dsh-llm-mlx v0.4.0

This release adds model-identity checks and bounded, fault-isolated streaming for the optional loopback compatibility proxy.

• Model identity: when modelPath is configured, verify server-declared metadata before reusing a healthy service. Reject a different or unknown identity and leave independently owned processes running.
• Proxy waits and buffering: separate TCP connection, first response body, and idle budgets (10 seconds / 5 minutes / 5 minutes by default), with a 1 MiB UTF-8 limit per SSE event. Continuing data renews the idle budget; active generation has no total duration limit. The limits can be configured.
• Failure isolation: return 504/502 before output starts; terminate incomplete responses after output has begun without appending incompatible error JSON. Cancel associated upstream requests on client disconnect or proxy disposal. Malformed request URLs return 400 while the host continues serving valid requests.

Custom RuntimeDependencies adapters now need to implement verifyModel. Server metadata checks do not attest to model weights or generation quality. The proxy remains optional, and existing direct model routes are preserved. Clients should treat interrupted answers as incomplete. See the versioned English and Chinese README for configuration details.

The attached tarball contains compiled lib/ output and has no installation lifecycle script. For a standard DSH Desktop profile:

dsh plugin --profile desktop add https://github.com/robbywang25/dsh-llm-mlx/releases/download/v0.4.0/dsh-llm-mlx-0.4.0.tgz

The dsh-market catalog is managed separately; verify the version referenced by its entry. Earlier release assets remain available.

Changes: model identity #1, proxy budgets #2, request isolation #3.

Validation for commit f61219ece704331e85ff54ad26f6cd317b03dfad: 69 tests, type checking, reproducible compiled output, and package checks pass on Node 22 and 24 CI. All 10 tarball files match that commit; Gitleaks found no leaks. The SHA-256 checksum is attached. Synthetic HTTP and isolated-process tests cover failure paths without model generation calls.

dsh-llm-mlx v0.3.0

Adds an optional loopback CC Switch / Claude Desktop compatibility proxy for MLX-VLM OpenAI streams. It normalizes duplicate reasoning aliases that CC Switch 3.20.x drops, propagates downstream cancellation, and offers an explicit chat-only mode that replaces Cowork agent instructions and removes tool traffic for least-privilege local-model evaluation. The proxy is disabled by default and never binds beyond 127.0.0.1.

dsh-llm-mlx v0.2.2

Adds an optional MLX-VLM maxNumSeqs setting so memory-constrained Macs can serialize concurrent agent requests with --max-num-seqs. The setting is rejected for MLX-LM; existing defaults remain unchanged. Includes the macOS Desktop PTY compatibility fix from v0.2.1.

dsh-llm-mlx v0.2.0

Adds managed MLX-VLM server support while preserving MLX-LM as the default. Both engines remain loopback-only, use argument-array process startup, and are covered by the same owned-child cleanup boundary.

dsh-llm-mlx 0.1.0

First public release: adds the local-mlx provider, loopback-only MLX-LM reuse, optional DSH-owned server startup, lifecycle cleanup, bilingual documentation, and nine safety tests. The release contains no model weights or user data.