text)
+        reply = result.text
     except Exception as e:
         print_formatted_text(
             HTML("<ansired>LLM error: {}</ansired>").format(str(e))
@@ -1005,4 +1008,10 @@ def _dispatch_to_llm(user_text, *, source, llm, stt, robot_client, logger):
         logger.error("TalkFailed", {"error": str(e)})
     stt.unmute()
     logger.info("MuteEnd", {})
-    logger.info("LLMTurn", {"user": user_text, "reply": reply, "source": source})
\ No newline at end of file
+    logger.info("LLMTurn", {
+        "user": user_text,
+        "reply": reply,
+        "source": source,
+        "config_hash": result.config_hash,
+        "config_name": result.config_name,
+    })
\ No newline at end of file
diff --git a/pepper_wizard/llm/client.py b/pepper_wizard/llm/client.py
index eceaa69..c0036a7 100644
--- a/pepper_wizard/llm/client.py
+++ b/pepper_wizard/llm/client.py
@@ -1,11 +1,16 @@
 import os
-from collections import deque
+import threading
+from collections import deque, namedtuple
+from .identity import resolve_config, config_fingerprint
 from typing import TYPE_CHECKING
 
 if TYPE_CHECKING:
     from .config_watcher import LLMConfigWatcher
 
 
+ReplyResult = namedtuple("ReplyResult", ["text", "config_hash", "config_name"])
+
+
 class LLMUnavailable(Exception):
     pass
 
@@ -20,11 +25,7 @@ class LLMClient:
 
     def __init__(self, watcher: "LLMConfigWatcher"):
         self._watcher = watcher
-
-        # Initialise with the default maxlen; reply() will resize on first call
-        # if the watcher's history_turns differs. We do not call watcher.current()
-        # here so that the watcher call-count seen by callers reflects only
-        # active turns, not construction overhead.
+        self._lock = threading.Lock()
         self._history = deque(maxlen=10 * 2)
 
         api_key = os.environ.get("ANTHROPIC_API_KEY")
@@ -48,33 +49,37 @@ def __init__(self, watcher: "LLMConfigWatcher"):
     def model(self) -> str:
         return self._watcher.current().get("model", "claude-haiku-4-5")
 
-    def reply(self, user_text: str) -> str:
-        """Send `user_text` with the rolling history and return the reply."""
-        config = self._watcher.current()
-
-        desired_maxlen = config.get("history_turns", 10) * 2
-        if self._history.maxlen != desired_maxlen:
-            self._history = deque(self._history, maxlen=desired_maxlen)
-
-        self._history.append({"role": "user", "content": user_text})
-
-        response = self._client.messages.create(
-            model=config.get("model", "claude-haiku-4-5"),
-            system=config.get(
-                "system_prompt",
-                "You are Pepper, a humanoid robot. Keep replies brief and conversational.",
-            ),
-            max_tokens=config.get("max_tokens", 256),
-            temperature=config.get("temperature", 0.7),
-            messages=list(self._history),
-        )
+    def reply(self, user_text: str) -> "ReplyResult":
+        """Send user_text with the rolling history; return a ReplyResult with the
+        reply text plus the identity (config_hash, config_name) of the config used
+        for THIS turn. A per-client lock serializes concurrent callers (e.g. the
+        auto-dispatch VAD thread and typed input) so attribution is captured atomically 
+        instead of read back from shared state."""
+        with self._lock:
+            raw = self._watcher.current()
+            resolved = resolve_config(raw)
+
+            desired_maxlen = resolved["history_turns"] * 2
+            if self._history.maxlen != desired_maxlen:
+                self._history = deque(self._history, maxlen=desired_maxlen)
+
+            self._history.append({"role": "user", "content": user_text})
+
+            response = self._client.messages.create(
+                model=resolved["model"],
+                system=resolved["system_prompt"],
+                max_tokens=resolved["max_tokens"],
+                temperature=resolved["temperature"],
+                messages=list(self._history),
+            )
 
-        reply_text = "".join(
-            block.text for block in response.content if block.type == "text"
-        ).strip()
+            reply_text = "".join(
+                block.text for block in response.content if block.type == "text"
+            ).strip()
 
-        self._history.append({"role": "assistant", "content": reply_text})
-        return reply_text
+            self._history.append({"role": "assistant", "content": reply_text})
+            return ReplyResult(reply_text, config_fingerprint(resolved), raw.get("name"))
 
     def reset(self):
-        self._history.clear()
+        with self._lock:
+            self._history.clear()
diff --git a/pepper_wizard/llm/identity.py b/pepper_wizard/llm/identity.py
new file mode 100644
index 0000000..4558abd
--- /dev/null
+++ b/pepper_wizard/llm/identity.py
@@ -0,0 +1,55 @@
+"""Source-agnostic identity for an LLM dialogue config.
+
+The identifier is a short content hash of the *resolved* config (the behavioural
+fields actually sent to the model), so a turn can be attributed to its config
+regardless of where the prompt came from: llm.json, a default, or a future
+orchestrator. `name` is a human label, kept out of the hash so renaming a persona
+does not change its behavioural fingerprint while editing a prompt does.
+"""
+
+import hashlib
+import json
+from typing import List, Optional
+
+_DEFAULTS = {
+    "model": "claude-haiku-4-5",
+    "system_prompt": "You are Pepper, a humanoid robot. Keep replies brief and conversational.",
+    "max_tokens": 256,
+    "temperature": 0.7,
+    "history_turns": 10,
+}
+
+# Fields that define model behaviour, and therefore the identity.
+_BEHAVIOURAL_KEYS = ("model", "system_prompt", "max_tokens", "temperature", "history_turns")
+
+
+def resolve_config(raw: dict) -> dict:
+    """Apply defaults, returning only the behavioural fields the model sees."""
+    return {key: raw.get(key, _DEFAULTS[key]) for key in _BEHAVIOURAL_KEYS}
+
+
+def config_fingerprint(config: dict) -> str:
+    """Short deterministic identity hash of a config.
+
+    Resolves defaults first, so the hash always reflects what the model
+    receives whether a raw or already-resolved dict is passed (resolve_config
+    is idempotent). This removes any raw-vs-resolved mis-attribution surface.
+    """
+    resolved = resolve_config(config)
+    canonical = json.dumps(resolved, sort_keys=True, separators=(",", ":"))
+    return hashlib.sha256(canonical.encode("utf-8")).hexdigest()[:12]
+
+
+def build_config_snapshot(raw: dict, reason: str, *, source: str = "file", changed: Optional[List[str]] = None) -> dict:
+    """Assemble an LLMConfigSnapshot data dict from a raw config dict."""
+    resolved = resolve_config(raw)
+    snapshot = {
+        "config_hash": config_fingerprint(resolved),
+        "config_name": raw.get("name"),
+        "config_source": source,
+        "reason": reason,
+        "config": resolved,
+    }
+    if changed is not None:
+        snapshot["changed"] = changed
+    return snapshot
diff --git a/tests/test_llm_client.py b/tests/test_llm_client.py
index 48df202..6e81a20 100644
--- a/tests/test_llm_client.py
+++ b/tests/test_llm_client.py
@@ -131,7 +131,7 @@ def test_reply_appends_user_and_assistant_turns(self):
         client = LLMClient(watcher)
         self._mock_messages.create.return_value = _make_anthropic_response("the reply")
         result = client.reply("the question")
-        self.assertEqual(result, "the reply")
+        self.assertEqual(result.text, "the reply")
         self.assertEqual(list(client._history), [
             {"role": "user", "content": "the question"},
             {"role": "assistant", "content": "the reply"},
@@ -159,6 +159,65 @@ def test_model_property_reflects_watcher(self):
         watcher.update(model="claude-sonnet-4-6")
         self.assertEqual(client.model, "claude-sonnet-4-6")
 
+    def test_reply_captures_config_hash_and_name(self):
+        from pepper_wizard.llm.client import LLMClient
+        from pepper_wizard.llm.identity import config_fingerprint, resolve_config
+
+        cfg = {
+            "name": "persona-a",
+            "model": "claude-haiku-4-5",
+            "system_prompt": "be brief",
+            "max_tokens": 50,
+            "temperature": 0.3,
+            "history_turns": 4,
+        }
+        watcher = FakeWatcher(cfg)
+        client = LLMClient(watcher)
+        result = client.reply("hi")
+        self.assertEqual(result.config_hash, config_fingerprint(resolve_config(cfg)))
+        self.assertEqual(result.config_name, "persona-a")
+
+    def test_config_name_none_when_absent(self):
+        from pepper_wizard.llm.client import LLMClient
+
+        watcher = FakeWatcher({"history_turns": 4})
+        client = LLMClient(watcher)
+        result = client.reply("hi")
+        self.assertIsNone(result.config_name)
+
+    def test_config_hash_changes_when_config_changes(self):
+        from pepper_wizard.llm.client import LLMClient
+
+        watcher = FakeWatcher({"system_prompt": "old", "history_turns": 4})
+        client = LLMClient(watcher)
+        first_hash = client.reply("first").config_hash
+        watcher.update(system_prompt="new")
+        second_hash = client.reply("second").config_hash
+        self.assertNotEqual(second_hash, first_hash)
+
+    def test_concurrent_replies_are_serialized(self):
+        import threading
+        import time
+        from pepper_wizard.llm.client import LLMClient
+
+        watcher = FakeWatcher({"history_turns": 100})
+        client = LLMClient(watcher)
+
+        def slow_create(*args, **kwargs):
+            time.sleep(0.005)  
+            return _make_anthropic_response("ok")
+
+        self._mock_messages.create.side_effect = slow_create
+        threads = [threading.Thread(target=client.reply, args=("msg-%d" % i,)) for i in range(6)]
+        for t in threads:
+            t.start()
+        for t in threads:
+            t.join()
+
+        history = list(client._history)
+        self.assertEqual(len(history), 12)
+        self.assertEqual([m["role"] for m in history], ["user", "assistant"] * 6)
+
 
 if __name__ == "__main__":
     unittest.main()
diff --git a/tests/test_llm_identity.py b/tests/test_llm_identity.py
new file mode 100644
index 0000000..8271976
--- /dev/null
+++ b/tests/test_llm_identity.py
@@ -0,0 +1,94 @@
+import unittest
+
+from pepper_wizard.llm.identity import (
+    resolve_config,
+    config_fingerprint,
+    build_config_snapshot,
+)
+
+_BEHAVIOURAL = {"model", "system_prompt", "max_tokens", "temperature", "history_turns"}
+
+
+class ResolveConfigTests(unittest.TestCase):
+    def test_applies_defaults_for_missing_fields(self):
+        resolved = resolve_config({"system_prompt": "hi"})
+        self.assertEqual(resolved["model"], "claude-haiku-4-5")
+        self.assertEqual(resolved["max_tokens"], 256)
+        self.assertEqual(resolved["temperature"], 0.7)
+        self.assertEqual(resolved["history_turns"], 10)
+        self.assertEqual(resolved["system_prompt"], "hi")
+
+    def test_drops_non_behavioural_keys(self):
+        resolved = resolve_config({"name": "persona-a", "model": "m", "extra": 1})
+        self.assertNotIn("name", resolved)
+        self.assertNotIn("extra", resolved)
+        self.assertEqual(set(resolved), _BEHAVIOURAL)
+
+
+class FingerprintTests(unittest.TestCase):
+    def test_deterministic_and_order_independent(self):
+        a = config_fingerprint(
+            {"model": "m", "system_prompt": "p", "max_tokens": 1,
+             "temperature": 0.1, "history_turns": 2}
+        )
+        b = config_fingerprint(
+            {"history_turns": 2, "temperature": 0.1, "max_tokens": 1,
+             "system_prompt": "p", "model": "m"}
+        )
+        self.assertEqual(a, b)
+
+    def test_length_is_twelve_hex(self):
+        h = config_fingerprint(resolve_config({}))
+        self.assertEqual(len(h), 12)
+        int(h, 16)  # raises if not hex
+
+    def test_prompt_edit_changes_hash(self):
+        base = resolve_config({"system_prompt": "old"})
+        edited = resolve_config({"system_prompt": "new"})
+        self.assertNotEqual(config_fingerprint(base), config_fingerprint(edited))
+
+    def test_fingerprint_ignores_non_behavioural_keys(self):
+        # A raw config (with name + extras) and its resolved form fingerprint
+        # identically: config_fingerprint resolves internally, so there is no
+        # raw-vs-resolved mis-attribution surface.
+        raw = {"name": "persona-a", "model": "m", "system_prompt": "p", "extra": 1}
+        self.assertEqual(config_fingerprint(raw), config_fingerprint(resolve_config(raw)))
+
+
+class BuildSnapshotTests(unittest.TestCase):
+    def test_session_start_shape(self):
+        snap = build_config_snapshot(
+            {"name": "a", "model": "m", "system_prompt": "p"}, "session_start"
+        )
+        self.assertEqual(snap["reason"], "session_start")
+        self.assertEqual(snap["config_source"], "file")
+        self.assertEqual(snap["config_name"], "a")
+        self.assertEqual(len(snap["config_hash"]), 12)
+        self.assertEqual(set(snap["config"]), _BEHAVIOURAL)
+        self.assertNotIn("changed", snap)
+
+    def test_config_name_none_when_absent(self):
+        snap = build_config_snapshot({"model": "m"}, "session_start")
+        self.assertIsNone(snap["config_name"])
+
+    def test_name_excluded_from_hash(self):
+        a = build_config_snapshot({"name": "a", "model": "m", "system_prompt": "p"}, "session_start")
+        b = build_config_snapshot({"name": "b", "model": "m", "system_prompt": "p"}, "session_start")
+        self.assertEqual(a["config_hash"], b["config_hash"])
+        self.assertNotEqual(a["config_name"], b["config_name"])
+
+    def test_reload_includes_changed(self):
+        snap = build_config_snapshot(
+            {"model": "m"}, "reload", changed=["temperature 0.7→0.9"]
+        )
+        self.assertEqual(snap["reason"], "reload")
+        self.assertEqual(snap["changed"], ["temperature 0.7→0.9"])
+
+    def test_hash_matches_fingerprint_of_resolved(self):
+        raw = {"model": "m", "system_prompt": "p"}
+        snap = build_config_snapshot(raw, "session_start")
+        self.assertEqual(snap["config_hash"], config_fingerprint(resolve_config(raw)))
+
+
+if __name__ == "__main__":
+    unittest.main()