You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 4f1f470
Browse filesBrowse the repository at this point in the historyBrowse files
* fix(test): isolate the copilot CLI snapshot from host XDG data
The unified report also scans opencode under XDG_DATA_HOME, so a developer's real opencode database leaked into the snapshot. Point it at an empty fixture directory like the other home directories.
* perf(nix): fetch only the LiteLLM pricing file instead of the repository
Every Nix CI job unpacked the whole 2.7 GB BerriAI/litellm repository (7-25s per job), and pricing bumps re-lock it every few commits. The input is now a single-file URL pinned to a commit; update-litellm-lock.nu rewrites the commit in that URL before re-locking, and non-Nix cargo builds read the locked URL directly.
* perf(nix): give tests, clippy and hawk a cached dev-profile dependency layer
The test derivation ran cargo test in the release profile on top of the release layers, which hold neither dev-dependencies nor the test-snapshots feature variants, so ~50s of dependencies recompiled on every run; hawk compiled the whole workspace uncached (~25s). A dev-profile deps-only layer now serves tests and clippy, hawk gets its own layer, and main's test job roots both on the sticky disk.
* perf(nix): skip crane's check pass in the cross dependency layers
buildDepsOnly runs cargo check before cargo build, but only the host layer's check artifacts are used (by clippy and hawk). The static Linux and darwin-x64 layers only need the build output, which saves ~30s whenever dependencies change.
* ci: start build and test jobs without waiting on preflight
preflight (~45s) sat ahead of the slowest build on the critical path. ci-gate still requires it, and the pkg.pr.new publish now waits for it so a leaked secret is never published.
* ci: build the Windows native package on a Dev Drive
The rustup install (44s) and Cargo cache extraction (36s) were dominated by Defender and NTFS small-file costs. CARGO_HOME, RUSTUP_HOME and the target directory now live on a ReFS Dev Drive. The cache paths change, so the first run after merge starts cold.
* perf(dev): link with lld on macOS and share dependencies via sccache
lld links the test binaries an edit relinks faster than the Nix cctools ld64 (about 7-10% on leaf and mid-depth rebuilds). sccache shares compiled dependencies, C build scripts included, across worktrees: a fresh worktree's cargo test --no-run went from ~60s to ~45s wall (CPU 315s to 200s) with a warm cache. Set RUSTC_WRAPPER, even empty, to opt out.
* perf(nix): drop the redundant check phase from the test deps layer
The build command already compiles the test binaries; crane's check phase only added another cargo test --no-run for the ccusage bin.
* ci: keep the Windows arm64 build off the Dev Drive
setup-dev-drive needs the Hyper-V VHD cmdlets, which the release workflow's windows-11-arm runner is not known to provide, so arm64 falls back to the default Cargo and target locations.
* ci: exclude Windows build directories from Defender instead of a Dev Drive
Blacksmith's Windows image lacks the Hyper-V cmdlets setup-dev-drive needs (New-VHD), so the Dev Drive step failed. Defender path exclusions for the Cargo home, rustup home and workspace target the same small-file scan cost without Hyper-V, and the step is a no-op where Defender is absent.
* ci: benchmark against the nearest base commit with a preview package
Pricing snapshot commits are pushed to main with GITHUB_TOKEN, so they trigger no CI and never get a pkg.pr.new package. A PR based on one made both perf comment jobs wait out the full 5-minute package timeout and then skip the base comparison. Walk back over pricing-only commits to the nearest one with a package; any other commit without one keeps the old wait.
* ci: drop the Windows Defender exclusion step
Blacksmith's Windows image has no Defender (Add-MpPreference fails with 'Invalid class'), so the step only cost ~7s; the slow cache extraction is not a Defender scan cost there.
Nix builds embed the LiteLLM pricing file from the locked `litellm` flake input, so sandboxed builds do not fetch pricing at build time. To update the locked pricing snapshot:
224
+
Nix builds embed the LiteLLM pricing file from the locked `litellm` flake input (a single-file URL pinned to a LiteLLM commit), so sandboxed builds do not fetch pricing at build time. To update the locked pricing snapshot:
225
225
226
-
Non-Nix Cargo builds read the same locked LiteLLM revision from `flake.lock` and fetch the pricing file from that revision at build time.
226
+
Non-Nix Cargo builds read the same locked LiteLLM revision URL from `flake.lock` and fetch the pricing file from that revision at build time.
0 commit comments