fix(agent): frame compaction handoff sections as historical context

agent/context_compressor.py

@@ -34,6 +34,12 @@ from agent.redact import redact_sensitive_text
 
 logger = logging.getLogger(__name__)
 
+HISTORICAL_TASK_HEADING = "## Historical Task Snapshot"
+HISTORICAL_IN_PROGRESS_HEADING = "## Historical In-Progress State"
+HISTORICAL_PENDING_ASKS_HEADING = "## Historical Pending User Asks"
+HISTORICAL_REMAINING_WORK_HEADING = "## Historical Remaining Work"
+
+
 SUMMARY_PREFIX = (
     "[CONTEXT COMPACTION — REFERENCE ONLY] Earlier turns were compacted "
     "into the summary below. This is a handoff from a previous context "
@@ -43,12 +49,14 @@ SUMMARY_PREFIX = (
     "Respond ONLY to the latest user message that appears AFTER this "
     "summary — that message is the single source of truth for what to do "
     "right now. "
-    "If the latest user message is consistent with the '## Active Task' "
+    f"If the latest user message is consistent with the '{HISTORICAL_TASK_HEADING}' "
     "section, you may use the summary as background. If the latest user "
     "message contradicts, supersedes, changes topic from, or in any way "
-    "diverges from '## Active Task' / '## In Progress' / '## Pending User "
-    "Asks' / '## Remaining Work', the latest message WINS — discard those "
-    "stale items entirely and do not 'wrap up the old task first'. "
+    f"diverges from '{HISTORICAL_TASK_HEADING}' / "
+    f"'{HISTORICAL_IN_PROGRESS_HEADING}' / "
+    f"'{HISTORICAL_PENDING_ASKS_HEADING}' / "
+    f"'{HISTORICAL_REMAINING_WORK_HEADING}', the latest message WINS — "
+    "discard those stale items entirely and do not 'wrap up the old task first'. "
     "Reverse signals in the latest message (e.g. 'stop', 'undo', 'roll "
     "back', 'just verify', 'don't do that anymore', 'never mind', a new "
     "topic) must immediately end any in-flight work described in the "
@@ -1155,7 +1163,7 @@ class ContextCompressor(ContextEngine):
             )
 
         reason_text = f" Summary failure reason: {reason}." if reason else ""
-        body = f"""## Active Task
+        body = f"""{HISTORICAL_TASK_HEADING}
 {active_task}
 
 ## Goal
@@ -1172,7 +1180,7 @@ Recovered from a deterministic fallback because the LLM context summarizer was u
 ## Active State
 Unknown from deterministic fallback. Inspect current repository/session state if needed.
 
-## In Progress
+{HISTORICAL_IN_PROGRESS_HEADING}
 {active_task}
 
 ## Blocked
@@ -1184,13 +1192,13 @@ None recoverable from deterministic fallback.
 ## Resolved Questions
 None recoverable from deterministic fallback.
 
-## Pending User Asks
+{HISTORICAL_PENDING_ASKS_HEADING}
 {active_task}
 
 ## Relevant Files
 {_bullets(relevant_files, limit=12)}
 
-## Remaining Work
+{HISTORICAL_REMAINING_WORK_HEADING}
 Continue from the most recent unfulfilled user ask and protected tail messages. Verify state with tools before making claims.
 
 ## Last Dropped Turns
@@ -1312,7 +1320,7 @@ Summary generation was unavailable, so this is a best-effort deterministic fallb
             _temporal_anchoring_rule = ""
 
         # Shared structured template (used by both paths).
-        _template_sections = f"""## Active Task
+        _template_sections = f"""{HISTORICAL_TASK_HEADING}
 [THE SINGLE MOST IMPORTANT FIELD. Capture the user's most recent unfulfilled
 input verbatim — the exact words they used. This includes:
 - Explicit task assignments ("refactor the auth module")
@@ -1359,7 +1367,7 @@ Be specific with file paths, commands, line numbers, and results.]
 - Any running processes or servers
 - Environment details that matter]
 
-## In Progress
+{HISTORICAL_IN_PROGRESS_HEADING}
 [Work currently underway — what was being done when compaction fired]
 
 ## Blocked
@@ -1371,13 +1379,13 @@ Be specific with file paths, commands, line numbers, and results.]
 ## Resolved Questions
 [Questions the user asked that were ALREADY answered — include the answer so it is not repeated]
 
-## Pending User Asks
+{HISTORICAL_PENDING_ASKS_HEADING}
 [Questions or requests from the user that have NOT yet been answered or fulfilled. If none, write "None."]
 
 ## Relevant Files
 [Files read, modified, or created — with brief note on each]
 
-## Remaining Work
+{HISTORICAL_REMAINING_WORK_HEADING}
 [What remains to be done — framed as context, not instructions]
 
 ## Critical Context

tests/agent/test_context_compressor.py

@@ -3,7 +3,11 @@
 import pytest
 from unittest.mock import patch, MagicMock
 
-from agent.context_compressor import ContextCompressor, SUMMARY_PREFIX
+from agent.context_compressor import (
+    ContextCompressor,
+    HISTORICAL_TASK_HEADING,
+    SUMMARY_PREFIX,
+)
 
 
 @pytest.fixture()
@@ -157,7 +161,7 @@ class TestCompress:
             result = c.compress(msgs)
 
         combined = "\n".join(str(m.get("content", "")) for m in result)
-        assert "## Active Task" in combined
+        assert HISTORICAL_TASK_HEADING in combined
         assert "Please fix the compression summary failure" in combined
         assert "read_file" in combined
         assert "agent/context_compressor.py" in combined
@@ -1213,7 +1217,8 @@ class TestCompressWithClient:
         """When the summary lands as standalone role='user' (e.g. head ends
         with assistant/tool), the message body must include the explicit
         '--- END OF CONTEXT SUMMARY ---' marker. Without it, weak models
-        read the verbatim past user request quoted in '## Active Task' as
+        read the verbatim past user request quoted in the historical task
+        snapshot as
         fresh input (#11475, #14521).
         """
         mock_response = MagicMock()

tests/agent/test_context_compressor_temporal_anchoring.py

@@ -15,7 +15,7 @@ from datetime import datetime, timezone
 from unittest.mock import MagicMock, patch
 
 import hermes_time
-from agent.context_compressor import ContextCompressor
+from agent.context_compressor import ContextCompressor, HISTORICAL_TASK_HEADING
 
 
 def _compressor() -> ContextCompressor:
@@ -98,7 +98,7 @@ def test_clock_failure_omits_rule_but_compaction_still_runs():
     prompt = mock_call.call_args.kwargs["messages"][0]["content"]
     assert "TEMPORAL ANCHORING" not in prompt
     # Structured template still intact.
-    assert "## Active Task" in prompt
+    assert HISTORICAL_TASK_HEADING in prompt
 
 
 def test_anchoring_rule_uses_date_from_hermes_time_now():

tests/agent/test_resume_stale_active_task.py

@@ -1,9 +1,9 @@
 """Regression coverage for #35344: a resumed session must not let a stale
-``## Active Task`` from an inherited compaction handoff hijack the reply to a
+historical task snapshot from an inherited compaction handoff hijack the reply to a
 new, unrelated user message.
 
 The failure mode (real report): a lineage was compacted, producing a handoff
-whose ``## Active Task`` described task A. The lineage was resumed later and
+whose historical task snapshot described task A. The lineage was resumed later and
 the user asked about an unrelated task B. The model answered with A because
 the handoff's resume directive outranked the fresh ask.
 
@@ -16,14 +16,15 @@ named reverse-signal verbs. Two invariants guard the resume path specifically:
      pre-fix stale handoff cannot keep its "resume exactly" directive forever.
 
   2. The current handoff prefix contains an unambiguous "latest message wins /
-     discard stale Active Task" rule, so an unrelated new ask is privileged over
-     the inherited ``## Active Task``.
+     discard stale historical task" rule, so an unrelated new ask is privileged over
+     the inherited task snapshot.
 
 These are content/structural assertions (no live model call) — they pin the
 mechanism that makes the stale task historical rather than active.
 """
 
 from agent.context_compressor import (
+    HISTORICAL_TASK_HEADING,
     SUMMARY_PREFIX,
     LEGACY_SUMMARY_PREFIX,
     ContextCompressor,
@@ -48,10 +49,10 @@ _OLD_CONFLICTING_PREFIX = (
 
 def test_latest_message_wins_over_inherited_active_task():
     """The handoff must explicitly privilege the latest user message over a
-    stale ``## Active Task`` — the core #35344 contract."""
+    stale historical task snapshot — the core #35344 contract."""
     lower = SUMMARY_PREFIX.lower()
     assert "latest user message" in lower
-    assert "## active task" in lower
+    assert HISTORICAL_TASK_HEADING.lower() in lower
     # Conflict-resolution must be explicit, not implied.
     assert "wins" in lower or "supersede" in lower
     assert "discard" in lower
@@ -69,7 +70,7 @@ def test_resumed_stale_handoff_gets_renormalized_to_current_prefix():
     prefix when re-normalized on re-compaction — so the "resume exactly"
     directive cannot survive into a resumed session."""
     stale_body = (
-        "## Active Task\n"
+        f"{HISTORICAL_TASK_HEADING}\n"
         "User asked: 'Migrate the billing module to Stripe'\n\n"
         "## Goal\nMigrate billing.\n"
     )
@@ -92,7 +93,7 @@ def test_resumed_stale_handoff_gets_renormalized_to_current_prefix():
 def test_legacy_prefix_handoff_also_renormalized():
     """The same upgrade applies to the oldest ``[CONTEXT SUMMARY]:`` handoff
     format that may sit in a long-lived resumed lineage."""
-    legacy = f"{LEGACY_SUMMARY_PREFIX} ## Active Task\nUser asked: 'task A'"
+    legacy = f"{LEGACY_SUMMARY_PREFIX} {HISTORICAL_TASK_HEADING}\nUser asked: 'task A'"
     renormalized = ContextCompressor._with_summary_prefix(legacy)
     assert renormalized.startswith(SUMMARY_PREFIX)
     assert LEGACY_SUMMARY_PREFIX not in renormalized
@@ -107,7 +108,7 @@ def test_inherited_handoff_detected_in_resumed_protected_head():
     Task read as live intent)."""
     messages = [
         {"role": "system", "content": "system prompt"},
-        {"role": "user", "content": f"{SUMMARY_PREFIX}\n## Active Task\nUser asked: 'task A'"},
+        {"role": "user", "content": f"{SUMMARY_PREFIX}\n{HISTORICAL_TASK_HEADING}\nUser asked: 'task A'"},
         {"role": "assistant", "content": "ok"},
         {"role": "user", "content": "Unrelated task B: what's the capital of France?"},
     ]
@@ -129,7 +130,7 @@ def test_historical_prefixed_handoff_detected_and_stripped():
     stale 'resume exactly' text as a fresh turn."""
     messages = [
         {"role": "system", "content": "system prompt"},
-        {"role": "user", "content": f"{_OLD_CONFLICTING_PREFIX}\n## Active Task\nUser asked: 'task A'"},
+        {"role": "user", "content": f"{_OLD_CONFLICTING_PREFIX}\n{HISTORICAL_TASK_HEADING}\nUser asked: 'task A'"},
         {"role": "assistant", "content": "ok"},
         {"role": "user", "content": "Unrelated task B"},
     ]

tests/agent/test_summary_prefix_semantics.py

@@ -18,7 +18,13 @@ the agent repeatedly re-surfacing already-cancelled work across turns.
 These tests pin the post-fix invariants so the conflict cannot regress.
 """
 
-from agent.context_compressor import SUMMARY_PREFIX
+from agent.context_compressor import (
+    HISTORICAL_IN_PROGRESS_HEADING,
+    HISTORICAL_PENDING_ASKS_HEADING,
+    HISTORICAL_REMAINING_WORK_HEADING,
+    HISTORICAL_TASK_HEADING,
+    SUMMARY_PREFIX,
+)
 
 
 def test_no_resume_exactly_directive():
@@ -30,10 +36,24 @@ def test_latest_message_wins_on_conflict():
     """The prefix must explicitly say latest user message wins on conflict."""
     lower = SUMMARY_PREFIX.lower()
     assert "latest user message" in lower
+    assert HISTORICAL_TASK_HEADING.lower() in lower
+    assert HISTORICAL_PENDING_ASKS_HEADING.lower() in lower
+    assert HISTORICAL_REMAINING_WORK_HEADING.lower() in lower
     # Must have an explicit conflict-resolution rule.
     assert "wins" in lower or "supersede" in lower or "discard" in lower
 
 
+def test_handoff_sections_are_framed_as_historical():
+    """The summary headings referenced in the prefix must sound historical,
+    not like live instructions for the current turn."""
+    lower = SUMMARY_PREFIX.lower()
+    assert "## active task" not in lower
+    assert "## pending user asks" not in lower
+    assert "## remaining work" not in lower
+    assert HISTORICAL_TASK_HEADING.lower() in lower
+    assert HISTORICAL_IN_PROGRESS_HEADING.lower() in lower
+
+
 def test_reverse_signals_called_out():
     """Reverse signals (stop/undo/never mind/topic change) must be named so
     the model recognizes them as cancellation triggers, not just background."""