[Planning][P3] skilllite-services shared orchestration layer (Phase 1 redesign)

[EXboys]

Summary

Architecture docs describe a future skilllite-services crate sitting between entry crates (skilllite, skilllite-assistant, future MCP entry) and domain crates, hosting shared application orchestration used by both CLI and Desktop.

Phase 0 decisions were recorded (TASK-2026-042), a bootstrap crate was created (TASK-2026-043), and a Phase 1A migration was attempted (TASK-2026-044) — then rolled back (TASK-2026-045) because coupling risk outweighed benefit before the Assistant split contract was complete.

This issue tracks a redesigned Phase 1 that respects the split-ready L1/L2/L3 model.

Problem

Today, shared flows (workspace resolution, evolution orchestration, skill discovery) are duplicated or accessed via:

• Direct path deps in Desktop (blocks repo split)
• Ad-hoc logic in skilllite-commands
• Growing --json CLI surface (preferred for split cleanliness)

We need a deliberate decision on when a shared Rust library layer adds value vs when L2 CLI JSON is sufficient.

Design constraints (from architecture)

1. Assistant must not path-dep services at split time — only subprocess to skilllite binary
2. Prefer L2 CLI JSON over in-process services for split cleanliness
3. If services are needed, expose as skilllite desktop-serve daemon rather than linking from Assistant
4. Dependencies flow one way: Entry → Services → Domain crates

Proposed scope (Phase 1 redesign)

Option	Description	Split-friendly?
A — CLI-only	Continue expanding --json subcommands; no skilllite-services crate	✅ Best for split
B — Engine-internal services	skilllite-services used by CLI + optional daemon; Assistant calls via subprocess	✅ If daemon/CLI only
C — Shared library in Assistant	Path-dep services from Desktop	❌ Blocks P4 extract

Recommendation: Start with Option A inventory; adopt Option B only for flows where CLI JSON is too slow or too fragmented (with caching/TTL in bridge).

Acceptance criteria

• Written ADR / CONTEXT doc comparing Options A/B/C with explicit decision
• Inventory of duplicated orchestration between CLI and Desktop (with file references)
• If Option B chosen: bootstrap skilllite-services with one vertical slice (e.g. workspace resolution) + deny.toml boundary rule
• If Option A chosen: close this issue with checklist of remaining --json gaps
• No regression to Assistant split P2 exit criterion (zero engine path deps)

References

• docs/en/ENTRYPOINTS-AND-DOMAINS.md — Dependency direction
• docs/en/ASSISTANT-SPLIT-ARCHITECTURE.md §7 — Interim skilllite-services
• TASK-2026-042 (Phase 0 decisions), TASK-2026-045 (rollback rationale)