omnius 1.0.648 → 1.0.650

package/dist/update-worker.js

@@ -293856,6 +293856,7 @@ async function getDaemonReportedIdentity(port) {
     const resp = await fetch(`http://127.0.0.1:${p}/health`, { signal: AbortSignal.timeout(2e3) });
     if (!resp.ok) return null;
     const data = await resp.json();
+    if (data.status !== "ok") return null;
     const version = data.boot_version ?? data.version;
     if (!version) return null;
     return {
@@ -293867,15 +293868,13 @@ async function getDaemonReportedIdentity(port) {
     return null;
   }
 }
-async function getDaemonReportedVersion(port) {
-  return (await getDaemonReportedIdentity(port))?.bootVersion ?? null;
-}
 async function waitForDaemonReady(port, expectedVersion, attempts = 24) {
   let observedVersion = null;
   for (let attempt = 0; attempt < attempts; attempt++) {
     await new Promise((resolve11) => setTimeout(resolve11, 500));
-    if (!await isDaemonRunning(port)) continue;
-    observedVersion = await getDaemonReportedVersion(port);
+    const identity = await getDaemonReportedIdentity(port);
+    if (!identity) continue;
+    observedVersion = identity.bootVersion;
     if (!expectedVersion || expectedVersion === "0.0.0" || observedVersion === expectedVersion) {
       return { ok: true, observedVersion };
     }

package/docs/OMNIUS-RELIABILITY-LONG-HORIZON-TODO.md

@@ -61,11 +61,63 @@ It records scope, dependencies, evidence, and closure rules.
   - Test an abort that forbids unary fallback.
 - [x] Run focused Telegram suites.
 - [x] Build the CLI package.
+- [x] Stop treating a clean admin-chat answer as an incomplete agentic task.
+  - Root cause: the live model returned a valid visible reply on each turn, but
+    `AgenticRunner` required a `task_complete` tool call on the conversational
+    Telegram surface. It therefore repeated the generation three times and
+    committed the run as incomplete after 55,193 tokens and zero tool calls.
+  - Add an explicit `acceptTextOnlyCompletion` runner contract. Enable it only
+    for admin-DM chat; action and mutation runs retain authoritative completion
+    and evidence gates.
+  - Retain the first completed stream answer before completion-boundary
+    filtering so Telegram cannot discard the terminal conversational reply.
+- [x] Separate the Telegram admin reply from internal run diagnostics.
+  - Render only the natural answer, concise work summary, or artifact
+    description on the Reply tab.
+  - Do not prepend `Admin run`, echo the request, append a page marker, expose
+    raw exceptions, or persist completion reports as assistant conversation.
+  - Put lifecycle status and mutations in Activity, tool results in Tools,
+    completion and verification reports in Evidence, and prompt/context data in
+    Context.
+  - Return the panel to Reply when a run finishes, while marking updated
+    internal tabs for inspection.
+  - Preserve and deliver a usable draft and generated artifacts when the runner
+    misses `task_complete`; keep the incomplete report in Evidence.
+- [x] Add Telegram reply-versus-evidence regressions.
+  - Assert exact clean Reply rendering for completed and failed panels.
+  - Assert that the literal `back at it?` incomplete-run failure delivers only
+    its usable conversational draft and stores `INCOMPLETE` behind Evidence.
+  - Assert that a one-turn admin chat delivery and persisted assistant history
+    contain the exact model answer, not the panel report.
+- [ ] QUEUED — Make automatic generated-artifact upload receipts authoritative.
+  - Return delivered and failed artifact receipts from the automatic sender.
+  - Put raw upload diagnostics in Tools or Evidence.
+  - If an attachment cannot be delivered, replace any success claim with one
+    concise user-facing blocker without exposing a local path.
+- [x] Make Telegram intake completion depend on a delivery receipt.
+  - Require `ok: true` and a positive Telegram `message_id`.
+  - Try the existing HTML-to-plain fallback, then fail the run if both sends
+    fail or neither returns a receipt.
+  - Mark the intake consumed only after a confirmed visible delivery; otherwise
+    mark it failed for replay and diagnosis.
 - [ ] VERIFY — Run live Telegram route inference.
   - Endpoint: `http://127.0.0.1:11434`.
   - Model: `robit/ornith-1.5:9b`.
   - Required GPU: a broker-selected A100 80 GB GPU.
   - Forbidden GPU: the GT 1030.
+  - Partial live evidence: Telegram intake accepted `back at it?`, created a task
+    epoch, and completed broker-backed `/v1/chat/completions` and `/api/chat`
+    requests without the former 30-second abort.
+  - The earlier open-ledger result is now diagnosed as a conversational
+    completion-policy mismatch, not a transport or model failure.
+  - Direct live proof after the causal fix: the exact Ornith model returned
+    `TELEGRAM_CHAT_TERMINAL_OK`; the runner completed in one turn with zero tool
+    calls and 30 completion tokens while resident on an A100 80 GB GPU.
+  - Open gate: receive and deliver one fresh real Telegram message through the
+    rebuilt installed CLI. Mocked Bot API coverage is complete, but it does not
+    replace an end-to-end Telegram delivery receipt.
+  - Ready state: the rebuilt `/usr/local/bin/omnius` TUI is running in
+    `telegram_test`; its Telegram poller owns the live workspace lock.
   - Required result: valid visible route JSON.
   - Required result: nonzero generated token telemetry.
   - Required result: no 30-second abort.
@@ -96,11 +148,100 @@ It records scope, dependencies, evidence, and closure rules.
 - [x] Confirm the exact model tag `robit/ornith-1.5:9b`.
 - [x] Confirm model context length `262144`.
 - [x] Confirm model capabilities for completion, vision, tools, and thinking.
-- [ ] VERIFY — Confirm that Omnius can request broker capacity without bypass.
-- [ ] VERIFY — Confirm parallel sub-agent requests on separate host lanes.
+- [x] Replace per-request capacity polling with one FIFO scheduler and one reconciler.
+- [x] Cancel disconnected queued clients before backend inference starts.
+- [x] Add queue IDs, positions, phases, depth, age, and aggregate timing metrics.
+- [x] Make queue time exclude backend generation and time to first response.
+- [x] Warm each managed model with zero generated tokens before lane readiness.
+- [x] Verify model residency before admitting a lane.
+- [x] Canonicalize only an omitted `:latest` tag.
+- [x] Preserve separately configured model aliases.
+- [x] Replace least-recently-used idle lanes for mixed-model demand.
+- [x] Remove the system-lane block on independent managed-lane expansion.
+- [x] Scope anonymous CUDA handling to selected GPUs and broker-external processes.
+- [x] Ignore foreign-process departure and unselected-GPU churn.
+- [x] Retire only affected managed lanes after stable allocation growth.
+- [x] Protect queued timeout and client-write paths from broken-pipe tracebacks.
+- [x] Keep `/api/tags`, `/api/ps`, `/api/version`, `/api/show`, and `/v1/models`
+  available during cooperative GPU-lease drains.
+- [x] Add broker regressions for lazy three-lane scaling, FIFO, cancellation, aliases,
+  mixed models, warm readiness, queue telemetry, and CUDA process churn.
+- [x] Add the negotiator suite to the broker repository CI workflow.
+- [x] Commit broker scheduler repair locally on `main` as `4fbad0c`.
+- [x] Commit drain-safe metadata repair locally on `main` as `cd344e3`.
+- [x] Commit host-memory restart preflight repair on `main` as `4a24eb5`.
+- [x] Commit endpoint-aware queue admission repair on `main` as `4349a5b`.
+- [x] Commit the public private-lane range contract on `main` as `c98492f`.
+- [x] Deploy the repaired broker through `./ollama-unify.sh --install-safety`.
+  - Installed helper SHA-256:
+    `39458260611a80f50d87f03e7567bcdcc632ebd1af421e9d94f69521410d80cd`.
+  - `ollama-unify-negotiator.service` and `ollama.service` are active.
+  - Public port `11434` and system backend port `11436` report Ollama `0.32.13`.
+- [x] Disable the conflicting legacy `ollama-env-update.service` unit.
+- [x] Confirm that Omnius can request broker capacity without bypass.
+  - Discovery published managed ports `11437` through `11468`.
+  - Omnius created no private `ollama serve` owner during the live run.
+- [x] Confirm parallel requests on separate host lanes.
+  - Live discovery showed two concurrent `robit/ornith-1.5:9b` lanes on separate
+    A100 80 GB GPUs and a separate embedding lane.
 - [ ] VERIFY — Confirm that broker queue time remains visible to Telegram.
-- [ ] EXTERNAL — Confirm the broker repository commit on `main`.
-- [ ] EXTERNAL — Confirm the required broker service restart state.
+- [x] Confirm broker commits `4fbad0c`, `cd344e3`, `4a24eb5`, `4349a5b`, and
+  `c98492f` are pushed to `origin/main`.
+- [x] Confirm local and remote broker `main` are identical at `c98492f`.
+- [x] Confirm the required broker service restart state.
+
+## P0: Broker ownership incident root cause
+
+- [x] Identify the duplicate scheduler and port-ownership race.
+  - Root cause: an older Omnius elastic-pool worker survived independently and
+    spawned detached Ollama servers on ports `11436` through `11438` while the
+    broker installer temporarily stopped the supervised services.
+  - Result: the system Ollama service could not reclaim `11436`, and the host
+    broker opened another managed set on later ports.
+- [x] Remove the stale Omnius worker and its detached Ollama process groups.
+- [x] Route broker-managed setup and recovery through systemd.
+- [x] Fail closed when local broker discovery is malformed or temporarily absent.
+- [x] Remove the unsafe `OMNIUS_OLLAMA_UNIFY_ALLOW_POOL` ownership bypass.
+- [x] Prevent `/parallel` from detached-spawning Ollama for a broker-managed endpoint.
+- [x] Keep one supervised system Ollama owner and one broker scheduler.
+
+## P0: Runtime package and daemon integrity
+
+- [x] Replace the Zod-4-only `z.json()` runtime contract with a recursive JSON
+  schema that works with the bundled Zod runtime.
+- [x] Add nested JSON acceptance and `undefined` rejection tests.
+- [x] Make `publish/dist/launcher.cjs` executable during every publish build.
+- [x] Confirm the source artifact and `/usr/local` installed CLI have the same
+  SHA-256 and report Omnius `1.0.649`.
+  - Current installed `dist/index.js` SHA-256:
+    `9e7a830d1370c0127f38fa4e24b8b15cc9069746904d80a26e41f9ee48a59b8b`.
+- [x] Remove the stale user-service drop-in that moved the shared daemon from
+  port `11435` to the Cygnus port `11535`.
+- [x] Confirm the shared Omnius daemon is active on `11435` and the separate
+  Cygnus daemon is active on `11535`, both at `1.0.649`.
+- [x] Replace the stale `telegram_test` main-model identifier with the exact tag
+  `robit/ornith-1.5:9b`.
+- [x] Stop the diagnostic TUI and remove its dead-PID Telegram ownership lock.
+
+## P0: Startup control-plane recovery
+
+- [x] Replace boolean model availability with `available`, `unavailable`, and
+  `unknown` outcomes.
+- [x] Treat timeout, HTTP 502, and connection failure as `unknown`.
+- [x] Preserve the configured model when availability is unknown.
+- [x] Run first-run setup only after a successful catalog proves the model absent.
+- [x] Match omitted Ollama `:latest` with strict canonical equality.
+- [x] Use one coalesced names-only `/api/tags` request for startup availability
+  and readline completion.
+- [x] Keep full per-model `/api/show` discovery out of startup checks.
+- [x] Classify Telegram startup failures as retryable or terminal.
+- [x] Retry network errors, HTTP/Bot API 429, 5xx, and webhook propagation delay.
+- [x] Do not retry invalid tokens or duplicate long-poller ownership conflicts.
+- [x] Add a bounded TUI Telegram auto-start supervisor.
+- [x] Cancel pending Telegram auto-start when the operator stops Telegram.
+- [x] Use one daemon `/health` response for liveness and immutable boot identity.
+- [x] Tolerate bounded transient daemon health-read failures before recovery.
+- [x] Require daemon health payload `status: ok` before accepting identity.
 
 ## P0: Live inference hardware gate
 
@@ -115,9 +256,18 @@ It records scope, dependencies, evidence, and closure rules.
 - [x] Inspect `nvidia-smi` model process placement.
 - [x] Confirm model runner PID `2199237` on an A100.
 - [x] Confirm that no test runner uses the GT 1030.
-- [ ] VERIFY — Send the first generated-token request.
-- [ ] VERIFY — Inspect GPU placement during generation.
-- [ ] VERIFY — Record response text, usage, timing, and graph validity.
+- [x] Send the first generated-token request.
+  - Exact response text: `BROKER_OK`.
+  - Provider usage: 5 evaluated tokens.
+- [x] Inspect GPU placement during generation.
+  - The Ornith runner was resident on an A100 80 GB GPU.
+  - No Ollama runner was resident on the GT 1030.
+- [x] Re-preflight the final installed live session.
+  - The broker admitted zero-token capacity on managed lane 10.
+  - Runner PID `3179075` is resident on A100 GPU 2.
+  - No Ollama runner is resident on the GT 1030.
+- [x] Record response text, provider usage, and broker admission.
+- [ ] VERIFY — Record live ontology graph validity.
 
 ## P1: Ontology long-horizon work graph
 
@@ -298,9 +448,41 @@ It records scope, dependencies, evidence, and closure rules.
 - [x] Telegram transport, observability, link, and Bot API suites: 161 passed.
 - [x] Native Ollama transport suite: 19 passed.
 - [x] Model-broker focused suite: 6 passed.
+- [x] Model availability and first-run setup regressions: 11 passed.
+- [x] Telegram Bot API, startup retry, and routing suite: 91 passed.
+- [x] Daemon identity, singleton, and listener-reclaim regressions: 23 passed.
+- [x] CLI typecheck passed after startup control-plane repairs.
+- [x] CLI build passed after startup control-plane repairs.
+- [x] Inference visibility, terminal links, Telegram transport, and GPU policy:
+  81 focused tests passed.
+- [x] Completion authority, evidence, feature verification, and broker ownership:
+  79 focused orchestrator tests passed.
+- [x] `task_complete`, broker execution, and todo persistence: 46 focused tests passed.
+- [x] Cross-version operational-world JSON schema: 15 tests passed.
+- [x] Published and installed launchers are mode `0755` and their bundled CLI
+  hashes match.
 - [x] Rerun Telegram focused suites after the timeout repair.
+- [x] Telegram routing, inference, observability, and receipt regressions after
+  the conversational completion repair: 158 passed.
+- [x] Functional admin-chat regression proves one tool-bearing runner inference,
+  exact visible reply delivery, positive message receipt, and one persisted
+  assistant history record.
+- [x] Telegram Reply/Evidence isolation suite: 101 focused tests passed, including
+  the literal incomplete `back at it?` delivery regression.
+- [x] CLI typecheck, all workspace builds, publish generation, and publish
+  artifact audit passed after Reply/Evidence isolation.
+- [ ] VERIFY — Restart the operator-owned `telegram_test` TUI so its in-memory
+  process loads the rebuilt bundle, then confirm one fresh private DM keeps
+  diagnostics behind the buttons.
+- [x] Direct live `AgenticRunner` inference with `robit/ornith-1.5:9b` completed
+  in one turn with 5,837 total provider tokens and no `task_complete` demand.
+- [x] Legacy `maxTurns` control passed alone with a 60-second harness timeout
+  (20.9-second test body); its earlier 30-second failure occurred only under
+  concurrent test-worker contention.
 - [x] Rerun CLI typecheck and build after link support.
 - [x] Rerun all workspace TypeScript builds after the transport repair.
+- [x] Rebuild all workspaces, regenerate `publish/`, pass the publish artifact
+  audit, install the rebuilt package, and restart both user daemons.
 - [ ] VERIFY — Run live Telegram inference.
 - [ ] VERIFY — Run live ontology patch inference.
 

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.648",
+  "version": "1.0.650",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.648",
+      "version": "1.0.650",
       "bundleDependencies": [
         "image-to-ascii"
       ],
@@ -4375,9 +4375,9 @@
       "license": "Apache-2.0 OR MIT"
     },
     "node_modules/ip-address": {
-      "version": "10.5.0",
-      "resolved": "https://registry.npmjs.org/ip-address/-/ip-address-10.5.0.tgz",
-      "integrity": "sha512-R5SnVLJmgYYvf2F2ZgwSBnelz5G4q5AxIC277GDfUaNbrZKNANcBC7RHqYYePlszf4kBolVkJauG0ZjHHFh55g==",
+      "version": "10.7.0",
+      "resolved": "https://registry.npmjs.org/ip-address/-/ip-address-10.7.0.tgz",
+      "integrity": "sha512-BGFsyJd5mpXp3rK6jIdADLNgpJUK1jnjzvYF8lK+VyDab9JAmqN0YOKDdP17HlgKb2+ehPgDc8EtnRLbGCAMhA==",
       "license": "MIT",
       "engines": {
         "node": ">= 12"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.648",
+  "version": "1.0.650",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/library.js",