Repository navigation
Tags
Tags
feat(docker): pass named build contexts to docker.image A Dockerfile can read a host directory, such as a restored Go build cache, only through a BuildKit named build context. docker.image had no way to pass one, so using it meant replacing the target with a hand-written exec. build_contexts passes each entry as --build-context. Files the Dockerfile copies with COPY --from=<name>, or builds FROM <name>, are inputs. A context used only by RUN --mount is not, so a cache directory changing on every run does not force a rebuild. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
feat!: rename blocks to targets and targets to groups People think of the things a build produces or runs as targets, so the unit bit builds is now a target, and a named list of them for the command line is a group (`group default = [...]`). The `target` statement is removed without an alias. State files keep their `block` key, because reading one under a new key fails and loses track of what it built. This repo's BUILD.bit stays on the old syntax until a release with `group` is pinned in Hermit, since the hooks and CI run the pinned bit. sync-ci now renders with that bit, and integration-test runs the debug build against its own project in tests/smoke. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
fix: give tools found through a Hermit link farm a stable cache key Hermit puts a link farm ahead of the project's bin/ on PATH, in a directory named by a hash of the installed packages. Keying an executable by its resolved path made the key differ between machines and CI runs, so blocks such as buf.generate reran every time and never hit the shared cache. Absolute link targets into the package cache also made the fingerprint depend on the home directory. Executables found outside the repository are now keyed by name, and hash_file skips absolute link targets while still hashing what they lead to. Executables inside the repository keep their path key so Git changes such as Hermit upgrades still map to the blocks that run them. The executable hash cache file is versioned so stale hashes are dropped. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
fix(pnpm): support pnpm 12 alongside pnpm 10 pnpm 12 reports settings in camel case and no longer lists its config file paths, so the provider ignored a custom modules or virtual store directory, the pnpmfile and patches, and stopped tracking the user and global config files. Read settings in either case, ask pnpm for the global config path, and track pnpm 12's auth.ini and auth .npmrc. Hermit's pnpm moves to 12 so CI exercises the new version. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
feat: treat TERM=dumb like a pipe Terminals that set TERM=dumb, such as Emacs shell-mode, can't interpret cursor movement, so the live region's redraw codes turned into garbage. Output is now plain and line-streamed, as with --long, whenever stdout is not a terminal or TERM=dumb. NO_COLOR still only removes colour. Child processes are left to detect the terminal themselves. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
fix: stop the block graph implying edges that don't exist The ASCII graph merged unrelated edges into shared strokes, so it showed edges the build doesn't have, such as s3-server feeding fmt-check, and could not tell a crossing from a junction. - Rows never carry both a source's stroke out to its lane and another source's stroke in to a target. - Pure crossings render as an unbroken │, keeping ┼ for junctions. - Edges converging on a target meet only in that target's own merge column, each reaching it on a private track row, with blank rows inserted when none is free. Track corners are rounded so they read as flowing into the merge. - Sources line up with their topmost child where they can, and rows left empty are removed, so targets sit beside their sources. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
feat: trim disk cache entries unused for five days The disk cache grew without bound: nothing ever removed receipts or blobs short of `--cache disk:// --clean`, which drops everything. Follow Go's build cache. Lookups refresh the mtime of the receipt and each blob it needs, at most hourly, and re-storing an existing blob refreshes it so a new receipt never points at an expiring one. Once a day a run deletes files older than five days. trim.txt records the last trim and is claimed under a non-blocking lock and rewritten before deleting, so concurrent runs neither repeat nor wait for a trim. Readers that lose a file see a miss and rebuild. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
feat: add --docs to print the embedded README An agent porting CI to `bit --ci` checked --help, --schema, SPEC.md and the binary, but never found the README section that already explained templates and large-job grouping. Nothing in the binary pointed to it. `bit --docs [topic]` prints the README compiled into the binary, so the text matches the installed version and needs no network. A topic selects one section by heading slug or leading words (`ci`). On a terminal it is rendered with styles, OSC 8 links and wrapping; piped output stays plain Markdown. --help now mentions --docs, and --ci and --cache point at their sections. pulldown-cmark (MIT, no default features) parses the Markdown. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
fix: make --ci workflows read like hand-written ones Generated workflows were hard to read and did not match what a person would type. MiniJinja's tojson escapes ' < > & for HTML, so commands were full of \u0027. Matrix blocks were selected slice by slice with bracketed names, instances repeated their default arguments, and job names listed internal node names. Replace tojson with a plain JSON encoder, select a matrix by its name when the job runs all its slices, omit arguments equal to their default, and name jobs by how the command selects their large blocks. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PreviousNext