Slipstream 0.3.5

Slipstream 0.3.5 — dense Metal catalog (Dolphin3-Cyber-8B)

Adds a dense (non-MoE) Metal path so small full-load GGUFs can Start without a PGRN sidecar, and catalogs Dolphin3-Cyber-8B.

Why

Slipstream’s Metal start previously required a .pgrn convert. That is correct for MoE expert streaming (Qwen 35B daily driver) and wrong for dense Llama-class models: there are no experts to stream.

Changes

• Catalog: dolphin3-cyber-8b → RavichandranJ/Dolphin3-Cyber-8B-GGUF Q4_K_M (optional Q5_K_M row), not soon
• dense: true → Start when GGUF present; skip convert messaging
• start_server: if dense and no PGRN → full Metal (--gpu-layers 99, --jinja, no --pgrn*); MoE path unchanged
• Dense sampler defaults: temp 0.7 / top_p 0.9 (conversational); coding MoE stays temp 0
• i18n EN+DE: note.dolphin3, st.denseReady
• Docs: docs/DOLPHIN3_CYBER_8B.md + smoke numbers

Smoke (internal SSD, owned PID)

GGUF	~/Modelle/dolphin3-cyber-8b-q4/…Q4_K_M.gguf (~4.6 GiB)
Load	~5 s to /health
RSS	~5.1 GiB
Prefill / decode	~145 / ~27.6 tok/s (short chat)
PGRN	no

Qwen 35B MoE path untouched. One serve only (CRASH_AVOIDANCE).

Verify

node app/scripts/test_a11y_click_contract.mjs
node app/scripts/test_api_parity_ai.mjs
# optional live smoke (no dual-serve):
app/scripts/live_dolphin_dense_smoke.sh

Slipstream 0.3.4

Slipstream 0.3.4 — the shell is clickable and announced

A usability and accessibility release. No engine, streaming, or performance
behaviour changed. The v0.3.3 ad-hoc bundle seal and the measured v0.3.2 prompt-cache
and tool-schema gains carry over untouched.

Why this exists

Every UI test in this repo is a regex over app/dist/app.js and
app/dist/index.html. That kind of test can prove a string is present; it cannot
prove a click does anything. Driving the real DOM instead surfaced four defects that
had survived every green run.

Fixed

Defect	Effect before
The four first-run path fields (pDir, pPgrn, pUrl, pServer) had <label>s that were never associated with them	The fields had no accessible name, and clicking the visible label text did not focus the field
The tab strip had no role="tablist" / role="tab" / aria-selected	Eight tabs announced as plain buttons with no indication of which one is current
showTab() moved only a CSS class	Even with the attributes present, the announced tab would never change
#toast had no role="status" / aria-live	Every message, including the first-run "choose a model folder" error, was silent for screen readers

A new data-i18n-aria channel in applyLang() keeps the tablist label localized, so
it follows the EN/DE switch the way tips and placeholders already did.

Verified

In a real browser against app/dist, with the HTTP cache disabled:

• 8 tabs render, no horizontal overflow, zero console or page errors
• aria-selected tracks every tab click and stays consistent with the visual state
• all four labels focus their input when clicked
• the tablist label switches Hauptbereiche ↔ Main areas
• the toast is a polite status region
• Tools / JSON / schema selections survive MLX → Metal → Auto and a full reload

Suites: 7/7 app/scripts/test_*.mjs, 8/8 app/scripts/test_*.py.

Re-runnable: app/scripts/browser_click_walk.js (serve app/dist, stub __TAURI__
via addInitScript, supply page). The same invariants are pinned statically in
app/scripts/test_a11y_click_contract.mjs, so they cannot silently regress.

Two traps that produced false findings first

Both are worth knowing before the next browser session, because each one made a
correct app look broken:

• Stub timing. Patching window.__TAURI__.core.invoke after app init leaves the
  readiness gate holding stale state. The Start button then looks enabled with no
  model and the first-run journey CTA looks dead. Install the stub with
  addInitScript before load. With a correct fresh-install stub the journey behaves
  properly: Start is disabled and reads "Modell nicht bereit", and step 2 toasts and
  routes to Models.
• Browser cache. After editing app/dist/app.js, page.reload() can keep the old
  script while a separate fetch() shows the new file — so a working fix reads as
  broken. Disable the cache over CDP (Network.setCacheDisabled) before asserting.

Artifacts

Artifact	Bytes	SHA-256
Slipstream_0.3.4_aarch64.dmg	67,325,951	f1e09b38ed98a39a2bbef5cb1a79b1a764ceec2a09facacd78eefd19b87afe3b
Slipstream_0.3.4_aarch64.zip	65,780,310	8f2106b66ac1f809f0d1d82d1f1b17ad404974f158ce3d5e361ce0eb11fb9ce7
slipstream-node_0.3.4_macos_aarch64.tar.gz	4,340,819	132da4aacdedac059d9664d0c39327d6a660e6812d337cd8572f37c79f7d8db5

The node archive reports version 0.3.4 and contains the arm64 binary, launchd,
systemd and non-root Docker templates plus the community-mesh operations guide —
the same layout as 0.3.3, verified by diffing the file lists. Linux stays supported
from source and through the documented Docker/systemd path; this release claims no
cross-compiled Linux binary and no CUDA result.

Bundle verification: every component the runtime manifest marks required is present
in the built .app, codesign --verify --deep --strict reports valid and
"satisfies its Designated Requirement", and spctl still rejects — the expected
outcome for an ad-hoc, unnotarized bundle.

A release gate was missing, and it nearly shipped a hole

The first 0.3.4 candidate built, signed, verified and launched cleanly — and was
incomplete. runtime-manifest.json declares omlx-pgrn/uv as required: true,
but uv was never staged, so the bundle shipped without the 44 MB binary the MLX
runtime bootstrap needs to install anything on a user's machine.

Nothing caught it. test_runtime_manifest.py validates the manifest's shape, not
whether the files it declares exist. codesign --deep --strict passed because a
correctly signed bundle can still be missing content. The only signal was the
.dmg coming out 22 MB smaller than 0.3.3.

app/scripts/test_staged_resources_complete.py now fails the build when a required
component is unstaged, when a declared-executable component is not +x, or when the
staged uv drifts from the pinned version. Verified in both directions: green with uv
staged, and red — naming omlx_uv -> omlx-pgrn/uv and pointing at
stage_uv_runtime.sh — with it removed.

Install

The app is signed ad-hoc, not with an Apple Developer ID, and is not
notarized. macOS may require right-click → Open on first launch.

Slipstream 0.3.3

Slipstream 0.3.3

Canonical product repository: Schero94/slipstream

Highlights

• Complete macOS bundle seal: the app and its resources are now ad-hoc
  signed as one bundle. The mounted DMG passes
  codesign --verify --deep --strict and satisfies its designated requirement.
• Honest trust boundary: ad-hoc signing detects bundle damage but is not an
  Apple Developer ID signature and is not notarization. Gatekeeper may still
  require right-click → Open on first launch.
• Warm, bounded native llama path: v0.3.2's 512 MiB prompt-cache ceiling and
  conditional tool-schema warm-up remain included. The qualified first visible
  tool TTFT is 4.73 s versus 23.36 s cold, with no swap growth.
• No runtime expansion: Slipstream remains llama.cpp/PGRN + oMLX/PGRN. It
  bundles no Ollama runtime and creates no second model store.

Why this patch exists

The previous app launched successfully, but strict whole-bundle verification
found only Rust's linker-signed Mach-O and no resource seal. Official Tauri 2
guidance specifies bundle.macOS.signingIdentity: "-" for ad-hoc signing. A
release-contract test now pins this configuration so an unsealed app cannot
quietly return.

The rebuilt app contains Contents/_CodeSignature/CodeResources, passes strict
deep verification both before packaging and after mounting the DMG, and passes
a real installed launch/quit smoke. spctl still rejects it because no Apple
Developer identity or notarization ticket is claimed.

Validation

• 183/183 complete Rust workspace tests pass.
• 96/96 Tauri/Rust tests pass.
• 40/40 Python runtime, storage, signing and qualification tests pass.
• All six Node UI contracts and the atomic oMLX bootstrap test pass.
• The v0.3.3 Pages source passes Chromium at 1440×1000 and 390×844 with the
  seal boundary and measured TTFT visible, zero overflow and no console errors.
• DMG CRC, mounted-app deep signature verification and ZIP integrity pass.
• The mounted app contains version 0.3.3, 477 resource files, all 16 required
  runtime components, no Python bytecode cache and no Ollama payload.
• App, llama server, PGRN converter, pinned uv and headless node are arm64.
• /Applications/Slipstream.app was upgraded from 0.3.2 to 0.3.3. Its deep
  signature verifies, a real launch/quit passed and ports 8080/8081 remained
  free. Both prior app versions remain recoverable from the user's Trash.

Release artifacts

Artifact	Bytes	SHA-256
Slipstream_0.3.3_aarch64.dmg	67,320,475	88e511a5dc83f1b64b6ebc9f7685376c145c3e66f74b3e57cd1b6cf936f490ef
Slipstream_0.3.3_aarch64.zip	65,758,148	cdb9b07c9d2c6663f1c762b0932c215838fe3aff26d24ce81deeecc172da6a6f
slipstream-node_0.3.3_macos_aarch64.tar.gz	4,341,131	6cc2e21332ed04218f7d2c86c7063884fee662f01c74a0ca3152388ed3d2bbf9

The node archive reports version 0.3.3 and contains the arm64 binary, launchd,
systemd and non-root Docker templates plus the community-mesh operations guide.
Linux remains supported from source and through the documented Docker/systemd
path; this release does not claim a cross-compiled Linux binary or CUDA result.

Known boundaries

The app is ad-hoc signed but not Developer ID signed or notarized. Community
traffic is encrypted in transit, but the chosen inference worker decrypts and
sees the prompt; sensitive and secret work stays local by default. Public
discovery, NAT traversal and relay fallback are not yet deployed, so community
peers still need a known, reachable address.

Slipstream 0.3.2

Slipstream 0.3.2

Canonical product repository: Schero94/slipstream

Highlights

• Pure native stack: Slipstream remains llama.cpp/PGRN plus oMLX/PGRN. It
  bundles no Ollama runtime and creates no second model store.
• Bounded llama prompt cache: the app launcher now supplies
  --cache-ram 512, replacing llama.cpp's 8192 MiB default with a measured
  ceiling that cannot silently consume most of a small Mac's free RAM.
• Conditional tool-schema warm-up: when Tools are enabled, llama.cpp warms
  Slipstream's exact get_current_time + calculator schema after readiness.
  The UI exposes warming, ready and soft-failure states; disabling Tools or
  stopping the server aborts the request.
• Evidence before rollout: the optimization is not copied to oMLX because
  its two qualified short-schema requests remained cold. Each engine keeps the
  path its measurements support.

Measured effect

The product configuration was Qwen3.6-35B-A3B on an external SSD, 10 GiB PGRN
expert cache, 3 GiB RAM reserve, 512 MiB llama prompt cache, MTP and four PGRN
I/O threads.

Request	TTFT	Cached prompt	Result
hidden one-token schema prime	23.36 s	0 / 355	PASS
first visible calculator(19+23)	4.73 s	337 / 361	exact tool + args
get_current_time()	2.46 s	337 / 355	exact tool + {}

First visible tool TTFT fell 79.7% relative to the cold schema prefill. RSS
was 15.18 to 15.42 GiB, free+inactive memory stayed near 6.6 GiB and swap stayed
exactly 1148.12 MiB. The hidden request consumes one completion token in
llama.cpp's accounting; its generated text is discarded and never enters chat
history. A max_tokens: 0 follow-up was rejected because the OpenAI chat
adapter still returned and accounted one token.

Evidence: bench/results/app-tool-prime-llamacpp-20260802.json, SHA-256
25658bfb45ea1a722ec6a69e86061b1f1b3b0a633402f204677bc6b597610e1b.

Validation

• 183/183 complete Rust workspace tests pass.
• 96/96 Tauri/Rust tests pass.
• 39/39 Python runtime, storage and qualification tests pass.
• All six Node UI contracts and the atomic oMLX bootstrap test pass.
• Chromium passed desktop 1440×1000 and mobile 390×844 tool warm-up gates:
  2/2 transitions, exact request contract, no horizontal overflow and no
  console warnings or errors.
• The updated GitHub Pages source passed the same desktop/mobile browser sizes:
  v0.3.2 and the measured TTFT are visible, horizontal overflow is zero, and
  the console has no warnings or errors.
• DMG CRC and ZIP integrity pass. The mounted app contains version 0.3.2, 477
  resource files, all 16 manifest-required components, no Python bytecode cache
  and arm64 app/server/converter/uv binaries.
• /Applications/Slipstream.app was upgraded from 0.3.1 to 0.3.2. A real launch
  produced the registered app process; quit released it and ports 8080/8081
  remained free. The prior app is recoverable from the user's Trash.

Release artifacts

Artifact	Bytes	SHA-256
Slipstream_0.3.2_aarch64.dmg	67,288,167	c89fa8a76301ba25172d34bf1d1d9ed511805ddb6fc4b75b2621a9adb027cc45
Slipstream_0.3.2_aarch64.zip	65,746,319	1a58805ac00957b136eea902ea93068434cdc0391d8a7caa56032ee1f990f388
slipstream-node_0.3.2_macos_aarch64.tar.gz	4,340,656	240e167f583df9fdf945ee33d4ddefd3c86e89abaa5198a6a4b60c1be1f0a47b

The node archive reports version 0.3.2 and contains the arm64 binary, launchd,
systemd and non-root Docker templates plus the community-mesh operations guide.
Linux remains supported from source and through the documented Docker/systemd
path; this release does not claim a cross-compiled Linux binary or CUDA result.

Known boundaries

The app is unsigned and unnotarized. Community traffic is encrypted in transit,
but the chosen inference worker decrypts and sees the prompt; sensitive and
secret work stays local by default. Public discovery, NAT traversal and relay
fallback are not yet deployed, so community peers still need a known, reachable
address.

Slipstream 0.3.1

Slipstream 0.3.1

Canonical product repository: Schero94/slipstream

Highlights

• Pure native stack: Slipstream remains llama.cpp/PGRN plus oMLX/PGRN. It
  does not bundle Ollama or create a second model store.
• One measured API contract: both engines are qualified with the same
  OpenAI-compatible streaming harness for exact chat, strict JSON Schema and a
  specifically selected tool call.
• Verified runtime capsule: the app reports every required bundled runtime
  component and refuses to start an incomplete native engine.
• Storage-aware by default: internal, external and network model locations
  are identified before placement; free-space reserves prevent unsafe copies.
• Bounded oMLX prefix cache: the automatic SSD cache budget now respects
  current free space and a 3 GiB reserve instead of using 10% of total volume
  capacity on a nearly full system disk.
• Headless Linux stays qualified: the exact release engine builds without
  Metal or curl and passes all portable PGRN tests.

Measured engine contract

Every case ran twice at temperature zero, completed its SSE stream through
[DONE], produced a stable semantic hash and left swap unchanged.

Engine / model location	Plain	Strict JSON	Forced add(19,23)
oMLX/PGRN, Qwen internal SSD	5.58 tok/s	4.07 tok/s	3.21 tok/s
llama.cpp/PGRN, Qwen external SSD	5.84 tok/s	4.90 tok/s	6.18 tok/s

The llama result comes from a clean static arm64 build of exact engine commit
0c716f30be270d1fb1077a1a0795684e6faeecf8. It fixes llama.cpp's rejection of
OpenAI's specific-function tool_choice object and validates the requested
function against the supplied schemas. The source patch reconstructs cleanly
from the pinned upstream base.

The large llama Qwen result is deliberately labelled external-SSD: only about
10 GiB was free internally, so copying a roughly 22 GiB GGUF would have violated
the reserve. A small internal Granite model proved endpoint, SSE, PGRN and JSON
mechanics but did not follow the exact-chat and tool protocol, so it is not
advertised as tool-compatible. The large oMLX Qwen model is qualified on the
internal SSD.

Validation

• 183/183 complete Rust workspace tests pass.
• 96/96 Tauri/Rust tests pass.
• All Python runtime/manifest/storage/qualification contracts, all six Node UI
  contracts and the atomic oMLX bootstrap shell test pass.
• Linux engine gate: 14 PGRN objects linked and 12/12 portable tests passed.
• Real engine requests: oMLX 6/6 and llama.cpp 6/6 passed with deterministic
  output and zero swap growth.
• GitHub Pages passed a real-browser gate at 1440×1000 and 390×844: v0.3.1 is
  present, horizontal overflow is zero and the console has no warnings/errors.
• DMG CRC and ZIP integrity pass. The mounted app contains version 0.3.1, 477
  resource files, zero Python bytecode caches and arm64 app/server/converter.
• /Applications/Slipstream.app was upgraded from 0.3.0 to 0.3.1. A real
  launch produced an on-screen 1000×800 main window; quit released the process
  and ports 8080/8081 remained free.

Dependency-audit boundary

The embedded upstream llama server UI reports 21 npm advisories in its complete
development lock tree. The five critical entries are Vitest/browser test tools;
tar, sharp, Sass/Immutable and PostCSS findings are build-only. SvelteKit's
reported remote-form issues do not apply to the static adapter output. The
direct DOMPurify advisory requires CUSTOM_ELEMENT_HANDLING, which this UI does
not enable. No applicable runtime exploit path was found; the raw count is
recorded here so build-only advisories are not mistaken for a clean audit.

Release artifacts

Artifact	Bytes	SHA-256
Slipstream_0.3.1_aarch64.dmg	67,287,899	4a58ce9770d68d9c2d5071e2ebb1a3d24d830c05c6695afefa3cafd50f3c54a6
Slipstream_0.3.1_aarch64.zip	65,743,482	439b818cb5484183b168dc39427b9a54fec363d966cd90adbd614c7f4423ba62
slipstream-node_0.3.1_macos_aarch64.tar.gz	4,340,200	b788a1f94f2fa1a47fedd9f5679c65f027c74066cb80e83f3ae6b60049fdad73

The node archive contains the arm64 slipstream-node binary, launchd/systemd
and non-root Docker templates, plus the community-mesh operations guide. Linux
is supported from source and by the provided Docker/systemd path; this release
does not claim a cross-compiled Linux binary or CUDA throughput.

Known boundaries

The app remains unsigned and unnotarized. Community traffic is encrypted in
transit, but the chosen inference worker decrypts and sees the prompt; sensitive
and secret work stays local by default. Public discovery, NAT traversal and
relay fallback are not yet deployed, so community peers still need a known,
reachable address.

Slipstream 0.3.0

Slipstream 0.3.0

Canonical product repository: Schero94/slipstream

Highlights

• Headless Linux and macOS node: slipstream-node ships with hardened
  systemd, launchd and non-root Docker templates.
• Direct authenticated QUIC: mesh-serve and mesh-send-job use a
  persistent libp2p Ed25519 identity bound to Slipstream's signed Hello and
  X25519 encryption key.
• Free, explicit community donation: public capacity is accepted only with
  --mode community --donate-capacity; it never uses the demo credit faucet.
• Safe Mac controls: local/private/community mode, donation and remote Chat
  are separate controls and default off. Local inference always wins when ready.
• Fail-closed security: process-wide bounded replay cache, one-shot
  challenges, encrypted results only, token/frame/concurrency/per-peer limits,
  and identity pinning.
• Evidence-first vendor queue: 15 current Colibri, oMLX and llama.cpp PRs are
  SHA-pinned but remain unqualified until deterministic A/B gates pass.

Privacy boundary

Transport and Slipstream payloads are encrypted, so passive observers cannot
read prompts. The selected inference worker decrypts and sees plaintext. This is
not FHE, blind inference or a trusted enclave. Sensitive and Secret requests stay
local by default; never send secrets or private code to an unknown worker.

Measured validation

• 183/183 complete Rust workspace tests pass at the released version.
• 83/83 Tauri/Rust tests, 5/5 Mac UI contracts, 3/3 packaged-resource
  checks and 6/6 vendor-harness tests pass.
• Real two-process QUIC community smoke: signed/pinned worker, sealed request,
  5/5 deterministic MockEngine tokens, sealed result; a wrong pin failed before
  inference.
• 4/4 mesh adapter tests cover stable transport identity, Local bind refusal,
  QUIC/Hello identity binding, one-shot challenge and sealed-only result.
• Headless asset, Mac server/privacy, prefer-local Chat and JavaScript syntax
  contracts pass.
• Vendor A/B harness: no vendor candidate is represented as accepted without a
  real model measurement.
• GitHub Pages was exercised in a real browser at 1440x1000 and 390x844: the
  v0.3.0 section is visible, horizontal overflow is zero and the browser console
  has no warnings or errors.
• The packaged app was inspected after mounting the DMG: version 0.3.0, arm64
  app/server/converter, 475 oMLX runtime files, 113 MiB installed bundle and
  88 MiB embedded oMLX runtime. The DMG CRC and ZIP integrity checks pass.
• /Applications/Slipstream.app was upgraded to 0.3.0, launched and quit
  cleanly without enabling a server or community donation.

Release artifacts

Artifact	Bytes	SHA-256
Slipstream_0.3.0_aarch64.dmg	47,067,086	ea93e177220a66e6505f30390cf2b84ae78c7817cda94803773a6294e64082d1
Slipstream_0.3.0_aarch64.zip	45,476,496	158358bd7e72834c294a1ae522c6e9e405499ae309f1f79adbd9fbc09655886d
slipstream-node_0.3.0_macos_aarch64.tar.gz	4,341,222	e8d9af8cc63b88d6ccbbe1883154e2d3efd1b8645f06200d95cd532ead97bc5a

The headless tarball includes the node binary, launchd/systemd/Docker templates
and the community-mesh operations guide. Linux is supported from source and by
the provided non-root Docker/systemd path; this release does not claim a
cross-compiled Linux binary artifact.

Use the direct mesh

See COMMUNITY_MESH_OPERATIONS.md for exact
Linux/Mac/Docker commands and firewall guidance.

Known boundary

Version 0.3.0 connects to a known reachable QUIC multiaddress. It does not yet
ship a public bootstrap fleet, Kademlia discovery, AutoNAT/DCUtR hole punching,
or relay fallback. NATed hosts need a UDP mapping, VPN/private overlay or a
directly reachable address. The macOS app remains unsigned and unnotarized.

Slipstream 0.2.34

Slipstream 0.2.34

Canonical product repository: Schero94/slipstream

Highlights

• One coding contract: Tools, tool_choice, JSON mode, and json_schema
  now use the same Slipstream Chat UI and OpenAI-compatible request shape on
  Metal/llama.cpp and MLX/oMLX.
• Bounded MLX structured requests: Tools/JSON automatically select the
  internal contract profile (capacity 2048, hot 1024, I/O 16). A running
  incompatible MLX profile asks for one safe restart instead of sending a request
  that can exceed the Metal working-set watermark.
• Grammar-complete runtime: the standalone MLX bootstrap installs and verifies
  xgrammar==0.2.3 and apache-tvm-ffi==0.1.11 without pulling the unused
  roughly 2-GiB PyTorch dependency.
• Expert-cache correctness: active-bank experts can no longer be selected as
  LRU victims while that bank is being assembled.
• Reproducible bundle inputs: the small launcher/bootstrap/lock/runtime text
  resources are now versioned in the public product repository.

Live evidence behind the release

Engine / check	Result
llama.cpp Granite SSE	TTFT 0.008 s; usage present
llama.cpp Granite JSON Schema	valid, 0.119 s
llama.cpp cancellation → recovery	0.030 s first chunk → 0.001 s recovery
oMLX Qwen JSON Schema	valid; 12.238 s wall; 1.9 tok/s server
oMLX Qwen tool call	calculator({"expression":"7*6"}); 16.475 s; 1.6 tok/s
oMLX process after both calls	about 10.0 GiB RSS; no memory-pressure abort

The bundled Granite model accepted Metal tool requests but did not choose a tool.
That is sufficient for protocol coverage, not a claim about Metal tool-selection
quality. oMLX performed the complete generated tool-call roundtrip.

Metal Q5 qualification

Qwen3.6-35B-A3B-UD-Q5_K_XL is qualified for the exact operating point
cache=10 GiB, headroom=3 GiB, I/O 4, compact slots, draft disabled:

• two consecutive PASS runs;
• 5.834–6.773 tok/s decode across 32/64-token requests;
• 14.494 GiB peak RSS;
• 0 swapouts and identical Thinking hashes.

The immediately preceding run failed the unchanged reclaim gate because 613
swap-ins were already occurring. It remains recorded as a failure; no safety gate
was weakened for this release.

Validation provenance

The source loop before canonical port completed:

• 22/22 Slipstream UI contract scripts;
• 191 MLX/PGRN tests;
• 83 Tauri/Rust tests;
• 14/14 llama.cpp/PGRN tests including model E2E;
• 810 oMLX grammar/tool/engine tests passing, 3 skipped;
• browser journeys at 1440×1000 and 1024×768 without overflow or console errors.

The canonical product repository then passed:

• 4/4 JavaScript engine-contract scripts and 3/3 packaged-oMLX resource tests;
• 83/83 Tauri/Rust tests and 154/154 P2P workspace tests;
• a clean arm64 DMG build, DMG CRC verification, and ZIP integrity test;
• installation over v0.2.33, version verification, and a successful v0.2.34
  application launch without starting an inference server;
• GitHub Pages journeys at 1440×1000 and 390×844 with no horizontal overflow
  or browser-console warnings.

Artifacts

The GitHub Release publishes both:

• Slipstream_0.2.34_aarch64.dmg
• Slipstream_0.2.34_aarch64.zip

Artifact	Bytes	SHA-256
Slipstream_0.2.34_aarch64.dmg	47,005,825	ae1f62a1b2aee316ab3b3ddf031133d0ed0c6228adc69b62675cb3c7f370c0a7
Slipstream_0.2.34_aarch64.zip	45,443,612	0700300c537d514334469c73d7a215c2863d8fa66faf0b0b715180e63e472e70

The bundle contains no local helper scripts or Python bytecode caches. Its staged
oMLX profile.py and store.py match the validated source inputs byte-for-byte.

Build

cd app/src-tauri
cargo tauri build

The release remains unsigned/unnotarized. On first launch use right-click → Open.

Slipstream 0.2.33

Slipstream 0.2.33

Canonical product repo. Releases ship here (Schero94/slipstream), not LLM-BOOM.

Highlights

• Peak free tolerance: preferred quiet window stays ≥22 GiB free+inactive, but admit band is ≥21.5 (PEAK_FREE_GIB=22 − PEAK_FREE_TOLERANCE_GIB=0.5). A 21.81 near-miss no longer hard-aborts peak.
• Smoke PROFILE=peak: same env knobs; hard refuse still if free < 17 (C admission) or free < cache+headroom+2 (post-load floor estimate).
• UI aligned: computeReco / applyPeak / Start with cache≥14 use admit 21.5; marginal confirm only below that. Copy still says prefer ≥22.

Measured this pass (internal Qwen Metal smoke)

free before	24.74 GiB
knobs	cache 14 · io 4 · headroom 3
decode (first short req, --no-warmup)	8.37 tok/s
hit-rate	77.69%
swap	flat ~1229 MiB used
Stop	OK → free end 24.54 GiB

Honest note: 8.37 is a cold/short first request, not a claim of historical warm peak ~18.9 tok/s.

Build

cargo tauri build --bundles app from app/src-tauri, then ditto into /Applications/Slipstream.app.

Slipstream 0.2.32

Slipstream 0.2.32

Canonical product repo. Releases ship here (Schero94/slipstream), not LLM-BOOM.

Highlights

• Metal Apply Best bands: free≥22 → cache 14; ≥17 → 10; else conservative — always io=4, headroom 3 (qualified PEREGRINE recipe).
• Metal Peak button + Start-time gate for cache≥14 (refuse <17 GiB free; confirm 17–22).
• I/O default 4 (was 8). Backend serde default Auto.
• Obs strip cfg: shows configured cache/io/headroom so path smokes are not misread as peak.
• Good tokens preset: thinking off (temp 0 on send).
• Smoke: PROFILE=path|warm|peak + --pgrn-io-threads when io>1.
• P2P: CLI --spawn-engine refuses live serve lock / healthy endpoint; JobResult sealed on wire; Cluster freeze tip; prefer-local Chat contract test.

Not claimed this release

• Live re-measure of ~18.9 tok/s peak (needs quiet ≥22 GiB free+inactive). Product path now exposes the qualified knobs safely.

Build

cargo tauri build --bundles app from app/src-tauri (stage resources/llama-server / omlx-pgrn as needed), then ditto into /Applications/Slipstream.app.

Slipstream 0.2.31 — Stop guarantee, internal Models, EN/DE

Canonical product repo

This release is published on Schero94/slipstream only. Schero94/LLM-BOOM is the dev monorepo — do not use its v0.2.29 / v0.2.30 tags as product releases.

What's new

• Stop guarantee: SIGTERM the child Slipstream started (Metal llama-server or MLX launcher / lockfile PID), clean /tmp/slipstream-omlx-pgrn.lock, clear HANDS_OFF. No broad pkill -f omlx-server.
• Internal Models path: Start defaults to ~/Modelle (MLX at ~/Modelle/mlx). Sticky /Volumes / Crucial roots are unstuck. External drives stay Advanced overflow — copy out when internal would fill, then free local space.
• i18n: EN + DE full key parity across every main view.
• Qwen 3.6 Metal smoke (internal SSD): Start → PGRN stream → Stop PASS — decode ~6.2 tok/s @ cache 6 GiB, hit-rate ~72%, free+inactive stayed ≥12 GiB, Stop returned RAM to ~22 GiB.

Storage policy

1. Active = internal ~/Modelle
2. Overflow = external only when internal would fill
3. Never sticky-default Crucial for Start

Next (not this tag)

• DeepSeek V4 Flash experiment (real Q4_K experts)
• Model-agnostic discovery under ~/Modelle

Install

Rebuild/stage resources/llama-server and install into /Applications/Slipstream.app, or copy a release-built binary. Ad-hoc signed; first launch may need right-click → Open.

See docs/RELEASE_0.2.31.md.