Cohesix 1.2.0 status

Use this page to find the right workflow and understand what its evidence means. The installed bundle’s VERSION.txt, RELEASE_NOTES.md and manifest identify its exact contents. The Build Plan owns task scope and the Test Plan owns acceptance checks; neither a source declaration nor a component result transfers proof to a different image or host.

Where work runs

The QEMU and Pi manifests each declare 256 passive Worker instances across Heartbeat, GPU and LoRA roles. Both selected production profiles enable authenticated MCP and A2A routes at the host gateway. Check the installed manifest and profile before using an agent client. Host protocol replies and target admission are distinct from independently verified native provider outcomes. See Host API and the Glossary for those terms.

Release B qualification

Milestone 28g is Complete for the owner’s selected 1.2.0 acceptance gate. At clean source 522463fa1ade, the canonical QEMU and physical Pi Test Plans passed Stages 01–05 under the owner’s renewed EX-2026-0024 exception. Those results retain that source identity. Medium and high QEMU pressure passed on Mac HVF. The corrected Jetson KVM replay passed medium and high loads with 256 READY executable Workers and zero request errors in each window at the b68a2c2a1a38 guest identity.

At clean source b68a2c2a1a38, the strict-production Mac and Jetson QEMU TCP replays passed against their selected native host tools and policy. The canonical factory assembled the 1.2.0 Mac, Linux ARM64 and Pi 4 archives. Both extracted host archives passed native-tool, packaged Python, authenticated TCP and SwarmUI UI qualification. The physical Pi booted the exact production image through TFTP/RAM with verified post-reset transfers, completed wired production TCP checks and passed a separate private Wi-Fi quick smoke. The Pi archive contains the same image bytes in its compact .img. The exact packaged image passed physical SD readback, one fresh boot and packaged-client authenticated TCP qualification. The distributed archives and installers are byte-identical to that five-part qualified set.

The Mac .pkg is Developer ID signed, notarized and stapled. The JetPack 7.2.1 / L4T 39.2.1 ARM64 controller and SwarmUI .deb files are bound by a detached manifest signature verified with the tracked Debian publisher key. Both native installer sets were installed and passed publisher, package-manager receipt and installed-byte readback checks. Lukas Bower personally completed the full adoption walkthrough and marked it PASS. The bounded verifier accepted his written report as owner_attested_host at the installed package identity; its step, time and download measurements are unrecorded. Separate MCP, A2A, NeMo, CUDA, PEFT and MLX results retain their original component and host identities. Three focused real-gateway tests passed the four effective protocol modes. The M28g implementation record indexes the exact archives, installers and selected test results. GitHub and PyPI publication are recorded separately after promotion.

The Python SDK and NeMo kit wheels are version 1.2.0. The SDK’s PyPI distribution is published as cohesix==1.2.0 from the owner-approved v1.2.0 tag; both public file hashes match the reviewed GitHub Actions build. The GitHub Release contains the three portable archives and signed native installers. Optional QEMU setup probes a real four-core startup and offers a pinned 10.1.0 build when the installed binary is incompatible. The selected QEMU and Pi production manifests enable host-side MCP and A2A under the compiler-controlled master switch.

The latest 60-minute Pi 4 GENET RAM-image diagnostic completed four MCP and six A2A original jobs with target-confirmed, independently checked Jetson CUDA outputs and 122 healthy samples. Its collector was amended during the run. Direct Jetson CUDA and NeMo and Mac MLX checks passed at their stated host-only scope.

Nine operator skills passed syntax, link and copied-location guidance checks on Mac and Jetson. Read-only and mock host operations ran from outside a source directory, and the guides now require verified client transport, two private protocol headers, one gateway owner and separate user credentials. The agent delegation and GPU guides now cover cumulative standing attempts, finite caller operations and publisher-epoch Worker replacement.

Milestone 28 is complete at its selected foundation scope. The source declares false-default, compiler-controlled MCP and A2A access switches and implements selected REST/CLI/Python jobs with a private standing ledger for GPU submit and service restart. Exact-source KVM target and native Linux AArch64 CUDA/service observations passed the selected jobs and standing authority cases, including stale request refusal and recovery of pending result delivery. At that foundation commit the gateway had no MCP or A2A routes. The M28 implementation record retains the evidence and limits. This component result does not change Release A acceptance or qualify Release B.

Milestone 28d is complete for the selected MCP-only Jetson Orin scope. A named MCP SDK client reached an authenticated gateway and exact-source KVM Queen, completed native CUDA and PEFT work, and recovered the original results after a lost response and gateway restart. The shared verifier confirmed the deployed PEFT generation; an installed Python wheel projected the same outcomes. Focused transport and manual checks covered disabled, scoped, authentication and refusal behavior. Cancellation, exhausted budget and revocation caused no new effect. The M28d implementation record retains identities and proof limits. This selection did not advertise mixed MLX/CUDA, weight distribution or vMLX MCP client compatibility; Pi and A2A stayed disabled. It does not qualify a physical Pi or Release B.

Milestone 28e is Complete for the selected Linux CUDA/PEFT A2A scope. The compiler-controlled A2A 0.3.0 gateway projected real Jetson CUDA and PEFT jobs through an exact-source KVM Queen. The standard A2A SDK recovered both original tasks after gateway restart; independent CUDA output verification and the shared signed PEFT verifier accepted their native results. NeMo Agent Toolkit’s native client resolved the scoped card, looked up the CUDA task and requested cancellation without changing its completed outcome. Focused tests and live refusals covered disabled access, other subjects, exhausted budget and revocation. The M28e implementation record retains source, image, package and evidence identities. Pi A2A remains disabled; mixed MLX/CUDA, weight distribution and vMLX composition were not selected. This component result does not qualify a physical Pi or Release B.

Milestone 28f is Complete for the selected Linux AArch64 NeMo Agent Toolkit 1.9.0 component. A digest-pinned kit wheel and version-pinned dependency lock installed into a fresh Jetson environment. Toolkit’s native MCP client submitted and recovered an independently verified CUDA job, while its native A2A client delegated a real HF PEFT release whose signed verifier accepted serving generation 9. A changed evaluation policy refused a second PEFT candidate before training or promotion. Separate model-backed NeMo agents connected to the authenticated MCP and A2A paths; native Toolkit evaluation and profiling compared the same completed task with and without governed lookup. The M28f implementation record retains the exact package, target, model and evidence identities, focused checks, adverse results and limits. One budgeted A2A CUDA attempt remains reserved with no confirmed target result; a cross-protocol retry was refused and no success or released allocation is claimed. This component result does not qualify a physical Pi or Release B.

Milestone 28a is complete at its selected Orin scope. The source contains a private digest-pinned workload registration path, version 2 request validation, native GPU diagnostics and an independently checked batch-edge example. An exact-source KVM Queen on Merlin2 admitted reference and independently configured adaptation jobs through both systemd and Docker GPU owners. Signed original terminals and independently verified output bytes bind each case to the selected Orin. Cancellation, lost-response, bridge-restart and native-timeout checks settled their reservations without replay. The M28a implementation record retains the focused evidence and limits; this component result does not qualify Release B.

Milestone 28b is complete for the selected private model reference. Cohesix can train or import a small model adapter, measure it against the running version, and reversibly promote it with a verified application request. A pinned Linux AArch64 NVIDIA host completed genuine LoRA training, independent compatible import, and full Trainer checkpoint resume through an exact-source KVM Queen and WorkerLora receipt path. Held out comparison rejected a worse adapter before load. An interrupted promotion restored its incumbent, which another application request observed. The M28b implementation record keeps the distinct source, profile, negative-result and network-observation limits. This is component evidence, not Pi 4 or integrated Release B qualification.

Milestone 28c is Complete for the narrowed Mac developer workflow on 25 September 2026. The completion record binds the supported macOS 27 Apple M4, focused tests, selected private KVM service work, pinned model/data/adapter, installed vMLX 1.6.65 and the final Developer ID app. Spoken Siri was removed from the developer value gate; macOS user-created Shortcuts supply the useful native action path. The subsequent governed release and serving work is Complete in 28c1 under separate admitted evidence.

Milestone 28c1 is Complete for the selected Mac component on 25 September 2026. An exact-source pinned QEMU Queen admitted real Apple M4 Metal training and an imported release interruption under original ticket identities. The shared verifier reported signed succeeded and recovered_failure outcomes; the latter restored accepted generation 1. The pinned signed vMLX 1.6.65 engine served four frozen responses from the content-bound fused model, refused a changed generation, and observed the verified incumbent. The implementation record retains source, image, profile, native and graph identities, focused checks, and the physical Pi/Release B proof limits.

The installed development-signed App Intents path enrolled a delegated Keychain connection and used user-created Shortcuts to start and inspect an approved private KVM service job. coh resolved its original admission and confirmed result. A later held cancellation settled refused_no_effect before dispatch, and revocation removed the native scope choice. Foundation Models explained the scoped job and proposed a typed inspect follow-up without submitting it. The actions record retains exact identities, the initial cancellation defect and correction. These are selected component observations, not Release B qualification.

The final SwarmUI release binary and App Intents extension were Developer ID signed under Team KB88FQXUX2; Apple accepted notarisation submission b14be979-3330-4770-94f1-73fce245bf6e. The stapled app installed at ~/Applications/SwarmUI-M28c-Build2.app passed strict signature verification and Gatekeeper assessment, and its two signed executable hashes matched the notarised stage. pluginkit registered only that final extension. A saved read-only status Shortcut received HTTP 403 after the earlier scope revocation; this is a live refusal, not a fresh authorised job on final bytes.

The final installed Local MLX panel ran bounded inference and 16-row held-out evaluation using the pinned Qwen2.5-1.5B instruction model and 48-step LoRA adapter on the observed Apple M4 Metal device. The panel displayed the model/adapter hashes, about 974 MB/1.14 GB peak allocation, a useful bounded answer and held-out loss 0.8091070055961609, labelled local observation, no Cohesix admission or promotion. The Python native component’s earlier real 48-step LoRA training and four frozen operational answers remain a narrow diagnostic quality result; answer templates repeat across the train/test split. The installed signed vMLX 1.6.65 engine served a disposable fused copy and reproduced four direct fused answers inside the frozen latency bound; its g1 label is diagnostic, with source and repaired loaded hashes retained separately in the MLX record.

Focused Python MLX (6), vMLX (12), frontend (3), SwarmUI workbench (12), native compile, Rust formatting and generated-consistency checks passed; the installed Metal inference/evaluation and final Apple signing path were exercised. The full suite was not run. M28c adds no Pi hardware, mixed MLX/CUDA, accepted Mac deployment, broad model-quality or integrated Release B claim.

Choose a workflow

Read evidence at its original scope

Selected component records cover the host job foundation, CUDA workload, private adapter, Mac MLX and governed serving, MCP, A2A and NeMo. The Mac app and Shortcuts record covers its separate installed client path. Each record identifies the exercised source, profile, host or target and limits; use the release’s own evidence to judge an assembled 1.2.0 distribution.

The earlier 1.1.0-beta release notes and qualification record retain their original Pi, host and package evidence, including disclosed limits. They remain useful for history and comparison, but do not qualify a new 1.2.0 image.

To verify a specific claim, match its source and selected manifest to the installed bundle, then use Hardware Bring-up for physical proof, the Test Plan for acceptance, and Benchmarks for measured performance. The exceptions register retains accepted historical gaps.