arch 0001 migration plan

ARCH-0001 migration plan

Status: accepted — implemented. Date: 2026-09-11 · Issue: ARCH-0001 Phased repo split: core library, deploy targets, dashboard Owner: Capsize Games (maintainer) · Depends on: plans/arch-0001-adr-repo-topology.md

Decision

The split was executed in four ordered phases. Phase 1 (multiple distributions in one repository) was mandatory and landed; Phases 2–4 fired only when their trigger (plans/arch-0001-decision-metrics.md) was met. Phase 2 (dashboard), Phase 3 (spikeforge-targets), and Phase 4 (spikeforge-hub) shipped; the Phase 4 spikeforge-server extraction did not fire (T4 no-go). Every phase was additive before subtractive and reversible: the new repository was created and pushed from git history before anything was deleted here, and the pre-phase commit is the rollback point.

Phase 1 — monorepo, multiple distributions (no repo moves)

Goal: make the boundaries real without moving a file.

1. Add the protocol/ contract authority, the protocol_version envelope field, the TS codegen/validation step, and tests/test_protocol_schema_parity.py (plans/arch-0001-protocol-contract.md).
2. Add the packages/ workspace with packages/spikeforge/pyproject.toml (excluding server/) and packages/spikeforge-server/pyproject.toml (plans/arch-0001-packaging-versioning.md).
3. Extend scripts/blocked_deps/sitecustomize.py with fastapi, pydantic, uvicorn; add scripts/check_core_boundary.py; add the headless CI step (plans/arch-0001-core-boundary.md).
4. Remove the web extra from core; make it the server distribution's base dependencies. Update docker-compose.yml and Dockerfile to install spikeforge-server.
5. Port tests/test_packaging_profiles.py to read the new packages/*/pyproject.toml files instead of setup.py.

No test moves. Tests stay where they are because nothing has moved.

Rollback: revert the Phase 1 PR. Core packaging returns to setup.py; the protocol work is additive and can remain.

Phase 2 — extracted the dashboard (trigger T1)

1. On a throwaway clone, git subtree split --prefix=client -b spikeforge-dashboard-split and push to a new capsize-games/spikeforge-dashboard.
2. Move the dashboard CI job (npm ci && npm run build) and the client/package-lock.json cache to the new repo.
3. In core, pin the dashboard bundle version in compatibility.json; the server serves a pinned prebuilt bundle instead of building client/ in-repo.
4. Keep client/ in core for one release as a read-only mirror, then delete it.

Test relocation: the parity tests (tests/test_client_animation_payload.py and the schema parity test) stay in core, because core owns protocol/ and is the cross-contract authority. Their inputs become the pinned generated TS artifact rather than a live client/ checkout.

Rollback: restore client/ from the mirror (or from the pre-phase tag); the server build reverts to in-repo client/.

Phase 3 — extracted spikeforge-targets (trigger T2)

1. On a throwaway clone, ran the git filter-repo multi-prefix extraction from plans/arch-0001-packaging-versioning.md and pushed to capsize-games/spikeforge-targets with import root spikeforge_targets.
2. Deleted the four legacy submodules (targets, energy, event_runtime, plus hub) and targets/backends from the core package, and added the spikeforge-targets pinned dev dependency. Because the project is pre-1.0 and unpublished, no re-export shim was added (see "No back-compat aliases").
3. Moved console-script ownership of spikeforge-energy and spikeforge-targets to the new distribution.
4. Moved the tests for the moved code into spikeforge-targets; the server's imports now use spikeforge_targets.* (spikeforge_targets.energy, spikeforge_targets.event_runtime).

Test relocation rule: tests move with their subject. Tests that reference targets, energy, or event_runtime relocated to capsize-games/spikeforge-targets. Tests that exercise the seam (the server driving a target through the protocol) stay in the cross-package integration suite described below.

Rollback: restore the four prefixes from the pre-phase tag and drop the pin.

Phase 4 — extracted spikeforge-hub (go) and spikeforge-server (no-go)

• spikeforge-hub (go, trigger T3; extracted). Single-directory git subtree split of spikeforge_hub to capsize-games/spikeforge-hub with import root spikeforge_hub; moved the spikeforge-hub console script and hub tests. No re-export shim was retained.
• spikeforge-server (conditional, trigger T4; no-go). T4 did not fire, so no capsize-games/spikeforge-server repository was created. The split would run only when the server must release on its own cadence — git subtree split --prefix=server to capsize-games/spikeforge-server. The server keeps import root server (plans/arch-0001-target-topology.md), and its 17 server-touching test files and the server/schemas/* modules would move with it.

Protocol ownership after Phase 4 (if it fires): protocol/ remains in capsize-games/spikeforge as the contract authority. The server and dashboard repos consume it as a pinned artifact or git submodule; they never fork it.

Cross-package integration suite

Regardless of how far extraction goes, one thin suite stays co-located with the protocol authority (core) and installs the pinned satellites. It guards the seams rather than the internals:

• tests/test_client_animation_payload.py — the client-facing payload contract.
• tests/test_protocol_schema_parity.py — pydantic models vs JSON Schema vs protocol_version.txt.
• one server-through-core smoke test that drives a real request over the protocol using the pinned spikeforge-server.

Everything else migrates to the owning repository.

Docs strategy

plans/ stays authoritative and in the core repository (plans/index.md); scripts/build_docs.sh keeps globbing plans/*.md and this migration does not move the docs pipeline. Satellite repositories carry only a minimal README.md and link back to the core docs site for shared design. If extraction reaches Phase 4 and the site must build per repo, the core site remains the hub page; no plans/ content is duplicated.

No back-compat aliases

The extraction shipped without legacy import-path aliases. Because the project is pre-1.0, unpublished, and has no external importers to protect, the old spikeforge.{targets,energy,event_runtime,hub} module paths were deleted outright rather than kept alive as a deprecated re-export shim:

Removed core submodule	Replacement root
targets	spikeforge_targets
targets.backends	spikeforge_targets.backends
energy	spikeforge_targets.energy
event_runtime	spikeforge_targets.event_runtime
hub	spikeforge_hub

There is no DeprecationWarning window, no "1 minor release" shim lifetime, and no alias to retire later. Legacy WebSocket payload keys, main.py, main_encodings.py, and the legacy checkpoint keys _fc1/_lif1/_fc2/_lif2 are not affected and remain stable throughout — this migration is about packaging topology, not the public runtime surface.

Reversibility summary

Every phase is guarded by its trigger (plans/arch-0001-decision-metrics.md), landed as one reviewable PR, and tagged beforehand. Because extraction pushes a new repository from history before deleting anything, a failed phase is reverted by checking out the pre-phase tag and deleting the pin — no data is lost. Risks and their mitigations are tracked in plans/arch-0001-risk-register.md.

Related decisions

• plans/arch-0001-adr-repo-topology.md — the topology decision.
• plans/arch-0001-packaging-versioning.md — pinning and the extraction commands.
• plans/arch-0001-core-boundary.md — what Phase 1 enforces.
• plans/arch-0001-decision-metrics.md — when each phase fires.
• plans/repo_topology_plan.md — the source analysis.