Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++
21.9 KB
Raw Blame

Terminal work against the 10 October 2026 implementation plan

Historical record (11 October 2026). This page was written while the Terminal was being built, before Strata 0.2.0. It is kept for the record and is not kept up to date: it describes six views, source-run commands and checks that had not yet been run on target hardware. For what the Terminal does now, see TERMINAL.md, TERMINAL_LAUNCH.md and TERMINAL_QUICKSTART.md.

Source baseline for the original milestone: Niko1221/Strata at fb58e0dbc8399662c0e47c76578c6e878b14f6cf. Branch feat/terminal-cli-monitor-current merges the public main fetched on 11 October 2026 at 61b3fb5dd3f1e8ec09cf7e4e05208bc6d3c46406. The plan lives at ../output/strata-terminal-implementation-plan.md in the parent design workspace. This is a working implementation, not a qualified release.

Current-main integration checks on macOS: 706 server tests passed with 11 skipped, all 218 terminal tests passed, and the C++20 batch/stage sampler compiled with warnings as errors and passed. The imported editable Desktop app passed TypeScript checking, a production build and 116 Node tests. The broad setup suite ran 465 tests with two failures and four errors; representative CUDA build and prebuilt-download failures reproduce on a clean public-main checkout, while HIP errors need a ROCm environment absent here. Fake-engine two-replica request/slot correlation and configured card assignment are covered; native GPU builds, native per-replica batch/stage qualification, Windows/Linux install, full Desktop authority and coding harness remain open. The Desktop native shell now supports an opt-in, owner-backed single-profile Serving Start/Stop/Restart/Verify preview, a read-only six-panel Monitor, a read-only Components screen with manual host GPU probe, a reviewed Profile editor for source run config and typed serving overrides, a read-only Models screen with per-config detail and GGUF preflight, a Benchmarks screen for reviewed managed runs, saved reports, workload-checked comparisons and private Markdown export, and a native Calibration screen for reviewed measure/apply/rollback through a pollable host job and shared journal; prototype mutations and its demo monitor sampler are disabled.

Desktop calibration now saves owner-only stage/sample progress, checks an exact worker lease across fresh host processes, and accepts an operator stop request only for the active measurement job. The request is passed into engine generation and the final receipt distinguishes cancellation from success or uncertain completion. The worker does not emit raw prompts or measurements to the browser. Apply and rollback are not cancellable in flight; abrupt worker death, real engine stop latency and Windows locking remain qualification gates.

The source-run Overview now follows YASD's asymmetric two-column composition: serving, prompt/answer speed and batch pressure on the left; four selected GPUs, live requests and activity on the right. The 80 x 24 view keeps its footer; the 110 x 30 view keeps the four card rows, request queue and pipeline groups; a taller view adds reported session totals and model/engine facts. Wide views show separate bounded load and VRAM histories for each selected card, plus batch and pre-admission histories with visible ranges; replica mode keeps child batch trends separate. VRAM history uses each card's reported used/total memory, records gaps for missing or ambiguous sensors and resets with the server instance. The 80-column view keeps current memory readings without VRAM graphs. Missing newest values and batch samples older than five seconds hide their trends. The Overview also hides current batch and host-span numbers when their engine samples are older than five seconds; the detail views retain the last reading with its age. The illustrative captures were regenerated from the current renderer. These layout checks do not establish real four-card measurements.

The replica-specific follow-up maps child slots and pipeline groups into distinct global IDs, attaches replica IDs to active requests, and labels configured GPU groups in the Overview and adds a local stage label to each card only when one engine-reported stage maps there. It shows each live child engine's validated BSTAT processed/active rows per window separately, with sample age in the Serving view. Child startup logs can also report per-replica stage/layer ranges and physical cards when the child's launch visibility order matches its configured card group. Each child now also reports its validated PSTAT host-stage sample, with local stage percentages and age in the Engine view and separate R0/R1 lines in the wide Overview. Server-wide rows/window and stage timing remain unavailable in replica mode because the children have independent sample windows. These host spans include polling delay and are not GPU utilization. Illustrative 110- and 80-column four-card replica screenshots are in the transfer packet; their values are fixtures, not hardware results.

Plan phase Current state Remaining gate
T0 references Upstream, YASD and current UI tokens recorded in the plan Receive/audit supplied TUI zip and DeepSeek harness repo; freeze UI revision and rights
T1 CLI/serve strata.py grammar; foreground server wrapper; named owner profiles with typed overrides, plan/edit/history/rollback, explicit use-config source selection and effective-value origins; separate saved monitor connections with editable URL/prefix/credential reference and config show --effective; typed local operation recipes with edit/plan/run/history/rollback and a bounded recipe-run journal; ephemeral concurrency and queue-cap overrides; offline request checks; non-loopback key preflight; opt-in ready file with actual port and instance/model state; opt-in graceful drain; POSIX owner host with start/status/restart/stop and a shared serving activity lease, prior-argv rollback, guarded hard-kill recovery, private launch attestation of config bytes and server arguments, previewed distinct-file source switching and in-place config migration with rollback, reviewed owner/instance restart preconditions checked before journaling and again inside the owner, authenticated generation readiness and explicit runtime verify with an optional reviewed owner/instance plan; one-shot JSON with monitor --once required for JSON output Windows owner IPC/ACL, shared journal and Desktop integration, real engine stop/restart qualification, Linux package tests
T2 telemetry Versioned /v1/telemetry; bounded read client with occupancy derived only from complete slot reports; 4-GPU per-card rows; engine slot groups and pseudonymous request correlation; read-only telemetry key; admission pressure, queue deadline and rejection counters; server first-engine-token timing; bounded engine-start-log layer ranges and stage/card mapping when visibility order is unambiguous; opt-in engine BSTAT rows/window, separate per-replica BSTAT rows/window and PSTAT stage host spans, plus grouped-pipeline PSTAT stage host spans (source integration, not hardware-qualified); bounded redacted request transitions in the snapshot, with client-observed sequence gaps GPU kernel occupancy, non-grouped stage timing, client-observed TTFT/cache and native-backend qualification; log-free topology source, key rotation/packaging, push event stream
T3 monitor Six-view optional Textual shell; YASD-inspired one-screen Overview with separate Serving, Performance plus bounded output trend, Batching/Parallel, selected-GPU histories with bounded current-card retention, missing-sensor placeholders without stale graphs, Requests and Activity panels at 80+ columns; matching visual Serving, Requests, Engine, Hardware and Logs pages at 128 × 36 and larger; compact text fallback at 60/40 columns with selected-row priority and same-event resize rendering; bounded local device history; freeze, filter, theme, refresh and detach; aggregate throughput restarts from a healthy post-outage baseline and unexpected collection failures show a redacted stale/disconnected state; scrollable live/recent request selection with bounded detail; read-only device/slot drilldowns; stable row identity and explicit retention loss across polls and instance restarts; Serving pane selection and page/home/end keyboard navigation at wide and compact widths; reported stage ranges and sampled pipeline stage in-flight fractions shown in Serving/Engine; Logs shows recent redacted request transitions, observed retention gaps and instance resets; slot phases and group occupancy require a complete, unambiguous slot report GPU kernel occupancy, visual QA and reconnect/resize soak on real Windows/Linux terminals, real single/four-GPU workloads and overhead qualification
T4 code Missing-component exit 5 avoids a second agent loop Integrate supplied shared harness/session API; Desktop parity, approvals, leases, scripted JSONL and policy exit 7
T5 components/release Optional requirements-monitor.txt, source-run documentation, standard-library saved monitor connections in the portable client, read-only doctor/components verify inventory, setup-config models list --installed and models inspect MODEL_ID, telemetry-backed models list --served, and read-only models import PATH --plan with bounded GGUF v3 metadata/tensor-directory and payload-extent and overlap checks, routed-expert encoding signals and known setup-selection hints, opt-in bounded NVIDIA host-card/driver probe that verifies selected indices Actual model adoption, signed component manifests, engine/backend compatibility, qualified HIP/other live probes, installer/update/rollback, cross-platform qualification, notices and measured performance budget
T6 operator integration Handoff contract documented in docs/DEVELOPER_INTEGRATION_HANDOFF.md; versioned owner-local runtime operation bridge in strata_control.management, called by CLI and by a single-profile Desktop Serving screen through an opt-in loopback route with review/execute Start/Stop/Restart/Verify and owner/instance preconditions for verification; a read-only native Monitor route and six-panel UI that check owner/telemetry instance identity, bound fields, keep the telemetry key on the host and omit the prototype monitor ticker; a read-only native Components route and screen using the same CLI inspection path with an explicit opt-in live GPU probe; a native Profile screen that previews and saves source config and typed serving overrides through the same revisioned store as CLI, with a 90-second server-side review and source digest check; a native Models screen that lists setup-written run configs, inspects one config and previews a GGUF file set through the CLI's bounded read-only paths; a native Benchmarks screen that reviews a bounded workload against the managed owner and loaded instance, launches the shared benchmark runner through a pollable host job with private phase progress and an exact-output saved-profile lease check, active/ended lease status and a durable journal receipt recoverable by job ID after host restart, overlap refusal, accepts exact-run operator cancellation from CLI or Desktop, saves a distinct cancelled outcome, reads saved reports, distinguishes incomplete suites, shows phase metrics, compares compatible workloads and downloads private Markdown; a native Calibration screen that reviews decode measurement, apply and exact rollback, requires exclusive-access acknowledgement, pins source/report digests and polls a host job with an owner-only receipt, exact worker lease, bounded decode-sweep progress and exact active measurement cancellation; saved reports reconcile finished measurements after host restart; native Restart separately reviews selected source switches and in-place config migrations before drain; prototype mutations disabled in that mode; local profile and recipe schemas, editors, typed set, revision/rollback, read-only component inventory, calibration measure/review/apply, explicit serving benchmark reports with read-only, workload-checked comparison, private Markdown export and optional pre/post owner-launch matching; a POSIX owner runtime host with macOS fake-engine hard-kill recovery, generation probe and previewed in-place config migration Full Desktop operation authority, model adoption, calibration abrupt-worker cleanup, real-engine stop latency and target-host recovery qualification; Windows control transport; ambiguous-state/manual recovery, real-GPU drain/restart, unmanaged/Windows calibration exclusivity and full recipe capability/UX audit

Tests are under tests/terminal/ and the existing server tests under serve/. No RTX 6000 hardware or model pack is present in this design checkout, so hardware acceptance must be run on the intended machine with the exact card generation and driver recorded. Telemetry marks missing values with null/--.

Latest local checks on macOS with Python 3.11 and a mock/fake engine: tests/terminal/ ran 218 tests with the optional Textual dependency installed (all passed), including the one-screen overview at 110x30 and 80x24 and bounded, cleaned filter paste with keyboard focus; the merged public-main serve/ suite ran 706 tests (11 skipped, all others passed). A live mock HTTP request produced a bounded redacted activity sequence visible through the read-only telemetry key, and the client kept malformed, untrusted event fields out of the Logs view. The source-run status --json and bench run --output json commands reached live mock servers on OS-assigned ports. Saved read-only connection profiles resolved a proxy prefix and environment credential reference in both the source CLI and a standard-library-only portable bundle; effective output omitted the secret. The managed runtime was exercised with a fake engine through start, read-only restart preview, an edited recipe restart, edited-profile restart, failed-activation rollback, stop, hard-killed host recovery, authenticated one-token generation verification, private config/argv launch attestation, a rejected config-read race and an explicit lazy-model load. POSIX managed and foreground serving leases excluded calibration from overlapping engine actions, including a different-port owned host; rollback gained the configured-port preflight. A profile source switch to a distinct config was previewed, digest-pinned and generation-verified against a fake engine; failed activation restored the prior owned launch. An edited run config was previewed and digest-pinned, then restarted with a private original-bytes snapshot; a failed activation restored the prior launch from that snapshot without overwriting the edited file. A managed benchmark matched the owned launch before and after a fake-engine workload; mismatches were refused before output or retained as failed evidence after the workload. An owner-local benchmark lease probe identified one exact output path as active, then ended after its separate holder process was killed; another path stayed untracked. An active mock HTTP stream was stopped through the owner-local cancel action and saved a distinct cancelled report and progress record. A new owner-bridge process read the same journal job ID, and a saved report reconciled a receipt left running by an interrupted worker. The standalone C++ batch/stage sampler compiled and passed with clang++ C++20 warnings as errors; a fake batched engine carried PSTAT through the server without contaminating replies. An earlier 60.29-second synthetic Monitor soak also passed two outages, one restart, bounded 32-sample output-rate history, single in-flight polling and detach without stopping the fake telemetry server; see monitor-soak-60s-current-2026-10-11.json. It remains client stability evidence. These checks do not establish CUDA/HIP/SYCL engine builds, Windows/Linux package behavior, four-card performance or Desktop harness parity.

A 30.31-second synthetic Textual soak on the compact-renderer revision exercised repeated 110x30, 80x24, 60x16 and 40x12 resizes, four selected GPU cards, missing and duplicate card sensors, two disconnects and one restart. It passed bounded device/output histories, one in-flight read and detach; see monitor-soak-30s-ui-fixed-2026-10-11.json. The compact and detail views now replace cached model, request and GPU work with a connection panel during initial attach, disconnect, credential rejection and incompatible telemetry. A later 30.34-second run passed on the palette/host-span display revision e885bf8; see monitor-soak-30s-ui-current-2026-10-11.json.

A later 30.32-second synthetic Textual soak on the batch/queue trend revision passed the same four resize sizes, two outages, one restart, one in-flight read and detach. Batch and pre-admission histories stayed below the 32-sample limit, including gap markers; see monitor-soak-30s-work-trends-2026-10-11.json. These are client stability checks with fixture telemetry, not GPU performance results.

A new 30-second synthetic Textual soak on the VRAM-trend revision passed the same resize, outage, restart and detach scenarios. Both selected-card load and VRAM series remained within four cards and 32 samples each and retained unavailable gaps for missing or ambiguous sensors; see monitor-soak-30s-vram-final-2026-10-11.json. The percentages are derived from fixture memory readings, not measured GPU occupancy or throughput.

The current r31 portable monitor bundle passed an isolated standard-library HTTP and digest-pinned local install/status/uninstall smoke with Python 3.10.20 and 3.11.15 on macOS ARM64. Each run made three telemetry GETs, including one malformed-response probe, and no other HTTP methods; the raw reports are portable-monitor-stdlib-smoke-py310-r31-2026-10-11.json and portable-monitor-stdlib-smoke-py311-r31-2026-10-11.json. Eight focused bundle lifecycle tests also passed under Python 3.10. This does not establish Linux/Windows installation.

The same r31 bundle passed an optional packaged Textual 8.2.8 headless check on Python 3.11.15: it imported the interactive Monitor from the .pyz, connected to loopback telemetry, rendered all six one-screen Overview panels at 110 x 30 and detached with q. The full check made four telemetry GETs and no other methods; see portable-monitor-interactive-smoke-py311-r31-2026-10-11.json. It is not a physical terminal, target OS or real-engine qualification.

A separate macOS POSIX pseudo-terminal smoke launched packaged r31 at 110 x 30, saw live telemetry and raw/no-echo mode, then detached with q and Ctrl-C in separate sessions. Both exited 0, restored the exact prior terminal settings, and left the fake telemetry server readable; see portable-monitor-pty-smoke-py311-r31-2026-10-11.json. The Ubuntu/Python 3.11 CI workflow includes this check but has not run; physical Linux/Windows terminal qualification remains open.

A deep-JSON loopback regression now checks that source and isolated portable status commands return a single redacted exit-3 error for a malformed nested response; the client no longer exposes a Python recursion traceback.

A Linux/Windows GitHub Actions matrix now defines bundle-build and standard-library smoke jobs on Python 3.10 and 3.11, plus optional packaged Textual Overview checks on 3.11. The matrix has not run in this local checkout; target-platform evidence remains open.

The same matrix now includes a source CLI -S smoke before optional terminal dependencies are installed. Local Python 3.10 and 3.11 runs passed help, one-shot status and monitor, offline serve's missing-config exit 5 and coding CLI's missing-harness exit 5, with exactly two loopback GETs and no other methods; see minimal-cli-no-site-py310-2026-10-11.json and minimal-cli-no-site-py311-2026-10-11.json. This proves those dispatch paths do not require Rich/Textual; actual engine startup and clean minimal release installation remain open.

A fresh Python 3.11 virtual environment with only requirements.txt and no Rich/Textual passed all 32 owned-runtime integration tests on macOS, including foreground JSONL ready/stop and POSIX owner start/restart/stop. See minimal-server-runtime-module-macos-py311-2026-10-11.json. A new Linux CI job provisions the same runtime-only environment, but it has not run. This proves source-checkout serving without the optional terminal framework in those tests; real-model, GPU, Windows and packaged installer gates remain open.

The full serve/ suite passed in that same runtime-only environment: 706 tests ran in 270.073 seconds, 11 skipped and no failures; see minimal-server-serve-suite-macos-py311-2026-10-11.json. The Linux CI job now runs both the owned-runtime and server suites; it remains unrun. This is fake-engine/source-checkout evidence, not a clean target installer or actual GPU result.

The sample-age Overview revision passed 215 terminal tests and a 30.35-second synthetic Textual soak with 20 resizes across the four sizes, two outages, one restart, at most 26 retained batch/queue samples, one read in flight and detach while the server remained readable; see monitor-soak-30s-current-work-age-2026-10-11.json. Focused renderer checks show stale single-engine and replica BSTAT/PSTAT samples as unavailable in Overview. Native engine timing remains unqualified.

The 30-minute synthetic UI stress run launched on monitor revision ba138f6 passed in 1,800.31 seconds, with 360 resizes across all four sizes, two outages, one instance restart, four selected-card histories, missing/duplicate sensor gaps, no more than 32 samples per history, one HTTP read in flight and detach without stopping the synthetic server. It handled 3,581 GETs, four expected HTTP failures and no non-GET methods; final traced-memory growth was 597,327 bytes from its post-start baseline. Its 26 RSS samples cover the final 25 minutes, from 44,880 to 45,904 KiB, with a 480 KiB first-to-last change. See monitor-soak-30m-ui-2026-10-11.json and monitor-soak-30m-ui-rss-2026-10-11.csv. Later palette and label changes do not alter the collection path; the shorter run covers their display revision. These are synthetic client checks, not target-terminal or inference-overhead qualification.