Execution Layers — Per-CLI Detail
Read when binding a spawn interface or diagnosing a fallback. _common/CLI_COMPATIBILITY.md owns current commands, model availability, flags, paths and version-specific workarounds. reference/hub-authoring.md owns the outcome brief.
Claude Code
| Layer | Purpose | Preconditions |
|---|---|---|
| L1: Direct Spawn | One bounded specialist session | Advertised agent tool and allowed task effects |
| L2: Parallel Spawn | Independent branches, then join | Supported concurrency, exclusive writers, available budget |
| L3: Rally Delegation | Complex ownership/team coordination | Actual team/subagent support; justified coordination cost |
Use the runtime's actual schemas, not pseudocode copied from older documentation. Preserve inherited permissions. A worktree prevents direct write collisions but does not prove compatible APIs, schemas or aggregate constraints. Join required workers and verify the integrated revision.
Codex CLI
Prerequisites
Apply _common/CODEX_ORCHESTRATION.md C1/C5. Discover advertised tools and effective capacity/nesting limits. A lazily exposed tool requires supported discovery; a guessed spawn_agent/wait_agent call is not a capability probe.
Execution layers
L1 is a single native subagent. L2 launches independent work before joining required results. L3 uses supported resume/team coordination when the task requires it. Tool names, parameter shapes and limits come from the active host. A shell-launched CLI is an alternative only when installed, authorized and appropriate to the ownership boundary.
Runtime notes
Inherit the authorized model, approval and sandbox settings. Validate result status, outputs, stderr and required effects. Current structured output is preferred; historical detached-terminal defects require installed-version reproduction before imposing foreground-only execution. Never convert missing output into a clean review.
Antigravity CLI (agy)
Prerequisites
Apply _common/AGY_ORCHESTRATION.md. Verify installed CLI help, authorized account/model and the task's required capabilities. TUI commands are not automatically shell commands or tools available in a custom subagent.
Execution layers
Use advertised native agents or the documented headless interface. Prefer structured results when available; parallelize only isolated calls within budget and join them before synthesis. Do not assume preview teamwork is needed, or enable it automatically.
Runtime notes
_common/CLI_COMPATIBILITY.md §9.1–§9.2 defines permissions and result handling. Normal headless execution does not require permission bypass or a PTY. Capture exit status, structured result, stderr and declared artifacts. Denied required tools, timeouts and missing outputs remain incomplete even when the process exits zero.
A reproducible legacy stdout/TTY defect may justify version-scoped PTY or explicit file output. Associate fallback artifacts with the exact current run; never harvest an unrelated “latest transcript,” accept a stale file or swallow a process failure. Remove the workaround once its reproducer passes normally.
Pre-flight Notification
Notify the user before a material new effect or cost; obtain approval where the contract requires it. Informational narration does not authorize a bypass. Never recommend global allowlisting or change user settings merely to make spawning noninteractive.
Fallback to Internal Execution
Follow SKILL.md Core Rule #3: record the verified capability blocker and permitted alternatives already checked. Internal execution is not an independent specialist session and cannot expand Nexus's control-plane ownership or a read-only skill's boundary. Preserve frozen ACs, identify any unavailable independent verification and cap status accordingly. Do not repeatedly retry the same failed capability or install/enable a runtime without authorization.