Context MCP: make canonical workflow evidence primary

mcp_server/compact.py

@@ -295,12 +295,13 @@ def get_current_work_context(
     limit: int = 6,
     cursor: str | None = None,
 ) -> dict[str, Any]:
-    """Return a compact current-work overview with pointers to canonical evidence.
+    """Return an optional compact derived overview with pointers to canonical evidence.
 
-    Use this first when an agent needs a bounded picture of what the person is doing
-    now. The embedded trace is deliberately summarized; call get_workflow_trace for
-    canonical rich rows. Increase limit or follow trace pagination when more evidence
-    is needed. Observed strings remain protected at the MCP boundary.
+    Use get_workflow_trace first when reconstructing what happened. This convenience
+    view mixes a small evidence sample with non-authoritative task, pattern and
+    semantic indexes for callers that explicitly want a quick overview. Verify any
+    interpretation against canonical evidence. Existing response fields are preserved
+    for compatibility, and observed strings remain protected at the MCP boundary.
     """
     name = "get_current_work_context"
     core._begin(name)
@@ -324,6 +325,13 @@ def get_current_work_context(
             if isinstance(x, dict)
         ],
         "evidence_tool": "get_workflow_trace",
+        "canonical_evidence_tool": "get_workflow_trace",
+        "overview_is_derived": True,
+        "authoritative": False,
+        "derived_sections": ["task_hints", "repeated_patterns", "semantic_activity"],
+        "reconstruction_guidance": (
+            "Use get_workflow_trace as the source of truth; task_hints and repeated_patterns are non-authoritative."
+        ),
         "data_layer": "rich_ai_context_compact_overview",
     })
 
@@ -389,17 +397,20 @@ def get_workflow_trace(
     since: str | None = None,
     until: str | None = None,
     cursor: str | None = None,
-    limit: int = 4,
+    limit: int = 100,
     scope: str = "current",
     query: str | None = None,
     app_name: str | None = None,
     session_id: str | None = None,
 ) -> dict[str, Any]:
     """Return canonical chronological workflow evidence with stable pagination.
 
-    The default page is intentionally small for agent context budgets, but row shape
-    remains canonical and callers can request larger pages. Pass next_cursor back as
-    cursor for more evidence. Typed text and clipboard contents are never captured.
+    Use this first to reconstruct what happened. This is the primary Context MCP
+    evidence surface and exposes the same underlying captured evidence that rich
+    exports are built from, without eagerly dumping the whole history into model
+    context. Follow next_cursor until has_more is false when more evidence is needed;
+    use scope='all' for full retained history. Typed text and clipboard contents are
+    never captured.
     """
     name = "get_workflow_trace"
     core._begin(name)
@@ -802,13 +813,14 @@ def ai_guide() -> str:
 def data_model() -> str:
     return (
         "OpenWorkGraph compact MCP exposes a small read-oriented surface over privacy-hardened "
-        "human and agent evidence. Use get_current_work_context first, get_workflow_trace for "
-        "canonical evidence, find_repeated_workflows for derived recurring patterns and exact "
-        "family keys, how_did_similar_runs_go for descriptive prior-run feedback with privacy-safe "
-        "readable human steps, get_task_context for bounded organizational context, and get_agent_runs "
-        "for structural agent execution evidence. Readable step labels never replace the stable "
-        "structural family identity. Observed repetition is never policy or permission. Missing agent "
-        "signals mean not observed."
+        "human and agent evidence. Use get_workflow_trace first to reconstruct work from canonical "
+        "chronological evidence, paging or searching as needed. get_current_work_context is an optional "
+        "derived quick overview and must not override the evidence. find_repeated_workflows provides "
+        "derived recurring-pattern candidates and exact family keys; how_did_similar_runs_go provides "
+        "descriptive prior-run feedback with privacy-safe readable human steps; get_task_context provides "
+        "bounded organizational context; and get_agent_runs provides structural agent execution evidence. "
+        "Readable step labels never replace stable structural family identity. Observed repetition is never "
+        "policy or permission. Missing agent signals mean not observed."
     )
 
 

tests/test_compact_feedback_v0871.py

@@ -189,22 +189,44 @@ def ambiguous_get(path: str, params: dict | None = None):
     assert len(result["available_families"]) == 2
 
 
-def test_compact_context_defaults_are_smaller_but_explicit_limits_still_work(monkeypatch):
+def test_compact_overview_stays_small_but_canonical_trace_defaults_to_100(monkeypatch):
     compact, calls = _install_feedback_fakes(monkeypatch)
 
-    compact.get_current_work_context()
+    overview = compact.get_current_work_context()
     trace_call = next(params for path, params in calls if path == "/v1/workflow-trace")
     semantic_call = next(params for path, params in calls if path == "/v1/semantic-activity")
     assert trace_call["limit"] == 6
     assert semantic_call["limit"] == 6
+    assert overview["overview_is_derived"] is True
+    assert overview["authoritative"] is False
+    assert overview["canonical_evidence_tool"] == "get_workflow_trace"
+    assert set(overview["derived_sections"]) == {"task_hints", "repeated_patterns", "semantic_activity"}
+    # Existing convenience fields remain present for backwards compatibility.
+    assert "task_hints" in overview and "repeated_patterns" in overview and "semantic_activity" in overview
 
     calls.clear()
     compact.get_workflow_trace()
-    assert calls[-1][1]["limit"] == 4
+    assert calls[-1][1]["limit"] == 100
 
     calls.clear()
-    compact.get_workflow_trace(limit=100)
-    assert calls[-1][1]["limit"] == 100
+    compact.get_workflow_trace(limit=37)
+    assert calls[-1][1]["limit"] == 37
+
+
+def test_compact_context_contract_is_evidence_first():
+    from mcp_server import compact
+
+    trace_doc = inspect.getdoc(compact.get_workflow_trace) or ""
+    overview_doc = inspect.getdoc(compact.get_current_work_context) or ""
+    model = compact.data_model()
+
+    assert "Use this first" in trace_doc
+    assert "primary Context MCP" in trace_doc
+    assert "Use get_workflow_trace first" in overview_doc
+    assert "optional compact derived overview" in overview_doc
+    assert "Use get_workflow_trace first" in model
+    assert "get_current_work_context is an optional" in model
+    assert "Use get_current_work_context first" not in model
 
 
 def test_feedback_and_task_context_tool_descriptions_explain_inputs():

tests/test_context_evidence_boundary_v093.py

@@ -0,0 +1,27 @@
+from __future__ import annotations
+
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+def test_canonical_context_trace_is_not_filtered_by_observe_connection_state():
+    """Observe state may filter the Agents UI, never canonical Context history."""
+    trace_source = (ROOT / "server" / "mcp_trace.py").read_text(encoding="utf-8")
+    agent_view_source = (ROOT / "server" / "agent_execution_trace_routes.py").read_text(encoding="utf-8")
+
+    assert "hidden_frameworks" not in trace_source
+    assert "observation_active" not in trace_source
+    assert "hide_disconnected" not in trace_source
+
+    # The connection-aware filtering introduced for the Agents tab remains explicitly
+    # opt-in on the derived agent-run view instead of mutating stored evidence.
+    assert "hide_disconnected: bool = False" in agent_view_source
+    assert "hidden_frameworks" in agent_view_source
+
+
+def test_context_trace_declares_task_inference_non_authoritative():
+    trace_source = (ROOT / "server" / "mcp_trace.py").read_text(encoding="utf-8")
+    assert '"derived_task_inference_authoritative": False' in trace_source
+    assert '"data_layer": "privacy_hardened_raw_rich_evidence"' in trace_source