feat(minimal): add min task run — declared tasks in ephemeral sessions - #1139

sequenceDiagram
  participant User
  participant CLI
  participant TaskRunner
  participant Session
  participant TaskProcess
  User->>CLI: min task run task
  CLI->>TaskRunner: dispatch task arguments
  TaskRunner->>Session: create and prepare ephemeral session
  TaskRunner->>Session: execute task
  Session->>TaskProcess: stream command input
  TaskProcess-->>Session: stream stdout, stderr, and exit status
  Session-->>TaskRunner: return task result
  TaskRunner->>Session: destroy or retain session
  TaskRunner-->>CLI: return task exit code
  CLI-->>User: print task output and exit

_{Comment @coderabbitai help to get the list of available commands.}

along with a TTL etc

Absolutely! That would be fantastic

Filed a follow-up tracking the ephemeral/task marking on the session record + TTL (with the reaper deliberately coordinated with the broader session-lifecycle design rather than ahead of it). Thanks @twitchyliquid64.

Addressed the exit-code review feedback in 8aabd89:

• Missing exit status now fails closed. The exec bridge tracked the remote status as a bare u32 initialized to 0, so a daemon or transport death that closed the channel before an ExitStatus message read as success. The bridge now tracks receipt explicitly (Option<u32>); a run that ends without a reported status prints task ended without reporting an exit status and exits 1. The status resolution is factored into a pure exit_outcome(Option<u32>) with unit coverage for the missing-status path.
• Ctrl-C behavior documented. The interrupt guard's doc comment and the task run help text now state that SIGINT is consumed on the host — the session is torn down (or kept with --keep) and the process exits 130 without relaying the interrupt into the box, so the task gets no chance to handle it, unlike a local exec.
• Non-interactive stdin documented. The bridge's doc comment and help text now spell out that tasks run non-interactively: a terminal stdin is half-closed immediately (the daemon exec channel has no PTY — only the shell path does), so input reaches the task only when piped; interactive work belongs in min session attach.

The other gap in the analysis — the daemon mapping signal deaths to unwrap_or(1) instead of 128+signo in crates/minimald/src/exec.rs — predates this PR and is tracked separately; not touched here.

Addressed the three CodeRabbit findings in 8a76e7a:

• Interrupt cleanup is now bounded and re-interruptible (task.rs guard). The Ctrl-C handler's connect + DestroySession runs under the same 10-second cleanup ceiling as best_effort_destroy (with a manual-cleanup hint if it times out), and is raced against a second ctrl_c() listener: a second interrupt during cleanup abandons it and exits 130 immediately. An unresponsive daemon can delay the exit, never prevent it, and a second Ctrl-C is no longer silently swallowed. The guard's doc comment now describes this behavior.
• Untrimmable task names rejected client-side (check_task_declared). The daemon trims the task name out of the exec command string, so a declared name with leading/trailing whitespace could never round-trip into the box. Such names are now rejected before any config lookup, with the name debug-quoted so the whitespace is visible (e.g. task name " build " has leading or trailing whitespace...). Unit-tested, including that the rejection fires ahead of the minimal.toml discovery.
• e2e leftover-session checks fixed for pipefail (scripts/session-e2e.sh). The two mnl ls | grep -q checks could falsely pass: grep -q's early exit SIGPIPEs the ls under pipefail, turning a genuine leftover into a non-zero pipeline. Both now capture mnl ls output to a variable and glob-match it — the same convention the script's attach-proof markers prescribe. Shellcheck stays clean.