omnius 1.0.856 → 1.0.857

package/docs/work-orders/WO-automatic-capacity-recovery-and-room-history-20261013.md

@@ -0,0 +1,134 @@
+# Automatic capacity recovery and consecutive room history
+
+Status: streamed-source recovery is implemented and verified through the
+isolated installed package; default eager-source transport preservation is
+under repair. Supersedes treating ordinary context pressure as an acceptable
+terminal result in the complete-source work order. Publication and active-runtime
+update remain with the operator.
+
+Room-history repair: implemented and verified. Consecutive completed requests
+reuse the same chronological block for the exact room/topic identity while the
+block remains at the scrollback tail and the idle gap is below 30 seconds.
+Other rooms, unrelated appended output, overlapping work and Stop break the
+sequence. Every inference retains its own callback identity, request boundary,
+output, failure history and usage; stale callbacks cannot modify a later request.
+Read mode remains available for the entire retained chronology.
+
+Verification: 67 selected TUI/registry/status-bar tests and CLI typecheck pass.
+Two unrelated broad observability assertions reproduce against the prior bridge
+and remain unchanged; they are not counted as passing.
+
+## Accepted requirements
+
+1. Preserve the complete original request, constraints, source identity and
+   evidence. Whole-source custody alone does not establish whole-source reading.
+2. Recover automatically from context pressure: release owned duplicates,
+   organize evidence hierarchically, and choose an authorized eligible model
+   with sufficient actual context before declaring an external blocker.
+3. Persist processing coverage and resumable progress. Recovery must change
+   capacity, representation, output allowance or route; never replay the same
+   failing request indefinitely. Preserve Stop, ownership, privacy and explicit
+   fallback settings.
+4. Handle large HTTP sources through durable streaming storage instead of an
+   arbitrary small successful-prefix or immediate size rejection. Retain exact
+   source bytes and verify the completed retrieval before granting source-read
+   evidence. Resource exhaustion must identify the real external limit.
+5. Keep every requested facet, contradictions, citations and unresolved details
+   through hierarchical evidence processing. Complete archived findings remain
+   recoverable; a reduced parent view cannot manufacture support or fulfillment.
+6. Validate the original constrained 32k RFC task through the installed native
+   Telegram path with real model/catalog/source responses and a complete cited
+   final answer. Distinguish simulated transport ACKs from actual Telegram sends.
+7. Consolidate consecutive TUI inference activity for the same Telegram room
+   into one chronological block. An interlaced room/activity or meaningful time
+   break starts a separate block. Preserve request identity, errors, read-mode
+   history, scrolling and room/topic separation.
+8. Present the bundled public models as built-in free inference in user-facing
+   model, route and status text. Keep actual provider identities in machine
+   audit receipts so transport failures remain diagnosable.
+9. Apply known endpoint/model schema restrictions before the first request.
+   Preserve validation against the original host contract. A transport adapter
+   must not weaken required fields, exclusivity, ownership or effect checks.
+10. Distinguish accepted task output from acknowledged Telegram delivery. Retry
+    a proven preconnection failure using the same immutable saved operation;
+    retain ambiguous sends for reconciliation without regenerating the task.
+    Public progress must reflect preparation, sending, retrying, unconfirmed
+    delivery or complete acknowledgement.
+
+## Implementation and acceptance record
+
+- Complete HTTP bodies stream to verified, indexed immutable files. The eager
+  threshold selects storage representation; it does not truncate or reject the
+  remaining response. CAS imports and extraction read bounded exact ranges.
+- Source workers automatically cover the finite source under the original full
+  request. Large retained findings use lossless encoding first, then isolated
+  hierarchical reduction with an independent preservation review and resumable
+  checkpoints. Coverage and reduced prose do not establish factual support.
+- Context negotiation binds a prepared remote route to the exact packet and
+  owner. Working local capacity remains unchanged; an explicit hard ceiling
+  remains binding. Consumed preparation tokens never reach child backends.
+- Installed validation found a second capacity disconnect: a fitting primary
+  packet could fall back successfully to a larger remote model without a prior
+  preparation receipt, leaving the following source read at 32k. The first
+  streamed run accounted for all 109,910 characters with HTTP 200 responses,
+  but exhausted its unchanged 32-generation budget reducing/reopening findings.
+  It failed acceptance. The repair must bind actual successful route capacity
+  to that response's exact returned tool calls, including ordinary fallback;
+  this receipt cannot authorize a later model dispatch or enlarge local context.
+- The installed live test must exercise actual catalog, source and model
+  responses at the original 32k working window. Its forced file-storage mode
+  tests the production streaming path on the complete RFC; separate real HTTP
+  producer tests exercise sources larger than 32 MiB. Neither implies a live
+  Telegram API send: transport acknowledgements are simulated in scratch state.
+- The latest production failure is a saved seven-item research answer followed
+  by an unacknowledged final send (`fetch failed`). Progress edits succeeded.
+  The old log omitted the transport cause, so that operation remains ambiguous;
+  the repair must not retroactively infer that it was never sent.
+- Final isolated recursive build, artifact audit (586 files, zero warnings),
+  compiler-free fresh npm install and 581 focused checks pass: execution 81,
+  orchestrator 391, delivery/activity 92, room continuity 13 and retained
+  timeout/rate-limit/Stop/5xx delivery invariants 4.
+- Installed bundle SHA-256:
+  `bf6ed73b5034f8d1463eb2cd93fabd2e3d5476e151ddb9760a1ed9a0a9c977cc`.
+  The original 32k working-window test used actual catalog, RFC retrieval and
+  model replies: 11 generations, all HTTP 200, two source cores covering all
+  109,910 characters, all three cited details, fulfilled goal, consumed intake
+  and complete simulated Telegram acknowledgement. Mandatory versus advisory
+  wording was manually checked in the actual unmodified answer.
+- A separate installed hierarchy run made three real HTTP 200 requests:
+  extraction, organization and independent preservation review. It retained
+  all six findings and exact caveats/Unicode/numbers/code in 1,802 characters
+  under a 1,900-character parent budget; checkpoint reopen required no inference.
+  Its fixture budget is selected below the genuine lossless representation so
+  concise extraction cannot accidentally skip the hierarchy being tested.
+- Known schema restrictions were applied before the first wire request; no
+  400/unsupported-schema repair occurred in the successful installed runs.
+  Curated receipts are in `evidence/automatic-capacity-stream-installed-20261013.json`
+  and `evidence/source-hierarchy-installed-20261013.json`. The failed prior
+  candidate remains explicitly recorded in
+  `evidence/automatic-capacity-before-source-route-20261013.json`.
+- Delivery transport regressions exercise the actual Bridge fetch boundary and
+  durable operation ledger. No actual Telegram message/typing call, GPU model
+  request, publication or active-session restart was performed by this repair.
+- Additional default-storage live acceptance exposed an older memory-compiler
+  handoff: a complete eager view admitted under the successful remote route was
+  replaced with a memory artifact reference before model transport. Repeated
+  fetches did not fix that missing input. Pending repair must preserve the exact
+  host-owned unread source until its successful model input, then permit normal
+  historical archiving; no marker-based exemption or global compaction disable.
+
+## Delivery
+
+Use scoped changes and commits, preserve existing unrelated working-tree and
+publication changes, rebuild and verify the isolated installed package, fetch
+the tracked branch and push verified commits. The operator publishes and
+updates the active runtime.
+
+## Prior failure to reproduce
+
+The complete-source 32k installed test accounted for every one of 109,910
+characters, then rejected a final request requiring 27,282 input tokens,
+1,967 safety tokens and 4,096 output tokens: 33,345 total. The runner rejected
+before model transport, preventing the normal remote route selection from
+recovering. The former test recorded an honest incomplete result; the new
+acceptance requires automatic recovery and completed final delivery.

package/docs/work-orders/WO-computer-use-loop-reconciliation-20261013.md

@@ -0,0 +1,104 @@
+# Reconcile legacy desktop tools with the rapid computer-use controller
+
+Status: reconciliation implemented; isolated build, focused regressions and real
+read-only desktop observation/capture verified. Live Kev/Moondream inference
+and visual input acceptance remain limited by the missing profile and host GPU
+health described below. No claim that every application/control is supported.
+
+## Observed production disconnect
+
+The current terminal run in `telegram_test`,
+`tui-3144597-mv5p6bvu-1791939658797-1`, was asked to open Firefox and search
+archive.org. At 2026-10-14T01:04:57Z it invoked `vision_action_loop` with
+`action=observe`. The result captured the desktop using `gnome-screenshot`, ran
+Tesseract OCR, reported `Vision analysis unavailable`, and performed no input
+action. The next recorded tool was a todo update. This is not evidence that
+Kev, Moondream, or the new rapid decision/executor loop ran.
+
+## Accepted work
+
+1. Identify the canonical new controller and trace the legacy tool, registration,
+   desktop discovery, prompt guidance and configured model routing into it.
+2. Reconcile existing desktop entry points with the new controller. Ordinary
+   authorized desktop tasks must discover the intended semantic observation,
+   finite-action policy, visual grounding and verified executor rather than an
+   independent OCR-only observe/planner loop.
+3. Preserve OS focus, exact current target identity, task authority, Stop and
+   public/private tool boundaries. An observation or tool success cannot claim
+   an action, model invocation or verified goal progress.
+4. Report actual policy/vision availability and invoked model identities. Use
+   the configured Kev/alternative decision and Moondream adapters when required;
+   semantic fast paths may correctly avoid model calls. Do not silently load a
+   different checkpoint or allocate an unintended GPU.
+5. Cover the actual runtime registration/discovery and legacy-to-controller
+   transition. Perform live model/desktop acceptance only after the repository's
+   exact endpoint, model and accelerator preflight; distinguish offline fixtures
+   from actual GUI actions and inference.
+
+## Scope and delivery
+
+Continue the already authorized capacity, source-processing and Telegram
+delivery repairs. Keep this migration separately reviewable and committed after
+its own checks; preserve unrelated work and the operator's publication staging.
+The operator publishes and updates the active runtime.
+
+Architecture anchor:
+`/home/roko/Documents/open_weight_rapid_computer_use_architecture_2026-10-11.md`.
+
+## Reconciled runtime
+
+- TUI, its subagents and Telegram admin construct one configured native tool per
+  scope and pass the same instance to the legacy wrapper. Browser and native
+  factories share operator-controlled pinned policy/vision clients, preflight
+  gates, request quarantine and explicit capability/model receipts.
+- `vision_action_loop` now returns canonical native observations and actionable
+  handoff to `native_ui`. Its old coordinates/keyboard/allow-actions parameters
+  cannot dispatch input. The previous manual utility is unregistered and absent
+  from default discovery. Existing camera/image tools retain their own purposes.
+- Desktop discovery and every shared model-tier route select `native_ui`;
+  browser UI selects `playwright_browser`. Telegram and visual-sensor guidance
+  no longer recommends the retired OCR loop or global keyboard injection.
+- Unique semantic controls skip inference. Ambiguous candidates can use pinned
+  Kev; missing semantic targets can use pinned Moondream. Selection/grounding
+  cannot dispatch: a separate host-admitted step checks the exact task effect,
+  current window/control, Stop and observed outcome.
+- Native visual capture matches AT-SPI's active PID/bounds against actual X11
+  XID/PID/geometry before and after capture. Input rechecks that binding, moves
+  to the finite point, verifies actual pointer arrival and rechecks ownership
+  before a targeted click. Accessibility IDs are never cast to XIDs and titles
+  never establish window identity. Capture failure retains real semantic state.
+- This visual input adapter targets Linux X11. Existing native semantic adapters
+  remain available; unavailable visual backends are explicit. There is no new
+  global raw-key fallback or claimed universal visual control.
+
+## Verification
+
+The CU-only isolated recursive build and artifact audit pass. Its focused checks
+pass 261 cases: execution 164, CLI runtime/actual Telegram discovery/guidance 60,
+and shared routing/discovery 37. The model transport tests use fixtures and are
+not counted as live inference.
+
+The final isolated bundle SHA-256 is
+`359772f23c1dbc48c3cac29253ba57115b7f024650180b8268997d2729a830da`.
+Its tarball audit passes with 588 files and zero warnings. A fresh global npm
+install also passes, including the required SQLite prebuild without a native
+compiler. The isolated package was not published; the operator owns publication.
+
+Actual read-only OS validation observed 84 AT-SPI nodes and captured the active
+786x557 window with verified X11/accessibility geometry. It made zero model
+requests, input actions or service changes. Both model capabilities correctly
+reported `not_configured`. Receipt:
+`evidence/native-reconciliation-readonly-live-20261014.json`; repeatable probe:
+`scripts/live-native-reconciliation-harness.mjs`.
+
+## Current runtime preflight
+
+Read-only checks at approximately 2026-10-14T01:15Z found no configured
+`~/.omnius/computer-use-runtime.json` and no resident Kev or Moondream process.
+GPU0 is an A100 80 GB with 6 MiB allocated; GPU1 hosts the existing voice
+workloads. `docker gpu discover` reports `gpu_health.admission_blocked=true`
+because selected GPU2 is missing and reports recovery required. The scoped
+lease policy therefore does not currently permit starting fresh model workloads.
+No GPU lease, model load, inference request, reboot or service change was made.
+This limits live model acceptance; it does not prevent implementing and testing
+the shared runtime, canonical discovery and guarded execution contracts.

package/docs/work-orders/WO-escape-crash-and-gpu-incident-20261013.md

@@ -0,0 +1,79 @@
+# Escape exit and NVIDIA device incident
+
+Status: parallel read-only investigation complete. GNOME Terminal crashed in
+GTK; one A100 is unqueryable and the broker correctly blocks new CUDA work.
+The exact GTK trigger and physical GPU failure cause remain unresolved. The
+computer-use reconciliation is delivered separately; a broad application
+refactor is not established as a repair for either failure.
+
+## User report and confirmed observations
+
+- User reports pressing Escape in the `telegram_test` TUI crashed the terminal
+  and that the A100s appeared to disappear. Requested full parallel agent
+  investigation and repair.
+- Process 3144597 is no longer present. Its interruption lifecycle records
+  `stopped`, `all owners acknowledged stop`, at 2026-10-14T01:30:11.580Z.
+- `~/.omnius/crash.log` has an approximately 01:30Z modification time but its
+  last recorded exception is an older 03:13Z packaging verifier entry. File
+  modification time is not evidence of a current exception or its cause.
+- At 01:15Z, before Escape, GPU2 was already unqueryable. At 01:34Z, GPU0 and
+  GPU1 still enumerate as A100 80 GB; GPU1 retains the existing voice workloads.
+  The broker reports a missing selected GPU and recovery required, blocking new
+  CUDA admissions. No current evidence establishes Escape as its cause.
+
+## Parallel investigations
+
+1. Trace Escape, Stop, cancellation ownership, rendering and process exit.
+   Recover actual terminal/error/exit evidence; distinguish an exception,
+   signal, intentional exit, terminal failure and OS resource exhaustion.
+2. Inspect NVIDIA/PCIe/kernel events and actual current device/runner placement.
+   Establish first fault time and affected UUIDs before proposing recovery.
+3. Inspect broker/service/container state, GPU health admission and existing
+   leases. Correlate endpoint failures and workload events with the other logs.
+
+## Repair and delivery constraints
+
+Preserve active workloads and durable user tasks. No GPU reset, reboot,
+unrelated workload termination or CUDA/model load during diagnosis. Every CUDA
+deployment follows the host broker policy and exact hardware preflight.
+Implement a scoped verified application repair when the cause is established;
+do not substitute a speculative broad rewrite for a reproducible fix.
+Checkpoint current native-loop, source and delivery work, then complete and
+push each coherent repair without including unrelated publication changes.
+
+## Established findings
+
+| Finding | Evidence | Limit |
+| --- | --- | --- |
+| GNOME Terminal's server received SIGSEGV | `/var/crash/_usr_libexec_gnome-terminal-server.1000.crash`, PID 6967, GNOME Terminal 3.52.0-1ubuntu2, dated 2026-10-13 18:30:11 PDT | The faulting stack is `gtk_main_do_event` → GDK → GLib. It does not establish which input caused the GTK fault. |
+| Omnius acknowledged Stop | Interruption receipt at 2026-10-14T01:30:11.580Z, all 209 interruption owners acknowledged | Receipt is cancellation evidence, not proof that terminal rendering cannot trigger a native library bug. |
+| Actual Escape handler survives its fixture | Compiled DirectInput and the actual handler source: first Escape pauses, second aborts; zero close, exit or signal calls, and subsequent text input accepted | CPU fixture, not a reproduction in the affected GNOME Terminal. |
+| GPU2 does not answer PCI configuration reads | A100 UUID `GPU-fc9d264d-8a20-bf8c-eb72-72a6dcbdaf2a`, PCI `0000:84:00.0`; configuration reads return all FF, header 7f, invalid link state | Cached PCI/proc identity alone does not prove a functioning GPU. Physical cause is unresolved. |
+| GPU0 and GPU1 still answer individual driver queries | A100 UUIDs `GPU-3538758b-6ba7-fa1b-a80c-e315a606c3f0` and `GPU-170a99ee-850f-2182-1050-4e8d3c87b6b0`; valid PCI configuration and individual `nvidia-smi` queries | Both report recovery action Reboot. Presence and a voice HTTP health check do not prove healthy CUDA inference. |
+| GPU degradation predates Escape | Broker health recorder already degraded at 2026-10-12T20:08Z and 23:57Z; GPU2 unqueryable at 2026-10-14T01:15Z | Earlier HTTP 503s alone would not date the physical failure. |
+| Broker admission is a deliberate health fence | `/usr/local/libexec/ollama-unify-gpu-negotiator`, fresh-driver-health checks at lines 874–955; missing selected GPU or required reboot returns 503 | Bypassing admission would not repair the device. Existing voice lease remains active. |
+
+The kernel warning near 01:30:46Z names Nautilus PID 548295 closing NVIDIA
+file descriptors (`nvidia_dev_put` / `nvidia_close`). It is not a recorded
+Omnius model-loading event. No new Xid/AER, incident OOM, current Omnius
+JavaScript exception, or Omnius PID segmentation-fault entry was found in the
+examined records. This is an absence of supporting evidence, not proof that
+all possible failure records were retained.
+
+## Evidence and next repair boundary
+
+The private extracted terminal core and backtrace remain in
+`/tmp/omnius-terminal-crash-20261014/`; the core is not checked into source.
+The actual-handler receipt is `escape-handler-proof.json` in that directory.
+Sanitized GPU-only command evidence is
+`/tmp/omnius-gpu-readonly-evidence-20261014.json`, SHA-256
+`36da195241d3ef7514cafaba36aaa29fec42d0d264905933a283d5ce1a9cb48a`.
+These temporary local artifacts can expire; the durable report records only
+the findings above and their diagnostic limits.
+
+For the terminal, retain the native crash report and reproduce the GTK event
+fault before choosing a terminal package repair or a narrowly evidenced TUI
+change. For GPUs, coordinate release of the active voice workload and host
+recovery before any reset/reboot, then verify each designated device and the
+broker's health admission before model loads. No reset, reboot, service change,
+CUDA allocation, or workload termination was performed during this investigation.

package/docs/work-orders/evidence/automatic-capacity-before-source-route-20261013.json

@@ -0,0 +1,72 @@
+{
+  "schema": "omnius.live-telegram-public-research.v1",
+  "bundleSha256": "5ec7e474ca581ce61f1773bec7a8723167d31ad74223e1711c4233968cf1b2dc",
+  "startedAt": "2026-10-14T00:56:45.290Z",
+  "finishedAt": "2026-10-14T01:01:33.584Z",
+  "bounds": {
+    "physicalGenerations": 32,
+    "contextWindowTokens": 32768,
+    "lifecycleMs": 300000,
+    "sourceUrls": [
+      "https://www.rfc-editor.org/rfc/rfc9112.txt"
+    ],
+    "sourceStorage": "file-backed host storage policy"
+  },
+  "result": {
+    "intakeLifecycle": "failed",
+    "runnerStatus": "incomplete",
+    "terminalReplyDisposition": null,
+    "admittedWebFetchReceiptCount": 1,
+    "sourceUrlsRead": [
+      "https://www.rfc-editor.org/rfc/rfc9112.txt"
+    ]
+  },
+  "physicalGenerations": 32,
+  "attemptedGenerations": 33,
+  "passed": false,
+  "state": "failed_acceptance",
+  "blocked": [
+    "Generation budget exhausted"
+  ],
+  "httpStatuses": [
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200,
+    200
+  ],
+  "completeSource": {
+    "chars": 109910,
+    "sha256": "e4f426bac6206b67fdf9e0da826154f70588db2133a0a86b15cde4ff725d8937",
+    "processedChunks": 6,
+    "totalChunks": 6
+  },
+  "diagnosis": "Actual successful nonprepared remote fallback did not expose its capacity to the source-read guard. Full source coverage still ended in parent reduction/reopen loops and exhausted the unchanged generation budget. This candidate failed acceptance."
+}