safe-subagent-spawn

**Initial release of safe-subagent-spawn — a reliability-focused skill for safe, standardized subagent creation and multi-round lifecycle management.** ### Why this skill exists OpenClaw's `sessions_spawn` tool accepts many parameters and supports multiple spawn modes (run, ACP, Discord session, etc.), but: 1. **Models frequently get the parameters wrong.** GPT-5.4 and other models commonly confuse `agentId` (a real agent id from `agents.list`) with model strings like `"google/gemini-2.5-pro"`, include unsupported fields, or pick incorrect default values — causing repeated trial-and-error failures that waste tokens and time. 2. **Non-run modes are fragile for most users.** ACP mode and Discord session mode depend on locally installed packages and additional infrastructure that most non-developer users don't have. These modes are a frequent source of hard-to-diagnose failures. This skill exclusively uses **run mode** — the most stable and universally available spawn mode — eliminating an entire category of setup-related errors. ### What's included - **safe_subagent_spawn.py** — Validated wrapper that generates a ready-to-use `sessions_spawn` JSON payload. Enforces correct invariants (`runtime: subagent`, `mode: run`, `thread: false`, `cleanup: keep`, no `streamTo`), validates that `agentId` and `model` are not confused, and ensures the context file contains at least one directive before allowing a spawn. The output is passed directly to `sessions_spawn` with no manual assembly required. - **create_context.py** — Creates a structured Markdown context file with metadata (timestamp, task slug, file path), standing instructions for the child, background information, and an optional first directive. Task slugs are validated (lowercase alphanumeric + hyphens, 1–48 chars). Supports deferred directives — create context with background only, then collect external information before formulating the first task. - **append_to_context.py** — Appends directives, child outputs, or external messages to an existing context file. Features include: - **Automatic round numbering** — tracks round numbers across directive/child-output pairs, keeping them in sync. - **Strict sequential ordering** — enforces that Directive Round N exists before Child Output Round N can be appended, and Child Output Round N-1 exists before Directive Round N+1. Prevents out-of-order spawns. - **External message merging** — consecutive external messages are merged into a single section without redundant headers, keeping context files clean. - **Content deduplication** — silently discards duplicate appends (identical content to the latest entry of the same role), preventing accidental double-writes. - **Flexible input** — accepts content via `--content` (inline), `--content-file` (from file), or stdin (pipe/heredoc), accommodating both short directives and large child outputs. ### Key design decisions - **Full verbatim history** — every child output is preserved in full; no summarization, no truncation. The next child in a multi-round workflow reads the entire context file and acts on the latest directive. - **Context files are permanent** — never deleted, providing a complete audit trail of all subagent interactions. - **Parent never reads back** — the parent agent tracks context files by path and task slug but never re-reads them after creation, avoiding context window pollution. - **No automatic retry** — if a subagent times out, the parent reports the failure to the user instead of silently retrying, ensuring transparency. - **No nested subagents** — standing instructions in every context file prohibit the child from spawning its own subagents, keeping the delegation tree flat and predictable.