v3.2.0

3.2.0 (2026-07-21)

Features

• mlx: KV-arch metadata + disable paged cache on qwen3-next 80B (db813c3)
• single-slot serving for all 40B+ MLX models (ffb98d2)

v3.1.1

3.1.1 (2026-07-21)

Bug Fixes

• mlx-cluster: anchor rank pgrep to the venv script path (#1333) (08a7ccb)

v3.1.0

3.1.0 (2026-07-21)

Features

• mlx: ping external deadman on healthy watchdog probe (#1328) (2319ae1)

v3.0.2

3.0.2 (2026-07-21)

Bug Fixes

• mlx: watchdog probe discrimination — stop kickstarting on busy (#1324) (a79ff67)

v3.0.1

3.0.1 (2026-07-21)

Bug Fixes

• mlx: make gpu-memory-utilization apply on the lifecycle engine path (#1317) (d04be98)
• mlx: serialize qwen3-next-80b requests (concurrency 1) (9235f59)
• mlx: watchdog escalation ladder — bootout+bootstrap on repeated failure (e2935c4)

v3.0.0

3.0.0 (2026-07-20)

⚠ BREAKING CHANGES

• mlx: rename cluster standalone-ceiling identifiers, trim option docs (#1313)

Features

• mlx: default llama-swap proxy logging to debug (#1310) (a51fd9e)
• mlx: patch vllm-mlx to honor VLLM_MLX_LOG_LEVEL, default debug (#1312) (a81a5c5)

Refactoring

• mlx: rename cluster standalone-ceiling identifiers, trim option docs (#1313) (931cfcd)

v2.4.0

2.4.0 (2026-07-20)

Features

• mlx: generation-parity preflight with auto-heal in cluster-join (#1298) (21bb181)

Bug Fixes

• mlx: drop reasoning parser from 80B-Instruct catalog entry (#1297) (8d19e34)
• mlx: generation auto-heal rebuilds from the remote flake ref (#1299) (b2b29ab)

v2.3.0

2.3.0 (2026-07-20)

Features

• mlx: clusterMode.keepResidentBackends — spare a standalone engine from cluster quiesce (#1291) (200d107)

v2.2.0

2.2.0 (2026-07-19)

Features

• mlx: cluster-join/cluster-detach lifecycle commands (#1285) (0eb0294)

Bug Fixes

• mlx: grep absolute paths + orchestrator robustness nits (#1288) (b7f3484)

v2.1.0

2.1.0 (2026-07-18)

Features

• automode: homelab-gated allow rules for the classifier (002e627)
• mcp: add Zammad MCP catalog entry (basher83/Zammad-MCP) (#1254) (a0bd3ec)
• mcp: enable zammad MCP server in catalog (ac308ae)
• mcp: inject Vikunja credentials via doppler-mcp (#1241) (a998434)
• mcp: source Splunk credentials from OpenBao (#1230) (0a6a610)
• mlx: add defaultRepetitionPenalty to keep batches uniform (#1262) (fb2afdb)
• mlx: add Qwen3-Next-80B-A3B-Instruct catalog entry (new fleet brain) (#1272) (e9e0569)
• opencode: render shared MCP servers via new shared mcp client helper (#1235) (71722ff)
• qwen-code: wire shared permissions and AGENTS.md context (#1237) (b0c1db9)

Bug Fixes

• cluster: pin mlx on the rank, readiness-probe the endpoint, quiesce on every kickstart (#1245) (8344def)
• deps: refresh gh-aw action SHA pins [aw:gh-aw-pin-refresh] (#1238) (3f9afa8)
• discover-models.py: activation-time run registers nothing (env quoting); manual run with identical inputs registers 27 (#1270) (ad0de9f)
• discover-models.py: activation-time run registers nothing (env quoting); manual run with identical inputs registers 27 (#1270) (260c8d5)
• docs: drop internal secrets topology, parametrize Doppler/OpenBao identifiers (826539e)
• mcp: inline the single-use version aliases in the catalog (#1261) (ecdbe9e)
• mcp: make splunk-mcp-connect pass shellcheck SC2016 (#1243) (c7208cb)
• mcp: stop disabling TLS verification in splunk-mcp-connect (#1263) (7e6f811)
• mlx: guard readiness age calculation (6791b68)
• mlx: reap orphaned workers on proxy start, probe serving not liveness (#1260) (fcf4562)

Performance

• warm the pipeline rank after readiness (first request pays the full cold load) (#1247) (#1250) (ae3cf60)