migrate-to-submodule
Plan and execute moving non-host-specific code (and its tests) from the OpenHuman core (crates/openhuman-core) into the vendored tiny* submodule libraries (vendor/tiny*), then release the submodule and re-pin the registry digests. Use when asked to "move X out of core", "upstream this into tinyfoo", "extract a crate", or to shrink the host to composition/policy only.
インストール方法を見る含まれるファイル(1)
- SKILL.md10.1 KB
SKILL.md(原文)
インストールする前に、エージェントに与えられる指示の中身を確認できます。
Migrate core code into a vendored submodule
OpenHuman is the host: composition, config, security policy, approvals,
lifecycle, and adapters. Behavior that is not specific to OpenHuman belongs in
the owning vendor/tiny* project (see "Submodule ownership" in CLAUDE.md).
This skill drives one migration end to end: plan, move code + tests, ship the
submodule release, re-pin the host.
Default to executing. Use plan mode only for step 1 (the split decision), and
only when the boundary is genuinely ambiguous or the move is large. Work in a
worktree (worktree <slug> from the repo root); never on main. Do not squash
commits; the auto-commit hook checkpoints for you.
How TinyBus fits
TinyBus is the seam that makes this migration possible and shapes its design.
- A loadable module is a first-party
cdylibspeaking the TinyBus module ABI, downloaded from a pinned GitHub release and attached to an in-process broker as a bus peer. The core calls it through a typed contract, not a Rust import. - Each module has a
*-buscontract crate (names, methods, request/response types, contract version). It is synchronous and free of I/O and runtime deps. Never redeclare a contract type in OpenHuman; call members through the contract constants. - Because the module is a separate binary in the same process, anything you move
must be reachable from the host through the bus contract. So a migration is
usually three pieces: (a) implementation in the module crate, (b) new/changed
methods and types in the
*-buscrate (bump its contract version), (c) a thin host adapter undercrates/openhuman-core/src/modules/<name>.rs(or the domain that calls it) that forwards to the bus and keeps policy/config/creds. - Data that cannot cross a boundary (task-local state, callbacks, secrets) is
passed explicitly in the request, or served to the module as a host callback
(see
modules/memory_host.rs). Credentials never ride in tool arguments; they go through module init/reinit config. - Module code must not depend on
openhuman. Host callbacks are traits/bus services the module defines and the host implements. - Not every submodule is a loadable module. Pure libraries (
tinyagents,tinytools,tinychannels,tinyskills,tinyvoice,tinywallet, ...) are linked via Cargo; a move there needs no bus contract or digest re-pin, only a gitlink bump. Check the module'sMODULE.md/README to know which kind it is.
Step 1: Plan (freedom to decide the split)
-
Identify the code to move and read its callers. Use Explore/Grep on
crates/openhuman-core/src/<domain>/. -
Classify every item as moves or stays:
Moves to the submodule Stays in OpenHuman Algorithms, parsers, protocol/wire types, engines, generic tools, provider clients with no product identity Config schema and loading, SecurityPolicy/approvals/sandbox decisions, credential resolution (resolve_backend_credential),Outcome<T>controllers and RPC schemas, event-bus publishing, product identity/attribution headers, Redux/Tauri glueTests that exercise only the moved logic Cross-module tests, host-adapter tests, JSON-RPC E2E, anything needing the mock backend or CoreContext -
Pick the destination using the ownership table in
CLAUDE.md. Put the change in the repo that owns the behavior, not where it is easiest to land. Read the target'sREADME.md,AGENTS.md, crate boundaries and*-buscrate first. -
You may create new crates in the submodule workspace when the code has no good home (e.g. a new
tinyfoo-<thing>library crate, or a new*-busif a new module needs its own contract). Prefer extending an existing crate. Only propose a brand-new submodule/repo when no existing project owns the concept; that one is a user decision, so stop and ask. -
Write the plan down briefly (in the PR description or
docs/plans/): what moves, what stays, new crates, bus contract changes, dependency direction, and the release order. Skip a formal doc for a small, obvious move.
Dependency rules to check while planning: the core must not depend on
tinyhumans-sdk (only openhuman-tinyhumans may); a module must not depend on
openhuman; there is one tinytools copy (via vendor/tinyagents/); tool-call
parsing/dialects/agent loop belong in tinyagents, not the host.
Step 2: Move the code, upstream first
Submodule PRs go first, raised against that repo's canonical upstream
(tinyhumansai/<name>), never a fork. Follow the submodule's own AGENTS.md/
CONTRIBUTING.md (fmt, clippy -D warnings, per-file coverage gates such as
90% in tinysearch's release workflow).
- Branch inside the submodule at the recorded gitlink (the
worktreehelper already put every submodule on branch<slug>). Do not create nested worktrees. - Move the implementation into the target crate with
git mv-style history where practical. Keep public API minimal and documented; keep files under about 500 lines. - Move tests with the code. Unit tests go beside the moved modules in the
submodule. Add bus-contract tests in the
*-buscrate for any new wire type. Tests that need the host stay in OpenHuman. - If a bus contract changed, bump its version and update both sides in the same change set. Keep the contract synchronous and I/O-free.
- Replace the removed core code with the thin host adapter. Delete the old
implementation and its old tests, do not leave a copy. Keep
Outcome<T>controllers and RPC namespace strings (wire contracts) unchanged. - Verify in the submodule:
cargo fmt --all -- --check,cargo clippy --all-targets --all-features -- -D warnings,cargo test --all-features(plus the coverage gate if the repo has one). - Open the submodule PR ready for review, babysit to green, merge. Unshallow submodule clones before merging upstream if needed.
If the change spans nested submodules (tinyagents/vendor/tinytools,
tinymemory/vendor/tinycortex, ...), release/merge the innermost first and bump
gitlinks outward.
Step 3: Release the submodule
Only for loadable modules (a cdylib with a registry record). Pure Cargo
libraries just need the merged commit.
- Trigger the submodule's
Releaseworkflow (workflow_dispatchonmain, inputbump: patch/minor/major/current), e.g.gh workflow run release.yml -R tinyhumansai/<name> -f bump=patch. It runs fmt/clippy/tests, builds per-platform archives, tagsvX.Y.Z, and publishes a GitHub release withchecksum.toml. - Wait for it to finish (
gh run watch). Confirm the release has one archive per host key the registry lists (macOS 15/26 arm64+x86_64, Ubuntu 22.04/24.04 x86_64+arm64, Windows) pluschecksum.toml. - Ask before triggering a release if the user has not already authorized it: it publishes tags and artifacts and is hard to undo.
Step 4: Re-pin the host
Modules are pinned twice, and both must describe the same release
(scripts/ci/check-module-pins.mjs):
- Submodule gitlink:
git -C vendor/<name> fetch --tags && git -C vendor/<name> checkout vX.Y.Z, then stage the gitlink in the superproject. - Registry record: edit
crates/openhuman-core/src/modules/registry/records_*.rsfor the module:version,release_url, and everyPlatformAssetarchivename andsha256.- Copy digests verbatim from the published release's
checksum.toml(gh release download vX.Y.Z -R tinyhumansai/<name> -p checksum.toml). - Never compute a digest from a local build. A locally recomputed hash agrees with itself no matter what was served, which defeats the pin.
- At load time tinybus fetches the release's own
checksum.toml, compares it with the host pin, hashes the archive, then extracts. A re-cut tag stops matching instead of silently replacing what runs in-process.
- Copy digests verbatim from the published release's
- Update anything else that names the version (for example the pinned
archive version and sha256 in the CI workflows). If a module is deliberately off-tag, that needs an entry in
scripts/ci/module-pin-exemptions.jsonwith a reason. Prefer a fresh release over an exemption. - Bump the OpenHuman-side contract crate reference if the
*-busversion moved, and updatecrates/openhuman-core/src/modules/README.mdandplatform/about_app/if the layout or user-visible capability changed. - Verify:
Test with a feature both enabled and disabled if a gate moved. Never exportnode scripts/ci/check-module-pins.mjs node scripts/ci/check-submodule-monotonic.mjs pnpm rust:layout cargo check --manifest-path Cargo.toml cargo test -p openhuman-cli --test <affected> # or: pnpm debug rust <filter>CARGO_TARGET_DIR. Run long CI-style commands throughscripts/ci-cancel-aware.sh.
Step 5: Open the OpenHuman PR
- PR against
upstream(tinyhumansai/openhuman), ready for review, not draft. If the submodule release is not out yet, say so and mark draft only then. - Body: what moved, what deliberately stayed and why, new crates, bus contract changes, links to the submodule PRs and release, and the digest source.
- The gitlink must point at a commit that exists upstream (merged, tagged).
- Hand off to
pr-babysitterfor CI and review threads.
Checklist
- Every moved item is non-host-specific; host policy/config/creds stayed
- Tests moved with the code; cross-module/host tests stayed
- Bus contract updated and versioned; no redeclared contract types
- Submodule PR merged upstream first; fmt/clippy/tests/coverage green
- Release published with
checksum.toml(loadable modules) - Gitlink and registry
version/release_url/digests agree, digests copied from the release -
check-module-pinsandcheck-submodule-monotonicpass - Old core code and tests deleted, docs updated
レビュー
まだレビューはありません。使ってみた感想をお寄せください。