omnius 1.0.813 → 1.0.816

package/docs/DISCOVERY.json

@@ -29473,6 +29473,42 @@
         }
       ]
     },
+    {
+      "id": "guide.engineering-kolibri-placement-and-empty-response-20261007",
+      "kind": "guide",
+      "title": "Kolibri placement and response investigation",
+      "summary": "Read-only runner placement, broker journals, and request metadata on October 7, 2026 established that robit/kolibri-1:q4km loaded 81/81 layers across the two operator-selected A100s at PCI 01:00.0 and 82:00.0. Its native context was 262,144 tokens. Both cards showed the same runner, roughly 37 GB allocated per card, and active utilization during generation. ",
+      "keywords": [
+        "engineering",
+        "kolibri",
+        "placement",
+        "and",
+        "empty",
+        "response",
+        "20261007",
+        "md"
+      ],
+      "maturity": "stable",
+      "audiences": [
+        "user",
+        "integrator",
+        "coding-agent"
+      ],
+      "layer": "documentation",
+      "interfaces": [
+        {
+          "type": "file",
+          "target": "docs/engineering/kolibri-placement-and-empty-response-20261007.md"
+        }
+      ],
+      "references": [
+        {
+          "type": "documentation",
+          "target": "docs/engineering/kolibri-placement-and-empty-response-20261007.md",
+          "relation": "canonical-artifact"
+        }
+      ]
+    },
     {
       "id": "guide.engineering-telegram-task-context-repair-20261007",
       "kind": "guide",

package/docs/DISCOVERY.md

@@ -428,6 +428,7 @@ Daemon equivalents are `GET /v1/discovery/bootstrap`, `GET /v1/discovery?q=<inte
 | `guide.discovery-markdown` | Omnius Discovery Catalog | Human-readable generated index of every discoverable Omnius capability and bundled documentation artifact. |
 | `guide.duplicate-calls-root-cause-analysis` | Duplicate Tool Calls — Root Cause Analysis | The agent repeatedly calls the same tool with the same arguments (e.g., fileread on the same path, grepsearch with the same pattern, shell with the same command). This wastes turns, burns context, and creates visible "looping" behavior. |
 | `guide.duplicate-calls-root-cause-deep` | Duplicate Tool Calls — Deep Root Cause Analysis | The agent repeatedly calls the same tool with the same arguments right after having already called it. This is visible even in this agent's own conversation — fileread and grepsearch calls get duplicated within 1-2 turns. |
+| `guide.engineering-kolibri-placement-and-empty-response-20261007` | Kolibri placement and response investigation | Read-only runner placement, broker journals, and request metadata on October 7, 2026 established that robit/kolibri-1:q4km loaded 81/81 layers across the two operator-selected A100s at PCI 01:00.0 and 82:00.0. Its native context was 262,144 tokens. Both cards showed the same runner, roughly 37 GB allocated per card, and active utilization during generation.  |
 | `guide.engineering-telegram-task-context-repair-20261007` | Telegram task context repair: evidence, research, and acceptance contract | Date: 2026-10-07. Status: design induction for an isolated repair; this document does not certify live Telegram behavior. |
 | `guide.ephemeral-skill-pack-small-context` | Ephemeral Skill Pack Small-Context Plan | AIWG and Omnius skills are valuable, but full SKILL.md bodies are too large for small and medium context windows. The main agent should see only a tiny, task-scoped skill manifest, then delegate full skill unpacking to a sub-agent that returns a compact extraction. |
 | `guide.explorations-context-window-todo-association` | Exploration: Associating Context-Window Entries with Todo-Scoped Run Periods | Goal (from user): Understand how granular context-window entries (messages / tool results) are associated with periods of an agentic run, and specifically whether they are tied to a todo item so that, when that todo completes, its context content can be compressed. |

package/docs/engineering/kolibri-placement-and-empty-response-20261007.md

@@ -0,0 +1,113 @@
+# Kolibri placement and response investigation
+
+## Observed behavior
+
+Read-only runner placement, broker journals, and request metadata on October 7,
+2026 established that `robit/kolibri-1:q4_k_m` loaded 81/81 layers across the two
+operator-selected A100s at PCI `01:00.0` and `82:00.0`. Its native context was
+262,144 tokens. Both cards showed the same runner, roughly 37 GB allocated per
+card, and active utilization during generation. Timings reported approximately
+50 generated tokens per second. The third A100 and display GPU were not used.
+
+The observed quiet periods included avoidable model replacement. Optional
+semantic-memory embeddings evicted a freshly warmed exclusive Kolibri group;
+Kolibri then had to load again. Later SNR relevance/noise evaluators requested
+the previous Qwen model because their engine retained the configuration object
+from before the parent's model selection changed. This was distinct from a
+failed tensor split, GPU fault, or Telegram model override.
+
+At the initial placement observation, the installed daemon and TUI had adopted
+version 1.0.813. A later automatic update adopted 1.0.814 with the same older
+CLI bundle; neither version contained the repairs described here. The affected
+news request originated in the TUI; the earlier daemon GUI session was not its
+request owner.
+
+## Repairs and acceptance boundaries
+
+- Resolve the SNR parent's configuration at each evaluation batch and share
+  that snapshot between its two evaluators. Model, endpoint, provider protocol,
+  credentials, and timeouts must follow the same parent selection. Existing
+  callers supplying a fixed configuration remain supported.
+- Mark optional memory embeddings as broker `background`/`yield` work. The
+  broker permits warm reuse and genuinely free capacity, but returns a
+  retryable `background_capacity_deferred` response before any resident lane
+  would need reclamation. Durable memory and text retrieval remain available
+  when embeddings are unavailable. Foreground and explicit background/wait
+  requests retain their existing admission behavior.
+- Preserve the actual six-empty-response terminal cause in a typed,
+  run/epoch-bound failure receipt. Successful read-only observations are not
+  command receipts, and their absence must not replace a known model-response
+  failure with a generic verification diagnosis. Failed checks remain failed;
+  an empty response never constitutes an accepted answer or completion.
+
+The broker change preserves exclusive peer groups, GPU health checks, scoped
+leases, context accounting, and reservations until the entire owning process
+group exits. Neither repair lowers context to avoid a failure or redirects
+inference to another GPU.
+
+## Remaining causal uncertainty
+
+The news run executed successful web tools and subsequently received six
+responses without usable visible text or tool calls. Existing journals retain
+token counts and timings but not the raw response envelopes. They cannot
+establish whether the model emitted hidden reasoning, control-only output, or
+output lost during parsing/normalization. A positive generated-token count is
+not evidence of a visible answer.
+
+Regression tests establish the repaired host behavior using isolated CPU
+fixtures and mocked transports. They do not certify a successful live news
+answer or completion of the earlier Telegram PDF task. A broker restart and
+supported client update are separate adoption boundaries; source commits and a
+fresh publication artifact alone do not update already running processes.
+
+## Later Telegram PDF findings
+
+The subsequent admin DM used the verified 262,144-token context with roughly
+20,000 estimated input tokens. Its budget nevertheless declared compaction
+eligible: a roughly 217,000-token optional generation ceiling was counted as
+occupied context, and the compiler then tried to reduce total occupancy to a
+fixed fraction of the window. The fixed output allowance alone made that
+target impossible. Generation allowance is not accumulated input; compaction
+pressure and its target must instead use input, the existing minimum reply
+allowance, admission safety, and the verified context capacity. The final
+transport admission guard remains independently authoritative.
+
+Two further host boundaries were implicated. Capability discovery satisfied a
+generic observation criterion before a PDF was read, so the objective
+controller blocked the actual extractor as delivery due. A deictic native reply
+also lost its exact referenced upload batch when the durable objective was
+created, leaving a coarse contract instead of per-source work. Preparation
+receipts must remain auditable without standing in for requested source
+evidence; native attachment lineage must reach both the task contract and the
+runner request.
+
+A later numerical reply was delivered despite an incomplete-verification
+terminal record, failed extraction evidence, and no PDF extraction in that
+run. Independent CPU extraction of all 23 hash-matched source PDFs established
+that the answer was wrong, including internal arithmetic contradictions.
+Private document values and conversation identities remain outside this
+repository. Provisional provider-delivery authority must reject failed work
+evidence and blockers, while allowing a genuinely ready answer whose only
+remaining obligation is the provider send. Action drafts are not accepted
+answers merely because the model produced them.
+
+## Passive Telegram history acceptance boundary
+
+Under genuine input pressure, passive Telegram chronology should be the first
+context externalized. The current native request, selected reply ancestors,
+task contract, attachment workset, and observed task evidence remain in the
+request. Ordinary history remains available when input fits; a large optional
+reply ceiling alone must not trigger history removal.
+
+The distinction must originate in the bridge's typed context composition,
+with exact scope, source hash, and authored offsets. Quoted headings or model
+text cannot declare a historical segment disposable. Before replacing a
+segment, the runner must durably save its exact bytes, flush storage, verify
+readback, and provide an exact retrieval reference. Storage failure must retain
+the original data. Externalization changes the request working set, never the
+durable conversation, task ownership, evidence ledger, or completion status.
+
+Acceptance requires production-boundary controls for preserved unpressured
+history, actual input reduction before task compaction, unchanged current-task
+and source bytes, valid retrieval, storage failure, stale/foreign manifests,
+and recovery of full history in a later unpressured request.

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.813",
+  "version": "1.0.816",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.813",
+      "version": "1.0.816",
       "bundleDependencies": [
         "image-to-ascii"
       ],
@@ -4893,9 +4893,9 @@
       }
     },
     "node_modules/node-abi": {
-      "version": "3.97.0",
-      "resolved": "https://registry.npmjs.org/node-abi/-/node-abi-3.97.0.tgz",
-      "integrity": "sha512-6Pu8F0eXGZlyz0EepPJrWgWDcBI8rKt2Wj8exOJU1tjVbu/VyZy5ZaLYS5N+Jp/Y3IP3sIRKcIOElTu7uC6KKA==",
+      "version": "3.98.0",
+      "resolved": "https://registry.npmjs.org/node-abi/-/node-abi-3.98.0.tgz",
+      "integrity": "sha512-Llh8noNw7s15m+rmIrZnZQbe4gLyasL9loT74rC92PkmwWKeEedKvc7o6pjp8B3MhrYT/Cl846NOYZbz1wmddw==",
       "license": "MIT",
       "dependencies": {
         "semver": "^7.3.5"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.813",
+  "version": "1.0.816",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/library.js",