ADR-0003: Primitive Index/Search/Harvest and Full Multi-Target Support In Scope

Status: Accepted

Context

The reference branch adds a hand-rolled BM25 primitive-index/search/harvest subsystem and multi-target support (vscode, vscode-insiders, copilot-cli, kiro, windsurf, claude-code) alongside its architecture refactor. Initial migration-plan drafting recommended deferring both as a separate epic, to avoid conflating "align architecture" with "ship new product surface." On review, the product argument for inclusion was judged stronger: without search/discovery, installing a bundle first requires already knowing it exists, which weakens the CLI's UX; multi-target support is also largely modeled already (core's Target union already includes all six types).

Investigation while planning surfaced a nuance worth recording: only the Kiro target has a real content transformer (app/transform/transformers/kiro-transformer.ts) in the reference branch — Windsurf and Claude-Code currently fall back to a no-op transformer, i.e. only the type/writer plumbing exists for them, not real content adaptation.

Decision

Include both in scope for feat/ai-primitives-hub-next, sequenced as their own dedicated commits/sub-phases (harvest+search infra in Phase 3, discovery/search orchestration and multi-target transforms in Phase 4, index/discover CLI commands alongside install/uninstall — not after — in Phase 5), rather than deferred. Windsurf and Claude-Code transformers are written for real during Phase 4, not copied as no-ops.

Consequences

• Positive: the CLI ships with real discoverability and full multi-target support from its first release, matching the reference branch's product ambition.
• Negative: materially larger scope than a pure architecture-alignment effort; Phases 3–5 carry more work than originally scoped. Mitigated by keeping this work in its own commits, separate from the pure domain-extraction commits, so review and git bisect aren't complicated by mixing the two concerns.