Move docstring spec-refs into # comments

src/openarmature/graph/events.py

@@ -709,7 +709,7 @@ class LlmRetryAttemptEvent:
       ``finish_reason`` / ``output_content`` / ``output_tool_calls``):
       populated on a successful attempt; ``None`` / empty list on a
       failed attempt. ``output_tool_calls`` is the source the OTel
-      observer renders the §5.5.1 / §5.5.10 output tool-call attributes
+      observer renders the output tool-call attributes
       from (this is the per-attempt event that drives the LLM span).
     - failure side (``error_category`` / ``error_message`` /
       ``error_type``): populated on a failed attempt; ``None`` on a
@@ -745,8 +745,9 @@ class LlmRetryAttemptEvent:
     # Proposal 0076: the attempt's output tool calls (ToolCall records),
     # mirroring LlmCompletionEvent.output_tool_calls. Populated on a
     # successful attempt; empty list on a failed one (no response). The
-    # OTel observer renders the output tool-call span attributes from
-    # this field (the per-attempt event is the LLM-span source).
+    # OTel observer renders the observability §5.5.1 / §5.5.10 output
+    # tool-call span attributes from this field (the per-attempt event is
+    # the LLM-span source).
     output_tool_calls: list["ToolCall"] = field(default_factory=list["ToolCall"])
 
 

src/openarmature/graph/parallel_branches.py

@@ -139,10 +139,12 @@ async def run(self, state: ParentT) -> Mapping[str, Any]:
             "compiled.invoke)."
         )
 
+    # Spec pipeline-utilities §11.10: a branch dispatches when its ``when``
+    # predicate is absent or returns true.
     def dispatched_branches(self, state: Any) -> list[tuple[str, BranchSpec[Any]]]:
         """Return the branches that dispatch for ``state``, in insertion
         order: every declared branch whose ``when`` predicate is absent or
-        returns true (§11.10).
+        returns true.
 
         ``when`` MUST be a pure function of the dispatch-time parent state,
         so this is deterministic and safe to evaluate both here (the
@@ -288,6 +290,11 @@ async def innermost(s: Any) -> Mapping[str, Any]:
             return await self._fail_fast(state, tasks, contributions)
         return await self._collect(state, tasks, contributions, errors)
 
+    # Spec pipeline-utilities §11.4 (the buffer-then-project subgraph path is
+    # bypassed for inline callables) + §11.7 (per-leg branch middleware). The
+    # callable emits its own graph-engine §6 started/completed pair keyed by
+    # ``branch_name``, rendered per observability §5.7 as a per-branch dispatch
+    # span with no inner-node span beneath it.
     async def _dispatch_callable_branch(
         self,
         branch_name: str,
@@ -300,14 +307,14 @@ async def _dispatch_callable_branch(
 
         The callable reads the parent state and returns a parent-shaped
         partial update directly — no subgraph, no ``inputs`` / ``outputs``
-        projection (§11.4). Branch ``middleware`` (e.g. a per-leg
-        ``FailureIsolationMiddleware``, §11.7) wraps the callable.
+        projection. Branch ``middleware`` (e.g. a per-leg
+        ``FailureIsolationMiddleware``) wraps the callable.
 
         A callable branch has no inner nodes, so it IS the unit of work:
         it emits one ``started`` / ``completed`` observer pair keyed by its
-        ``branch_name`` (graph-engine §6), which the observers render as the
-        branch's per-branch dispatch span with NO inner-node span beneath it
-        (observability §5.7). To render that way the pair is emitted at the
+        ``branch_name``, which the observers render as the
+        branch's per-branch dispatch span with NO inner-node span beneath it.
+        To render that way the pair is emitted at the
         parallel-branches NODE's own namespace (not a descended branch
         namespace), tagged with ``branch_name`` (set on the ContextVar in
         ``run_branch``): an event at a pb-node's own namespace carrying a

src/openarmature/observability/langfuse/client.py

@@ -185,11 +185,13 @@ def trace(
         Implementations track Traces internally; observation calls
         pass `trace_id` to associate.
 
-        `session_id` / `user_id` (proposal 0064 §8.4.1) populate
+        `session_id` / `user_id` populate
         Langfuse's cross-trace grouping fields (the Sessions / Users
         dashboards); each is unset when its source is absent.
         """
         # Spec §8.4.1: the Trace id is the OA invocation_id verbatim.
+        # session_id / user_id (proposal 0064) populate Langfuse's
+        # cross-trace grouping (the Sessions / Users dashboards).
         ...
 
     # The current observer doesn't invoke this method — it sets the
@@ -248,6 +250,8 @@ def generation(
         start_time: datetime | None = None,
     ) -> LangfuseGenerationHandle: ...
 
+    # Spec proposal 0063: dedicated Tool observation (Langfuse asType="tool"),
+    # not a Generation with a discriminator.
     def tool(
         self,
         *,
@@ -261,7 +265,7 @@ def tool(
         output: Any = None,
         start_time: datetime | None = None,
     ) -> LangfuseSpanHandle:
-        """Open a dedicated Tool observation (proposal 0063, Langfuse
+        """Open a dedicated Tool observation (Langfuse
         ``asType="tool"``). Like :meth:`generation` minus the model /
         usage / prompt surface: a Tool carries ``input`` (arguments) /
         ``output`` (result) / metadata / level. Returns a minimal

src/openarmature/observability/langfuse/observer.py

@@ -1821,9 +1821,10 @@ def _handle_typed_llm_failed(self, event: LlmFailedEvent) -> None:
             status_message=event.error_category,
         )
 
+    # Spec proposal 0063: dedicated Langfuse Tool observation (asType="tool").
     def _handle_tool_call(self, event: ToolCallEvent | ToolCallFailedEvent) -> None:
         """Open + close a dedicated Tool observation (Langfuse
-        ``asType="tool"``, proposal 0063) under the calling node's Span
+        ``asType="tool"``) under the calling node's Span
         observation. DEFAULT level on a ToolCallEvent; ERROR (with
         ``error_type`` / ``error_message`` in metadata and as the status
         message) on a ToolCallFailedEvent. ``input`` (arguments) /

src/openarmature/observability/llm_event.py

@@ -147,9 +147,9 @@ class LlmEventPayload(BaseModel):
     caller_invocation_metadata: dict[str, Any] = Field(default_factory=dict)
 
 
+# Spec observability §5.5.5: tool-call serialization.
 def serialize_tool_calls(tool_calls: Sequence[ToolCall]) -> list[dict[str, Any]]:
-    """The observability §5.5.5 tool-call serialization,
-    ``[{id, name, arguments}, ...]``.
+    """Serialize tool calls to ``[{id, name, arguments}, ...]``.
 
     The single home for the encoding, shared by the input-message
     payload (the provider's ``input.messages`` serialization, where the

src/openarmature/observability/otel/observer.py

@@ -359,7 +359,7 @@ def _serialize_for_attribute(value: Any) -> str:
     renders via its ``str()`` rather than raising inside the observer
     (which would lose the whole span). It is a no-op for the
     already-encodable payloads (messages, params); it matters for the
-    proposal 0063 tool ``result``, which is an opaque, language-idiomatic
+    tool ``result``, which is an opaque, language-idiomatic
     value the tool produced (a model, dataclass, datetime, ...)."""
     return json.dumps(value, sort_keys=True, separators=(",", ":"), ensure_ascii=False, default=str)
 
@@ -1389,17 +1389,19 @@ def _emit_checkpoint_save_span(self, event: NodeEvent) -> None:
     # active_prompt_group snapshots taken at dispatch time — NOT the
     # ContextVar. The dispatch worker's task-local Context doesn't see
     # node-body ContextVar writes.
+    # Spec observability §11 GenAI metrics. §11.2: the attempt index is
+    # deliberately not a metric dimension (cardinality control).
     def _record_llm_metrics(self, event: LlmRetryAttemptEvent) -> None:
-        """Record the §11 GenAI metric observations for one LLM-call
+        """Record the GenAI metric observations for one LLM-call
         attempt.
 
         Duration is recorded for every attempt (including a failed one,
         carrying ``error.type``); token usage only when the attempt
         returned a usage record (a failed attempt has none). Sourced from
         the per-attempt ``LlmRetryAttemptEvent`` — one duration sample per
         attempt under call-level retry, matching the per-attempt span
-        model. The attempt index is deliberately not a dimension (§11.2
-        cardinality).
+        model. The attempt index is deliberately not a dimension
+        (cardinality control).
         """
         if self._duration_histogram is None or self._token_histogram is None:
             return
@@ -1615,16 +1617,18 @@ def _handle_typed_llm_retry_attempt(self, event: LlmRetryAttemptEvent) -> None:
         self._run_enrichers(span, event)
         span.end(end_time=end_time_ns)
 
+    # Spec proposal 0063 tool span. Payload attributes (arguments / result)
+    # follow observability §5.5.4 opt-out + §5.5.5 truncation.
     def _handle_tool_call(self, event: ToolCallEvent | ToolCallFailedEvent) -> None:
-        """Emit an ``openarmature.tool.call`` span for a tool execution
-        (proposal 0063), parented under the calling node.
+        """Emit an ``openarmature.tool.call`` span for a tool execution,
+        parented under the calling node.
 
         A ``ToolCallEvent`` renders OK with the result attribute; a
         ``ToolCallFailedEvent`` renders ERROR with the standard OTel
         ``error.type`` attribute + an exception event carrying the
         message. ``arguments`` / ``result`` are payload, gated by
-        ``disable_provider_payload`` (§5.5.4) + the §5.5.5 truncation
-        contract. The OA-namespace ``openarmature.tool.*`` attributes
+        ``disable_provider_payload`` and truncated. The OA-namespace
+        ``openarmature.tool.*`` attributes
         mirror the Development ``gen_ai.tool.*`` surface, which is NOT
         emitted in v1. ``disable_llm_spans`` does not gate this span.
         """

tests/conformance/adapter.py

@@ -1063,11 +1063,12 @@ def _add_fan_out_node(
     )
 
 
+# Spec proposal 0075: callable branch ``call:`` directive.
 def _build_call_fn(
     branch_name: str,
     call_cfg: Mapping[str, Any],
 ) -> Callable[[Any], Awaitable[Mapping[str, Any]]]:
-    """Translate a callable branch's ``call:`` directive (proposal 0075)
+    """Translate a callable branch's ``call:`` directive
     into an async function over the parent state.
 
     Reuses the node-behavior factories (``update`` / ``flaky`` /
@@ -1087,8 +1088,9 @@ def _build_call_fn(
     raise ValueError(f"callable branch {branch_name!r}: unsupported call directive {dict(call_cfg)!r}")
 
 
+# Spec proposal 0075 §11.10: branch ``when:`` predicate.
 def _build_when_predicate(when_cfg: Mapping[str, Any]) -> Callable[[Any], bool]:
-    """Translate a branch ``when:`` directive (proposal 0075 §11.10) into
+    """Translate a branch ``when:`` directive into
     a parent-state predicate. Supports ``{field: <name>}`` — a truthy
     check on a parent-state field at dispatch time."""
     if "field" in when_cfg:
@@ -1115,7 +1117,7 @@ def _add_parallel_branches_node(
 
     A branch is either a ``subgraph`` (name resolved against the shared
     ``subgraphs`` registry, with optional ``inputs`` / ``outputs``) or an
-    inline ``call`` (proposal 0075), and may carry a ``when`` predicate.
+    inline ``call``, and may carry a ``when`` predicate.
     ``branch_middleware`` maps branch-name to a pre-translated middleware
     list; the test driver populates it from each branch's ``middleware:``
     block.

tests/conformance/harness/directives.py

@@ -305,13 +305,14 @@ def _normalize_instance_middleware(cls, data: Any) -> Any:
         return data_dict
 
 
+# Spec proposal 0075: parallel-branches branch entry.
 class ParallelBranchSpec(_AllowExtras):
     """One entry inside a ``parallel_branches.branches`` mapping.
 
     A branch's work is given by exactly one of ``subgraph`` (a compiled
     subgraph referenced by name, with optional ``inputs`` / ``outputs``)
     or ``call`` (an inline node-behavior directive — ``update`` / ``flaky``
-    / ``raises`` — run as a function over the parent state, proposal 0075).
+    / ``raises`` — run as a function over the parent state).
     An optional ``when`` directive (``{field: <name>}``) skips the branch
     at dispatch. Permissive on extras because fixtures may carry extra
     knobs (e.g., per-branch annotations the harness ignores).
@@ -385,8 +386,9 @@ class MockToolSpec(_AllowExtras):
     raises: dict[str, Any] | None = None
 
 
+# Spec proposal 0063: tool-using node directive.
 class CallsToolSpec(_AllowExtras):
-    """Tool-using node (proposal 0063): the node body enters the
+    """Tool-using node: the node body enters the
     ``with_tool_call`` instrumentation scope, runs the ``mock_tool``, and
     OA emits the ToolCallEvent / ToolCallFailedEvent. ``tool_call_id`` is
     optional (null = a standalone instrumented function);
@@ -400,11 +402,12 @@ class CallsToolSpec(_AllowExtras):
     stores_result_in: str | None = None
 
 
+# Spec proposal 0059: embedding-using node directive.
 class CallsEmbedSpec(_AllowExtras):
     """Embedding-using node: sends ``input`` to a mock embedding provider
     and stores the result in ``stores_response_in``. Used by the
     GenAI-metrics embedding fixture (089). The embedding capability
-    (proposal 0059) is unimplemented in python until v0.16.0, so 089 is
+    is unimplemented in python until v0.16.0, so 089 is
     deferred at the runner; this directive is modelled here only so the
     fixture parses + round-trips through the harness schema."""
 

tests/integration/test_langfuse_sdk_adapter.py

@@ -249,16 +249,19 @@ async def test_sdk_adapter_generation_timestamps_round_trip_through_langfuse() -
         client.shutdown()
 
 
+# Spec proposal 0064 §8.4.1: trace(session_id=, user_id=) populates the live
+# Trace's sessionId / userId grouping. proposal 0020 (sessions capability) will
+# later supply openarmature.session_id along the same passthrough.
 @pytest.mark.integration
 async def test_sdk_adapter_populates_session_and_user_id_on_live_langfuse() -> None:
-    """End-to-end (proposal 0064 §8.4.1): trace(session_id=, user_id=)
+    """End-to-end: trace(session_id=, user_id=)
     populates the live Trace's sessionId / userId grouping fields.
 
     The observer leaves session_id dormant until the sessions capability
-    (proposal 0020) supplies openarmature.session_id, but the adapter
+    supplies openarmature.session_id, but the adapter
     passes whatever it is given, so this exercises BOTH passthroughs at
-    the SDK boundary: when 0020 lands, the session_id rides the same
-    propagate_attributes path the userId promotion uses today.
+    the SDK boundary: when the sessions capability lands, the session_id
+    rides the same propagate_attributes path the userId promotion uses today.
     """
     from langfuse import Langfuse
 

tests/unit/test_tool_call.py

@@ -1,4 +1,5 @@
-"""Unit tests for the tool-call instrumentation scope (proposal 0063).
+# Spec proposal 0063: tool-call instrumentation scope.
+"""Unit tests for the tool-call instrumentation scope.
 
 Exercise ``with_tool_call`` directly by installing a collecting dispatch
 callback + a node-scope identity into the correlation ContextVars, the