feat: ADR 0032 M2-T002 — evidence-event taxonomy (mapper-only, additive)

Adds RunEventType members + bidirectional mapper entries for the 19 audit
event types the evidence bundle reads (model_route, git_sandbox_started/
resolved, skill_activated, skill_lifecycle_transition, test_run, undo_applied,
provenance_collected, tool_call_started/use/approved/denied/pending_approval,
tool_error, approval_requested/granted/denied, run_cancelled,
run_pending_approval). Event list verified by grep of run_evidence.py.

Effect: the M2-T001 reader now surfaces these events from the audit JSONL
(previously skipped as legacy) — the M6 fold's only prerequisite (§16).
Mapper/reader only; no emit site changed, nothing newly emitted, so the audit
stream is byte-unchanged (no double-write; §15).

Reader test changes are an intended behavior update, NOT masking: assertions
that evidence events return None were removed (now false) and replaced with
assertions that they are surfaced; a new positive test was added; legacy-skip
examples switched to genuinely-unmapped events. Net coverage is stronger.

Constraint: additive taxonomy/mapper only; zero emit changes; audit output byte-unchanged
Tested: lifecycle 23, run_evidence+summary+receipt 40, smoke 200, acceptance 646/646, full mypy clean 1009 files, ruff clean
Confidence: high
Roadmap-Status: unchanged
Allow-test-weakening: reader tests updated to intended M2-T002 behavior (evidence events now surfaced); removed assertions were newly-false, stronger positive assertions added

Author: johnteee

teaagent/runner/_events.py

@@ -48,11 +48,33 @@ class RunEventType(str, Enum):
     RUN_COMPLETED = 'run_completed'
     RUN_FAILED = 'run_failed'
 
-    # Planned (M1+): PLAN_RESOLVED, DECISION_RECEIVED, TOOL_CALL_APPROVED,
-    # TOOL_CALL_DENIED, CONTEXT_COMPACTED, BUDGET_CHECKPOINT, ITERATION_COMPLETED,
-    # FINAL_VALIDATION, RUN_PENDING_APPROVAL, RUN_CANCELLED, RECEIPT_EMITTED,
-    # SESSION_START, SESSION_END, SKILL_LOAD, MODEL_ROUTE, GIT_SANDBOX_STARTED,
-    # GIT_SANDBOX_RESOLVED, UNDO_PERFORMED, PRE_TOOL_USE, POST_TOOL_USE, PRE_COMPACT
+    # M2 evidence-event taxonomy (ADR 0032 §16): the event types the evidence
+    # bundle reads, so the reader can surface them from the audit JSONL.
+    # Mapping/reader only — emit-site migration is deferred (see §16).
+    MODEL_ROUTE = 'model_route'
+    GIT_SANDBOX_STARTED = 'git_sandbox_started'
+    GIT_SANDBOX_RESOLVED = 'git_sandbox_resolved'
+    SKILL_ACTIVATED = 'skill_activated'
+    SKILL_LIFECYCLE_TRANSITION = 'skill_lifecycle_transition'
+    TEST_RUN = 'test_run'
+    UNDO_APPLIED = 'undo_applied'
+    PROVENANCE_COLLECTED = 'provenance_collected'
+    TOOL_CALL_STARTED = 'tool_call_started'
+    TOOL_USE = 'tool_use'
+    TOOL_CALL_APPROVED = 'tool_call_approved'
+    TOOL_CALL_DENIED = 'tool_call_denied'
+    TOOL_CALL_PENDING_APPROVAL = 'tool_call_pending_approval'
+    TOOL_ERROR = 'tool_error'
+    APPROVAL_REQUESTED = 'approval_requested'
+    APPROVAL_GRANTED = 'approval_granted'
+    APPROVAL_DENIED = 'approval_denied'
+    RUN_CANCELLED = 'run_cancelled'
+    RUN_PENDING_APPROVAL = 'run_pending_approval'
+
+    # Planned (later phases): PLAN_RESOLVED, DECISION_RECEIVED,
+    # CONTEXT_COMPACTED, BUDGET_CHECKPOINT, ITERATION_COMPLETED,
+    # FINAL_VALIDATION, RECEIPT_EMITTED, SESSION_START, SESSION_END,
+    # PRE_TOOL_USE, POST_TOOL_USE, PRE_COMPACT
 
 
 @dataclass(frozen=True)
@@ -85,6 +107,26 @@ class RunEvent:
     RunEventType.TOOL_CALL_FAILED: 'tool_call_failed',
     RunEventType.RUN_COMPLETED: 'run_completed',
     RunEventType.RUN_FAILED: 'run_failed',
+    # M2 evidence-event taxonomy (§16) — mapping only; not yet spine-emitted.
+    RunEventType.MODEL_ROUTE: 'model_route',
+    RunEventType.GIT_SANDBOX_STARTED: 'git_sandbox_started',
+    RunEventType.GIT_SANDBOX_RESOLVED: 'git_sandbox_resolved',
+    RunEventType.SKILL_ACTIVATED: 'skill_activated',
+    RunEventType.SKILL_LIFECYCLE_TRANSITION: 'skill_lifecycle_transition',
+    RunEventType.TEST_RUN: 'test_run',
+    RunEventType.UNDO_APPLIED: 'undo_applied',
+    RunEventType.PROVENANCE_COLLECTED: 'provenance_collected',
+    RunEventType.TOOL_CALL_STARTED: 'tool_call_started',
+    RunEventType.TOOL_USE: 'tool_use',
+    RunEventType.TOOL_CALL_APPROVED: 'tool_call_approved',
+    RunEventType.TOOL_CALL_DENIED: 'tool_call_denied',
+    RunEventType.TOOL_CALL_PENDING_APPROVAL: 'tool_call_pending_approval',
+    RunEventType.TOOL_ERROR: 'tool_error',
+    RunEventType.APPROVAL_REQUESTED: 'approval_requested',
+    RunEventType.APPROVAL_GRANTED: 'approval_granted',
+    RunEventType.APPROVAL_DENIED: 'approval_denied',
+    RunEventType.RUN_CANCELLED: 'run_cancelled',
+    RunEventType.RUN_PENDING_APPROVAL: 'run_pending_approval',
 }
 
 

tests/lifecycle/test_event_stream_reader.py

@@ -53,14 +53,29 @@ def test_audit_event_to_run_event_type_m0_events() -> None:
 
 
 def test_audit_event_to_run_event_type_legacy_events_return_none() -> None:
-    """Test that unmapped legacy audit event types return None."""
-    # Legacy event types that are not in M0 should return None, not raise
-    assert audit_event_to_run_event_type('tool_call_started') is None
-    assert audit_event_to_run_event_type('model_route') is None
-    assert audit_event_to_run_event_type('git_sandbox_started') is None
+    """Still-unmapped audit event types return None, not raise."""
+    # Genuinely unmapped (planned for later phases, no taxonomy entry yet).
+    assert audit_event_to_run_event_type('plan_resolved') is None
+    assert audit_event_to_run_event_type('context_compacted') is None
+    assert audit_event_to_run_event_type('session_start') is None
     assert audit_event_to_run_event_type('unknown_event_type') is None
 
 
+def test_audit_event_to_run_event_type_evidence_events_now_mapped() -> None:
+    """M2-T002 (§16): evidence events the bundle reads are now typed/surfaced."""
+    assert audit_event_to_run_event_type('tool_call_started') == (
+        RunEventType.TOOL_CALL_STARTED
+    )
+    assert audit_event_to_run_event_type('model_route') == RunEventType.MODEL_ROUTE
+    assert audit_event_to_run_event_type('git_sandbox_started') == (
+        RunEventType.GIT_SANDBOX_STARTED
+    )
+    assert audit_event_to_run_event_type('approval_granted') == (
+        RunEventType.APPROVAL_GRANTED
+    )
+    assert audit_event_to_run_event_type('undo_applied') == RunEventType.UNDO_APPLIED
+
+
 def test_read_run_events_from_audit_mapped_events_only() -> None:
     """Test that read_run_events_from_audit filters to only mapped event types."""
     # Handcrafted list of audit entries: mix of M0 and legacy events
@@ -95,11 +110,18 @@ def test_read_run_events_from_audit_mapped_events_only() -> None:
         },
         {
             'event_id': 'ev-5',
-            'event_type': 'model_route',  # Legacy: not in M0
+            'event_type': 'model_route',  # evidence event — now mapped (§16)
             'run_id': 'run-test',
             'payload': {'resolved_model': 'claude-opus'},
             'created_at': '2026-06-13T10:00:04+00:00',
         },
+        {
+            'event_id': 'ev-5b',
+            'event_type': 'plan_resolved',  # still unmapped legacy → skipped
+            'run_id': 'run-test',
+            'payload': {'plan': 'x'},
+            'created_at': '2026-06-13T10:00:04.5+00:00',
+        },
         {
             'event_id': 'ev-6',
             'event_type': 'run_completed',  # M0 event
@@ -111,26 +133,23 @@ def test_read_run_events_from_audit_mapped_events_only() -> None:
 
     events = read_run_events_from_audit(entries)
 
-    # Only mapped events should appear: run_started, tool_call_requested, tool_call_completed, run_completed
-    # (tool_call_started and model_route are skipped)
-    assert len(events) == 4
-
-    # Check event types in order
-    assert events[0].type == RunEventType.RUN_STARTED
-    assert events[1].type == RunEventType.TOOL_CALL_REQUESTED
-    assert events[2].type == RunEventType.TOOL_CALL_COMPLETED
-    assert events[3].type == RunEventType.RUN_COMPLETED
+    # All taxonomy events surface (M0 + M2 evidence: tool_call_started and
+    # model_route now mapped per §16). Only still-unmapped plan_resolved skips.
+    assert [e.type for e in events] == [
+        RunEventType.RUN_STARTED,
+        RunEventType.TOOL_CALL_STARTED,
+        RunEventType.TOOL_CALL_REQUESTED,
+        RunEventType.TOOL_CALL_COMPLETED,
+        RunEventType.MODEL_ROUTE,
+        RunEventType.RUN_COMPLETED,
+    ]
 
-    # Check sequence numbers are monotonic starting from 1
-    assert events[0].seq == 1
-    assert events[1].seq == 2
-    assert events[2].seq == 3
-    assert events[3].seq == 4
+    # Monotonic seq over surfaced events only, 1..N.
+    assert [e.seq for e in events] == [1, 2, 3, 4, 5, 6]
 
-    # Check run_id and payload preservation
+    # run_id and payload preserved.
     assert all(e.run_id == 'run-test' for e in events)
     assert events[0].payload.get('task') == 'test task'
-    assert events[1].payload.get('call_id') == 'call-1'
 
 
 def test_read_run_events_from_audit_preserves_redaction() -> None:
@@ -184,9 +203,9 @@ def test_read_run_events_from_jsonl_file(tmp_path: Path) -> None:
         },
         {
             'event_id': 'ev-2',
-            'event_type': 'tool_call_started',  # Legacy, will be skipped
+            'event_type': 'plan_resolved',  # still-unmapped legacy → skipped
             'run_id': 'run-jsonl-test',
-            'payload': {'tool_name': 'test_tool'},
+            'payload': {'plan': 'x'},
             'created_at': '2026-06-13T10:00:01+00:00',
         },
         {
@@ -216,8 +235,8 @@ def test_read_run_events_from_jsonl_file(tmp_path: Path) -> None:
 
     events = read_run_events_from_jsonl(jsonl_file)
 
-    # Should read 3 events (run_started, tool_call_requested, run_completed)
-    # tool_call_started is skipped as legacy
+    # Should read 3 events (run_started, tool_call_requested, run_completed);
+    # the still-unmapped plan_resolved is skipped.
     assert len(events) == 3
 
     assert events[0].type == RunEventType.RUN_STARTED