omnius 1.0.861 → 1.0.863

package/docs/work-orders/native-tui-renderer.md

@@ -1,93 +1,115 @@
-# Native terminal rendering migration
-
-Omnius now uses `@opentui/core@0.6.6`, the native terminal engine used by
-OpenCode. The agent/CLI remains on Node >=22; a packaged Bun 1.3.14 process
-owns OpenTUI and the physical terminal. This avoids loading Bun/Node-26 FFI in
-the existing Node backend or changing its native SQLite and supervisor runtime.
-
-The original Earth palette, default terminal transparency, header, tool cards,
-tasks area, expandable input, metrics, agent views, and session representations
-remain the component contract. This migration uses OpenTUI's imperative
-renderables, rather than copying OpenCode's application or Solid component tree.
-
-## Rendering and ownership
-
-- `native/screen.ts` composes existing formatted components into a bounded
-  retained cell surface. ANSI styling and geometry become native styled rows;
-  physical cursor/scroll-region/screen paint controls are consumed locally.
-- `native/host.ts` coalesces changes at 16 ms, sends changed rows, and permits
-  one unacknowledged frame. Updates during terminal backpressure replace state
-  instead of growing an output queue.
-- `native/worker.ts` renders native `TextRenderable` rows and acknowledges after
-  native output settles. Only allowlisted title, clipboard, and input-mode
-  effects pass to the terminal. Default/indexed color intents are preserved.
-- The parent acquires raw input before native capability queries. Protocol
-  replies reach OpenTUI, while keyboard/mouse/paste reach the editor and menus.
-  Modal listeners use the same filtered stream and restore prior listeners.
-- Inherited commands await native suspension. Resume forces a complete retained
-  frame. Stop/update/exit close the worker; failure restores terminal modes and
-  triggers the client's existing session-preserving shutdown.
-
-## History cost
-
-The old renderer rebuilt its prefix row index during streamed updates and
-cleared the entire wrap cache above 8,000 entries. The replacement uses an
-incremental Fenwick index, bounded LRU wrapping, shared dynamic-block snapshots,
-linear ANSI-aware wrapping, and viewport-only result allocation. Wheel bursts
-coalesce on local terminals as well as remote ones. End with an empty prompt
-returns to live. Queue hints occupy their own row above the input box.
-
-The recorded CPU benchmark compares commit
-`e54e7942700cf575f00b2142626168ed8c929f69` with this implementation, using 9,900
-unique styled lines and 120 append/reflow frames crossing the 10,000-line limit:
-
-| Measurement | Before | After |
-| --- | ---: | ---: |
-| Streamed append/reflow median | 43.93 ms | 0.023 ms |
-| Streamed append/reflow p95 | 48.19 ms | 0.112 ms |
-| Static history count/wrap calls during those frames | 1,196,840 | 120 |
-| Initial row-index construction | 72.44 ms | 83.77 ms |
-
-These are shared-host CPU history/layout measurements, excluding native paint,
-IPC, input latency, and inference. Initial index construction is not faster.
-
-## Validation evidence
-
-- [CPU benchmark](evidence/native-tui-history-benchmark-20261015.json)
-- [Real PTY and independent xterm screen checks](evidence/native-tui-pty-live-20261015.json)
-- [GPU placement preflight](evidence/native-tui-gpu-preflight-20261015.json)
-- [Live model streaming in the PTY](evidence/native-tui-model-live-20261015.json)
-- [Clean tarball runtime smoke](evidence/native-tui-package-smoke-20261015.json)
-
-All 264 targeted regression tests across 28 files passed, as did a clean
-workspace rebuild, publish bundling/artifact audit, and tarball content checks.
-An isolated tarball installation launched the Node CLI, opened SQLite, and
-painted/restored a real PTY using the packaged Bun/OpenTUI worker. Package install
-scripts were disabled; only Bun and SQLite were rebuilt for the smoke test.
-
-The terminal suite exercises 20,000 history lines, tool clicks, 200-event wheel
-bursts, multiline editing, menus, streaming, Escape, narrow/wide resizing,
-exclusive subprocess handoff, terminal restoration, and native-worker cleanup.
-The synthetic test sent zero inference requests. The optional live run streamed
-868 response chunks from `robit/qwen3.8-27b-obliterated-e03:27b` while scrolling
-and typing, through the same `StreamRenderer`, `StatusBar`, and native renderer.
-
-For that live run, a broker lease reserved GPU 0
-(`GPU-3538758b-6ba7-fa1b-a80c-e315a606c3f0`, A100 80 GB). A dedicated pinned
-Ollama 0.34.4 server at `127.0.0.1:11572` exposed only that UUID. A zero-token
-load preceded actual `nvidia-smi` runner placement verification and the
-generation request. The test server/model was stopped, the lease released, and
-the operator inference hold restored. The recorded endpoint is no longer live.
+# Native OpenTUI terminal frontend
+
+Every interactive terminal surface uses `@opentui/core@0.6.6`. Node owns the
+agent and services; a packaged Bun 1.3.14 process owns OpenTUI, raw terminal
+input, focus, layout, scrolling, selection and painting. The October 16
+[instrument interface](instrument-tui.md) replaces the former Earth palette with
+terminal defaults and introduces lightweight geometry and collapsible telemetry.
+Header actions, agent tabs, history, pinned tasks, expandable input, enhancement
+controls and metrics retain their functional structure.
+The implementation uses OpenTUI's [imperative native renderables](https://github.com/anomalyco/opentui/tree/main/packages/core).
+
+## Component contract
+
+`native/protocol.ts` carries semantic documents and actions. The host sends
+stable-ID content patches, coalesces updates and permits one unacknowledged
+frame. Missing acknowledgements fail visibly and restore terminal ownership.
+Keyboard, paste, pointer and capability bytes go directly to OpenTUI. The Node
+input facade receives editor changes, submissions and backend shortcuts.
+
+The worker builds native Box, Text, Markdown, Code, ScrollBox, Textarea,
+Select, TabSelect, Image and EmbeddedTerminal renderables. Menus, nested
+prompts, secret input, onboarding, memory maps, camera controls, progress,
+agent command center and Neovim share this component tree and focus model.
+External interactive programs use an explicit suspended terminal handoff.
+Machine-readable and piped output remain plain streams.
+
+There is no terminal-cell screen compositor, physical cursor painter, manual
+wrapped-row index, mouse escape parser or second raw keyboard parser in the
+application. Semantic text preserves attributes and links while following the
+operator's terminal colors; cursor/screen controls are discarded. Persisted tool data and live inference,
+shell, trajectory, action-tree, media and sub-agent data become native cards.
+Session restoration preserves semantic documents and old stored transcripts.
+
+## October 15 verification
+
+- `pnpm --filter omnius exec vitest run tests/native-components.test.ts` runs
+  actual OpenTUI under packaged Bun, including long-history scrolling and
+  concurrent editing, Unicode paste, independent stack/tree scroll positions,
+  native menus, secret masking, cards, themes, media controls, image fallback,
+  footer actions, and an embedded terminal.
+- `node scripts/benchmark-tui-history.mjs --output FILE` records those native
+  scenario timings. They include native layout/render/input work; they are
+  not a claim about per-keystroke latency or a speedup over the old renderer.
+- `node scripts/test-tui-production-terminal.mjs --report FILE --transcript FILE`
+  launches the actual CLI in a real PTY, observes the screen independently,
+  uses isolated settings/data, permits local metadata and blocks all inference.
+  Add `--entry /path/to/installed/omnius/dist/index.js` to test a clean tarball
+  installation using observation of the actual worker IPC.
+- `node scripts/test-tui-production-local.mjs --preflight FILE` validates an
+  already-resident, verified GPU/model scope, then exercises the actual CLI's
+  local answer, tool execution, streaming and Stop paths. Its transport observer
+  records request scope and test markers without replacing the backend.
+
+Earlier `native-tui-*-20261015.json` evidence describes the previous row adapter
+and a synthetic fixture. Those measurements are historical and do not validate
+this complete semantic frontend. Current results are recorded separately.
+
+The final targeted run passed 484 CLI tests across 63 files and 189 orchestrator
+tests across five files, including the mandatory daemon web transport tests.
+All 15 actual native component scenarios and six metadata-only production PTY
+checks passed. A separate production CLI run passed eight checks with the real
+local model: an answer, a file-read tool round trip, streaming with 200 wheel
+events, continued editing, resize, Stop, editable recovery and clean shutdown.
+Input painted in 26 ms during that streaming wheel burst. This is one observed
+live measurement, not a general latency percentile.
+
+Current receipts:
+
+- [Targeted tests and broader-suite limits](evidence/native-tui-targeted-verification-20261015.json)
+- [Actual native component scenarios](evidence/native-tui-semantic-components-20261015.json)
+- [Production terminal checks](evidence/native-tui-production-terminal-20261015.json)
+- [Verified GPU 0 and actual local-model CLI checks](evidence/native-tui-local-inference-repair-20261015.json)
+- [Clean package artifact checks](evidence/native-tui-package-artifact-20261015.json)
+- [Installed package terminal and runtime checks](evidence/native-tui-package-install-20261015.json)
+- [Independent committed-source build and 484 tests](evidence/native-tui-committed-scope-20261015.json)
+
+The broader CLI suite did not finish: its interrupted run observed 257 failures
+in Telegram tests. Those failures were not all investigated or established as
+pre-existing. A full-suite pass is not claimed.
+
+## Local inference repair
+
+The broker's explicit operator hold caused local requests to fail and automatic
+fallback to select a remote route. Omnius now retains a sanitized local failure
+reason beside the effective route and clears it after local recovery. GPU
+selection initializes consistently for interactive, single-task and daemon
+entry points, and native Ollama requests retain the exact GPU allowlist.
+
+API model routing preserves organization namespaces such as `robit/`; it strips
+only known routing prefixes or resolves an advertised route. Global config,
+settings, default database and terminal history honor `OMNIUS_CONFIG_HOME`.
+CLI test setup assigns an isolated configuration home so tests cannot change
+the operator's live routing settings.
+
+The operator explicitly approved normal local inference on GPU 0. The supported
+health-checked resume released the hold; actual server and runner placement were
+verified on GPU 0 before generation. The hold remains released after successful
+testing. Other GPU workloads and host device selections are preserved.
+The live run made 12 local physical requests: 11 returned HTTP 200 and a repeated
+helper rerank operation returned a nonfatal HTTP 409. All user-facing checks
+passed. Tool-bearing requests were observed with `think: false` without a forced
+environment override.
 
 ## Packaging
 
-The publish builder emits `dist/tui/native/worker.js` separately, keeps OpenTUI
-external, and carries exact OpenTUI/Bun/tree-sitter dependencies into the publish
-manifest. The artifact audit requires the native worker. The Bun executable and
-native platform library add installation size; a normal npm install provisions
-them automatically. No separate Bun installation is required.
-
-Reproduce terminal validation after building with
-`node scripts/test-tui-native-live.mjs --report /tmp/omnius-tui-live.json`.
-Use `node scripts/benchmark-tui-history.mjs --help` for CPU benchmark options.
-Real inference remains subject to the repository's exact GPU preflight gate.
+The builder emits `dist/tui/native/worker.js` separately and keeps exact
+OpenTUI/Bun runtime dependencies. It rebuilds every workspace package before
+bundling. Clean builds remove deleted legacy renderer modules from distributable
+outputs. Tests and tarball installation must disable unrelated postinstall
+daemon/service startup while validating the native runtime assets.
+The final tarball passed all six production terminal checks in a fresh isolated
+installation. Node CLI help, a real SQLite round trip, packaged Bun/OpenTUI and
+the separately bundled native worker were also verified. The tarball contains
+the README and injected README metadata, with no tests, coverage, caches,
+sourcemaps, build caches or deleted legacy renderer modules.

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.861",
+  "version": "1.0.863",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.861",
+      "version": "1.0.863",
       "bundleDependencies": [
         "image-to-ascii"
       ],
@@ -2783,13 +2783,13 @@
       }
     },
     "node_modules/asn1js": {
-      "version": "3.0.10",
-      "resolved": "https://registry.npmjs.org/asn1js/-/asn1js-3.0.10.tgz",
-      "integrity": "sha512-S2s3aOytiKdFRdulw2qPE51MzjzVOisppcVv7jVFR+Kw0kxwvFrDcYA0h7Ndqbmj0HkMIXYWaoj7fli8kgx1eg==",
+      "version": "3.1.0",
+      "resolved": "https://registry.npmjs.org/asn1js/-/asn1js-3.1.0.tgz",
+      "integrity": "sha512-HWYi1THkOqguCiO+vwj09ZkQFD2kKEsejYHAYCtX0D80WVdoDnvRc9IukXX/jeb0T9NRWJnRtqwdbHltnFjwXA==",
       "license": "BSD-3-Clause",
       "dependencies": {
         "pvtsutils": "^1.3.6",
-        "pvutils": "^1.1.5",
+        "pvutils": "^1.2.0",
         "tslib": "^2.8.1"
       },
       "engines": {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omnius",
-  "version": "1.0.861",
+  "version": "1.0.863",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/library.js",