Skip to content

v41_v50_implementation.md

Latest commit

 

History

History
114 lines (82 loc) · 7.52 KB

v41_v50_implementation.md

File metadata and controls

114 lines (82 loc) · 7.52 KB
Raw

v41.0.0–v50.0.0 Implementation Record

This document records the implemented development scope for MLX Server Manager v41.0.0 through v50.0.0. The app version in the development tree is 50.0.0. The latest published binary remains v40.0.0 until a separate commit, tag, push, packaging, and release workflow is explicitly requested.

Direct Mode remains unchanged:

OpenAI-compatible client -> selected managed backend or adopted external endpoint

v41.0.0 — Platform Compatibility Baseline

  • Lowers the application deployment target from macOS 26.5 to macOS 15.0.
  • Keeps the application target Apple Silicon (arm64) only.
  • Adds an explicit Platform Compatibility result with current OS, architecture, declared minimum version, and blocking status for a non-Apple-Silicon process.
  • Adds release-asset checks for minimum macOS and arm64 architecture.
  • Does not claim real macOS 15 launch verification until a macOS 15 device or virtual machine has run the release candidate.

v42.0.0 — Evidence Freshness

  • Captures bounded deterministic fingerprints for runtime, profile, model, endpoint, memory, benchmark, and Hermes Agent inputs.
  • Marks evidence Current, Stale, or Unavailable when the user explicitly compares current inputs.
  • Uses configured executable/model/Hermes paths only and does not add broad filesystem monitoring.
  • Running all Operations Center checks establishes a new baseline; changing settings or files does not silently rerun checks.

v43.0.0 — Runtime Environment Doctor

  • Inspects both configured managed backends through a testable actor service.
  • Checks path configuration, file existence, executable permission, script interpreter availability, thin Mach-O architecture, bounded --version output, bounded --help output, and Python dependency-import failures.
  • Provides copyable command previews for selected manual recovery actions.
  • Does not install, update, activate, delete, or repair Python environments.

v44.0.0 — Model Metadata Inspector

  • Inspects only the local path referenced by the selected registered profile.
  • Reads model type, architecture, quantization metadata, context length, vocabulary size, tokenizer presence, chat template, reasoning/tool candidates, safetensors file count, indexed shard count, and total direct-weight size.
  • Reports incomplete or review states for missing/undecodable metadata, tokenizer data, weights, and referenced shards.
  • Does not execute remote code, modify metadata, scan unrelated model folders, or claim guaranteed runtime compatibility.

v45.0.0 — Post-Launch Stability Verification

  • Requires a currently running app-owned selected profile.
  • Runs five explicit short checks of PID continuity and GET /v1/models readiness.
  • Includes current app-observed managed memory and degraded runtime state.
  • Reports Stabilizing, Stable, Degraded, Unstable, Exited, or Not Running.
  • Sends no generation request and does not monitor adopted external processes.

v46.0.0 — Resource Calibration

  • Records an explicit point-in-time local snapshot for a running app-owned selected profile.
  • Stores hardware-memory context, current forecast, observed managed memory, execution fingerprint, and compatible benchmark aggregate values when available.
  • Bounds persisted calibration history to the newest 100 records.
  • Does not change profile options or claim that one point-in-time measurement predicts every future workload.

v47.0.0 — Profile Groups and Launch Plans

  • Adds persisted named Launch Plan definitions.
  • Provides a custom editor for plan name, arbitrary registered-profile inclusion, removal, order, and per-step Required or Optional state.
  • Creates quick plans from the selected profile or all registered profiles without starting a server.
  • Supports explicit editing, reordering, required-state changes, deletion, validation, execution, cancellation, and plan-owned stop requests.
  • Runs profiles sequentially through the existing app-managed start path and waits for each profile to become ready.
  • Preserves profiles that were already running before the plan.

v48.0.0 — Capacity-Aware Multi-Model Operations

  • Assesses the active plan using registered model size estimates, current app-managed memory, current system memory, and profiles already running.
  • Reports Safe, Caution, Critical, or Unavailable.
  • Displays a conservative smaller-estimate-first recommended order for profiles not already running without changing the saved plan order.
  • Blocks plan execution when the projection crosses the emergency memory floor.
  • Does not stop a model automatically, change a profile, or route inference between profiles.

v49.0.0 — Architecture, Supportability, and Distribution

  • Isolates Operations Center state, services, AppViewModel intents, and SwiftUI rendering from the existing large core files.
  • Keeps all new long-running work under replaceable request coordinators rather than raw ViewModel tasks.
  • Adds a bounded, memory-only, redacted Support Bundle containing platform, runtime, evidence, model, stability, capacity, launch diagnosis, and the newest operational-history context.
  • Excludes prompts, responses, credentials, raw environment variables, model files, and full home paths.
  • Adds scripts/verify_release_asset.sh for read-only verification of release-asset SHA-256, archive root, forbidden entries, app version, minimum macOS, arm64 architecture, and code signature.

v50.0.0 — Unified Operations Center 2.0

  • Adds one explicit modal Operations Center accessible from Settings.
  • Combines Platform Compatibility, existing Operational Readiness, Evidence Freshness, Runtime Environment Doctor, Model Metadata Inspector, Post-Launch Stability, Resource Calibration, Launch Plans, Capacity Assessment, and Support Bundle.
  • Keeps existing Operational Readiness available as the narrower non-generating check.
  • Adds English/Japanese labels, light/dark rendering coverage, standard macOS controls, keyboard-focusable buttons and toggles, and stable accessibility identifiers.
  • Does not add a proxy, request router, background startup, automatic repair, external-process ownership, telemetry, cloud logging, or a general-purpose Chat UI.

Persistence

  • Advances the authoritative settings-state schema from 2 to 3.
  • Migrates schema 1 and schema 2 snapshots to schema 3.
  • Persists Launch Plan definitions and bounded Resource Calibration records.
  • Retains authoritative-state backup recovery and refuses unsupported future schemas without overwriting them.

Launch Plan Safety

Before execution, a Launch Plan rejects:

  • an empty plan;
  • duplicate profile entries;
  • a missing required profile;
  • duplicate plan endpoints;
  • non-loopback profiles that require separate network-exposure confirmation;
  • profiles with options known to be unsupported by the configured runtime;
  • profiles requiring separate critical-memory confirmation;
  • a total capacity projection below the emergency memory floor.

A plan never bypasses existing Start guardrails. Stopping plan-started servers affects only app-owned profiles recorded as started by that plan execution.

Verification Boundary

Automated verification covers models, services, persistence migrations, ViewModel ownership boundaries, launch-plan definition editing, secret redaction, macOS 15 compilation, localization, and rendering. Real runtime/model behavior remains dependent on the user's installed executables, model files, ports, and local memory conditions.

A real macOS 15 device or virtual machine is still required before claiming that a v50 release asset has been launch-tested on the declared minimum operating system.