G10 durable runtime observability correlation

coding-agent/.opencode/prompts/tasks/durable-runtime-observability-correlation.md

@@ -0,0 +1,48 @@
+Goal:
+Propagate durable runtime correlation identifiers into observability spans.
+
+Scope:
+- Add safe, low-cardinality runtime correlation attributes to AgentKit pipeline
+  stage and LLM generation spans.
+- Include `tape_id` from the active tape.
+- Forward only whitelisted `AgentRunContext.trace_metadata` correlation keys.
+- Bind HTTP root runs to `turn_id` and `tape_id` trace metadata from
+  `SessionManager`.
+- Keep OTLP/Langfuse exporter privacy filtering intact.
+
+Out of scope:
+- Do not add raw prompt, message, content, tool result, secret, or text values
+  to span attributes.
+- Do not change trace id derivation.
+- Do not change observability configuration defaults.
+- Do not change runtime persistence or replay APIs.
+
+Context:
+- ADRs:
+  - `docs/adr/0029-durable-runtime-identity.md`
+  - `docs/adr/0028-observability-and-langfuse-integration.md`
+- Postmortem patterns consulted:
+  - `postmortem/patterns/PM-0001-address-code-review-issues.md`
+  - `postmortem/patterns/PM-0006-add-usage-event-fields-and-fix-tool-name-kwarg-in-pipeline.md`
+  - `postmortem/patterns/PM-0009-preserve-neutral-bare-anchor-semantics.md`
+  - `postmortem/patterns/PM-0010-route-incremental-context-append-through-tapeview.md`
+- Relevant files:
+  - `src/agentkit/runtime/pipeline.py`
+  - `src/coding_agent/ui/session_manager.py`
+  - `tests/agentkit/runtime/test_pipeline.py`
+  - `tests/ui/test_session_manager_runtime.py`
+  - `tests/coding_agent/test_observability.py`
+
+Target tests:
+- `uv run pytest tests/agentkit/runtime/test_pipeline.py -k "span" -v`
+- `uv run pytest tests/agentkit/runtime/test_pipeline.py -v`
+- `uv run pytest tests/agentkit/observability/test_core.py -v`
+- `uv run pytest tests/ui/test_session_manager_runtime.py -k "run_id or approval or message_snapshot" -v`
+- `uv run pytest tests/coding_agent/test_observability.py -v`
+- `uv run ruff check src/agentkit/runtime/pipeline.py src/coding_agent/ui/session_manager.py tests/agentkit/runtime/test_pipeline.py tests/ui/test_session_manager_runtime.py tests/coding_agent/test_observability.py`
+- `uv run ruff format --check src/agentkit/runtime/pipeline.py src/coding_agent/ui/session_manager.py tests/agentkit/runtime/test_pipeline.py tests/ui/test_session_manager_runtime.py tests/coding_agent/test_observability.py`
+
+Stop conditions:
+- Stop with a blocker if the work requires exporting raw prompt/message/result
+  data or weakening the existing OTLP denylist.
+- Stop with a blocker if trace metadata needs a broader public API decision.

coding-agent/docs/durable_runtime/GOAL_PROGRESS.md

@@ -14,7 +14,7 @@ Last updated: 2026-05-19
 | G07 | Complete | `SessionManager` saves `{run_id}:latest` compacted message snapshots from `ctx.messages` when a runtime store is configured. |
 | G08 | Complete | HTTP replay APIs expose runtime run records, latest message snapshots, and runtime events with `last_event_id` filtering. |
 | G09 | Complete | `SessionManager` persists durable approval interaction records for pending requests, applied decisions, session auto-approvals, and approval timeouts when a runtime store is configured. |
-| G10 | Pending | Not started. |
+| G10 | Complete | AgentKit pipeline spans include safe runtime correlation attributes, and HTTP root runs bind `turn_id`/`tape_id` trace metadata without weakening OTLP privacy filtering. |
 | G11 | Pending | Not started. |
 
 ## G00 Verification
@@ -128,3 +128,24 @@ Last updated: 2026-05-19
   - `uv run pytest tests/approval/test_coordinator.py tests/approval/test_store.py -v`
   - `uv run pytest tests/coding_agent/test_pg_runtime_store.py -k "interaction" -v`
   - `uv run ruff check src/coding_agent/ui/session_manager.py tests/ui/test_session_manager_runtime.py`
+
+## G10 Verification
+
+- Added `.opencode/prompts/tasks/durable-runtime-observability-correlation.md`.
+- Updated `src/agentkit/runtime/pipeline.py`.
+- Updated `src/coding_agent/ui/session_manager.py`.
+- Updated `tests/agentkit/runtime/test_pipeline.py`.
+- Updated `tests/ui/test_session_manager_runtime.py`.
+- Updated `tests/coding_agent/test_observability.py`.
+- Postmortem patterns consulted:
+  - `postmortem/patterns/PM-0001-address-code-review-issues.md`
+  - `postmortem/patterns/PM-0006-add-usage-event-fields-and-fix-tool-name-kwarg-in-pipeline.md`
+  - `postmortem/patterns/PM-0009-preserve-neutral-bare-anchor-semantics.md`
+  - `postmortem/patterns/PM-0010-route-incremental-context-append-through-tapeview.md`
+- Target tests:
+  - `uv run pytest tests/agentkit/runtime/test_pipeline.py -k "span" -v`
+  - `uv run pytest tests/agentkit/runtime/test_pipeline.py -v`
+  - `uv run pytest tests/agentkit/observability/test_core.py -v`
+  - `uv run pytest tests/ui/test_session_manager_runtime.py -k "run_id or approval or message_snapshot" -v`
+  - `uv run pytest tests/coding_agent/test_observability.py -v`
+  - `uv run ruff check src/agentkit/runtime/pipeline.py src/coding_agent/ui/session_manager.py tests/agentkit/runtime/test_pipeline.py tests/ui/test_session_manager_runtime.py tests/coding_agent/test_observability.py`

coding-agent/src/agentkit/runtime/pipeline.py

@@ -68,6 +68,16 @@
     "render": "runtime.stage.apply_directives",
     "dispatch": "runtime.stage.dispatch",
 }
+_TRACE_METADATA_ATTRIBUTE_KEYS = frozenset(
+    {
+        "turn_id",
+        "tape_id",
+        "tool_call_id",
+        "interaction_id",
+        "event_id",
+        "checkpoint_id",
+    }
+)
 
 StructuredToolResultScopeFactory = Callable[[bool], AbstractContextManager[None]]
 
@@ -108,8 +118,31 @@ def _stage_span_attributes(
     }
     if ctx.session_id:
         attributes["session_id"] = ctx.session_id
+    attributes.update(_runtime_correlation_attributes(ctx))
+    return attributes
+
+
+def _safe_trace_metadata_attributes(run_context: AgentRunContext) -> dict[str, Any]:
+    attributes: dict[str, Any] = {}
+    for key, value in run_context.trace_metadata.items():
+        if key not in _TRACE_METADATA_ATTRIBUTE_KEYS:
+            continue
+        if isinstance(value, bool | int | float | str):
+            attributes[key] = value
+    return attributes
+
+
+def _runtime_correlation_attributes(ctx: "PipelineContext") -> dict[str, Any]:
+    attributes: dict[str, Any] = {}
     if ctx.run_context is not None:
+        attributes.update(_safe_trace_metadata_attributes(ctx.run_context))
         attributes["run_id"] = ctx.run_context.run_id
+        if ctx.run_context.agent_id is not None:
+            attributes["agent_id"] = ctx.run_context.agent_id
+        if ctx.run_context.parent_run_id is not None:
+            attributes["parent_run_id"] = ctx.run_context.parent_run_id
+    if ctx.tape.tape_id:
+        attributes["tape_id"] = ctx.tape.tape_id
     return attributes
 
 
@@ -125,8 +158,7 @@ def _llm_generation_attributes(
     }
     if ctx.session_id:
         attributes["session_id"] = ctx.session_id
-    if ctx.run_context is not None:
-        attributes["run_id"] = ctx.run_context.run_id
+    attributes.update(_runtime_correlation_attributes(ctx))
     model = ctx.config.get("model")
     if isinstance(model, str) and model:
         attributes["model"] = model

coding-agent/src/coding_agent/ui/session_manager.py

@@ -1073,11 +1073,15 @@ def _bind_root_run_identity(self, session: Session, ctx: Any, run_id: str) -> No
         if run_context is not None:
             if not isinstance(run_context, AgentRunContext):
                 raise TypeError("runtime context run_context must be AgentRunContext")
+            trace_metadata = dict(run_context.trace_metadata)
+            trace_metadata["turn_id"] = run_id
+            trace_metadata["tape_id"] = ctx.tape.tape_id
             ctx.run_context = replace(
                 run_context,
                 session_id=session.id,
                 run_id=run_id,
                 parent_run_id=None,
+                trace_metadata=trace_metadata,
             )
 
     def _run_metadata_for_session(self, session: Session) -> JSONObject:

coding-agent/tests/agentkit/runtime/test_pipeline.py

@@ -376,14 +376,19 @@ async def mock_stream(messages, tools=None, **kwargs):
 
         plugin._mock_llm.stream = mock_stream
         ctx = PipelineContext(
-            tape=Tape(),
+            tape=Tape(tape_id="tape-1"),
             session_id="session-1",
             run_context=AgentRunContext(
                 session_id="session-1",
-                run_id="turn-1",
-                agent_id=None,
+                run_id="run-1",
+                agent_id="root",
                 environment=RuntimeContextEnvironment(),
                 context_budget=ContextBudget(),
+                trace_metadata={
+                    "turn_id": "run-1",
+                    "interaction_id": "interaction-1",
+                    "prompt": "do-not-record",
+                },
             ),
             config={"observation_sink": sink},
         )
@@ -407,10 +412,15 @@ async def mock_stream(messages, tools=None, **kwargs):
         assert build_context_span.attributes == {
             "stage": "build_context",
             "session_id": "session-1",
-            "run_id": "turn-1",
+            "run_id": "run-1",
+            "agent_id": "root",
+            "tape_id": "tape-1",
+            "turn_id": "run-1",
+            "interaction_id": "interaction-1",
             "entry_count_before": 0,
             "entry_count_after": 0,
         }
+        assert "prompt" not in build_context_span.attributes
         model_span = stage_spans[3]
         assert model_span.attributes["entry_count_before"] == 0
         assert model_span.attributes["entry_count_after"] == 1
@@ -436,14 +446,15 @@ async def mock_stream(messages, tools=None, **kwargs):
 
         plugin._mock_llm.stream = mock_stream
         ctx = PipelineContext(
-            tape=Tape(),
+            tape=Tape(tape_id="tape-1"),
             session_id="session-1",
             run_context=AgentRunContext(
                 session_id="session-1",
-                run_id="turn-1",
+                run_id="run-1",
                 agent_id=None,
                 environment=RuntimeContextEnvironment(),
                 context_budget=ContextBudget(),
+                trace_metadata={"turn_id": "run-1"},
             ),
             config={
                 "model": "test-model",
@@ -458,7 +469,9 @@ async def mock_stream(messages, tools=None, **kwargs):
         assert llm_spans[0].status == "ok"
         assert llm_spans[0].attributes == {
             "session_id": "session-1",
-            "run_id": "turn-1",
+            "run_id": "run-1",
+            "tape_id": "tape-1",
+            "turn_id": "run-1",
             "message_count": 2,
             "tool_schema_count": 0,
             "model": "test-model",

coding-agent/tests/coding_agent/test_observability.py

@@ -71,7 +71,12 @@ def test_otlp_sink_posts_span_without_prompt_or_output_content() -> None:
             attributes={
                 "session_id": "session-1",
                 "input_tokens": 10,
+                "content": "content-do-not-send",
+                "message": "message-do-not-send",
                 "prompt": "do-not-send",
+                "result": "result-do-not-send",
+                "secret": "secret-do-not-send",
+                "text": "text-do-not-send",
             },
             start_time=1.0,
             end_time=2.0,
@@ -86,6 +91,20 @@ def test_otlp_sink_posts_span_without_prompt_or_output_content() -> None:
     assert "llm.generation" in body
     assert "input_tokens" in body
     assert "do-not-send" not in body
+    assert "content-do-not-send" not in body
+    assert "message-do-not-send" not in body
+    assert "result-do-not-send" not in body
+    assert "secret-do-not-send" not in body
+    assert "text-do-not-send" not in body
+    span = _first_span(json.loads(body))
+    exported_keys = {
+        item["key"]
+        for item in span.get("attributes", [])
+        if isinstance(item, dict) and isinstance(item.get("key"), str)
+    }
+    assert {"content", "message", "prompt", "result", "secret", "text"}.isdisjoint(
+        exported_keys
+    )
 
 
 def test_otlp_sink_groups_spans_by_session_and_run() -> None:

coding-agent/tests/ui/test_session_manager_runtime.py

@@ -249,6 +249,10 @@ def fake_create_agent(**kwargs):
     assert session.runtime_ctx.run_context.session_id == session_id
     assert session.runtime_ctx.run_context.run_id == run_id
     assert session.runtime_ctx.run_context.parent_run_id is None
+    assert session.runtime_ctx.run_context.trace_metadata == {
+        "turn_id": run_id,
+        "tape_id": session.runtime_ctx.tape.tape_id,
+    }
 
 
 @pytest.mark.asyncio