feat(agent): strip raw tool-scaffolding from streamed answers — deterministic filter with one bounded recovery

tests/unit/test_chat_gate.py

@@ -11,7 +11,7 @@ from __future__ import annotations
 import json
 import uuid
 from collections.abc import Iterator
-from typing import Any
+from typing import TYPE_CHECKING, Any
 
 import pytest
 from fastapi.testclient import TestClient
@@ -25,6 +25,9 @@ from app.rag.llm import StreamPiece
 from app.rag.retriever import RetrievedChunk, weak_hit_titles
 from app.rag.suggestions import MAX_SUGGESTIONS, derive_suggestions
 
+if TYPE_CHECKING:
+    from app.rag.scaffolding import ScaffoldingFilter
+
 ANSWER = "I haven't done anything like that — try one of these instead!"
 
 #: A small KB outline standing in for the lite-generated one (phase 31).
@@ -332,6 +335,11 @@ def test_low_prompt_has_titles_only_no_content() -> None:
     assert "<relevance>LOW</relevance>" in prompt
     assert "DEFLECT_MODE" in prompt
     assert "HONESTY GATE" in prompt  # the LOW rule is what the model follows
+    # Phase 71 (owner-permitted 2026-09-03): the deflection plain-text
+    # line — the LOW turn offers no tools, so any tool markup there is
+    # always wrong (prevention at the prompt; the filter + recovery is
+    # the backstop).
+    assert "Reply in plain text only — you have no tools in this mode." in prompt
     assert "- Kubernetes Homelab Cluster" in prompt
     assert "- Backup Strategy" in prompt
     assert "ALPHA_DOC_CONTENT" not in prompt
@@ -347,6 +355,8 @@ def test_high_path_unaffected() -> None:
     assert plan.suggestions == []
     assert "<relevance>HIGH</relevance>" in plan.system_prompt
     assert "DEFLECT_MODE" not in plan.system_prompt
+    # Phase 71: the deflection plain-text line never leaks into HIGH.
+    assert "Reply in plain text only" not in plan.system_prompt
     assert "ALPHA_DOC_CONTENT" in plan.system_prompt
     assert "BETA_DOC_CONTENT" in plan.system_prompt
     assert [d.title for d in plan.docs] == ["Kubernetes Homelab Cluster", "Backup Strategy"]
@@ -438,6 +448,7 @@ class _CannedLLM:
         self,
         messages: list[dict[str, str]],
         tools: list[dict[str, Any]] | None = None,
+        scaffolding: ScaffoldingFilter | None = None,  # phase 71 pass-through
     ):
         self.seen.append(messages)
         self.seen_tools.append(tools)