spawn

Spawn a subagent to run work off-thread. The subagent either invokes a target skill or executes a raw prompt directly. The harness infers which from the input — no explicit mode flag. In both paths the subagent's full context stays out of the parent; the parent gets the result.

Inputs

• <input> — a target skill identifier (as the harness addresses skills) optionally followed by arguments, OR a raw prompt for the subagent to execute directly. The harness detects which: if <input>'s leading token matches a known skill name, treat it as a skill invocation with the remainder as arguments; otherwise treat the whole <input> as a raw prompt.

Procedure

1. Inspect <input>. If its leading token matches a registered skill name (via the harness's skill-listing mechanism), treat it as a skill invocation: <skill-name> = that token, <arguments> = the remainder. Otherwise treat the entire <input> as a raw prompt. Do NOT read the target skill's manifest from the parent.
2. Discover subagent profiles via the harness's native mechanism. Default: general-purpose profile, full tool access, foreground.
3. Spawn a subagent via the harness's native mechanism:
   • Skill invocation: prompt the subagent to invoke the target skill via the harness's skill-invocation mechanism with <skill-name> and <arguments>, then execute under that skill's directives. Front-load any context the target needs beyond <arguments>.
   • Raw prompt: prompt the subagent to execute <input> directly. Front-load all context the task needs; subagents are stateless.
4. Collect the subagent's return via the harness's native result-reading mechanism and output it. The target skill's body and the subagent's working context never enter parent context.

Output

The subagent's raw return.