Run agent actions and workspace search together instead of one or the other

application/single_app/config.py

@@ -96,7 +96,7 @@
 EXECUTOR_TYPE = 'thread'
 EXECUTOR_MAX_WORKERS = 30
 SESSION_TYPE = 'filesystem'
-VERSION = "0.260.024"
+VERSION = "0.260.025"
 IS_DEVELOPMENT = is_development_env_enabled()
 
 SESSION_COOKIE_SAMESITE = os.getenv('SESSION_COOKIE_SAMESITE', 'Lax')

application/single_app/functions_mixed_source_orchestration.py

@@ -412,7 +412,15 @@
 
 
 def should_run_tabular_evidence(user_question, has_narrative_sources=False):
-    """Return whether a mixed-source question needs tabular data or schema evidence."""
+    """Return whether an in-scope tabular source should be computed for this question.
+
+    Evidence gathering is additive. When an authorized tabular source is in scope the
+    tabular engine runs unless the question unambiguously targets a narrative artifact,
+    because deciding which evidence is relevant belongs to the synthesis step rather
+    than to this gate. Indexed tabular chunks carry only a truncated schema preview, so
+    skipping computation leaves the model with a handful of preview rows that can never
+    support a numeric conclusion.
+    """
     normalized_question = " ".join(str(user_question or "").strip().lower().split())
     if not normalized_question:
         return True
@@ -430,24 +438,22 @@
         "across the files", "across the documents", "across the sources",
         "mixed sources",
     )
-    narrative_markers = (
+    # Only artifact markers suppress computation. Topic words such as "report" or
+    # "policy" describe subject matter, not which engine can answer, and previously
+    # suppressed computation over spreadsheets that held the requested values.
+    narrative_artifact_markers = (
         "pdf", "docx", "word document", "presentation", "powerpoint",
-        "paragraph", "section", "policy", "procedure", "contract",
-        "agreement", "memo", "letter", "narrative", "prose", "report",
+        "paragraph", "section",
     )
 
     if any(marker in normalized_question for marker in tabular_markers):
         return True
     if any(marker in normalized_question for marker in collective_markers):
         return True
     if has_narrative_sources and any(
-        marker in normalized_question for marker in narrative_markers
+        marker in normalized_question for marker in narrative_artifact_markers
     ):
         return False
-    if normalized_question in {"summarize", "summary", "summarize the selected sources"}:
-        return True
-    if has_narrative_sources:
-        return False
     return True
 
 
@@ -1413,7 +1419,14 @@
                 source_kind=SOURCE_KIND_TABULAR,
                 engine=EVIDENCE_ENGINE_TABULAR_TOOLS,
                 status=EVIDENCE_STATUS_SKIPPED,
-                summary="Tabular processing was not needed for this narrative-only request.",
+                summary=(
+                    "Tabular computation was not run for this source, so its full table was "
+                    "never read. Any indexed excerpt from this source contains only a "
+                    "truncated schema preview of the first few rows. Do not derive counts, "
+                    "totals, averages, minimums, maximums, trends, or any other numeric "
+                    "conclusion from those preview rows. Call the tabular analysis action if "
+                    "values from this source are required."
+                ),
                 coverage={
                     "selection_mode": normalized_selection_mode,
                     "terminal": True,
@@ -1921,6 +1934,11 @@
             "and tabular tool citations; do not convert computed table facts into unsupported narrative claims. "
             "When selection_mode is selected, current selected-source evidence supersedes prior document "
             "grounding; do not use prior source claims to fill missing current coverage. "
+            "This handoff is your starting evidence, not your only means of gathering evidence: if you have "
+            "actions available and this handoff does not contain what the question needs, call the appropriate "
+            "action to obtain it and reason over the handoff and the action results together before answering. "
+            "Never derive numeric conclusions from an indexed preview of a tabular source whose evidence status "
+            "is not completed; obtain those values from a computed tabular result instead. "
             f"{partial_coverage_instruction}\n\n{serialized_payload}"
         ),
         "mixed_source_coverage": coverage,

application/single_app/route_backend_chats.py

@@ -4523,7 +4523,9 @@
                         Retrieved Excerpts:
                         {retrieved_content}
 
-                        Base your answer only on information supported by the retrieved excerpts and any computed tool-backed results included elsewhere in this conversation context. If the answer is not supported by that information, say so.
+                        These excerpts are your starting evidence, not your only means of gathering evidence. If you have actions or tools available and the excerpts do not contain what the question needs, call the appropriate action to obtain it, then reason over the retrieved excerpts and the action results together. Gather the evidence you are capable of gathering before declining to answer.
+                        Ground every claim in a retrieved excerpt, in computed tool-backed results included elsewhere in this conversation context, or in a result you obtained by calling an action. Never estimate, infer, or fabricate values that none of those sources support; if the evidence is still missing after you have used the actions available to you, say so.
+                        Excerpts drawn from a spreadsheet or other tabular source contain only a truncated schema preview of that file, not its data. Never derive counts, totals, averages, minimums, maximums, trends, or any other numeric conclusion from those preview rows; obtain such values from a computed tabular result instead.
                         If computed tabular results are provided in another system message, treat them as authoritative for row-level values, calculations, and numeric conclusions. Do not say that you lack direct access to the data when those computed results are present.
 
                         Example

docs/explanation/fixes/AGENT_ACTIONS_WITH_WORKSPACE_EVIDENCE_FIX.md

@@ -0,0 +1,175 @@
+# Agent Actions Ignored When Workspace Evidence Is Present
+
+**Fixed in version: 0.260.025**
+
+**Issue:** [#1332](https://github.com/microsoft/simplechat/issues/1332)
+
+**Related:** [#1021](https://github.com/microsoft/simplechat/issues/1021) — turn-level orchestration across chat capabilities, the strategic solution to this class of problem. This fix addresses the concrete symptom and does not close that initiative.
+
+## Issue
+
+Selecting an agent that has actions, enabling a workspace, and asking a specific
+quantitative question produced an answer that:
+
+1. never invoked any of the agent's actions, and
+2. reported numbers that were not actually present in the spreadsheet it cited.
+
+The reported case was a telemetry question against a workspace containing an
+Excel file. Workspace search retrieved a narrative document plus the spreadsheet,
+and the assistant answered from retrieved text alone, fabricating values.
+
+The turn behaved as "retrieval **or** actions" instead of "retrieval **and**
+actions". Evidence gathering should be additive: gather everything the turn is
+capable of gathering, then reason over the union and decide what is relevant.
+
+## Root Cause
+
+Three independent defects combined to produce the symptom.
+
+### 1. Tabular computation was suppressed by the presence of any narrative source
+
+`should_run_tabular_evidence()` in `functions_mixed_source_orchestration.py` was
+a keyword heuristic that ended with a blanket rule:
+
+```python
+if has_narrative_sources:
+    return False
+```
+
+A single PDF landing in the relevance results suppressed computation over an
+authorized spreadsheet. The heuristic also treated topic words — `report`,
+`policy`, `procedure`, `contract`, `agreement`, `memo`, `letter`, `narrative`,
+`prose` — as evidence-type signals. Those words describe subject matter, not
+which engine can answer a question, so they misfired frequently.
+
+When the gate returned `False`, `execute_tabular_evidence_sources(...,
+execute=False)` emitted a `skipped` evidence envelope and the tabular engine
+never ran.
+
+Note that when `enable_mixed_source_chat_search` is disabled, the legacy path in
+`route_backend_chats.py` computes workspace tabular sources unconditionally. The
+mixed-source path had regressed that behavior; this fix restores parity.
+
+### 2. Only a truncated preview of a spreadsheet is indexed
+
+`_build_tabular_schema_summary()` in `functions_documents.py` indexes a single
+schema chunk holding at most `TABULAR_SCHEMA_SUMMARY_MAX_PREVIEW_ROWS` (3) rows
+per sheet. This is intentional — the full file lives in blob storage and the
+tabular engine reads it directly.
+
+However, when defect 1 skipped computation, that preview was still handed to the
+model as ordinary retrieved text. The model then derived counts and averages from
+three rows, which is the direct source of the incorrect values.
+
+The indexed chunk even ends with "This file is available for detailed analysis
+via the Tabular Processing plugin" — the model read the advertisement for the
+tool while being instructed not to use it.
+
+### 3. The retrieval augmentation prompt forbade using actions
+
+`build_search_augmentation_system_prompt()` in `route_backend_chats.py`
+instructed:
+
+> Base your answer only on information supported by the retrieved excerpts and
+> any computed tool-backed results included elsewhere in this conversation
+> context.
+
+The mixed-source evidence handoff built by
+`build_mixed_source_evidence_handoff()` was likewise a closed "synthesize one
+answer" instruction.
+
+Agent actions were in fact available. Agents are constructed with
+`FunctionChoiceBehavior.Auto()` in `semantic_kernel_loader.py`, and the agent is
+invoked with the augmented history through `selected_agent.invoke_stream(...)`.
+No code disables tools when documents are in scope. The model simply obeyed the
+instruction not to look anywhere else, and the retrieved excerpts appeared
+sufficient, so it never called an action.
+
+## Files Modified
+
+| File | Change |
+|---|---|
+| `application/single_app/functions_mixed_source_orchestration.py` | Inverted the tabular gate; narrowed narrative markers; rewrote the skipped-envelope summary; added action permission and a preview-row guard to the evidence handoff |
+| `application/single_app/route_backend_chats.py` | Rewrote `build_search_augmentation_system_prompt()` |
+| `application/single_app/config.py` | Version `0.260.024` -> `0.260.025` |
+| `functional_tests/test_agent_actions_with_workspace_evidence.py` | New regression test |
+| `functional_tests/test_mixed_source_chat_search_consistency.py` | Updated the gating contract a generic question now computes rather than skips |
+
+## Code Changes
+
+### Additive tabular gating
+
+`should_run_tabular_evidence()` now defaults to running. Computation is skipped
+only when narrative sources are present **and** the question unambiguously names
+a narrative artifact:
+
+```python
+narrative_artifact_markers = (
+    "pdf", "docx", "word document", "presentation", "powerpoint",
+    "paragraph", "section",
+)
+```
+
+Topic words no longer suppress computation. Explicit tabular intent and
+collective phrasing still short-circuit to `True`.
+
+### Skipped sources are now self-correcting
+
+A skipped tabular envelope previously read "Tabular processing was not needed for
+this narrative-only request", which told the model the source was irrelevant. It
+now states that the full table was never read, that any indexed excerpt is a
+truncated preview, that numeric conclusions must not be drawn from it, and that
+the tabular analysis action should be called if values are required.
+
+### Prompt contract permits and expects action use
+
+The retrieval augmentation prompt now frames excerpts as starting evidence rather
+than the only permitted evidence, directs the model to call an available action
+when the excerpts lack what the question needs, keeps a hard no-fabrication rule,
+and forbids deriving any numeric conclusion from tabular preview rows.
+
+The mixed-source handoff carries the same permission plus a guard against
+numeric conclusions drawn from an indexed preview of a source whose evidence
+status is not `completed`.
+
+No new setting was introduced. These are correctness fixes and apply
+unconditionally.
+
+## Validation
+
+```powershell
+python .\functional_tests\test_agent_actions_with_workspace_evidence.py
+python -m pytest .\functional_tests\test_mixed_source_manifest_contracts.py .\functional_tests\test_tabular_computed_results_prompt_priority.py -q
+```
+
+The new test asserts:
+
+- a quantitative question with narrative sources present now computes the
+  tabular source (the reported regression);
+- an unambiguous narrative-artifact question still skips computation;
+- topic words such as "report" no longer suppress computation;
+- the skipped envelope warns against numeric conclusions from preview rows;
+- the search prompt permits action invocation and no longer says "only";
+- the handoff instruction permits action invocation.
+
+Existing contracts in `test_tabular_computed_results_prompt_priority.py` and
+`test_mixed_source_manifest_contracts.py` continue to pass unchanged.
+
+### Known unrelated failures
+
+Three tests in `test_mixed_source_chat_search_consistency.py` fail both before
+and after this change, at identical assertions. Their harness builds a synthetic
+namespace for `_execute_mixed_source_tabular_evidence` that is missing
+`maybe_queue_search_tabular_generated_output`, so the stubbed tabular runner
+raises and every source reports `failed`. A third failure originates in
+`foundry_agent_runtime.py`. Both are pre-existing harness drift and are out of
+scope for this fix.
+
+## Before / After
+
+| | Before | After |
+|---|---|---|
+| Spreadsheet + PDF in scope, quantitative question | Tabular engine skipped | Tabular engine runs |
+| Model's view of a skipped spreadsheet | "not needed for this narrative-only request" | Explicit warning that the table was not read and the action should be called |
+| Retrieved excerpts insufficient | Model answers from excerpts or declines | Model calls an available action, then reasons over both |
+| Numbers from a 3-row preview | Permitted implicitly | Explicitly forbidden |

docs/explanation/fixes/index.md

@@ -39,3 +39,4 @@ category: Version History
 - [Generated Artifact Paging, Truncation, and Guidance Carry-Forward Fix](GENERATED_ARTIFACT_PAGING_AND_GUIDANCE_FIX.md)
 - [Admin Settings Pane Variable Scope Fix](ADMIN_SETTINGS_PANE_VARIABLE_SCOPE_FIX.md)
 - [Inline Media Cited-Only Gating Fix](INLINE_MEDIA_CITED_ONLY_GATING_FIX.md)
+- [Agent Actions With Workspace Evidence Fix](AGENT_ACTIONS_WITH_WORKSPACE_EVIDENCE_FIX.md)