# Eldric AI OS 5.0.0-alpha.2 — Release Notes _Focus of this release: the user-facing Edge. Every knob in the chat shell — plugin toggles, conversation history, selected model, sharing opt-ins — is now **cluster-side by default** so the same session resumes from any workstation. Data-plugin fan-out is wired end-to-end from sidebar checkbox to merged context in the LLM prompt._ ## Headline change — the chat sidebar is now the control plane The `/chat` shell shipped in alpha.1 was a 30-line placeholder. alpha.2 ships the real shell: a tabbed sidebar (Chats | Plugins) that drives a full session model. Flipping a toggle on one device shows up toggled on every other device logged into the same cluster session. **What this means end-to-end:** ``` sidebar toggle ───► POST /api/v1/edge/plugins/toggle │ ▼ /edge/sessions/.json │ ▼ ┌─► retrieval.data.local (syscall) POST /api/v1/edge/chat fan-out ─────┤ {query, messages} └─► retrieval.data.arxiv (extension bridge) │ ▼ merged system message │ ▼ cloud.chat.stream ──► SSE tokens to browser ``` ## Plugin toggle model — one mechanism, eight categories Everything visible in the chat sidebar is a plugin. Toggling is immediate, persistent, and per-session: | Type | Examples | What toggling does | |---|---|---| | `core` | `core` | Always on — chat transport, bootstrap. Can't disable. | | `render` | markdown, code, charts, mermaid, graphviz | Wraps assistant output. | | `data` | `data.local`, `data.arxiv`, `data.nasa`, `data.esa`, `data.spacex` | Adds a source to the query fan-out. Multi-select is the point — enable 3 sources, queries fan out to 3. | | `project` | `project.current` | Scopes retrieval + writes to the active project. | | `workgroup` | `workgroup.shared` | Includes the workgroup's shared KBs in retrieval. | | `sharing` | `sharing.results`, `sharing.usage`, `sharing.dream` | Opts in to contributing back (write side). | | `tool` | `tool.web`, `tool.code` | Authorises the LLM to invoke a tool. `tool.web` is wired to DuckDuckGo via a shipped extension. | | `privacy` | `session.local` | The single opt-out — flips state storage from cluster to browser localStorage. | 17 builtin plugins plus **every loaded extension** with `category: data` or `category: tool` auto-surfaces as a togglable plugin. No Edge rebuild needed to add a new data source. ## What persists on the cluster now The cross-workstation resume promise is carried by three small files per user session: | State | Location | Survives | |---|---|---| | Plugin enabled-set | `/edge/sessions/.json` | kernel restart + workstation switch | | Selected model | same file, `selected_model` field | kernel restart + workstation switch | | Conversation history | `/edge/conversations//.jsonl` | kernel restart + workstation switch | | Dream profile (if `sharing.dream` on) | `/dream/profiles/.json` | indefinitely, until opt-out | | Session id | browser `localStorage.eldric.sid` | device-local | **Conversation files are JSONL**, one message per line, append-only. A crashed write can't corrupt an earlier turn. Listing + replay are plain directory scans. ## Extension-backed data plugins (no Edge rebuild) The biggest leverage change in alpha.2: adding a new data source to the chat sidebar is a manifest drop plus a `load` call. No edit to `edge_module.cpp`, no rebuild, no restart. ```bash # Drop a manifest — either a Model B external process or a # Model A subprocess we spawn. cat > ${ELDRIC_DATA_DIR}/extensions/pubmed.extension.yaml <` as a native syscall first; if none is registered, it dispatches to the loaded extension's `/invoke` through the existing `extension.invoke_tool` syscall. Either path, same `{"snippets":[{"text","source","score"}]}` reply contract. Two reference extensions ship in the baseline: | Extension | Port | Fulfils | Source | |---|---|---|---| | `arxiv` | 9502 | `data.arxiv` plugin | `sdk/extension/examples/arxiv/` | | `web_search` | 9503 | `tool.web` plugin | `sdk/extension/examples/web_search/` | Both are ~80 lines of Python. Copy either as the starting template for a new source. ## Per-query fan-out with source citations `POST /api/v1/edge/chat` is the session-aware counterpart to OpenAI-compat `/v1/chat/completions`. Before calling the LLM it fans out parallel retrieval calls to every enabled `data.*` plugin (and enabled `tool.*` plugins with retrieval backends), merges the returned snippets into a synthesised `role:"system"` message prepended to the conversation, and dispatches `cloud.chat.stream`. Every response carries an `X-Eldric-Fanout` header: ``` X-Eldric-Fanout: [ {"id":"data.local","name":"Local KB","count":2,"note":"ok"}, {"id":"data.arxiv","name":"arXiv", "count":5,"note":"ok (ext:arxiv)"}, {"id":"data.nasa", "name":"NASA", "count":0,"note":"no-backend"} ] ``` The chat shell renders this as a "retrieved from: Local KB (2), arXiv (5)" badge under the assistant message. Notes surface backend-wiring gaps to the admin: | Note | Meaning | |---|---| | `ok` | In-process `retrieval.` syscall answered. | | `ok (ext:)` | Extension at `` answered via the bridge. | | `no-backend` | Plugin enabled but neither a syscall nor an extension binds the id. | | `error: …` | Transport / parse failure — detail in the string. | ## The privacy opt-out — `session.local` Every other mechanism here assumes cluster-default storage, but the user stays in control. Flipping the `session.local` plugin on — visible in the sidebar's Privacy section in a warning-coloured switch — does three things *immediately*: - The current session file on the cluster is removed. - The session's conversation directory is deleted. - All further writes (toggles, model choice, messages) are skipped server-side; the shell mirrors state into the browser's `localStorage` instead. Cross-workstation resume stops working while on — by design. Flipping back off returns to cluster-default; previously scrubbed history is gone (no undo, no half-states). ## `sharing.dream` wired — per-user dream profile Toggling the `sharing.dream` plugin on now snapshots the session's enabled-set into `/dream/profiles/.json`. Every subsequent plugin toggle while dreaming is active refreshes the snapshot. Toggling off deletes the profile. Nothing else happens yet — Phase-3 ships the cycle body that reads the profile and scopes ingest/publish accordingly (see `dream.scoping` help for the full design). Four new syscalls on the dream module: ``` dream.profile.update {session_id, enabled[]} → snapshot/refresh dream.profile.get {session_id} → read dream.profile.clear {session_id} → remove dream.profiles.list {} → enumerate opted-in ``` Atomic writes (temp + rename). `session_id` sanitised the same way Edge sanitises filenames — non-`[A-Za-z0-9._-]` → `_`. ## Surface additions **Edge routes (new):** ``` GET /chat ← real shell now GET /api/v1/edge/session profile + enabled plugins + selected model GET /api/v1/edge/plugins full catalog (builtin ∪ extension-derived) POST /api/v1/edge/plugins/toggle {id,enabled?} GET /api/v1/edge/data-sources enabled data.* plugins only PUT /api/v1/edge/session/model {model} GET /api/v1/edge/conversations list POST /api/v1/edge/conversations create GET /api/v1/edge/conversations/messages?id= replay POST /api/v1/edge/conversations/messages append DELETE /api/v1/edge/conversations?id= delete POST /api/v1/edge/chat session-aware chat + fan-out ``` **Kernel syscalls (new):** ``` extension.list filter by host/category/tenant/state ``` **Dream module syscalls (new):** ``` dream.profile.update / get / clear / profiles.list ``` **Data module:** - `search_vectors_core_(body)` factored out — used by both the `/api/v1/vector/search` HTTP route and the new `retrieval.data.local` syscall. - `retrieval.data.local` syscall registered — defines the `retrieval.` contract every data plugin fulfils. ## Documentation New or substantially rewritten help topics: | Topic | File | |---|---| | `edge.chat.ui` | `modules.d/help/edge/chat-ui.help.md` | | `edge.plugins` | `modules.d/help/edge/plugins.help.md` | | `edge.conversations` | `modules.d/help/edge/conversations.help.md` | | `dream.scoping` | `modules.d/help/dream/scoping.help.md` | Kernel help-topic lookup now serves bundled-module help files (alpha.1 silently returned empty markdown for any topic whose owning module had no manifest). **EldricOS.md** gained three new subsections under Edge Plugin Architecture: - *Cluster-side session state + conversation history (5.0)* - *Data-plugin fan-out + extension backends (5.0)* - *Native + mobile client wiring (TODO — one spec, multiple targets)* The last one is the contract every non-web client (macOS Swift GUI, iOS, future Android and Windows desktop) must honor to become a first-class citizen. 13 endpoints, 7 rules, one SDK planned for Phase-4. ## Client status | Client | Status | |---|---| | Web (`modules.d/edge/chat.html`) | **Reference implementation — alpha.2** | | macOS Swift GUI (`gui/Eldric/`) | Pending rewire to the 5.0 Edge endpoints | | iOS (`EldricMobile/`) | Pending rewire | | Android | Not started | | Windows desktop | Not started | Until the Phase-4 Eldric Client SDK lands, each client reimplements the 13 endpoints manually against the spec in EldricOS.md. ## Commits that ship in this release (22) **Control-plane arc (10 commits):** ``` 84249048 5.0 kernel: edge chat shell + plugin toggle model ffa730ee 5.0 edge: persist session plugin toggles + ship sharing.dream 290637bd 5.0 edge: cluster-side conversation history + session.local opt-out db4c79ea 5.0 edge: wire chat shell to cluster-side conversation store 58446d83 5.0 edge/data: data-plugin fan-out → retrieval syscall → cloud.chat 1c342378 5.0 edge: extension-backed data plugins — arxiv + tool.web eea59307 5.0 edge: auto-surface loaded extensions as data/tool plugins 88ecf44a 5.0 dream: wire sharing.dream plugin → per-user profile snapshot adbaf43b 5.0 edge: persist selected model cluster-side 33eb325a 5.0 docs: alpha.2 release notes + native-client wiring spec ``` **Completion + hardening arc (12 commits):** ``` fe753b7e 5.0 dream: cycle resolves opted-in scopes at ingest 298d5bc2 5.0 cloud: echo (debug) backend — zero-config UX demos ac5bef3d 5.0 test: alpha.2 smoke harness — 33 checks, ~10s, stdlib only 41fbdd75 5.0 edge: session.local scrubs the dream profile too 23c43d6e 5.0 dream: ingest + checkpoint phases wired to syscalls 2fa7d8fa 5.0 edge: conversation pagination — ?before= + ?limit= 1f413c70 5.0 edge: SSE event stream at /api/v1/edge/events 0ff87a7a 5.0 kernel: /api/v1/catalog enriches with install_state 93dbe522 5.0 sdk: NASA APOD reference extension — 4.x→5.0 port template a466e32d 5.0 agent: fall back to cloud.chat when inference.chat absent f583b9d4 5.0 edge: conversation search + markdown/json export 03209e0b 5.0 edge: plugin/conversation/chat usage counters → /metrics f5600f2d 5.0 kernel: extension manifest schema — stricter validation 1424b1de 5.0 scripts: deploy-repo.sh (saved, not runnable yet) 3bdb9111 5.0 edge: conversation rename + audit trail for privacy toggles ``` ## Spec-audit closure arc (8 additional commits) Re-reading EldricOS.md against the code surfaced four Phase-1 deliverables that were compiled-in but not exposed, plus ops runbook + repo gaps. All eight commits validated by the harness. ``` 7999f250 5.0 data: expose Matrix Memory maintenance + observability surface 6791db29 5.0 kernel: rich Brain tool metadata on extension manifests 2aa1d558 5.0 kernel: relax manifest category whitelist + auth smoke test 60c2be80 5.0 router: HTTP fallback for headless controller discovery 81808042 5.0 docs: kernel.installing — repo.eldric.ai admin runbook a3a666f6 5.0 docs: kernel.smoke-test — harness runbook for ops c8df5c43 5.0 test: ELDRIC_EDGE_BASE — point smoke harness at a remote install d740ed5e 5.0 kernel: enrich /health with version / uptime / auth / module_count d04a7c92 5.0 packaging: bump to 2.alpha2 release + fix tool_specs YAML→JSON fidelity a32f6b5b 5.0 docs: audit help cross-links + fix 2 broken See-also refs ``` **repo.eldric.ai live.** `https://repo.eldric.ai/5.0/` serves - `eldric-aios-5.0.0-2.alpha2.fc43.x86_64.rpm` (1.4 MB CPU baseline) - `eldric-aios-cuda-5.0.0-2.alpha2.fc42.x86_64.rpm` (512 MB GPU add-on, `GGML_CUDA=ON`, sm_75..89) both signed with a 4096-bit RSA GPG key (`repo@eldric.ai`, 5-year expiry). The CPU RPM declares a `Supplements:` weak dep on `cuda-drivers`, so `dnf install eldric-aios` on a GPU host auto-pulls the cuda add-on and the `/usr/bin/eldric-aios` wrapper execs the CUDA binary when `/dev/nvidia0` is present. Full round-trip validated end-to-end on a fresh Fedora 42 + RTX 4070 Ti host and a fresh Fedora 43 CPU-only host. **Smoke harness: 86 checks + a sibling help-link auditor.** `scripts/audit-help-links.py` spawns a short-lived kernel, scrapes the topic registry, walks every `.help.md` for broken `See also` refs. CI drop-in next to the primary harness. - **Matrix Memory surface** — `/api/v1/memory/{health,matrices, checkpoint,verify}` wired to the already-compiled `MatrixMemoryService`. The class shipped in alpha.1 but the HTTP surface had only store + recall. - **Rich Brain tool metadata** — `tool_specs:` map on manifests carrying `description`, `input_schema` (JSON Schema), `examples`, `usage_policy`, `resource_hints` per tool. Per EldricOS.md §7982 the Brain needs these to learn tools at runtime without retraining. Surfaced on `extension.list`. - **Manifest category relaxed** — category still required but no longer enum-gated; legacy names (`inference_adapter`, `oa_papers`, `private_knowledge`, `customer_backend`) load cleanly, name-traversal still blocked. - **Router HTTP fallback** — headless `eldric-routerd` on a separate host now discovers workers via `GET /api/v1/workers` when the syscall isn't loaded. - **`/health` enriched** — `version`, `uptime_s`, `auth`, `module_count` in one round-trip. - **`ELDRIC_EDGE_BASE=`** — harness runs against any remote install; skips the four filesystem-bound tests, exercises the other 18 identically over HTTP. - **repo.eldric.ai live** — on 10.19.0.31, TLS via Let's Encrypt, GPG-signed RPMs (4096-bit RSA `repo@eldric.ai`), 7 baseline extension manifests under `/catalog/`. `dnf install eldric-aios` works on any Fedora 42+ host. Admin runbook at help topic `kernel.installing`. - **`kernel.smoke-test` runbook** — 16-group table of what the harness actually exercises + how to run remote-mode. Harness grew 66 → **81 checks** (`scripts/smoke-test-alpha2.py`, ~15 s, stdlib only, CI-drop-in). ## Extended additions after the initial release-notes commit The first alpha.2 notes covered the control-plane arc. The hardening / completion arc above adds (all validated by the `scripts/smoke-test-alpha2.py` harness — 81 checks, <15 s): - **Dream cycle actually consumes the profile.** `trigger_cycle()` resolves every opted-in session's scope at ingest and fires `retrieval.` per enabled data plugin; `checkpoint` writes consolidated payloads to matrix memory only for scopes with `sharing.results` enabled. Phase-3 just replaces the still-stubbed extract/probe/distill phases with LLM calls. - **Echo (debug) backend.** Pick model `eldric-echo` and `/api/v1/edge/chat` pretty-prints what the LLM would receive (including the retrieval-context system message). Zero deps. - **Session persistence completeness.** `session.local` opt-in now also scrubs the dream profile; model choice persists cluster-side (`selected_model` in the session file). A user moving between workstations finds their plugins, their model, their conversations, and their dream state exactly as they left them. - **Live catalog reload.** Edge serves SSE at `/api/v1/edge/events`; kernel publishes `extension.loaded` / `extension.unloaded`. Clients see newly-loaded data plugins without a page refresh. - **Conversation pagination, search, export, rename.** Full CRUD surface for long chat histories — `?before=`, `?limit=`, `?q=`, `?format=md|json`, `PATCH {id,title}`. - **Catalog enrichment.** `/api/v1/catalog` now reports `install_state` per entry (`not_installed` / `available` / `loaded`) so dashboards render install/enable/disable with one request. - **Agent fallback.** `agent.invoke` tries `inference.chat` first, falls back to `cloud.chat`; default model is `eldric-echo` so a zero-config install answers out of the box. - **Usage metrics.** `/metrics` now carries `eldric_edge_plugin_toggles_total`, `eldric_edge_conversations_created_total`, `eldric_edge_chat_requests_total` with plugin labels but never user content. - **Audit trail for privacy toggles.** `session.local` and `sharing.*` flips land in the hash-chained audit log; ephemeral toggles (`data.*`, `render.*`, `tool.*`) don't. - **Manifest schema tightening.** Stricter validation in `parse_manifest` — required version + category, name safety charset (blocks traversal), Model A entry_point required, license tier whitelisted. Bad manifests rejected at scan rather than at load. - **Reference extension template.** NASA APOD ports as the reference 4.x→5.0 extension — ~60 LOC Python Model B, fulfils the `retrieval.` contract, auto-surfaces as `data.nasa_apod` when loaded. Template for the rest of the 4.x science surface. - **Smoke harness.** `scripts/smoke-test-alpha2.py` — 66 assertions across every path this release touches. Stdlib only. CI-drop-in. - **Deploy script.** `scripts/deploy-repo.sh` for `repo.eldric.ai` (10.19.0.31, `/data/repo/eldric`). Saved for when NAT + DNS land; the pre-flight SSH probe fails fast until then. ## Upgrade from alpha.1 Drop-in. The `/data/eldric` layout is compatible — alpha.2 adds two new subdirectories (`edge/sessions/`, `edge/conversations/`, plus `dream/profiles/` if any user opts in) but doesn't touch existing alpha.1 paths (`vectors/`, `memory/`, `models/`, `cloud/`, `sources.jsonl`, `extensions/`). RPM layout unchanged. `dnf upgrade` picks up the new baseline chat shell at `/usr/share/eldric/modules.d/edge/chat.html`; overlay customisations continue to work. Existing alpha.1 clients (OpenAI SDK pointing at `/v1/chat/completions`) keep working — that path stays stateless and OpenAI-compat. Session-aware behaviour is opt-in via `/api/v1/edge/chat`. ## What's still Phase-3 - **Dream cycle body** consuming the profile: per-user ingest scoped to enabled data plugins, publish gated by `sharing.*`. - **Identity service** behind `/api/v1/edge/session`: the current profile is hardcoded `anon/default/viewer`. Phase-3 wires real users, tenants, projects, workgroups, and plugs them into the scope plumbing that's already in place. - **Scoped skills / agents** — the Phase-3.5 extension that turns user-authored skill and agent declarations into additional plugin categories (spec in `dream.scoping`). - **Eldric Client SDK** in Swift / Kotlin / Rust / TypeScript so each non-web client is ~100 lines of UI glue. - **Authentication policy default flip** — `--auth=enforce` still ships off by default; alpha.3 ships an `eldric-setup` wizard that provisions the first admin token.