Repository navigation
Highlights
- Auto-sync release automation overhaul. The proposal pipeline now chases one upstream release line at a time (#309), splits research into validated analysis and proposal sessions (#289), renders stage prompts from versioned templates with a synced output contract (#317, #318), supports bounded repair sessions that correct rejected outputs (#290, #298), enforces a prerelease whitelist with own-component sources (#310), ingests failed Pack runs through a dispatch input (#311), and injects frozen per-recipe upgrade guidance extracted from Dockerfile headers (#355).
- Patch governance. Proposals declare patch dispositions and patch version coverage, with repair sessions able to correct them (#321, #327).
- runner-release skill. Documented procedure for cutting releases and aligning support docs (#356).
Features
- Split service Dockerfiles and automate release proposals (2ad02b5)
- Announce runner.py.json changes to gpustack-ota (4f69feb)
- Auto-sync manual dispatch inputs and ubuntu-24.04 runners (#292)
- Structured proposal report with headings and tables (#325)
- Advisory draft checker for stage sessions (#319)
- Per-reply output ceiling exposed as llm-max-tokens (#320)
Fixes
- cann: move LMCache-Ascend to the whitelisted 0.5.6rc4 line (#352)
- auto-sync: gate PR-comment revisions on open state and author association (#351)
- cuda sglang: install into the runtime venv (#341, #342)
- cuda: drop the broken 12.9 vllm 0.30.0 variant and skip the sglang lmcache dependency check (#338)
- pack: configurable collect-record timeout, unblocked collection and merge on cann base images (#331, #336)
- Sign off workflow-generated pull request commits (#334)
- vllm: carry the Mooncake store pending-load fix into new builds (80034d1, f0264ed)
- auto-sync hardening: persistent proposal patch files assembled on the controller (#302, #306), evidence-key validation and repair handoff (#293, #295, #301, #303), group patches applied against the frozen head (#296), contained patch file reads (#288), bare-URL source lists (#307), pinned revise-mode old-engine semantics (#347)
- dependabot: bump urllib3 and virtualenv (#357)
Docs
- Engine-upgrade guide in every active recipe (#346), covering VLLM_TORCHCODEC_VERSION for cuda vllm (#348) and sharper prerelease selection rules (#353)
- README restructure with a contributing section (#294)
Other
- Routine runner catalog updates from scheduled auto-sync and Pack runs (#322, #332, #333, #335, #337, #340, #343, #345, #349, #350, #354)
Full Changelog: v0.1.30.post1...v0.1.31