feat(desktop): ABI Desktop standalone local app

.gitignore

@@ -107,6 +107,9 @@ storage/vectorstore/*
 !libs/naas-abi/naas_abi/apps/nexus/demo/**
 !libs/naas-abi/naas_abi/apps/nexus/apps/*/public/**
 
+# ABI Desktop bundled UI assets (logo, favicon)
+!libs/naas-abi/naas_abi/apps/desktop/gui/web/assets/**
+
 # `abi dev` runtime artefacts (per-worktree instance/pid/log files)
 .abi/
 # Dagster local home dir created by `abi dev up`

Makefile

@@ -105,6 +105,7 @@ help:
 	@echo "  chat-support-agent       Start customer support specialized agent"
 	@echo ""
 	@echo "DEVELOPMENT SERVERS & TOOLS:"
+	@echo "  dev-desktop              Start ABI Desktop dev server (reload + supervisor)"
 	@echo "  api                      Start API server for local development (port 9879)"
 	@echo "  api-prod                 Build and run production API server in Docker"
 	@echo "  api-local                Start local API server in Docker with volume mounting"
@@ -117,6 +118,8 @@ help:
 	@echo ""
 	@echo "TESTING & QUALITY ASSURANCE:"
 	@echo "  test                     Run all Python tests using pytest"
+	@echo "  test-desktop             Run ABI Desktop tests with coverage KPI"
+	@echo "  dev-desktop              Start ABI Desktop dev server (reload + supervisor)"
 	@echo "  test-coverage            Run tests with coverage reporting and badge generation"
 	@echo "  test-ci                  Run basic tests for CI (no external dependencies)"
 	@echo "  test-abi                 Run tests specifically for the abi library"
@@ -546,6 +549,14 @@ path=tests/
 test: deps
 	@ uv run python -m pytest .
 
+# ABI Desktop unit tests + coverage KPI
+test-desktop: deps
+	@ libs/naas-abi/naas_abi/apps/desktop/scripts/test.sh
+
+# ABI Desktop browser dev server (detached, auto-reload, crash restart)
+dev-desktop: deps
+	@ libs/naas-abi/naas_abi/apps/desktop/scripts/dev.sh
+
 # Run API-specific tests
 test-api: deps
 	@ uv run python -m pytest libs/naas-abi-core/naas_abi_core/apps/api/api_test.py -v -s

docs/adr/20260710_desktop-abi-model-router.md

@@ -0,0 +1,69 @@
+# Desktop ABI model router
+
+## Status
+
+Accepted
+
+## Date
+
+2026-07-10
+
+## Context
+
+ABI Desktop users bring their own models: local Ollama installs and cloud provider APIs. Today they pick a model manually from a dropdown. The product vision is **ABI as an intelligent model router**: the user types intent in chat, and the system suggests the best local vs cloud model based on TTL that describes each model's realizable BFO7 processes and hosting site.
+
+Iteration 1 delivered BFO7 section routing (chat → plan, code → build) over Oxigraph. Iteration 3 adds task-aware **model** routing on the same graph stack, without importing `naas_abi` or `naas_abi_core`.
+
+## Decision
+
+1. **Router layer over Oxigraph + BFO7 realizabilities**
+
+   - Extend `ontology/desktop-routing.ttl` with `abi:LanguageModel`, `abi:hostedAt`, `abi:supportsTools`, `abi:canRealize`, `abi:modelRef`, and site individuals `abi:SiteLocal` / `abi:SiteCloud` (aligned with `BFO7Buckets.ttl` `abi:` namespace).
+   - Scaffolded `instances.ttl` seeds example `LanguageModel` individuals per org/model context.
+   - `DesktopGraph.suggest_models(intent_tags, prefer_local)` runs SPARQL over the active context graph, filters tool-capable models, ranks by process match count and hosting preference.
+
+2. **Intent tagging (phased)**
+
+   - **MVP**: rule-based `tag_intent_from_text()` maps composer keywords to intent tags (`code`, `plan`, `summarize`, `general`), then to BFO7 process IRIs (`bfo:BFO_0000015` process, `bfo:BFO_0000023` role).
+   - **V2**: LLM intent tagger replaces keyword rules; same SPARQL router and TTL contract.
+
+3. **API and UI**
+
+   - `POST /api/router/suggest` body `{ "text": "refactor this python file", "prefer_local": true }` returns `{ intent_tags, suggestions: [{ model_ref, label, hosted_at, score, reason, matched_processes }] }`.
+   - Chat/Code composer debounces input and shows suggestion chips above the composer; click applies the model chip and `@ollama:…` or cloud ref.
+
+## Architecture
+
+```mermaid
+flowchart LR
+  User[User intent in composer]
+  Tagger[IntentTagger]
+  Router[SPARQL Router]
+  Suggest[ModelSuggestion]
+  Harness[Harness send]
+
+  User --> Tagger
+  Tagger -->|intent_tags| Router
+  Router -->|instances.ttl + BFO7| Suggest
+  Suggest -->|model_ref| Harness
+```
+
+## Phased rollout
+
+| Phase | IntentTagger | Router | UI |
+|-------|----------------|--------|-----|
+| MVP (this iteration) | Keyword rules in `graph.py` | SPARQL over `instances.ttl` | Debounced chips in `app.js` |
+| V2 | LLM classification | Same SPARQL + richer process TTL | Persist user overrides, graph panel |
+
+## Consequences
+
+- Model routing is data-driven: edit `instances.ttl` to add/remove models without code changes.
+- Local vs cloud tradeoff is explicit via `abi:hostedAt` and `prefer_local`.
+- Only models with `abi:supportsTools true` are suggested (opencode agents require tools).
+- V2 can add per-process subclasses from Nexus process TTL without changing the API shape.
+- Bundle stays self-contained: router vocab ships in `desktop/ontology/`.
+
+## Related
+
+- `docs/adr/20260710_desktop-bfo7-routing-graph.md`
+- `docs/adr/20260710_desktop-org-model-workspace.md`

docs/adr/20260710_desktop-bfo7-routing-graph.md

@@ -0,0 +1,82 @@
+# Desktop BFO7 routing knowledge graph
+
+## Status
+
+Accepted
+
+## Date
+
+2026-07-10
+
+## Context
+
+ABI Desktop already scaffolds per org/model `ontology.ttl` and `instances.ttl` files and loads them into an embedded Oxigraph store. Chat prompts prepend `AGENTS.md` and `MEMORY.md`, but agent selection still comes only from SQLite settings (`chat_agent`, `code_agent`). Nexus ships BFO7 bucket ontologies that classify processes, agents, and documents along seven BFO continuant/occurrent buckets. The desktop app must stay importable without `naas_abi` / `naas_abi_core`, yet still benefit from the same ontology base for routing hints.
+
+## Decision
+
+1. **System ontology config**
+
+   - `desktop_config.resolve_system_ontology_paths()` resolves TTL files in this order:
+     - `ABI_DESKTOP_SYSTEM_ONTOLOGY_PATHS` (explicit override)
+     - bundled `ontology/desktop-routing.ttl`
+     - repo or bundled `BFO7BucketsProcessOntology.ttl`
+   - Optional full `BFO7Buckets.ttl` is discovered when present in the repo; bundled copy is not required for v1.
+
+2. **Graph loading**
+
+   - `DesktopGraph.load_system_ontology()` loads system TTL into named graph `http://ontology.naas.ai/abi/desktop#system`.
+   - `load_org_model_context()` always ensures the system graph is loaded, then loads org/model `ontology.ttl` and `instances.ttl` into `http://ontology.naas.ai/abi/desktop#context/{org}/{model}`.
+
+3. **Routing vocabulary**
+
+   Shared classes in `ontology/desktop-routing.ttl`:
+
+   - `abid:SectionRoute` with `abid:forSection`, `abid:harnessAgent`, `abid:mapsToBfoProcess`
+   - `abid:Organization`, `abid:ModelContext`, `abid:ContextDocument`
+
+   Per org/model scaffold templates declare section route subclasses and seed instances:
+
+   - chat → `plan` agent, maps to `bfo:BFO_0000015` (process / WHAT)
+   - code → `build` agent, maps to `bfo:BFO_0000015`
+   - `AGENTS.md` and `MEMORY.md` as `ContextDocument` individuals
+
+4. **SPARQL routing query**
+
+   ```sparql
+   PREFIX abid: <http://ontology.naas.ai/abi/desktop#>
+
+   SELECT ?agent ?bucketLabel WHERE {
+     GRAPH <http://ontology.naas.ai/abi/desktop#context/{org}/{model}> {
+       ?route abid:forSection ?section ;
+              abid:harnessAgent ?agent .
+       FILTER(?section = "chat")
+       OPTIONAL { ?route abid:mapsToBfoProcess ?bucket . }
+     }
+     OPTIONAL {
+       GRAPH <http://ontology.naas.ai/abi/desktop#system> {
+         ?bucket rdfs:label ?bucketLabel .
+       }
+     }
+   }
+   LIMIT 1
+   ```
+
+5. **Chat send wiring (v1)**
+
+   On `POST /api/chats/{id}/messages`:
+
+   - `graph.resolve_route_agent(active_org, active_model, section)` overrides settings when instances define a route.
+   - `graph.build_routing_prompt_hint(...)` prepends a short routing block to the harness prompt (stored user message stays raw).
+
+## Consequences
+
+- Routing is data-driven: editing `instances.ttl` changes agent selection without code changes.
+- BFO7 process bucket labels appear in prompts as lightweight context for the harness.
+- Bundle stays lean: only routing vocab + process ontology are vendored under `desktop/ontology/`.
+- `owl:imports` in scaffolded ontologies document intent; Oxigraph loads explicit file paths, not remote imports.
+- Iteration 2 enriched instances (distinct BFO7 buckets, model URI, harness, model hints), added `resolve_route()`, wired model hints on send, and surfaced active routing in `/api/health` + Graph UI.
+- Iteration 3 can add per-tool routes, settings↔TTL sync, and a seven-bucket visualization.
+
+## Related
+
+- `docs/adr/20260710_desktop-org-model-workspace.md`

docs/adr/20260710_desktop-org-model-workspace.md

@@ -0,0 +1,66 @@
+# Desktop org/model workspace layout
+
+## Status
+
+Accepted
+
+## Date
+
+2026-07-10
+
+## Context
+
+ABI Desktop needs a canonical way to organize agent instructions, persistent memory, and RDF knowledge per organization and model context. The workspace root may contain multiple organizations (for example a monorepo like `/Users/jrvmac/abi`), and each organization may host multiple model-specific contexts.
+
+## Decision
+
+1. **Path schema**
+
+   ```
+   {workspace_root}/{org}/{model}/AGENTS.md
+   {workspace_root}/{org}/{model}/MEMORY.md
+   {workspace_root}/{org}/{model}/ontology.ttl
+   {workspace_root}/{org}/{model}/instances.ttl
+   ```
+
+   Organizations are directories directly under `workspace_root`. Model contexts are directories under each org. Infrastructure directories (`.git`, `node_modules`, virtualenvs, caches) are excluded from org discovery.
+
+2. **Settings keys** (SQLite `settings` table via `DesktopStore`)
+
+   - `active_org` — selected organization folder name (default: `default`)
+   - `active_model` — selected model context folder name (default: `default`)
+
+   These are separate keys, not a combined `org/model` path string.
+
+3. **Scaffolding**
+
+   - `workspace_layout.scaffold_org_model()` creates missing dirs and template files idempotently.
+   - `ensure_workspace()` scaffolds `default/default` on first workspace creation.
+   - Settings save and `POST /api/workspace/orgs/{org}/models/{model}/scaffold` also scaffold on demand.
+
+4. **API**
+
+   - `GET /api/workspace/orgs` — list orgs + active org/model
+   - `GET /api/workspace/orgs/{org}/models` — list models for an org
+   - `POST /api/workspace/orgs/{org}/models/{model}/scaffold` — create templates + load graph
+   - `PUT /api/settings` accepts `active_org` and `active_model`
+
+5. **Chat context**
+
+   On `POST /api/chats/{id}/messages`, the server reads `AGENTS.md` and `MEMORY.md` from the active org/model path and prepends them to the harness prompt. The user message stored in SQLite remains the raw user text.
+
+6. **Graph loading**
+
+   `DesktopGraph.load_org_model_context()` loads `ontology.ttl` and `instances.ttl` into a named Oxigraph graph `http://ontology.naas.ai/abi/desktop#context/{org}/{model}`. Switching active context clears the previous context graph before loading the new one. Activity triples (chats/messages) remain in the default graph.
+
+7. **UI**
+
+   Settings → General exposes org/model selectors and a scaffold button. The Code explorer auto-expands the active org/model folders.
+
+## Consequences
+
+- Multiple org/model contexts can coexist inside one git workspace without replacing the workspace root.
+- Agent instructions are versionable alongside code when the workspace is a git repo.
+- SPARQL queries can target context graphs with `GRAPH <...#context/{org}/{model}> { ... }`.
+- Org/model names are restricted to safe path segments (`[A-Za-z0-9][A-Za-z0-9._-]*`).
+- Legacy workspaces without org folders continue to work; `default/default` is created automatically.