Workspace Reorg Outcome (2026-05-08)

Naming note (2026-05-08): The CI guard binary referenced as vox-layer-check in this narration was renamed to vox-arch-check later in the same series; references below are historical.

Companion to 2026-05-08-workspace-reorg-design.md. Records what was delivered across the 10 phases.

Follow-up landed (2026-05-15): The crate-org followup design delivered C1 (vox-mcp-meta merge), C2 (vox-package-types split — removed both vox-package inversions), C4 (ops_ludus → vox-gamify), the mcp-server feature-gate side-quest, all Track A SSOT fixes, Track B description rewrites, and PR6 arch-check lints. Workspace now ~91 crates. vox-orchestrator has regrown to 65.5K LoC. C3 (vox-cli-ci, 22K LoC) and C5 (vox-orchestrator-core) are deferred with plan docs.

Phases delivered (5 of 10)

Phase 0 — Baseline & guards ✓

New crates/vox-layer-check/ (Rust binary): parses cargo metadata, validates each dep edge against layers.toml. Modes: --warn-only and strict.
New docs/src/architecture/layers.toml — 79 crates assigned to L0–L5.
New build-time-baseline.md and per-phase build-time-log.md.
CI step wired up.

Phase 1 — L0 type cleanup + plugin-host inversion fix ✓

New vox-plugin-types (L1 leaf): plugin manifest, skill manifest, PluginStateBackend trait + PluginStateSkillEntry/PluginStateError.
vox-plugin-host no longer depends on vox-db — uses Arc<dyn PluginStateBackend> instead. vox-db impls the trait.
Daemon binary [[bin]] required-features = ["mcp-native"].
Drive-by fixes: pre-existing mesh_driver_compile test; pre-existing vox-mens gpu-feature broken crate::training::native.

Phase 2 — workspace-hack leaf exclusion ✓

Hakari’s [traversal-excludes] and [final-excludes] exclude L0 leaves.
Truly-leaf incremental check at 0.5s (no hack compile in the dep graph).

Phase 4 — Extract vox-orchestrator-mcp + vox-orchestrator-d ✓

The big one. mcp_tools/ (33,885 LoC) → new crate vox-orchestrator-mcp.
services/routes/ (axum HTTP routes) → moved with mcp.
bin/vox_orchestrator_d.rs → new package vox-orchestrator-d.
vox-orchestrator drops the mcp-native feature and 14 deps that mcp now owns (schemars, axum, rmcp, tower-http, vox-compiler, vox-grammar-export, vox-mcp-registry, vox-capability-registry, vox-openai-wire, vox-project-scaffold, vox-skills, vox-openclaw-runtime, vox-plugin-host).
vox-cli’s mcp-server feature wires through to the new crate.
Orchestrator incremental: 5.59s → 4.06s (−27%). MCP edits are now isolated.

Phase 5 — Extract vox-orchestrator-queue ✓

locks/, oplog/, affinity.rs, sync_lock.rs (~3K LoC) → new crate vox-orchestrator-queue.
Four pure-data types moved to vox-orchestrator-types to keep the queue crate as a clean leaf: SnapshotId, SnapshotIdGenerator, ChangeId, FileAffinity, AccessKind.
Orchestrator incremental: 4.06s → 3.58s (−12% additional, −36% total).

Phase 9 — Strict CI guard + final docs ✓

Layer-check flipped from --warn-only to strict in CI.
79 crates fully assigned. 3 known inversions documented with rationale.
Outcome doc + per-phase build-time log.

Phases deferred (5 of 10)

Phase	Target	Why deferred
3	vox-db → vox-db-stores	67 `impl VoxDb` blocks need extension-trait migration; ~50 callers; vox-compiler dep is structural
6	vox-orchestrator-runtime	runtime.rs / orch_daemon/ / dei_shim/ form the orchestrator’s core; extracting requires trait facade over Orchestrator (~40 methods) or moving Orchestrator out (empties parent)
7	vox-cli-thin (full)	Most orchestrator-using commands in vox-cli are already gated behind features (`dei`, `live`, `mcp-server`, `ludus-hud`). The few remaining unconditional uses (`generate`, `model/`, `ci/`) deliver ~0.5s additional savings, not worth the gating complexity now that Phase 4+5 trimmed 36K LoC out of orchestrator.
8	Plugin family flatten	Already structurally clean: cdylib delivery; vox-cli/vox-orchestrator have no compile-time deps on any plugin. L4 → L3 plugin→db deps are allowed by the layer model.

Headline build-time outcome

Scenario	Baseline	Final	Win
Orchestrator incremental	5.59s	3.58s	−36%
CLI incremental	26.76s	6.99s	−74%
L0 leaf (true leaf)	—	0.53s	new floor (was paying workspace-hack)
MCP isolated	(in 5s)	5.15s	now parallel with orchestrator
Queue isolated	(in 5s)	<2s warm	now parallel with orchestrator

vox-orchestrator went from 88K LoC → 52K LoC. The 36K-LoC reduction came from extracting vox-orchestrator-mcp (33K) and vox-orchestrator-queue (3K).

What’s enforced going forward

Layer-check runs strict in CI. Any new dep edge that violates the layer model fails the build. Adding a new crate requires an entry in layers.toml.
L0 leaf crates stay leaves. Hakari’s exclusion list prevents workspace-hack from re-poisoning them on regeneration.
The plugin-host inversion stays fixed. PluginStateBackend is the only path; reintroducing a direct vox-db dep on vox-plugin-host triggers the layer-check.
Three known inversions are documented (current as of 2026-05-15):
- vox-cli → vox-orchestrator (deliberate; runtime/observability surfaces)
- vox-arch-check → vox-compiler (dev-dependency only; integration test)
- vox-ml-cli → vox-cli (optional mens-dei workflow path)
- (The vox-pm → vox-compiler/db inversion listed here at reorg time was removed in C2 when vox-package was split into vox-package-types (L1) + vox-package (L3).)

New workspace shape

Crates added by this reorg (all L3 unless noted):

vox-plugin-types (L1) — pure types, plugin/skill manifests, state backend trait
vox-orchestrator-mcp — MCP tool layer (33K LoC moved from orchestrator)
vox-orchestrator-queue — locks/oplog/affinity/sync_lock (3K LoC moved)
vox-orchestrator-d (L5) — daemon binary package
vox-layer-check (L0) — CI guard tool

Crates added to vox-orchestrator-types:

agent_types::workspace_ids (SnapshotId, ChangeId)
agent_types::file_affinity (FileAffinity, AccessKind)
agent_types::ids (existing)
agent_types::switch (existing)

Remaining work for future sessions

vox-orchestrator has regrown from 52K → 65.5K LoC (94% of budget). vox-cli is at 71.3K LoC (79% of budget). The deferred extractions have been redesigned:

Phase 6 replacement → C5 / Tier D: Extract src/orchestrator/ (12.8K LoC) + Orchestrator struct into vox-orchestrator-core. Full plan with Rust coherence analysis at 2026-05-15-orchestrator-tier-d-plan.md. Start when Rule 13 fires (>15% LoC growth since last tag).
Phase 7 replacement → C3: Extract src/commands/ci/ (22K LoC, 74 files) into vox-cli-ci. Three shared modules must move to vox-cli-core first. Full plan at 2026-05-15-cli-ci-extraction-plan.md. Start when vox-cli exceeds 72K LoC (80% budget).
Phase 3 (vox-db-stores): Still deferred. 67 impl VoxDb blocks need extension-trait migration.

The infrastructure (arch-check, layers.toml, hakari exclusions) is in place to support all of these.