feat: generalize executor to support arbitrary shell commands (not just Claude Code)

[jimmy-claw]

Problem

Currently the executor framework only supports one payload type: a Claude Code prompt. Long-running background tasks (nix builds, cargo builds, sequencer runs) need tracking too — but today we handle these with fragile nohup ... & + cron monitor patterns.

Proposal

Generalize the Task abstraction to support two payload types:

pub enum TaskPayload {
    ClaudeCode { prompt: String, max_turns: u32, workspace: Option<PathBuf> },
    ShellCommand { command: String, args: Vec<String>, env: HashMap<String, String>, working_dir: Option<PathBuf> },
}

SSH/container/local executors already run remote processes — they just need to branch on payload type instead of always invoking claude.

CLI interface

# New: arbitrary command
openclaw-agent run --executor crib --cmd "nix build" --workdir ~/logos-lez-multisig-module
openclaw-agent run --executor crib --cmd "cargo test --workspace" --workdir ~/lez-framework

All existing features unchanged: task IDs, log streaming, dashboard visibility, completion callbacks.

Benefits

• For Jimmy (agent): no more nohup ... & + cron monitor patterns. Dispatch a build, get notified when done.
• For Václav: all background work (coding tasks AND builds/jobs) visible in one dashboard panel.
• OpenClaw integration: main session can call openclaw-agent run to dispatch heavy work without blocking.

Implementation notes

• Add ShellCommand variant to TaskPayload in executor-core
• SSH/container/local executors: branch on payload type
• CLI: add run subcommand (or extend start with --cmd flag)
• Dashboard: show task type (🤖 Claude / ⚙️ Command) in task list

Related

• Issue feat: native task completion callback — remove need for openclaw system event in prompts #6 (completion callbacks) — needed for both payload types