Integrate private Otter archive and durable conversation images

docs/design_index.md

@@ -88,6 +88,7 @@ live.
 | Prompts, models, routing, optimisation, or fine-tuning | [prompt programmes and model routing](engineering/prompt_programs_and_model_routing_playbook.md) |
 | Retrieval, memory, RAG, KB growth, or long-horizon state | [agent memory and enduring knowledge](engineering/agent_memory_and_enduring_knowledge.md) |
 | Represented, experience- or discussion-derived learning for role, capability, workflow, tool, acquisition, introspection, message, or task decisions | Draft [represented advice design](engineering/represented_advice_design.md), including Phase 0 subtraction, Phase 0.5 non-active candidates, and the [completed bounded negative learning cycle](engineering/jvnautosci_2720_learning_cycle_2026-09-05.md) under JVNAUTOSCI-2720; then the applicable prompt, memory, evaluation, workflow, or security guidance. The cycle links source-derived formation, experimental exposure, outcome evidence, rejection and subsequent verified absence. It does not establish positive advice efficacy, an ordinary runtime consumer, production activation, represented Arm C or autonomous promotion. |
+| Private Otter archive access, conversation images or research-slide reading | [Private Otter archive and conversation images](engineering/otter_archive_and_conversation_images.md) |
 | Importing external human/agent transcripts into conversation carriers | [External conversation import](engineering/external_conversation_import.md) |
 | Durable assertions, text/logical assertion typing, propositions, context-sensitive retrieval, hypotheses, publication or promotion, or user-, organisation-, project-, source-, theory-, or time-relative knowledge | [Assertion and propositional-sentence ontology](engineering/assertion_and_propositional_sentence_ontology.md), then [contextual knowledge evolution](engineering/contextual_knowledge_evolution.md); for canonical ontology publication or scope change, the active [ontology publication authority boundary](engineering/ontology_publication_authority.md) |
 | Evaluation, benchmarks, or research-sensitive architecture | [agent evaluation and research uptake](engineering/agent_evaluation_and_research_uptake.md) |

docs/engineering/otter_archive_and_conversation_images.md

@@ -0,0 +1,75 @@
+# Private Otter archive and conversation images
+
+Implementation and operating reference for [JVNAUTOSCI-2726](https://naoinstitute.atlassian.net/browse/JVNAUTOSCI-2726).
+Delivery evidence and current release decision belong in Jira. This page describes the bounded interface;
+it does not establish that a particular deployment has been configured or activated.
+
+## Private archive configuration
+
+Provision the separate private OtterArchiveMCP checkout, its SQLite index, and the original backup
+on the machine running Von. Keep that server on stdio. Set these variables in the runtime environment:
+
+| Variable | Meaning |
+| --- | --- |
+| `VON_OTTER_ARCHIVE_OWNER_USER_CONCEPT_ID` | Verified owning Von user; required before access |
+| `VON_OTTER_ARCHIVE_RESOURCE_ID` | Deployment selector; defaults to `personal_otter_archive` |
+| `VON_OTTER_ARCHIVE_MCP_PROJECT_DIR` | Private server checkout; required |
+| `OTTER_ARCHIVE_DB` | Existing archive index; required |
+| `VON_OTTER_ARCHIVE_MCP_COMMAND` | Optional executable override; otherwise the checkout's virtual environment entry point |
+| `VON_OTTER_ARCHIVE_MCP_TIMEOUT_SEC` | Stdio process liveness bound (default 30 seconds); adjust from actual operation evidence |
+
+The owner is resolved before launching the process. Ordinary turns receive a hidden trusted resource
+binding; model arguments cannot choose a path, executable or owner. The subprocess receives only a
+small operating environment and the database path, without provider credentials. Changing configuration
+requires restarting the affected runtime. Removing the owner binding also denies subsequent reads of
+staged private screenshots and derived crops. Existing LinkedIn and hosted Otter integrations remain
+independent.
+
+All ten upstream tools are available with the `otter_archive_` prefix. Orient broad questions with
+status/overview, then search or retrieve bounded sources. `find_entities` frequencies describe source
+records and observed labels, not reconciled meetings or verified identities. Unresolved export-group
+associations stay unresolved. Ordinary reads do not start OCR, vision inference or corpus maintenance.
+
+`get_artifact`/search results supply an authenticated `source_url`. That view displays original images,
+source text, labelled OCR and cached inference, with extraction provenance available for inspection.
+`get_screenshot` stages a checksum-verified original in Von's existing blob/file-copy store. Original
+chunk retrieval returns bounded metadata and an authenticated download URL, keeping base64 out of model
+text. Consumers reassembling originals must check the complete upstream hash.
+
+## Reusable image handling
+
+The existing drop, paste and picker routes feed one image pipeline. Up to eight still PNG, JPEG or WebP
+images may accompany a message (8 MiB and 25 million pixels per image). Media is decoded rather than
+trusted by filename. The composer shows progress, previews, removal and failures; unresolved preparation
+blocks submission. Removing an unsent image removes its message association, not its durable source.
+
+Original bytes live in the existing durable blob store. File-copy records retain detected media type,
+dimensions, checksum and source provenance. New image records explicitly use `attributes.v1` as their
+file metadata authority to avoid duplicate per-field text-relation writes and reads; existing file-copy
+records retain their earlier metadata contract. The normal concept name and type remain represented.
+
+Messages and history carry small attachment descriptors. Only the provider boundary hydrates access-
+checked original bytes into native image inputs, with source IDs, dimensions and citation URLs alongside
+the images. Retained continuations and ordinary telemetry contain descriptors rather than base64.
+OpenAI Chat Completions/Responses and Ollama have image transport; the Gemini structured transport
+currently reports unsupported image input explicitly. A transport does not establish that every model
+on that provider supports vision; use a vision-capable configured model and inspect failures.
+
+`crop_conversation_image` creates a labelled crop/enlargement for small details. It preserves parent
+checksum, pixel rectangle, scale and source audience. Resizing cannot reconstruct destroyed information.
+Originals remain independently inspectable. Conversation image reads require the owning actor; an Otter
+`shared` label or organisation membership does not grant archive access.
+
+## Validation and limits
+
+The CC0 fixtures in `tests/fixtures/research_images` and their generator/reference file cover columns,
+symbols, an exact recurrence, a logarithmic chart, table/code structure, directed feedback and degradation.
+Reference criteria precede tuning. Tests cover authority before proxy launch, hidden bindings, original
+checksums, SDK image bytes and references, crop provenance, history projection and composer failures.
+Use the browser protocol for live upload/reload/follow-up and source-view acceptance.
+
+Private live receipts stay outside Git. Report the selected model, input/preprocessing, source hashes,
+actual crop/cache use, latency and any failed attempts. Sparse OCR or model enrichment is not evidence
+that a topic is absent. Cached model interpretations can be wrong; scientific claims must remain tied
+to inspectable original evidence. This integration does not reconcile the whole archive, prove
+corpus-wide OCR accuracy, cancel Otter, deploy a network archive service or grant wider access.

src/backend/integrations/internal_mcp/catalogue.py

@@ -41325,6 +41325,8 @@ def _build_default_catalogue_knowledge_io_definitions() -> List[MethodDefinition
                 "inclusion for direct jira_add_attachment calls."
             ),
         ),
+        *otter_archive_definitions(),
+        *conversation_image_definitions(),
         # Owner-scoped LinkedIn export search via a local external MCP server.
         MethodDefinition(
             name="linkedin_index_status",
@@ -46111,6 +46113,10 @@ def _register_dynamic_catalogue_methods(
         )
 
 
+from .conversation_image_tools import definitions as conversation_image_definitions
+from .otter_archive_tools import definitions as otter_archive_definitions
+
+
 def build_default_catalogue() -> MethodCatalogue:
     """Return a catalogue pre-populated with the baseline method set."""
 

src/backend/integrations/internal_mcp/conversation_image_tools.py

@@ -0,0 +1,43 @@
+"""Reusable detail recovery for conversation images from any supported source."""
+
+from .gateway import MethodDefinition
+from .schemas import Schema
+
+
+def crop_conversation_image(**kwargs):
+    from ...security.access_control import get_effective_user_concept_id
+    from ...services.conversation_image_service import crop_image
+    from .catalogue import make_error_response
+
+    try:
+        image = crop_image(user_concept_id=get_effective_user_concept_id(), **kwargs)
+        return {
+            "success": True,
+            "image_attachments": [image],
+            "note": "Derived detail view. Resizing does not restore information absent from the original.",
+        }
+    except (ValueError, PermissionError) as exc:
+        return make_error_response("image_detail_unavailable", str(exc))
+
+
+def definitions():
+    return [
+        MethodDefinition(
+            name="crop_conversation_image",
+            handler=crop_conversation_image,
+            input_schema=Schema(
+                required={
+                    "concept_id": str,
+                    "x": int,
+                    "y": int,
+                    "width": int,
+                    "height": int,
+                },
+                optional={"scale": int},
+                allow_unknown=False,
+            ),
+            output_schema=Schema(required={}, optional={}, allow_unknown=True),
+            category="read",
+            description="Inspect a crop of a conversation image, preserving its source and private audience. Supply pixel x/y/width/height and optional integer scale 1–4 (default 2). Use the resulting actual image to recover small text or diagram detail; report what remains unreadable. Original bytes remain unchanged.",
+        )
+    ]