From 0a53c8029f043767dc6967911fde5a42373fc446 Mon Sep 17 00:00:00 2001
From: zeroshotmind <zero.shot.mind@gmail.com>
Date: Tue, 7 Jul 2026 14:45:35 -0700
Subject: [PATCH] test(redteam): cover RedTeamCase metadata sync and
 _build_system_prompt assembly
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

RedTeamCase._sync_metadata_from_config (case.py) and _build_system_prompt
(strategies/_common.py) had no direct unit tests — only incidental exercise
via fixture setup in other strategy test files. Neither behavior was
actually asserted:

- RedTeamCase: metadata=None population, setdefault-style preservation of
  caller-provided keys over attack_goal values, partial backfill, and the
  empty-dict (falsy but not None) branch.
- _build_system_prompt: that it includes and correctly orders all six
  shared prompt blocks, joins them with a blank line, and leaves the
  {actor_profile}/{max_turns} placeholders for callers to fill in later.
---
 .../experimental/redteam/test_case.py         | 59 +++++++++++++++++++
 .../experimental/redteam/test_common.py       | 47 +++++++++++++++
 2 files changed, 106 insertions(+)
 create mode 100644 tests/strands_evals/experimental/redteam/test_case.py
 create mode 100644 tests/strands_evals/experimental/redteam/test_common.py

diff --git a/tests/strands_evals/experimental/redteam/test_case.py b/tests/strands_evals/experimental/redteam/test_case.py
new file mode 100644
index 00000000..c4fac1f5
--- /dev/null
+++ b/tests/strands_evals/experimental/redteam/test_case.py
@@ -0,0 +1,59 @@
+"""Tests for RedTeamCase metadata sync behavior."""
+
+from strands_evals.experimental.redteam.case import RedTeamCase
+from strands_evals.experimental.redteam.types import AttackGoal, RedTeamConfig
+
+
+def _config(**attack_goal_kwargs) -> RedTeamConfig:
+    kwargs = {"risk_category": "guideline_bypass", "actor_goal": "goal"}
+    kwargs.update(attack_goal_kwargs)
+    return RedTeamConfig(attack_goal=AttackGoal(**kwargs))
+
+
+def test_metadata_none_is_populated_from_attack_goal():
+    """A case with no metadata gets the full attack_goal dump as its metadata."""
+    case = RedTeamCase(name="c0", input="hello", config=_config())
+
+    assert case.metadata == case.config.attack_goal.model_dump()
+    assert case.metadata["risk_category"] == "guideline_bypass"
+    assert case.metadata["actor_goal"] == "goal"
+    assert case.metadata["severity"] == "medium"
+
+
+def test_existing_metadata_keys_are_preserved_not_overwritten():
+    """Caller-provided metadata wins over the attack_goal dump for overlapping keys."""
+    case = RedTeamCase(
+        name="c0",
+        input="hello",
+        config=_config(severity="critical"),
+        metadata={"severity": "low", "custom": "value"},
+    )
+
+    # setdefault semantics: caller's "severity" survives, config's does not override it.
+    assert case.metadata["severity"] == "low"
+    # Caller-only keys are untouched.
+    assert case.metadata["custom"] == "value"
+    # Keys absent from caller metadata are still filled in from attack_goal.
+    assert case.metadata["risk_category"] == "guideline_bypass"
+    assert case.metadata["actor_goal"] == "goal"
+
+
+def test_partial_metadata_is_filled_in_from_attack_goal():
+    """Keys the caller didn't set are backfilled from attack_goal without touching the rest."""
+    case = RedTeamCase(
+        name="c0",
+        input="hello",
+        config=_config(context="ctx", success_criteria="criteria"),
+        metadata={"unrelated": "kept"},
+    )
+
+    assert case.metadata["unrelated"] == "kept"
+    assert case.metadata["context"] == "ctx"
+    assert case.metadata["success_criteria"] == "criteria"
+
+
+def test_empty_dict_metadata_is_filled_in_from_attack_goal():
+    """An explicit empty dict (falsy but not None) still gets backfilled, exercising the else branch."""
+    case = RedTeamCase(name="c0", input="hello", config=_config(), metadata={})
+
+    assert case.metadata == case.config.attack_goal.model_dump()
diff --git a/tests/strands_evals/experimental/redteam/test_common.py b/tests/strands_evals/experimental/redteam/test_common.py
new file mode 100644
index 00000000..ef046e1f
--- /dev/null
+++ b/tests/strands_evals/experimental/redteam/test_common.py
@@ -0,0 +1,47 @@
+"""Tests for shared prompt-block assembly in strategies/_common.py."""
+
+from strands_evals.experimental.redteam.strategies._common import (
+    EXIT_BLOCK,
+    FORMAT_BLOCK,
+    PROFILE_BLOCK,
+    ROLE_BLOCK,
+    RULES_BLOCK,
+    _build_system_prompt,
+)
+
+
+def test_build_system_prompt_includes_all_common_blocks():
+    prompt = _build_system_prompt("PLAYBOOK CONTENT")
+
+    assert ROLE_BLOCK in prompt
+    assert PROFILE_BLOCK in prompt
+    assert "PLAYBOOK CONTENT" in prompt
+    assert RULES_BLOCK in prompt
+    assert EXIT_BLOCK in prompt
+    assert FORMAT_BLOCK in prompt
+
+
+def test_build_system_prompt_orders_blocks_role_profile_playbook_rules_exit_format():
+    """Downstream .format() calls fill {actor_goal}/{max_turns} placeholders positionally
+    within this fixed layout, so the block order is load-bearing, not incidental."""
+    prompt = _build_system_prompt("PLAYBOOK")
+
+    positions = [
+        prompt.index(ROLE_BLOCK),
+        prompt.index(PROFILE_BLOCK),
+        prompt.index("PLAYBOOK"),
+        prompt.index(RULES_BLOCK),
+        prompt.index(EXIT_BLOCK),
+        prompt.index(FORMAT_BLOCK),
+    ]
+
+    assert positions == sorted(positions)
+
+
+def test_profile_and_exit_blocks_retain_format_placeholders():
+    """These blocks are filled in later via str.format(actor_profile=..., max_turns=...);
+    _build_system_prompt itself must not consume those placeholders."""
+    prompt = _build_system_prompt("PLAYBOOK")
+
+    assert "{actor_profile}" in prompt
+    assert "{max_turns}" in prompt