feat(rag): retry a failed LLM request before the first token lands — BOR_LLM_RETRIES/BOR_LLM_RETRY_DELAY with a live 'retrying' status

app/config.py

@@ -73,6 +73,13 @@ class Settings(BaseSettings):
     #: pieces are still counted for the per-turn log line but never
     #: emitted — the answer stream itself is unchanged.
     stream_thinking: bool = True
+    #: Retries of a failed LLM request when the endpoint stops responding
+    #: (phase 67, ``BOR_LLM_RETRIES``); ``0`` = no retries (the turn fails
+    #: on the first error, pre-phase-67 behavior).
+    llm_retries: int = 3
+    #: Flat seconds to wait between attempts (phase 67,
+    #: ``BOR_LLM_RETRY_DELAY``); the TODO-locked 5 s, no backoff.
+    llm_retry_delay: float = 5.0
 
     # --- RAG tuning ---
     embedding_dim: int = 768  # verified against aipi /v1 (embed model)
@@ -261,6 +268,24 @@ class Settings(BaseSettings):
             raise ValueError("agent_max_rounds must be >= 0 (0 = no tools)")
         return v
 
+    @field_validator("llm_retries")
+    @classmethod
+    def _llm_retries_non_negative(cls, v: int) -> int:
+        """``0`` is the no-retry kill switch (pre-phase-67 behavior) — a
+        negative value is a typo (the ``agent_max_rounds`` pattern)."""
+        if v < 0:
+            raise ValueError("llm_retries must be >= 0 (0 = no retries)")
+        return v
+
+    @field_validator("llm_retry_delay")
+    @classmethod
+    def _llm_retry_delay_non_negative(cls, v: float) -> float:
+        """A negative delay is a typo — fail loud at startup (the
+        ``agent_max_rounds`` pattern)."""
+        if v < 0:
+            raise ValueError("llm_retry_delay must be >= 0 (seconds)")
+        return v
+
     @field_validator("upload_max_mb")
     @classmethod
     def _upload_max_mb_positive(cls, v: int) -> int: