Skip to content
Conduit

skill

cross-cli-handshake

Cross-CLI Role Handshake: one delegation policy turning claude, codex, and agy into each other's pseudo-subagents

jpbaking/cross-cli-handshakecommunitylistedunreviewed

Set up

Skills are installed by placing their folder where your agent loads skills.

The download comes from github.com/jpbaking/cross-cli-handshake itself, not from our server, so what you get is exactly what its author published.

Install instructions come from the README

Skills do not use an MCP config. Follow the steps in the README below.

README

synced Oct 22, 2026

cross-cli-handshake

One delegation policy, three orchestrators.

The Cross-CLI Role Handshake turns the coding-agent CLIs installed on one machine — claude (Claude Code), codex (OpenAI Codex), and agy (Google Antigravity) — into each other's pseudo-subagents. Whichever CLI you are driving is the orchestrator; it delegates suitable work to the other two via shell, with a role marker (CROSS_CLI_ROLE=DELEGATE / CROSS_CLI_ORCHESTRATOR=<name>) that prevents delegation loops while keeping every delegate's own skills, rules, and hooks active.

Previously this policy lived as three hand-synchronized documents (one per CLI). This toolkit replaces them with a single canonical, orchestrator-neutral source:

AGENTS.md              # authoritative repository guidance
CLAUDE.md              # generated @AGENTS.md bridge
rules/shared/
  core.md                # handshake, routing principles, write preconditions,
                         # budgets, cooldowns, verification, destructive-action policy
  delegates/claude.md    # per-delegate profile: exact invocation, resume
  delegates/codex.md     # pattern, quirks, and "route here when" criteria
  delegates/agy.md
AGENT-INSTALL.md         # the agent-guided, merge-aware install contract

The same four policy files install identically for all three harnesses; each CLI reads the policy as addressed to itself and skips its own delegate profile. The repository and install layout follows Portable Agent Authoring: shared policy bodies stay neutral, while each host receives one marker-guarded pointer in its global instruction file.

What the policy covers

  • Role handshake — orchestrator/delegate markers, prompt-injection resistance (markers only count in the first two lines of a direct invocation), role persistence across resumed sessions.
  • Task routing — capability profiles per delegate (strong tier: Claude, Codex; fast tier: Agy), failure fallbacks, unavailability handling.
  • Write-task preconditions — clean git baseline, one writer and one orchestrator per working tree, secrets scoping.
  • Runtime budgets and conversation-local usage-limit cooldowns.
  • Verification — the orchestrator reviews every delegate diff and sends findings back to the same delegate session before ever fixing directly.
  • Destructive actions — never delegated; delegates must stop and report.

Install

This is a user-global, user-machine toolkit — it configures the CLIs on your machine and never touches projects. There is no project-level install.

Paste this prompt into any of your coding agents:

Fetch https://raw.githubusercontent.com/jpbaking/cross-cli-handshake/main/AGENT-INSTALL.md and follow it end to end.

Installs are merge-aware: your existing ~/.gemini/GEMINI.md, ~/.codex/AGENTS.md, and ~/.claude/CLAUDE.md content is preserved, and any legacy hand-maintained copy of this policy is only removed with your approval. Re-run the same procedure to update; it is per-user, per-machine, and single-version.

License

0BSD

↑↓ navigate openesc close