diff --git a/application/agents/tools/artifact_ref.py b/application/agents/tools/artifact_ref.py
index 830c622d..24751624 100644
--- a/application/agents/tools/artifact_ref.py
+++ b/application/agents/tools/artifact_ref.py
@@ -1,8 +1,11 @@
 """Virtual short artifact handles (``A1``, ``A2``, ...) the model can type to reference an artifact.
 
-A ref is NOT persisted: ``A{n}`` is the n-th artifact (1-based, created_at asc) within the
-caller's parent (``conversation_id`` or ``workflow_run_id``). Refs resolve only inside that
-parent, never cross-tenant; resolution still goes through the parent-scoped authz gate.
+A ref is NOT a stored column: ``A{n}`` is the artifact's STABLE per-parent ``ref_seq``, assigned at
+creation and kept in the artifact's ``metadata``, so deleting an earlier artifact no longer
+re-points a later ref the model already holds. Artifacts created before ``ref_seq`` existed have
+none, so resolution falls back to the legacy positional (n-th by created_at) lookup. Refs resolve
+only inside the caller's parent (``conversation_id`` or ``workflow_run_id``), never cross-tenant;
+resolution still goes through the parent-scoped authz gate.
 """
 
 from __future__ import annotations
@@ -41,6 +44,19 @@ def resolve_artifact_id(
     """Resolve a short ref or a uuid to an artifact id, scoped to the caller's parent; None otherwise."""
     position = parse_ref(raw)
     if position is not None:
+        # A ref is the artifact's stable per-parent ``ref_seq``: resolve by it first so a
+        # deletion of an earlier artifact never re-points this ref. Legacy rows (created
+        # before ref_seq) and repos without the newer method fall back to the positional
+        # (n-th by created_at) lookup.
+        by_seq = getattr(repo, "resolve_id_by_ref_seq", None)
+        if callable(by_seq):
+            resolved = by_seq(
+                position,
+                conversation_id=conversation_id,
+                workflow_run_id=workflow_run_id,
+            )
+            if resolved is not None:
+                return resolved
         return repo.artifact_id_at_position(
             position,
             conversation_id=conversation_id,
diff --git a/application/agents/tools/code_executor.py b/application/agents/tools/code_executor.py
index 62e4fe60..e71c18d7 100644
--- a/application/agents/tools/code_executor.py
+++ b/application/agents/tools/code_executor.py
@@ -20,6 +20,7 @@ from application.sandbox.artifacts_capture import (
     infer_mime as _infer_mime,
     kind_for_mime as _kind_for_mime,
     snapshot_signatures,
+    unique_input_path,
 )
 from application.sandbox.base import ExecResult
 from application.sandbox.sandbox_creator import SandboxCreator
@@ -224,6 +225,9 @@ class CodeExecutorTool(Tool):
         if not inputs:
             return {"loaded": loaded}
         storage = StorageCreator.get_storage()
+        # Two inputs whose current versions share a filename would clobber each other at
+        # the same ``inputs/{name}`` path; track used paths and disambiguate deterministically.
+        used_paths: set = set()
         for raw_id in inputs:
             raw = str(raw_id).strip()
             if not raw:
@@ -295,12 +299,13 @@ class CodeExecutorTool(Tool):
                 return {"error": f"failed to read input artifact {artifact_id}."}
             if max_bytes and len(data) > max_bytes:
                 return {"error": f"input artifact {artifact_id} exceeds the {max_bytes}-byte sandbox input limit."}
+            rel_path = unique_input_path(f"inputs/{filename}", used_paths)
             try:
-                manager.put_file(session_id, f"inputs/{filename}", data)
+                manager.put_file(session_id, rel_path, data)
             except Exception:
                 logger.exception("code_executor: put_file failed for input artifact")
                 return {"error": f"failed to stage input artifact {artifact_id} into the workspace."}
-            loaded.append(f"inputs/{filename}")
+            loaded.append(rel_path)
         return {"loaded": loaded}
 
     def _bridge_chat_attachment(self, raw: str) -> Any:
diff --git a/application/agents/tools/read_document.py b/application/agents/tools/read_document.py
index 6f031fa0..7acdf60b 100644
--- a/application/agents/tools/read_document.py
+++ b/application/agents/tools/read_document.py
@@ -14,6 +14,8 @@ from __future__ import annotations
 import logging
 from typing import Any, Dict, List, Optional
 
+from celery import current_task
+
 from application.agents.tools.artifact_ref import resolve_artifact_id
 from application.agents.tools.attachment_bridge import (
     AttachmentBridgeError,
@@ -188,31 +190,44 @@ class ReadDocumentTool(Tool):
         return result
 
     def _dispatch(self, artifact_id: str, options: Dict[str, Any]) -> Dict[str, Any]:
-        """Enqueue ``parse_document`` on the dedicated parsing queue and await with a timeout.
+        """Parse INLINE inside a Celery worker, else dispatch to the parsing queue and await.
 
-        Deadlock note: this tool may run in the WEB process (/stream) OR inside a
-        Celery worker (headless/scheduled agents). Routing to the dedicated
-        ``parsing`` queue + separate parsing workers avoids worker-self-deadlock;
-        the await degrades (returns an error) on timeout/failure rather than hanging.
+        This tool runs in the WEB process (/stream) OR inside a Celery worker
+        (headless/scheduled/workflow agents). When it already runs inside a worker that also
+        serves the ``parsing`` queue (the shipped default ``-Q docsgpt,parsing``), dispatching
+        and blocking on ``get()`` would self-deadlock: concurrent agent tasks each hold a pool
+        slot blocked in ``get()`` so ``parse_document`` never gets a free slot. So parse INLINE
+        in-process inside a worker; only dispatch+await (degrading on timeout/failure) from web.
         """
+        parent = self._parent()
+
+        # ``current_task`` is a Celery proxy: truthy only while this runs inside a worker task,
+        # falsy in the web process (the bare proxy is NOT identity-None, so test truthiness).
+        if current_task:
+            from application.worker import run_parse_document
+
+            try:
+                result = run_parse_document(artifact_id, parent, self.user_id, options)
+            except Exception as exc:
+                logger.exception("read_document: inline parse failed")
+                return {"status": "error", "error": f"document parsing failed: {type(exc).__name__}: {exc}"}
+            if not isinstance(result, dict):
+                return {"status": "error", "error": "document parsing produced an unexpected result."}
+            return result
+
         from celery.exceptions import TimeoutError as CeleryTimeoutError
 
         from application.api.user.tasks import parse_document
 
-        parent = self._parent()
         timeout = float(getattr(settings, "DOCUMENT_PARSE_TIMEOUT", 120))
         queue = getattr(settings, "DOCUMENT_PARSE_QUEUE", "parsing")
         try:
             async_result = parse_document.apply_async(
                 args=[artifact_id, parent, self.user_id, options], queue=queue
             )
-            # ``read_document`` runs both from the web request path AND from inside a
-            # Celery task (headless/scheduled agents). In a prefork worker
-            # ``task_join_will_block()`` is process-wide, so the default
-            # ``disable_sync_subtasks=True`` makes ``get()`` raise RuntimeError
-            # ("Never call result.get() within a task!"). The dedicated ``parsing``
-            # queue + separate workers already avoid the real self-deadlock, so opt
-            # out of the blanket guard explicitly.
+            # The web process (not a worker) awaits here; ``disable_sync_subtasks=False`` keeps
+            # the call correct if invoked from a non-prefork (eventlet/gevent) worker where the
+            # inline branch above still ran but the blanket guard would otherwise raise.
             result = async_result.get(timeout=timeout, disable_sync_subtasks=False)
         except (CeleryTimeoutError, TimeoutError):
             return {"status": "error", "error": f"document parsing timed out after {int(timeout)}s."}
diff --git a/application/agents/workflows/workflow_engine.py b/application/agents/workflows/workflow_engine.py
index 713a2fa8..adc20b83 100644
--- a/application/agents/workflows/workflow_engine.py
+++ b/application/agents/workflows/workflow_engine.py
@@ -345,6 +345,16 @@ class WorkflowEngine:
         has_structured_response = False
         first_chunk = True
         for event in node_agent.gen(formatted_prompt):
+            # A tool that pauses for approval makes the LLM handler yield
+            # ``tool_calls_pending`` and end. An ephemeral node agent has no resume path,
+            # so silently continuing would leave the node with empty output (or a confusing
+            # "Structured output was expected" when it has a json_schema). Fail visibly.
+            if event.get("type") == "tool_calls_pending":
+                raise ValueError(
+                    f'Node "{node.title}" uses a tool that requires approval, which is not '
+                    "supported inside a workflow. Disable require_approval for tools used in "
+                    "workflow nodes."
+                )
             if "answer" in event:
                 chunk = str(event["answer"])
                 full_response_parts.append(chunk)
@@ -496,6 +506,7 @@ class WorkflowEngine:
         """Stage referenced input artifacts (run-scoped, never cross-tenant) into the workspace."""
         from application.agents.tools.artifact_ref import resolve_artifact_id
         from application.core.settings import settings
+        from application.sandbox.artifacts_capture import unique_input_path
         from application.storage.db.repositories.artifacts import ArtifactsRepository
         from application.storage.db.session import db_readonly
         from application.storage.storage_creator import StorageCreator
@@ -507,6 +518,9 @@ class WorkflowEngine:
             return loaded
         max_bytes = int(getattr(settings, "SANDBOX_MAX_INPUT_BYTES", 0) or 0)
         storage = StorageCreator.get_storage()
+        # Two inputs whose current versions share a filename would clobber each other at the
+        # same ``inputs/{name}`` path; track used paths and disambiguate deterministically.
+        used_paths: set = set()
         for raw in raw_ids:
             with db_readonly() as conn:
                 repo = ArtifactsRepository(conn)
@@ -541,8 +555,9 @@ class WorkflowEngine:
                 raise ValueError(
                     f"input artifact {artifact_id} exceeds the {max_bytes}-byte sandbox input limit."
                 )
-            manager.put_file(session_id, f"inputs/{filename}", data)
-            loaded.append(f"inputs/{filename}")
+            rel_path = unique_input_path(f"inputs/{filename}", used_paths)
+            manager.put_file(session_id, rel_path, data)
+            loaded.append(rel_path)
         return loaded
 
     def _materialize_node_attachments(
diff --git a/application/api/user/artifacts/routes.py b/application/api/user/artifacts/routes.py
index 03e8c318..96911794 100644
--- a/application/api/user/artifacts/routes.py
+++ b/application/api/user/artifacts/routes.py
@@ -384,6 +384,15 @@ class DownloadArtifact(Resource):
                         ),
                         500,
                     )
+                # A 302 to a cross-origin S3 URL can't be read by the app's authed
+                # fetch (the bucket has no CORS grant for the app origin). When the
+                # client opts in via ?disposition=url (or Accept: application/json)
+                # hand the presigned URL back as JSON so it can navigate to it
+                # top-level (no CORS). Default stays a 302 so nothing else breaks.
+                if request.args.get("disposition") == "url" or (
+                    "application/json" in request.headers.get("Accept", "")
+                ):
+                    return make_response(jsonify({"success": True, "url": url}), 200)
                 return redirect(url, code=302)
 
             # Stream the bytes in chunks instead of buffering the whole object in
diff --git a/application/parser/document_reader.py b/application/parser/document_reader.py
index 1d18ac5b..4524394d 100644
--- a/application/parser/document_reader.py
+++ b/application/parser/document_reader.py
@@ -22,10 +22,10 @@ from application.utils import safe_filename
 
 logger = logging.getLogger(__name__)
 
-# Cap the text returned to the LLM so a huge document can't flood context; the
-# full result is still persisted as a ``data`` artifact. When the text exceeds
-# the cap a head+tail window keeps both the document's beginning AND end (e.g.
-# totals/signatures) within the byte budget.
+# Default cap for the LLM-facing VIEW of the text (applied in ``bound_parse_payload``,
+# NOT during parsing) so a huge document can't flood context; the full result is still
+# persisted as a ``data`` artifact. When the text exceeds the cap a head+tail window keeps
+# both the document's beginning AND end (e.g. totals/signatures) within the byte budget.
 _TEXT_MAX_BYTES = 8000
 _MAX_TABLES_RETURNED = 20
 _MAX_TABLE_ROWS = 50
@@ -55,18 +55,26 @@ def truncate_text_head_tail(text: str, max_bytes: Optional[int] = None) -> str:
     return f"{head_text}\n\n...[truncated {dropped} bytes]...\n\n{tail_text}"
 
 
-def bound_parse_payload(payload: Dict[str, Any]) -> Dict[str, Any]:
+def bound_parse_payload(payload: Dict[str, Any], max_chars: Optional[int] = None) -> Dict[str, Any]:
     """Bound every shape of a parse payload so the Redis-backed result stays small.
 
-    ``content`` is re-windowed and ``chunks`` is capped in count and per-chunk
-    length. ``structured`` is left as-is: it rides back so json_schema validation
-    in the tool can run against it, and it is already bounded by the input byte
-    cap plus the table caps (``_compact_table`` / ``summary``); the full result is
-    also persisted as a ``data`` artifact. The dict is mutated in place.
+    This is where ALL view-bounding happens: parsing now returns the FULL content and the
+    persisted ``data`` artifact keeps it, while the view ridden back through the Redis result
+    backend is bounded here. ``content`` is capped to ``max_chars`` when given, else re-windowed
+    to a head+tail byte window; ``chunks`` is capped in count and per-chunk length. ``structured``
+    is left as-is: it rides back so json_schema validation in the tool can run against it, and it
+    is already bounded by the input byte cap plus the table caps (``_compact_table`` /
+    ``summary``). ``payload['truncated']`` is set when the content view actually cut. The dict is
+    mutated in place.
     """
     content = payload.get("content")
     if isinstance(content, str):
-        payload["content"] = truncate_text_head_tail(content)
+        if max_chars and int(max_chars) > 0:
+            capped = content[: int(max_chars)]
+        else:
+            capped = truncate_text_head_tail(content)
+        payload["content"] = capped
+        payload["truncated"] = capped != content
 
     chunks = payload.get("chunks")
     if isinstance(chunks, list):
@@ -336,7 +344,7 @@ def _shape(
             extracted = _docling_structured(path, ocr_enabled=ocr_enabled, include_tables=include_tables)
         except Exception as exc:
             return {"error": f"structured parsing requires Docling: {type(exc).__name__}: {exc}"}
-        bounded, truncated = _bounded(extracted["markdown"], max_chars)
+        bounded, truncated = _bounded(extracted["markdown"])
         return {
             "output": "structured",
             "content": bounded,
@@ -364,7 +372,7 @@ def _shape(
         except Exception:
             text, tables = _parse_to_text(parser, path), []
         text = _apply_pages(text, pages)
-        bounded, truncated = _bounded(text, max_chars)
+        bounded, truncated = _bounded(text)
         payload: Dict[str, Any] = {"output": output, "content": bounded, "truncated": truncated}
         if tables:
             payload["tables"] = tables
@@ -388,17 +396,17 @@ def _shape(
                 path, ocr_enabled=ocr_enabled, include_tables=True)["tables"]]
         except Exception:
             tables = []
-    bounded, truncated = _bounded(text, max_chars)
+    bounded, truncated = _bounded(text)
     payload: Dict[str, Any] = {"output": output, "content": bounded, "truncated": truncated}
     if tables:
         payload["tables"] = tables
     return payload
 
 
-def _bounded(text: str, max_chars: Optional[int]) -> tuple[str, bool]:
-    """Bound text to ``max_chars`` (chars) or the default byte window; flag truncation."""
-    if max_chars and int(max_chars) > 0:
-        capped = text[: int(max_chars)]
-        return capped, len(capped) < len(text)
-    bounded = truncate_text_head_tail(text)
-    return bounded, bounded != text
+def _bounded(text: str) -> tuple[str, bool]:
+    """Return the FULL extracted text (never truncated here); the view is bounded in ``bound_parse_payload``.
+
+    Parsing keeps the complete text so the persisted ``data`` artifact is the full parse;
+    ``max_chars`` and the default head+tail window now bound only the LLM-facing view.
+    """
+    return text, False
diff --git a/application/sandbox/artifacts_capture.py b/application/sandbox/artifacts_capture.py
index 581202c8..07270a84 100644
--- a/application/sandbox/artifacts_capture.py
+++ b/application/sandbox/artifacts_capture.py
@@ -13,7 +13,8 @@ import hashlib
 import io
 import logging
 import mimetypes
-from typing import Any, Dict, List, Optional, Tuple
+import os
+from typing import Any, Dict, List, Optional, Set, Tuple
 
 from sqlalchemy import text
 
@@ -63,6 +64,25 @@ _KIND_BY_MIME_PREFIX: Dict[str, str] = {
 }
 
 
+def unique_input_path(rel_path: str, used: Set[str]) -> str:
+    """Return ``rel_path`` (reserving it) or a ``-2``/``-3``... suffixed variant when already staged.
+
+    Two inputs whose current versions share a filename would otherwise clobber each other at the
+    same ``inputs/{name}`` path; the numeric suffix is inserted before the extension. Mutates
+    ``used`` to record the returned path.
+    """
+    if rel_path not in used:
+        used.add(rel_path)
+        return rel_path
+    base, ext = os.path.splitext(rel_path)
+    n = 2
+    while f"{base}-{n}{ext}" in used:
+        n += 1
+    unique = f"{base}-{n}{ext}"
+    used.add(unique)
+    return unique
+
+
 def infer_mime(filename: str) -> str:
     """Infer a mime type from a filename, falling back to a generic binary type."""
     mime, _ = mimetypes.guess_type(filename)
@@ -271,16 +291,36 @@ def _enforce_user_quota(repo: ArtifactsRepository, user_id: str, added_bytes: in
         raise QuotaExceeded(f"artifact storage quota reached ({max_total} bytes); delete artifacts to free space")
 
 
+def _ref_from_metadata(metadata: Any) -> Optional[str]:
+    """Return the short ref (``A{ref_seq}``) from a stored metadata dict, or None when absent/non-numeric."""
+    if not isinstance(metadata, dict):
+        return None
+    try:
+        seq = int(metadata.get("ref_seq"))
+    except (TypeError, ValueError):
+        return None
+    return make_ref(seq) if seq >= 1 else None
+
+
 def _ref_for(
     repo: ArtifactsRepository,
     artifact_id: str,
     *,
     conversation_id: Optional[str],
     workflow_run_id: Optional[str],
+    metadata: Any = None,
 ) -> Optional[str]:
-    """Compute the short ref (``A{n}``) for an artifact from its position in its parent; None on failure."""
+    """Compute the short ref for an artifact: its STABLE stored ``ref_seq``, else positional fallback.
+
+    The ref_seq is assigned at creation and kept in ``metadata`` so it survives deletions of
+    earlier artifacts. Legacy rows created before ref_seq existed have none, so they fall back to
+    the (mutable) position-in-parent; ``resolve_artifact_id`` mirrors this fallback.
+    """
     if conversation_id is None and workflow_run_id is None:
         return None
+    seq_ref = _ref_from_metadata(metadata)
+    if seq_ref is not None:
+        return seq_ref
     try:
         position = repo.position_in_parent(
             artifact_id, conversation_id=conversation_id, workflow_run_id=workflow_run_id
@@ -339,8 +379,13 @@ def persist_new_artifact(
             artifact_id = str(artifact["id"])
             storage_path = _storage_key(user_id, artifact_id, 1, safe_name)
             _set_version_storage_path(conn, artifact_id, 1, storage_path)
+            # The freshly-created row carries the stable ``ref_seq`` in its metadata.
             ref = _ref_for(
-                repo, artifact_id, conversation_id=conversation_id, workflow_run_id=workflow_run_id
+                repo,
+                artifact_id,
+                conversation_id=conversation_id,
+                workflow_run_id=workflow_run_id,
+                metadata=artifact.get("metadata"),
             )
             storage.save_file(io.BytesIO(data), storage_path)
             saved_key = storage_path
@@ -402,8 +447,15 @@ def append_artifact_version(
             version_number = int(version["version"])
             storage_path = _storage_key(user_id, artifact_id, version_number, safe_name)
             _set_version_storage_path(conn, artifact_id, version_number, storage_path)
+            # The identity's stable ``ref_seq`` was assigned at creation; read it so an
+            # edit hands back the SAME ref the model already holds.
+            existing = repo.get_artifact(str(artifact_id))
             ref = _ref_for(
-                repo, str(artifact_id), conversation_id=conversation_id, workflow_run_id=workflow_run_id
+                repo,
+                str(artifact_id),
+                conversation_id=conversation_id,
+                workflow_run_id=workflow_run_id,
+                metadata=(existing or {}).get("metadata"),
             )
             storage.save_file(io.BytesIO(data), storage_path)
             saved_key = storage_path
diff --git a/application/storage/db/repositories/artifacts.py b/application/storage/db/repositories/artifacts.py
index c3a515be..ebe26aa6 100644
--- a/application/storage/db/repositories/artifacts.py
+++ b/application/storage/db/repositories/artifacts.py
@@ -57,7 +57,21 @@ class ArtifactsRepository:
         preview_text: Optional[str] = None,
         produced_by: Any = None,
     ) -> dict:
-        """Create the identity row and its version 1 atomically; return the artifact dict."""
+        """Create the identity row and its version 1 atomically; return the artifact dict.
+
+        When a parent is given, a STABLE per-parent ``ref_seq`` (max existing + 1) is
+        computed BEFORE the insert and stored into ``metadata`` so the short ``A{n}`` ref
+        the model receives survives later deletions. See ``next_ref_seq`` for the (benign)
+        concurrent-create dup caveat.
+        """
+        metadata_payload = metadata
+        if conversation_id is not None or workflow_run_id is not None:
+            ref_seq = self.next_ref_seq(
+                conversation_id=conversation_id, workflow_run_id=workflow_run_id
+            )
+            merged = dict(metadata) if isinstance(metadata, dict) else {}
+            merged["ref_seq"] = ref_seq
+            metadata_payload = merged
         artifact = self._conn.execute(
             text(
                 """
@@ -84,7 +98,7 @@ class ArtifactsRepository:
                 "message_id": message_id,
                 "kind": kind,
                 "title": title,
-                "metadata": json.dumps(metadata) if metadata is not None else None,
+                "metadata": json.dumps(metadata_payload) if metadata_payload is not None else None,
             },
         ).fetchone()
         artifact_dict = _artifact_to_dict(artifact)
@@ -273,6 +287,61 @@ class ArtifactsRepository:
         ).fetchone()
         return str(row[0]) if row is not None else None
 
+    def next_ref_seq(
+        self,
+        *,
+        conversation_id: Optional[str] = None,
+        workflow_run_id: Optional[str] = None,
+    ) -> int:
+        """Return the next stable ``ref_seq`` for a parent (max stored + 1; 1 when empty).
+
+        Caveat: two concurrent creates in one parent can read the same MAX and get a
+        duplicate ``ref_seq`` (creates within a parent are normally serial from one agent
+        turn); a dup resolves to the earliest — acceptable, and strictly better than the
+        re-point bug. A DB unique index is a possible future hardening (no migration here).
+        """
+        if conversation_id is None and workflow_run_id is None:
+            raise ValueError("next_ref_seq requires conversation_id or workflow_run_id")
+        clauses, params = self._parent_clauses(conversation_id, workflow_run_id)
+        # Guard the cast: legacy rows have no ``ref_seq`` (NULL, ignored), and the regex
+        # match keeps any non-numeric value from erroring the cast.
+        row = self._conn.execute(
+            text(
+                "SELECT COALESCE(MAX(CASE WHEN metadata ->> 'ref_seq' ~ '^[0-9]+$' "
+                "THEN (metadata ->> 'ref_seq')::int END), 0) + 1 "
+                f"FROM artifacts WHERE {' AND '.join(clauses)}"
+            ),
+            params,
+        ).fetchone()
+        return int(row[0]) if row is not None else 1
+
+    def resolve_id_by_ref_seq(
+        self,
+        seq: int,
+        *,
+        conversation_id: Optional[str] = None,
+        workflow_run_id: Optional[str] = None,
+    ) -> Optional[str]:
+        """Return the id of the artifact whose stored ``ref_seq`` == ``seq`` in a parent, or None.
+
+        Earliest ``created_at`` wins on the rare concurrent-create tie (see ``next_ref_seq``).
+        """
+        if conversation_id is None and workflow_run_id is None:
+            raise ValueError("resolve_id_by_ref_seq requires conversation_id or workflow_run_id")
+        if not isinstance(seq, int) or seq < 1:
+            return None
+        clauses, params = self._parent_clauses(conversation_id, workflow_run_id)
+        params["seq"] = str(seq)
+        row = self._conn.execute(
+            text(
+                f"SELECT id FROM artifacts WHERE {' AND '.join(clauses)} "
+                "AND metadata ->> 'ref_seq' = :seq "
+                "ORDER BY created_at ASC, id ASC LIMIT 1"
+            ),
+            params,
+        ).fetchone()
+        return str(row[0]) if row is not None else None
+
     @staticmethod
     def _parent_clauses(
         conversation_id: Optional[str], workflow_run_id: Optional[str], alias: str = ""
diff --git a/application/worker.py b/application/worker.py
index 6cca0728..d7a05a67 100755
--- a/application/worker.py
+++ b/application/worker.py
@@ -1611,11 +1611,18 @@ def attachment_worker(self, file_info, user):
 
 
 def parse_document_worker(self, artifact_id, parent, user_id, options):
-    """Parse an input artifact's bytes to a shaped result on the dedicated parsing queue.
+    """Thin Celery-task wrapper; delegates to the process-agnostic ``run_parse_document``."""
+    return run_parse_document(artifact_id, parent, user_id, options)
 
-    Security: the artifact is re-resolved through the run-scoped gate IN THE WORKER
-    (never trusting a raw storage path) so authz is enforced independently here, in
-    addition to the pre-enqueue check in the tool.
+
+def run_parse_document(artifact_id, parent, user_id, options):
+    """Parse an input artifact's bytes to a shaped result; runnable inline OR on the parsing queue.
+
+    Security: the artifact is re-resolved through the run-scoped gate here (never trusting a
+    raw storage path) so authz is enforced independently, in addition to the pre-enqueue check
+    in the tool. ``read_document`` calls this directly (in-process) when it already runs inside a
+    Celery worker; the web process dispatches ``parse_document`` to the parsing queue, which lands
+    here via ``parse_document_worker``.
     """
     from application.agents.tools.artifact_ref import resolve_artifact_id
     from application.parser.document_reader import bound_parse_payload, parse_document_bytes
@@ -1628,7 +1635,7 @@ def parse_document_worker(self, artifact_id, parent, user_id, options):
         return {"status": "error", "error": "parse_document requires a conversation_id or workflow_run_id."}
 
     # Re-resolve through the parent-scoped gate so a forged/cross-run id is rejected
-    # in the worker too; resolve a short ref to an id within this parent only.
+    # here too; resolve a short ref to an id within this parent only.
     try:
         with db_readonly() as conn:
             repo = ArtifactsRepository(conn)
@@ -1646,7 +1653,7 @@ def parse_document_worker(self, artifact_id, parent, user_id, options):
                 return {"status": "error", "error": f"input artifact {artifact_id} not found in this conversation/run."}
             version = repo.get_version(resolved_id, artifact["current_version"])
     except Exception:
-        logging.error("parse_document_worker: failed to resolve input artifact", exc_info=True)
+        logging.error("run_parse_document: failed to resolve input artifact", exc_info=True)
         return {"status": "error", "error": f"failed to load input artifact {artifact_id}."}
 
     if not version or not version.get("storage_path"):
@@ -1657,9 +1664,11 @@ def parse_document_worker(self, artifact_id, parent, user_id, options):
     try:
         data = StorageCreator.get_storage().get_file(version["storage_path"]).read()
     except Exception:
-        logging.error("parse_document_worker: failed to read input artifact bytes", exc_info=True)
+        logging.error("run_parse_document: failed to read input artifact bytes", exc_info=True)
         return {"status": "error", "error": f"failed to read input artifact {artifact_id}."}
 
+    # Parse returns the FULL content (no max_chars/window here) so the persisted artifact is the
+    # complete parse; all view-bounding is applied by ``bound_parse_payload`` below.
     result = parse_document_bytes(
         data,
         filename,
@@ -1667,7 +1676,6 @@ def parse_document_worker(self, artifact_id, parent, user_id, options):
         ocr=options.get("ocr", "auto"),
         pages=options.get("pages"),
         engine=options.get("engine", "auto"),
-        max_chars=options.get("max_chars"),
         include_tables=bool(options.get("include_tables", True)),
     )
     if result.get("error"):
@@ -1675,17 +1683,17 @@ def parse_document_worker(self, artifact_id, parent, user_id, options):
 
     payload = {"status": "ok", **result}
     if options.get("persist"):
-        # The full shaped result is persisted by reference; only a bounded view rides
-        # back through the Redis result backend (the bytes live in the artifact).
+        # Persist the FULL shaped result by reference (bytes live in the artifact); only the
+        # bounded view computed below rides back through the Redis result backend.
         artifact_ref = _persist_parse_result(result, display_name, user_id, parent, options)
         if isinstance(artifact_ref, dict) and artifact_ref.get("error"):
             payload["artifact_error"] = artifact_ref["error"]
         elif artifact_ref is not None:
             payload["artifact"] = artifact_ref
-    # Bound the Redis-backed view across all shapes: content is re-windowed, chunks are
-    # capped, and structured (needed for json_schema validation) is bounded by the input
-    # cap + table caps. The FULL result already lives in the persisted artifact above.
-    payload = bound_parse_payload(payload)
+    # Bound the Redis-backed VIEW across all shapes: content is capped by max_chars (else the
+    # default head+tail window), chunks are count/length-capped, and structured (needed for
+    # json_schema validation) rides back as-is. The FULL result already lives in the artifact.
+    payload = bound_parse_payload(payload, max_chars=options.get("max_chars"))
     return payload
 
 
diff --git a/deployment/k8s/deployments/sandbox-deploy.yaml b/deployment/k8s/deployments/sandbox-deploy.yaml
index 941c99ef..24f92e81 100644
--- a/deployment/k8s/deployments/sandbox-deploy.yaml
+++ b/deployment/k8s/deployments/sandbox-deploy.yaml
@@ -34,6 +34,10 @@ spec:
       # runtimeClassName: gvisor
       securityContext:
         runAsNonRoot: true
+        # Numeric UID the kubelet can verify as non-root; MUST match the sandbox
+        # image's `USER` (deployment/sandbox/Dockerfile). Without it the pod
+        # fails to start with CreateContainerConfigError.
+        runAsUser: 10001
         seccompProfile:
           type: RuntimeDefault
       containers:
diff --git a/deployment/k8s/network-policies/sandbox-egress-policy.yaml b/deployment/k8s/network-policies/sandbox-egress-policy.yaml
index 489cb6eb..bf9ca8d2 100644
--- a/deployment/k8s/network-policies/sandbox-egress-policy.yaml
+++ b/deployment/k8s/network-policies/sandbox-egress-policy.yaml
@@ -40,6 +40,7 @@ spec:
   egress:
     # DNS to cluster resolvers (kube-dns). Restricted to UDP/TCP 53 so the
     # broad-egress rule below does not need to whitelist the resolver IP.
+    # Kept for clusters WITHOUT NodeLocal DNSCache, where the resolver is a pod.
     - to:
         - namespaceSelector: {}
       ports:
@@ -47,6 +48,22 @@ spec:
           port: 53
         - protocol: TCP
           port: 53
+    # NodeLocal DNSCache: on clusters running it, the stub resolver is a
+    # link-local IP (commonly 169.254.20.10), NOT a pod, and it sits inside the
+    # 169.254.0.0/16 range carved out of the broad-egress rule below -- so
+    # without this explicit allow ALL DNS is dropped and every hostname lookup
+    # (pip install, public APIs) times out. NetworkPolicy rules are
+    # additive-allow, so this narrowly re-permits port 53 to that one IP.
+    # Operators whose NodeLocal DNSCache uses a different link-local IP must
+    # adjust this cidr.
+    - to:
+        - ipBlock:
+            cidr: 169.254.20.10/32   # NodeLocal DNSCache link-local resolver
+      ports:
+        - protocol: UDP
+          port: 53
+        - protocol: TCP
+          port: 53
     # Broad public-internet egress, with every private / link-local / ULA /
     # carrier-grade-NAT / documentation range carved out via `except`. The
     # cloud metadata IP 169.254.169.254 is inside the 169.254.0.0/16 hole, so
diff --git a/deployment/optional/docker-compose.optional.sandbox-egress.yaml b/deployment/optional/docker-compose.optional.sandbox-egress.yaml
index 4af85004..eca41f23 100644
--- a/deployment/optional/docker-compose.optional.sandbox-egress.yaml
+++ b/deployment/optional/docker-compose.optional.sandbox-egress.yaml
@@ -11,31 +11,68 @@
 # (control plane to backend/worker) and `sandbox-egress` (its internet route).
 # This overlay flips `sandbox-egress` to `internal: true`, which removes the
 # runner's direct route to the internet / host / RFC1918 / metadata entirely --
-# so even raw sockets in arbitrary sandbox code have nowhere to go except the
-# proxy on `sandbox-net`. The proxy is the ONLY container with an internet route
+# so raw sockets in arbitrary sandbox code have no route OFF the host except the
+# deny-private proxy. The proxy is the ONLY container with an internet route
 # (on its own `sandbox-egress-out` network) and its ACL denies private
 # destinations. We flip a network's scalar property by KEY (a well-defined
 # Compose merge); we do NOT try to remove an item from the service's `networks`
 # LIST, which Compose cannot express (it unions lists) -- the reason a naive
 # `networks: [sandbox-net]` override silently leaves the runner on its old net.
 #
-# Two viable approaches; pick one:
+# WHAT THIS OVERLAY DOES NOT CONTAIN (read before enabling the sandbox):
+# The runner STILL shares `sandbox-net` with `backend` and `worker` -- that is
+# its control path and it cannot be removed without breaking code execution. A
+# shared Docker network is bidirectional and Compose cannot sever it
+# one-directionally, so arbitrary sandbox code can still open sockets to
+# `backend:7091` and reach the worker. The internal flip contains raw sockets to
+# the internet / host / RFC1918 / metadata, NOT to backend/worker on sandbox-net.
+# (The Kubernetes NetworkPolicy DOES block this, via its RFC1918 egress
+# carve-out; compose has no equivalent.)
 #
-#  (1) Host / cloud firewall (simplest, belt-and-suspenders for single-host):
-#      Also drop egress to RFC1918 (10/8, 172.16/12, 192.168/16), link-local
-#      (169.254/16, incl. the 169.254.169.254 metadata IP), and ULA on the
-#      docsgpt-sandbox container's interface using iptables/nftables on the
-#      Docker host. Example (host root):
-#        SBX=$(docker inspect -f '{{.NetworkSettings.Networks.docsgpt-oss_sandbox-egress.IPAddress}}' docsgpt-sandbox)
-#        iptables -I DOCKER-USER -s "$SBX" -d 169.254.0.0/16 -j DROP
-#        iptables -I DOCKER-USER -s "$SBX" -d 10.0.0.0/8     -j DROP
-#        iptables -I DOCKER-USER -s "$SBX" -d 172.16.0.0/12  -j DROP
-#        iptables -I DOCKER-USER -s "$SBX" -d 192.168.0.0/16 -j DROP
-#      (Allow established/return traffic and DNS as needed for your setup.)
+# MITIGATION -- required when enabling the sandbox (do the first, ideally both):
+#  - Run the backend with real authentication (`AUTH_TYPE` != none / a real auth
+#    provider) so a reachable API rejects unauthenticated requests. Without it,
+#    runner->backend reach is a free control-plane bypass.
+#  - Add a host-firewall DROP for runner->backend/worker on sandbox-net
+#    (approach (1) below).
+#
+# Two layers below: (2) is what this overlay wires up (internet egress via the
+# deny-private proxy); (1) is the host-firewall DROP that closes the
+# runner->backend/worker gap. Apply both for the strongest posture.
+#
+#  (1) Host-firewall DROP for runner->backend/worker (closes the gap above). Run
+#      as root on the Docker host AFTER `... up`. The project name is pinned to
+#      `docsgpt-oss` (see `name:` in docker-compose.yaml), so the shared control
+#      net is `docsgpt-oss_sandbox-net`. `docker compose ... ps -q` resolves each
+#      container id regardless of its generated name (there is no fixed
+#      `container_name`), and the `index` function is REQUIRED because a
+#      Go-template dotted key cannot contain hyphens -- the old
+#      `{{.NetworkSettings.Networks.docsgpt-oss_sandbox-egress.IPAddress}}` form
+#      never parsed. Both IPs are read on sandbox-net so source+dest match the
+#      shared path:
+#        CF="-f deployment/docker-compose.yaml -f deployment/optional/docker-compose.optional.sandbox-egress.yaml --profile sandbox"
+#        NET=docsgpt-oss_sandbox-net
+#        SBX=$(docker inspect -f "{{(index .NetworkSettings.Networks \"$NET\").IPAddress}}" "$(docker compose $CF ps -q docsgpt-sandbox)")
+#        BE=$( docker inspect -f "{{(index .NetworkSettings.Networks \"$NET\").IPAddress}}" "$(docker compose $CF ps -q backend)")
+#        WK=$( docker inspect -f "{{(index .NetworkSettings.Networks \"$NET\").IPAddress}}" "$(docker compose $CF ps -q worker)")
+#        iptables -I DOCKER-USER -s "$SBX" -d "$BE" -j DROP   # runner -> backend (incl. :7091)
+#        iptables -I DOCKER-USER -s "$SBX" -d "$WK" -j DROP   # runner -> worker
+#      (These container-to-container DROPs require the host's
+#      `bridge-nf-call-iptables=1`, Docker's default, so bridged traffic
+#      traverses DOCKER-USER.) If a lookup is empty (renamed project, or the
+#      container is not up yet), resolve by hand instead of running a broken
+#      command: `docker compose $CF ps` to list ids, then `docker inspect
+#      <runner-id>` / `<backend-id>`, read the sandbox-net
+#      `.NetworkSettings.Networks.*.IPAddress`, and add the same DROP rules.
+#      Re-run after any recreate -- container IPs are not stable across `up`.
+#      (Internet / RFC1918 / metadata egress is already gone via the internal
+#      flip, so no separate SSRF drop is needed here.)
 #
 #  (2) Egress-gateway sidecar (this overlay): the runner has NO direct internet
 #      route (see above); its only path out is the deny-private forward proxy.
 #      Point the runner's HTTP(S) client at the proxy via the env vars below.
+#      This does NOT stop runner->backend/worker on sandbox-net -- pair it with
+#      backend auth and/or approach (1).
 #
 # Apply alongside the base stack:
 #   docker compose -f deployment/docker-compose.yaml \
@@ -43,11 +80,14 @@
 # (also works layered on docker-compose-hub.yaml / docker-compose-azure.yaml,
 # which share the same service/network names).
 #
-# NOTE: because `sandbox-egress` is internal here, non-HTTP raw egress is blocked
-# at L3 regardless of what the code does (there is simply no route). HTTP(S) that
-# honors the proxy env is filtered by the proxy ACL. For belt-and-suspenders on a
-# multi-tenant or untrusted deployment, also apply approach (1) or the Kubernetes
-# NetworkPolicy.
+# NOTE: because `sandbox-egress` is internal here, non-HTTP raw egress TO THE
+# INTERNET / HOST / RFC1918 / METADATA is blocked at L3 regardless of what the
+# code does (there is simply no route). HTTP(S) that honors the proxy env is
+# filtered by the proxy ACL. This does NOT block raw egress to `backend:7091` /
+# the worker on the shared `sandbox-net` (see WHAT THIS OVERLAY DOES NOT CONTAIN
+# / MITIGATION above): enable backend auth and/or apply approach (1). On a
+# multi-tenant or untrusted deployment, prefer the Kubernetes NetworkPolicy,
+# which blocks the internal path too.
 
 services:
   # Forward proxy that denies private/link-local/metadata destinations and
diff --git a/deployment/sandbox/Dockerfile b/deployment/sandbox/Dockerfile
index 39f25748..b1ac1816 100644
--- a/deployment/sandbox/Dockerfile
+++ b/deployment/sandbox/Dockerfile
@@ -58,7 +58,10 @@ COPY kernel-launch.sh /opt/docsgpt/kernel-launch.sh
 RUN chmod 0555 /opt/docsgpt/kernel-launch.sh
 COPY kernels/docsgpt-python/kernel.json /usr/local/share/jupyter/kernels/docsgpt-python/kernel.json
 
-USER sandbox
+# Numeric UID (not the name) so a kubelet with `runAsNonRoot: true` can verify
+# the user is non-root without resolving /etc/passwd. This uid MUST match
+# `runAsUser` in deployment/k8s/deployments/sandbox-deploy.yaml.
+USER 10001
 WORKDIR /home/sandbox
 
 EXPOSE 8888
diff --git a/deployment/sandbox/README.md b/deployment/sandbox/README.md
index bb63d8e9..e68a80e8 100644
--- a/deployment/sandbox/README.md
+++ b/deployment/sandbox/README.md
@@ -228,11 +228,20 @@ The hardened container runs **without `NET_ADMIN`**, so it cannot self-apply
   [`deployment/optional/docker-compose.optional.sandbox-egress.yaml`](../optional/docker-compose.optional.sandbox-egress.yaml),
   which flips `sandbox-egress` to `internal: true` (removing the runner's direct
   internet/RFC1918/metadata route entirely) and forces egress through a
-  deny-private **egress-gateway proxy** sidecar; for belt-and-suspenders on
-  untrusted/multi-tenant hosts also add a **host/cloud firewall rule** (drop the
-  four private ranges on the sandbox container's interface), since a forward
-  proxy only filters code that honors `HTTP(S)_PROXY` — the `internal` flip is
-  what contains raw sockets. Note the broker/DB published ports are bound to
+  deny-private **egress-gateway proxy** sidecar. That `internal` flip is what
+  contains **raw sockets to the internet/host/RFC1918/metadata** (a forward proxy
+  only filters code that honors `HTTP(S)_PROXY`).
+
+  **What compose canNOT contain:** the runner stays on `sandbox-net` with the
+  backend and worker — that is its control path, and a shared Docker network is
+  bidirectional, so Compose cannot sever it one-directionally. Sandbox code can
+  therefore still open sockets to `backend:7091` and the worker. This is a real
+  gap the Kubernetes NetworkPolicy closes (via its RFC1918 egress carve-out) but
+  compose cannot. **Mitigate it** when enabling the sandbox: run the backend with
+  real authentication (`AUTH_TYPE` != none / a real auth provider) so a reachable
+  API rejects unauthenticated requests — **required** — and/or add a host-firewall
+  `DROP` for runner→backend/worker on `sandbox-net` (see approach (1) in the
+  overlay file's header comment). Note the broker/DB published ports are bound to
   `127.0.0.1` so the runner cannot reach them via the host gateway either.
 
 ## Other hardening (deployment-level)
diff --git a/frontend/src/agents/workflow/WorkflowBuilder.tsx b/frontend/src/agents/workflow/WorkflowBuilder.tsx
index b3c81fce..753586da 100644
--- a/frontend/src/agents/workflow/WorkflowBuilder.tsx
+++ b/frontend/src/agents/workflow/WorkflowBuilder.tsx
@@ -723,25 +723,34 @@ function WorkflowBuilderInner() {
   }, [isPublishing]);
 
   useEffect(() => {
+    // A keystroke targeting a text field / editable popover (the Code / JSON
+    // schema textareas, the Output Variable input) must edit the field, never
+    // fire a canvas shortcut.
+    const isEditableTarget = (target: EventTarget | null): boolean => {
+      const el = target as HTMLElement | null;
+      return (
+        el?.tagName === 'INPUT' ||
+        el?.tagName === 'TEXTAREA' ||
+        el?.isContentEditable === true
+      );
+    };
+    // Shared guard for both canvas shortcuts (Delete/Backspace to remove the
+    // selected node, Escape to close the config panel): ignore the keystroke
+    // while typing in a field, while the Preview Sheet is open (its own inputs
+    // own the keys — a stray Delete there must not delete the node behind it),
+    // or once another handler has already consumed the event. Kept in one place
+    // so the two branches can't drift apart.
+    const shouldIgnoreShortcut = (e: KeyboardEvent): boolean =>
+      e.defaultPrevented || showPreview || isEditableTarget(e.target);
+
     const handleKeyDown = (e: KeyboardEvent) => {
-      if (e.key === 'Delete' && selectedNode) {
+      if (shouldIgnoreShortcut(e)) return;
+
+      if ((e.key === 'Delete' || e.key === 'Backspace') && selectedNode) {
         handleDeleteNode();
+        return;
       }
       if (e.key === 'Escape') {
-        // While the Preview Sheet is open, leave Escape to its own dismissal
-        // (popovers, the Sheet itself) so a reflexive Escape can't tear down
-        // the node-config panel and lose the run/attachments/typed prompt.
-        if (showPreview) return;
-        // Ignore Escape originating from text fields / editable popovers so it
-        // dismisses the field rather than the whole config panel.
-        const target = e.target as HTMLElement | null;
-        if (
-          target?.tagName === 'INPUT' ||
-          target?.tagName === 'TEXTAREA' ||
-          target?.isContentEditable
-        ) {
-          return;
-        }
         setShowNodeConfig(false);
         setSelectedNode(null);
       }
diff --git a/frontend/src/agents/workflow/WorkflowPreview.tsx b/frontend/src/agents/workflow/WorkflowPreview.tsx
index 3b8151c5..fbc2e92b 100644
--- a/frontend/src/agents/workflow/WorkflowPreview.tsx
+++ b/frontend/src/agents/workflow/WorkflowPreview.tsx
@@ -32,6 +32,7 @@ import {
   addQuery,
   fetchWorkflowPreviewAnswer,
   handleWorkflowPreviewAbort,
+  previewSendBlockReason,
   resendQuery,
   resetWorkflowPreview,
   selectActiveNodeId,
@@ -443,6 +444,9 @@ export default function WorkflowPreview({
   const hasCompletedAttachment = completedAttachments.length > 0;
 
   const [lastQueryReturnedErr, setLastQueryReturnedErr] = useState(false);
+  const [sendBlockedMessage, setSendBlockedMessage] = useState<string | null>(
+    null,
+  );
   const [openDetailsIndex, setOpenDetailsIndex] = useState<number | null>(null);
   const [openArtifactsIndex, setOpenArtifactsIndex] = useState<number | null>(
     null,
@@ -493,6 +497,20 @@ export default function WorkflowPreview({
       isRetry?: boolean;
       index?: number;
     }) => {
+      // An unsaved draft can't bridge uploaded documents into the run (no
+      // persisted workflow_id), so block the send rather than run with the docs
+      // silently dropped. The uploads are kept (not cleared) for a retry after
+      // the workflow is saved.
+      const blockReason = previewSendBlockReason(
+        workflowId,
+        hasCompletedAttachment,
+      );
+      if (blockReason) {
+        setSendBlockedMessage(blockReason);
+        return;
+      }
+      setSendBlockedMessage(null);
+
       const trimmedQuestion = question.trim();
       // Doc-driven nodes read ``input_documents`` rather than the query, so an
       // attachment-only run is allowed to proceed with an empty question.
@@ -509,9 +527,15 @@ export default function WorkflowPreview({
         handleFetchAnswer({ question: trimmedQuestion, index: undefined });
       }
     },
-    [dispatch, handleFetchAnswer, hasCompletedAttachment],
+    [dispatch, handleFetchAnswer, hasCompletedAttachment, workflowId],
   );
 
+  // Clear the block message once it no longer applies (the workflow was saved,
+  // or the attachments were removed) so a stale warning doesn't linger.
+  useEffect(() => {
+    if (workflowId || !hasCompletedAttachment) setSendBlockedMessage(null);
+  }, [workflowId, hasCompletedAttachment]);
+
   const handleQuestionSubmission = (
     question?: string,
     updated?: boolean,
@@ -710,6 +734,11 @@ export default function WorkflowPreview({
             )}
           </div>
           <div className="bg-card absolute right-0 bottom-0 left-0 flex w-full flex-col gap-2 px-4 pt-2 pb-4">
+            {sendBlockedMessage && (
+              <p className="text-xs text-red-500" role="alert">
+                {sendBlockedMessage}
+              </p>
+            )}
             <MessageInput
               onSubmit={(text) => handleQuestionSubmission(text)}
               loading={status === 'loading'}
diff --git a/frontend/src/agents/workflow/WorkflowRunArtifacts.tsx b/frontend/src/agents/workflow/WorkflowRunArtifacts.tsx
index 7a17259b..53703468 100644
--- a/frontend/src/agents/workflow/WorkflowRunArtifacts.tsx
+++ b/frontend/src/agents/workflow/WorkflowRunArtifacts.tsx
@@ -47,9 +47,15 @@ export default function WorkflowRunArtifacts({
       .then(async (res: Response) => {
         if (cancelled) return;
         if (!res.ok) {
-          // The run row is missing/unauthorized (e.g. an unsaved-draft preview):
-          // show an informational empty state rather than a hard error.
-          setArtifacts([]);
+          // A 403 is the expected unsaved-draft / unauthorized case (no
+          // persisted run row): show the informational empty state. Any other
+          // non-OK status (500, an expired 401, ...) is a real failure — surface
+          // an error + Retry instead of masquerading as "no artifacts".
+          if (res.status === 403) {
+            setArtifacts([]);
+          } else {
+            setError('Failed to load artifacts');
+          }
           setLoading(false);
           return;
         }
diff --git a/frontend/src/agents/workflow/components/PromptTextArea.tsx b/frontend/src/agents/workflow/components/PromptTextArea.tsx
index c85a7221..ce133990 100644
--- a/frontend/src/agents/workflow/components/PromptTextArea.tsx
+++ b/frontend/src/agents/workflow/components/PromptTextArea.tsx
@@ -14,6 +14,12 @@ export interface WorkflowVariable {
   label: string;
   templatePath: string;
   section: string;
+  // True when the variable resolves to artifact reference(s) at run time
+  // (uploaded input_documents, code-node outputs) rather than plain LLM/state
+  // TEXT. Only artifact-bearing variables may be picked as node Documents;
+  // selecting a text output there makes the engine append the literal variable
+  // name and the node hard-fails. Heuristic keyed on the producing node type.
+  producesArtifact?: boolean;
 }
 
 const GLOBAL_CONTEXT_VARIABLES: WorkflowVariable[] = [
@@ -118,41 +124,54 @@ export function extractUpstreamVariables(
       label: 'agent.input_documents',
       templatePath: 'agent.input_documents',
       section: 'Workflow input',
+      // Uploaded documents are artifact references.
+      producesArtifact: true,
     },
     ...GLOBAL_CONTEXT_VARIABLES,
   ];
   const seen = new Set(variables.map((variable) => variable.templatePath));
   const upstreamIds = getUpstreamNodeIds(selectedNodeId, edges);
 
+  const pushNodeOutput = (
+    node: Node,
+    outputName: string,
+    sectionFallback: string,
+    producesArtifact: boolean,
+  ) => {
+    const templatePath = toAgentTemplatePath(outputName);
+    if (seen.has(templatePath)) return;
+    seen.add(templatePath);
+    variables.push({
+      label: templatePath,
+      templatePath,
+      section: node.data.title || node.data.label || sectionFallback,
+      producesArtifact,
+    });
+  };
+
   for (const node of nodes) {
     if (!upstreamIds.has(node.id)) continue;
 
-    if (node.type === 'agent') {
-      const defaultOutputTemplatePath = toAgentTemplatePath(
+    if (node.type === 'agent' || node.type === 'code') {
+      // Agent and code nodes both expose `node_<id>_output` and an optional
+      // `output_variable`, but only code-node outputs resolve to artifact
+      // references in the engine — agent outputs are LLM TEXT. So both feed the
+      // prompt-variable popover, while only code outputs are offered as
+      // Documents (see toDocumentVariableOptions / producesArtifact).
+      const producesArtifact = node.type === 'code';
+      const sectionFallback = node.type === 'code' ? 'Code' : 'Agent';
+      pushNodeOutput(
+        node,
         `node_${node.id}_output`,
+        sectionFallback,
+        producesArtifact,
       );
-      if (!seen.has(defaultOutputTemplatePath)) {
-        seen.add(defaultOutputTemplatePath);
-        variables.push({
-          label: defaultOutputTemplatePath,
-          templatePath: defaultOutputTemplatePath,
-          section: node.data.title || node.data.label || 'Agent',
-        });
-      }
 
       const outputVariable = String(
         node.data?.config?.output_variable || '',
       ).trim();
       if (outputVariable) {
-        const templatePath = toAgentTemplatePath(outputVariable);
-        if (!seen.has(templatePath)) {
-          seen.add(templatePath);
-          variables.push({
-            label: templatePath,
-            templatePath,
-            section: node.data.title || node.data.label || 'Agent',
-          });
-        }
+        pushNodeOutput(node, outputVariable, sectionFallback, producesArtifact);
       }
     }
 
diff --git a/frontend/src/agents/workflow/documentConfig.test.ts b/frontend/src/agents/workflow/documentConfig.test.ts
index a4ce2ce8..7066f873 100644
--- a/frontend/src/agents/workflow/documentConfig.test.ts
+++ b/frontend/src/agents/workflow/documentConfig.test.ts
@@ -88,27 +88,33 @@ describe('stripAgentPrefix', () => {
 });
 
 describe('toDocumentVariableOptions', () => {
-  const make = (templatePath: string): WorkflowVariable => ({
+  const make = (
+    templatePath: string,
+    producesArtifact = false,
+  ): WorkflowVariable => ({
     label: templatePath,
     templatePath,
     section: 'x',
+    producesArtifact,
   });
 
-  it('keeps input_documents and upstream outputs as bare names', () => {
+  it('offers only artifact-producing variables (input_documents, code outputs)', () => {
     const options = toDocumentVariableOptions([
-      make('agent.input_documents'),
-      make('agent.wire_doc'),
-      make("agent['weird name']"),
+      make('agent.input_documents', true),
+      make('agent.node_code1_output', true),
+      make("agent['weird name']", true),
     ]);
     expect(options).toEqual([
       { value: 'input_documents', label: 'input_documents' },
-      { value: 'wire_doc', label: 'wire_doc' },
+      { value: 'node_code1_output', label: 'node_code1_output' },
       { value: 'weird name', label: 'weird name' },
     ]);
   });
 
-  it('excludes query, chat_history and global context variables', () => {
+  it('excludes plain agent/state TEXT outputs that are not artifact refs', () => {
     const options = toDocumentVariableOptions([
+      make('agent.node_agent1_output'),
+      make('agent.some_state_var'),
       make('agent.query'),
       make('agent.chat_history'),
       make('source.content'),
@@ -120,8 +126,8 @@ describe('toDocumentVariableOptions', () => {
 
   it('deduplicates by bare name', () => {
     const options = toDocumentVariableOptions([
-      make('agent.wire_doc'),
-      make('agent.wire_doc'),
+      make('agent.wire_doc', true),
+      make('agent.wire_doc', true),
     ]);
     expect(options).toEqual([{ value: 'wire_doc', label: 'wire_doc' }]);
   });
diff --git a/frontend/src/agents/workflow/documentConfig.ts b/frontend/src/agents/workflow/documentConfig.ts
index c524c07e..191d76fe 100644
--- a/frontend/src/agents/workflow/documentConfig.ts
+++ b/frontend/src/agents/workflow/documentConfig.ts
@@ -59,23 +59,25 @@ export function stripAgentPrefix(templatePath: string): string {
   return trimmed;
 }
 
-/** Multiselect options of upstream document variables, stored as bare names. */
+/**
+ * Multiselect options of upstream document variables, stored as bare names.
+ *
+ * Only variables that carry artifact references at run time are offered
+ * (`producesArtifact`): the uploaded `input_documents` and code-node outputs.
+ * Plain agent/state TEXT outputs are excluded — the engine can't resolve them
+ * to artifacts, so picking one appends the literal variable name and the node
+ * hard-fails. The distinction is assigned upstream in `extractUpstreamVariables`
+ * keyed on the producing node type.
+ */
 export function toDocumentVariableOptions(
   variables: WorkflowVariable[],
 ): { value: string; label: string }[] {
   const options: { value: string; label: string }[] = [];
   const seen = new Set<string>();
   for (const variable of variables) {
-    const path = variable.templatePath;
-    const isInputDocuments = path === 'agent.input_documents';
-    const isAgentOutput =
-      path.startsWith('agent.') &&
-      path !== 'agent.query' &&
-      path !== 'agent.chat_history';
-    const isAgentBracket = path.startsWith('agent[');
-    if (!isInputDocuments && !isAgentOutput && !isAgentBracket) continue;
+    if (!variable.producesArtifact) continue;
 
-    const bareName = stripAgentPrefix(path);
+    const bareName = stripAgentPrefix(variable.templatePath);
     if (!bareName || seen.has(bareName)) continue;
     seen.add(bareName);
     options.push({ value: bareName, label: bareName });
diff --git a/frontend/src/agents/workflow/workflowPreviewSlice.test.ts b/frontend/src/agents/workflow/workflowPreviewSlice.test.ts
index ebadefc3..c329652c 100644
--- a/frontend/src/agents/workflow/workflowPreviewSlice.test.ts
+++ b/frontend/src/agents/workflow/workflowPreviewSlice.test.ts
@@ -4,9 +4,12 @@ import { Attachment } from '../../upload/uploadSlice';
 import reducer, {
   addQuery,
   collectCompletedAttachmentIds,
+  collectRunAttachmentIds,
+  previewSendBlockReason,
   resetWorkflowPreview,
   setPreviewOpen,
   setWorkflowRunId,
+  UNSAVED_DRAFT_ATTACHMENTS_MESSAGE,
 } from './workflowPreviewSlice';
 
 const seedState = () => reducer(undefined, { type: '@@INIT' });
@@ -39,6 +42,40 @@ describe('collectCompletedAttachmentIds', () => {
   });
 });
 
+describe('collectRunAttachmentIds', () => {
+  const completed = [att({ id: 'done', status: 'completed' })];
+
+  it('returns completed ids for a saved workflow', () => {
+    expect(collectRunAttachmentIds(completed, 'wf-1')).toEqual(['done']);
+  });
+
+  it('returns [] for an unsaved draft so uploads are neither sent nor cleared', () => {
+    expect(collectRunAttachmentIds(completed, null)).toEqual([]);
+    expect(collectRunAttachmentIds(completed, undefined)).toEqual([]);
+    expect(collectRunAttachmentIds(completed, '')).toEqual([]);
+  });
+});
+
+describe('previewSendBlockReason', () => {
+  it('blocks an unsaved draft that has completed attachments', () => {
+    expect(previewSendBlockReason(null, true)).toBe(
+      UNSAVED_DRAFT_ATTACHMENTS_MESSAGE,
+    );
+    expect(previewSendBlockReason(undefined, true)).toBe(
+      UNSAVED_DRAFT_ATTACHMENTS_MESSAGE,
+    );
+    expect(previewSendBlockReason('', true)).toBe(
+      UNSAVED_DRAFT_ATTACHMENTS_MESSAGE,
+    );
+  });
+
+  it('allows a saved workflow, or an unsaved draft with no attachments', () => {
+    expect(previewSendBlockReason('wf-1', true)).toBeNull();
+    expect(previewSendBlockReason('wf-1', false)).toBeNull();
+    expect(previewSendBlockReason(null, false)).toBeNull();
+  });
+});
+
 describe('setWorkflowRunId', () => {
   it('stores the run id on the addressed query', () => {
     let state = seedState();
diff --git a/frontend/src/agents/workflow/workflowPreviewSlice.ts b/frontend/src/agents/workflow/workflowPreviewSlice.ts
index 061a4178..b7888885 100644
--- a/frontend/src/agents/workflow/workflowPreviewSlice.ts
+++ b/frontend/src/agents/workflow/workflowPreviewSlice.ts
@@ -76,6 +76,40 @@ export function collectCompletedAttachmentIds(
     .map((att) => att.id);
 }
 
+/** User-facing message shown when a Preview send is blocked. */
+export const UNSAVED_DRAFT_ATTACHMENTS_MESSAGE =
+  'Save the workflow before attaching documents in Preview.';
+
+/**
+ * Reason a Preview send must be blocked, or null when it may proceed. An
+ * unsaved draft has no persisted ``workflow_id``, so the backend can't bridge
+ * uploaded documents into the run — it would execute with the docs silently
+ * dropped. Block instead, keeping the uploads for a retry after the save.
+ */
+export function previewSendBlockReason(
+  workflowId: string | null | undefined,
+  hasCompletedAttachment: boolean,
+): string | null {
+  if (!workflowId && hasCompletedAttachment) {
+    return UNSAVED_DRAFT_ATTACHMENTS_MESSAGE;
+  }
+  return null;
+}
+
+/**
+ * Completed attachment ids eligible for a run. Attachments only survive as
+ * run-scoped artifacts when the run persists a ``workflow_runs`` row, which
+ * needs a saved ``workflow_id``; on an unsaved draft none are eligible (the UI
+ * blocks that send upstream), so they are neither sent nor cleared.
+ */
+export function collectRunAttachmentIds(
+  attachments: Attachment[],
+  workflowId: string | null | undefined,
+): string[] {
+  if (!workflowId) return [];
+  return collectCompletedAttachmentIds(attachments);
+}
+
 export const fetchWorkflowPreviewAnswer = createAsyncThunk<
   void,
   {
@@ -100,8 +134,12 @@ export const fetchWorkflowPreviewAnswer = createAsyncThunk<
     if (state.preference) {
       // Doc-driven workflows read uploaded files as run-scoped artifacts; pass
       // the completed attachment ids so the backend bridges them (user-scoped).
-      const attachmentIds = collectCompletedAttachmentIds(
+      // Gated on a saved workflow_id: an unsaved draft can't bridge them, so we
+      // neither send nor clear the uploads (the Preview UI blocks that send and
+      // keeps them for a retry after saving).
+      const attachmentIds = collectRunAttachmentIds(
         state.upload.attachments,
+        workflowId,
       );
       if (attachmentIds.length > 0) dispatch(clearAttachments());
 
diff --git a/frontend/src/api/endpoints.ts b/frontend/src/api/endpoints.ts
index 9b634cdd..8f23f896 100644
--- a/frontend/src/api/endpoints.ts
+++ b/frontend/src/api/endpoints.ts
@@ -114,10 +114,17 @@ const endpoints = {
       `/api/artifacts/${artifactId}`,
     LIST_WORKFLOW_RUN_ARTIFACTS: (workflowRunId: string) =>
       `/api/artifacts?workflow_run_id=${encodeURIComponent(workflowRunId)}`,
-    DOWNLOAD_ARTIFACT: (artifactId: string, version?: number) =>
-      version != null
-        ? `/api/artifacts/${artifactId}/download?version=${version}`
-        : `/api/artifacts/${artifactId}/download`,
+    DOWNLOAD_ARTIFACT: (
+      artifactId: string,
+      version?: number,
+      disposition?: string,
+    ) => {
+      const params = new URLSearchParams();
+      if (version != null) params.set('version', String(version));
+      if (disposition) params.set('disposition', disposition);
+      const qs = params.toString();
+      return `/api/artifacts/${artifactId}/download${qs ? `?${qs}` : ''}`;
+    },
     RESTORE_ARTIFACT: (artifactId: string) =>
       `/api/artifacts/${artifactId}/restore`,
     WORKFLOWS: '/api/workflows',
diff --git a/frontend/src/api/services/userService.ts b/frontend/src/api/services/userService.ts
index 5e283034..bd65491c 100644
--- a/frontend/src/api/services/userService.ts
+++ b/frontend/src/api/services/userService.ts
@@ -339,8 +339,16 @@ const userService = {
     artifactId: string,
     token: string | null,
     version?: number,
+    disposition?: 'url',
   ): Promise<Response> =>
-    apiClient.get(endpoints.USER.DOWNLOAD_ARTIFACT(artifactId, version), token),
+    apiClient.get(
+      endpoints.USER.DOWNLOAD_ARTIFACT(artifactId, version, disposition),
+      token,
+      // ?disposition=url asks the s3 strategy to return the presigned URL as
+      // JSON (for a top-level navigation) instead of a CORS-blocked 302; the
+      // Accept header is the same opt-in via content negotiation.
+      disposition === 'url' ? { Accept: 'application/json' } : {},
+    ),
   restoreArtifactVersion: (
     artifactId: string,
     version: number,
diff --git a/frontend/src/components/DocumentArtifactView.tsx b/frontend/src/components/DocumentArtifactView.tsx
index d2edc6fe..61db9692 100644
--- a/frontend/src/components/DocumentArtifactView.tsx
+++ b/frontend/src/components/DocumentArtifactView.tsx
@@ -279,6 +279,7 @@ export default function DocumentArtifactView({
         artifact.id,
         token,
         selectedVersion,
+        'url',
       );
       const ok = await triggerResponseDownload(response, filename);
       if (!ok && isMountedRef.current) setActionError('Download failed');
diff --git a/frontend/src/components/artifactViewUtils.test.ts b/frontend/src/components/artifactViewUtils.test.ts
index 33f6794a..5f08d253 100644
--- a/frontend/src/components/artifactViewUtils.test.ts
+++ b/frontend/src/components/artifactViewUtils.test.ts
@@ -1,4 +1,4 @@
-import { describe, expect, it } from 'vitest';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
 
 import {
   buildPreviewDocument,
@@ -12,7 +12,9 @@ import {
   PREVIEW_CSP,
   PREVIEW_CSP_MERMAID,
   previewModeForKind,
+  readPresignedUrlEnvelope,
   sortVersionsDesc,
+  triggerResponseDownload,
   type ArtifactVersion,
 } from './artifactViewUtils';
 
@@ -234,6 +236,104 @@ describe('buildPreviewDocument', () => {
   });
 });
 
+describe('readPresignedUrlEnvelope', () => {
+  const json = (body: unknown, contentType = 'application/json') =>
+    new Response(JSON.stringify(body), {
+      headers: { 'Content-Type': contentType },
+    });
+
+  it('extracts the presigned url from the s3 JSON envelope', async () => {
+    const url = await readPresignedUrlEnvelope(
+      json({ success: true, url: 'https://signed.example/x?sig=1' }),
+    );
+    expect(url).toBe('https://signed.example/x?sig=1');
+  });
+
+  it('returns null for a byte stream and leaves the body readable', async () => {
+    const res = new Response('<p>hi</p>', {
+      headers: { 'Content-Type': 'text/html' },
+    });
+    expect(await readPresignedUrlEnvelope(res)).toBeNull();
+    expect(await res.text()).toBe('<p>hi</p>');
+  });
+
+  it('returns null for a JSON artifact whose bytes are not the envelope', async () => {
+    const res = json({ foo: 1 });
+    expect(await readPresignedUrlEnvelope(res)).toBeNull();
+    // clone() was used to peek, so the original body is still consumable.
+    expect(await res.json()).toEqual({ foo: 1 });
+  });
+
+  it('rejects a non-http(s) url in the envelope', async () => {
+    expect(
+      await readPresignedUrlEnvelope(
+        json({ success: true, url: 'javascript:alert(1)' }),
+      ),
+    ).toBeNull();
+    expect(
+      await readPresignedUrlEnvelope(json({ success: true, url: '' })),
+    ).toBeNull();
+  });
+});
+
+describe('triggerResponseDownload', () => {
+  let anchors: HTMLAnchorElement[];
+
+  beforeEach(() => {
+    anchors = [];
+    vi.spyOn(URL, 'createObjectURL').mockReturnValue('blob:mock');
+    vi.spyOn(URL, 'revokeObjectURL').mockImplementation(() => undefined);
+    const orig = document.createElement.bind(document);
+    vi.spyOn(document, 'createElement').mockImplementation((tag: string) => {
+      const el = orig(tag);
+      if (String(tag).toLowerCase() === 'a') {
+        vi.spyOn(el as HTMLAnchorElement, 'click').mockImplementation(
+          () => undefined,
+        );
+        anchors.push(el as HTMLAnchorElement);
+      }
+      return el;
+    });
+  });
+
+  afterEach(() => {
+    vi.restoreAllMocks();
+  });
+
+  it('returns false for a non-OK response', async () => {
+    const res = new Response('', { status: 500 });
+    expect(await triggerResponseDownload(res, 'f.bin')).toBe(false);
+  });
+
+  it('navigates top-level to the presigned url for the s3 envelope', async () => {
+    const res = new Response(
+      JSON.stringify({ success: true, url: 'https://signed.example/x' }),
+      { headers: { 'Content-Type': 'application/json' } },
+    );
+    expect(await triggerResponseDownload(res, 'f.bin')).toBe(true);
+    // No blob path for the presigned envelope.
+    expect(URL.createObjectURL).not.toHaveBeenCalled();
+    const anchor = anchors.at(-1)!;
+    expect(anchor.href).toBe('https://signed.example/x');
+    expect(anchor.target).toBe('_blank');
+    expect(anchor.click).toHaveBeenCalled();
+  });
+
+  it('saves streamed bytes via a blob url for the backend strategy', async () => {
+    const res = new Response('BINARY', {
+      headers: {
+        'Content-Type': 'application/octet-stream',
+        'Content-Disposition': 'attachment; filename="deck.pptx"',
+      },
+    });
+    expect(await triggerResponseDownload(res, 'fallback.bin')).toBe(true);
+    expect(URL.createObjectURL).toHaveBeenCalled();
+    const anchor = anchors.at(-1)!;
+    expect(anchor.download).toBe('deck.pptx');
+    expect(URL.revokeObjectURL).toHaveBeenCalledWith('blob:mock');
+  });
+});
+
 describe('filenameFromContentDisposition', () => {
   it('reads the plain quoted filename', () => {
     expect(
diff --git a/frontend/src/components/artifactViewUtils.ts b/frontend/src/components/artifactViewUtils.ts
index b5b12ee6..75b7420b 100644
--- a/frontend/src/components/artifactViewUtils.ts
+++ b/frontend/src/components/artifactViewUtils.ts
@@ -258,14 +258,71 @@ export function filenameFromContentDisposition(
 }
 
 /**
- * Turn a fetched download Response into a browser file-save, honoring
- * `Content-Disposition` when present. Returns false if the response is not OK.
+ * Detect the `URL_STRATEGY=s3` presigned-URL envelope and return its `url`, or
+ * null when the response is a byte stream (`URL_STRATEGY=backend`). The download
+ * endpoint returns `{ success: true, url }` as JSON only when asked via
+ * `?disposition=url`; a plain byte stream (even a JSON artifact's own bytes)
+ * carries the file's own content-type and never this exact shape. The response
+ * is `clone()`d before reading so the caller can still consume the original body
+ * as bytes when this is not the envelope. The `url` must be an absolute http(s)
+ * URL to further guard against a JSON artifact coincidentally matching.
+ */
+export async function readPresignedUrlEnvelope(
+  response: Response,
+): Promise<string | null> {
+  const contentType = response.headers.get('Content-Type') ?? '';
+  if (!contentType.toLowerCase().includes('application/json')) return null;
+  const data = await response
+    .clone()
+    .json()
+    .catch(() => null);
+  if (
+    data &&
+    typeof data === 'object' &&
+    (data as { success?: unknown }).success === true &&
+    typeof (data as { url?: unknown }).url === 'string' &&
+    /^https?:\/\//i.test((data as { url: string }).url)
+  ) {
+    return (data as { url: string }).url;
+  }
+  return null;
+}
+
+/** Trigger a top-level browser navigation to `url` (opens a new tab). */
+function triggerUrlNavigation(url: string, fallbackName: string): void {
+  const link = document.createElement('a');
+  link.href = url;
+  // `download` is ignored for cross-origin URLs, so the object's own
+  // Content-Disposition drives the save; open a new tab so the app isn't
+  // navigated away if the object renders inline.
+  link.download = fallbackName;
+  link.target = '_blank';
+  link.rel = 'noopener noreferrer';
+  document.body.appendChild(link);
+  link.click();
+  document.body.removeChild(link);
+}
+
+/**
+ * Turn a fetched download Response into a browser file-save. Supports both URL
+ * strategies from a single call: the s3 strategy returns a presigned-URL JSON
+ * envelope (fetching its bytes cross-origin would be CORS-blocked, so navigate
+ * to it top-level), while the backend strategy streams the bytes (saved via a
+ * blob URL, honoring `Content-Disposition`). Returns false if the response is
+ * not OK, or is a malformed JSON envelope with no usable url.
  */
 export async function triggerResponseDownload(
   response: Response,
   fallbackName: string,
 ): Promise<boolean> {
   if (!response.ok) return false;
+
+  const presignedUrl = await readPresignedUrlEnvelope(response);
+  if (presignedUrl) {
+    triggerUrlNavigation(presignedUrl, fallbackName);
+    return true;
+  }
+
   const name =
     filenameFromContentDisposition(
       response.headers.get('Content-Disposition'),
diff --git a/frontend/src/components/useArtifactBytes.ts b/frontend/src/components/useArtifactBytes.ts
index afa62f06..d58b77e7 100644
--- a/frontend/src/components/useArtifactBytes.ts
+++ b/frontend/src/components/useArtifactBytes.ts
@@ -4,6 +4,7 @@ import userService from '../api/services/userService';
 import {
   type BytesPreviewMode,
   MAX_INLINE_TEXT_BYTES,
+  readPresignedUrlEnvelope,
 } from './artifactViewUtils';
 
 type BytesState =
@@ -49,12 +50,22 @@ export function useArtifactBytes(
     };
 
     userService
-      .downloadArtifact(artifactId, token, version)
+      .downloadArtifact(artifactId, token, version, 'url')
       .then(async (response: Response) => {
         if (!response.ok) throw new Error('download failed');
 
+        // Under URL_STRATEGY=s3 the endpoint returns a presigned URL; fetch the
+        // bytes directly from it (a bucket without a CORS grant for the app
+        // origin blocks this read → caught below → download-card fallback). The
+        // backend strategy streams the bytes here, so this is a no-op for it.
+        const presignedUrl = await readPresignedUrlEnvelope(response);
+        const bytesResponse = presignedUrl
+          ? await fetch(presignedUrl)
+          : response;
+        if (!bytesResponse.ok) throw new Error('bytes fetch failed');
+
         if (mode === 'image') {
-          const blob = await response.blob();
+          const blob = await bytesResponse.blob();
           if (cancelled) return;
           revoke();
           const url = URL.createObjectURL(blob);
@@ -64,7 +75,7 @@ export function useArtifactBytes(
         }
 
         // html / svg / text families render as a (bounded) string.
-        const blob = await response.blob();
+        const blob = await bytesResponse.blob();
         if (cancelled) return;
         if (blob.size > MAX_INLINE_TEXT_BYTES) throw new Error('too large');
         const text = await blob.text();
diff --git a/tests/agents/test_workflow_agent_graph.py b/tests/agents/test_workflow_agent_graph.py
index f5f17c3a..f2726a3c 100644
--- a/tests/agents/test_workflow_agent_graph.py
+++ b/tests/agents/test_workflow_agent_graph.py
@@ -7,8 +7,11 @@ against the ephemeral ``pg_conn`` fixture. Agent construction is bypassed via
 
 from contextlib import contextmanager
 from datetime import datetime, timezone
+from types import SimpleNamespace
 from unittest.mock import MagicMock, patch
 
+import pytest
+
 
 
 def _make_agent(*, workflow_id=None, workflow=None, workflow_owner=None,
@@ -334,3 +337,48 @@ class TestGen:
         assert results == [
             {"type": "error", "error": "Failed to load workflow configuration."}
         ]
+
+
+class TestAgentNodeApprovalPause:
+    """A node agent whose tool pauses for approval must fail the node visibly, not emit empty output."""
+
+    def test_tool_calls_pending_raises_clear_error(self, monkeypatch):
+        from application.agents.workflows import workflow_engine as we
+        from application.agents.workflows.schemas import (
+            NodeType,
+            Workflow,
+            WorkflowGraph,
+            WorkflowNode,
+        )
+        from application.agents.workflows.workflow_engine import WorkflowEngine
+
+        # An ephemeral node agent whose LLM handler yields the pause signal and ends,
+        # emitting no "answer". Previously the engine dropped it and the node completed
+        # with empty output (or raised a confusing "Structured output was expected").
+        class _PendingAgent:
+            attachments = None
+
+            def gen(self, prompt):
+                yield {"type": "tool_calls_pending", "data": {"pending_tool_calls": [{"id": "x"}]}}
+
+        monkeypatch.setattr(
+            we.WorkflowNodeAgentFactory, "create", lambda **kw: _PendingAgent()
+        )
+
+        graph = WorkflowGraph(workflow=Workflow(name="Approval Pause"), nodes=[], edges=[])
+        agent = SimpleNamespace(
+            endpoint="stream", llm_name="openai", model_id="gpt-4o-mini", api_key="k",
+            chat_history=[], user="u", decoded_token={"sub": "u"},
+        )
+        engine = WorkflowEngine(
+            graph, agent, workflow_run_id="22222222-2222-2222-2222-222222222222"
+        )
+
+        node = WorkflowNode(
+            id="agent_1", workflow_id="wf-1", type=NodeType.AGENT, title="Extractor",
+            position={"x": 0, "y": 0},
+            config={"llm_name": "openai", "system_prompt": "s", "tools": ["t"]},
+        )
+
+        with pytest.raises(ValueError, match="requires approval"):
+            list(engine._execute_agent_node(node))
diff --git a/tests/agents/test_workflow_code_node.py b/tests/agents/test_workflow_code_node.py
index 5f693961..a1961ba7 100644
--- a/tests/agents/test_workflow_code_node.py
+++ b/tests/agents/test_workflow_code_node.py
@@ -396,6 +396,57 @@ def test_materialize_code_inputs_rejects_oversize(monkeypatch):
     assert manager.put_files == []  # nothing staged
 
 
+def test_materialize_code_inputs_dedupes_same_filename(monkeypatch):
+    """Two code-node inputs sharing a filename stage to DISTINCT inputs/ paths (no clobber)."""
+    from contextlib import contextmanager
+
+    monkeypatch.setattr(
+        "application.agents.tools.artifact_ref.resolve_artifact_id",
+        lambda repo, raw, **k: str(raw),
+    )
+
+    class _Repo:
+        def __init__(self, conn):
+            pass
+
+        def get_artifact_in_parent(self, artifact_id, *, workflow_run_id=None, conversation_id=None):
+            return {"id": artifact_id, "current_version": 1}
+
+        def get_version(self, artifact_id, version):
+            # Same filename, distinct stored bytes per input.
+            return {"filename": "data.csv", "storage_path": f"p/{artifact_id}.csv"}
+
+    @contextmanager
+    def _readonly():
+        yield object()
+
+    class _Storage:
+        def get_file(self, path):
+            import io
+
+            return io.BytesIO(path.encode())
+
+    monkeypatch.setattr(
+        "application.storage.db.repositories.artifacts.ArtifactsRepository", _Repo
+    )
+    monkeypatch.setattr("application.storage.db.session.db_readonly", _readonly)
+    monkeypatch.setattr(
+        "application.storage.storage_creator.StorageCreator.get_storage",
+        staticmethod(lambda: _Storage()),
+    )
+
+    engine = _engine()
+    manager = _FakeManager(_Result(ok=True, stdout="ok"))
+    loaded = engine._materialize_code_inputs(
+        manager, engine._session_id(), ["art-a", "art-b"], "user-code"
+    )
+
+    assert loaded == ["inputs/data.csv", "inputs/data-2.csv"]
+    staged = dict(manager.put_files)
+    assert set(staged) == {"inputs/data.csv", "inputs/data-2.csv"}
+    assert staged["inputs/data.csv"] != staged["inputs/data-2.csv"]  # no clobber
+
+
 # ---------------------------------------------------------------------------
 # Pass-by-reference: survives serialization + CEL branches on the metadata.
 # ---------------------------------------------------------------------------
diff --git a/tests/agents/tools/test_artifact_ref.py b/tests/agents/tools/test_artifact_ref.py
index a7d2b505..429ffb4c 100644
--- a/tests/agents/tools/test_artifact_ref.py
+++ b/tests/agents/tools/test_artifact_ref.py
@@ -102,3 +102,62 @@ def test_resolve_ref_under_workflow_run_parent():
     repo = _FakeRepo(mapping={2: target}, run="run-9")
     assert resolve_artifact_id(repo, "A2", workflow_run_id="run-9") == target
     assert repo.calls == [(2, None, "run-9")]
+
+
+# ---------------------------------------------------------------------------
+# Stable per-parent ref_seq: a ref no longer re-points after an earlier delete
+# ---------------------------------------------------------------------------
+class _SeqRepo:
+    """Fake repo backing resolve_id_by_ref_seq (stable) with the legacy positional fallback."""
+
+    def __init__(self, *, seq_map=None, pos_map=None, conv=None, run=None):
+        self.seq_map = seq_map or {}
+        self.pos_map = pos_map or {}
+        self.conv = conv
+        self.run = run
+        self.seq_calls = []
+        self.pos_calls = []
+
+    def _in_scope(self, conversation_id, workflow_run_id) -> bool:
+        if self.conv is not None and conversation_id != self.conv:
+            return False
+        if self.run is not None and workflow_run_id != self.run:
+            return False
+        return True
+
+    def resolve_id_by_ref_seq(self, seq, *, conversation_id=None, workflow_run_id=None):
+        self.seq_calls.append((seq, conversation_id, workflow_run_id))
+        if not self._in_scope(conversation_id, workflow_run_id):
+            return None
+        return self.seq_map.get(seq)
+
+    def artifact_id_at_position(self, n, *, conversation_id=None, workflow_run_id=None):
+        self.pos_calls.append((n, conversation_id, workflow_run_id))
+        if not self._in_scope(conversation_id, workflow_run_id):
+            return None
+        return self.pos_map.get(n)
+
+
+def test_ref_resolves_by_stable_seq_not_shifted_position():
+    b, c = str(uuid.uuid4()), str(uuid.uuid4())
+    # After deleting the earlier A(seq 1): ref_seq keeps B=2/C=3; positions shifted to B=1/C=2.
+    repo = _SeqRepo(seq_map={2: b, 3: c}, pos_map={1: b, 2: c}, conv="conv-1")
+    assert resolve_artifact_id(repo, "A2", conversation_id="conv-1") == b
+    assert resolve_artifact_id(repo, "A3", conversation_id="conv-1") == c
+    # The stable ref_seq hit; the (shifted) positional fallback was never consulted.
+    assert repo.pos_calls == []
+
+
+def test_ref_falls_back_to_position_for_legacy_rows():
+    legacy = str(uuid.uuid4())
+    # Legacy row has no ref_seq -> resolve_id_by_ref_seq misses -> positional fallback resolves it.
+    repo = _SeqRepo(seq_map={}, pos_map={1: legacy}, conv="conv-1")
+    assert resolve_artifact_id(repo, "A1", conversation_id="conv-1") == legacy
+    assert repo.seq_calls == [(1, "conv-1", None)]
+    assert repo.pos_calls == [(1, "conv-1", None)]
+
+
+def test_ref_seq_does_not_cross_parents():
+    b = str(uuid.uuid4())
+    repo = _SeqRepo(seq_map={2: b}, pos_map={2: b}, conv="conv-1")
+    assert resolve_artifact_id(repo, "A2", conversation_id="conv-OTHER") is None
diff --git a/tests/agents/tools/test_read_document_tool.py b/tests/agents/tools/test_read_document_tool.py
index f4160b58..31192687 100644
--- a/tests/agents/tools/test_read_document_tool.py
+++ b/tests/agents/tools/test_read_document_tool.py
@@ -302,3 +302,55 @@ def test_malformed_json_schema_rejected_before_enqueue(monkeypatch):
     )
     out = _tool().execute_action("read_document", input=_ART_ID, json_schema={"properties": {}}, persist=False)
     assert out["status"] == "error" and "invalid json_schema" in out["error"]
+
+
+# ---------------------------------------------------------------------------
+# Worker-context: parse INLINE inside a Celery worker; dispatch from the web
+# ---------------------------------------------------------------------------
+@pytest.mark.unit
+def test_dispatch_inline_when_in_worker(monkeypatch):
+    _stub_repo(monkeypatch, found=True, conv="conv-1", run=None)
+    # Inside a worker current_task is truthy -> parse inline, never enqueue (else the
+    # parsing queue self-deadlocks the worker that also serves it).
+    monkeypatch.setattr(rd, "current_task", object())
+
+    import application.api.user.tasks as tasks
+    monkeypatch.setattr(
+        tasks.parse_document, "apply_async",
+        lambda *a, **k: (_ for _ in ()).throw(AssertionError("must not enqueue inside a worker")),
+    )
+
+    import application.worker as worker
+    called: Dict[str, Any] = {}
+
+    def _fake_run(artifact_id, parent, user_id, options):
+        called["args"] = (artifact_id, parent, user_id)
+        called["options"] = options
+        return {"status": "ok", "content": "inline", "truncated": False}
+
+    monkeypatch.setattr(worker, "run_parse_document", _fake_run)
+
+    out = _tool().execute_action("read_document", input=_ART_ID, persist=False)
+    assert out["status"] == "ok" and out["content"] == "inline"
+    # Same auth re-resolution + parent shape as the dispatch path.
+    assert called["args"] == (_ART_ID, {"conversation_id": "conv-1"}, "u-1")
+    assert called["options"]["persist"] is False
+
+
+@pytest.mark.unit
+def test_dispatch_enqueues_when_not_in_worker(monkeypatch):
+    _stub_repo(monkeypatch, found=True, conv="conv-1", run=None)
+    # Web process: current_task falsy -> dispatch to the parsing queue, never inline.
+    monkeypatch.setattr(rd, "current_task", None)
+    captured = _patch_task(monkeypatch, payload={"status": "ok", "content": "queued", "truncated": False})
+
+    import application.worker as worker
+    monkeypatch.setattr(
+        worker, "run_parse_document",
+        lambda *a, **k: (_ for _ in ()).throw(AssertionError("web path must dispatch, not inline")),
+    )
+
+    out = _tool().execute_action("read_document", input=_ART_ID, persist=False)
+    assert out["status"] == "ok" and out["content"] == "queued"
+    assert captured["args"][0] == _ART_ID
+    assert captured["args"][1] == {"conversation_id": "conv-1"}
diff --git a/tests/api/user/test_artifacts_routes.py b/tests/api/user/test_artifacts_routes.py
index 23efd0b1..a796b0d2 100644
--- a/tests/api/user/test_artifacts_routes.py
+++ b/tests/api/user/test_artifacts_routes.py
@@ -580,6 +580,36 @@ class TestDownloadArtifact:
         assert resp.status_code == 302
         assert resp.headers["Location"].startswith("https://signed.example/")
 
+    def test_s3_strategy_disposition_url_returns_json(
+        self, _patch_db, flask_app, token_owner, monkeypatch
+    ):
+        # ?disposition=url opts into a JSON envelope (for a top-level browser
+        # navigation) instead of the CORS-blocked cross-origin 302.
+        from application.api.user.artifacts.routes import DownloadArtifact
+
+        art = self._seed(_patch_db)
+        storage = _FakeStorage(b"unused")
+        monkeypatch.setattr(
+            "application.api.user.artifacts.routes.StorageCreator.get_storage",
+            lambda: storage,
+        )
+        monkeypatch.setattr(
+            "application.api.user.artifacts.routes.settings.URL_STRATEGY",
+            "s3",
+            raising=False,
+        )
+
+        resp = _call(
+            flask_app,
+            DownloadArtifact,
+            art["id"],
+            token=token_owner,
+            query={"disposition": "url"},
+        )
+        assert resp.status_code == 200
+        assert resp.json["success"] is True
+        assert resp.json["url"].startswith("https://signed.example/")
+
     def test_stranger_denied(self, _patch_db, flask_app, monkeypatch):
         from application.api.user.artifacts.routes import DownloadArtifact
 
diff --git a/tests/parser/test_document_reader.py b/tests/parser/test_document_reader.py
index 06019a7b..248a6bd8 100644
--- a/tests/parser/test_document_reader.py
+++ b/tests/parser/test_document_reader.py
@@ -61,19 +61,21 @@ def test_markdown_output_default():
 
 
 @pytest.mark.unit
-def test_max_chars_truncates_and_flags():
+def test_max_chars_does_not_truncate_parse_output():
+    # max_chars now bounds only the VIEW (bound_parse_payload); parse returns the FULL text
+    # so the persisted artifact is the complete parse.
     out = parse_document_bytes(("A" * 100).encode(), "note.txt", output="text", max_chars=10, include_tables=False)
-    assert out["truncated"] is True
-    assert len(out["content"]) == 10
+    assert out["content"] == "A" * 100
+    assert out["truncated"] is False
 
 
 @pytest.mark.unit
-def test_default_window_truncates_large_text():
-    big = ("A" * (dr._TEXT_MAX_BYTES * 3)).encode()
-    out = parse_document_bytes(big, "note.txt", output="text", include_tables=False)
-    assert out["truncated"] is True
-    assert "...[truncated" in out["content"]
-    assert len(out["content"].encode("utf-8")) <= dr._TEXT_MAX_BYTES + 64
+def test_large_text_is_full_in_parse_output():
+    # The default head+tail window moved to bound_parse_payload; parse keeps the full text.
+    big = "A" * (dr._TEXT_MAX_BYTES * 3)
+    out = parse_document_bytes(big.encode(), "note.txt", output="text", include_tables=False)
+    assert out["content"] == big
+    assert out["truncated"] is False
 
 
 # ---------------------------------------------------------------------------
@@ -348,3 +350,28 @@ def test_bound_parse_payload_keeps_structured_for_validation():
     out = bound_parse_payload({"output": "structured", "content": "# ok", "structured": structured})
     # structured must survive so the tool's json_schema validation can run on it.
     assert out["structured"] == structured
+
+
+@pytest.mark.unit
+def test_bound_parse_payload_max_chars_bounds_view_only():
+    # max_chars caps the returned view; the parse itself is unaffected (see parse tests above).
+    content = "A" * 100
+    out = bound_parse_payload({"output": "text", "content": content}, max_chars=10)
+    assert out["content"] == "A" * 10
+    assert out["truncated"] is True
+
+
+@pytest.mark.unit
+def test_bound_parse_payload_default_window_when_no_max_chars():
+    big = "A" * (dr._TEXT_MAX_BYTES * 3)
+    out = bound_parse_payload({"output": "text", "content": big})
+    assert "...[truncated" in out["content"]
+    assert len(out["content"].encode("utf-8")) <= dr._TEXT_MAX_BYTES + 64
+    assert out["truncated"] is True
+
+
+@pytest.mark.unit
+def test_bound_parse_payload_small_content_not_flagged():
+    out = bound_parse_payload({"output": "text", "content": "hi"}, max_chars=10)
+    assert out["content"] == "hi"
+    assert out["truncated"] is False
diff --git a/tests/storage/db/repositories/test_artifacts.py b/tests/storage/db/repositories/test_artifacts.py
index 31302d15..3673ca01 100644
--- a/tests/storage/db/repositories/test_artifacts.py
+++ b/tests/storage/db/repositories/test_artifacts.py
@@ -41,7 +41,8 @@ class TestCreateArtifact:
         assert artifact["user_id"] == "user-1"
         assert artifact["kind"] == "presentation"
         assert artifact["title"] == "Q3 deck"
-        assert artifact["metadata"] == {"source": "chat"}
+        # A stable per-parent ref_seq is merged into the caller's metadata at creation.
+        assert artifact["metadata"] == {"source": "chat", "ref_seq": 1}
         assert artifact["current_version"] == 1
         assert artifact["id"] is not None
         assert artifact["_id"] == artifact["id"]
@@ -460,6 +461,97 @@ class TestVirtualRefPositions:
             repo.artifact_id_at_position(1)
 
 
+class TestStableRefSeq:
+    """A stable per-parent ``ref_seq`` (stored in metadata) backs the ``A{n}`` refs and survives deletes."""
+
+    def test_create_assigns_incrementing_ref_seq(self, pg_conn):
+        repo = _repo(pg_conn)
+        conv = _conversation_id()
+        a = repo.create_artifact("u", "document", conversation_id=conv)
+        b = repo.create_artifact("u", "document", conversation_id=conv)
+        c = repo.create_artifact("u", "document", conversation_id=conv)
+        assert a["metadata"]["ref_seq"] == 1
+        assert b["metadata"]["ref_seq"] == 2
+        assert c["metadata"]["ref_seq"] == 3
+
+    def test_ref_seq_is_parent_scoped(self, pg_conn):
+        repo = _repo(pg_conn)
+        a1 = repo.create_artifact("u", "document", conversation_id=_conversation_id())
+        b1 = repo.create_artifact("u", "document", workflow_run_id=_conversation_id())
+        # Each parent numbers ref_seq from 1 independently.
+        assert a1["metadata"]["ref_seq"] == 1
+        assert b1["metadata"]["ref_seq"] == 1
+
+    def test_ref_seq_merges_without_clobbering_caller_metadata(self, pg_conn):
+        repo = _repo(pg_conn)
+        a = repo.create_artifact(
+            "u", "document", conversation_id=_conversation_id(), metadata={"k": "v"}
+        )
+        assert a["metadata"] == {"k": "v", "ref_seq": 1}
+
+    def test_next_ref_seq_is_monotonic_across_delete(self, pg_conn):
+        repo = _repo(pg_conn)
+        conv = _conversation_id()
+        a = repo.create_artifact("u", "document", conversation_id=conv)
+        repo.create_artifact("u", "document", conversation_id=conv)  # ref_seq 2
+        repo.create_artifact("u", "document", conversation_id=conv)  # ref_seq 3
+        assert repo.next_ref_seq(conversation_id=conv) == 4
+        # Deleting an earlier artifact must NOT lower the next seq (no reuse -> no re-point).
+        repo.delete_artifact(a["id"])
+        assert repo.next_ref_seq(conversation_id=conv) == 4
+
+    def test_resolve_id_by_ref_seq_is_stable_after_delete(self, pg_conn):
+        repo = _repo(pg_conn)
+        conv = _conversation_id()
+        a = repo.create_artifact("u", "document", conversation_id=conv)
+        b = repo.create_artifact("u", "document", conversation_id=conv)
+        c = repo.create_artifact("u", "document", conversation_id=conv)
+        assert repo.resolve_id_by_ref_seq(2, conversation_id=conv) == b["id"]
+        assert repo.resolve_id_by_ref_seq(3, conversation_id=conv) == c["id"]
+        # Delete the FIRST artifact; B/C refs must NOT shift.
+        repo.delete_artifact(a["id"])
+        assert repo.resolve_id_by_ref_seq(2, conversation_id=conv) == b["id"]
+        assert repo.resolve_id_by_ref_seq(3, conversation_id=conv) == c["id"]
+        assert repo.resolve_id_by_ref_seq(1, conversation_id=conv) is None  # A is gone
+
+    def test_resolve_id_by_ref_seq_missing_and_invalid(self, pg_conn):
+        repo = _repo(pg_conn)
+        conv = _conversation_id()
+        repo.create_artifact("u", "document", conversation_id=conv)
+        assert repo.resolve_id_by_ref_seq(99, conversation_id=conv) is None
+        assert repo.resolve_id_by_ref_seq(0, conversation_id=conv) is None
+        with pytest.raises(ValueError):
+            repo.next_ref_seq()
+        with pytest.raises(ValueError):
+            repo.resolve_id_by_ref_seq(1)
+
+    def test_resolver_resolves_a_n_by_stable_seq_after_delete(self, pg_conn):
+        from application.agents.tools.artifact_ref import resolve_artifact_id
+
+        repo = _repo(pg_conn)
+        conv = _conversation_id()
+        a = repo.create_artifact("u", "document", conversation_id=conv)
+        b = repo.create_artifact("u", "document", conversation_id=conv)
+        repo.delete_artifact(a["id"])
+        # The tool path resolves A2 to B via ref_seq even though B is now positionally first.
+        assert resolve_artifact_id(repo, "A2", conversation_id=conv) == b["id"]
+
+    def test_legacy_row_without_ref_seq_falls_back_to_position(self, pg_conn):
+        from application.agents.tools.artifact_ref import resolve_artifact_id
+
+        repo = _repo(pg_conn)
+        conv = _conversation_id()
+        art = repo.create_artifact("u", "document", conversation_id=conv)
+        # Simulate a pre-migration row that never got a ref_seq.
+        pg_conn.execute(
+            text("UPDATE artifacts SET metadata = NULL WHERE id = CAST(:id AS uuid)"),
+            {"id": art["id"]},
+        )
+        assert repo.resolve_id_by_ref_seq(1, conversation_id=conv) is None
+        # No ref_seq -> positional fallback still resolves A1.
+        assert resolve_artifact_id(repo, "A1", conversation_id=conv) == art["id"]
+
+
 class TestCascadeDelete:
     def test_deleting_artifact_removes_versions(self, pg_conn):
         repo = _repo(pg_conn)
diff --git a/tests/test_code_executor_tool.py b/tests/test_code_executor_tool.py
index 333afc80..f2a33ecf 100644
--- a/tests/test_code_executor_tool.py
+++ b/tests/test_code_executor_tool.py
@@ -428,3 +428,54 @@ def test_materialize_inputs_rejects_oversize_by_declared_size(monkeypatch):
     out = _tool()._materialize_inputs(manager, "conv-1", [_ART_ID])
     assert "exceeds" in out["error"] and "sandbox input limit" in out["error"]
     assert manager.put_files == {}  # nothing staged
+
+
+def test_materialize_inputs_dedupes_same_filename(monkeypatch):
+    """Two inputs whose current versions share a filename stage to DISTINCT inputs/ paths."""
+    from application.agents.tools import code_executor as ce
+
+    id_a = str(uuid.uuid4())
+    id_b = str(uuid.uuid4())
+
+    class _Repo:
+        def __init__(self, conn):
+            pass
+
+        def artifact_id_at_position(self, n, *, conversation_id=None, workflow_run_id=None):
+            return None
+
+        def get_artifact_in_parent(self, artifact_id, *, conversation_id=None, workflow_run_id=None):
+            if conversation_id != "conv-1":
+                return None
+            return {"id": artifact_id, "current_version": 1, "title": "seed.csv"}
+
+        def get_version(self, artifact_id, version):
+            # Both inputs carry the SAME current filename but distinct stored bytes.
+            return {"filename": "seed.csv", "storage_path": f"p/{artifact_id}.csv"}
+
+    class _Conn:
+        def __enter__(self):
+            return object()
+
+        def __exit__(self, *exc):
+            return False
+
+    class _Storage:
+        def get_file(self, path):
+            import io
+
+            return io.BytesIO(path.encode())  # distinct bytes per artifact path
+
+    monkeypatch.setattr(ce, "db_readonly", lambda: _Conn())
+    monkeypatch.setattr(ce, "ArtifactsRepository", _Repo)
+    monkeypatch.setattr(ce.StorageCreator, "get_storage", staticmethod(lambda: _Storage()))
+
+    manager = _InputManager()
+    out = _tool()._materialize_inputs(manager, "conv-1", [id_a, id_b])
+
+    assert "error" not in out
+    # The colliding second input is suffixed before the extension; both are reported.
+    assert out["loaded"] == ["inputs/seed.csv", "inputs/seed-2.csv"]
+    assert set(manager.put_files) == {"inputs/seed.csv", "inputs/seed-2.csv"}
+    # Each path holds its own artifact's bytes (no clobber).
+    assert manager.put_files["inputs/seed.csv"] != manager.put_files["inputs/seed-2.csv"]
diff --git a/tests/worker/test_parse_document_worker.py b/tests/worker/test_parse_document_worker.py
index 62e36829..1a762587 100644
--- a/tests/worker/test_parse_document_worker.py
+++ b/tests/worker/test_parse_document_worker.py
@@ -206,3 +206,61 @@ def test_result_payload_chunks_are_bounded(monkeypatch):
     assert out["total_chunks"] == dr._MAX_CHUNKS_RETURNED * 3
     assert all("...[truncated" in c for c in out["chunks"])
     assert all(len(c) < len(huge_chunk) for c in out["chunks"])
+
+
+def _capture_persist(monkeypatch):
+    """Patch persist_new_artifact to record what would be stored; return the capture dict."""
+    import application.sandbox.artifacts_capture as ac
+
+    captured: Dict[str, Any] = {}
+
+    def _fake_persist(**kwargs):
+        captured.update(kwargs)
+        return {"artifact_id": "new-art", "version": 1, "filename": "x.json",
+                "mime_type": "application/json", "size": len(kwargs.get("data", b""))}
+
+    monkeypatch.setattr(ac, "persist_new_artifact", _fake_persist)
+    return captured
+
+
+@pytest.mark.unit
+def test_persist_keeps_full_content_while_view_is_bounded(monkeypatch):
+    # A >8000-byte doc: the persisted artifact must keep the FULL text, while the returned
+    # (Redis/LLM) view is head+tail windowed.
+    import json
+
+    _patch_repo(monkeypatch, found=True, run="run-1")
+    big = "Z" * 12000
+    _patch_parse(monkeypatch, {"output": "markdown", "content": big, "truncated": False})
+    captured = _capture_persist(monkeypatch)
+
+    out = worker.parse_document_worker(
+        None, _ART_ID, {"workflow_run_id": "run-1"}, "u-1", {"output": "markdown", "persist": True}
+    )
+
+    persisted = json.loads(captured["data"].decode("utf-8"))
+    assert persisted["content"] == big  # FULL parse persisted
+    assert len(persisted["content"]) == 12000
+    assert len(out["content"]) < len(big)  # bounded view
+    assert "...[truncated" in out["content"]
+    assert out["truncated"] is True
+
+
+@pytest.mark.unit
+def test_persist_full_but_view_respects_max_chars(monkeypatch):
+    # max_chars bounds only the returned view; the persisted artifact stays full.
+    import json
+
+    _patch_repo(monkeypatch, found=True, run="run-1")
+    big = "Z" * 5000
+    _patch_parse(monkeypatch, {"output": "markdown", "content": big, "truncated": False})
+    captured = _capture_persist(monkeypatch)
+
+    out = worker.parse_document_worker(
+        None, _ART_ID, {"workflow_run_id": "run-1"}, "u-1",
+        {"output": "markdown", "persist": True, "max_chars": 100},
+    )
+
+    assert json.loads(captured["data"].decode("utf-8"))["content"] == big  # full persisted
+    assert len(out["content"]) == 100  # view capped by max_chars
+    assert out["truncated"] is True