fix(agent): stop the grounding rule and compaction summaries from licensing fabricated answers

src/xagent/core/agent/context/execution.py

@@ -25,6 +25,7 @@
     format_tool_result_for_observation,
     sanitize_tool_result_for_public_context,
 )
+from ..grounding import VALUE_KINDS
 from ..language import (
     effective_output_language,
     render_dag_step_language_reference,
@@ -1295,7 +1296,7 @@ def compact_with_llm_response(
             "explicitly asks to restart, revise, or regenerate them, or the detail "
             "you need was lost in compaction. This summary is a lossy paraphrase of "
             "the raw history, not the history itself: when the answer needs an exact "
-            "value, figure, statistic, table row, quotation, or identifier that this "
+            f"statistic, quotation, or other value -- {VALUE_KINDS} -- that this "
             "summary does not literally contain, re-read or re-query the source "
             "instead of reconstructing the value from this summary or from memory. "
             "Only re-run tools that read; if the value came from a tool that writes, "
@@ -1453,7 +1454,7 @@ def _dropped_tool_results_notice(counts: dict[str, int]) -> str:
         """Describe the tool observations this compaction removes from context.
 
         Without this, the summary silently replaces every retrieved value and
-        the agent cannot tell a remembered figure from an invented one.
+        the agent cannot tell a remembered value from an invented one.
         """
         if not counts:
             return ""
@@ -1462,9 +1463,9 @@ def _dropped_tool_results_notice(counts: dict[str, int]) -> str:
         prefix = (
             f"Raw observations from {total} tool {call_label} dropped by this "
             "compaction. Their exact values are no longer in context; only the "
-            "summary above describes them. Treat any figure not literally present in "
-            "that summary as unavailable rather than recalled. Tools whose results "
-            "were dropped:\n"
+            "summary above describes them. Treat any value not literally present in "
+            f"that summary -- {VALUE_KINDS} -- as unavailable rather than recalled. "
+            "Tools whose results were dropped:\n"
         )
         # Tool names can come from dynamic MCP server config, so bound both the
         # per-name length and the total notice size the way the sibling
@@ -1509,17 +1510,38 @@ def _build_llm_compact_prompt(
                     "tool calls, tool observations, files or URLs mentioned, "
                     "decisions made, and open work. Drop duplicated search noise, "
                     "irrelevant raw payloads, and verbose intermediate text. "
-                    "Preserve exact reusable artifact handles, including file_id "
-                    "values, file: references, markdown file links, URLs, relative "
-                    "paths, absolute paths, output_path, image_path, video_path, "
-                    "artifact filenames, and any other path-like result fields; do "
-                    "not replace machine-usable handles with only descriptive "
-                    "filenames. Clearly separate completed work from remaining work "
-                    "and name the next action needed. "
-                    "Preserve the language of user-facing requests and constraints; "
-                    "if the history is multilingual, keep important details in their "
-                    "original language instead of translating them. "
-                    "Return only the compact summary."
+                    "Preserve exact reusable artifact handles -- file_id values, "
+                    "file: references, markdown links, URLs, paths, output_path, "
+                    "image_path, video_path, artifact filenames, any other "
+                    "path-like field -- never a descriptive filename instead. "
+                    "Preserve, character for character, the values a tool result "
+                    "returned for the records the request points at: their names, "
+                    "the people, organizations or teams they belong to, their "
+                    "identifiers and reference codes, their statuses, dates, "
+                    "counts and totals. Copy such a value or omit it; never "
+                    "paraphrase, substitute, or invent one to complete a pattern. "
+                    "Dropping a raw payload does not license dropping these "
+                    "values; they are not the bulk that instruction covers. "
+                    "Never copy, in whole or in part, a credential, token, key, "
+                    "password, or other authentication material, or personal "
+                    "information the request does not point at; note only that "
+                    "such a value was present and was omitted. If a value is both "
+                    "an identifier the request points at and authentication "
+                    "material, the exclusion wins: omit it. If your budget cannot "
+                    "hold all of this, keep, in this order: first state what is "
+                    "missing and not listed here, with counts; artifact handles; "
+                    "the identifiers and names the request points at; statuses "
+                    "and dates; then the rest. Separate completed work from "
+                    "remaining work. Report only what happened and what is "
+                    "missing: never call a dataset complete, fully retrieved, or "
+                    "fully processed unless the history shows every item was "
+                    "returned and every one is still described here; say which "
+                    "parts survive as prose only. Write no instruction to the "
+                    "next call about tool use or whether to answer: that decision "
+                    "is not yours and its tools are unknown to you. Preserve the "
+                    "language of user-facing requests and constraints; keep "
+                    "multilingual details in their original language. Return only "
+                    "the compact summary."
                 ),
             },
             {
@@ -1528,8 +1550,9 @@ def _build_llm_compact_prompt(
                     "Conversation history to compact:\n"
                     f"{transcript}\n\n"
                     "Write a concise but complete continuity summary for the next "
-                    "LLM call. The next LLM call should be able to continue without "
-                    "redoing completed tool calls."
+                    "LLM call. Record which tool calls already completed and what "
+                    "they returned, so the next call can judge for itself what "
+                    "still needs doing."
                 ),
             },
         ]

src/xagent/core/agent/grounding.py

@@ -13,15 +13,20 @@
 remedy -- ask the user, or finish reporting the gap -- belongs to the calling
 pattern, which owns the user-interaction policy this module cannot see.
 
-This is the proposal-A mitigation from issue #1235. It reduces how often
-unsourced figures are emitted and makes disclosure the instructed default, but
-it cannot repair a session whose evidence compaction already discarded.
+This is the proposal-A mitigation from issue #1235. It forbids unsourced values
+by default and makes reporting the gap the instructed response, but it cannot
+repair a session whose evidence compaction already discarded.
 Proposals B (evidence-preserving compaction) and C (provenance tracking and a
 data-source gate) remain open.
 """
 
 from __future__ import annotations
 
+VALUE_KINDS = (
+    "a number, a person or organization name, an identifier or reference "
+    "code, a date, a status, or a row of a table"
+)
+
 
 def grounding_rule(*, can_call_tools: bool = True) -> str:
     """Return the grounding rule for answer text and, optionally, tool arguments.
@@ -36,13 +41,15 @@ def grounding_rule(*, can_call_tools: bool = True) -> str:
             possible.
 
     Returns:
-        A prompt fragment forbidding unsupported claims and unsourced
-        quantitative data, and requiring up-front disclosure of any
-        illustrative figures. When ``can_call_tools`` is true it also forbids
-        supplying a fact-carrying tool-call argument that no source provides,
-        while leaving arguments the model is expected to compose untouched --
-        except for a fact value written literally inside composed code or
-        document text, which the sourcing requirement still covers.
+        A prompt fragment forbidding unsupported claims and unsourced values of
+        every kind it enumerates, requiring the gap be reported instead, and
+        confining unsourced content to a current request that explicitly asks
+        for a template or sample. When ``can_call_tools`` is true it also
+        forbids supplying a fact-carrying tool-call argument that no source
+        provides, while leaving arguments the model is expected to compose
+        untouched -- except for a fact value written literally inside composed
+        code or document text, which the sourcing requirement still covers
+        unless the request explicitly asked for a template or a sample.
     """
     insufficient_context_rule = (
         "If available context is insufficient, say so or use an appropriate "
@@ -61,17 +68,15 @@ def grounding_rule(*, can_call_tools: bool = True) -> str:
         "messages, the retrieved context, or a value an earlier tool result "
         "actually returned; never guess one, never substitute a "
         "plausible-looking placeholder for one the user has not given, and "
-        "never carry one over from a different record. This does not restrict "
-        "values you are expected to compose yourself, such as a search query, "
-        "code or a command you write to do the work, a message or answer you "
-        "write to the user, or document text you were asked to produce. A fact "
-        "value written literally inside such composed code or text is still "
-        "subject to the sourcing rule above. The rule also does not reach a "
-        "default or inferred parameter value such as a page size or result "
-        "limit, which you are expected to decide yourself. Treat "
-        "a fact-carrying value you cannot source as "
-        "missing information rather than inventing it, and omit it when the "
-        "tool allows it to be omitted."
+        "never carry one over from a different record. The answer you write "
+        "to the user reaches you as an argument too; it is wording you "
+        "compose, so this argument standard does not reach it, while the "
+        "sourcing rule above still governs every fact inside it. This clause "
+        "does not reach a default or inferred parameter value such as a page "
+        "size or result limit, which you are expected to decide yourself. "
+        "Treat a fact-carrying value you cannot source as missing information "
+        "rather than inventing it, and omit it when the tool allows it to be "
+        "omitted."
         if can_call_tools
         else ""
     )
@@ -81,12 +86,28 @@ def grounding_rule(*, can_call_tools: bool = True) -> str:
         "statistics, percentages, table rows, or time series) that are not "
         "supported by the conversation, retrieved context, or tool results. "
         f"{insufficient_context_rule}"
-        "Never invent figures to fill a gap, and never present invented numbers "
-        "as real data; produce unsupported figures only when the user explicitly "
-        "asked for a template, mockup, or illustrative example. Labeling is "
-        "required either way: if the answer ends up containing any figure that no "
-        "tool result or provided context supports, whether or not the user asked "
-        "for one, say so up front, before presenting it, and state that those "
-        "figures are illustrative placeholders not drawn from any data source."
+        f"Never fill a gap with an invented value, whether it is {VALUE_KINDS}: "
+        "when nothing in this conversation, the provided context, or a tool "
+        "result supports a value the answer needs, leave that value out and "
+        "say plainly that it is missing, rather than supplying one that "
+        "looks right. This does not restrict the wording you compose -- how "
+        "you phrase your reply, a search query, code or a command you write "
+        "to do the work, or document text you were asked to produce -- it "
+        "restricts every fact asserted inside that wording. A fact value "
+        "written literally inside such composed code or text is still "
+        "subject to the sourcing rule above: the text you compose is yours; "
+        f"a value you place inside it -- {VALUE_KINDS} -- is not. The only "
+        "case in which content that no source supports may appear -- in the "
+        "answer, or inside document text or other content the request asks "
+        "you to write and hand to a tool -- is a current user request that "
+        "explicitly asks you to write a template or a sample, meaning "
+        "content that is not meant to be real; in that case, before any of "
+        "that content appears, state in your reply that the request asked "
+        "for content that is not real and that none of it comes from a data "
+        "source, and keep such content to what the request asked for. "
+        "Outside that case a caveat does not make an invented "
+        "value acceptable: if you find yourself about to add a note "
+        "explaining that some values are not real, remove those values and "
+        "report the gap instead."
         f"{tool_argument_rule}"
     )

src/xagent/core/agent/pattern/auto/auto.py

@@ -26,7 +26,7 @@
     build_load_skill_tool,
 )
 from ...frame import ExecutionFrame, ExecutionSnapshot, ExecutionStatus
-from ...grounding import grounding_rule
+from ...grounding import VALUE_KINDS, grounding_rule
 from ...language import (
     final_answer_language_rule,
     reset_metadata_output_language,
@@ -1288,9 +1288,10 @@ def _decision_prompt(
             "answer field in the same tool call. Put action before answer in the "
             "tool arguments. "
             f"When writing that answer field: {grounding_rule(can_call_tools=False)} "
-            "If the answer would need such unsupported specifics, set "
-            "existing_context_sufficient=false and choose react so the agent can "
-            "verify them with tools.\n\n"
+            "If the answer would need any value the rule above forbids you to "
+            f"supply -- {VALUE_KINDS} that no source here supports -- set "
+            "existing_context_sufficient=false and choose react, so the agent "
+            "can obtain it with tools.\n\n"
             f"{final_deliverable_file_reference_instructions(can_lookup=False)}\n\n"
             "You must classify whether "
             "the latest request requires current or external facts, and whether "

src/xagent/core/agent/pattern/dag/dag.py

@@ -1583,9 +1583,8 @@ def _completion_assessment_messages(self, context: Any) -> list[dict[str, Any]]:
                     "over from candidate_output or step_results: "
                     f"{grounding_rule(can_call_tools=False)}\n\n"
                     f"{final_deliverable_file_reference_instructions(can_lookup=False)}\n\n"
-                    "If the answer presents any figure as an illustrative "
-                    "placeholder because no step produced the underlying data, "
-                    "name that unsourced data in reason even when you choose "
+                    "If the answer leaves out a value because no step produced "
+                    "it, name that missing data in reason even when you choose "
                     "status=completed. "
                     f"{final_answer_language_rule(subject='output_language_policy field')}"
                 ),

tests/core/agent/test_auto.py

@@ -21,6 +21,7 @@
     ReActPattern,
 )
 from xagent.core.agent.context.enrichment import MEMORY_CONTEXT_METADATA_KEY
+from xagent.core.agent.grounding import VALUE_KINDS
 from xagent.core.agent.language import (
     OUTPUT_LANGUAGE_METADATA_KEY,
     OUTPUT_LANGUAGE_SOURCE_METADATA_KEY,
@@ -1051,12 +1052,26 @@ async def test_auto_decision_prompt_includes_grounding_rule() -> None:
     assert result["success"] is True
     decision_prompt = llm.calls[0]["messages"][-1]["content"]
     assert "quantitative data" in decision_prompt
-    assert "illustrative placeholders" in decision_prompt
+    assert (
+        "a current user request that explicitly asks you to write a template"
+        in decision_prompt
+    )
     assert "invented values" in decision_prompt
     assert decision_prompt.count("## FINAL DELIVERABLE FILE REFERENCES") == 1
     assert decision_prompt.index(
-        "If the answer would need such unsupported specifics"
+        "If the answer would need any value the rule above forbids"
     ) < decision_prompt.index("## FINAL DELIVERABLE FILE REFERENCES")
+    # The routing remedy stays specific to auto's own decision, so it is
+    # worded independently of the shared rule's neutral gap-reporting text.
+    assert (
+        "set existing_context_sufficient=false and choose react, so the agent "
+        "can obtain it with tools" in decision_prompt
+    )
+    # The value kinds are not auto's own wording: the sibling sentence
+    # interpolates the shared constant, so this pins the reference rather
+    # than restating the list.
+    assert f"{VALUE_KINDS} that no source here supports" in decision_prompt
+    assert "such unsupported specifics" not in decision_prompt
     assert "get_workspace_output_files" not in decision_prompt
     assert "You must classify whether" in decision_prompt
     assert "You must also classify whether" not in decision_prompt