Cohesix Documentation
Documentation View source

Milestone 28c native MLX component record

Scope update (25 September 2026): The revised M28c accepts actual local Metal training plus installed SwarmUI inference and evaluation as developer work. Its completion record binds that UI and the final app. The admitted Mac release, accepted generation, governed vMLX canary and rollback remain 28c1 requirements. Earlier “In Progress” labels below describe the original deployment scope and are not current M28c completion gates.

At the time of this component record, m28c-mlx-metal-provider was In Progress. The native Python component now rejects remote or changed model and dataset inputs, pins MLX 0.32.2 and MLX-LM 0.31.3, requires the observed Metal GPU, caps unified memory, and returns hashed prompt/output and resource observations. Training checks a deadline and cancellation at every loss report; MLX-LM’s final adapter config is normalized to content and settings rather than scratch output paths. The MLX code does not yet run under a selected Cohesix job or promote a generation. Gateway release admission and a release-agent standing dispatch fence now have focused source tests, but the selected native executor remains Linux/CUDA; no Mac release has been admitted.

Title/ID: m28c-mlx-metal-provider
Milestone: 28c / macOS 27: Siri/App Intents, Apple AI and Metal-Backed Workflows
Goal: Run selected local MLX inference and LoRA through the shared admitted release lifecycle.
Inputs: accepted M28b lifecycle; M28c Apple feasibility; selected M4, MLX and local model/data.
Changes: tools/cohesix-py/cohesix/mlx_native.py + tools/cohesix-py/tests/test_mlx_native.py implement and check bounded native primitives. The gateway and agent now check selected release admission and dispatch against the exact accepted baseline, while the selected Mac native executor and provider graph still need integration.
Commands: PYTHONPATH=tools/cohesix-py out/m28c/mlx-venv/bin/python -m pytest -q tools/cohesix-py/tests/test_mlx_native.py; local content-bound MLX inference, train, held-out evaluate and repeatability probes in ignored out/m28c.
Checks: Four focused pure refusal tests pass. The selected Mac completed genuine four-step LoRA on Metal and two same-seed runs with the 16-row held-out dataset produced identical adapter bundle digests. The held-out comparison improved loss versus the base; no predeclared 28b policy or independent admitted comparison has been run.
Deliverables: Local MLX primitives and diagnostic compute measurements, not an accepted M28c provider or developer journey.

Title/ID: m28c-swarmui-mlx-workbench
Milestone: 28c / macOS 27: Siri/App Intents, Apple AI and Metal-Backed Workflows
Goal: Provide the installed app's terminal-free route to the selected MLX release and its signed evidence.
Inputs: 28b release report and exact selected job JSON; current SwarmUI session and installed coh tool.
Changes: SwarmUI's Local MLX desk routes plan/follow/verify/recover through coh peft release, reviewed start through coh job submit, and original-ID status/reconcile/cancel through coh job. The host parser now forwards parent command gateway connection arguments. A signed-journal projection shows Metal, held-out comparison, canary and rollback separately.
Commands: cargo test --locked -p swarmui --lib; cargo test --locked -p swarmui --test workbench; node --test apps/swarmui/tests/mlx_frontend.test.mjs; targeted WebKit Desktop Playwright Local MLX desk case.
Checks: 18 SwarmUI library tests, 11 workbench tests, two pure frontend projection tests and the one targeted browser form test pass. The browser fixture proves form wiring, not an admitted MLX release or installed app acceptance.
Deliverables: A source-level guided UI that preserves gateway admission and original job identity. Native Mac MLX execution, installed-byte validation and m28c-developer-live remain open.
Local diagnostic observation Result Proof limit
Model and data Model tree SHA-256 8b05c3c6097752da054b392d5abab559452fd5c14ae191ee5386130adca4677d; later dataset tree SHA-256 0911cc05194749d0045d590e59333bec4d393ca5802c795dc31f9b726ac39961. Inputs are ignored local copies. The later fixture has 16 training, four validation and 16 test rows. Row count alone does not establish useful model quality.
Metal training Two four-step LoRA runs on the Apple M4 with the later dataset produced the same adapter tree SHA-256 98e6b1ff936d2614c1e32a12853f328ab2dc4ac3340d19be7408d95703446e8e. Peak Metal allocation was about 310–316 MB. Local MLX-LM process only; no Root ticket, host-agent journal, signed outcome, cancellation/interrupt or Cohesix deployment.
Held-out evaluation Base loss 6.320817947387695; later adapter loss 6.016604900360107 on the same 16 test rows. Improvement is diagnostic, not predeclared acceptance. Semantic response quality and policy compatibility remain unqualified.
Inference The base and later adapter both produced nonempty bounded output through the on-device chat template. Later adapter output SHA-256 38add7b12b21120a4860ef45a855d1de7d5173ef7351ac2a3a3921a420826c9d, with 292,800,832 bytes peak Metal allocation. No served generation, vMLX canary, quality assertion or signed Cohesix result. Prompt and response text are excluded from status evidence.
Diagnostic vMLX conversion MLX-LM fused the adapter into a separate model tree SHA-256 4c6876a4c0a78e6b9e902f652af442e8763b3b844e6c3c55a69e8de6ca7ccff2. Direct MLX held-out loss was 6.0157470703125 on the same 16 rows and its bounded output digest matched the adapter. The installed vMLX 1.6.65 engine served a disposable copy at loopback; the bounded Cohesix client returned the same output SHA-256 38add7b12b21120a4860ef45a855d1de7d5173ef7351ac2a3a3921a420826c9d. vMLX’s text adapter load is unsupported by its documented --lora-paths image-model option; this proves a conversion and transport route only. The response interpreted “canary release” as a bird and exhausted its token cap, so it fails useful response quality. No Cohesix canary generation was admitted.
vMLX mutation and incumbent switch vMLX repaired 272 tensor alignments only in the disposable fused serving copy; model.safetensors changed from SHA-256 7f31bedbc3de3647f4cb3220e3373303fab2323870215f918af07563a75cdb25 to 9360693cf724310a0a98af35a2b82d494cec0a8981ae9794a4a06cfdd5fe57da. The fused source tree stayed unchanged. After stopping the candidate server, the same loopback port served a copied incumbent tree with unchanged SHA-256 8b05c3c6097752da054b392d5abab559452fd5c14ae191ee5386130adca4677d; the old candidate model ID was refused, and the incumbent output digest 474b6d26060740ef5c0763435e9947ee80b4b37ad5b23cd322e586043b6f8990 matched direct base inference. This is a manual diagnostic server switch, not Cohesix generation-fenced rollback or a quality-qualified incumbent. Neither server imported Cohesix credentials or enabled MCP tool execution.
Client switch refusal The bounded client now reads health and model listing again after a response; a focused test switches the server’s advertised model during inference and the client refuses to return the output. This detects an observed model switch across a request, not a same-ID weight replacement or an exact process/generation binding.
Content-bound vMLX serving session cohesix.vmlx_runtime.VmlxSession verified the installed net.vmlx.app 1.6.65 signature, Team 55KGF2S5AY and engine commit 22f9c77711fb580df32f4d40bbaea989c2d5421b with entrypoint SHA-256 6ae7d9f0b5db2035b623fc0cacecc3f572bc46db18b69cc8fd611f71b0c2ac7d. It copied the unchanged fused source tree into a private disposable directory, served it with a minimal environment and no inherited Cohesix credentials on loopback port 18085 as cohesix-g3-4c6876a4c0a78e6b, observed process 23689 and loaded tree SHA-256 c41783392b1d884e8915cba48f822a73eae0061315f33f080e174ecd8d5291fa, returned one bounded response, and stopped the process. The source tree remained SHA-256 4c6876a4c0a78e6b9e902f652af442e8763b3b844e6c3c55a69e8de6ca7ccff2. Generation 3 was a local diagnostic label, not an admitted Cohesix deployment. The response exhausted its 32-token bound; no predeclared useful-quality policy or root ticket applied. VmlxSession records bytes and process while running but does not retain a signed outcome or arrange recovery after its own process exits.
Stronger-model quality attempt The MLX Community Qwen2.5-1.5B-Instruct-4bit model was pinned to revision 8b403126fc14f14cfc99bb4cfa72ecbc129ea677, Apache-2.0, and local tree SHA-256 8b40b6d325ea432fda5d9d9612813d203d44bde7c232b465b0af32454b1d4791. Before inference or training, ignored policy out/m28c/qwen-quality-policy.md was written with SHA-256 bb0eff0788be397e2d443d623063da3195a995e977ba5081e8e5509826ee54e0. With the same 16-row held-out dataset, base loss was 7.088871955871582; 16-step, rank-4, seed-41 Metal LoRA created adapter tree SHA-256 d64ca54360acfc8d1e20094972b1d089ec94e9330388f16d036aff7c16010f29 and candidate loss 3.7540886402130127, with peak Metal allocation below 1 GiB. The predeclared held-out and resource bounds passed, but the response gate failed: the canary and worse-candidate answers met their criteria, the HTTP answer did not explicitly require both observed generation and quality/outcome evidence, and the cancellation answer refused and repeated punctuation. This is a negative quality result and cannot be promoted or counted as useful-work acceptance. No admitted ticket or signed outcome was involved.
Instruction-formatted quality attempt A second ignored policy was frozen at SHA-256 5f9cf02413345d1c2a9f7c6fc89d12141f9223b3e8f7d17b90b729e23bdb64b3 before compute. Its 16/4/16-row local dataset tree is 6b2ef5031d278330c0e4db1a40f5de3ddcc2647df74e2fa7a1c2d0f3229670b7. A 48-step, rank-8, seed-42 Metal LoRA created adapter tree a9b837751e207b3c2c52d166f4fb341921cd8d5082197a82915d88f1ea498850, with 1,182,363,920 bytes peak allocation. Baseline loss was 5.796783924102783 and candidate loss 0.8091069459915161 on 16 test rows. All four frozen operational questions met their explicit semantic criteria within 2,045 ms each and 1,006,469,604 bytes maximum peak allocation. Ignored detailed record out/m28c/qwen-instruction-attempt.json has SHA-256 de93b5bcdb6d5f88d9ff2fd75236379b5d11d6b368619279f51ca1e742e1080b. The four answer templates are repeated verbatim across disjoint question phrasings in train, validation and test rows. The loss and responses show the selected narrow facts can be learned; they do not measure broad, independent developer usefulness. This remains an unadmitted local diagnostic: no shared phase journal, target ticket, signed result, accepted serving generation or vMLX canary/rollback.
Direct MLX loopback transport cohesix.mlx_service bound one fixed model and the instruction-formatted adapter to 127.0.0.1:18086 under a diagnostic ID containing the full adapter SHA-256. A real bounded chat request observed Apple M4 Metal, 984,493,744 bytes peak allocation and output SHA-256 d54caf7c83dd0f7f18a72bab8339bf62449e77e67ba63665021ca7d1b1da8f3b, matching the direct MLX question-2 response from the frozen quality attempt. The process exited after the check. This API has no Cohesix credential, admitted generation, durable supervisor or signed result. It does not replace the 28b load/canary/promote/rollback phases, and the local g0 label is not accepted deployment generation zero.
Instruction-adapter vMLX compatibility Before serving, ignored policy out/m28c/qwen-instruction-vmlx-policy.md was frozen at SHA-256 ff1b3a525da6cd239ef8b69734fedbf3583c1b4649dd988d2e8aea9988a7c872: four operational questions, exact model, no tools or stream, 64-token cap, each answer within 5,000 ms. MLX-LM fused the selected Qwen model and 48-step adapter into a separate source tree SHA-256 b75efc50e56aec7c671357ee6eec7d4df55b8c82896a4308d941e0b86e629777. Direct fused inference produced four useful responses; its first three output digests matched the unfused adapter and the fourth differed in wording while retaining the required decision. Installed signed vMLX 1.6.65 engine commit 22f9c77711fb580df32f4d40bbaea989c2d5421b, executable SHA-256 6ae7d9f0b5db2035b623fc0cacecc3f572bc46db18b69cc8fd611f71b0c2ac7d, served a disposable copy as diagnostic cohesix-g1-b75efc50e56aec7c, process 83239. All four vMLX texts exactly matched direct fused inference and took 4,161/2,774/2,840/2,710 ms. The source model and adapter trees stayed unchanged; the loaded copy tree was a2e3c00622388097a0e8ca3c2b8e600ba765ff44be7d56c88aca5bca43f7cec4 after local alignment repair. Ignored detailed report out/m28c/qwen-vmlx-session.json has SHA-256 76f43119c53feb37f1ff9b9cc59591a5dcefd6b899dbfb0180e717fc5d0b1acb. This qualifies a real serving format and narrow response comparison, not a Cohesix admitted generation: g1 is only a diagnostic label. The trained adapter was produced outside the shared phase journal, no signed result or governed canary exists, and no Cohesix rollback was run. The answer templates remain narrow.

m28c-vmlx-compatibility is Complete at the revised diagnostic scope. The later instruction adapter’s fused copy passed a predeclared, narrow four-question quality and latency policy on the installed engine, including process and repaired-byte binding. An admitted generation, signed Cohesix result, same-generation canary and governed incumbent rollback are separate 28c1 requirements. The earlier manual server switch does not supply them.

The host-tool/Python/benchmark compatibility review finds the native MLX and vMLX modules additive. The standing-authority parser recognises peft.release as an optional compiler-selected action. The gateway now admits it only with an exact private request, helper and accepted baseline, while the agent rechecks the incumbent and standing dispatch barrier before native effects. Forward phases also recheck a cancellation request; rollback remains permitted under valid recovery authority. The current selected manifest does not enable that action, and the agent’s phase executor remains Linux/CUDA, so this is not Mac MLX admission. The new SwarmUI desk uses the existing coh parser and gateway job route; it does not add a transport or provider. Python and raw/REST benchmark schemas and target interfaces are unchanged. Integration must preserve M28b’s durable phase journal and signed result custody; a call to these modules alone cannot satisfy that contract.

A later M28c native-action source change added a gateway-derived start for an already selected systemd.restart standing scope, with matching coh, Python REST and App Intents clients. It supplies a useful governed start path without changing target interfaces, benchmarks, the selected peft.release ceiling or MLX admission. The Mac release executor, admitted generation and live cross-surface evidence above remain outstanding.

Material AI assistance contributed the implementation, tests and this record. The ignored local artifacts are diagnostic and must be replaced by exact-source live evidence after the admitted implementation is complete.