refactor(ai,python): author voRequest/voResponse columns; remove derive_trace_fields + pre_freeze hook

Trace entities now AUTHOR their typed voRequest/voResponse jsonb columns
explicitly (field.object + @objectref + @storage:jsonb) rather than having
the loader DERIVE them by injecting synthetic field.object nodes at load
time. This removes the only load-time metadata mutation in the Python
AI-trace feature.

- Delete loader/derive_trace_fields.py + its unit test.
- Remove the generic pre_freeze loader hook from MetaDataLoader (__init__
  param/assignment/comment, the load() invocation, the from_directory
  passthrough, and the now-unused Callable import).
- cli._load_root reverts to plain MetaDataLoader.from_directory(metadata_dir).
- trace_helper_generator is unchanged (already defines a local LLM_CALL_BASE;
  never imported from the deriver) — it emits the typed columns off the
  prompt's @responseRef/@payloadRef, no derivation needed.
- Integration fixture restores the Slice-1 authored voResponse shape; the
  round-trip test loads with plain MetaDataLoader(strict=True) and drops the
  no-longer-authored voRequest DDL column.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: dmealing

server/python/src/metaobjects/cli.py

@@ -41,7 +41,6 @@
 from pathlib import Path
 
 from metaobjects import MetaDataLoader
-from metaobjects.loader.derive_trace_fields import derive_trace_fields
 from metaobjects.agent_context import (
     AGENT_CONTEXT_MANIFEST_PATH,
     Manifest,
@@ -107,10 +106,7 @@ def _default_generators() -> list[Generator]:
 
 def _load_root(metadata_dir: str) -> tuple[MetaData | None, list[str]]:
     """Load metadata; return ``(root, error_messages)``. ``root`` is None on error."""
-    # AI-trace pre-pass: derive typed voRequest/voResponse jsonb columns onto
-    # LlmCallBase-derived entities so the trace-helper generator (and gen/verify)
-    # see them without the author restating them. No-op when no trace entities.
-    result = MetaDataLoader.from_directory(metadata_dir, pre_freeze=derive_trace_fields)
+    result = MetaDataLoader.from_directory(metadata_dir)
     if result.errors:
         msgs = [f"{e.code}: {e.message}" for e in result.errors]
         return None, msgs

server/python/src/metaobjects/loader/derive_trace_fields.py

@@ -15,7 +15,6 @@
 from __future__ import annotations
 
 import json
-from collections.abc import Callable
 from dataclasses import dataclass, field
 from pathlib import Path
 
@@ -83,17 +82,10 @@ def __init__(
         self,
         providers: list[Provider] | None = None,
         strict: bool = False,
-        pre_freeze: Callable[[MetaData], None] | None = None,
     ) -> None:
         self._registry: TypeRegistry = compose_registry(
             providers if providers is not None else list(core_providers)
         )
-        # Optional pre-freeze enrichment hook (mirrors the TS loader's `preFreeze`
-        # option and the Java MetaDataLoader.setPreFreeze). Runs inside `load`
-        # after extends-resolution and before the validation passes, so any nodes
-        # it injects (e.g. AI-trace voRequest/voResponse jsonb columns) are
-        # validated like authored ones and reach BOTH codegen and the runtime.
-        self._pre_freeze = pre_freeze
         # ADR-0023 — strict load closes the open-attr policy: an authored own
         # @-attr matching no per-type schema and no commonAttr → ERR_UNKNOWN_ATTR
         # (alongside Python's always-on unknown TYPE/SUBTYPE rejection). Defaults
@@ -140,12 +132,6 @@ def load(self, sources: list[MetaDataSource]) -> LoadResult:
             )
             resolve_supers(result.root, result.errors)
 
-        # Pre-freeze enrichment hook (mirrors TS/Java): runs after extends-
-        # resolution and before validation, so DERIVED nodes (e.g. AI-trace
-        # voRequest/voResponse jsonb columns) are validated + reach codegen+runtime.
-        if self._pre_freeze is not None:
-            self._pre_freeze(result.root)
-
         run_validations(
             result.root,
             self._registry,
@@ -167,15 +153,13 @@ def from_directory(
         exclude: list[str] | None = None,
         recurse: bool = True,
         strict: bool = False,
-        pre_freeze: Callable[[MetaData], None] | None = None,
     ) -> LoadResult:
         """Load every JSON/YAML file under ``directory`` (recursive by default).
 
         ``strict`` (ADR-0023) — when True, an undeclared own ``@-attr`` →
         ``ERR_UNKNOWN_ATTR``. Defaults False (downstream-friendly open policy).
-        ``pre_freeze`` — optional in-load tree-enrichment hook (see ``__init__``).
         """
-        loader = cls(providers=providers, strict=strict, pre_freeze=pre_freeze)
+        loader = cls(providers=providers, strict=strict)
         sources = list(
             DirectorySource(directory, exclude=exclude, recurse=recurse).expand()
         )

server/python/src/metaobjects/loader/meta_data_loader.py

@@ -1,35 +1,26 @@
-# Test-only trace entity for the AI-trace recorder PG round-trip (Python).
-# Loaded alongside the shipped library/ai/llm-call.yaml (LlmCallBase). The typed
-# voRequest/voResponse jsonb columns are NOT declared here — they are DERIVED by
-# derive_trace_fields (Slice 3) from the nested template.prompt's
-# @payloadRef/@responseRef when the loader runs with the pre_freeze hook. This
-# proves derivation reaches the metadata-driven ObjectManager runtime.
+# Test-only trace entity for the AI-trace recorder PG round-trip (Python, Slice 1).
+# Loaded alongside the shipped library/ai/llm-call.yaml (LlmCallBase). Declares a
+# typed voResponse jsonb column by hand (Slice 3 will switch this to derivation
+# from a nested template.prompt).
 metadata:
   package: metaobjects::ai
   children:
-    # Typed request payload VO (referenced by @payloadRef → derived voRequest).
-    - object.value:
-        name: GreetingRequest
-        children:
-          - field.string: { name: prompt }
-
-    # Typed response VO (referenced by @responseRef → derived voResponse).
+    # Typed value object stored in the voResponse jsonb column.
     - object.value:
         name: GreetingResponse
         children:
           - field.string: { name: greeting }
           - field.int:    { name: score }
 
     # Concrete trace entity: inherits the 18 LlmCallBase fields, declares its
-    # table + primary identity + a template.prompt. The deriver injects
-    # voRequest/voResponse field.object jsonb columns from the prompt refs.
+    # table + primary identity + a typed voResponse jsonb column.
     - object.entity:
         name: GreetingCall
         extends: metaobjects::ai::LlmCallBase
         children:
           - source.rdb:       { table: llm_call, role: primary }
           - identity.primary: { name: primary, fields: ["spanId"] }
-          - template.prompt:
-              name: greet
-              payloadRef: metaobjects::ai::GreetingRequest
-              responseRef: metaobjects::ai::GreetingResponse
+          - field.object:
+              name: voResponse
+              storage: jsonb
+              objectRef: metaobjects::ai::GreetingResponse

server/python/tests/integration/meta_ai_trace.yaml

@@ -17,7 +17,6 @@
 from pathlib import Path
 
 from metaobjects import FileSource, MetaDataLoader
-from metaobjects.loader.derive_trace_fields import derive_trace_fields
 from metaobjects.runtime import (
     LlmCallInput,
     ObjectManager,
@@ -50,7 +49,6 @@
   "startedAt" timestamp,
   "llmRequest" jsonb,
   "llmResponse" jsonb,
-  "voRequest" jsonb,
   "voResponse" jsonb
 )
 """
@@ -73,10 +71,10 @@ def _connect(info):
 
 
 def test_typed_trace_round_trips_through_postgres() -> None:
-    # voRequest/voResponse are NOT authored in the fixture — they are DERIVED
-    # in-load from the prompt's @payloadRef/@responseRef via the pre_freeze hook,
-    # proving derivation reaches the metadata-driven ObjectManager runtime.
-    loader = MetaDataLoader(strict=True, pre_freeze=derive_trace_fields)
+    # voResponse is AUTHORED in the fixture (field.object + @objectRef + jsonb).
+    # No derivation: this proves the recorder persists the authored typed column
+    # through the metadata-driven ObjectManager runtime.
+    loader = MetaDataLoader(strict=True)
     result = loader.load([
         FileSource(_repo_file("library/ai/llm-call.yaml")),
         FileSource(Path(__file__).parent / "meta_ai_trace.yaml"),