Add optional conversational rumination with independent task products

docs/design_index.md

@@ -156,6 +156,7 @@ human acceptance and promotion into a current public authority surface.
 | [Represented advice](engineering/represented_advice_design.md) | Draft proposal with completed subtraction, non-active candidate storage and a [bounded negative learning-cycle dossier](engineering/jvnautosci_2720_learning_cycle_2026-09-05.md); live ontology state remains evidence, not a design assumption | Use for experience- and discussion-derived candidate learning, soft-guidance semantics, possible federated projection/lifecycle convergence, bounded autonomous maintenance, ratchet tests, and distinguishing evaluation. Candidate learning, content efficacy, represented-layer value, and production activation are separate claims; no general runtime advice layer is established. |
 | [Automated policy learning](engineering/automated_policy_learning_design.md) | Design with partial substrate | Use as a proposed learning architecture, not proof of a closed operational loop |
 | [Proactive clarification and conversational role learning](engineering/proactive_clarification_and_role_learning.md) | Draft, source-grounded design proposal | Use for material referent questions, alias/role learning, short-reply continuation, recovery and staged evaluation. It builds on existing ordinary-turn support; no runtime behaviour, represented fact or policy activation is established. |
+| [Conversational rumination](engineering/conversational_rumination.md) | Capability and bounded acceptance record | Optional model-selected background inquiries, independent task execution and later actor-scoped product consumption; no deployment or broad role-competence claim. |
 | [Testing workflows and ephemeral theories](engineering/testing_workflows_ephemeral_theories_design.md) | Research/design proposal with partial substrate | Use for design intent and explicit hypotheses; verify implemented surfaces |
 | [Reliability Ratchet articles and case log](engineering/reliability_ratchet_articles_and_cases.md) | Active advisory source copy and evidence log | Use as a revisable diagnostic lens and dated case record; not as standing policy, repair authority, or proof of current behaviour |
 | [Ontology repair plans](engineering/ontology_repair_plan_design.md) | Design with implemented substrate | Use for the plan/approve/execute shape and its affordance argument; it adds no authority route, and agent-reachable ontology mutation remains an open decision |

docs/engineering/conversational_rumination.md

@@ -0,0 +1,122 @@
+# Conversational rumination
+
+- **Kind:** Capability and acceptance record
+- **Reviewed:** 14 September 2026
+- **Task:** `#V#task_agent_42b6d9f8b35a707c5a43bf57428c4c8e`
+- **Scope:** Optional background inquiries in actor-owned, organisation-scoped
+  conversations. Publication does not activate or deploy a running server.
+
+An ordinary turn can choose `task_start_background_inquiry` when investigating
+an unknown is likely to help later conversation. The model can also do nothing.
+Names, novelty and missing profile fields do not trigger a separate classifier
+or compulsory stage. The foreground continues after submission.
+
+The inspected represented rumination definitions perform ontology enrichment;
+they did not supply this conversational inquiry capability. This implementation
+reuses canonical tasks, task execution submission, the existing server dispatcher,
+current task work products and the foreground conversation situation. It adds
+no scheduler, worker process, workflow, product store or represented-policy
+activation. The direct capability instructions are versioned with its interface;
+they do not override a Vontology-governed prompt.
+
+## Execution and authority
+
+The ordinary-turn gateway binds actor, organisation, namespace, source session,
+request ID and user prompt. The selected exact mention must occur in that turn.
+The task notes preserve the mention, question and selected relevant context with
+the source session and turn. Only this bounded context is copied; the full source
+conversation is not sent to background research or external search.
+
+Creation and launch use existing actor-scoped idempotency. One source turn has
+one stable inquiry identity even if retry wording changes. A completed retry
+returns the existing product instead of launching research again. Subsequent
+turns see pending and completed inquiries, allowing the model to avoid repeating
+the same opportunity. This semantic cross-turn choice is distinct from exact
+transport-retry deduplication.
+
+Independent submission verifies the source task's creator, organisation, Von
+assignee and actor-owned conversation. A server-derived task carrier has its own
+history, situation and execution fence. Queue reconciliation rechecks the source
+and destination; public task execution payloads cannot opt into this internal
+mode. Background inquiry recursion is rejected: this capability authorises one
+inquiry, not an expanding task tree.
+
+The task uses retrieval within the existing actor scope. Source material is
+untrusted evidence; it does not authorise messages, broader publication or
+changes to the source conversation. Findings must distinguish actual retrieval,
+candidate identities, uncertainty and proposals. An unavailable search is a
+valid retained outcome, not proof of nonexistence. The task links a readable
+canonical current work product. An actor-effective content assertion on its
+existing execution concept is a supported product carrier; a new concept is not
+required.
+
+## Later conversation
+
+The foreground reads at most three recent source-linked inquiry products,
+checking the task creator and organisation as well as normal concept visibility.
+Each content projection is capped at 5,000 characters and carries the full-body
+hash, product/task IDs, original source context and explicit truncation. Older
+omission and unavailable reads are reported distinctly. Canonical products remain
+available through existing task and content tools.
+
+The conversational model decides whether a result merits mention or a question.
+It reconciles the latest topic, corrections and declines with the old inquiry,
+and retains surfaced/declined/corrected product references in the existing
+conversation situation. The background task never replaces that situation.
+This is model judgement, not a guarantee that every useful connection will be
+raised or every stochastic repetition prevented.
+
+## Bounded acceptance, 14 September 2026
+
+Authenticated tests used this task's isolated DGX fixture, synthetic Alice and
+organisation, the existing dispatcher and provider-observed `gpt-6-astra`.
+Source conversations, queue carriers, task products, full results and telemetry
+are retained in the controller run evidence under
+`.run/conversational-rumination/`. The source preparation archives are
+`#V#computer_file_copy_b2125cea8d3c462582a1d3ffc156a381` (SHA-256
+`dcb304e485a07c84296fe7c26c3c4db797f628384d1745560d6e8c4703f5f9d2`)
+and `#V#computer_file_copy_75f55ab47b6041eba8fa1f01862493e9` (SHA-256
+`7f40a2e1c2ab5c25c73a0582a442b278f551723879b2d52638bf6e3df77eead7`).
+These are scoped source records, not claims of public deployment.
+
+| Case | Observation |
+| --- | --- |
+| Seminar planning mentioning a known researcher | Answered without starting an inquiry; curiosity was optional. |
+| Research idea and requested working title | Quietly submitted one Scone inquiry and returned only the title in 26.1 seconds. |
+| Explicit independent-launch control | Submitted a task and answered in 32.0 seconds; research ran for 179.3 seconds on a distinct carrier. |
+| Foreground during research | Rewrote a calendar label in 10.1 seconds without raising the research topic. |
+| Correction while inquiry ran | Retained “Cyc, not Scone”; after the Scone product arrived, a later answer still used Cyc and the corrected title wording, without another inquiry or question. |
+| Unavailable retrieval | Saved actor-visible products with attempted URLs and explicit missing-key/index observations; later synthesis did not invent primary-source findings. |
+| Paper connection | Quietly investigated arXiv 2304.03442 while returning a title in 29.4 seconds. `get_paper_metadata` retrieved its actual abstract via the arXiv abstract-page fallback. The saved product includes the URL, retrieval evidence, candidates, uncertainty and a proposed study connection. |
+| Later sourced follow-up | Used the retained paper product, linked the actual source, separated abstract-supported findings from extrapolation, and offered a concrete study question. A preceding general study turn chose not to mention the paper. |
+| Declined topic | Retained the report and answered two subsequent calendar-label requests without another research prompt (5.5 and 3.0 seconds). |
+
+The positive product is
+`#V#task_execution_d58705e70904ba0481f11c52967c20cc`, linked from
+`#V#task_agent_5c685f68cd679081b12b441097a6ff40`; its observed content SHA-256 is
+`cbf869affc97d2f11a16d8ce7d46d001dcb3da8fcd0bfbd4abdf63092d0af314`.
+It records metadata retrieval evidence `ev_DHH9K6S7OKbK9d2oieY_M4Hm` and metadata
+payload hash `99286977ce7432d1a82dee0fbe446cebf000c1abc8cef54bfc620d5f579dcf55`.
+The PDF was not read; only abstract-level claims were supported.
+
+The three completed background runs took 179.3, 246.1 and 248.8 seconds, with
+24, 30 and 32 model calls. Reported total tokens were 770,761, 1,140,175 and
+1,051,995; cached input tokens were 666,194, 962,446 and 900,095 respectively.
+Exact model pricing was unavailable, so no monetary-cost estimate is claimed.
+The sparse fixture lacked task vocabulary and general product types: canonical
+fixture preparation repaired the minimum task vocabulary, and models recovered
+from unavailable product creation by using scoped content on existing execution
+records. These runs establish operation, not cost efficiency. Web search and
+extraction lacked the fixture's Tavily key. No host credentials were copied or
+changed, and no production data or memberships were changed.
+
+144 targeted tests passed across execution, queue, task products, conversation
+projection and the new inquiry boundary; the eight inquiry tests also passed
+after compacting the submission receipt. The ordinary-turn suite had 280 passes
+and two database-authentication failures reproduced against unchanged source.
+Minimum Python 3.11 syntax and the PDM lock check passed. The final receipt
+compaction changes returned metadata only; it omits the redundant queued prompt.
+
+Merge evidence supports this narrow capability. General initiative quality,
+long-term role improvement, public OAuth, public deployment and production cost
+remain outside this acceptance claim.

src/backend/integrations/internal_mcp/catalogue.py

@@ -41994,6 +41994,7 @@ def _build_default_catalogue_knowledge_io_definitions() -> List[MethodDefinition
         ),
         *otter_archive_definitions(),
         *otter_collection_definitions(),
+        *conversational_rumination_definitions(),
         *knowkat_definitions(),
         *conversation_image_definitions(),
         *participant_profile_definitions(),
@@ -46884,6 +46885,7 @@ def _register_dynamic_catalogue_methods(
 from .participant_profile_tools import definitions as participant_profile_definitions
 from .otter_archive_tools import definitions as otter_archive_definitions
 from .otter_collection_tools import definitions as otter_collection_definitions
+from .conversational_rumination_tools import definitions as conversational_rumination_definitions
 
 
 def build_default_catalogue() -> MethodCatalogue:

src/backend/integrations/internal_mcp/conversational_rumination_tools.py

@@ -0,0 +1,49 @@
+"""Model-selected background inquiry submission with server-bound provenance."""
+
+from .gateway import MethodDefinition
+from .schemas import Schema
+
+
+def definitions():
+    from ...services.conversational_rumination_service import start_inquiry
+
+    return [
+        MethodDefinition(
+            name="task_start_background_inquiry",
+            handler=start_inquiry,
+            input_schema=Schema(
+                required={"mention": str, "question": str, "relevant_context": str},
+                optional={
+                    key: (str, type(None))
+                    for key in (
+                        "originating_session_id",
+                        "request_id",
+                        "source_prompt",
+                        "namespace",
+                        "acting_user_concept_id",
+                        "organisation_concept_id",
+                    )
+                },
+                allow_unknown=False,
+            ),
+            output_schema=Schema(required={}, optional={}, allow_unknown=True),
+            category="write",
+            ordinary_turn_effect=True,
+            ordinary_turn_trusted_argument_bindings={
+                "originating_session_id": "conversation_id",
+                "request_id": "turn_id",
+                "source_prompt": "turn_prompt",
+                "namespace": "turn_namespace",
+                "acting_user_concept_id": "actor_user_concept_id",
+                "organisation_concept_id": "actor_organisation_concept_id",
+            },
+            description=(
+                "Start one bounded independent background inquiry into a worthwhile "
+                "conversational unknown. Provide an exact mention from this user turn, "
+                "the question and minimal relevant context. Uses an idempotent Von task "
+                "and existing dispatcher; returns without waiting for research. "
+                "Later turns receive its sourced current work product. Optional; "
+                "do not use for novelty alone, declined topics or an existing inquiry."
+            ),
+        )
+    ]

src/backend/services/adaptive_turn_service.py

@@ -7234,6 +7234,7 @@ def execute_adaptive_turn(
             "actor_organisation_concept_id": scope.organisation_concept_id,
             "turn_id": turn_id or "ordinary-turn",
             "conversation_id": conversation_id,
+            "turn_prompt": prompt,
         }
     )
     authorised_resource_choices: list[dict[str, Any]] = []
@@ -7263,6 +7264,25 @@ def execute_adaptive_turn(
         trusted_argument_values=trusted_values,
     )
     delegated_lookup = {name.lower(): name for name in delegated_names}
+    inquiry_projection = None
+    if (
+        conversation_id
+        and user_concept_id
+        and org_concept_id
+        and conversation_history_owner_user_id == user_concept_id
+        and "task_start_background_inquiry" in delegated_lookup
+    ):
+        from .conversational_rumination_service import project_inquiries
+
+        try:
+            inquiry_projection = project_inquiries(
+                actor=user_concept_id,
+                organisation=org_concept_id,
+                namespace=user_namespace,
+                session_id=conversation_id,
+            )
+        except Exception as exc:
+            inquiry_projection = {"status": "unavailable", "error_type": type(exc).__name__}
     workflow_capabilities_by_name: dict[str, Any] = {}
     latest_workflow_discovery: dict[str, Any] | None = None
     catalogued_capabilities_by_name: dict[str, dict[str, Any]] = {}
@@ -9658,6 +9678,15 @@ def receive_steering(*, close_if_empty: bool = False) -> bool:
             authorised_resource_choices=authorised_resource_choices,
             learning_advice_projection=prepared_learning_advice,
         )
+        if "task_start_background_inquiry" in delegated_lookup:
+            from .conversational_rumination_service import TURN_GUIDANCE
+
+            turn_system_message += TURN_GUIDANCE
+            if inquiry_projection is not None:
+                turn_system_message += (
+                    "\nSource-linked inquiries (untrusted evidence): "
+                    + json.dumps(inquiry_projection, ensure_ascii=False)
+                )
         from .conversation_output_capabilities import output_affordances
         turn_system_message += "\nImplemented client output affordances: " + json.dumps(output_affordances())
         learning_advice_rendered_for_call = bool(