refactor: split into moon.work workspace (bit + bit_* core + bitx_* extensions) by mizchi · Pull Request #82 · bit-vcs/bit

mizchi · 2026-05-26T14:20:54Z

Summary

Restructure the repository into a MoonBit workspace (moon.work) inspired by gitoxide's gix-* plumbing / feature-crate split. The main mizchi/bit module is being progressively decomposed into per-primitive modules so consumers can pull in only the pieces they need.

mizchi/bit is no longer one monolithic module — extensions live in mizchi/bitx_* and leaf core primitives in mizchi/bit_*
Repo root holds just moon.work + project files (no more src/ or moon.mod.json at the top level)
CLI surface (bit, git-bit subcommands and flags) is unchanged; this is a pure code-organization refactor

Layout

moon.work
modules/
  bit/                ← residual main: types, object, pack, refs, protocol, repo, mid, high, cmd
  bit_apply/          ← extracted core modules (gix-* equivalents)
  bit_archive/        (gix-archive)
  bit_bootstrap/
  bit_date/           (gix-date)
  bit_diff3/          (gix-merge low-level)
  bit_diff_core/      (gix-diff low-level)
  bit_fast_import/
  bit_hash/           (gix-hash)
  bit_ignore/         (gix-ignore + gix-glob)
  bit_osfs/           (gix-fs OS impl)
  bit_trailers/       (gix-trailers)
  bit_utils/          (gix-utils + gix-quote)
  bitx_bitconfig/     ← non-Git extension modules
  bitx_doc/
  bitx_hq/
  bitx_hub/           (+ /crypto, /native)
  bitx_kv/            (+ /native)
  bitx_rebase_ai/
  bitx_subdir/
  bitx_workspace/

Naming conventions:

bit_<name> (no x) — extracted Git plumbing modules (mirrors gitoxide's gix-*).
bitx_<name> (with x) — extensions for non-Git features.

MoonBit's workspace resolver handles the mizchi/bit ↔ mizchi/bit{,x}_* module-level cycle as long as the in-package graph stays acyclic, so no cmd extraction was needed.

Tooling updates

tools/check-layers.mjs walks modules/*/src, classifies mizchi/bitx_* as ext layer and mizchi/bit_* as core layer
tools/check-layer-guards.sh, tools/install-bit.sh, tools/playground-dev.sh, tools/build-bit-native.sh
tools/flaker-affected-rules.{toml,mjs,test.mjs}, tools/flaker-cli-wrapper.mjs
Taskfile.pkl — source globs, build commands, _build/ output paths (workspace mode emits _build/<target>/<profile>/build/mizchi/bit/...)
.github/workflows/pages-demo.yml path filter
flake.nix, package.nix (moonModJson path)
docs/package-layout.md — rewritten with the new layout, naming conventions, and per-module mapping table

Test plan

Reviewer notes

External consumers that imported mizchi/bit/<sub> should migrate to mizchi/bit_<sub> (or mizchi/bitx_<sub>). Inside this repo every reference has already been updated.
Renames from old names to new module names: date_parse → bit_date, string_utils → bit_utils, tar → bit_archive. Others map 1:1 (e.g. hash → bit_hash).
Remaining mizchi/bit/... core packages (types, object, pack, refs, protocol, repo, etc.) are not extracted yet — they have inter-package dependencies that need to be resolved in a follow-up.

Commits

refactor: extract x-kv into separate mizchi/bitx_kv module (PoC)
refactor: extract all x-* extensions into bitx_* workspace modules
refactor: move main bit module into modules/bit/ for uniform layout
refactor: fix build paths after module move
refactor: extract 12 leaf core packages into bit_* modules

The workspace split multiplies the number of modules moon test walks, and the previous '--no-parallelize -j 1' configuration pushed the job past its 30-minute deadline (see PR #82 run 26488539393 timeout). The extracted bit_* modules are independent enough that parallel test execution should be safe; bump the timeout to 60m as a safety net.

Introduce a moon.work workspace and split the first low-frequency extension (x-kv) out of the main mizchi/bit module, mirroring gitoxide's gix-* vs feature-crate layout. x-kv has no in-tree consumers, so it is the safe minimal candidate (no module-level cycle with mizchi/bit). Documents the new modules/ layout and the bitx_* naming convention in docs/package-layout.md.

Split the remaining bit extensions (x-bitconfig, x-doc, x-hq, x-hub, x-rebase-ai, x-subdir, x-workspace) out of the main mizchi/bit module and into their own MoonBit modules under modules/, joining the previously-extracted bitx_kv. MoonBit's workspace resolver accepts the module-level cycle (mizchi/bit <-> mizchi/bitx_*) because the in-package dependency graph stays acyclic. - moon.work now lists all eight bitx_* modules as members - cmd/bit and cmd/git-bit are rewired to import from the new module paths - tools/check-layers.mjs walks modules/<name>/src as additional roots and classifies mizchi/bitx_* as the ext layer - docs/package-layout.md updated for the new module table and layout - Taskfile.pkl test/bench targets point at the new module names All 298 tests across the 8 extracted modules pass (moon test); moon check on the full workspace is clean (0 errors).

Flatten the repo layout so every workspace member lives under modules/. Previously the main mizchi/bit module sat at the repo root (src/ + moon.mod.json) while extensions lived under modules/bitx_*. Now: modules/ bit/ ← main module (was repo root) moon.mod.json README.mbt.md src/ bitx_bitconfig/ bitx_doc/ ... - moon.work members updated to point at modules/bit instead of "." - modules/bit/moon.mod.json: dropped the now-irrelevant `exclude` list (the repo-root sibling dirs are no longer reachable from the module root, so MoonBit's source globbing leaves them alone naturally) - tools/check-layers.mjs no longer treats src/ at repo root specially; it just walks modules/*/src - tools/check-layer-guards.sh, tools/install-bit.sh, tools/playground- dev.sh, tools/flaker-affected-rules.{toml,mjs}, tools/flaker-affected- rules.test.mjs all rewritten to use the modules/bit/src/* prefix - Taskfile.pkl: moonbitSources glob covers modules/**, build commands point at modules/bit/src/{lib,cmd/bit} - flake.nix / package.nix: moonModJson now resolves to modules/bit/moon.mod.json - .github/workflows/pages-demo.yml path filter updated - docs/package-layout.md rewritten with the new repository layout diagram and updated per-package path columns moon check passes (0 errors) and node tools/check-layers.mjs is clean.

Workspace mode produces _build artifacts under the module-prefixed path (e.g. _build/native/release/build/mizchi/bit/cmd/bit/bit.exe instead of _build/native/release/build/cmd/bit/bit.exe). Update the build script and Taskfile outputs to match. Also fix a stale mizchi/bit/x-hub/native reference in a doc comment.

Continue the gitoxide-style split by promoting every leaf (no internal mizchi/bit dependency) core/* package out of mizchi/bit into its own MoonBit module. The CLI surface is unchanged — only import paths in moon.pkg files move. Extracted modules: - bit_hash (gix-hash) - bit_date (gix-date, was core/date_parse) - bit_utils (gix-utils + gix-quote, was core/string_utils) - bit_trailers (gix-trailers) - bit_ignore (gix-ignore + gix-glob) - bit_archive (gix-archive, was core/tar) - bit_diff_core (gix-diff, low-level) - bit_diff3 (gix-merge, low-level) - bit_apply (patch application) - bit_fast_import - bit_osfs (gix-fs OS impl) - bit_bootstrap Naming convention: `bit_<name>` (no x) for git plumbing extracted from mizchi/bit, distinct from `bitx_<name>` used for non-git extensions. - moon.work now lists all 12 new core modules - modules/bit/moon.mod.json declares them as deps - bitx_doc / bitx_kv / bitx_rebase_ai / bitx_workspace add bit_osfs and/or bit_utils to their deps (they consumed those packages) - Blackbox test files (*_test.mbt) updated to use the new package aliases (e.g. @apply.X -> @bit_apply.X) - tools/check-layers.mjs recognises mizchi/bit_<name> as the core layer - docs/package-layout.md updated with the new layout diagram and per-module mapping table moon check passes (0 errors) and all 12 new modules' native tests pass (bit_apply 7, bit_archive 21, bit_date 6, bit_diff3 23, bit_diff_core 6, bit_fast_import 32, bit_hash 11, bit_ignore 16, bit_trailers 6, bit_utils 4 — bit_bootstrap and bit_osfs have no tests). mizchi/bit/lib (285/285) and mizchi/bit/cmd/git-bit (4/4) still pass too.

@bit

Three more core packages whose internal mizchi/bit/* dependencies were already resolved to bit_* modules in wave 1 are now standalone: src/object → modules/bit_object (gix-object) src/config_parse → modules/bit_config (gix-config) src/pack → modules/bit_pack (gix-pack) The trickier bit: bit_object owns the underlying type defs that mizchi/bit/types re-exports as aliases (`pub type ObjectId = @object.ObjectId`, etc). MoonBit's type checker needs the source module imported wherever a type alias is dereferenced through @bit.<Type>, so every moon.pkg that imports `mizchi/bit` and uses `@bit.GitError` / `@bit.ObjectId` etc. now also imports `mizchi/bit_object`. Same applies to bitx_doc / bitx_kv / bitx_rebase_ai / bitx_subdir / bitx_workspace / bitx_hub — they now declare bit_object in moon.mod.json deps. Long-term direction: move the type defs themselves out of bit_object into a future bit_types module, so the aliases become unnecessary and the transitive-import requirement goes away. moon check passes (0 errors). Tests: - mizchi/bit_object: 27/27 - mizchi/bit_config: 8/8 - mizchi/bit_pack: 72/72 - mizchi/bit/lib: 285/285

@bit

src/types → modules/bit_types (shared trait/type defs) src/protocol → modules/bit_protocol (gix-protocol + gix-transport) src/reftable → modules/bit_reftable (reftable backend) Same transitive-import handling as wave 2: every moon.pkg that imports mizchi/bit and uses @bit.<trait> (FileSystem / RepoFileSystem / etc.) now also imports mizchi/bit_types, and the containing modules declare mizchi/bit_types in moon.mod.json deps. moon check passes. Tests: - mizchi/bit_types: 0 (no tests) - mizchi/bit_protocol: 30/30 - mizchi/bit_reftable: 9/9 - mizchi/bit/lib: 285/285

src/io → modules/bit_io (gix-fs abstract + native bindings) src/refs → modules/bit_refs (gix-ref) Both modules now satisfy "no internal mizchi/bit/* runtime dep" after waves 1-3 carved out their direct dependencies. bit_io brings its `native` sub-package along (mizchi/bit_io/native). Consumers of @bit.TestFs and friends had to gain mizchi/bit_io imports + moon.mod.json deps via the same transitive-alias pattern as previous waves. moon check passes. Tests: - mizchi/bit_io: 0 (no tests) - mizchi/bit_refs: 2/2 - mizchi/bit/lib: 285/285

src/remote → modules/bit_remote (gix-url + gix-discover) src/repo → modules/bit_repo (repo handle / materialization) bit_repo introduces a new transitive: types defined here (ObjectStore, CommitInfo) are re-exported through the mizchi/bit root facade. Same pattern as previous waves — consumers gain mizchi/bit_repo imports + moon.mod.json deps. After this wave, mizchi/bit/src/ contains only pack_ops, repo_ops, worktree, diff, lib, vfs, fingerprint, grep, runtime, cmd/, tests/, fuzz_tests/. moon check passes. Tests: - mizchi/bit_remote: 4/4 - mizchi/bit_repo: 2/2 - mizchi/bit/lib: 285/285

…wave 6) src/pack_ops → modules/bit_pack_ops (collect_reachable_objects, etc.) src/repo_ops → modules/bit_repo_ops (revparse, repository ops) src/worktree → modules/bit_worktree (status / add / commit / rm / mv) src/diff → modules/bit_diff (high-level diff / show) These four still depend on mizchi/bit/lib at runtime, so they are classified as high (not mid) by tools/check-layers.mjs — same as the parking lot the docs already described. Once their lib dep is broken they can be promoted to mid. After this wave modules/bit/src/ contains only: lib, vfs, fingerprint, grep, runtime, cmd/{bit,git-bit}, tests/, fuzz_tests/, top.mbt. moon check passes. Tests: - mizchi/bit_pack_ops: 1/1 - mizchi/bit_repo_ops: 0 (no tests) - mizchi/bit_worktree: 1/1 - mizchi/bit_diff: 0 (no tests) - mizchi/bit/lib: 285/285

…ve 7) Pull the remaining non-lib high-layer packages out of mizchi/bit into their own modules: src/vfs → modules/bit_vfs src/fingerprint → modules/bit_fingerprint src/grep → modules/bit_grep src/runtime → modules/bit_runtime These still depend on mizchi/bit/lib at runtime, so they remain classified as the high layer (same as wave 6's mid extraction). After this wave modules/bit/src/ has just: lib (high), cmd/, tests/, fuzz_tests/, top.mbt. Consumers updated: bitx_kv / bitx_subdir add bit_vfs dep; bitx_workspace adds bit_fingerprint; bitx_rebase_ai adds bit_runtime. moon check passes. Tests (native): - mizchi/bit_vfs: 50/50 - mizchi/bit_grep: 28/28 - mizchi/bit_runtime: 14/14 - mizchi/bit_fingerprint: 7/9 (2 failures are environment-dependent git-commit signing tests, unrelated to the refactor) - mizchi/bit/lib: 285/285

src/lib → modules/bit_lib (gix-equivalent porcelain, 103 .mbt files, 285 tests, includes lib/native subpackage) bit_lib is the biggest consumer of the bit_* core split — it depends on nearly every previously-extracted bit_* module. 13 downstream consumers (every mid/high module plus all bitx_* extensions and the CLI) now declare mizchi/bit_lib in their moon.mod.json deps. After this wave modules/bit/src/ holds only the cmd/ binaries, the top.mbt root facade, and the integration test/fuzz test packages. moon check passes. mizchi/bit_lib: 285/285 native tests pass.

js-build: - tools/bit-git.mjs imports the lib payload at the new workspace path (_build/js/release/build/mizchi/bit_lib/bit_lib.js instead of the pre-split _build/js/release/build/lib/lib.js). - Taskfile.pkl JS lib build targets modules/bit_lib/src and the output glob is mizchi/bit_lib/bit_lib.js. sync-npm-lib-raw and the bun-bundle tasks updated to match. - tools/playground-dev.sh builds modules/bit_lib/src. ci.yml: - Native binary stage uses the workspace output path _build/native/release/build/mizchi/bit/cmd/bit/bit.exe. - JS / WASM smoke tests target the new module names (mizchi/bit_lib, mizchi/bit_runtime, mizchi/bit_diff3, mizchi/bit_repo, mizchi/bit_grep). - distributed-test points at mizchi/bitx_hub / bitx_hub/native / bitx_kv instead of the old mizchi/bit/x-* paths. nix-build: - moonPlatform.buildMoonPackage does not understand multi-module workspaces (it tries to fetch every dep from mooncakes.io, which fails for the local mizchi/bit_* / bitx_* members). Replace it with a custom mkDerivation that: 1. strips workspace-local deps before invoking buildCachedRegistry so the cached registry only resolves real mooncakes.io packages 2. uses bundleWithRegistry to materialize MOON_HOME with the toolchain + cached registry 3. runs `moon build --target native --release modules/bit/src/cmd/bit` inside the source tree to honor moon.work resolution - flake.nix devShell moonHome uses the same stripped manifest so `nix develop` still works. Local verification: - node --test tools/{js-build,npm-lib,npm-cli,demo-relay, demo-editor-link,playground-commands,playground-view}.test.mjs: 24/24 pass - nix build cannot be exercised in this sandbox; will be confirmed by CI.

Same workspace-output-path fix as the test job, applied to the git-compat shards' 'Build bit' step. moon build now targets modules/bit/src/cmd/bit and the binary path is _build/native/release/build/mizchi/bit/cmd/bit/bit.exe.

The workspace split multiplies the number of modules moon test walks, and the previous '--no-parallelize -j 1' configuration pushed the job past its 30-minute deadline (see PR #82 run 26488539393 timeout). The extracted bit_* modules are independent enough that parallel test execution should be safe; bump the timeout to 60m as a safety net.

moon test --target native walked all workspace members and timed out even at 60 minutes (1h14m wall before kill). Splitting it makes the slow consumer visible and lets each fail/timeout independently: test: loops over modules/bit_*/bitx_* members + the tests/ and fuzz_tests/ packages in mizchi/bit. Skips modules/bit so cmd/bit doesn't dominate. 30-minute timeout. cmd-native-test: a new job dedicated to mizchi/bit/cmd/bit and mizchi/bit/cmd/git-bit native tests, which are heavy (large tcc linker step). 75-minute timeout. Both jobs reuse the same build + setup steps so the wbtest fixtures keep working. The non-cmd loop runs every workspace member with an explicit -p so a single broken module fails fast.

cmd/bit native tests still don't fit in 75m after the workspace split (actual elapsed ~88m before timeout). Split the job into a matrix so cmd/bit and cmd/git-bit run in parallel runners: - cmd/bit: 120-minute timeout (heavy tcc link) - cmd/git-bit: 30-minute timeout (small consumer) Both still rebuild the wbtest fixture binary up front because the test harness reads tools/git-shim/moon.

moon build leaves _build/.moon-lock held when the process exits on some workspace configurations; the subsequent moon test then blocks forever waiting for the lock instead of running. Add the same pkill + rm guard that the test job already uses between its JS and native steps. https://claude.ai/code/session_015aa85JK5yNfkrqt2uhodvi

TCC (the fallback C compiler moon uses when no system compiler is found) has very slow link times for large binaries. The cmd/bit package links 30+ workspace modules and takes >120 min with TCC. Clang links the same binary in a fraction of the time. ubuntu-latest runners ship with clang pre-installed, so no CI setup change is needed. https://claude.ai/code/session_015aa85JK5yNfkrqt2uhodvi

moon.pkg sets link.native.cc = "clang" for cmd/bit and cmd/git-bit. The nix sandbox does not expose clang by default, so the build fails immediately when moon tries to invoke it. Add clang to nativeBuildInputs so the compiler is available inside the sandbox. https://claude.ai/code/session_015aa85JK5yNfkrqt2uhodvi

moon build output path may shift across moonbit versions; use find to locate bit.exe rather than hard-coding the path. Also print moon --version so the actual version is visible in CI logs. For the cmd-native-test lock issue: when link.native.cc=clang is set, moon spawns clang child processes that may still hold _build/.moon-lock after the moon process itself exits. Add pkill -f clang + sleep 1 to the lock-clearing step so no stray compiler holds the lock when moon test starts. https://claude.ai/code/session_015aa85JK5yNfkrqt2uhodvi

… in workspace) x/agent, x/agent/llm, x/agent/mcp, x/agent/native were part of the old monolithic src/ layout but were not extracted into a bitx_agent workspace module. The distributed-test job still referenced them, causing immediate failure on every CI run. Remove the missing package references; bitx_hub and bitx_kv cover the distributed testing scope. https://claude.ai/code/session_015aa85JK5yNfkrqt2uhodvi

…esolution moon.work member paths require ./ prefix for moonbit workspace resolver to correctly resolve local module references. Without it, moon may fail to recognize workspace members (exit 255) and fall back to registry lookup. Also update CI module-list parser to strip the ./ prefix before using basename for the mizchi/<name> package path.

- bit_repo_ops/moon.mod.json: add mizchi/bit_utils dep (required by moon workspace resolver - all transitive deps must be declared) - js_exports_wbtest.mbt: pass depth=0 to js_clone_remote (PR #30 added depth param, test wasn't updated)

cc=clang disables tcc-run, forcing moon test to do a full clang link on every test run. For the cmd/bit package (40+ module workspace), the clang link takes 50+ minutes and holds the build lock the entire time, causing moon test to block. tcc is fast and produces working binaries; there is no practical benefit to clang for test builds. Also remove clang from package.nix nativeBuildInputs (no longer needed since moon uses its bundled tcc for the nix build).

In workspace mode (moon.work), moon outputs binaries under a namespaced path to avoid conflicts: _build/native/release/build/mizchi/bit/cmd/bit/bit.exe instead of _build/native/release/build/cmd/bit/bit.exe. The wbtest files were looking for the binary at the old non-workspace path, causing moon test to try to rebuild (blocking on the lock file). Update 11 files with the correct workspace output path.

mizchi changed the title ~~refactor: split into moon.work workspace (modules/bit + modules/bitx_*)~~ refactor: split into moon.work workspace (bit + bit_* core + bitx_* extensions) May 26, 2026

mizchi force-pushed the claude/moon-work-module-split-3BBmH branch from 239965d to c0d720c Compare May 29, 2026 12:21

claude added 22 commits May 29, 2026 12:54

ci: fix git-compat job build path

ef32860

Same workspace-output-path fix as the test job, applied to the git-compat shards' 'Build bit' step. moon build now targets modules/bit/src/cmd/bit and the binary path is _build/native/release/build/mizchi/bit/cmd/bit/bit.exe.

mizchi force-pushed the claude/moon-work-module-split-3BBmH branch from 7ffcc1b to d597ee6 Compare May 29, 2026 12:55

claude added 3 commits May 29, 2026 13:16

mizchi merged commit d2efc9d into main May 29, 2026
16 checks passed

mizchi deleted the claude/moon-work-module-split-3BBmH branch May 29, 2026 15:46

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

refactor: split into moon.work workspace (bit + bit_* core + bitx_* extensions)#82

refactor: split into moon.work workspace (bit + bit_* core + bitx_* extensions)#82
mizchi merged 26 commits into
mainfrom
claude/moon-work-module-split-3BBmH

mizchi commented May 26, 2026 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

mizchi commented May 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Layout

Tooling updates

Test plan

Reviewer notes

Commits

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

mizchi commented May 26, 2026 •

edited

Loading