JVNAUTOSCI-2613: Carry shared conversation situations across turns

AGENTS.md

@@ -26,7 +26,11 @@ a reasonable interpretation, take a bounded action, inspect the result, and
 repair mistakes. Uncertainty alone is not a reason to refuse, interrupt the
 user, or demand confirmation. Ask when authority is missing or when plausible
 choices differ materially in privacy exposure, external commitment, cost, or
-harm that cannot readily be detected and recovered.
+harm that cannot readily be detected and recovered. Also ask one focused
+question when a material fact, preference, or constraint cannot be obtained
+from the shared situation, accessible evidence, tools, or a reasonable
+recoverable assumption. That is information acquisition, not a permission
+ritual; continue independent useful work where possible.
 
 Von is also a research platform for testing whether represented knowledge and
 behavioural authority improve reliability, adaptability, inspectability, and
@@ -129,19 +133,34 @@ never be forced onto newer evidence.
    semantics, validation, execution, persistence, integrations, safety, and
    performance-critical mechanisms. It must not silently become the durable
    home of task-specific semantic policy that should be represented.
-9. **Share situation without requiring stages.** When more than one consumer or
-   model call is justified, share the canonical situation and give each only
+9. **Let conversations carry their shared situations.** A conversation is a
+   distinguished carrier of the participants' evolving situation across turns,
+   not merely a message transcript. Its situation is a provisional, revisable
+   account of material objectives, referents, commitments, assumptions,
+   unknowns, questions, effects, and outcome criteria—not canonical domain
+   truth or an authority grant. Bring relevant projections and exact
+   observations from other carriers into it while leaving canonical task,
+   effect, document, and knowledge stores authoritative. When more than one
+   consumer or model call is justified, share that situation and give each only
    the context it needs. Record material additions, omissions, summaries, and
-   provenance in proportion to the claim. A simple or direct turn need not
-   instantiate a universal turn-state object or stage projections.
+   provenance in proportion to the claim, and do not widen actor-scoped
+   diagnostics merely because the carrier has a wider audience. Visibility or
+   contribution to a shared conversation does not by itself grant authority to
+   erase or replace its owner-scoped carrier. Use the weakest adequate
+   representation; inspectable text is a good default, not a hard requirement.
+   A simple or direct turn need not instantiate a universal state object,
+   schema, summarisation stage, or stage projections.
 10. **Preserve opportunity.** Support layers should expose typed facts,
     bounded actions, evidence, retry/recovery options, and partial progress so a
     capable model can still succeed. A safety requirement may close an unsafe
     path; it should not erase safe alternatives.
 11. **Minimal imposition.** Use available context and tools before interrupting
     the user. Prefer low-burden, reversible progress, and treat needless
     refusal, clarification, or confirmation as real failures. Ask when
-    ambiguity is decision-relevant or authority is missing.
+    ambiguity is decision-relevant, material information is otherwise
+    unavailable, or authority is missing. Preserve an unanswered material
+    question in the shared situation and continue independent useful work; do
+    not turn elicitation into performative confirmation.
 12. **Evidence proportional to the claim.** Do not demand release-grade proof
     for a mechanical change, and do not claim end-to-end success from a unit
     test. Match validation cost to risk and asserted scope.
@@ -182,6 +201,7 @@ Use these defaults:
 | Ambiguous action choice or adaptive recovery | Model judgement with bounded tools, read-back, and evaluation |
 | Semantic judgement that may change with models or evidence | Prompt/programme plus evaluation |
 | Reusable, inspectable multi-step behaviour, especially durable or recoverable | VWL workflow |
+| Evolving conversation objectives, referents, commitments, unknowns, questions, and effect observations | Inspectable conversation-situation text, supplemented by structure only where it has a clear advantage |
 | Durable typed knowledge, provenance, policy identity, or cross-session state | Vontology/KB |
 | Raw document, trace, event, blob, cache, or derived index | Fit-for-purpose store with represented manifest/provenance where needed |
 

docs/engineering/contextual_knowledge_evolution.md

@@ -70,6 +70,53 @@ In particular:
 Qualify “canonical” when using it. Say canonical **concept identity**, live
 **authority**, base **publication**, or durable **read-back**, as applicable.
 
+### 2.1 Conversation situations as lightweight theories
+
+A conversation is a distinguished carrier of what its participants currently
+take their shared undertaking to be. That situation is not identical to the
+transcript: it may carry resolved referents, continuing objectives,
+commitments, assumptions, material unknowns, outstanding questions, observed
+effects, outcome criteria, and links to relevant sub-situations or other
+carriers. In this broad contextual sense it is a theory, but it need not begin
+as a formal logic, Vontology graph, workflow, or universal state schema.
+
+An inspectable, bounded plain-text description on the conversation is the
+default while it remains adequate. Treat it as provisional and revisable, not
+as canonical domain truth, durable publication, enduring memory, or an
+authority grant. Exact machine observations may accompany it as a small
+structured projection when identity, idempotence, or effect reconciliation
+benefits from structure. The originating task, workflow, document, or
+knowledge store remains authoritative; the conversation imports the
+observation needed for continuity and does not redispatch an effect merely to
+rediscover its recorded outcome.
+
+The initial structured observation projection is intentionally a recent,
+bounded suffix rather than a second event store. It reports how many earlier
+observations were omitted, keeps the durable producer's projection watermark,
+and carries only audience-safe outcome fields into a shared conversation.
+Raw actor-scoped diagnostic text remains in its authoritative diagnostic
+carrier. Situation revisions retain the producing request identity so an
+invisible sidecar change can be traced to the turn that authored it.
+
+Carrier visibility, contribution, and destructive authority are different.
+An accepted participant may contribute turns without thereby acquiring the
+right to reset the owner's transcript and shared situation. Treat erase,
+replacement, and other high-blast-radius carrier operations as separate
+authority decisions.
+
+Text and inspectability are preferences, not hard criteria. A stronger or
+different representation earns its place when a current capability needs a
+clear advantage such as field-level concurrent revision, stable typed
+identity, cross-conversation querying, inference, independent governance, or
+more reliable effect reconciliation. Where useful, retain an inspectable
+projection, but do not preserve text ceremony when another representation is
+materially better.
+
+Do not turn the carrier into a compulsory summarisation stage. A direct turn
+may read the current situation and revise it in the same model call. Add
+sub-situation identity, promotion, inheritance, conflict resolution, or
+cross-session reuse only when actual work needs those semantics.
+
 ## 3. Four questions for a coding agent
 
 For a triggered change, ask only:
@@ -98,6 +145,9 @@ claim.
 
 As observed on 29 July 2026, related capabilities are distributed across:
 
+- inspectable conversation-situation text and bounded exact observations on
+  chat-history sessions, used as provisional per-conversation context rather
+  than publication or canonical knowledge;
 - base concept and text relations in `concept_service.py`,
   `concept_relation_service.py`, and `text_value_service.py`;
 - actor- and organisation-visible assertion deltas in

docs/engineering/minimal_imposition_design_principle.md

@@ -5,7 +5,7 @@
 - **Authority:** Explanatory guidance under `AGENTS.md`; adds no independent
   per-task gate or control taxonomy
 - **Created:** 2026-04-03
-- **Last reviewed:** 2026-07-25
+- **Last reviewed:** 2026-07-29
 - **Freshness boundary:** Principle, not current implementation evidence
 
 ## 1. Purpose
@@ -108,6 +108,15 @@ Von should normally exhaust the context it already has access to before asking h
 
 Questions should normally be **narrow, low-effort, and high-value**. If the remaining ambiguity is not materially decision-relevant, interruption is often the wrong default.
 
+The converse matters too: a focused question is a legitimate information-
+acquisition action when a material fact, preference, or constraint is not
+available from those sources and a recoverable assumption would change the
+undertaking. This is not the same as asking for permission Von already has.
+Record the outstanding question in the conversation's shared situation so it
+survives the turn, ask only what unlocks a materially better next step, and
+continue any independent useful work rather than making the whole undertaking
+wait.
+
 The same rule applies after a failed specialised route. If a selected workflow
 fails before meaningful tool progress or durable effect verification, Von
 should normally try the next bounded machine-side recovery route from the