01
这个 skill 做什么
Use this skill when bounded coding work should be routed through the independent cursor-delegate CLI for inspect-only, proposal, or apply-mode execution. The skill owns reviewed task packets, orchestration, global authorization barriers, user link handoffs, acceptance, and a marker-validated private session plus standard-library lease wrapper for temporary packets/logs; the separate TypeScript CLI owns all Cursor SDK integration, model resolution, the packaged Next.js authorization/dashboard frontend, workspace safety, and status artifacts.
02
什么时候使用
- 01Dispatching implementation slices through delegation_session.py run with an exact session marker, unique log name, foreground lease, explicit scope, non-goals, stop conditions, and verification commands.
- 02Allowing Cursor task/Agent-tool internal subagents only under a packet policy with requested-model limits, concurrency, purpose, and write limits.
- 03Running planned single-stream work, or coordinating dependency-aware hierarchical workstreams, while preserving parent-agent ownership of shared contracts, integration, and final approval.
- 04Having non-root workstreams fail closed on missing authorization, then letting root stop the complete descendant tree, start the packaged Next.js frontend/control plane, and present its local authorization URL plus Cursor API Keys together without automatically opening either link before resuming the recorded agents after verification.
- 05Monitoring long Cursor SDK runs through status.json, including active and recent internal subagents.
- 06Following the real step/turn/request/tool/task sequence in a bounded execution tape while preserving streamed assistant and reasoning text across terminal reconciliation and replay gaps.
- 07Running focused repair loops when Cursor output needs narrow follow-up rather than broad re-planning.
- 08Keeping copied packets and ordinary skill-managed logs outside the target repository, while an explicit retained-log directory remains caller-owned and is never cleaned, then safely deleting only the root-owned temporary session after final acceptance or reconciled abandonment.
03
如何工作
- 01
Classify the task as direct Cursor, planned single-stream, hierarchical orchestration, or blocked; build a dependency ledger and effective-concurrency bound only for hierarchical work.
- 02
Create one private marker-owned session, then prepare the bounded task packet and run every dispatch through delegation_session.py with the exact marker and a unique log name; use an absolute caller-owned retained-log directory only for durable evidence that cleanup must never delete.
- 03
Resolve the catalog-defined Grok 4.5 High non-Fast preset for new top-level Cursor runs; validate any structured model evidence the SDK reports without inventing optional evidence, and treat internal task/Agent-tool model labels as requests because the current SDK does not expose their structured parameters.
- 04
Check that cursor-delegate is installed; non-root agents use fail-only auth, and root recursively confirms every descendant and owned process stopped before starting the CLI-owned loopback flow and presenting both status URLs to the user.
- 05
Review Cursor reports, task/Agent tool-call evidence, diffs, verification results, scope boundaries, and repair output before final acceptance.
- 06
After all CLI/frontend processes, authorization paths, follow-ups, and workstreams are terminal, preserve bounded evidence and run the allowlisted cleanup helper for the exact owned session.
04
你会得到什么
- 01A routing decision with mode, risk gates, workspace strategy, Cursor mode, model defaults, internal subagent policy, and hierarchical scheduling limits when applicable.
- 02A Cursor task packet or repair packet ready for reviewed CLI dispatch.
- 03A final review verdict grounded in logs, diffs, verification evidence, and any internal subagent reports.
- 04A cleanup disposition showing that the exact skill-owned session was removed or retained for a specific continuation/ownership reason.
05
重要边界
- 01Do not use this skill unless the user explicitly invokes delegate-to-cursor-sdk or $delegate-to-cursor-sdk.
- 02Treat the independently installed cursor-delegate command as a prerequisite; installing the skill does not install the CLI.
- 03Do not install SDK dependencies in the skill directory or recreate the CLI with ad hoc code when cursor-delegate is missing.
- 04Do not allow Cursor internal subagents unless the task packet includes a Cursor Internal Subagent Policy.
- 05Treat needs_authorization as a global barrier: do not leave other agents running, and do not resume them after cancellation or failure.
- 06Keep loopback authorization root-owned; never interrupt a child with an unaccounted CLI/frontend process, and never automatically open either authorization URL.
- 07Keep session cleanup ownership with root throughout; non-root agents hand over evidence and process disposition only, never ownership.
- 08Never delete caller-owned packets, retained audit directories, repository files, branches/worktrees, or ambiguous sessions; cleanup is marker-bound, allowlisted, root-owned, and gated on final acceptance or explicitly reconciled abandonment.
- 09Treat cleanup as a one-time ownership-proven action: after the marker is deleted, a repeated call must refuse because ownership cannot be proven again.
- 10Do not treat support, planning, workstream, or internal subagent output as final authority without upstream review.
- 11Do not commit, push, deploy, rotate credentials, alter billing, or broaden scope unless explicitly requested and reviewed.
06