Alternatives and rejected directions

Per-repo daemon

Rejected for v1. kiki uses one user-scoped daemon so cross-repo visibility and daemon lifecycle are simple for the user.

Security boundary via workspaces

Rejected as a v1 claim. Workspaces are useful cooperative separation, but they do not constrain same-UID agents.

Mirroring jj flags

Rejected. kiki provides thread-aware porcelain. Users should run jj directly for arbitrary jj behavior.

Transcript-fed publishing

Rejected. Transcript prose is locally stored and often contains sensitive or dead-end context. Publishing and auto metadata features must not consume it. Consented catch-up or transcript-MCP output to a configured model provider is a separate, explicitly disclosed egress boundary with a thread/provider/purpose-bound grant.

General-purpose hook dispatcher in v1

Rejected. Claude Code runs matching hooks concurrently. Kiki requires exclusive PreToolUse control for managed v1 sessions instead of reimplementing Claude's hook merging, timeout, output, and failure semantics. A dispatcher may be reconsidered only after the cascade boundary is proven.

Hidden live-head bookmark

Rejected. The managed workspace's @ is the live thread head; the human bookmark is an explicit checkpoint and publication projection. A second hidden bookmark would create another pointer that can drift without solving a v1 requirement.

Arbitrary owned-stack DAGs in v1

Rejected. V1 explicitly rebases only a provable single-parent chain from synchronized base to thread head. Merge commits, foreign descendants, or multiple roots become TopologyDiverged; kiki does not infer a broader source revset.

Mirrored audit sinks and tamper-evident chaining

Rejected for v1. Each parseable attempt has exactly one authoritative SQLite destination: the target repo's audit_log when its repo resolves, otherwise the user database's user_audit_log. Mirroring the same attempt to JSONL or both databases, and adding hash chaining, would create another consistency protocol without defending against the documented same-UID threat model.

Automatic transcript redaction

Rejected for v1. Heuristic or model-based redaction creates false confidence and may erase the context recall is meant to preserve. Kiki discloses provider egress and asks for consent instead.