Sync from GitHub main @ b2c3db43f778c7feaaf7d149697958f5cc69ef21

README.md

@@ -13,17 +13,17 @@ pinned: false
 [![Docker](https://img.shields.io/badge/docker--compose-demo-blue?logo=docker)](docs/runbook.md)
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
 
-A **production-oriented, multi-stage Natural Language → SQL system** built around
-**explicit safety guarantees, verification, evaluation, and observability** — **not a prompt-only demo**.
+**NL2SQL Copilot** is a **safety-first NL→SQL backend system** built as a **multi-stage, testable pipeline**.
+It treats LLM outputs as **proposals** and enforces execution safety via **verification, bounded repair, evaluation, and observability**.
 
-This project treats LLMs as **untrusted components** inside a constrained, measurable system,
-and operates them through an explicit pipeline with verification and a **bounded repair loop**.
+The system is designed to operate LLMs as **untrusted components** inside a constrained, measurable backend —
+with explicit guarantees around what can be generated, executed, and returned.
 
 ---
 
 ## At a glance
 
-* **Agentic NL2SQL pipeline** with explicit planning, verification, and a bounded repair loop
+* **Multi-stage NL2SQL pipeline** with explicit planning, verification, and a bounded repair loop
 * **Safety-first execution** enforced at the system level (single-statement, SELECT-only)
 * **Failure-aware design** with verifier-driven repair and a structured error taxonomy
 * **Built-in evaluation**: lightweight smoke runs and Spider-based benchmarks with artifacts
@@ -68,6 +68,7 @@ A live interactive demo is available on Hugging Face Spaces:
 > This is a minimal end-to-end demo of the deployed system.
 > Detailed UI inspection views (trace, repair, timings) are documented in
 > [docs/ui.md](docs/ui.md).
+
 ---
 
 ## Quickstart (Local)
@@ -158,6 +159,8 @@ Detector
 → Repair (bounded)
 ```
 
+>**LLM outputs are treated as proposals; the system decides what is safe to execute.**
+
 The pipeline is designed so that **failures are explicit, classified, and observable** —
 not hidden behind retries or prompt heuristics.
 

nl2sql/pipeline.py

@@ -391,26 +391,6 @@ class Pipeline:
         details: List[str] = []
         exec_result: Dict[str, Any] = {}
 
-        def _tag_last_trace_attempt(stage_name: str, attempt: int) -> None:
-            # Attach attempt metadata to the most recent trace entry for this stage.
-            for t in reversed(traces):
-                if t.get("stage") == stage_name:
-                    notes = t.get("notes") or {}
-                    if not isinstance(notes, dict):
-                        notes = {}
-                    notes["attempt"] = attempt
-                    t["notes"] = notes
-                    return
-
-        def _fallback_trace(stage_name: str, dt_ms: float, ok: bool) -> None:
-            traces.append(
-                self._mk_trace(
-                    stage=stage_name,
-                    duration_ms=dt_ms,
-                    summary=("ok" if ok else "failed"),
-                )
-            )
-
         schema_preview = schema_preview or ""
         clarify_answers = clarify_answers or {}