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.
Strata developer integration handoff
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.
11 October 2026 - execution brief for joining the current Strata engine, Desktop UI, serving CLI, monitor TUI and coding harness.
Purpose. The developer should use the current UI handover zip, the owner's forthcoming DeepSeek harness repository and TUI zip, and the current GitHub repository to deliver one working product. This document adds the missing operator workflow: editable ways to restart a model, calibrate the engine, verify components and repeat a serving setup. Read it with strata-development-work-plan.md, strata-architecture-plan.md and strata-terminal-implementation-plan.md. Their release, security, model support and hardware gates still apply.
1. What exists now
| Input | Verified checkpoint | Developer action |
|---|---|---|
| Current Strata source | Public Niko1221/Strata main fetched on 11 October 2026 at 61b3fb5dd3f1e8ec09cf7e4e05208bc6d3c46406, including the newer replica, queue keepalive and request-validation work. The developer must refresh this identity at intake. |
Keep the current engine scheduler and API behavior; review later upstream changes against the integration branch. |
| Terminal milestone | Separate Strata-TUI-CLI checkout, branch feat/terminal-cli-monitor-current, merges that public main with the terminal milestone. It adds named serving profiles and separate monitor connections, editable recipes and journal, component checks with an optional live NVIDIA card inventory, calibration, benchmarks, POSIX owner runtime, a versioned owner-local runtime operation bridge, ready file and optional monitor bundle. code is a missing-component stub. |
Split the integrated result into reviewable PRs. Qualify shared authority, recovery and Windows transport before release. |
| Serving monitor | Versioned /v1/telemetry and optional six-view TUI. Its YASD-inspired Overview shows Serving, Performance, Batching/Parallel, four GPU histories, Requests and Activity together at 80+ columns; detailed views show request/slot/group/sensor rows, admission and queue counts, server first-token timing, reported layer ranges when a startup log is available, opt-in grouped-pipeline stage in-flight host spans and a bounded redacted request-activity snapshot for Logs. |
Qualify batch/stage samples on native hardware, add GPU kernel occupancy and client TTFT, obtain a log-free topology source, and run real hardware qualification. Preserve /metrics and API clients. |
| Existing calibration | The source CLI and native Desktop bridge offer separate decode-sweep plan, measure, apply and exact rollback operations. Legacy setup.py --calibrate still updates config and may start the engine; macOS hardware identity now reads RAM through sysctl. |
Migrate the legacy setup path to the reviewed workflow; qualify progress, stop latency, receipt recovery and GPU exclusivity on target hosts. |
| UI handover | Editable React and Node source from the 11 October ZIP is staged in desktop/. Opt-in, single-profile native mode uses shared Python operations for Serving Start/Stop/Restart/Verify, six-panel Monitor, Models and GGUF preflight, Components, Profile edits, reviewed managed Benchmark runs, exact-run cancellation, saved reports and reviewed Calibration measure/apply/rollback. The host checks local Origin/Host, bounds browser fields and retains telemetry credentials. It disables prototype mutations and inference. Chat, the original Tuning design and browser profile store remain prototype; benchmark jobs now have owner-only journal receipts, and their reports persist. |
Confirm font rights, preserve the visual design and map remaining screens to shared services. |
| Owner-supplied assets | The UI handover ZIP is available locally; the DeepSeek harness repository and owner's TUI ZIP have not been received or audited in this checkout. | Hash, license-check and map each asset at M0. Do not claim harness or final UI parity until integration passes. |
UI handover identity: strata-ui-handover-2026-10-11.zip in the parent output/ directory has SHA-256 9491197bc5512a501a6ed5cbd2fcbce482c6cee9eb9bd1bb8aab1216fc2e4182. Its ZIP CRC and all 684 file sizes and SHA-256 digests in CONTENTS-MANIFEST.json were verified. Recheck this identity when the developer receives the copy; license and runtime integration review remain separate gates.
Verified local checks: 215 terminal tests passed on this macOS worktree with mock/fake engines and Textual installed. The merged public-main server suite passed 706 tests with 11 skipped. The terminal tests include private saved monitor connections, proxy-prefix and credential-reference resolution in both the source CLI and portable bundle, bounded engine-stage log parsing, untrusted map rejection, opt-in batch-window sample validation and profile override precedence, responsive live/recent request detail and scrolling, device/slot drilldowns, keyboard paging, stable selection and restart invalidation, prompt-reading progress, server first-engine-token display, frozen queue-age behavior and a portable monitor bundle exercised without site packages. A live mock request produced redacted, sequenced activity through the read-only telemetry key; malformed event fields were excluded from terminal rendering. The source and portable logs commands read the same bounded activity with a read-only credential; follow polling deduplicates transitions, reports sequence gaps and resets on an instance change. The live TUI now marks observed sequence gaps in Overview and Logs, preserves the display while frozen and clears an old gap when the serving instance changes. An offline monitor-only install, hash-pinned update, rollback, commit and removal were verified without changing the server; failure injection restored the active client, and unexpected files block removal. The parallel tests cover serial and batched queue deadlines, API error formats, waiting age, overload, active, queued and partial-prompt disconnects, and recovery. The benchmark CLI produced privacy-safe JSON reports against both mock and fake batched servers. A live mock stream stopped through the exact-run owner-local cancel action; the saved report and progress marked the run cancelled, and the lease released. A benchmark job receipt was read from a fresh owner-bridge process, and a final report reconciled an unfinished receipt. A distinct-file profile source switch and an in-place edit were previewed, digest-pinned and generation-verified on a POSIX-owned fake engine; failed activation restored the prior launch, including from a private old-config snapshot. A managed benchmark matched a POSIX-owned fake-engine launch before and after the workload; mismatches were refused before output or retained as failed evidence after the workload. Calibration rollback tests covered absent and existing setup settings, stale records, file-change races, symlink refusal and write-failure recovery. Direct saved-profile management records were checked alongside recipe records, including failed and uncertain outcomes and compact privacy-safe results. The POSIX host previewed a restart, then started, generation-verified, restarted and stopped an owned fake-engine instance, including failed-activation rollback and hard-killed host recovery after its child exited. Authenticated, lazy-model and failed-generation checks covered the new verification command and restart rollback. A typed recipe preview and run also restarted that owned instance. The standalone C++ batch/stage sampler passed a C++20 clang build/test on macOS, and a fake batched engine carried PSTAT through telemetry; native CUDA/HIP/SYCL builds and four-RTX-6000 performance remain unverified; record the exact card generation, VRAM, driver, model and topology for acceptance.
Desktop calibration now creates an owner-only journal receipt before launching measure, apply or rollback. The browser can retrieve that receipt by job ID after a Desktop host restart. Measurement reports carry the exact job ID, so a saved report can reconcile a receipt left running after report write; a different report cannot. Preflight failures finish the receipt as failed. An exact owner-local worker lease now distinguishes an active job from an orphaned receipt. The measurement writes bounded stages and sample counts; Desktop can request a stop only for the exact active measurement, and the final receipt records whether cancellation occurred. Abrupt worker-death cleanup, stop latency on real hardware and target-host qualification remain open; see docs/MANAGEMENT_PROTOCOL.md for the wire contract.
The normal setup dependency list now excludes CMake and Ninja. A local engine or image-encoder compile installs only missing pinned helpers from requirements-build.txt; existing system tools can be used offline. The dependency boundary passed focused Python 3.10 and 3.11 tests, and 92 hardware-independent setup checks passed on macOS. doctor reports the optional Textual monitor as missing, pinned-version present or version-unverified without importing it into serving. A fresh ready-made-engine install and local compile still need Windows/Linux qualification.
Foreground strata serve --output jsonl now emits versioned starting, ready, stopping, draining and stopped records while human diagnostics remain on stderr. The fake-engine subprocess test generated one response and checked JSON-only stdout and exit codes for both SIGTERM (0) and Ctrl-C (130). All 43 parser leaves offering human/JSON output expose --json; focused source and portable-client tests check the shortcut and streaming rejection. Both clients reject JSON monitor output without --once before opening a connection. --quiet suppresses successful human output, keeps failures on stderr and retains JSON/JSONL records; a fake-engine process test checked quiet JSONL serving through one generation request. The direct legacy server entry point retains its existing human output. Windows console-stop and real-engine lifecycle qualification remain open.
The terminal renderer now clips sanitized remote labels by display cells and pads fixed columns with display width. Focused cases cover combining marks, CJK device and request labels, regional-indicator flags and joined emoji across 110-, 80-, 60- and 40-column layouts. Bidi controls remain excluded. This is a stdlib fallback; ambiguous-width rendering still depends on the operator's terminal.
A 30-minute synthetic 0.5-second Monitor soak also passed: it observed two disconnect windows and one instance restart, returned to a healthy snapshot, retained at most 32 activity and 50 request rows, kept one HTTP collection in flight and left the synthetic server readable after detach. The raw JSON and a 45-sample combined-process RSS series are in the transfer packet. This checks client stability; real-server and four-card overhead remain release gates.
The interactive Monitor now opens in a connecting state without a blocking initial request. A failed poll labels the retained snapshot stale immediately and hides cached answer, prompt and aggregate output rates; repeated transport failures become disconnected. Authentication and telemetry-schema failures require a manual refresh after correction, while transport failures retry with bounded backoff and jitter. A reachable older server without telemetry or metrics keeps its accessible model/status fields under a clear NO TELEMETRY label; live work and rates remain unavailable. Headless and portable-client tests cover these states and recovery.
A separate 60-second synthetic run on an earlier Monitor lifecycle revision opened through CONNECTING, observed STALE during two deliberate outages, recovered after a synthetic server restart, kept one HTTP collection in flight and detached without stopping the test server. The raw report is included separately from the earlier 30-minute soak.
A fresh 60.29-second synthetic run with the output trend enabled also passed: two outages, one restart, 32 bounded valid output-rate samples, one maximum in-flight collection, recovered monitor state and a readable server after detach. Its raw report is monitor-soak-60s-trend-2026-10-11.json; it is client stability evidence, not GPU throughput evidence.
Ten fresh fake batched-engine active/queued cancellation cycles also passed, with generation and recovery each time, engine child exit after unload, and closed HTTP listeners. The raw per-cycle report is in the transfer packet. Real-model/driver cycle qualification remains open.
Current public-main merge qualification. The separate feat/terminal-cli-monitor-current branch merges public main at 61b3fb5d. Its conflict resolution keeps upstream replica serving, serial queue keepalives, logit bias, request validation and reasoning defaults alongside the terminal ownership, bounded queue deadlines and read-only telemetry. On this macOS checkout, the merged server suite passed 706 tests with 11 skipped. The standalone batch/stage sampler passed a C++20 clang build with warnings treated as errors. A broad setup suite ran 465 tests with two failures and four errors; representative CUDA build and prebuilt-download failures were reproduced on a clean checkout of the same public main, and the HIP errors require a ROCm environment unavailable on this Mac. Treat setup and backend packaging as open qualification work. A fake two-replica HTTP test now correlates requests to global slots and globally distinct pipeline groups, with configured GPU-card labels; the 110- and 80-column Overview fixtures plus Engine and Hardware replica detail captures show that layout. The monitor now shows each child’s opt-in BSTAT rows/window and PSTAT host-stage span independently, with sample age. A child’s startup log can report its stage/layer ranges and selected card when launch visibility matches the configured replica group. Server-wide rows/window and stage timing stay unavailable because replica sample windows are separate; host spans are not GPU utilization. Native four-card correlation and performance remain unqualified.
2. Required final behavior
The Desktop UI and CLI must act on the same installed profiles, runtime instances, component inventory and operation records. The TUI is an attachable, read-only monitor for Server users. The coding CLI and Desktop chat must use the same harness session/approval service. The engine is not copied into each front end.
Owner's UI controls CLI operator commands Monitor TUI
| | read only
+----------- typed local management API ------------+
|
profile store + operation journal + inventory
|
runtime worker / Python API / existing C++ engine
|
model pack + backend + telemetry
Desktop Chat -------- shared harness session/approval service -------- strata code
The source-run CLI and a trusted Desktop host can call the same typed runtime path through strata_control.management; see docs/MANAGEMENT_PROTOCOL.md for the one-shot local JSON contract. Desktop start must take the profile revision and source digest from runtime.start.plan; the bridge checks both before journaling and the owner host checks them again before engine launch. Reviewed Stop and Restart commands pin the owner and instance IDs shown by their plans, checked again inside the owner before drain. Restart also pins the candidate profile revision and source digest. The imported Node host now has an internal native-runtime.mjs child-process adapter and an opt-in /api/native/runtime route for one host-configured saved profile. Native mode shows Serving, Monitor, Models, Profile and Components. Serving presents owner status and reviewed Start/Stop/Restart/Verify; a 90-second single-use server-side plan pins the action. Verify also pins the reviewed owner and instance IDs, rechecked inside the owner before an optional one-token generation; loading an unloaded model is a separate explicit choice. Monitor reads bounded /v1/telemetry from the owned local instance, checks its instance ID and keeps the read-only key on the host. Components calls components.inspect through the same Python management bridge as the CLI, bounds fields before browser delivery and runs the host GPU probe only after a manual request. Models calls models.installed, models.inspect and models.import.plan through that bridge; the browser sees bounded report fields and never receives a claim of verified engine compatibility. Profile reads the saved profile and installed run configs, then uses a single-use server-side review; Python rechecks the old revision and source digest before saving, and the running instance remains unchanged until a separate reviewed restart. A changed source or edited config triggers an explicit source-change review in Serving before the owner drains the old instance. The route requires a loopback bind and local Host/Origin, and action POSTs need JSON plus a custom header. This is a local preview boundary, not audited Desktop authentication. With STRATA_NATIVE_PROFILE set, all prototype HTTP mutations and /v1 inference are rejected, the prototype monitor sampler is not started, and STRATA_REAL_ENGINE=1 is incompatible. Its separate browser profile store and remaining prototype views must be mapped or retired after parity checks. There is one lifecycle authority per runtime instance. Every operation has an instance/profile ID, owner, plan, progress events, result, cancellation policy and audit record. The UI and CLI are clients of that authority; neither builds a second installer or calls arbitrary shell strings from a browser request. A monitor credential grants /v1/telemetry only. An inference credential cannot install, calibrate or run commands. Partner hosts such as Unsloth own their registered Strata runtime lifecycle through the separate host adapter contract.
3. Easy-to-edit operational commands
User experience
An owner should be able to choose a saved serving profile, edit its model/device/parallel/batch parameters, preview the effect, then run the same action from UI or CLI. Save profile revisions so a successful configuration can be restored. Surface the exact effective engine arguments, component versions and resolved paths in the preview with secrets redacted.
Proposed public grammar (target, not all implemented):
strata profile list
strata profile create four-gpu --from-config CONFIG.json
strata profile show four-gpu --effective --output json
strata profile edit four-gpu
strata profile set four-gpu parallel=8 max_queue=32 queue_timeout=15 batch_telemetry=true
strata profile set four-gpu parallel=8 --plan
strata profile use-config four-gpu --from-config /absolute/path/next-model.json --plan
strata profile use-config four-gpu --from-config /absolute/path/next-model.json --if-revision PROFILE_REVISION
strata profile validate four-gpu
strata profile history four-gpu
strata profile rollback four-gpu --to REVISION
Read-only monitor connections are separate from serving profiles. They store a server root, an optional proxy prefix and a credential environment-variable name, never its value:
strata connection create work --url https://HOST --prefix /strata --credential-env STRATA_TELEMETRY_KEY
strata connection set work prefix=/new-strata --plan
strata connection set work prefix=/new-strata --if-revision CONNECTION_REVISION
strata config show --effective --connection work --output json
strata status --connection work --output json
strata monitor --connection work
strata logs --connection work --source runtime --follow --output jsonl
Runtime and calibration:
strata runtime start --profile four-gpu --plan
strata runtime start --profile four-gpu
strata runtime restart --profile four-gpu --plan
strata runtime restart --profile four-gpu --startup-wait 300
strata runtime restart --profile four-gpu --allow-source-switch --plan
strata runtime restart --profile four-gpu --allow-source-switch --if-profile-revision NEW_PROFILE_REVISION --if-source-sha256 CANDIDATE_SOURCE_SHA256 --if-owner-id CURRENT_OWNER_ID --if-instance-id CURRENT_INSTANCE_ID
strata runtime restart --profile four-gpu --allow-source-migration --plan
strata runtime restart --profile four-gpu --allow-source-migration --if-profile-revision PROFILE_REVISION --if-source-sha256 CANDIDATE_SOURCE_SHA256 --if-owner-id CURRENT_OWNER_ID --if-instance-id CURRENT_INSTANCE_ID
strata runtime status --profile four-gpu --output json
strata runtime verify --profile four-gpu --wait 300 --output json
strata runtime verify --profile four-gpu --load --wait 300 --output json
strata runtime stop --profile four-gpu --plan
strata runtime stop --profile four-gpu --if-owner-id OWNER_ID --if-instance-id INSTANCE_ID
strata runtime stop --profile four-gpu
strata engine calibrate --profile four-gpu --plan
strata engine calibrate --profile four-gpu --measure --exclusive --out CALIBRATION_REPORT
strata engine calibrate --profile four-gpu --review-result CALIBRATION_REPORT
strata engine calibrate --profile four-gpu --apply-result CALIBRATION_REPORT --exclusive
strata engine calibrate --profile four-gpu --review-rollback CALIBRATION_REPORT
strata engine calibrate --profile four-gpu --rollback-result CALIBRATION_REPORT --exclusive
strata doctor --profile four-gpu --components --output json
strata doctor --profile four-gpu --live --output json
strata components verify --profile four-gpu
strata models list --installed --output json
strata models inspect MODEL_ID --output json
strata models list --served --connection work --output json
strata models import /absolute/path/to/GGUF-directory --plan --output json
strata bench run --profile four-gpu --require-managed-match --suite serving-smoke --concurrency 8 --out RESULTS
strata bench run --profile four-gpu --suite parallel-overload-v1 --concurrency 48 --out OVERLOAD_RESULTS
strata bench run --profile four-gpu --suite parallel-cancel-v1 --concurrency 8 --out CANCEL_RESULTS
strata bench run --profile four-gpu --suite parallel-prefill-cancel-v1 --concurrency 8 --out PREFILL_CANCEL_RESULTS
strata bench report RESULTS --output json
strata bench compare BASELINE_RESULTS MONITORED_RESULTS --output json
strata bench report RESULTS --output markdown --out NEW_REPORT.md
strata bench compare BASELINE_RESULTS MONITORED_RESULTS --output markdown --out NEW_COMPARISON.md
Operation recipes:
strata ops init --profile four-gpu
strata ops list --profile four-gpu
strata ops show restart --profile four-gpu
strata ops edit restart --profile four-gpu
strata ops set restart --profile four-gpu startup_wait_s=45
strata ops plan restart --profile four-gpu
strata ops run restart --profile four-gpu \
--if-revision RECIPE_REV --if-profile-revision PROFILE_REV \
--if-owner-id CURRENT_OWNER_ID --if-instance-id CURRENT_INSTANCE_ID \
--if-source-sha256 CANDIDATE_SOURCE_SHA256
strata ops run check --profile four-gpu
strata ops set check --profile four-gpu live=true
strata ops plan check --profile four-gpu
strata ops run check --profile four-gpu
strata ops set calibrate --profile four-gpu report_dir=/private/path/reports
strata ops plan calibrate --profile four-gpu
strata ops run calibrate --profile four-gpu --exclusive
strata ops runs --profile four-gpu --limit 20 --output json
strata ops result OPERATION_ID --profile four-gpu --output json
Implemented now (source run): serve, status, monitor, logs (bounded redacted request activity), saved profile edits/revisions/rollback, typed ops recipes with a bounded local recipe/direct-operation journal, doctor, components verify, local setup-config models list --installed and models inspect MODEL_ID, read-only models list --served and models import PATH --plan, calibration plan/measure/review/apply/rollback, bench run/bench report/bench compare, and POSIX runtime start/status/verify/restart/stop. The model commands read only top-level strata-*.json configs, include malformed candidates, and return local component presence separately from unverified engine compatibility. The served list reports only the current instance from telemetry. The import preview checks shard names, presence, bounded GGUF v3 architecture/tokenizer metadata, tensor-directory payload extents/overlaps and observed routed-expert encodings, then matches names to supported setup selections without adopting or copying files. A candidate selection still does not prove whole-file integrity, provenance, model-pack suitability or an engine load; actual adoption remains to be implemented. Serve previews settings. Optional doctor --live and components verify --live query a bounded NVIDIA host inventory and flag selected card indices that are absent. The typed components.verify recipe can save live=true for the same check. Card inventory and driver strings do not prove engine/backend compatibility, physical topology or model hashes; HIP and other live probes remain unqualified. The optional max_queue cap and queue_timeout deadline bound admitted calls and their wait for the first engine turn; the monitor shows the cap, rejections, deadline, timeout count, oldest wait and observed client-disconnect cancellations. The deadline requires a queue cap. Nonstreaming timeouts return HTTP 429; streams receive API-specific error events after headers. A disconnected serial request now leaves the bounded FIFO wait promptly. Calibration measures decode speed, then smoke-tests and saves the prior config and setup settings before apply. A separate rollback review checks exact applied-state hashes; rollback smoke-tests the previous config and restores both stores when they are unchanged. --exclusive remains an owner assertion. On POSIX, managed owner hosts and foreground strata serve hold a shared local activity lease; calibration measure, apply and rollback take its exclusive lease, scan owner records including uncertain state, and check the configured port. This prevents overlap among those new launch paths even on different ports. Direct legacy server launches, unrelated GPU processes and Windows control remain outside this lease. A calibration recipe only creates a new measurement report; review/apply/rollback remain separate explicit commands. The native Desktop Benchmarks screen now reviews an exact suite, concurrency, repeats, timeout, condition and new output directory against a loaded managed instance. A single-use review pins profile revision, source digest, owner and telemetry instance; execution rechecks them before calling the shared bench.run implementation. The screen polls a host job and opens the saved report, including failed and incomplete phase evidence. Reports retain the planned phase sequence; an interrupted or short run cannot claim success or enter a successful comparison. A malformed postflight telemetry response also leaves a failed report instead of a stale running record. A private phase progress record and the final report persist across a Desktop host restart. The browser session retains their path and can read them through the shared management bridge; the exact-run saved-profile lease check distinguishes a process still holding that output path from one that released it. Stale progress alone does not establish liveness or work advancing. Saved-profile CLI and Desktop benchmark runs share that local lease to refuse overlapping workloads. Exact-run operator cancellation and durable Desktop benchmark receipts are implemented locally; Windows recovery needs target qualification. The serving benchmark can export a local Markdown summary and a workload-checked comparison with explicit measurement limits; these exports omit the server address, credentials, prompts and answers and never publish automatically. The serving benchmark records client latency/TTFT, a short operator condition label, and sampled slots, groups, queues and devices without prompt/answer content. Read-only bench compare pairs saved runs only when suite, model, request fingerprints, concurrency, source digest and available reported environment agree. Its median phase deltas remain provisional. An opt-in bench run --require-managed-match checks the local POSIX owner host, saved profile/source revision, launch attestation, API endpoint and telemetry instance before and after the workload. A preflight mismatch refuses the workload; a postflight mismatch leaves a failed report. A successful managed report is labelled managed_launch_checked, while ordinary foreground/remote runs remain unverified. The checkpoint match is not continuous attestation, and monitor attachment, driver, engine and model hashes remain unverified. An explicit overload suite synchronizes up to 64 clients, separates expected rejections/timeouts from unexpected failures and checks recovery. The active/queued cancellation suite closes an active stream after first content and a queued stream behind occupied slots. A separate prefill cancellation suite closes a target only after correlated batch-slot telemetry shows partial prompt reading, while a companion decodes. Both check disconnect observation, companion completion, settled resources and recovery. The active/queued suite passed with mock and fake batched engines; the prefill suite passed with the fake batched engine. These source-run checks are not sustained four-card qualification. The POSIX host controls only its own child and restores the prior launch arguments after failed activation. After a hard host kill, its server receives an inherited-pipe parent-loss signal; a subsequent start verifies the host's lease is gone and the recorded child has exited before reclaiming stale private state. A missing/ambiguous ownership record or active socket still needs inspection. An owner can now point a saved profile to a separate run-config file with profile use-config, preview the source switch and restart with --allow-source-switch plus the candidate SHA-256. For an ordinary file switch, the old regular source must remain byte-for-byte unchanged for rollback. An in-place edit needs --allow-source-migration and the previewed candidate SHA-256: before drain the POSIX owner stores the running config bytes in its private directory, and failed activation relaunches from that snapshot without overwriting the edited file. A successful activation retires the snapshot. A hard-linked source remains disallowed; a changed candidate digest is rejected before drain. Start and restart verify the ready-file identity through authenticated local telemetry and a generation request with at least one completion token before reporting a loaded model ready; a deliberately lazy model is reported as unloaded without forcing a load. runtime verify rechecks an owned instance, and --load explicitly tests an unloaded model. These checks use a local API key in memory and put no generated text in the operation journal; the optional API monitor can capture the fixed probe request. A one-token smoke check does not establish answer quality or sustained throughput. The managed POSIX host now compares a private server-reported MAC of the config bytes and launch arguments with the launch it requested before reporting ready. This binds the running server to that launch, but does not prove model-file hashes, GPU topology or post-start external changes. Still target: Windows owner IPC/ACL, journal coverage for remaining direct management commands and the shared Desktop service, sustained hardware overload qualification and signed cross-platform release packaging. No real-model or four-card result exists.
profile edit and ops edit use the configured editor on temporary JSON copies; validation and atomic replacement protect saved revisions. Invalid drafts remain available for repair. profile set and ops set supply typed edits and no-write previews. Recipe history and rollback are implemented locally. The developer must connect Desktop to the same schema and authority. Effective previews must show provenance, executable/argv and effects. The current preview covers local serving and recipe operations. The local journal records recipe runs and saved-profile direct runtime start/verify/restart/stop and calibration measure/apply/rollback with typed intent, start/end events and compact results; unfinished records have unverified completion. Explicit run-config path calibration and Desktop still need the shared lifecycle authority and journal.
The source-run TUI opens on a YASD-inspired one-screen Overview. The narrow left column stacks serving state, side-by-side prompt/answer speed and batching/parallel work; the wider right column keeps per-card history, requests and recent activity together. At taller terminals, the left column also shows reported lifetime session totals and bounded model/engine facts. The 110 x 30 and 80 x 24 captures preserve the navigation footer and all four selected cards. Number keys open the detail views. The reference is Antani Technologies' MIT-licensed yasd repository (https://github.com/AntaniTechnologies/yasd). This implementation draws only from the versioned Strata telemetry contract; it does not synthesize YASD metrics such as KV quantization or expert-cache occupancy that Strata cannot yet report. The source-run TUI can show live request detail with slot, group, queue wait, prompt count, newly read prompt tokens and generated output. Requests scrolls through live and recent rows; Enter shows bounded detail for either kind without prompt bodies. Selected requests, devices and slots remain attached to their reported IDs across polls. A live request follows into recent history; if a selected row leaves retention, detail closes with an explicit notice; an instance restart clears selection even if an ID is reused. Serving tabs select Devices, Requests or Groups even at wide and compact widths; Enter opens bounded read-only device/slot detail, and Page Up/Down plus Home/End navigate rows.
Compact layouts retain the selected row through a 40 x 12 resize. The selected pane precedes optional metrics, and the footer fits the terminal. It uses the snapshot time for queue age, so freeze holds that number steady. Live and recent server first-engine-token timing is now reported from engine queue entry; it includes queue and prefill. The Engine view and device detail show engine-reported inclusive layer ranges only when a current startup log supplies a valid plan; physical-card mapping needs an unambiguous visibility order. This is allocation, not measured stage occupancy. Opt-in batch_telemetry=true yields low-rate engine samples of processed verifier rows and active request slots per batch window, with a sample age; the default and older engines keep those values unavailable. This native path remains unbuilt and unqualified on CUDA, HIP and SYCL. Grouped-pipeline stage in-flight fractions now come from opt-in engine host spans; they include poll delay and are not GPU kernel occupancy. Client-observed TTFT, GPU kernel occupancy, non-grouped stage timing and a complete request lifecycle remain unavailable in the monitor and must stay labelled as such until instrumented.
The Overview shows prompt-reading progress only when a request reports both a read count and prompt length. It summarizes reading, decoding and idle slot phases only when the effective slot report is complete and unambiguous. Exact pipeline group occupancy additionally requires a valid group on every effective slot; a partial or ambiguous report shows unavailable instead of a numeric fraction. The normalizer and bounded client also leave occupied slots unavailable when the source slot list is truncated or ambiguous; the client derives the number from validated slots instead of trusting a conflicting field. A bounded output-throughput trend shows the range over the displayed polling samples; it resets on an instance change and disappears with stale or missing rates. At wide sizes, separate recent-poll trends show processed batch rows/window and pre-admission queue count; replica mode shows each child batch series separately, never a cross-replica average. These trends have bounded sample counts and visible min/max ranges, reset on an instance change, and disappear when their newest source value is missing, its batch sample is more than five seconds old, or the connection is stale. The Overview hides batch and host-span numbers from engine samples older than five seconds; detail views retain the last value with its age. They are display histories, not extra engine measurements. After any failed collection, the first healthy snapshot establishes a new rate baseline while activity sequence tracking remains intact; the next consecutive healthy sample can report aggregate throughput. It marks a reachable unloaded model SLEEPING and suppresses stale speed readings. On initial attach, disconnect, rejected credentials or incompatible telemetry, a dedicated connection panel replaces cached request and GPU values in all six views, including compact 40-column layouts. Disconnects retry automatically; credential and schema failures require manual refresh after correction. Automatic glyph selection uses Unicode block history graphs when the output encoding supports them and ASCII on limited streams. --glyphs braille offers denser cells on terminals with a supporting font; --glyphs ascii forces a plain fallback.
A portable strata-monitor.pyz can be built from packaging/monitor_bundle.py. It contains only the read-only client, renderer, one-screen dashboard, optional TUI module, shared status/monitor/logs grammar and Strata MIT notice; no engine or model files. status, monitor --once and logs work with standard-library Python, while interactive mode requires an explicit Textual install. The version-3 manifest checks SHA-256, bundle members, Python 3.10 minimum, telemetry schema 1 and the license notice; version-1 and version-2 source bundles remain verifiable for update/rollback. Eight focused bundle tests passed on the current Python 3.11 worktree; the version-3 artifact was built and verified with Python 3.10, and its standard-library connection command ran there without site packages. The earlier version-2 build passed the same tests on Python 3.10 and 3.11 on macOS. The manifest is unsigned. A local install command verifies the bundle, requires the candidate SHA-256 from its preview, copies it into a new directory and verifies the copy. A hash-pinned update retains one previous client; rollback exchanges the two verified copies, and commit removes the previous copy. uninstall requires the active SHA-256 from its preview and refuses a directory with unexpected files or an uncommitted previous copy. None of these commands installs Textual or changes the server. Treat this as a source-run transfer artifact; signed optional components and clean Windows/Linux install and update qualification remain release gates.
Editable recipe format
The source-run CLI stores one versioned recipe file per operation under the owner-only profile directory. Its current schema uses schema_version, profile, id, kind and typed params. The three supported kinds are runtime.restart with startup_wait_s, engine.calibrate.measure with an absolute report_dir, and components.verify with optional boolean live (false by default). The calibration suite is still the fixed decode sweep. Recipes are not shell scripts and cannot embed arbitrary command text. The developer must connect UI forms and CLI edits to one validated schema and operation service. The example below is a future target schema, not a current Strata config file:
{
"schema_version": 1,
"profile": "four-gpu",
"model_ref": "installed-model-id",
"devices": [0, 1, 2, 3],
"parallel": 8,
"batch_groups": "auto",
"max_queue": 32,
"queue_timeout": 15,
"operations": {
"restart": { "kind": "runtime.restart", "drain_timeout_s": 30 },
"calibrate": { "kind": "engine.calibrate", "suite": "throughput-and-latency", "trials": 3 },
"check": { "kind": "components.verify", "scope": ["engine", "backend", "model", "tokenizer", "api"] }
}
}
Keep canonical model profiles, user overrides and temporary command flags separate. The effective preview shows which layer supplied each value; the active runtime records a snapshot of what actually launched. Changes to the file do not silently restart a running model. A profile and recipe have stable IDs and a schema version. Validate ranges, model/backend capabilities and incompatible sampling or batch modes before GPU allocation. Use argument arrays and typed values when invoking owned processes. Do not splice raw command text into a shell.
For expert developers who truly need a custom local tool, a separate local-owner-only exec recipe may accept an absolute executable path, an argv array, a fixed working directory and named environment references. It is opt-in, never served through inference or telemetry APIs, never constructed from model output, and displays the exact command and effects before execution. First-class restart, calibration, doctor and benchmark flows should use typed operations instead of this escape hatch.
Operation semantics
| Operation | Preview / checks | Execution and result |
|---|---|---|
| Restart or rerun model | Resolve installed model, backend binary, devices, pack/schema, context, requested slots/groups and effective args. Check port and available resources. | Stop new admissions, drain/cancel within deadline, start the owned instance from a saved profile revision, wait for generation readiness, and report actual slots/groups. On failed activation, restore the previous working profile/runtime selection where safe. Never kill unrelated GPU processes. |
| Calibrate engine | Show scope, expected duration, exact hardware/model key, parameters and benchmark suite. --plan performs no mutation or network access. |
--measure obtains exclusive access or an explicitly drained instance, runs repeated trials and saves a signed or hash-bound report containing baseline, raw trials, candidate and hardware/model identity; it does not change the active config. --apply-result checks report freshness and identity, saves the prior config, applies the candidate, then verifies a smoke generation. On failure restore previous settings. Report latency and throughput rather than only fastest tok/s. |
| Doctor / component check | Read installed component manifests and inspect engine binary/version/protocol, backend and driver, selected GPUs, model shards and hashes, tokenizer/template, optional vision, TUI and harness compatibility. Label unverified checks. | Read-only report with stable issue codes, severity and repair suggestions. A separate explicit repair plan may download or replace artifacts; doctor never repairs by itself. |
| Benchmark | Resolve a named, versioned workload and output directory; show model/config/hardware identity. | Run only on explicit request, record JSON evidence and environment, include four-GPU mixed long/short prompts, cancels and overload, and compare with a pinned baseline. No benchmark runs when opening Monitor. |
The operator surfaces must show command state (planned, waiting, running, cancelling, succeeded, failed), timestamps, log location, resulting instance ID and whether any change applied. CLI --output json/jsonl must remain machine-clean. Secrets are referenced by owner-controlled storage, not embedded in recipes, shell histories, ready files or operation logs.
Four-GPU serving visibility
A 4 x RTX 6000 operator should see one aggregate line and four card rows in the selected-device order, updated from measured telemetry. A missing or duplicate card sensor leaves that selected row visible with unavailable readings and no stale utilization graph instead of substituting another host GPU. Client device history is capped at the currently selected card indices. Keep running requests, occupied decode/prefill slots, engine waiting, pre-admission queue, rejections and timeouts, batch groups, batch size/window, throughput, TTFT and per-card memory/utilization as distinct fields. A missing backend signal is unavailable, not zero or a guessed mapping. Group membership must identify the engine’s actual slot groups; stage placement is a separate engine-reported mapping, and time-series views should reveal whether batching improves throughput while harming tail latency. Provide terminal widths around 110 and 80 columns plus a compact 60-column view; preserve a clean JSON snapshot for scripts. Validate all labels against the engine scheduler before using the TUI as a performance dashboard.
One acceptance scenario should run mixed short and long requests at a recorded parallel setting, increase arrival rate until queue admission is exercised, cancel an in-flight request, then drain and restart the model. Save synchronized CLI JSON, TUI capture, backend counters and benchmark output. The expected invariant is that aggregate and per-card counts reconcile with the scheduler and that no request disappears between admitted, waiting, active, cancelled, completed and rejected states. Record the exact RTX 6000 generation and memory size with the driver, topology and model; the card name alone is insufficient.
4. UI, CLI and TUI mapping
| Surface | Connect to shared operation | Display rule |
|---|---|---|
| Desktop Overview / Models | Start, stop, restart, model selection and readiness | Show saved versus running revision and exact failure/rollback state. |
| Desktop Tuning | Edit profile and operation recipes; validate; preview calibration; apply and restart | Preserve the supplied UI's visual language. Explain duration and exclusivity. Clearly mark predictions versus measurements. |
| Desktop Settings / Components | doctor, component versions, check and explicit repair plan |
Show installed, missing, incompatible and optional components separately. |
| Desktop Serving / Monitor | Same versioned telemetry used by terminal; API connection details | Show batch group occupancy, in-flight/occupied/waiting/queued counts and four per-card readings without invented stage mappings. |
| CLI | Scriptable typed operations above | Same validation and results as UI, with stable exit codes and clean JSON/JSONL. |
| TUI monitor | Read-only telemetry, request/group/device panes | Show operation status from telemetry if exposed, but never execute a restart/calibration from a monitor key. Offer copyable CLI guidance when useful. |
strata code |
Shared harness session and approval service | Local workspace/tool authority is separate from remote inference credentials. |
Terminal styling should match the supplied UI's restrained green/neutral hierarchy and terminology. The terminal font is user-controlled; no font installation is required. The current wide, four-stage, sampled-batching, 80-column serving, recent-detail and slot-detail previews under Strata-TUI-CLI/docs/terminal-previews/, plus the six-view screenshots in output/terminal-screenshots/, use fixture values and are not hardware results. The six-view captures use the dark theme with color enabled; the actual monitor still honors NO_COLOR when set by an operator.
5. Developer implementation order
- Intake and reconcile. Record hashes, source revisions, rights, UI/TUI/harness behavior and the chosen current Strata base. Compare this terminal milestone against that base; port it without replacing newer engine changes. Return a screen/action/contract matrix and supported model/backend matrix.
- Shared runtime and profile contract. Extract pure validate/start from
setup.pyandserve/server.py; define profile schema and effective-value resolver. Preserve legacy launchers and APIs. Add owner-only management access, operation journal and ready/instance identity. - Operator operations. Implement profile edit/validate, restart, calibration plan/apply, doctor/components and benchmark hooks as typed operations. Split
setup.pycalibration measurement from its current update/config-write/start side effects. Add atomic config revisions and rollback. - Front-end wiring. Connect the supplied UI's Tuning, Overview, Models and Serving screens to the same operation and telemetry contracts. Connect the milestone CLI to management and package the TUI independently. Integrate the supplied TUI only after comparing behavior and dependencies; do not ship two terminal frameworks.
- Harness and release. Wire the supplied DeepSeek-derived harness to Desktop and
strata codewith shared sessions/approvals. Build verified Desktop and lean Server distributions and optional components. Keep host-native runtime work in its own adapter milestone. - Qualification. Run clean Windows/Linux installs, offline starts, management permissions, interruption/rollback and hardware tests. On the intended 4 x RTX 6000 host, record exact card generation, driver, model and topology; compare a defined mixed workload with and without the monitor attached.
Required developer PR evidence
- Before/after contract and migration notes; exact source and asset revisions.
- Tests for preview versus apply, config edit conflicts, active requests during restart, failed calibration rollback, stale ready file, component corruption, lost GPU and cancelled operation.
- UI and CLI parity for the same profile/action, with both success and failure screenshots or machine outputs.
- Server without Desktop, harness, Textual or build-tool dependencies; TUI installed and removed independently.
- Four-GPU telemetry where occupied slots, engine waiting, pre-admission queue, group mapping, aggregate throughput and device sensors have distinct measured meanings. Batch-window rows and grouped-pipeline stage in-flight spans require the opt-in engine protocol and native-backend qualification; GPU kernel occupancy remains unavailable. Startup layer allocation is labelled as reported topology.
- Benchmark raw results, hardware/software identity, repeated trials and overhead evidence. Do not describe fixture screenshots as RTX 6000 measurements.
6. Transfer checklist
Use the terminal telemetry field matrix when wiring the supplied UI to /v1/telemetry: it records units, source, request sensitivity, sample age and legacy fallbacks. A missing signal stays unavailable rather than becoming a zero or a four-card estimate.
The qualification matrix separates existing mock/fake-engine evidence from hardware, cross-platform, shared Desktop and harness gates. Carry its raw evidence requirements into the integration PRs.
Deliver the three plan documents, this handoff, the supplied asset originals with hashes, the selected GitHub base revision and the terminal milestone commit or patch. The milestone was developed in a separate local checkout, not merged into the user's current Strata GitHub repository. The developer should identify a current integration base, review the patch, then merge/cherry-pick in reviewable pieces. Do not treat the UI prototype's demo model or restart stub as production engine integration.
At first review, the developer should return: (1) a file/action map, (2) exact command schemas and profile/recipe storage location, (3) a migration plan for the current setup/calibration behavior, (4) dependency and license inventory, (5) acceptance environment including the exact RTX 6000 variant, and (6) an ordered PR schedule with remaining owner decisions.
Current compact monitor check. All 215 terminal tests pass, including a live 40x12 disconnect screen and all six views at 40/60/80 columns. Synthetic 30.31-second and 30.34-second UI runs passed on the compact and display revisions, respectively; see their raw reports. The 1,800.31-second UI soak on monitor revision ba138f6 passed 360 resizes, two outages, a restart, bounded four-card histories, one in-flight read and detach; raw JSON and late-run RSS are in the packet. Target-host and real-model qualification remain open.