#244 Connector research memo: Cursor interception surfaces

[Anonymous]

Originally created by: fu351

Doberman mediates via the MCP proxy, Claude Code host hooks, and the turn gate. Widening to another agent is a real maintenance commitment, so the next connector gets chosen with evidence rather than enthusiasm.

Codex CLI has since been built (src/doberman/hosthooks/codex.py, adapters/codex/), so this memo is now scoped to Cursor alone.

Deliverable — a decision memo, not code

• A capability matrix for Cursor hooks: interception points, what verdict envelope each accepts, whether an ask-the-human confirm is expressible, payload limits, and how stable the surface has been across releases.
• The fail-closed honesty test: can a deny be expressed such that the client cannot mistake silence for a crash? A surface where an uninstalled state cannot emit an explicit allow, or where a deny is indistinguishable from a failure, is disqualifying.
• A recommendation: whether Cursor is worth building a connector for, and the envelope contract it would use.

Memo lands in docs/ (or the PR description) with citations to Cursor's current hook documentation. Research issue — strong writeups welcome even without a code contribution.

[Anonymous]

Originally posted by: fu351

Narrowed to Cursor. The Codex CLI half is moot now that Codex is a shipped connector (src/doberman/hosthooks/codex.py), so the memo just needs the Cursor capability matrix and its fail-closed honesty test. Dropped to level-4 since it's now a single-surface writeup.