feat(memory): add incremental transcript compaction

tests/unit/agents/test_flash_scrub_edge.py

@@ -0,0 +1,325 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Unit tests for the Flash scrub-edge compressor (history redesign §3.2, M1).
+
+Covers the M1 acceptance items: the legacy-equivalent success-path product
+(asserted directly since the legacy ``compress_flash_messages`` reference was
+removed in M5), the three race regressions (grace recovery, grace exhaustion
+without backfill, failed-summary placeholder), the freeze invariant (bytes
+behind the scrub edge never change), and depth-1 XML stripping equivalence.
+"""
+
+import json
+from unittest.mock import Mock
+
+import pytest
+from langchain_core.messages import AIMessage, HumanMessage, ToolMessage
+
+from artemis.agents.flash.context_compressor import ScrubEdgeCompressor
+from artemis.agents.flash.summarizer import VisualStepSummarizer
+from artemis.context import ArtemisContext
+
+
+@pytest.fixture
+def mock_context():
+    ctx = Mock(spec=ArtemisContext)
+    ctx.data_engine = None
+    return ctx
+
+
+def _initial_observation():
+    return HumanMessage(
+        content=[
+            {"type": "text", "text": "Task: Search flights"},
+            {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,IMG_0"}},
+            {"type": "text", "text": "--- UI Element List ---\nTree 0"},
+        ]
+    )
+
+
+def _step(i: int) -> ToolMessage:
+    return ToolMessage(
+        tool_call_id=f"tc{i}",
+        name="click",
+        content=[
+            {"type": "text", "text": f"Action {i} completed."},
+            {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,IMG_{i}"}},
+            {"type": "text", "text": f"--- UI Element List ---\nTree {i}"},
+        ],
+    )
+
+
+def _simulate_turns(compressor, messages, new_messages):
+    """Mimic FlashRunner's ordering: compress at turn start, then append."""
+    for msg in new_messages:
+        compressor.compress(messages)
+        messages.append(msg)
+    compressor.compress(messages)
+
+
+def _content_fingerprint(msg) -> str:
+    return json.dumps(msg.content, sort_keys=True, default=str)
+
+
+def _has_image(msg) -> bool:
+    return any(
+        isinstance(b, dict) and b.get("type") in ("image_url", "image") for b in msg.content
+    )
+
+
+def test_success_path_product_behind_the_scrub_edge(mock_context):
+    """With summaries ready in time, every message behind the scrub edge is the
+    legacy-shaped product: action text kept, XML stripped, screenshot replaced
+    by the ``--- Historical Visual Transition ---`` summary block.
+
+    (Asserted directly; the legacy ``compress_flash_messages`` reference this
+    was originally diffed against was removed in M5.) Intended M1 timing
+    change: the swap happens at depth K, so the newest K-1 historical steps
+    still carry their screenshots.
+    """
+    summarizer = VisualStepSummarizer(mock_context)
+    for i in range(1, 7):
+        summarizer._summaries[f"tc{i}"] = f"Visual summary {i}."
+
+    new_messages: list = []
+    compressor = ScrubEdgeCompressor(
+        summarizer=summarizer,
+        prune_history_xml=True,
+        image_scrub_depth=3,
+        pending_grace_steps=3,
+    )
+    _simulate_turns(
+        compressor, new_messages, [_initial_observation()] + [_step(i) for i in range(1, 7)]
+    )
+
+    # Initial observation behind the edge: image silently dropped (no summary
+    # job), task text kept, XML stripped.
+    assert new_messages[0].content == [{"type": "text", "text": "Task: Search flights"}]
+
+    # Steps 1-4 behind the scrub edge: exact legacy-shaped swap product.
+    for i in range(1, 5):
+        assert new_messages[i].content == [
+            {"type": "text", "text": f"Action {i} completed."},
+            {
+                "type": "text",
+                "text": f"--- Historical Visual Transition ---\nVisual summary {i}.",
+            },
+        ], f"message {i} diverged from the legacy-shaped product"
+
+    # The live observation is intact.
+    assert _has_image(new_messages[6])
+    assert "Tree 6" in str(new_messages[6].content)
+
+    # Intended timing change: step 5 (depth 2 < K) keeps its screenshot under
+    # the scrub edge; its XML was still stripped at depth 1.
+    assert _has_image(new_messages[5])
+    assert "Visual summary 5." not in str(new_messages[5].content)
+    assert "Tree 5" not in str(new_messages[5].content)
+
+
+def test_race_pending_at_edge_recovers_within_grace(mock_context):
+    """A step whose summary is late keeps its image at the scrub edge and is
+    swapped to the summary once it arrives within the grace window."""
+    summarizer = VisualStepSummarizer(mock_context)
+    summarizer._step_inputs["tc1"] = {"step_number": 1}  # job pending
+
+    messages: list = []
+    compressor = ScrubEdgeCompressor(
+        summarizer=summarizer,
+        image_scrub_depth=3,
+        pending_grace_steps=3,
+    )
+    _simulate_turns(
+        compressor, messages, [_initial_observation()] + [_step(i) for i in range(1, 4)]
+    )
+
+    # Step 1 sits at the scrub edge with a pending summary: image retained.
+    assert _has_image(messages[1])
+
+    # Summary arrives within grace; the next pass applies it.
+    summarizer._summaries["tc1"] = "Recovered visual summary 1."
+    _simulate_turns(compressor, messages, [_step(4)])
+
+    assert not _has_image(messages[1])
+    assert {
+        "type": "text",
+        "text": "--- Historical Visual Transition ---\nRecovered visual summary 1.",
+    } in messages[1].content
+
+
+def test_race_grace_exhausted_placeholder_never_backfilled(mock_context):
+    """After the grace window a pending step becomes a placeholder, and a late
+    summary never mutates the frozen message again."""
+    summarizer = VisualStepSummarizer(mock_context)
+    summarizer._step_inputs["tc1"] = {"step_number": 1}  # pending forever (for now)
+
+    messages: list = []
+    compressor = ScrubEdgeCompressor(
+        summarizer=summarizer,
+        image_scrub_depth=2,
+        pending_grace_steps=1,
+    )
+    _simulate_turns(compressor, messages, [_step(i) for i in range(1, 5)])
+
+    # Depth 4 > K + grace = 3: grace exhausted, placeholder applied.
+    step1_text = str(messages[0].content)
+    assert not _has_image(messages[0])
+    assert "[visual summary pending; evidence at DataEngine step 1]" in step1_text
+    frozen_fingerprint = _content_fingerprint(messages[0])
+
+    # The summary arriving later must NOT backfill the frozen message.
+    summarizer._summaries["tc1"] = "Too late summary."
+    _simulate_turns(compressor, messages, [_step(5)])
+    compressor.compress(messages)
+
+    assert _content_fingerprint(messages[0]) == frozen_fingerprint
+    assert "Too late summary." not in str(messages[0].content)
+
+
+def test_race_failed_summary_becomes_unavailable_placeholder(mock_context):
+    """A failed summary is replaced by an unavailable placeholder at the scrub
+    edge without waiting out the grace window."""
+    summarizer = VisualStepSummarizer(mock_context)
+    summarizer._step_inputs["tc1"] = {"step_number": 1}
+    summarizer._failed.add("tc1")
+
+    messages: list = []
+    compressor = ScrubEdgeCompressor(
+        summarizer=summarizer,
+        image_scrub_depth=2,
+        pending_grace_steps=5,
+    )
+    _simulate_turns(compressor, messages, [_step(1), _step(2), _step(3)])
+
+    assert not _has_image(messages[0])
+    assert "[visual summary unavailable; evidence at DataEngine step 1]" in str(
+        messages[0].content
+    )
+
+
+def test_freeze_invariant_bytes_never_change(mock_context):
+    """Once a message crosses the scrub edge its content bytes never change,
+    across many turns and late-arriving summaries."""
+    summarizer = VisualStepSummarizer(mock_context)
+    # Odd steps have summaries ready from the start; even steps stay pending.
+    for i in range(1, 11):
+        summarizer._step_inputs[f"tc{i}"] = {"step_number": i}
+        if i % 2 == 1:
+            summarizer._summaries[f"tc{i}"] = f"Ready summary {i}."
+
+    messages: list = []
+    compressor = ScrubEdgeCompressor(
+        summarizer=summarizer,
+        image_scrub_depth=2,
+        pending_grace_steps=1,
+    )
+
+    frozen_fingerprints: dict[int, str] = {}
+
+    def check_frozen():
+        for idx in compressor._frozen:
+            fingerprint = _content_fingerprint(messages[idx])
+            if idx in frozen_fingerprints:
+                assert frozen_fingerprints[idx] == fingerprint, (
+                    f"frozen message {idx} mutated"
+                )
+            else:
+                frozen_fingerprints[idx] = fingerprint
+
+    for i in range(1, 11):
+        compressor.compress(messages)
+        check_frozen()
+        messages.append(AIMessage(content=f"Thinking about step {i}."))
+        messages.append(_step(i))
+        if i == 6:
+            # Late summaries for already-frozen even steps: must be ignored.
+            summarizer._summaries["tc2"] = "Late summary 2."
+            summarizer._summaries["tc4"] = "Late summary 4."
+    compressor.compress(messages)
+    check_frozen()
+
+    assert len(frozen_fingerprints) >= 6
+    # A frozen pending step carries the placeholder, not the late summary.
+    assert "Late summary 2." not in " ".join(str(m.content) for m in messages)
+
+
+def test_xml_depth1_strip_matches_legacy_semantics(mock_context):
+    """Depth-1 XML stripping keeps combined-block prefixes and the live list."""
+    messages: list = []
+    compressor = ScrubEdgeCompressor(summarizer=None, prune_history_xml=True)
+
+    combined = ToolMessage(
+        tool_call_id="tc1",
+        name="click",
+        content=[
+            {
+                "type": "text",
+                "text": "Tapped Settings.\n--- UI Element List ---\n[huge tree]",
+            },
+            {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,A"}},
+        ],
+    )
+    _simulate_turns(compressor, messages, [combined, _step(2)])
+
+    # Historical combined block: prefix retained, tree stripped (the exact
+    # legacy ``compress_flash_messages`` product for this input).
+    assert messages[0].content[0] == {"type": "text", "text": "Tapped Settings."}
+    # Live observation keeps its UI list untouched.
+    assert "Tree 2" in str(messages[1].content)
+
+
+def test_prune_history_xml_disabled_keeps_ui_lists(mock_context):
+    """With the switch off, UI lists survive both depth-1 and freeze scrubs."""
+    summarizer = VisualStepSummarizer(mock_context)
+    summarizer._summaries["tc1"] = "Summary 1."
+
+    messages: list = []
+    compressor = ScrubEdgeCompressor(
+        summarizer=summarizer,
+        prune_history_xml=False,
+        image_scrub_depth=2,
+        pending_grace_steps=0,
+    )
+    _simulate_turns(compressor, messages, [_step(1), _step(2), _step(3)])
+
+    # Step 1 froze (summary applied) but its UI list is preserved.
+    assert not _has_image(messages[0])
+    assert "Tree 1" in str(messages[0].content)
+    assert "Summary 1." in str(messages[0].content)
+
+
+def test_untracked_messages_are_never_touched(mock_context):
+    """Injected instructions and AI turns pass through untouched; a frozen
+    image-only tool message falls back to non-empty content."""
+    messages: list = [
+        HumanMessage(content="[REAL-TIME INJECTED INSTRUCTION from user]: stay put"),
+    ]
+    compressor = ScrubEdgeCompressor(summarizer=None, image_scrub_depth=2)
+
+    image_only = ToolMessage(
+        tool_call_id="tc1",
+        name="click",
+        content=[{"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,A"}}],
+    )
+    _simulate_turns(
+        compressor,
+        messages,
+        [image_only, AIMessage(content="thinking"), _step(2), _step(3)],
+    )
+
+    assert messages[0].content == "[REAL-TIME INJECTED INSTRUCTION from user]: stay put"
+    assert messages[2].content == "thinking"
+    # Image-only tool message froze with the defensive fallback text.
+    assert messages[1].content == [{"type": "text", "text": "Action completed."}]

tests/unit/agents/test_flash_step_summarizer.py

@@ -14,12 +14,12 @@
 
 """Unit tests for VisualStepSummarizer and Context Compressor in Flash profile."""
 
-from unittest.mock import AsyncMock, Mock, call
+from unittest.mock import ANY, AsyncMock, Mock, call
 from uuid import uuid4
 import pytest
 from langchain_core.messages import AIMessage, HumanMessage, ToolMessage
 
-from artemis.agents.flash.context_compressor import compress_flash_messages
+from artemis.agents.flash.context_compressor import ScrubEdgeCompressor
 from artemis.agents.flash.summarizer import VisualStepSummarizer
 from artemis.context import ArtemisContext
 from artemis.sdk.builders import Builders
@@ -100,8 +100,94 @@ async def test_summarizer_retries_independently(mock_context):
     assert attempt_count == 2
 
 
+@pytest.mark.asyncio
+async def test_summarizer_bounded_retry_marks_failed(mock_context):
+    """Persistent failures exit after 1 + retry_limit attempts with an explicit failed state."""
+    mock_context.data_engine = Mock()
+    summarizer = VisualStepSummarizer(mock_context, retry_limit=2)
+    summarizer._retry_delays = (0.0,)
+    attempt_count = 0
+
+    async def mock_ainvoke(messages):
+        nonlocal attempt_count
+        attempt_count += 1
+        raise TimeoutError("Simulated permanent failure")
+
+    summarizer._llm = Mock()
+    summarizer._llm.ainvoke = AsyncMock(side_effect=mock_ainvoke)
+
+    step_id = uuid4()
+    summarizer.dispatch(
+        step_number=1,
+        action_name="click",
+        action_args={"target": [100, 100]},
+        pre_img_bytes=b"pre1",
+        post_img_bytes=b"post1",
+        exec_outcome="Outcome 1",
+        action_key="tc-fail",
+        data_engine_step_id=step_id,
+    )
+    await summarizer.flush()
+
+    # Bounded exit: initial attempt + retry_limit retries, then explicit failure.
+    assert attempt_count == 3
+    assert summarizer.has_failed("tc-fail")
+    assert not summarizer.has_summary("tc-fail")
+    # Lossless semantics: a failed step still reads as pending so the
+    # compressor keeps its source image instead of dropping evidence.
+    assert summarizer.is_pending("tc-fail")
+
+    # DataEngine observability: pending on dispatch, failed on exhaustion.
+    mock_context.data_engine.update_step_summary.assert_has_calls(
+        [
+            call(step_id, None, status="pending", source="visual_transition", model=ANY),
+            call(step_id, None, status="failed", source="visual_transition", model=ANY),
+        ]
+    )
+
+
+@pytest.mark.asyncio
+async def test_summarizer_success_within_retry_budget_not_failed(mock_context):
+    """A summary that succeeds on a retry within the budget never enters failed state."""
+    summarizer = VisualStepSummarizer(mock_context, retry_limit=3)
+    summarizer._retry_delays = (0.0,)
+    attempt_count = 0
+
+    async def mock_ainvoke(messages):
+        nonlocal attempt_count
+        attempt_count += 1
+        if attempt_count < 3:
+            raise TimeoutError("Transient failure")
+        return AIMessage(content="[Step 1 State: Finally worked.]")
+
+    summarizer._llm = Mock()
+    summarizer._llm.ainvoke = AsyncMock(side_effect=mock_ainvoke)
+
+    summarizer.dispatch(
+        step_number=1,
+        action_name="click",
+        action_args={"target": [100, 100]},
+        pre_img_bytes=b"pre1",
+        post_img_bytes=b"post1",
+        exec_outcome="Outcome 1",
+        action_key="tc-ok",
+    )
+    await summarizer.flush()
+
+    assert attempt_count == 3
+    assert summarizer.has_summary("tc-ok")
+    assert not summarizer.has_failed("tc-ok")
+
+
 def test_context_compressor_with_ready_summaries(mock_context):
-    """Verify that ready summaries replace past image blocks and prune outdated XML."""
+    """Ready summaries replace past image blocks and outdated XML is pruned.
+
+    Rewritten against ``ScrubEdgeCompressor`` (M5; the legacy
+    ``compress_flash_messages`` was removed) with ``image_scrub_depth=1`` so
+    every historical message is scrubbed immediately, mirroring the legacy
+    timing this test was written for. Summaries are keyed by legacy ordinals
+    to keep the older-caller fallback path covered.
+    """
     summarizer = VisualStepSummarizer(mock_context)
     summarizer._summaries[1] = "[Step 1 State: Tapped Search bar. Keyboard appeared.]"
     summarizer._summaries[2] = "[Step 2 State: Typed 'flights'. Suggestions displayed.]"
@@ -134,7 +220,13 @@ def test_context_compressor_with_ready_summaries(mock_context):
         ),
     ]
 
-    compress_flash_messages(messages, summarizer=summarizer, prune_history_xml=True)
+    compressor = ScrubEdgeCompressor(
+        summarizer=summarizer,
+        prune_history_xml=True,
+        image_scrub_depth=1,
+        pending_grace_steps=0,
+    )
+    compressor.compress(messages)
 
     # Initial HumanMessage: Historical initial image pruned, task text preserved
     assert len(messages[0].content) == 1
@@ -155,88 +247,15 @@ def test_context_compressor_with_ready_summaries(mock_context):
     assert "Live Latest Tree 3" in messages[2].content[2]["text"]
 
 
-def test_context_compressor_graceful_fallback(mock_context):
-    """Verify that when a summary is not ready yet, it cleanly falls back to standard pruned state."""
-    summarizer = VisualStepSummarizer(mock_context)
-    # Step 1 has no summary in cache
-
-    messages = [
-        HumanMessage(
-            content=[
-                {"type": "text", "text": "Task: Search"},
-                {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,IMG_1"}},
-            ]
-        ),
-        ToolMessage(
-            tool_call_id="tc1",
-            name="click",
-            content=[
-                {"type": "text", "text": "Action 'click' completed."},
-                {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,IMG_2"}},
-            ],
-        ),
-    ]
-
-    compress_flash_messages(messages, summarizer=summarizer, prune_history_xml=True)
-
-    # Turn 1 should simply have image omitted (standard fallback)
-    assert len(messages[0].content) == 1
-    assert messages[0].content[0]["text"] == "Task: Search"
-
-    # Turn 2 (Latest) preserves image
-    assert len(messages[1].content) == 2
-    assert messages[1].content[1]["type"] == "image_url"
-
-
-def test_context_compressor_backfills_late_summary_idempotently(mock_context):
-    """A pending image remains until its late summary replaces it."""
-    summarizer = VisualStepSummarizer(mock_context)
-    summarizer._step_inputs["action-1"] = {"step_number": 1}
-    messages = [
-        HumanMessage(
-            content=[
-                {"type": "text", "text": "Task: Search"},
-                {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,IMG_1"}},
-            ]
-        ),
-        ToolMessage(
-            tool_call_id="action-1",
-            name="click",
-            content=[
-                {"type": "text", "text": "Tapped Search."},
-                {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,IMG_2"}},
-            ],
-        ),
-        ToolMessage(
-            tool_call_id="action-2",
-            name="input_text",
-            content=[
-                {"type": "text", "text": "Typed query."},
-                {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,IMG_3"}},
-            ],
-        ),
-    ]
-
-    compress_flash_messages(messages, summarizer=summarizer)
-    assert any(block["type"] == "image_url" for block in messages[1].content)
-    assert len(messages[1].content) == 2
-
-    summarizer._summaries["action-1"] = "Search field focused; keyboard appeared."
-    compress_flash_messages(messages, summarizer=summarizer)
-    compress_flash_messages(messages, summarizer=summarizer)
-
-    generated_blocks = [
-        block
-        for block in messages[1].content
-        if block.get("text", "").startswith("--- Historical Visual Transition ---")
-    ]
-    assert len(generated_blocks) == 1
-    assert generated_blocks[0]["text"].endswith("Search field focused; keyboard appeared.")
-    assert all(block["type"] != "image_url" for block in messages[1].content)
-
-
 def test_context_compressor_uses_tool_call_id_over_message_ordinal(mock_context):
-    """Generic tool messages cannot shift action summaries onto the wrong result."""
+    """Generic tool messages cannot shift action summaries onto the wrong result.
+
+    (Rewritten against ``ScrubEdgeCompressor`` in M5. The legacy tests for
+    pending-summary fallback and late-summary backfill were removed with
+    ``compress_flash_messages``: the scrub edge covers pending handling in
+    ``test_flash_scrub_edge.py``, and late backfill is now forbidden by the
+    freeze invariant rather than supported.)
+    """
     summarizer = VisualStepSummarizer(mock_context)
     summarizer._summaries["click-id"] = "Click-specific visual memory."
     summarizer._summaries["type-id"] = "Typing-specific visual memory."
@@ -265,7 +284,12 @@ def test_context_compressor_uses_tool_call_id_over_message_ordinal(mock_context)
         ),
     ]
 
-    compress_flash_messages(messages, summarizer=summarizer)
+    compressor = ScrubEdgeCompressor(
+        summarizer=summarizer,
+        image_scrub_depth=1,
+        pending_grace_steps=0,
+    )
+    compressor.compress(messages)
 
     assert "Click-specific visual memory." in str(messages[0].content)
     assert "Typing-specific visual memory." not in str(messages[0].content)
@@ -274,35 +298,6 @@ def test_context_compressor_uses_tool_call_id_over_message_ordinal(mock_context)
     assert messages[2].content[-1]["type"] == "image_url"
 
 
-def test_context_compressor_preserves_text_before_embedded_ui_list(mock_context):
-    """Pruning a combined XML block must retain its action result prefix."""
-    messages = [
-        ToolMessage(
-            tool_call_id="old",
-            name="click",
-            content=[
-                {
-                    "type": "text",
-                    "text": "Tapped Settings.\n--- UI Element List ---\n[huge tree]",
-                },
-                {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,A"}},
-            ],
-        ),
-        ToolMessage(
-            tool_call_id="live",
-            name="click",
-            content=[
-                {"type": "text", "text": "Opened page."},
-                {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,B"}},
-            ],
-        ),
-    ]
-
-    compress_flash_messages(messages, prune_history_xml=True)
-
-    assert messages[0].content == [{"type": "text", "text": "Tapped Settings."}]
-
-
 @pytest.mark.asyncio
 async def test_summarizer_keys_actions_independently_of_turn_number(mock_context):
     """Multiple actions from one LLM turn keep independent inputs and summaries."""
@@ -344,17 +339,229 @@ async def test_summarizer_keys_actions_independently_of_turn_number(mock_context
 
     assert summarizer.get_summary("tc-1") == "First action memory."
     assert summarizer.get_summary("tc-2") == "Second action memory."
-    assert summarizer._step_inputs["tc-1"]["data_engine_step_id"] == first_step_id
-    assert summarizer._step_inputs["tc-2"]["data_engine_step_id"] == second_step_id
+    # M1 keying: jobs are canonical under the DataEngine step id; the
+    # tool_call_id stays queryable as an alias.
+    assert summarizer.get_job_payload("tc-1")["data_engine_step_id"] == first_step_id
+    assert summarizer.get_job_payload("tc-2")["data_engine_step_id"] == second_step_id
+    assert summarizer.resolve_key("tc-1") == str(first_step_id)
     mock_context.data_engine.update_step_summary.assert_has_calls(
         [
-            call(first_step_id, "First action memory."),
-            call(second_step_id, "Second action memory."),
+            call(
+                first_step_id,
+                "First action memory.",
+                status="ready",
+                source="visual_transition",
+                model=ANY,
+            ),
+            call(
+                second_step_id,
+                "Second action memory.",
+                status="ready",
+                source="visual_transition",
+                model=ANY,
+            ),
         ],
         any_order=True,
     )
 
 
+@pytest.mark.asyncio
+async def test_single_frame_dispatch_uses_single_image_prompt_variant(mock_context):
+    """A missing after-frame (every Pro step, by design) selects the
+    single-image prompt: decision frame + red marker, no AFTER ACTION section."""
+    summarizer = VisualStepSummarizer(mock_context)
+    captured: list = []
+
+    async def mock_ainvoke(messages):
+        captured.append(messages)
+        return AIMessage(
+            content=(
+                "In Step 1, I tapped the 'Battery' row marked in red in the"
+                " Settings list, which showed 'Network & internet' and 'Display'."
+            )
+        )
+
+    summarizer._llm = Mock()
+    summarizer._llm.ainvoke = AsyncMock(side_effect=mock_ainvoke)
+
+    summarizer.dispatch(
+        step_number=1,
+        action_name="click",
+        action_args={"target": [100, 200]},
+        pre_img_bytes=b"decision_frame",
+        post_img_bytes=None,
+        exec_outcome="executed",
+    )
+    await summarizer.flush()
+
+    assert summarizer.has_summary(1)
+    system_text = str(captured[0][0].content)
+    human_text = str(captured[0][1].content)
+    assert "Exactly ONE screenshot" in system_text
+    assert "no independent post-action evidence" in system_text
+    assert "DECISION FRAME" in human_text
+    assert "AFTER ACTION SCREEN" not in human_text
+    assert "--- [2]" not in human_text
+
+
+@pytest.mark.asyncio
+async def test_dual_frame_dispatch_keeps_transition_prompt(mock_context):
+    """Both frames present (Flash path): the classic BEFORE/AFTER transition
+    prompt and section labels are unchanged."""
+    summarizer = VisualStepSummarizer(mock_context)
+    captured: list = []
+
+    async def mock_ainvoke(messages):
+        captured.append(messages)
+        return AIMessage(
+            content=(
+                "In Step 1, I tapped the 'Settings' gear icon marked in red;"
+                " the main Settings menu opened showing 'Apps'."
+            )
+        )
+
+    summarizer._llm = Mock()
+    summarizer._llm.ainvoke = AsyncMock(side_effect=mock_ainvoke)
+
+    summarizer.dispatch(
+        step_number=1,
+        action_name="click",
+        action_args={"target": [100, 200]},
+        pre_img_bytes=b"pre",
+        post_img_bytes=b"post",
+        exec_outcome="executed",
+    )
+    await summarizer.flush()
+
+    assert summarizer.has_summary(1)
+    system_text = str(captured[0][0].content)
+    human_text = str(captured[0][1].content)
+    assert "Before/After screenshots" in system_text
+    assert "--- [1] BEFORE ACTION SCREEN (Action Marked Visually in Red) ---" in human_text
+    assert "--- [2] AFTER ACTION SCREEN ---" in human_text
+    assert "DECISION FRAME" not in human_text
+
+
+@pytest.mark.asyncio
+async def test_echo_output_fails_attempt_then_recovers_on_retry(mock_context):
+    """An input-marker echo is a failed attempt (not a summary): the bounded
+    retry regenerates and the recovered text lands normally."""
+    summarizer = VisualStepSummarizer(mock_context)
+    summarizer._retry_delays = (0.0,)
+    attempts = 0
+
+    async def mock_ainvoke(messages):
+        nonlocal attempts
+        attempts += 1
+        if attempts == 1:
+            return AIMessage(content="--- [2] AFTER ACTION SCREEN ---")
+        return AIMessage(
+            content=(
+                "In Step 1, I tapped the search field marked in red; the field"
+                " showed placeholder text 'Search here'."
+            )
+        )
+
+    summarizer._llm = Mock()
+    summarizer._llm.ainvoke = AsyncMock(side_effect=mock_ainvoke)
+
+    summarizer.dispatch(
+        step_number=1,
+        action_name="click",
+        action_args={"target": [1, 2]},
+        pre_img_bytes=b"pre",
+        post_img_bytes=b"post",
+        exec_outcome="executed",
+    )
+    await summarizer.flush()
+
+    assert attempts == 2
+    assert summarizer.has_summary(1)
+    assert "Search here" in summarizer.get_summary(1)
+
+
+@pytest.mark.asyncio
+async def test_degenerate_outputs_exhaust_to_failed(mock_context):
+    """Echo/too-short/too-long outputs never land as summaries; persistent
+    degeneration exhausts the bounded retries into the explicit failed state."""
+    summarizer = VisualStepSummarizer(mock_context, retry_limit=1)
+    summarizer._retry_delays = (0.0,)
+
+    async def mock_ainvoke(messages):
+        text = str(messages)
+        if "Step 1" in text:
+            return AIMessage(content="--- [1] DECISION FRAME: SCREEN AT ACTION TIME ---")
+        if "Step 2" in text:
+            return AIMessage(content="x" * 2000)
+        return AIMessage(content="ok.")
+
+    summarizer._llm = Mock()
+    summarizer._llm.ainvoke = AsyncMock(side_effect=mock_ainvoke)
+
+    for step in (1, 2, 3):
+        summarizer.dispatch(
+            step_number=step,
+            action_name="click",
+            action_args={"target": [1, 2]},
+            pre_img_bytes=b"pre",
+            post_img_bytes=None,
+            exec_outcome="executed",
+        )
+    await summarizer.flush()
+
+    for step in (1, 2, 3):
+        assert summarizer.has_failed(step), f"step {step} should have failed"
+        assert not summarizer.has_summary(step)
+
+
+@pytest.mark.asyncio
+async def test_raw_lens_call_meters_llm_usage_without_touching_context_base(mock_context):
+    """The raw-model lens bypass records an llm_usage trace, but its tiny
+    prompt never overwrites the session's last_prompt_tokens (the compaction
+    thresholds' live context base)."""
+    from artemis.services.token_meter import get_meter
+
+    engine = Mock()
+    engine.current_session_id = "lens-meter-session"
+    engine.current_step_id = None
+    mock_context.data_engine = engine
+    meter = get_meter("lens-meter-session")
+    meter.last_prompt_tokens = 777
+
+    summarizer = VisualStepSummarizer(mock_context)
+    summarizer._llm = Mock()
+    summarizer._llm.ainvoke = AsyncMock(
+        return_value=AIMessage(
+            content=(
+                "In Step 1, I tapped the 'Battery' row marked in red; the list"
+                " showed 'Display' and 'Battery'."
+            ),
+            usage_metadata={"input_tokens": 120, "output_tokens": 30, "total_tokens": 150},
+        )
+    )
+
+    summarizer.dispatch(
+        step_number=1,
+        action_name="click",
+        action_args={"target": [1, 2]},
+        pre_img_bytes=b"pre",
+        post_img_bytes=b"post",
+        exec_outcome="executed",
+    )
+    await summarizer.flush()
+
+    usage_calls = [
+        c
+        for c in engine.record_trace.call_args_list
+        if c.kwargs.get("name") == "llm_usage"
+    ]
+    assert usage_calls, "the raw lens bypass must record an llm_usage trace"
+    payload = usage_calls[0].kwargs["payload"]
+    assert payload["source"].startswith("lens:visual_transition:")
+    assert payload["prompt_tokens"] == 120
+    assert meter.last_prompt_tokens == 777
+
+
 def test_flash_config_and_builder():
     """Verify Flash profile configuration model and SDK builder fluent API."""
     cfg = Builders.AgentConfig.with_flash_config(
@@ -395,7 +602,13 @@ async def test_data_engine_integration_with_step_summarizer(mock_context):
     )
     await summarizer.flush()
 
-    mock_de.update_step_summary.assert_called_once_with(1, "[Step 1 State: Input text confirmed.]")
+    mock_de.update_step_summary.assert_called_once_with(
+        1,
+        "[Step 1 State: Input text confirmed.]",
+        status="ready",
+        source="visual_transition",
+        model=ANY,
+    )
 
 
 def test_draw_action_overlay_rendering():

tests/unit/agents/test_historical_state_hint.py

@@ -0,0 +1,201 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""HistoricalStateHintPromptComponent (M4): perceptual-hash hint + silence rules.
+
+Division of labor under test: the hint fires only for matches *older* than the
+recent window (returned-to-an-early-state regime); any match within the last 3
+steps silences it entirely — that regime belongs to the pixel-level
+ScreenshotSimilarityPromptComponent.
+"""
+
+import base64
+import io
+from types import SimpleNamespace
+from unittest.mock import MagicMock, patch
+
+import pytest
+from PIL import Image, ImageDraw
+
+from artemis.agents.operator.prompts import HistoricalStateHintPromptComponent, PromptBuilder
+from artemis.utils.image_hash import dhash_hex
+
+
+def _jpeg(draw_fn) -> bytes:
+    img = Image.new("RGB", (270, 600), "white")
+    draw_fn(ImageDraw.Draw(img))
+    buf = io.BytesIO()
+    img.save(buf, format="JPEG", quality=85)
+    return buf.getvalue()
+
+
+def _screen_a(d):
+    d.rectangle([20, 40, 250, 90], fill="#3355ff")
+    d.rectangle([60, 450, 210, 510], fill="#22aa66")
+
+
+def _screen_b(d):
+    for y in range(60, 560, 80):
+        d.rectangle([10, y, 260, y + 50], fill="#dddddd")
+
+
+SCREEN_A = _jpeg(_screen_a)
+SCREEN_B = _jpeg(_screen_b)
+HASH_A = dhash_hex(SCREEN_A)
+HASH_B = dhash_hex(SCREEN_B)
+B64_A = base64.b64encode(SCREEN_A).decode("utf-8")
+
+
+def _step(number: int, post_hash: str | None) -> dict:
+    meta = {"post_image_dhash": post_hash} if post_hash else {}
+    return {"step_number": number, "extra_metadata": meta}
+
+
+def _cfg(similarity_hint=True, max_distance=8):
+    return SimpleNamespace(
+        memory=SimpleNamespace(
+            transcript=SimpleNamespace(
+                similarity_hint=similarity_hint,
+                similarity_max_distance=max_distance,
+            )
+        )
+    )
+
+
+async def _run(steps, b64=B64_A, cfg=None):
+    builder = PromptBuilder()
+    component = HistoricalStateHintPromptComponent()
+    with patch("artemis.config.load_agent_config", return_value=cfg or _cfg()):
+        await component(
+            builder, MagicMock(), MagicMock(), latest_screenshot_b64=b64, steps=steps
+        )
+    return [p for p in builder.human_parts if isinstance(p, str)]
+
+
+@pytest.mark.asyncio
+async def test_hint_fires_on_old_matching_step():
+    steps = [
+        _step(1, HASH_B),
+        _step(2, HASH_A),  # old match
+        _step(3, HASH_B),
+        _step(4, HASH_B),
+        _step(5, HASH_B),
+        _step(6, HASH_B),
+    ]
+    parts = await _run(steps)
+    assert len(parts) == 1
+    assert "Historical state hint" in parts[0]
+    assert "Step 2" in parts[0]
+    assert "recall_history" in parts[0]
+
+
+@pytest.mark.asyncio
+async def test_silent_when_recent_step_also_matches():
+    steps = [
+        _step(1, HASH_A),  # old match exists...
+        _step(2, HASH_B),
+        _step(3, HASH_B),
+        _step(4, HASH_B),
+        _step(5, HASH_A),  # ...but a recent (last-3) match silences the hint
+        _step(6, HASH_B),
+    ]
+    assert await _run(steps) == []
+
+
+@pytest.mark.asyncio
+async def test_silent_without_any_match():
+    steps = [_step(n, HASH_B) for n in range(1, 7)]
+    assert await _run(steps) == []
+
+
+@pytest.mark.asyncio
+async def test_silent_when_config_disabled():
+    steps = [
+        _step(1, HASH_A),
+        _step(2, HASH_B),
+        _step(3, HASH_B),
+        _step(4, HASH_B),
+        _step(5, HASH_B),
+        _step(6, HASH_B),
+    ]
+    assert await _run(steps, cfg=_cfg(similarity_hint=False)) == []
+
+
+@pytest.mark.asyncio
+async def test_silent_with_too_few_steps():
+    # Everything within the recent window belongs to the pixel-level note.
+    assert await _run([_step(1, HASH_A), _step(2, HASH_A)]) == []
+
+
+@pytest.mark.asyncio
+async def test_tie_prefers_most_recent_old_step():
+    steps = [
+        _step(1, HASH_A),
+        _step(2, HASH_A),  # same distance, more recent
+        _step(3, HASH_B),
+        _step(4, HASH_B),
+        _step(5, HASH_B),
+        _step(6, HASH_B),
+    ]
+    parts = await _run(steps)
+    assert len(parts) == 1 and "Step 2" in parts[0]
+
+
+@pytest.mark.asyncio
+async def test_steps_without_hashes_are_skipped():
+    steps = [
+        _step(1, None),
+        _step(2, None),
+        _step(3, None),
+        _step(4, None),
+        _step(5, None),
+        _step(6, None),
+    ]
+    assert await _run(steps) == []
+
+
+def test_record_step_stamps_perceptual_hashes(tmp_path):
+    from artemis.context import ArtemisContext
+    from artemis.data_engine.engine import DataEngine
+
+    mock_ctx = MagicMock(spec=ArtemisContext)
+    setup = MagicMock()
+    setup.traces_path = str(tmp_path)
+    mock_ctx.execution_setup = setup
+    mock_ctx.device = None
+    engine = DataEngine(mock_ctx)
+    engine.start_session("dhash stamping test")
+
+    engine.record_step(
+        pre_screenshot_bytes=SCREEN_A,
+        post_screenshot_bytes=SCREEN_B,
+        summary="two distinct screens",
+    )
+    # Same pre/post bytes: post_image_name is nulled but the post-action
+    # screen IS the pre screen, so the post hash mirrors the pre hash.
+    engine.record_step(
+        pre_screenshot_bytes=SCREEN_A,
+        post_screenshot_bytes=SCREEN_A,
+        summary="screen unchanged",
+    )
+    for t in list(engine._pending_threads):
+        t.join()
+
+    steps = engine.storage.get_steps(engine.current_session_id)
+    meta1 = steps[0].extra_metadata
+    assert meta1["pre_image_dhash"] == HASH_A
+    assert meta1["post_image_dhash"] == HASH_B
+    meta2 = steps[1].extra_metadata
+    assert meta2["pre_image_dhash"] == HASH_A
+    assert meta2["post_image_dhash"] == HASH_A

tests/unit/agents/test_history_analyzer.py

@@ -344,7 +344,7 @@ async def test_history_analyzer_integration_with_task_tree():
     mock_path.exists.return_value = True
     mock_path.read_text.return_value = fake_plan
 
-    # Mock build_plan_and_history
+    # Mock the policy-table compiled-history renderer (M4)
     with (
         patch(
             "artemis.agents.history_analyzer.history_analyzer.get_llm",
@@ -355,23 +355,23 @@ async def test_history_analyzer_integration_with_task_tree():
             return_value=mock_path,
         ),
         patch(
-            "artemis.agents.history_analyzer.history_analyzer.build_plan_and_history"
+            "artemis.agents.history_analyzer.history_analyzer.build_history_for"
         ) as mock_build,
     ):
         mock_build.return_value = "Mocked operation history"
         analyzer = HistoryAnalyzer(mock_ctx)
         await analyzer.run("query")
 
-        # Verify build_plan_and_history call args
+        # Verify build_history_for call args
         mock_build.assert_called_once()
         args, kwargs = mock_build.call_args
-        assert args[0] == fake_plan
-        assert args[1] == steps
+        assert args[0] == "history_analyzer"
+        assert args[1] == fake_plan
+        assert args[2] == steps
 
         import hashlib
 
         expected_hash = hashlib.md5(b"Active Subgoal").hexdigest()
-        assert args[2] == expected_hash
+        assert args[3] == expected_hash
 
-        assert kwargs.get("last_n_detailed") == 1
-        assert kwargs.get("min_summaries") == len(steps)
+        assert kwargs.get("engine") is mock_ctx.data_engine

tests/unit/agents/test_operator.py

@@ -16,9 +16,15 @@ import json
 from unittest.mock import AsyncMock, MagicMock, Mock, patch
 
 from artemis.agents.operator.operator import OperatorNode
+from artemis.config.agent import MemoryTranscriptConfig
 from artemis.context import ArtemisContext
 import pytest
 
+# These tests exercise the legacy 2-message prompt path. Since M5 the
+# transcript path is the default, so flag-off semantics must be requested
+# explicitly (never via the config default).
+LEGACY_TRANSCRIPT = MemoryTranscriptConfig(enabled=False)
+
 
 @pytest.mark.asyncio
 async def test_operator_node_fast_path():
@@ -58,7 +64,7 @@ async def test_operator_node_fast_path():
 
     # Patch dependencies
     with patch("artemis.agents.operator.operator.get_llm", return_value=mock_llm):
-        node = OperatorNode(mock_ctx)
+        node = OperatorNode(mock_ctx, transcript_config=LEGACY_TRANSCRIPT)
         node_update = await node(mock_state)
 
         # Verify state was updated with action
@@ -77,7 +83,7 @@ async def test_perform_action_validation():
     from artemis.graph.state import State
 
     mock_ctx = MagicMock()
-    node = OperatorNode(mock_ctx)
+    node = OperatorNode(mock_ctx, transcript_config=LEGACY_TRANSCRIPT)
 
     mock_state = MagicMock(spec=State)
     mock_state.indexed_points = [[200, 300], [400, 500]]
@@ -114,6 +120,7 @@ async def test_perform_action_validation():
             "target_bounds": "[100,100][300,500]",
             "target_resource_id": "btn_1",
             "target_class": "android.widget.Button",
+            "target_label_source": "index",
         }
     ]
 
@@ -143,6 +150,7 @@ async def test_perform_action_validation():
         "target_bounds": "[300,400][500,600]",
         "target_resource_id": "input_1",
         "target_class": "android.widget.EditText",
+        "target_label_source": "index",
     }
 
     # Test swipe direction
@@ -170,6 +178,7 @@ async def test_perform_action_validation():
             "target_bounds": "[300,400][500,600]",
             "target_resource_id": "input_1",
             "target_class": "android.widget.EditText",
+            "target_label_source": "index",
         }
     ]
 
@@ -222,6 +231,59 @@ async def test_perform_action_validation():
     assert "Unsupported app management action" in err
 
 
+@pytest.mark.asyncio
+async def test_bare_coordinate_click_enriched_by_hit_test():
+    """A coordinate click hit tests the pre-action frame for element semantics."""
+    from artemis.agents.operator.operator import OperatorNode
+    from unittest.mock import MagicMock
+    from artemis.graph.state import State
+
+    mock_ctx = MagicMock()
+    node = OperatorNode(mock_ctx, transcript_config=LEGACY_TRANSCRIPT)
+
+    mock_state = MagicMock(spec=State)
+    mock_state.indexed_elements = [
+        {
+            "center": [550, 1440],
+            "text": "Confirm",
+            "bounds": [500, 1400, 600, 1480],
+            "class": "android.widget.Button",
+            "resource_id": "btn_confirm",
+            "is_ocr": False,
+        }
+    ]
+
+    # Normalized (500, 600) on the default 1080x2400 device -> pixel (540, 1440),
+    # which falls inside the "Confirm" button bounds.
+    actions, err = node._translate_and_validate_tool(
+        {"name": "click", "args": {"target": [500, 600]}}, mock_state
+    )
+    assert err is None
+    assert actions == [
+        {
+            "action": "tap",
+            "coordinates": [540, 1440],
+            "normalized_coordinates": [500, 600],
+            "times": 1,
+            "delay_ms": 100,
+            "target_text": "Confirm",
+            "target_bounds": [500, 1400, 600, 1480],
+            "target_resource_id": "btn_confirm",
+            "target_class": "android.widget.Button",
+            "target_label_source": "hit_test",
+        }
+    ]
+
+    # Graceful degradation: no perception data at all -> bare coordinates, "none".
+    mock_state.indexed_elements = []
+    actions, err = node._translate_and_validate_tool(
+        {"name": "click", "args": {"target": [500, 600]}}, mock_state
+    )
+    assert err is None
+    assert actions[0]["target_text"] is None
+    assert actions[0]["target_label_source"] == "none"
+
+
 @pytest.mark.asyncio
 async def test_operator_node_multiple_actions():
     from artemis.agents.operator.operator import OperatorNode
@@ -258,7 +320,7 @@ async def test_operator_node_multiple_actions():
     mock_llm.bind_tools.return_value = mock_llm
 
     with patch("artemis.agents.operator.operator.get_llm", return_value=mock_llm):
-        node = OperatorNode(mock_ctx)
+        node = OperatorNode(mock_ctx, transcript_config=LEGACY_TRANSCRIPT)
         node_update = await node(mock_state)
 
         update_dict = node_update
@@ -302,7 +364,7 @@ async def test_operator_node_no_record_step():
     mock_llm.bind_tools.return_value = mock_llm
 
     with patch("artemis.agents.operator.operator.get_llm", return_value=mock_llm):
-        node = OperatorNode(mock_ctx)
+        node = OperatorNode(mock_ctx, transcript_config=LEGACY_TRANSCRIPT)
         node_update = await node(mock_state)
 
         mock_ctx.data_engine.record_step.assert_not_called()
@@ -352,7 +414,7 @@ async def test_operator_dynamic_prompt():
     mock_llm.bind_tools.return_value = mock_llm
 
     with patch("artemis.agents.operator.operator.get_llm", return_value=mock_llm):
-        node = OperatorNode(mock_ctx, prompt_components=[MockComponent()])
+        node = OperatorNode(mock_ctx, prompt_components=[MockComponent()], transcript_config=LEGACY_TRANSCRIPT)
         node_update = await node(mock_state)
 
 
@@ -488,7 +550,7 @@ async def test_operator_history_compression():
         patch("pathlib.Path.exists", return_value=True),
         patch("pathlib.Path.read_text", mock_read_text),
     ):
-        node = OperatorNode(mock_ctx, last_n_detailed=5)
+        node = OperatorNode(mock_ctx, last_n_detailed=5, transcript_config=LEGACY_TRANSCRIPT)
         node_update = await node(mock_state)
 
 
@@ -506,7 +568,7 @@ async def test_operator_loop_detection():
     mock_state.subagent_calls = ["ask_explorer", "ask_explorer", "ask_explorer"]
     mock_state.initial_goal = "Test goal"
 
-    node = OperatorNode(mock_ctx)
+    node = OperatorNode(mock_ctx, transcript_config=LEGACY_TRANSCRIPT)
 
     with pytest.raises(RuntimeError) as exc_info:
         node_update = await node(mock_state)
@@ -568,7 +630,7 @@ async def test_operator_tracks_subagent_calls():
             side_effect=lambda t, ctx: t,
         ),
     ):
-        node = OperatorNode(mock_ctx, tools=[mock_diagnoser_tool])
+        node = OperatorNode(mock_ctx, tools=[mock_diagnoser_tool], transcript_config=LEGACY_TRANSCRIPT)
         node_update = await node(mock_state)
 
         update_dict = node_update
@@ -660,7 +722,7 @@ async def test_operator_defer_action_for_gathering():
             side_effect=lambda t, ctx: t,
         ),
     ):
-        node = OperatorNode(mock_ctx, tools=[mock_diagnoser_tool])
+        node = OperatorNode(mock_ctx, tools=[mock_diagnoser_tool], transcript_config=LEGACY_TRANSCRIPT)
         node_update = await node(mock_state)
 
         # Subagent tool should have been executed
@@ -683,7 +745,7 @@ async def test_wait_actions_translation():
     from artemis.graph.state import State
 
     mock_ctx = MagicMock()
-    node = OperatorNode(mock_ctx)
+    node = OperatorNode(mock_ctx, transcript_config=LEGACY_TRANSCRIPT)
     mock_state = MagicMock(spec=State)
 
     # 1. Test wait_for_delay translation
@@ -708,7 +770,7 @@ async def test_long_press_action_translation():
     from artemis.graph.state import State
 
     mock_ctx = MagicMock()
-    node = OperatorNode(mock_ctx)
+    node = OperatorNode(mock_ctx, transcript_config=LEGACY_TRANSCRIPT)
     mock_state = MagicMock(spec=State)
     mock_state.indexed_points = [[200, 300], [400, 500]]
     mock_state.indexed_elements = [
@@ -744,6 +806,7 @@ async def test_long_press_action_translation():
             "target_bounds": "[100,100][300,500]",
             "target_resource_id": "btn_1",
             "target_class": "android.widget.Button",
+            "target_label_source": "index",
         }
     ]
 
@@ -765,6 +828,7 @@ async def test_long_press_action_translation():
             "target_bounds": None,
             "target_resource_id": None,
             "target_class": None,
+            "target_label_source": "none",
         }
     ]
 
@@ -844,7 +908,7 @@ async def test_operator_no_defer_for_note_updating():
             side_effect=lambda t, ctx: t,
         ),
     ):
-        node = OperatorNode(mock_ctx, tools=[mock_update_note_tool])
+        node = OperatorNode(mock_ctx, tools=[mock_update_note_tool], transcript_config=LEGACY_TRANSCRIPT)
         node_update = await node(mock_state)
 
         # The update_note tool should have been executed immediately
@@ -934,7 +998,7 @@ async def test_operator_defer_for_note_reading():
             side_effect=lambda t, ctx: t,
         ),
     ):
-        node = OperatorNode(mock_ctx, tools=[mock_read_note_tool])
+        node = OperatorNode(mock_ctx, tools=[mock_read_note_tool], transcript_config=LEGACY_TRANSCRIPT)
         node_update = await node(mock_state)
 
         # The read_note tool should have been executed
@@ -1022,7 +1086,7 @@ async def test_operator_no_defer_for_note_appending():
             side_effect=lambda t, ctx: t,
         ),
     ):
-        node = OperatorNode(mock_ctx, tools=[mock_append_note_tool])
+        node = OperatorNode(mock_ctx, tools=[mock_append_note_tool], transcript_config=LEGACY_TRANSCRIPT)
         node_update = await node(mock_state)
 
         # The append_note tool should have been executed immediately
@@ -1111,7 +1175,7 @@ async def test_operator_background_tasks_prompt_injection():
                 side_effect=mock_invoke_llm_loop,
             ),
         ):
-            node = OperatorNode(mock_ctx)
+            node = OperatorNode(mock_ctx, transcript_config=LEGACY_TRANSCRIPT)
             node_update = await node(mock_state)
 
             # Assert that the captured messages (prompts) contain the background tasks section
@@ -1202,7 +1266,7 @@ async def test_operator_task_plan_warning_prompt_injection():
             side_effect=mock_invoke_llm_loop_1,
         ),
     ):
-        node = OperatorNode(mock_ctx)
+        node = OperatorNode(mock_ctx, transcript_config=LEGACY_TRANSCRIPT)
         node_update = await node(mock_state)
 
         full_content_1 = ""
@@ -1255,7 +1319,7 @@ async def test_operator_task_plan_warning_prompt_injection():
             side_effect=mock_invoke_llm_loop_2,
         ),
     ):
-        node = OperatorNode(mock_ctx)
+        node = OperatorNode(mock_ctx, transcript_config=LEGACY_TRANSCRIPT)
         node_update = await node(mock_state)
 
         full_content_2 = ""
@@ -1322,7 +1386,7 @@ async def test_operator_tool_limit_exceeded_warning():
             side_effect=mock_invoke_llm_loop,
         ),
     ):
-        node = OperatorNode(mock_ctx)
+        node = OperatorNode(mock_ctx, transcript_config=LEGACY_TRANSCRIPT)
         node_update = await node(mock_state)
 
         full_content = ""
@@ -1380,7 +1444,7 @@ async def test_operator_fallback_function_call_parsing():
     mock_llm.bind_tools.return_value = mock_llm
 
     with patch("artemis.agents.operator.operator.get_llm", return_value=mock_llm):
-        node = OperatorNode(mock_ctx)
+        node = OperatorNode(mock_ctx, transcript_config=LEGACY_TRANSCRIPT)
         node_update = await node(mock_state)
 
         # Verify state was updated with parsed click action
@@ -1402,7 +1466,7 @@ async def test_operator_swipe_translation():
     mock_ctx = MagicMock(spec=ArtemisContext)
     mock_state = MagicMock()
     mock_state.operator_raw_data = {"width": 1080, "height": 2400}
-    node = OperatorNode(mock_ctx)
+    node = OperatorNode(mock_ctx, transcript_config=LEGACY_TRANSCRIPT)
 
     # 1. Directional swipe
     actions, err = node._translate_and_validate_tool(

tests/unit/agents/test_operator_prompts.py

@@ -24,7 +24,7 @@ def test_operator_prompt_contract_matches_note_runtime_semantics():
 
     assert "write-through tools `update_note` and `append_note`" in prompt
     assert "may accompany at most one Turn-Ending Action" in prompt
-    assert "`read_note`, `list_notes`, and `save_note`" in prompt
+    assert "`read_note`, `list_notes`, `save_note`, and `recall_history`" in prompt
     assert "Do NOT submit a Turn-Ending Action at the same time" not in prompt
     assert "memory note tools (`read_note`, `list_notes`, `save_note`, `update_note`" not in prompt
 

tests/unit/agents/test_operator_transcript.py

@@ -0,0 +1,311 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Operator dual prompt path (M2): legacy golden parity and transcript mode.
+
+Flag off (``agent.memory.transcript.enabled=false``): the legacy 2-message
+build must stay byte-for-byte identical — pinned by SHA-256 goldens recorded
+against the pre-split implementation.
+
+Flag on: ``_build_prompt`` renders ``S + F + A + tail`` from the session
+transcript ledger; the turn's messages are staged after the tool loop and
+committed (with step key and validator result) at the next build.
+"""
+
+import base64
+import hashlib
+import re
+from types import SimpleNamespace
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+from langchain_core.messages import HumanMessage, SystemMessage
+
+from artemis.agents.operator.operator import OperatorNode
+from artemis.agents.operator.prompt_builder import load_operator_prompts
+from artemis.agents.operator.prompts import (
+    PLAN_HISTORY_STATIC_POINTER,
+    PLAN_HISTORY_TEMPLATE_SECTION,
+    PromptBuilder,
+    TemplatePromptComponent,
+    render_transcript_static_system,
+)
+from artemis.config.agent import MemoryTranscriptConfig
+from artemis.context import ArtemisContext
+from artemis.memory.step_memory import StepMemoryService
+from artemis.memory.transcript import (
+    EXECUTION_RESULT_MARKER,
+    PLAN_RECITATION_MARKER,
+    RESTORED_HISTORY_HEADER,
+    TranscriptLedger,
+)
+
+# SHA-256 of the legacy system message rendered by TemplatePromptComponent
+# with the fixed inputs below. Originally recorded BEFORE the M2 template
+# split; re-recorded once in M5 (2026-09-01) after the only intended template
+# change since — removing the Short-Term Memory instruction block from
+# operator.json (the <short_term_memory> relay was removed end to end per
+# pro-context-memory-redesign §3.3). Any other drift here means the flag-off
+# path is no longer byte-identical.
+GOLDEN_EMPTY_PLAN = "c45aec7d67069dad3a195534dd3c97dadb48454ca186a40eb471d5f9136589fa"
+GOLDEN_SENTINEL_PLAN = "670580efd48f237c3257d996a814145fbef4ec0bb467aa4dfa6eb2db4a613826"
+
+SCREENSHOT_B64 = base64.b64encode(b"fake-jpeg-bytes").decode("utf-8")
+
+
+async def _render_legacy_system(plan_and_history: str) -> str:
+    builder = PromptBuilder()
+    component = TemplatePromptComponent()
+    ctx = SimpleNamespace(execution_setup=None, actuator=None, data_engine=None)
+    state = SimpleNamespace(initial_goal="GOLDEN_FIXED_GOAL")
+    await component(
+        builder,
+        state,
+        ctx,
+        prompts=load_operator_prompts(),
+        template_name="main_template",
+        plan_and_history=plan_and_history,
+    )
+    return "".join(builder.system_parts)
+
+
+@pytest.mark.asyncio
+async def test_legacy_prompt_bytes_match_pre_split_golden():
+    empty = await _render_legacy_system("No plan or history yet.")
+    sentinel = await _render_legacy_system("PLAN_AND_HISTORY_SENTINEL\nline2")
+    assert hashlib.sha256(empty.encode("utf-8")).hexdigest() == GOLDEN_EMPTY_PLAN
+    assert hashlib.sha256(sentinel.encode("utf-8")).hexdigest() == GOLDEN_SENTINEL_PLAN
+
+
+def test_main_template_contains_exactly_one_plan_history_section():
+    template = load_operator_prompts()["main_template"]
+    assert template.count(PLAN_HISTORY_TEMPLATE_SECTION) == 1
+
+
+def test_transcript_static_system_is_stable_and_carries_no_history():
+    ctx = SimpleNamespace(execution_setup=None, actuator=None, data_engine=None)
+    state = SimpleNamespace(initial_goal="STATIC_GOAL")
+    prompts = load_operator_prompts()
+
+    first = render_transcript_static_system(prompts, ctx, state)
+    second = render_transcript_static_system(prompts, ctx, state)
+    assert first == second  # byte-stable S region
+
+    assert "STATIC_GOAL" in first
+    # The volatile section is replaced by the static pointer...
+    assert PLAN_HISTORY_STATIC_POINTER.splitlines()[1] in first
+    # ...and no unrendered slot or trailing observation header survives.
+    assert "{{" not in first
+    assert not first.endswith("# CURRENT OBSERVATION\n")
+
+
+def _transcript_ctx():
+    ctx = MagicMock(spec=ArtemisContext)
+    ctx.execution_setup = None
+    ctx.actuator = None
+    ctx.data_engine = None
+    ctx.step_memory = StepMemoryService(ctx=None)
+    ctx.transcript_ledger = None
+    return ctx
+
+
+def _transcript_state(**overrides):
+    state = MagicMock()
+    state.subagent_calls = []
+    state.initial_goal = "Transcript goal"
+    state.operator_feedback = None
+    state.injected_instruction = None
+    state.operator_tool_limit_exceeded = False
+    state.structured_decisions = None
+    state.last_execution_result = None
+    state.current_step_id = None
+    state.latest_ui_hierarchy = []
+    state.operator_raw_data = {
+        "screenshot_b64": SCREENSHOT_B64,
+        "xml_hierarchy": [],
+        "ocr_results": None,
+        "width": 1080,
+        "height": 2400,
+    }
+    for key, value in overrides.items():
+        setattr(state, key, value)
+    return state
+
+
+def _no_action_llm(captured: list):
+    mock_llm = MagicMock()
+    mock_response = MagicMock()
+    mock_response.tool_calls = []
+    mock_response.content = "no action this turn"
+
+    async def mock_ainvoke(*args, **kwargs):
+        captured.append(list(args[0]))
+        return mock_response
+
+    mock_llm.ainvoke = AsyncMock(side_effect=mock_ainvoke)
+    mock_llm.bind_tools.return_value = mock_llm
+    return mock_llm
+
+
+@pytest.mark.asyncio
+async def test_transcript_mode_two_turns_build_four_regions(tmp_path):
+    ctx = _transcript_ctx()
+    # Regression data shape (M2 cold-start bug, on-device A/B 2026-09-01):
+    # turn 1 runs before any step record exists; by turn 2 the first step HAS
+    # been recorded, so `steps` is non-empty while the previous turn is still
+    # staged (turn_count == 0). The cold-start check must not read that
+    # live-session state as a process restart — seeding would be rejected by
+    # the empty-ledger invariant and silently demote every turn to the legacy
+    # 2-message path.
+    ctx.data_engine = MagicMock()
+    ctx.data_engine.base_dir = str(tmp_path)
+    ctx.data_engine.global_base_dir = str(tmp_path)
+    ctx.data_engine.get_agent_friendly_steps.return_value = []
+    captured: list = []
+    mock_llm = _no_action_llm(captured)
+
+    with patch("artemis.agents.operator.operator.get_llm", return_value=mock_llm):
+        node = OperatorNode(ctx, transcript_config=MemoryTranscriptConfig(enabled=True))
+
+        # Turn 1
+        await node(_transcript_state())
+        ledger = ctx.transcript_ledger
+        assert isinstance(ledger, TranscriptLedger)
+        assert ledger.has_static_prefix
+        assert ledger.has_staged_turn
+        assert ledger.turn_count == 0
+
+        turn1_messages = captured[0]
+        assert isinstance(turn1_messages[0], SystemMessage)
+        assert isinstance(turn1_messages[-1], HumanMessage)
+        tail1_text = str(turn1_messages[-1].content)
+        assert "# CURRENT OBSERVATION [T+" in tail1_text
+        assert PLAN_RECITATION_MARKER in tail1_text
+
+        # Turn 2: step 1 is now recorded, and the previous turn is committed
+        # with its step id and result.
+        ctx.data_engine.get_agent_friendly_steps.return_value = [
+            {
+                "step_id": "step-1",
+                "step_number": 1,
+                "summary": "Step 1 summary",
+                "relative_time": "1.0s",
+                "extra_metadata": {"subgoal_hash": "default"},
+            }
+        ]
+        await node(
+            _transcript_state(
+                current_step_id="step-1",
+                structured_decisions='[{"action": "tap"}]',
+                last_execution_result={"status": "success"},
+            )
+        )
+
+        # The staged-but-uncommitted turn 2 state never triggers cold-start
+        # seeding (the transcript path stayed live instead of falling back).
+        assert not ledger.has_restored_history
+        assert ledger.turn_count == 1
+        turn2_messages = captured[1]
+        # S region is the same object with the same bytes.
+        assert turn2_messages[0] is turn1_messages[0]
+
+        active = list(ledger.active_messages)
+        # Committed turn-1 observation followed by its validator result.
+        assert any(EXECUTION_RESULT_MARKER in str(m.content) for m in active)
+        committed_obs = active[0]
+        # Depth-1 scrub removed the old plan recitation copy.
+        assert PLAN_RECITATION_MARKER not in str(committed_obs.content)
+        # The live tail still recites the plan.
+        assert PLAN_RECITATION_MARKER in str(turn2_messages[-1].content)
+        # Message layout is S + A + tail.
+        assert turn2_messages[-1] is not turn1_messages[-1]
+        assert re.search(r"T\+\d{2,}:\d{2}", str(turn2_messages[-1].content))
+
+
+@pytest.mark.asyncio
+async def test_transcript_actionless_turn_commits_without_validator_message():
+    ctx = _transcript_ctx()
+    captured: list = []
+    mock_llm = _no_action_llm(captured)
+
+    with patch("artemis.agents.operator.operator.get_llm", return_value=mock_llm):
+        node = OperatorNode(ctx, transcript_config=MemoryTranscriptConfig(enabled=True))
+        await node(_transcript_state())
+        # Planner-rejected turns clear structured_decisions: no validator ran.
+        await node(
+            _transcript_state(
+                current_step_id="step-1",
+                structured_decisions="",
+                last_execution_result={"status": "success"},
+            )
+        )
+
+    ledger = ctx.transcript_ledger
+    assert ledger.turn_count == 1
+    assert not any(
+        EXECUTION_RESULT_MARKER in str(m.content) for m in ledger.active_messages
+    )
+
+
+@pytest.mark.asyncio
+async def test_transcript_cold_start_builds_restored_history_block(tmp_path):
+    ctx = _transcript_ctx()
+    ctx.data_engine = MagicMock()
+    ctx.data_engine.base_dir = str(tmp_path)
+    ctx.data_engine.global_base_dir = str(tmp_path)
+    ctx.data_engine.get_agent_friendly_steps.return_value = [
+        {
+            "step_id": f"step_{i}",
+            "step_number": i,
+            "summary": f"Step {i} summary",
+            "relative_time": f"{i}.0s",
+            "extra_metadata": {"subgoal_hash": "default"},
+        }
+        for i in range(1, 4)
+    ]
+
+    captured: list = []
+    mock_llm = _no_action_llm(captured)
+
+    with patch("artemis.agents.operator.operator.get_llm", return_value=mock_llm):
+        node = OperatorNode(ctx, transcript_config=MemoryTranscriptConfig(enabled=True))
+        await node(_transcript_state())
+
+    ledger = ctx.transcript_ledger
+    assert ledger.has_restored_history
+
+    messages = captured[0]
+    restored_text = str(messages[1].content)
+    assert RESTORED_HISTORY_HEADER in restored_text
+    assert "--- Execution History ---" in restored_text
+    assert "Step 1 summary" in restored_text
+    # The live tail still follows normally and A stays appendable.
+    assert "# CURRENT OBSERVATION [T+" in str(messages[-1].content)
+    assert ledger.has_staged_turn
+
+
+@pytest.mark.asyncio
+async def test_flag_off_keeps_legacy_two_message_build_and_no_ledger():
+    ctx = _transcript_ctx()
+    captured: list = []
+    mock_llm = _no_action_llm(captured)
+
+    with patch("artemis.agents.operator.operator.get_llm", return_value=mock_llm):
+        node = OperatorNode(ctx, transcript_config=MemoryTranscriptConfig(enabled=False))
+        await node(_transcript_state())
+
+    assert len(captured[0]) == 2
+    assert isinstance(captured[0][0], SystemMessage)
+    assert isinstance(captured[0][1], HumanMessage)
+    assert ctx.transcript_ledger is None

tests/unit/agents/test_outputter.py

@@ -286,6 +286,9 @@ async def test_outputter_builds_concise_history(mock_get_llm):
         }
     ]
     mock_data_engine.get_agent_friendly_steps.return_value = steps
+    # M5: the transcript flag defaults on, so the outputter's compiled view
+    # consults the chunk store; no persisted chunks keeps the classic path.
+    mock_data_engine.get_history_chunks.return_value = []
     mock_data_engine.base_dir = "/tmp/mock_session"
     ctx.data_engine = mock_data_engine
 

tests/unit/agents/test_prompt_assembly.py

@@ -92,7 +92,7 @@ def test_full_toolset_is_the_default():
         )
 
 
-@pytest.mark.parametrize("tool", OPERATOR_DEVICE_TOOLS + ("video_analyzer",))
+@pytest.mark.parametrize("tool", OPERATOR_DEVICE_TOOLS + ("video_analyzer", "recall_history"))
 def test_removed_tool_leaves_no_reference(tool):
     """The executable definition of 'an absent tool costs the model nothing'."""
     for name, template in load_operator_prompts().items():
@@ -154,8 +154,8 @@ def test_full_set_enum_slots_render_verbatim():
     )
     assert (
         "(`ask_explorer`, `ask_diagnoser`, `video_analyzer`, `run_adb_command`,"
-        " `manage_task`, `analyze_task_output`, `read_note`, `list_notes`, and"
-        " `save_note`)" in out
+        " `manage_task`, `analyze_task_output`, `read_note`, `list_notes`,"
+        " `save_note`, and `recall_history`)" in out
     )
 
 

tests/unit/agents/test_summarizer.py

@@ -12,11 +12,21 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import json
-from unittest.mock import MagicMock, Mock, patch
+"""SummarizerNode tests (M2: visual-transition lens dispatch).
+
+Per the history redesign §5 the Pro SummarizerNode no longer runs its own LLM
+call to produce a text capsule: it dispatches the step to the shared
+step-memory service (``ctx.step_memory``) with the pre/post screenshot bytes
+pulled from the DataEngine, and degrades gracefully when images are missing.
+"""
+
+from types import SimpleNamespace
+from unittest.mock import MagicMock, Mock
+
+import pytest
+
 from artemis.agents.summarizer.summarizer import SummarizerNode
 from artemis.context import ArtemisContext
-import pytest
 
 
 class DummyState:
@@ -36,181 +46,107 @@ class DummyState:
 
 
 @pytest.fixture
-def mock_context():
+def mock_context(tmp_path):
     ctx = Mock(spec=ArtemisContext)
     ctx.device = None
-    ctx.data_engine = Mock()
     ctx.background_tasks = []
+    ctx.step_memory = MagicMock()
+
+    pre_path = tmp_path / "pre.jpg"
+    post_path = tmp_path / "post.jpg"
+    pre_path.write_bytes(b"PRE_IMAGE")
+    post_path.write_bytes(b"POST_IMAGE")
+
+    engine = Mock()
+    engine.get_step_number.return_value = 7
+    engine.get_step_record.return_value = SimpleNamespace(
+        action_taken=[{"action": "tap", "coordinates": [500, 600], "target_text": "Search"}]
+    )
+    engine.get_step_image_path.side_effect = lambda step_number, which="pre": (
+        pre_path if which == "pre" else post_path
+    )
+    ctx.data_engine = engine
     return ctx
 
 
 @pytest.mark.asyncio
-async def test_summarizer_success(mock_context):
-    """Test that SummarizerNode successfully generates summary from raw thinking and execution result."""
-    decisions = json.dumps([{"action": "tap", "coordinates": [500, 600]}])
-    thinking = "The operator clicked the search button."
-    result = {
-        "status": "success",
-        "executed_actions": [{"action": "tap", "coordinates": [500, 600]}],
-    }
-
+async def test_summarizer_dispatches_visual_lens(mock_context):
+    """The node dispatches to the shared service instead of calling an LLM."""
     state = DummyState(
-        structured_decisions=decisions,
-        operator_raw_thinking=thinking,
-        last_execution_result=result,
+        last_execution_result={"status": "success"},
         current_step_id="12345678-1234-5678-1234-567812345678",
     )
 
-    # Mock LLM response
-    mock_response = Mock()
-    mock_response.content = "Step completed successfully."
-    mock_llm = MagicMock()
+    node = SummarizerNode(mock_context)
+    result = await node(state)
 
-    async def mock_astream(*args, **kwargs):
-        yield mock_response
-
-    mock_llm.astream.side_effect = mock_astream
-
-    with (
-        patch(
-            "artemis.agents.summarizer.summarizer.get_llm", return_value=mock_llm
-        ) as mock_get_llm,
-        patch("asyncio.create_task") as mock_create_task,
-    ):
-        # Run the node call
-        node = SummarizerNode(mock_context)
-        await node(state)
-
-        # Verify task was registered in context
-        assert len(mock_context.background_tasks) == 1
-        assert mock_context.background_tasks[0] == mock_create_task.return_value
-
-        # Verify background task was spawned
-        assert mock_create_task.called
-        background_coroutine = mock_create_task.call_args[0][0]
-
-        # Await the background task synchronously for test verification
-        await background_coroutine
-
-    # Verify LLM was invoked with proper structured messages
-    assert mock_llm.astream.called
-    invoked_messages = mock_llm.astream.call_args[0][0]
-
-    system_msg = invoked_messages[0].content
-    human_msg = invoked_messages[1].content
-
-    assert "Summarizer Agent" in system_msg
-    assert "Decisions made" in human_msg
-    assert "The operator clicked the search button." in human_msg
-
-    # Verify Data Engine updated the step summary
-    from uuid import UUID
-
-    mock_context.data_engine.update_step_summary.assert_called_once_with(
-        UUID("12345678-1234-5678-1234-567812345678"),
-        "Step completed successfully.",
+    assert result == {}
+    mock_context.step_memory.dispatch.assert_called_once_with(
+        step_number=7,
+        action_name="tap",
+        action_args={"coordinates": [500, 600], "target_text": "Search"},
+        pre_img_bytes=b"PRE_IMAGE",
+        post_img_bytes=b"POST_IMAGE",
+        exec_outcome="success",
+        data_engine_step_id="12345678-1234-5678-1234-567812345678",
     )
+    # No direct capsule write from the node: the lens owns the versioned
+    # summary write when its background job completes.
+    mock_context.data_engine.update_step_summary.assert_not_called()
 
 
 @pytest.mark.asyncio
-async def test_summarizer_with_native_thinking(mock_context):
-    """Test that SummarizerNode successfully generates summary from both explicit and native thoughts."""
-    decisions = json.dumps([{"action": "tap", "coordinates": [500, 600]}])
-    raw_thinking = "Explicit monologue."
-    native_thinking = "Silent core reasoning."
-    result = {"status": "success"}
+async def test_summarizer_degrades_gracefully_without_images(mock_context):
+    """Missing DataEngine screenshots dispatch with None bytes, no exception."""
+    mock_context.data_engine.get_step_image_path.side_effect = None
+    mock_context.data_engine.get_step_image_path.return_value = None
 
     state = DummyState(
-        structured_decisions=decisions,
-        operator_raw_thinking=raw_thinking,
-        operator_native_thinking=native_thinking,
-        last_execution_result=result,
+        last_execution_result={"status": "failed", "error": "tap missed"},
         current_step_id="12345678-1234-5678-1234-567812345678",
     )
 
-    mock_response = Mock()
-    mock_response.content = "Summary result."
-    mock_llm = MagicMock()
+    node = SummarizerNode(mock_context)
+    await node(state)
 
-    async def mock_astream(*args, **kwargs):
-        yield mock_response
-
-    mock_llm.astream.side_effect = mock_astream
-
-    with (
-        patch(
-            "artemis.agents.summarizer.summarizer.get_llm", return_value=mock_llm
-        ) as mock_get_llm,
-        patch("asyncio.create_task") as mock_create_task,
-    ):
-        node = SummarizerNode(mock_context)
-        await node(state)
-
-        # Verify task was registered in context
-        assert len(mock_context.background_tasks) == 1
-        assert mock_context.background_tasks[0] == mock_create_task.return_value
-
-        background_coroutine = mock_create_task.call_args[0][0]
-        await background_coroutine
-
-    assert mock_llm.astream.called
-    invoked_messages = mock_llm.astream.call_args[0][0]
-    human_msg = invoked_messages[1].content
-
-    assert "Operator explicit thoughts:" in human_msg
-    assert "Explicit monologue." in human_msg
-    assert "Operator native thoughts:" in human_msg
-    assert "Silent core reasoning." in human_msg
+    kwargs = mock_context.step_memory.dispatch.call_args.kwargs
+    assert kwargs["pre_img_bytes"] is None
+    assert kwargs["post_img_bytes"] is None
+    assert kwargs["exec_outcome"] == "Error: tap missed"
 
 
 @pytest.mark.asyncio
-async def test_summarizer_with_chronological_trace(mock_context):
-    """Test that SummarizerNode generates a prompt containing the chronological step trace when available."""
-    step_id = "12345678-1234-5678-1234-567812345678"
+async def test_summarizer_falls_back_to_structured_decisions(mock_context):
+    """Without a step record the action comes from structured_decisions, and
+    extra actions ride along as additional_actions."""
+    mock_context.data_engine.get_step_record.return_value = None
+
     state = DummyState(
-        current_step_id=step_id,
+        structured_decisions=(
+            '[{"action": "swipe", "coordinates": [1, 2, 3, 4]},'
+            ' {"action": "tap", "coordinates": [9, 9]}]'
+        ),
+        last_execution_result={"status": "success"},
+        current_step_id="12345678-1234-5678-1234-567812345678",
     )
 
-    mock_step_data = {
-        "step_id": step_id,
-        "step_number": 1,
-        "relative_time": "5.0s",
-        "action_taken": [{"action": "click", "target_text": "Search"}],
-        "last_execution_result": {"status": "success"},
-        "interleaved_events": [
-            {"type": "thought", "content": "Let's click search."},
-        ],
-    }
+    node = SummarizerNode(mock_context)
+    await node(state)
 
-    mock_context.data_engine.get_agent_friendly_steps.return_value = [mock_step_data]
+    kwargs = mock_context.step_memory.dispatch.call_args.kwargs
+    assert kwargs["action_name"] == "swipe"
+    assert kwargs["action_args"]["coordinates"] == [1, 2, 3, 4]
+    assert kwargs["action_args"]["additional_actions"] == [
+        {"action": "tap", "coordinates": [9, 9]}
+    ]
 
-    # Mock LLM response
-    mock_response = Mock()
-    mock_response.content = "Summary using chronological trace."
-    mock_llm = MagicMock()
 
-    async def mock_astream(*args, **kwargs):
-        yield mock_response
+@pytest.mark.asyncio
+async def test_summarizer_noop_without_step_id(mock_context):
+    state = DummyState(current_step_id=None)
 
-    mock_llm.astream.side_effect = mock_astream
+    node = SummarizerNode(mock_context)
+    result = await node(state)
 
-    with (
-        patch("artemis.agents.summarizer.summarizer.get_llm", return_value=mock_llm),
-        patch("asyncio.create_task") as mock_create_task,
-    ):
-        node = SummarizerNode(mock_context)
-        await node(state)
-
-        # Verify task was registered in context
-        assert len(mock_context.background_tasks) == 1
-        assert mock_context.background_tasks[0] == mock_create_task.return_value
-
-        background_coroutine = mock_create_task.call_args[0][0]
-        await background_coroutine
-
-    assert mock_llm.astream.called
-    invoked_messages = mock_llm.astream.call_args[0][0]
-    human_msg = invoked_messages[1].content
-
-    assert "--- Execution History ---" in human_msg
-    assert "Let's click search." in human_msg
+    assert result == {}
+    mock_context.step_memory.dispatch.assert_not_called()

tests/unit/data_engine/test_engine.py

@@ -380,3 +380,62 @@ def test_update_step_summary_includes_step_number_in_sse(tmp_path):
     assert last_update["step_number"] == 1
 
 
+
+def _foreground_engine(tmp_path):
+    mock_ctx = MagicMock(spec=ArtemisContext)
+    mock_execution_setup = MagicMock()
+    mock_execution_setup.traces_path = str(tmp_path)
+    mock_ctx.execution_setup = mock_execution_setup
+    mock_ctx.device = None
+    engine = DataEngine(mock_ctx)
+    engine.start_session("foreground_app persistence")
+    return engine
+
+
+def test_record_step_persists_explicit_foreground_app(tmp_path):
+    """M5: the historically dropped foreground_app parameter is persisted."""
+    engine = _foreground_engine(tmp_path)
+    engine.record_step(
+        foreground_app="com.android.settings",
+        ui_tree=[{"package": "com.other.app", "bounds": "[0,0][10,10]"}],
+        summary="Explicit parameter wins",
+    )
+    for t in list(engine._pending_threads):
+        t.join()
+
+    step = engine.storage.get_steps(engine.current_session_id)[0]
+    assert step.extra_metadata["foreground_app"] == "com.android.settings"
+
+
+def test_record_step_derives_foreground_app_from_ui_tree(tmp_path):
+    """Without the parameter, the dominant non-overlay package is stamped."""
+    engine = _foreground_engine(tmp_path)
+    engine.record_step(
+        ui_tree=[
+            {"package": "com.android.systemui", "bounds": "[0,0][10,10]"},
+            {"package": "com.google.android.deskclock"},
+            {
+                "packageName": "com.google.android.deskclock",
+                "children": [{"package": "com.google.android.deskclock"}],
+            },
+        ],
+        summary="Derived from UI tree",
+    )
+    for t in list(engine._pending_threads):
+        t.join()
+
+    step = engine.storage.get_steps(engine.current_session_id)[0]
+    assert step.extra_metadata["foreground_app"] == "com.google.android.deskclock"
+
+
+def test_record_step_without_app_data_stamps_nothing(tmp_path):
+    engine = _foreground_engine(tmp_path)
+    engine.record_step(
+        ui_tree=[{"package": "com.android.systemui"}],
+        summary="Only overlay packages",
+    )
+    for t in list(engine._pending_threads):
+        t.join()
+
+    step = engine.storage.get_steps(engine.current_session_id)[0]
+    assert "foreground_app" not in (step.extra_metadata or {})

tests/unit/data_engine/test_history_chunk_storage.py

@@ -0,0 +1,105 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""History-chunk storage (M3 §6.1): append-only versioned rows keyed by step range."""
+
+from uuid import uuid4
+
+import pytest
+
+from artemis.data_engine.models import HistoryChunkRecord
+from artemis.data_engine.storage import StorageManager
+
+SESSION = "11111111-1111-1111-1111-111111111111"
+
+
+@pytest.fixture
+def storage(tmp_path):
+    return StorageManager(tmp_path / "test.db", tmp_path / "traces")
+
+
+def _chunk(start: int, end: int, version: int = 1, **overrides) -> HistoryChunkRecord:
+    fields = {
+        "chunk_id": uuid4(),
+        "session_id": SESSION,
+        "start_step_id": f"sid-{start}",
+        "end_step_id": f"sid-{end}",
+        "start_step_number": start,
+        "end_step_number": end,
+        "source_step_ids": [f"sid-{n}" for n in range(start, end + 1)],
+        "subgoal_hash": "hash-a",
+        "version": version,
+        "status": "pending",
+        "band1": {},
+        "band2": None,
+        "band3": f"- Step {start} (T+00:10): Tapped 'x' at [1, 2] -> executed",
+        "rendered_text": "[Chunk 1 | ...]",
+    }
+    fields.update(overrides)
+    return HistoryChunkRecord(**fields)
+
+
+def test_history_chunk_roundtrip(storage):
+    storage.create_history_chunk(_chunk(1, 4))
+    chunks = storage.get_history_chunks(SESSION)
+    assert len(chunks) == 1
+    rec = chunks[0]
+    assert (rec.start_step_number, rec.end_step_number) == (1, 4)
+    assert rec.source_step_ids == [f"sid-{n}" for n in range(1, 5)]
+    assert rec.subgoal_hash == "hash-a"
+    assert rec.status == "pending"
+    assert "T+00:10" in rec.band3
+
+
+def test_history_chunk_versions_are_append_only_latest_wins(storage):
+    storage.create_history_chunk(_chunk(1, 4, version=1, status="pending"))
+    storage.create_history_chunk(
+        _chunk(
+            1,
+            4,
+            version=2,
+            status="ready",
+            band1={"doing": "d", "verified_facts": ["f1"]},
+            band2="  - Steps 1–4: did things",
+        )
+    )
+
+    latest = storage.get_history_chunks(SESSION)
+    assert len(latest) == 1
+    assert latest[0].version == 2
+    assert latest[0].status == "ready"
+    assert latest[0].band1["verified_facts"] == ["f1"]
+
+    trail = storage.get_history_chunks(SESSION, all_versions=True)
+    assert [r.version for r in trail] == [1, 2]
+
+
+def test_history_chunks_ordered_by_step_range(storage):
+    storage.create_history_chunk(_chunk(9, 12))
+    storage.create_history_chunk(_chunk(1, 4))
+    storage.create_history_chunk(_chunk(5, 8))
+    latest = storage.get_history_chunks(SESSION)
+    assert [(r.start_step_number, r.end_step_number) for r in latest] == [
+        (1, 4),
+        (5, 8),
+        (9, 12),
+    ]
+
+
+def test_history_chunks_are_session_scoped(storage):
+    storage.create_history_chunk(_chunk(1, 4))
+    other = "22222222-2222-2222-2222-222222222222"
+    storage.create_history_chunk(_chunk(1, 4, session_id=other))
+    assert len(storage.get_history_chunks(SESSION)) == 1
+    assert len(storage.get_history_chunks(other)) == 1

tests/unit/data_engine/test_step_summary_versioning.py

@@ -0,0 +1,171 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Tests for the versioned, status-carrying step summary write (M0 §6.1)."""
+
+from unittest.mock import MagicMock
+
+from artemis.context import ArtemisContext
+from artemis.data_engine.engine import DataEngine
+
+
+def _make_engine(tmp_path):
+    mock_ctx = MagicMock(spec=ArtemisContext)
+    mock_execution_setup = MagicMock()
+    mock_execution_setup.traces_path = str(tmp_path)
+    mock_ctx.execution_setup = mock_execution_setup
+    mock_ctx.device = None
+    engine = DataEngine(mock_ctx)
+    engine.start_session("summary versioning test")
+    return engine
+
+
+def _flush(engine):
+    for t in list(engine._pending_threads):
+        t.join()
+
+
+def _get_step(engine, step_id):
+    steps = engine.storage.get_steps(engine.current_session_id)
+    for step in steps:
+        if str(step.step_id) == str(step_id):
+            return step
+    return None
+
+
+def test_ready_write_persists_summary_and_metadata(tmp_path):
+    engine = _make_engine(tmp_path)
+    step_id = engine.record_step(summary="initial")
+    _flush(engine)
+
+    applied = engine.storage.update_step_summary(
+        step_id,
+        "Tapped the login button.",
+        source="visual_transition",
+        model="gemini-2.5-flash-lite",
+        status="ready",
+    )
+    assert applied is True
+
+    step = _get_step(engine, step_id)
+    assert step.summary == "Tapped the login button."
+    assert step.extra_metadata["summary_status"] == "ready"
+    assert step.extra_metadata["summary_source"] == "visual_transition"
+    assert step.extra_metadata["summary_version"] == 1
+    assert step.extra_metadata["summary_model"] == "gemini-2.5-flash-lite"
+
+
+def test_auto_version_increments_per_write(tmp_path):
+    engine = _make_engine(tmp_path)
+    step_id = engine.record_step(summary="initial")
+    _flush(engine)
+
+    engine.storage.update_step_summary(step_id, "v1", status="ready")
+    engine.storage.update_step_summary(step_id, "v2", status="ready")
+
+    step = _get_step(engine, step_id)
+    assert step.summary == "v2"
+    assert step.extra_metadata["summary_version"] == 2
+
+
+def test_stale_explicit_version_is_dropped(tmp_path):
+    engine = _make_engine(tmp_path)
+    step_id = engine.record_step(summary="initial")
+    _flush(engine)
+
+    assert engine.storage.update_step_summary(step_id, "newer", version=5, status="ready")
+    applied = engine.storage.update_step_summary(step_id, "older", version=3, status="ready")
+    assert applied is False
+
+    step = _get_step(engine, step_id)
+    assert step.summary == "newer"
+    assert step.extra_metadata["summary_version"] == 5
+
+
+def test_pending_does_not_downgrade_ready(tmp_path):
+    engine = _make_engine(tmp_path)
+    step_id = engine.record_step(summary="initial")
+    _flush(engine)
+
+    engine.storage.update_step_summary(step_id, "done", status="ready")
+    applied = engine.storage.update_step_summary(step_id, None, status="pending")
+    assert applied is False
+
+    step = _get_step(engine, step_id)
+    assert step.summary == "done"
+    assert step.extra_metadata["summary_status"] == "ready"
+
+
+def test_failed_marks_status_without_touching_summary(tmp_path):
+    engine = _make_engine(tmp_path)
+    step_id = engine.record_step(summary="original text")
+    _flush(engine)
+
+    engine.storage.update_step_summary(step_id, None, status="pending", source="visual_transition")
+    applied = engine.storage.update_step_summary(
+        step_id, None, status="failed", source="visual_transition"
+    )
+    assert applied is True
+
+    step = _get_step(engine, step_id)
+    assert step.summary == "original text"
+    assert step.extra_metadata["summary_status"] == "failed"
+
+    # A failed write must never clobber an already-ready summary.
+    engine.storage.update_step_summary(step_id, "recovered", status="ready")
+    applied = engine.storage.update_step_summary(step_id, None, status="failed")
+    assert applied is False
+    step = _get_step(engine, step_id)
+    assert step.extra_metadata["summary_status"] == "ready"
+
+
+def test_missing_step_returns_false(tmp_path):
+    from uuid import uuid4
+
+    engine = _make_engine(tmp_path)
+    assert engine.storage.update_step_summary(uuid4(), "text", status="ready") is False
+
+
+def test_engine_level_write_carries_metadata_and_publishes_ready_only(tmp_path):
+    engine = _make_engine(tmp_path)
+    published = []
+    engine.subscribe(lambda event_type, data: published.append((event_type, data)))
+
+    step_id = engine.record_step(summary="initial")
+    _flush(engine)
+    published.clear()
+
+    # Non-ready statuses persist metadata but publish no SSE.
+    engine.update_step_summary(step_id, None, status="pending", source="visual_transition")
+    _flush(engine)
+    assert not [e for e in published if e[0] == "step_updated"]
+    step = _get_step(engine, step_id)
+    assert step.extra_metadata["summary_status"] == "pending"
+
+    engine.update_step_summary(
+        step_id,
+        "Ready text.",
+        status="ready",
+        source="visual_transition",
+        model="test-model",
+    )
+    _flush(engine)
+    ready_events = [e for e in published if e[0] == "step_updated"]
+    assert len(ready_events) == 1
+    assert ready_events[0][1]["summary"] == "Ready text."
+
+    step = _get_step(engine, step_id)
+    assert step.summary == "Ready text."
+    assert step.extra_metadata["summary_status"] == "ready"
+    assert step.extra_metadata["summary_model"] == "test-model"

tests/unit/mcp/test_action_executor_semantics.py

@@ -0,0 +1,158 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Tests for record-time target semantics enrichment in McpActionExecutor."""
+
+from unittest.mock import Mock
+
+from artemis.mcp.action_executor import McpActionExecutor
+
+
+def _make_executor(width=1080, height=2400):
+    ctx = Mock()
+    ctx.device.device_width = width
+    ctx.device.device_height = height
+    actuator = Mock()
+    actuator.controller = Mock()
+    return McpActionExecutor(ctx, actuator=actuator)
+
+
+def _state_with_elements(elements):
+    state = Mock()
+    state.indexed_elements = elements
+    return state
+
+
+def test_click_target_semantics_from_pre_action_frame():
+    executor = _make_executor()
+    # Normalized (500, 600) on a 1080x2400 device -> pixel (540, 1440).
+    state = _state_with_elements(
+        [
+            {
+                "text": "Confirm",
+                "bounds": [500, 1400, 600, 1480],
+                "class": "android.widget.Button",
+                "resource_id": "btn_confirm",
+                "is_ocr": False,
+            }
+        ]
+    )
+    semantics = executor._target_semantics("click", {"target": [500, 600]}, state)
+    assert semantics == {
+        "target_label_source": "hit_test",
+        "target_text": "Confirm",
+        "target_class": "android.widget.Button",
+        "target_resource_id": "btn_confirm",
+    }
+
+
+def test_semantics_degrade_to_none_without_element_data():
+    executor = _make_executor()
+
+    # Empty perception data
+    semantics = executor._target_semantics(
+        "click", {"target": [500, 600]}, _state_with_elements([])
+    )
+    assert semantics == {"target_label_source": "none"}
+
+    # No state at all
+    semantics = executor._target_semantics("click", {"target": [500, 600]}, None)
+    assert semantics == {"target_label_source": "none"}
+
+
+def test_non_targeted_actions_and_missing_targets_skip_enrichment():
+    executor = _make_executor()
+    state = _state_with_elements([{"text": "X", "bounds": [0, 0, 10, 10]}])
+
+    assert executor._target_semantics("press_key", {"key": "BACK"}, state) is None
+    # input_text without a coordinate target (focused field typing)
+    assert executor._target_semantics("input_text", {"text": "hi", "target": None}, state) is None
+    # Multi-point actions without resolvable points skip enrichment too.
+    assert executor._target_semantics("click_sequence", {"sequence": []}, state) is None
+    assert executor._target_semantics("swipe", {"start": None, "end": None}, state) is None
+
+
+def test_click_sequence_hit_tests_every_point_first_point_is_main_label():
+    """M5: multi-point actions get per-point best-effort semantics; the first
+    point's fields are hoisted as the action's main label."""
+    executor = _make_executor()
+    state = _state_with_elements(
+        [
+            {
+                "text": "Digit 1",
+                "bounds": [500, 1400, 600, 1480],
+                "class": "android.widget.Button",
+                "is_ocr": False,
+            },
+            {
+                "text": "Digit 2",
+                "bounds": [700, 1400, 800, 1480],
+                "class": "android.widget.Button",
+                "is_ocr": False,
+            },
+        ]
+    )
+    # Normalized points -> pixels: (500,600)->(540,1440) hits Digit 1;
+    # (690,600)->(745,1440) hits Digit 2; (10,10)->(10,24) hits nothing.
+    semantics = executor._target_semantics(
+        "click_sequence",
+        {"sequence": [[500, 600], [690, 600], [10, 10]], "delay_ms": 50},
+        state,
+    )
+    assert semantics["target_text"] == "Digit 1"
+    assert semantics["target_label_source"] == "hit_test"
+    per_point = semantics["points_semantics"]
+    assert len(per_point) == 3
+    assert per_point[0]["target_text"] == "Digit 1"
+    assert per_point[1]["target_text"] == "Digit 2"
+    assert per_point[2] == {"target_label_source": "none"}
+
+
+def test_swipe_hit_tests_start_and_end_start_is_main_label():
+    executor = _make_executor()
+    state = _state_with_elements(
+        [
+            {
+                "text": "Brightness slider",
+                "bounds": [100, 1400, 900, 1480],
+                "class": "android.widget.SeekBar",
+                "is_ocr": False,
+            }
+        ]
+    )
+    semantics = executor._target_semantics(
+        "swipe", {"start": [200, 600], "end": [800, 600], "duration_ms": 400}, state
+    )
+    assert semantics["target_text"] == "Brightness slider"
+    assert semantics["start_semantics"]["target_text"] == "Brightness slider"
+    assert semantics["end_semantics"]["target_text"] == "Brightness slider"
+    assert semantics["target_label_source"] == "hit_test"
+
+
+def test_long_press_uses_ocr_fallback():
+    executor = _make_executor()
+    state = _state_with_elements(
+        [
+            {
+                "text": "OCR Word",
+                "bounds": [520, 1420, 560, 1460],
+                "class": None,
+                "resource_id": None,
+                "is_ocr": True,
+            }
+        ]
+    )
+    semantics = executor._target_semantics("long_press", {"target": [500, 600]}, state)
+    assert semantics["target_label_source"] == "ocr"
+    assert semantics["target_text"] == "OCR Word"

tests/unit/memory/test_context_policy.py

@@ -0,0 +1,304 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""ContextPolicy table (M4): byte-exact parity with the eight historical call
+sites, config overrides, and the chunk hand-off to the compiled view."""
+
+from types import SimpleNamespace
+from unittest.mock import patch
+
+import pytest
+
+from artemis.memory.context_policy import (
+    CONTEXT_POLICIES,
+    DIGEST_LEDGER_MARKER_TEMPLATE,
+    build_history_for,
+    load_chunk_blocks,
+    resolve_policy,
+)
+from artemis.utils.task_tree import build_plan_and_history
+
+HASH_A = "hash-a"
+HASH_B = "hash-b"
+
+TASK_PLAN = f"- [x] First milestone <!-- {HASH_A} -->\n- [/] Second milestone <!-- {HASH_B} -->"
+
+
+def _step(number: int, subgoal_hash: str, summary: str | None = None) -> dict:
+    return {
+        "step_id": f"id-{number}",
+        "step_number": number,
+        "relative_time": f"{number * 10}.0s",
+        "summary": summary or f"Step {number} summary text.",
+        "action_taken": {"action": "click", "target_text": f"button {number}"},
+        "operator_raw_thinking": f"thinking {number}",
+        "last_execution_result": {"status": "success"},
+        "extra_metadata": {"subgoal_hash": subgoal_hash},
+    }
+
+
+STEPS = [
+    _step(1, HASH_A),
+    _step(2, HASH_A),
+    _step(3, HASH_A),
+    _step(4, HASH_B),
+    _step(5, HASH_B),
+    _step(6, HASH_B),
+    _step(7, HASH_B),
+]
+
+
+# ---------------------------------------------------------------------------
+# Golden parity: the policy table must reproduce every historical hard-coded
+# call byte for byte (kwargs recorded verbatim from the 2026-09-01 call sites).
+# ---------------------------------------------------------------------------
+
+
+GOLDEN_CALLS = {
+    "operator": lambda: build_plan_and_history(
+        TASK_PLAN,
+        STEPS,
+        HASH_B,
+        last_n_detailed=1,
+        strict_milestone_pruning=True,
+        recent_window_size=3,
+    ),
+    "operator_cold_start": lambda: build_plan_and_history(
+        TASK_PLAN,
+        STEPS,
+        HASH_B,
+        min_summaries=len(STEPS),
+        last_n_detailed=1,
+    ),
+    "failure_analyzer": lambda: build_plan_and_history(
+        TASK_PLAN,
+        STEPS,
+        HASH_B,
+        last_n_detailed=1,
+        strict_milestone_pruning=True,
+        recent_window_size=3,
+        for_failure_analyzer=True,
+    ),
+    "planner": lambda: build_plan_and_history(
+        "",
+        STEPS,
+        "default",
+        last_n_detailed=2,
+        strict_milestone_pruning=True,
+        recent_window_size=5,
+    ),
+    "outputter": lambda: build_plan_and_history(
+        TASK_PLAN,
+        STEPS,
+        HASH_B,
+        last_n_detailed=0,
+        min_summaries=len(STEPS),
+        strict_milestone_pruning=False,
+    ),
+    "history_analyzer": lambda: build_plan_and_history(
+        TASK_PLAN,
+        STEPS,
+        HASH_B,
+        last_n_detailed=1,
+        min_summaries=len(STEPS),
+    ),
+    "diagnoser": lambda: build_plan_and_history(
+        TASK_PLAN,
+        STEPS,
+        HASH_B,
+        keep_subgoal_hashes={HASH_B},
+    ),
+    "committee": lambda: build_plan_and_history(
+        TASK_PLAN,
+        STEPS,
+        HASH_B,
+        keep_subgoal_hashes={HASH_B},
+        last_n_detailed=3,
+    ),
+}
+
+POLICY_CALLS = {
+    "operator": lambda: build_history_for(
+        "operator", TASK_PLAN, STEPS, HASH_B, last_n_detailed=1
+    ),
+    "operator_cold_start": lambda: build_history_for(
+        "operator_cold_start", TASK_PLAN, STEPS, HASH_B
+    ),
+    "failure_analyzer": lambda: build_history_for(
+        "failure_analyzer", TASK_PLAN, STEPS, HASH_B
+    ),
+    "planner": lambda: build_history_for("planner", "", STEPS, "default"),
+    "outputter": lambda: build_history_for("outputter", TASK_PLAN, STEPS, HASH_B),
+    "history_analyzer": lambda: build_history_for(
+        "history_analyzer", TASK_PLAN, STEPS, HASH_B
+    ),
+    "diagnoser": lambda: build_history_for(
+        "diagnoser", TASK_PLAN, STEPS, HASH_B, keep_subgoal_hashes={HASH_B}
+    ),
+    "committee": lambda: build_history_for(
+        "committee", TASK_PLAN, STEPS, HASH_B, keep_subgoal_hashes={HASH_B}
+    ),
+}
+
+
+@pytest.mark.parametrize("agent", sorted(GOLDEN_CALLS))
+def test_policy_table_reproduces_call_site_bytes(agent):
+    assert POLICY_CALLS[agent]() == GOLDEN_CALLS[agent]()
+
+
+def test_every_declared_policy_is_golden_tested():
+    assert set(GOLDEN_CALLS) == set(CONTEXT_POLICIES)
+
+
+# ---------------------------------------------------------------------------
+# Config overrides
+# ---------------------------------------------------------------------------
+
+
+def _config_with_policies(policies):
+    return SimpleNamespace(memory=SimpleNamespace(policies=policies))
+
+
+def test_resolve_policy_applies_config_override():
+    cfg = _config_with_policies({"planner": {"last_n_detailed": 3, "bogus_field": 1}})
+    with patch("artemis.config.load_agent_config", return_value=cfg):
+        policy = resolve_policy("planner")
+    assert policy.last_n_detailed == 3
+    assert policy.strict_milestone_pruning is True  # untouched fields survive
+
+
+def test_resolve_policy_unknown_agent_raises():
+    with pytest.raises(KeyError):
+        resolve_policy("no_such_agent")
+
+
+# ---------------------------------------------------------------------------
+# Chunk hand-off (§10 decision 4)
+# ---------------------------------------------------------------------------
+
+
+RENDERED_CHUNK = (
+    "[Chunk 1 | Steps 1–3 | T+00:10 → T+00:30]\n"
+    "\n"
+    "① Synopsis & effects\n"
+    "  What this segment was doing: warming up.\n"
+    "\n"
+    "② Compressed step summary\n"
+    "  - Steps 1–3: opened the app and logged in\n"
+    "\n"
+    "③ Step action ledger\n"
+    "- Step 1 (T+00:10): click['button 1'] -> executed\n"
+    "- Step 2 (T+00:20): click['button 2'] -> executed\n"
+    "- Step 3 (T+00:30): click['button 3'] -> executed"
+)
+
+
+class _FakeEngine:
+    def __init__(self, rows):
+        self._rows = rows
+
+    def get_history_chunks(self, **kwargs):
+        return self._rows
+
+
+def _chunk_row():
+    return SimpleNamespace(
+        start_step_number=1,
+        end_step_number=3,
+        rendered_text=RENDERED_CHUNK,
+        status="ready",
+    )
+
+
+def _transcript_cfg(enabled: bool):
+    return SimpleNamespace(
+        memory=SimpleNamespace(
+            transcript=SimpleNamespace(enabled=enabled), policies={}
+        )
+    )
+
+
+def test_flag_off_output_is_byte_identical_even_with_engine():
+    engine = _FakeEngine([_chunk_row()])
+    with patch("artemis.config.load_agent_config", return_value=_transcript_cfg(False)):
+        with_engine = build_history_for(
+            "outputter", TASK_PLAN, STEPS, HASH_B, engine=engine
+        )
+        without_engine = build_history_for("outputter", TASK_PLAN, STEPS, HASH_B)
+    assert with_engine == without_engine
+    assert "[Chunk 1" not in with_engine
+
+
+def test_flag_on_full_view_renders_chunk_block_and_drops_covered_steps():
+    engine = _FakeEngine([_chunk_row()])
+    with patch("artemis.config.load_agent_config", return_value=_transcript_cfg(True)):
+        out = build_history_for("outputter", TASK_PLAN, STEPS, HASH_B, engine=engine)
+    # Full view: the whole three-band block including the ③ ledger.
+    assert "[Chunk 1 | Steps 1–3" in out
+    assert "③ Step action ledger" in out
+    assert "- Step 2 (T+00:20)" in out
+    # Covered per-step summary lines are replaced by the chunk block.
+    assert "*Step 2 (Start: 20.0s)" not in out
+    # Steps outside the chunk range still render.
+    assert "Step 5" in out
+
+
+def test_flag_on_digest_view_replaces_ledger_with_recall_marker():
+    engine = _FakeEngine([_chunk_row()])
+    cfg = _transcript_cfg(True)
+    with patch("artemis.config.load_agent_config", return_value=cfg):
+        out = build_history_for(
+            "diagnoser",
+            TASK_PLAN,
+            STEPS,
+            HASH_B,
+            keep_subgoal_hashes={HASH_A, HASH_B},
+            engine=engine,
+        )
+    assert "[Chunk 1 | Steps 1–3" in out
+    assert "② Compressed step summary" in out
+    marker = DIGEST_LEDGER_MARKER_TEMPLATE.format(start=1, end=3)
+    assert marker in out
+    assert "- Step 2 (T+00:20)" not in out  # digest carries no ③ rows
+
+
+def test_chunk_view_none_never_loads_chunks():
+    engine = _FakeEngine([_chunk_row()])
+    with patch("artemis.config.load_agent_config", return_value=_transcript_cfg(True)):
+        assert load_chunk_blocks(engine, None) is None
+        # The cold-start policy declares chunk_view=None: even with an engine
+        # and the flag on, its output carries no chunk blocks (the restored
+        # F region must stay a pure step rendering).
+        out = build_history_for(
+            "operator_cold_start", TASK_PLAN, STEPS, HASH_B, engine=engine
+        )
+    assert CONTEXT_POLICIES["operator_cold_start"].chunk_view is None
+    assert "[Chunk 1" not in out
+
+
+def test_most_recent_step_is_never_suppressed_by_a_chunk():
+    engine = _FakeEngine(
+        [
+            SimpleNamespace(
+                start_step_number=1,
+                end_step_number=7,  # covers every step incl. the newest
+                rendered_text=RENDERED_CHUNK,
+                status="ready",
+            )
+        ]
+    )
+    with patch("artemis.config.load_agent_config", return_value=_transcript_cfg(True)):
+        out = build_history_for("outputter", TASK_PLAN, STEPS, HASH_B, engine=engine)
+    # The most-recent step must keep its own (detailed) rendering.
+    assert "Step 7" in out

tests/unit/memory/test_step_memory_service.py

@@ -0,0 +1,187 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Unit tests for the shared StepMemoryService scheduling runtime (M1)."""
+
+import asyncio
+from unittest.mock import Mock
+
+import pytest
+
+from artemis.memory.step_memory import StepMemoryService
+
+
+class _ScriptedService(StepMemoryService):
+    """StepMemoryService with a scriptable attempt function and status log."""
+
+    def __init__(self, attempt_fn, **kwargs):
+        super().__init__(Mock(), **kwargs)
+        self._attempt_fn = attempt_fn
+        self.status_events: list[tuple] = []
+
+    async def _attempt(self, key):
+        return await self._attempt_fn(self, key)
+
+    def _on_status(self, key, status):
+        self.status_events.append((key, status))
+
+
+@pytest.mark.asyncio
+async def test_zero_blocking_submit_and_flush():
+    """submit() returns immediately; flush() drains the background job."""
+    release = asyncio.Event()
+
+    async def attempt(svc, key):
+        await release.wait()
+        svc._summaries[key] = f"summary-for-{key}"
+        return True
+
+    service = _ScriptedService(attempt)
+    service.submit("step-1", {"step_number": 1})
+
+    # Zero-blocking: the job cannot have completed yet (its gate is closed).
+    assert service.is_pending("step-1")
+    assert not service.has_summary("step-1")
+
+    release.set()
+    await service.flush()
+
+    assert service.get_summary("step-1") == "summary-for-step-1"
+    assert not service.is_pending("step-1")
+
+
+@pytest.mark.asyncio
+async def test_bounded_retry_enters_failed_state():
+    """A permanently failing job stops after 1 + retry_limit attempts."""
+    attempts = 0
+
+    async def attempt(svc, key):
+        nonlocal attempts
+        attempts += 1
+        return False
+
+    service = _ScriptedService(attempt, retry_limit=2)
+    service._retry_delays = (0.0,)
+    service.submit("step-x", {"step_number": 7})
+    await service.flush()
+
+    assert attempts == 3
+    assert service.has_failed("step-x")
+    assert not service.has_summary("step-x")
+    # The failed job still reads as pending (lossless semantics downstream).
+    assert service.is_pending("step-x")
+    assert service.status_events == [("step-x", "pending"), ("step-x", "failed")]
+
+
+@pytest.mark.asyncio
+async def test_step_id_keying_with_tool_call_alias():
+    """Jobs keyed by step id remain queryable through tool_call_id aliases."""
+
+    async def attempt(svc, key):
+        svc._summaries[key] = "canonical summary"
+        return True
+
+    service = _ScriptedService(attempt)
+    service.submit("uuid-123", {"step_number": 4}, aliases=("tc-abc",))
+    await service.flush()
+
+    assert service.resolve_key("tc-abc") == "uuid-123"
+    assert service.has_job("tc-abc")
+    assert service.has_summary("tc-abc")
+    assert service.get_summary("tc-abc") == "canonical summary"
+    assert service.get_summary("uuid-123") == "canonical summary"
+    assert service.get_job_payload("tc-abc") == {"step_number": 4}
+    assert not service.has_failed("tc-abc")
+
+
+@pytest.mark.asyncio
+async def test_max_concurrency_bounds_parallel_attempts():
+    """With max_concurrency=1, two jobs never run their attempts concurrently."""
+    active = 0
+    max_active = 0
+
+    async def attempt(svc, key):
+        nonlocal active, max_active
+        active += 1
+        max_active = max(max_active, active)
+        await asyncio.sleep(0.01)
+        active -= 1
+        svc._summaries[key] = "done"
+        return True
+
+    service = _ScriptedService(attempt, max_concurrency=1)
+    service.submit("a", {"step_number": 1})
+    service.submit("b", {"step_number": 2})
+    await service.flush()
+
+    assert service.has_summary("a") and service.has_summary("b")
+    assert max_active == 1
+
+
+@pytest.mark.asyncio
+async def test_flush_timeout_cancels_stragglers():
+    """flush() cancels jobs that outlive the timeout without failing them."""
+
+    async def attempt(svc, key):
+        await asyncio.sleep(30)
+        return True
+
+    service = _ScriptedService(attempt)
+    service.submit("slow", {"step_number": 1})
+    await service.flush(timeout_seconds=0.05)
+
+    assert service.is_pending("slow")
+    assert not service.has_failed("slow")
+    assert service._pending_tasks["slow"].done()
+
+
+@pytest.mark.asyncio
+async def test_flush_uses_configured_default_timeout():
+    """flush() with no argument honors the constructor's flush_timeout_s."""
+
+    async def attempt(svc, key):
+        await asyncio.sleep(30)
+        return True
+
+    service = _ScriptedService(attempt, flush_timeout_s=0.05)
+    service.submit("slow", {"step_number": 1})
+    await asyncio.wait_for(service.flush(), timeout=5)
+
+    assert service._pending_tasks["slow"].done()
+
+
+@pytest.mark.asyncio
+async def test_resubmit_clears_failed_state():
+    """Re-submitting a failed job gives it a fresh retry budget."""
+    behaviors = {"fail": True}
+
+    async def attempt(svc, key):
+        if behaviors["fail"]:
+            return False
+        svc._summaries[key] = "second time lucky"
+        return True
+
+    service = _ScriptedService(attempt, retry_limit=0)
+    service._retry_delays = (0.0,)
+    service.submit("job", {"step_number": 1})
+    await service.flush()
+    assert service.has_failed("job")
+
+    behaviors["fail"] = False
+    service._retry_counts.clear()
+    service.submit("job", {"step_number": 1})
+    await service.flush()
+
+    assert not service.has_failed("job")
+    assert service.get_summary("job") == "second time lucky"

tests/unit/memory/test_transcript_ledger.py

@@ -0,0 +1,298 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Unit tests for the Pro session transcript ledger (history redesign §3.2, M2).
+
+Covers the four-region discipline: S-region byte stability across turns, the
+append-only active region, depth-1 stripping of old UI lists and plan
+recitations, the depth-K screenshot scrub in the Pro message shape (grace /
+placeholder / freeze — the M1 race regressions re-run against HumanMessage
+observations keyed by step id), ``T+mm:ss`` session-offset timestamps, the
+tool-call/response pairing invariant, and the cold-start restored-history
+block.
+"""
+
+import json
+import re
+
+import pytest
+from langchain_core.messages import AIMessage, HumanMessage, SystemMessage, ToolMessage
+
+from artemis.memory.step_memory import StepMemoryService
+from artemis.memory.transcript import (
+    EXECUTION_RESULT_MARKER,
+    PLAN_RECITATION_MARKER,
+    PRO_UI_LIST_MARKER,
+    RESTORED_HISTORY_HEADER,
+    TranscriptLedger,
+    format_session_offset,
+)
+
+OFFSET_RE = re.compile(r"T\+\d{2,}:\d{2}")
+
+
+def _service() -> StepMemoryService:
+    return StepMemoryService(ctx=None)
+
+
+def _observation(i: int) -> HumanMessage:
+    return HumanMessage(
+        content=[
+            {"type": "text", "text": f"# CURRENT OBSERVATION [T+00:0{i % 10}]"},
+            {"type": "text", "text": f"{PLAN_RECITATION_MARKER}\n- [/] milestone {i}"},
+            {"type": "text", "text": "--- Current Screenshot ---"},
+            {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,IMG_{i}"}},
+            {"type": "text", "text": f"{PRO_UI_LIST_MARKER}\n[1] button {i}"},
+        ]
+    )
+
+
+def _turn(i: int, tool_call_count: int = 1) -> list:
+    messages: list = [_observation(i)]
+    if tool_call_count:
+        tool_calls = [
+            {"name": "click", "args": {"target": n + 1}, "id": f"tc{i}-{n}", "type": "tool_call"}
+            for n in range(tool_call_count)
+        ]
+        messages.append(AIMessage(content=f"thinking {i}", tool_calls=tool_calls))
+        for tc in tool_calls:
+            messages.append(ToolMessage(tool_call_id=tc["id"], content="Action Recorded"))
+    return messages
+
+
+def _play_turn(ledger: TranscriptLedger, i: int, *, prev_key=None, prev_result=None, **turn_kwargs):
+    """Mimic the operator's real ordering: commit previous, render, stage."""
+    ledger.commit_staged(step_key=prev_key, validator_result=prev_result)
+    turn = _turn(i, **turn_kwargs)
+    rendered = ledger.render([turn[0]])
+    ledger.stage_turn(turn)
+    return rendered
+
+
+def _fingerprint(msg) -> str:
+    return json.dumps(msg.content, sort_keys=True, default=str)
+
+
+def _has_image(msg) -> bool:
+    return any(
+        isinstance(b, dict) and b.get("type") in ("image_url", "image") for b in msg.content
+    )
+
+
+def test_format_session_offset():
+    assert format_session_offset(0) == "T+00:00"
+    assert format_session_offset(83) == "T+01:23"
+    # Minutes never wrap into hours: byte-stable monotonic labels.
+    assert format_session_offset(3725) == "T+62:05"
+    assert format_session_offset(-5) == "T+00:00"
+
+
+def test_static_prefix_set_once_and_byte_stable():
+    ledger = TranscriptLedger(step_memory=_service())
+    system = SystemMessage(content="STATIC SYSTEM PROMPT")
+    ledger.set_static_prefix([system])
+
+    with pytest.raises(RuntimeError):
+        ledger.set_static_prefix([SystemMessage(content="other")])
+
+    fingerprints = set()
+    for i in range(1, 6):
+        rendered = _play_turn(ledger, i, prev_key=f"step-{i - 1}" if i > 1 else None)
+        assert rendered[0] is system
+        fingerprints.add(rendered[0].content)
+    assert fingerprints == {"STATIC SYSTEM PROMPT"}
+
+
+def test_active_region_is_append_only():
+    ledger = TranscriptLedger(step_memory=_service())
+    seen_ids: list[int] = []
+    for i in range(1, 5):
+        _play_turn(ledger, i, prev_key=f"step-{i - 1}" if i > 1 else None)
+        active = ledger.active_messages
+        # The previously observed prefix is unchanged in identity and order.
+        assert [id(m) for m in active[: len(seen_ids)]] == seen_ids
+        seen_ids = [id(m) for m in active]
+    assert ledger.turn_count == 3  # 3 committed, 1 still staged
+
+
+def test_validator_result_message_carries_session_offset():
+    ledger = TranscriptLedger(step_memory=_service())
+    _play_turn(ledger, 1)
+    _play_turn(
+        ledger,
+        2,
+        prev_key="step-1",
+        prev_result={"status": "success", "execution": [{"action": "tap", "attempts": ["ok"]}]},
+    )
+
+    result_messages = [
+        m
+        for m in ledger.active_messages
+        if isinstance(m, HumanMessage)
+        and str(m.content).find(EXECUTION_RESULT_MARKER) >= 0
+    ]
+    assert len(result_messages) == 1
+    text = result_messages[0].content[0]["text"]
+    assert OFFSET_RE.search(text), text
+    assert "Status: success" in text
+    assert "ago" not in text
+
+
+def test_validator_result_skipped_for_actionless_turn():
+    ledger = TranscriptLedger(step_memory=_service())
+    _play_turn(ledger, 1)
+    _play_turn(ledger, 2, prev_key="step-1", prev_result=None)
+    assert not any(
+        EXECUTION_RESULT_MARKER in str(m.content) for m in ledger.active_messages
+    )
+
+
+def test_old_turn_recitation_and_ui_list_stripped_at_depth_1():
+    ledger = TranscriptLedger(step_memory=_service())
+    _play_turn(ledger, 1)
+    rendered = _play_turn(ledger, 2, prev_key="step-1")
+
+    committed_obs = ledger.active_messages[0]
+    committed_text = str(committed_obs.content)
+    # Depth-1 scrub: the previous turn's plan recitation and UI list are gone.
+    assert PLAN_RECITATION_MARKER not in committed_text
+    assert PRO_UI_LIST_MARKER not in committed_text
+    # Its screenshot is still inside the K-depth active window.
+    assert _has_image(committed_obs)
+    # The live tail keeps both.
+    tail_text = str(rendered[-1].content)
+    assert PLAN_RECITATION_MARKER in tail_text
+    assert PRO_UI_LIST_MARKER in tail_text
+
+
+def test_depth_k_image_resolved_to_ready_visual_summary():
+    service = _service()
+    service._step_inputs["step-1"] = {"step_number": 1}
+    service._summaries["step-1"] = "Objective visual transition 1."
+
+    ledger = TranscriptLedger(step_memory=service, image_scrub_depth=3)
+    for i in range(1, 5):
+        _play_turn(ledger, i, prev_key=f"step-{i - 1}" if i > 1 else None)
+
+    first_obs = ledger.active_messages[0]
+    assert not _has_image(first_obs)
+    assert {
+        "type": "text",
+        "text": "--- Historical Visual Transition ---\nObjective visual transition 1.",
+    } in first_obs.content
+
+
+def test_pending_grace_then_placeholder_never_backfilled():
+    service = _service()
+    service._step_inputs["step-1"] = {"step_number": 1}  # pending job, no summary
+
+    ledger = TranscriptLedger(
+        step_memory=service, image_scrub_depth=2, pending_grace_steps=1
+    )
+    # Enough turns to push turn 1 past K + grace.
+    for i in range(1, 7):
+        _play_turn(ledger, i, prev_key=f"step-{i - 1}" if i > 1 else None)
+
+    first_obs = ledger.active_messages[0]
+    assert not _has_image(first_obs)
+    assert "[visual summary pending; evidence at DataEngine step 1]" in str(
+        first_obs.content
+    )
+    frozen = _fingerprint(first_obs)
+
+    # A late summary must never mutate the frozen message.
+    service._summaries["step-1"] = "Too late."
+    _play_turn(ledger, 7, prev_key="step-6")
+    assert _fingerprint(ledger.active_messages[0]) == frozen
+
+
+def test_failed_summary_becomes_unavailable_placeholder():
+    service = _service()
+    service._step_inputs["step-1"] = {"step_number": 1}
+    service._failed.add("step-1")
+
+    ledger = TranscriptLedger(
+        step_memory=service, image_scrub_depth=2, pending_grace_steps=5
+    )
+    for i in range(1, 4):
+        _play_turn(ledger, i, prev_key=f"step-{i - 1}" if i > 1 else None)
+
+    first_obs = ledger.active_messages[0]
+    assert not _has_image(first_obs)
+    assert "[visual summary unavailable; evidence at DataEngine step 1]" in str(
+        first_obs.content
+    )
+
+
+def test_tool_call_response_pairs_are_never_split():
+    ledger = TranscriptLedger(step_memory=_service(), image_scrub_depth=2)
+    for i in range(1, 7):
+        _play_turn(
+            ledger,
+            i,
+            prev_key=f"step-{i - 1}" if i > 1 else None,
+            prev_result={"status": "success"} if i > 1 else None,
+            tool_call_count=2,
+        )
+
+    active = list(ledger.active_messages)
+    for idx, msg in enumerate(active):
+        if isinstance(msg, AIMessage) and msg.tool_calls:
+            following = active[idx + 1 : idx + 1 + len(msg.tool_calls)]
+            assert [getattr(m, "tool_call_id", None) for m in following] == [
+                tc["id"] for tc in msg.tool_calls
+            ], f"tool-call pairing split at active index {idx}"
+
+
+def test_restored_history_only_seeds_an_empty_ledger():
+    ledger = TranscriptLedger(step_memory=_service())
+    ledger.set_restored_history(f"{RESTORED_HISTORY_HEADER} steps 1-9 ...")
+    assert ledger.has_restored_history
+
+    with pytest.raises(RuntimeError):
+        ledger.set_restored_history("again")
+
+    ledger.set_static_prefix([SystemMessage(content="S")])
+    rendered = _play_turn(ledger, 1)
+    # Order: S region, then the frozen restored block, then the live tail.
+    assert rendered[0].content == "S"
+    assert RESTORED_HISTORY_HEADER in str(rendered[1].content)
+    assert "# CURRENT OBSERVATION" in str(rendered[-1].content)
+
+    ledger2 = TranscriptLedger(step_memory=_service())
+    _play_turn(ledger2, 1)
+    ledger2.commit_staged(step_key="step-1")
+    with pytest.raises(RuntimeError):
+        ledger2.set_restored_history("late")
+
+
+def test_stage_twice_commits_the_forgotten_turn():
+    ledger = TranscriptLedger(step_memory=_service())
+    ledger.stage_turn(_turn(1))
+    ledger.stage_turn(_turn(2))
+    assert ledger.turn_count == 1
+    assert ledger.has_staged_turn
+
+
+def test_no_ago_wording_in_ledger_output():
+    ledger = TranscriptLedger(step_memory=_service())
+    for i in range(1, 4):
+        _play_turn(
+            ledger,
+            i,
+            prev_key=f"step-{i - 1}" if i > 1 else None,
+            prev_result={"status": "success"} if i > 1 else None,
+        )
+    blob = " ".join(str(m.content) for m in ledger.active_messages)
+    assert " ago" not in blob

tests/unit/test_memory_config.py

@@ -0,0 +1,134 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Unit tests for the agent.memory configuration block (M1/M2)."""
+
+from artemis.config import AgentGlobalConfig
+
+
+def test_memory_defaults():
+    cfg = AgentGlobalConfig.model_validate({})
+    # 2 per the 2026-09-01 on-device baseline (serial queue at 1 pushed
+    # summary-ready P90 past the scrub-edge grace window).
+    assert cfg.memory.runtime.max_concurrency == 2
+    assert cfg.memory.runtime.retry_limit == 3
+    assert cfg.memory.runtime.flush_timeout_s == 30.0
+    # M5 (2026-09-01): the transcript path ships ON by default; explicit
+    # false is the byte-for-byte rollback switch (see the flip test below).
+    assert cfg.memory.transcript.enabled is True
+    assert cfg.memory.transcript.image_scrub_depth == 3
+    assert cfg.memory.transcript.pending_grace_steps == 3
+    assert cfg.memory.transcript.xml_scrub_depth == 1
+
+
+def test_memory_transcript_flag_explicit_disable_is_rollback_switch():
+    cfg = AgentGlobalConfig.model_validate(
+        {"memory": {"transcript": {"enabled": False}}}
+    )
+    assert cfg.memory.transcript.enabled is False
+    # Sibling scrub defaults are unaffected by flipping the flag.
+    assert cfg.memory.transcript.image_scrub_depth == 3
+
+
+def test_memory_explicit_values():
+    cfg = AgentGlobalConfig.model_validate(
+        {
+            "memory": {
+                "runtime": {"max_concurrency": 4, "retry_limit": 5, "flush_timeout_s": 10},
+                "transcript": {
+                    "image_scrub_depth": 5,
+                    "pending_grace_steps": 1,
+                    "xml_scrub_depth": 2,
+                },
+            }
+        }
+    )
+    assert cfg.memory.runtime.max_concurrency == 4
+    assert cfg.memory.runtime.retry_limit == 5
+    assert cfg.memory.runtime.flush_timeout_s == 10.0
+    assert cfg.memory.transcript.image_scrub_depth == 5
+    assert cfg.memory.transcript.pending_grace_steps == 1
+    assert cfg.memory.transcript.xml_scrub_depth == 2
+
+
+def test_legacy_step_summarizer_retry_limit_seeds_memory_runtime():
+    """An explicit legacy flash.step_summarizer.retry_limit keeps working."""
+    cfg = AgentGlobalConfig.model_validate(
+        {"flash": {"step_summarizer": {"retry_limit": 7}}}
+    )
+    assert cfg.flash.step_summarizer.retry_limit == 7
+    assert cfg.memory.runtime.retry_limit == 7
+
+
+def test_explicit_memory_runtime_wins_over_legacy_alias():
+    cfg = AgentGlobalConfig.model_validate(
+        {
+            "flash": {"step_summarizer": {"retry_limit": 7}},
+            "memory": {"runtime": {"retry_limit": 2}},
+        }
+    )
+    assert cfg.memory.runtime.retry_limit == 2
+    # The legacy key itself is untouched for back-compat readers.
+    assert cfg.flash.step_summarizer.retry_limit == 7
+
+
+def test_legacy_default_does_not_override_memory_default():
+    """A step_summarizer block without retry_limit must not seed anything."""
+    cfg = AgentGlobalConfig.model_validate(
+        {"flash": {"step_summarizer": {"enabled": True}}}
+    )
+    assert cfg.memory.runtime.retry_limit == 3
+
+
+def test_recall_and_similarity_defaults():
+    """M4 defaults: recall block on, bounded; similarity hint on. The distance
+    threshold was calibrated to 5 in M5 (2026-09-01 on-device dHash data:
+    same screen <=4, different screens >=7, valley at 5-6)."""
+    cfg = AgentGlobalConfig()
+    assert cfg.memory.recall.enabled is True
+    assert cfg.memory.recall.max_results == 5
+    assert cfg.memory.recall.max_text_tokens == 2000
+    assert cfg.memory.recall.max_image_steps == 1
+    assert cfg.memory.transcript.similarity_hint is True
+    assert cfg.memory.transcript.similarity_max_distance == 5
+    assert cfg.memory.policies == {}
+
+
+def test_chunking_era_cap_follows_max_chunks_by_default():
+    """M5: max_eras is an independent key; None (default) follows max_chunks."""
+    cfg = AgentGlobalConfig()
+    assert cfg.memory.chunking.max_chunks == 8
+    assert cfg.memory.chunking.max_eras is None
+
+    overridden = AgentGlobalConfig.model_validate(
+        {"memory": {"chunking": {"max_chunks": 12, "max_eras": 4}}}
+    )
+    assert overridden.memory.chunking.max_chunks == 12
+    assert overridden.memory.chunking.max_eras == 4
+
+
+def test_recall_and_policies_overridable():
+    cfg = AgentGlobalConfig.model_validate(
+        {
+            "memory": {
+                "recall": {"enabled": False, "max_results": 3},
+                "transcript": {"similarity_hint": False},
+                "policies": {"planner": {"last_n_detailed": 3}},
+            }
+        }
+    )
+    assert cfg.memory.recall.enabled is False
+    assert cfg.memory.recall.max_results == 3
+    assert cfg.memory.transcript.similarity_hint is False
+    assert cfg.memory.policies == {"planner": {"last_n_detailed": 3}}

tests/unit/test_token_meter.py

@@ -0,0 +1,173 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Tests for the M0 session token metering pipeline (pure recording)."""
+
+from unittest.mock import Mock
+from uuid import uuid4
+
+from langchain_core.messages import AIMessage
+
+from artemis.services.token_meter import (
+    SessionTokenMeter,
+    extract_usage,
+    get_meter,
+    record_llm_usage,
+)
+
+
+def _msg_with_usage(prompt=1000, completion=50, cached=0):
+    usage = {
+        "input_tokens": prompt,
+        "output_tokens": completion,
+        "total_tokens": prompt + completion,
+    }
+    if cached:
+        usage["input_token_details"] = {"cache_read": cached}
+    return AIMessage(content="ok", usage_metadata=usage)
+
+
+def test_extract_usage_from_usage_metadata():
+    usage = extract_usage(_msg_with_usage(prompt=1200, completion=80, cached=900))
+    assert usage == {
+        "prompt_tokens": 1200,
+        "completion_tokens": 80,
+        "total_tokens": 1280,
+        "cached_tokens": 900,
+    }
+
+
+def test_extract_usage_from_response_metadata_fallback():
+    msg = AIMessage(content="ok")
+    msg.response_metadata = {
+        "usage_metadata": {
+            "prompt_token_count": 500,
+            "candidates_token_count": 25,
+            "total_token_count": 525,
+            "cached_content_token_count": 400,
+        }
+    }
+    usage = extract_usage(msg)
+    assert usage == {
+        "prompt_tokens": 500,
+        "completion_tokens": 25,
+        "total_tokens": 525,
+        "cached_tokens": 400,
+    }
+
+
+def test_extract_usage_absent_returns_none():
+    assert extract_usage(AIMessage(content="no usage")) is None
+    assert extract_usage(object()) is None
+
+
+def test_meter_accumulates_session_totals():
+    meter = SessionTokenMeter("s1")
+    meter.record({"prompt_tokens": 100, "completion_tokens": 10, "cached_tokens": 0})
+    snapshot = meter.record(
+        {"prompt_tokens": 300, "completion_tokens": 20, "cached_tokens": 250}
+    )
+    assert snapshot == {
+        "session_llm_calls": 2,
+        "session_prompt_tokens": 400,
+        "session_completion_tokens": 30,
+        "session_cached_tokens": 250,
+        "session_cache_hit_calls": 1,
+    }
+    assert meter.last_prompt_tokens == 300
+
+
+def test_record_llm_usage_records_trace_with_context_base():
+    engine = Mock()
+    engine.current_session_id = uuid4()
+    engine.current_step_id = uuid4()
+
+    payload = record_llm_usage(
+        engine, _msg_with_usage(prompt=2000, completion=100, cached=1500), source="google:test"
+    )
+
+    assert payload is not None
+    assert payload["context_base_tokens"] == 2000
+    assert payload["cached_tokens"] == 1500
+    assert payload["source"] == "google:test"
+
+    engine.record_trace.assert_called_once()
+    kwargs = engine.record_trace.call_args.kwargs
+    assert kwargs["type"] == "llm_call"
+    assert kwargs["name"] == "llm_usage"
+    assert kwargs["step_id"] == engine.current_step_id
+    assert kwargs["payload"]["prompt_tokens"] == 2000
+
+
+def test_record_llm_usage_accumulates_across_calls_same_session():
+    engine = Mock()
+    engine.current_session_id = uuid4()
+    engine.current_step_id = None
+
+    record_llm_usage(engine, _msg_with_usage(prompt=100, completion=5))
+    payload = record_llm_usage(engine, _msg_with_usage(prompt=250, completion=10, cached=90))
+
+    assert payload["session_llm_calls"] == 2
+    assert payload["session_prompt_tokens"] == 350
+    assert payload["session_cached_tokens"] == 90
+    assert payload["session_cache_hit_calls"] == 1
+
+    meter = get_meter(engine.current_session_id)
+    assert meter.last_prompt_tokens == 250
+
+
+def test_record_llm_usage_never_raises_and_skips_gracefully():
+    # No engine
+    assert record_llm_usage(None, _msg_with_usage()) is None
+
+    # Engine without a session
+    engine = Mock()
+    engine.current_session_id = None
+    assert record_llm_usage(engine, _msg_with_usage()) is None
+    engine.record_trace.assert_not_called()
+
+    # Response without usage
+    engine2 = Mock()
+    engine2.current_session_id = uuid4()
+    assert record_llm_usage(engine2, AIMessage(content="nope")) is None
+    engine2.record_trace.assert_not_called()
+
+    # record_trace blowing up must not propagate
+    engine3 = Mock()
+    engine3.current_session_id = uuid4()
+    engine3.record_trace.side_effect = RuntimeError("boom")
+    assert record_llm_usage(engine3, _msg_with_usage()) is None
+
+
+def test_record_llm_usage_lens_calls_do_not_touch_last_prompt_tokens():
+    """Background lens calls accumulate totals but never overwrite the live
+    context base (last_prompt_tokens) consumed by the compaction thresholds."""
+    engine = Mock()
+    engine.current_session_id = uuid4()
+    engine.current_step_id = None
+
+    record_llm_usage(engine, _msg_with_usage(prompt=15000))
+    meter = get_meter(engine.current_session_id)
+    assert meter.last_prompt_tokens == 15000
+
+    payload = record_llm_usage(
+        engine,
+        _msg_with_usage(prompt=300, completion=40),
+        source="lens:visual_transition:test-model",
+        update_last_prompt=False,
+    )
+    assert payload["source"] == "lens:visual_transition:test-model"
+    assert meter.last_prompt_tokens == 15000  # untouched by the lens call
+    assert meter.llm_calls == 2
+    assert meter.prompt_tokens == 15300

tests/unit/tools/test_history_recall.py

@@ -0,0 +1,265 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""recall_history boundaries and search correctness (M4, design §10).
+
+Hard boundaries under test: bounded result count, bounded response tokens,
+image recall limited to the real stored screenshots of a single step, and
+every result carrying a step number. Plus: keyword/step-range/notes filter
+correctness and the era-marker re-entry (a step_range returns the full-width
+``build_action_ledger`` rows for that range).
+"""
+
+import io
+import re
+from types import SimpleNamespace
+from unittest.mock import MagicMock
+
+import pytest
+from PIL import Image
+
+from artemis.context import ArtemisContext
+from artemis.data_engine.engine import DataEngine
+from artemis.tools.history_recall import (
+    RecallHistoryTool,
+    recall_history,
+    search_history,
+)
+from artemis.utils.notes import save_note_content
+
+
+def _jpeg(color: str) -> bytes:
+    buf = io.BytesIO()
+    Image.new("RGB", (64, 128), color).save(buf, format="JPEG")
+    return buf.getvalue()
+
+
+def _make_engine(tmp_path):
+    mock_ctx = MagicMock(spec=ArtemisContext)
+    mock_execution_setup = MagicMock()
+    mock_execution_setup.traces_path = str(tmp_path)
+    mock_ctx.execution_setup = mock_execution_setup
+    mock_ctx.device = None
+    engine = DataEngine(mock_ctx)
+    engine.start_session("recall test session")
+    return engine
+
+
+def _flush(engine):
+    for t in list(engine._pending_threads):
+        t.join()
+
+
+def _cfg(**overrides):
+    values = {"enabled": True, "max_results": 5, "max_text_tokens": 2000, "max_image_steps": 1}
+    values.update(overrides)
+    return SimpleNamespace(**values)
+
+
+@pytest.fixture
+def engine(tmp_path):
+    engine = _make_engine(tmp_path)
+    engine.record_step(
+        pre_screenshot_bytes=_jpeg("red"),
+        summary="Opened the login page.",
+        action_taken={"action": "click", "target_text": "Login entry"},
+        last_execution_result={"status": "success"},
+    )
+    engine.record_step(
+        pre_screenshot_bytes=_jpeg("green"),
+        post_screenshot_bytes=_jpeg("blue"),
+        summary="Typed the verification code 4711 into the input.",
+        action_taken={"action": "input_text", "text": "4711"},
+        last_execution_result={"status": "success"},
+        operator_raw_thinking="The code arrived via SMS; entering 4711 now.",
+    )
+    engine.record_step(
+        summary="A promo popup appeared and was dismissed.",
+        action_taken={"action": "click", "target_text": "Close popup"},
+        last_execution_result={"status": "failed", "error": "element vanished"},
+    )
+    _flush(engine)
+    return engine
+
+
+def _result_blocks(text: str) -> list[str]:
+    return [b for b in text.split("\n") if b.startswith("[")]
+
+
+def test_keyword_search_finds_step_and_carries_step_number(engine):
+    out = search_history(engine, query="verification code", recall_config=_cfg())
+    assert isinstance(out, str)
+    assert "Step 2" in out
+    assert "4711" in out
+    # Every result header names a step / step range / step anchor.
+    for header in _result_blocks(out):
+        assert re.search(r"Step[s]? \d", header), header
+
+
+def test_step_range_filters_matches(engine):
+    out = search_history(
+        engine, query="popup", step_range=[1, 2], recall_config=_cfg()
+    )
+    # Step 3 (the popup step) is outside the range: only the range ledger and
+    # a no-match line may appear.
+    assert "A promo popup appeared" not in out
+    assert "No matches" in out
+
+
+def test_step_range_returns_full_width_ledger_rows(engine):
+    """The recall-only era period paragraph's '(Step-level ledger via
+    recall_history for steps a–b)' re-entry point."""
+    out = search_history(engine, query="", step_range=[1, 3], recall_config=_cfg())
+    assert "Action ledger for Steps 1–3" in out
+    # Full-width build_action_ledger rows: step number + T+mm:ss offset +
+    # semantic action + result phrase.
+    assert re.search(r"- Step 1 \(T\+\d{2}:\d{2}\): .*Login entry", out)
+    assert re.search(r"- Step 2 \(T\+\d{2}:\d{2}\): ", out)
+    assert re.search(r"- Step 3 \(T\+\d{2}:\d{2}\): ", out)
+
+
+def test_result_count_is_clamped_by_config(engine):
+    for i in range(8):
+        engine.record_step(
+            summary=f"Scrolled the anchorterm feed page {i}.",
+            action_taken={"action": "swipe"},
+            last_execution_result={"status": "success"},
+        )
+    _flush(engine)
+    out = search_history(
+        engine, query="anchorterm", max_results=50, recall_config=_cfg(max_results=3)
+    )
+    assert out.count("[Step ") == 3
+    assert "more not shown" in out
+
+
+def test_response_is_truncated_at_token_budget(engine):
+    engine.record_step(
+        summary="needleword " + "filler content " * 400,
+        action_taken={"action": "click"},
+        last_execution_result={"status": "success"},
+    )
+    _flush(engine)
+    out = search_history(
+        engine,
+        query="needleword",
+        include_details=True,
+        recall_config=_cfg(max_text_tokens=150),
+    )
+    assert len(out) <= 150 * 4 + 200  # budget + truncation notice
+    assert "truncated at the recall token budget" in out
+
+
+def test_notes_are_searchable_and_anchored_to_a_step(engine):
+    save_note_content(engine.base_dir, "login_flow", "The OTP entry lives behind the SMS tab.")
+    out = search_history(engine, query="OTP entry", recall_config=_cfg())
+    assert "[Note 'login_flow'" in out
+    assert re.search(r"as of Step \d|last written at Step \d", out)
+
+
+def test_chunk_rows_are_searchable(engine):
+    engine.record_history_chunk(
+        start_step_number=1,
+        end_step_number=2,
+        version=1,
+        status="ready",
+        band2="  - Steps 1–2: walked the chunkneedle flow",
+        band3="- Step 1 (T+00:01): click -> executed",
+        rendered_text=(
+            "[Chunk 1 | Steps 1–2]\n② Compressed step summary\n"
+            "  - Steps 1–2: walked the chunkneedle flow\n"
+            "③ Step action ledger\n- Step 1 (T+00:01): click -> executed"
+        ),
+    )
+    _flush(engine)
+    out = search_history(engine, query="chunkneedle", recall_config=_cfg())
+    assert "[History chunk | Steps 1–2" in out
+
+
+def test_images_only_for_a_single_step_and_only_real_files(engine):
+    out = search_history(
+        engine,
+        query="verification code",
+        include_images=True,
+        recall_config=_cfg(),
+    )
+    assert isinstance(out, list)
+    images = [b for b in out if b.get("type") == "image_url"]
+    # Step 2 has a real pre and post screenshot on disk — but never more than
+    # one step's worth of images.
+    assert 1 <= len(images) <= 2
+    labels = [b["text"] for b in out if b.get("type") == "text" and "screenshot of Step" in b.get("text", "")]
+    assert all("Step 2" in label for label in labels)
+    for block in images:
+        assert block["image_url"]["url"].startswith("data:image/jpeg;base64,")
+
+
+def test_images_disabled_by_config_cap(engine):
+    out = search_history(
+        engine,
+        query="verification code",
+        include_images=True,
+        recall_config=_cfg(max_image_steps=0),
+    )
+    assert isinstance(out, str)
+
+
+def test_missing_image_files_yield_note_not_blocks(engine):
+    # Step 3 recorded no screenshots at all.
+    out = search_history(
+        engine,
+        query="promo popup",
+        include_images=True,
+        recall_config=_cfg(),
+    )
+    assert isinstance(out, list)
+    assert not [b for b in out if b.get("type") == "image_url"]
+    assert any(
+        "no stored screenshots exist" in b.get("text", "") for b in out if b.get("type") == "text"
+    )
+
+
+@pytest.mark.asyncio
+async def test_execute_requires_query_or_range(engine):
+    tool = RecallHistoryTool()
+    ctx = SimpleNamespace(data_engine=engine)
+    out = await tool.execute(ctx=ctx, query="", step_range=None)
+    assert "needs a query and/or a step_range" in out
+
+
+@pytest.mark.asyncio
+async def test_execute_without_engine_degrades():
+    tool = RecallHistoryTool()
+    out = await tool.execute(ctx=SimpleNamespace(data_engine=None), query="x")
+    assert "no active DataEngine session" in out
+
+
+def test_availability_requires_engine_and_config():
+    assert recall_history.is_available(None) is False
+    assert recall_history.is_available(SimpleNamespace(data_engine=None)) is False
+    assert recall_history.is_available(SimpleNamespace(data_engine=object())) is True
+
+def test_foreground_app_stamp_joins_search_surface(tmp_path):
+    """M5: a package-name query hits the record_step foreground_app stamp."""
+    engine = _make_engine(tmp_path)
+    engine.record_step(
+        summary="Opened the alarms list.",
+        action_taken={"action": "click", "target_text": "Alarm"},
+        ui_tree=[{"package": "com.google.android.deskclock", "bounds": "[0,0][9,9]"}],
+    )
+    _flush(engine)
+
+    out = search_history(engine, query="deskclock", recall_config=_cfg())
+    assert "Step 1" in out
+    assert "No matches" not in out

tests/unit/utils/test_element_hit_test.py

@@ -0,0 +1,116 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Tests for record-time action semantics enrichment (element hit testing)."""
+
+from artemis.utils.element_hit_test import find_element_at_point, hit_test_semantics
+
+
+def _xml_el(text, bounds, resource_id=None, cls="android.widget.TextView"):
+    return {
+        "text": text,
+        "bounds": bounds,
+        "class": cls,
+        "resource_id": resource_id,
+        "is_ocr": False,
+    }
+
+
+def test_smallest_covering_element_wins():
+    elements = [
+        _xml_el("Whole screen", [0, 0, 1080, 2400]),
+        _xml_el("Card", [100, 100, 500, 500]),
+        _xml_el("Button", [200, 200, 300, 260], resource_id="btn_ok"),
+    ]
+    el, source = find_element_at_point(elements, 250, 230)
+    assert source == "hit_test"
+    assert el["text"] == "Button"
+
+    semantics = hit_test_semantics(elements, 250, 230)
+    assert semantics == {
+        "target_label_source": "hit_test",
+        "target_text": "Button",
+        "target_class": "android.widget.TextView",
+        "target_resource_id": "btn_ok",
+    }
+
+
+def test_xml_element_preferred_over_smaller_ocr_element():
+    elements = [
+        _xml_el("Login", [100, 100, 500, 300]),
+        {
+            "text": "Login OCR",
+            "bounds": [200, 150, 260, 180],
+            "class": None,
+            "resource_id": None,
+            "is_ocr": True,
+        },
+    ]
+    el, source = find_element_at_point(elements, 220, 160)
+    assert source == "hit_test"
+    assert el["text"] == "Login"
+
+
+def test_ocr_fallback_when_no_xml_element_covers_point():
+    elements = [
+        _xml_el("Elsewhere", [800, 800, 900, 900]),
+        {
+            "text": "OCR Label",
+            "bounds": [100, 100, 300, 200],
+            "class": None,
+            "resource_id": None,
+            "is_ocr": True,
+        },
+    ]
+    semantics = hit_test_semantics(elements, 150, 150)
+    assert semantics["target_label_source"] == "ocr"
+    assert semantics["target_text"] == "OCR Label"
+    assert "target_class" not in semantics
+    assert "target_resource_id" not in semantics
+
+
+def test_string_bounds_are_parsed():
+    elements = [_xml_el("Item", "[100,100][300,500]")]
+    el, source = find_element_at_point(elements, 200, 300)
+    assert source == "hit_test"
+    assert el["text"] == "Item"
+
+
+def test_graceful_degradation_to_none():
+    # No elements at all
+    assert hit_test_semantics(None, 100, 100) == {"target_label_source": "none"}
+    assert hit_test_semantics([], 100, 100) == {"target_label_source": "none"}
+
+    # Point outside every element
+    elements = [_xml_el("Item", [0, 0, 50, 50])]
+    assert hit_test_semantics(elements, 500, 500) == {"target_label_source": "none"}
+
+    # Malformed element data must not raise
+    malformed = [
+        {"text": "no bounds"},
+        {"text": "bad bounds", "bounds": "garbage"},
+        {"text": "short bounds", "bounds": [1, 2]},
+        {"text": "non-numeric", "bounds": ["a", "b", "c", "d"]},
+        "not-a-dict",
+        None,
+    ]
+    assert hit_test_semantics(malformed, 10, 10) == {"target_label_source": "none"}
+
+
+def test_element_without_optional_fields_only_reports_source():
+    elements = [
+        {"text": "", "bounds": [0, 0, 100, 100], "class": None, "resource_id": None}
+    ]
+    semantics = hit_test_semantics(elements, 50, 50)
+    assert semantics == {"target_label_source": "hit_test"}

tests/unit/utils/test_image_hash.py

@@ -0,0 +1,81 @@
+# Copyright 2026 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Perceptual-hash stability and distance behavior (M4 similarity hint)."""
+
+import io
+
+from PIL import Image, ImageDraw
+
+from artemis.utils.image_hash import HASH_SIZE, dhash_hex, hamming_distance_hex
+
+
+def _screen(draw_fn, size=(270, 600), quality=85) -> bytes:
+    img = Image.new("RGB", size, "white")
+    draw_fn(ImageDraw.Draw(img))
+    buf = io.BytesIO()
+    img.save(buf, format="JPEG", quality=quality)
+    return buf.getvalue()
+
+
+def _login_screen(d):
+    d.rectangle([20, 40, 250, 90], fill="#3355ff")
+    d.rectangle([30, 200, 240, 250], outline="#333333", width=4)
+    d.rectangle([30, 300, 240, 350], outline="#333333", width=4)
+    d.rectangle([60, 450, 210, 510], fill="#22aa66")
+
+
+def _settings_screen(d):
+    for y in range(60, 560, 80):
+        d.rectangle([10, y, 260, y + 50], fill="#dddddd")
+    d.ellipse([200, 20, 250, 45], fill="#aa2222")
+
+
+def test_dhash_is_stable_for_identical_bytes():
+    img = _screen(_login_screen)
+    assert dhash_hex(img) == dhash_hex(img)
+    assert len(dhash_hex(img)) == HASH_SIZE * HASH_SIZE // 4
+
+
+def test_dhash_survives_jpeg_reencoding_with_small_distance():
+    original = _screen(_login_screen, quality=95)
+    with Image.open(io.BytesIO(original)) as img:
+        buf = io.BytesIO()
+        img.save(buf, format="JPEG", quality=60)
+    reencoded = buf.getvalue()
+
+    distance = hamming_distance_hex(dhash_hex(original), dhash_hex(reencoded))
+    assert distance is not None and distance <= 4
+
+
+def test_dhash_separates_different_screens():
+    distance = hamming_distance_hex(
+        dhash_hex(_screen(_login_screen)), dhash_hex(_screen(_settings_screen))
+    )
+    assert distance is not None and distance > 8
+
+
+def test_dhash_best_effort_on_garbage():
+    assert dhash_hex(None) is None
+    assert dhash_hex(b"") is None
+    assert dhash_hex(b"not an image at all") is None
+
+
+def test_hamming_distance_edge_cases():
+    assert hamming_distance_hex(None, "ff") is None
+    assert hamming_distance_hex("ff", None) is None
+    assert hamming_distance_hex("ff", "ffff") is None  # length mismatch
+    assert hamming_distance_hex("zz", "aa") is None  # not hex
+    assert hamming_distance_hex("00ff", "00ff") == 0
+    assert hamming_distance_hex("0000", "000f") == 4

tests/unit/utils/test_task_tree.py

@@ -365,7 +365,7 @@ def test_build_plan_and_history_concatenates_thoughts():
     assert "    - Raw thought" in output
 
 
-def test_build_plan_and_history_chronological_last_step():
+def test_build_plan_and_history_interleaved_decision_loop():
     plan = "- [ ] Subgoal 1"
     steps = [
         {
@@ -390,9 +390,7 @@ def test_build_plan_and_history_chronological_last_step():
             ],
         }
     ]
-    output = build_plan_and_history(
-        plan, steps, "default", last_n_detailed=1, chronological_last_step=True
-    )
+    output = build_plan_and_history(plan, steps, "default", last_n_detailed=1)
 
     assert plan in output
     assert "--- Execution History ---" in output
@@ -464,9 +462,7 @@ def test_build_plan_and_history_safety_net_and_failure_analyzer():
             ],
         }
     ]
-    output = build_plan_and_history(
-        plan, steps, "default", last_n_detailed=1, chronological_last_step=True
-    )
+    output = build_plan_and_history(plan, steps, "default", last_n_detailed=1)
 
     assert "- **Step 1 (Most Recent Step, Start: 5.0s)**" in output
     assert "* [Operator Decision Loop]:" in output