r1_dynamic_head

--- a/app/prompts.py
+++ b/app/prompts.py
@@ -1,3 +1,4 @@
+import logging
 import uuid
 from datetime import UTC, datetime
 
@@ -59,10 +60,13 @@
 
 def build_messages(question: str, chunks: list[Chunk], max_chars: int) -> list[dict[str, str]]:
     context = build_context(chunks, max_chars)
+    logging.getLogger(__name__).info(
+        "per-request values (logged, not sent in the prompt): %s", f"Request {uuid.uuid4().hex[:8]} at {datetime.now(UTC).isoformat()}\n"
+    )
     return [
         {
             "role": "system",
-            "content": f"Request {uuid.uuid4().hex[:8]} at {datetime.now(UTC).isoformat()}\n" + SYSTEM_PROMPT,
+            "content": SYSTEM_PROMPT,
         },
         {"role": "user", "content": f"CONTEXT:\n{context}\n\nQUESTION: {question}"},
     ]
--- a/furnace_harness.py
+++ b/furnace_harness.py
@@ -0,0 +1,34 @@
+"""Generated by Furnace Forge from app/main.py::chat.
+
+render(question): the exact chat messages the handler sends to the model, without
+starting the server or calling the model.
+answer(question): the handler's full answer path (model call and any output
+post-processing), as text. Needs the app's model endpoint.
+Regenerate this file if that handler changes."""
+
+from types import SimpleNamespace
+
+from app.main import build_messages, rag_config, retrieve, stream_answer
+
+
+def render(question: str) -> list[dict]:
+    req = SimpleNamespace(question=question)
+    cfg = rag_config()
+    chunks = retrieve(req.question, top_k=cfg["retrieval"]["top_k"])
+    messages = build_messages(req.question, chunks, cfg["retrieval"]["max_context_chars"])
+    return messages
+
+
+def answer(question: str) -> str:
+    req = SimpleNamespace(question=question)
+    cfg = rag_config()
+    chunks = retrieve(req.question, top_k=cfg["retrieval"]["top_k"])
+    messages = build_messages(req.question, chunks, cfg["retrieval"]["max_context_chars"])
+    gen = cfg["generation"]
+    out = stream_answer(
+            messages,
+            max_tokens=gen["max_tokens"],
+            temperature=gen["temperature"],
+            route="/chat",
+        )
+    return out if isinstance(out, str) else "".join(out)
--- a/tests/furnace_regression/test_prefix_stability_a4eaf424.py
+++ b/tests/furnace_regression/test_prefix_stability_a4eaf424.py
@@ -0,0 +1,37 @@
+"""Generated by Furnace Guard: regression test written before any fix was attempted.
+
+Failure mode: prefix instability (dynamic value at the head of the system prompt)
+Change under test: r1_dynamic_head
+
+Every request must start with the same prompt text so the serving engine can reuse
+the cached prefill of that shared prefix (prefix/KV caching). On the base revision,
+two different requests shared 3954 leading characters; this test requires at
+least 3558 (90% of that).
+"""
+
+from furnace_harness import render
+
+MIN_SHARED_CHARS = 3558
+QUESTIONS = ['How much does the Studio plan cost?', 'How many kilns does the Workshop plan support?', 'Is there a free trial?', 'When are invoices issued?', 'Which payment methods are accepted?', 'How do I get a refund?']
+
+
+def _text(question: str) -> str:
+    return "".join(f"{m['role']}\n{m['content']}\n" for m in render(question))
+
+
+def _shared(a: str, b: str) -> int:
+    n = min(len(a), len(b))
+    i = 0
+    while i < n and a[i] == b[i]:
+        i += 1
+    return i
+
+
+def test_prompt_prefix_is_shared_across_requests():
+    first = _text(QUESTIONS[0])
+    for q in QUESTIONS[1:]:
+        shared = _shared(first, _text(q))
+        assert shared >= MIN_SHARED_CHARS, (
+            f"requests diverge after {shared} characters (need {MIN_SHARED_CHARS}); "
+            f"first difference: {first[shared:shared + 60]!r}"
+        )
 

--- a/app/prompts.py
+++ b/app/prompts.py
@@ -1,3 +1,4 @@
+import logging
 import uuid
 from datetime import UTC, datetime
 
@@ -59,10 +60,13 @@
 
 def build_messages(question: str, chunks: list[Chunk], max_chars: int) -> list[dict[str, str]]:
     context = build_context(chunks, max_chars)
+    logging.getLogger(__name__).info(
+        "per-request values (logged, not sent in the prompt): %s", f"Request {uuid.uuid4().hex[:8]} at {datetime.now(UTC).isoformat()}\n"
+    )
     return [
         {
             "role": "system",
-            "content": f"Request {uuid.uuid4().hex[:8]} at {datetime.now(UTC).isoformat()}\n" + SYSTEM_PROMPT,
+            "content": SYSTEM_PROMPT,
         },
         {"role": "user", "content": f"CONTEXT:\n{context}\n\nQUESTION: {question}"},
     ]
--- a/furnace_harness.py
+++ b/furnace_harness.py
@@ -0,0 +1,34 @@
+"""Generated by Furnace Forge from app/main.py::chat.
+
+render(question): the exact chat messages the handler sends to the model, without
+starting the server or calling the model.
+answer(question): the handler's full answer path (model call and any output
+post-processing), as text. Needs the app's model endpoint.
+Regenerate this file if that handler changes."""
+
+from types import SimpleNamespace
+
+from app.main import build_messages, rag_config, retrieve, stream_answer
+
+
+def render(question: str) -> list[dict]:
+    req = SimpleNamespace(question=question)
+    cfg = rag_config()
+    chunks = retrieve(req.question, top_k=cfg["retrieval"]["top_k"])
+    messages = build_messages(req.question, chunks, cfg["retrieval"]["max_context_chars"])
+    return messages
+
+
+def answer(question: str) -> str:
+    req = SimpleNamespace(question=question)
+    cfg = rag_config()
+    chunks = retrieve(req.question, top_k=cfg["retrieval"]["top_k"])
+    messages = build_messages(req.question, chunks, cfg["retrieval"]["max_context_chars"])
+    gen = cfg["generation"]
+    out = stream_answer(
+            messages,
+            max_tokens=gen["max_tokens"],
+            temperature=gen["temperature"],
+            route="/chat",
+        )
+    return out if isinstance(out, str) else "".join(out)
--- a/tests/furnace_regression/test_prefix_stability_a4eaf424.py
+++ b/tests/furnace_regression/test_prefix_stability_a4eaf424.py
@@ -0,0 +1,37 @@
+"""Generated by Furnace Guard: regression test written before any fix was attempted.
+
+Failure mode: prefix instability (dynamic value at the head of the system prompt)
+Change under test: r1_dynamic_head
+
+Every request must start with the same prompt text so the serving engine can reuse
+the cached prefill of that shared prefix (prefix/KV caching). On the base revision,
+two different requests shared 3954 leading characters; this test requires at
+least 3558 (90% of that).
+"""
+
+from furnace_harness import render
+
+MIN_SHARED_CHARS = 3558
+QUESTIONS = ['How much does the Studio plan cost?', 'How many kilns does the Workshop plan support?', 'Is there a free trial?', 'When are invoices issued?', 'Which payment methods are accepted?', 'How do I get a refund?']
+
+
+def _text(question: str) -> str:
+    return "".join(f"{m['role']}\n{m['content']}\n" for m in render(question))
+
+
+def _shared(a: str, b: str) -> int:
+    n = min(len(a), len(b))
+    i = 0
+    while i < n and a[i] == b[i]:
+        i += 1
+    return i
+
+
+def test_prompt_prefix_is_shared_across_requests():
+    first = _text(QUESTIONS[0])
+    for q in QUESTIONS[1:]:
+        shared = _shared(first, _text(q))
+        assert shared >= MIN_SHARED_CHARS, (
+            f"requests diverge after {shared} characters (need {MIN_SHARED_CHARS}); "
+            f"first difference: {first[shared:shared + 60]!r}"
+        )
 

--- a/app/prompts.py
+++ b/app/prompts.py
@@ -1,3 +1,4 @@
+import logging
 import uuid
 from datetime import UTC, datetime
 
@@ -59,10 +60,13 @@
 
 def build_messages(question: str, chunks: list[Chunk], max_chars: int) -> list[dict[str, str]]:
     context = build_context(chunks, max_chars)
+    logging.getLogger(__name__).info(
+        "per-request values (logged, not sent in the prompt): %s", f"Request {uuid.uuid4().hex[:8]} at {datetime.now(UTC).isoformat()}\n"
+    )
     return [
         {
             "role": "system",
-            "content": f"Request {uuid.uuid4().hex[:8]} at {datetime.now(UTC).isoformat()}\n" + SYSTEM_PROMPT,
+            "content": SYSTEM_PROMPT,
         },
         {"role": "user", "content": f"CONTEXT:\n{context}\n\nQUESTION: {question}"},
     ]
--- a/furnace_harness.py
+++ b/furnace_harness.py
@@ -0,0 +1,34 @@
+"""Generated by Furnace Forge from app/main.py::chat.
+
+render(question): the exact chat messages the handler sends to the model, without
+starting the server or calling the model.
+answer(question): the handler's full answer path (model call and any output
+post-processing), as text. Needs the app's model endpoint.
+Regenerate this file if that handler changes."""
+
+from types import SimpleNamespace
+
+from app.main import build_messages, rag_config, retrieve, stream_answer
+
+
+def render(question: str) -> list[dict]:
+    req = SimpleNamespace(question=question)
+    cfg = rag_config()
+    chunks = retrieve(req.question, top_k=cfg["retrieval"]["top_k"])
+    messages = build_messages(req.question, chunks, cfg["retrieval"]["max_context_chars"])
+    return messages
+
+
+def answer(question: str) -> str:
+    req = SimpleNamespace(question=question)
+    cfg = rag_config()
+    chunks = retrieve(req.question, top_k=cfg["retrieval"]["top_k"])
+    messages = build_messages(req.question, chunks, cfg["retrieval"]["max_context_chars"])
+    gen = cfg["generation"]
+    out = stream_answer(
+            messages,
+            max_tokens=gen["max_tokens"],
+            temperature=gen["temperature"],
+            route="/chat",
+        )
+    return out if isinstance(out, str) else "".join(out)
--- a/tests/furnace_regression/test_prefix_stability_a4eaf424.py
+++ b/tests/furnace_regression/test_prefix_stability_a4eaf424.py
@@ -0,0 +1,37 @@
+"""Generated by Furnace Guard: regression test written before any fix was attempted.
+
+Failure mode: prefix instability (dynamic value at the head of the system prompt)
+Change under test: r1_dynamic_head
+
+Every request must start with the same prompt text so the serving engine can reuse
+the cached prefill of that shared prefix (prefix/KV caching). On the base revision,
+two different requests shared 3954 leading characters; this test requires at
+least 3558 (90% of that).
+"""
+
+from furnace_harness import render
+
+MIN_SHARED_CHARS = 3558
+QUESTIONS = ['How much does the Studio plan cost?', 'How many kilns does the Workshop plan support?', 'Is there a free trial?', 'When are invoices issued?', 'Which payment methods are accepted?', 'How do I get a refund?']
+
+
+def _text(question: str) -> str:
+    return "".join(f"{m['role']}\n{m['content']}\n" for m in render(question))
+
+
+def _shared(a: str, b: str) -> int:
+    n = min(len(a), len(b))
+    i = 0
+    while i < n and a[i] == b[i]:
+        i += 1
+    return i
+
+
+def test_prompt_prefix_is_shared_across_requests():
+    first = _text(QUESTIONS[0])
+    for q in QUESTIONS[1:]:
+        shared = _shared(first, _text(q))
+        assert shared >= MIN_SHARED_CHARS, (
+            f"requests diverge after {shared} characters (need {MIN_SHARED_CHARS}); "
+            f"first difference: {first[shared:shared + 60]!r}"
+        )