Enforce bounded operation guardrails

Assistant: codex
Assistant-Model: gpt-5.6-sol
Assistant-Session: 01a028de-e2c8-7732-8521-46a7fc5db82f

src/activity_core/rules/executor.py

@@ -38,12 +38,20 @@ class InstructionResult:
     prompt_hash: str | None = None
     model: str | None = None
     output_validated: bool = False
-    review_required: bool = False
+    review_advisory: bool = False
     condition_matched: str | None = None
     validation_error: str | None = None
     llm_response_metadata: dict[str, Any] | None = None
 
 
+def _review_advisory(instr: Any) -> bool:
+    """Read the renamed advisory flag while accepting legacy caller objects."""
+    legacy = getattr(instr, "review_required", None)
+    if legacy is not None:
+        return bool(legacy)
+    return bool(getattr(instr, "review_advisory", False))
+
+
 def _resolve_path(obj: Any, path: str) -> Any:
     """Walk a dot-separated path on obj or dict. Returns None if not found."""
     parts = path.split(".")
@@ -134,7 +142,7 @@ def execute_instruction_with_audit(
                 prompt_hash=None,
                 model=getattr(instr, "model", None),
                 output_validated=False,
-                review_required=True,
+                review_advisory=True,
                 condition_matched=getattr(instr, "condition", "") or None,
                 validation_error=str(exc),
             )
@@ -149,7 +157,7 @@ def execute_instruction_with_audit(
                 prompt_hash=None,
                 model=getattr(instr, "model", None),
                 output_validated=False,
-                review_required=True,
+                review_advisory=True,
                 condition_matched=getattr(instr, "condition", "") or None,
                 validation_error=str(exc),
             )
@@ -200,13 +208,10 @@ def _execute(
         response_metadata = _llm_response_metadata(llm_client)
         task_specs, report, error = _validate_output(raw_output, instr, allow_list)
         if error:
-            # Truncate to keep log volume bounded but long enough to see the
-            # actual JSON shape mismatch (typical reports are <2KB).
-            preview = (raw_output or "")[:2000]
             logger.warning(
                 "instruction_output_error: instruction=%r, prompt_hash=%s, "
-                "error=%s, raw_output_preview=%r",
-                instr.id, prompt_hash, error, preview,
+                "error=%s",
+                instr.id, prompt_hash, error,
             )
             # Posture B (WP-0016-T03): try to recover a partial-but-usable
             # report from individually-parseable items before declaring total
@@ -227,7 +232,7 @@ def _execute(
                     prompt_hash=prompt_hash,
                     model=instr.model,
                     output_validated=False,
-                    review_required=True,
+                    review_advisory=True,
                     condition_matched=instr.condition or None,
                     validation_error=error,
                     llm_response_metadata=response_metadata,
@@ -240,7 +245,7 @@ def _execute(
         prompt_hash=prompt_hash,
         model=instr.model,
         output_validated=True,
-        review_required=bool(getattr(instr, "review_required", False)),
+        review_advisory=_review_advisory(instr),
         condition_matched=instr.condition or None,
         llm_response_metadata=response_metadata,
     )
@@ -272,7 +277,7 @@ def _empty_result(
         prompt_hash=prompt_hash,
         model=getattr(instr, "model", None),
         output_validated=False,
-        review_required=bool(getattr(instr, "review_required", False)),
+        review_advisory=_review_advisory(instr),
         condition_matched=getattr(instr, "condition", "") or None,
         validation_error=validation_error,
     )
@@ -281,47 +286,28 @@ def _empty_result(
 def _invalid_output_report(
     instr: Any,
     validation_error: str,
-    raw_output: Any,
+    _raw_output: Any,
     response_metadata: dict[str, Any] | None = None,
 ) -> dict[str, Any] | None:
     """Build a durable diagnostic report for invalid report-sink output.
 
     Task-only instructions keep the legacy empty-result behavior. Instructions
-    with report sinks should leave operators a bounded artifact that preserves
-    the partial model output without marking it as schema-valid.
+    with report sinks leave a bounded diagnostic without retaining unvalidated
+    provider output.
     """
     if not getattr(instr, "report_sinks", None):
         return None
 
-    partial_output: Any
-    raw_preview: str | None = None
-    if isinstance(raw_output, str):
-        try:
-            partial_output = _parse_json_output(raw_output)
-        except json.JSONDecodeError:
-            partial_output = None
-            raw_preview = raw_output[:_RAW_OUTPUT_PREVIEW_LIMIT]
-    else:
-        partial_output = raw_output
-
     report: dict[str, Any] = {
         "summary": (
             f"Instruction {instr.id} produced output that failed validation; "
-            "partial output was preserved for operator review."
+            "unvalidated provider output was discarded."
         ),
         "status": "validation_failed",
         "validation_error": validation_error,
     }
     if response_metadata:
         report["llm_response_metadata"] = response_metadata
-    if isinstance(partial_output, dict):
-        if isinstance(partial_output.get("summary"), str):
-            report["partial_summary"] = partial_output["summary"]
-        report["partial_report"] = partial_output
-    elif isinstance(partial_output, list):
-        report["partial_report"] = partial_output
-    elif raw_preview is not None:
-        report["raw_output_preview"] = raw_preview
     return report
 
 
@@ -343,7 +329,6 @@ _SNIPPET_LIMIT = 200
 # fail the whole report or flow unbounded into a downstream consumer.
 _MAX_STRING_LEN = 4000
 _MAX_DEPTH = 8
-_RAW_OUTPUT_PREVIEW_LIMIT = 12000
 _SUMMARY_RE = re.compile(r'"summary"\s*:\s*"((?:[^"\\]|\\.)*)"')
 
 
@@ -670,7 +655,7 @@ def _resilient_report(
         prompt_hash=prompt_hash,
         model=getattr(instr, "model", None),
         output_validated=True,
-        review_required=True,
+        review_advisory=True,
         condition_matched=getattr(instr, "condition", "") or None,
         validation_error=None,
         llm_response_metadata=response_metadata,
@@ -745,7 +730,7 @@ def _deterministic_context_report(instr: Any, context: dict) -> InstructionResul
         prompt_hash=None,
         model=getattr(instr, "model", None),
         output_validated=True,
-        review_required=bool(getattr(instr, "review_required", False)),
+        review_advisory=_review_advisory(instr),
         condition_matched=getattr(instr, "condition", "") or None,
     )
 
@@ -806,7 +791,7 @@ def _sbom_catchup_report(instr: Any, catchup: dict) -> InstructionResult:
         prompt_hash=None,
         model=getattr(instr, "model", None),
         output_validated=True,
-        review_required=bool(getattr(instr, "review_required", False)),
+        review_advisory=_review_advisory(instr),
         condition_matched=getattr(instr, "condition", "") or None,
     )