acq run: acq run --name X -- CMD (no agent positional) silently no-ops instead of failing closed #253

Summary

acq run with a --name but no known-agent positional (e.g. acq run --name s2verify -- sh -lc '…') produces no output, creates no sandbox, and drops the post--- command entirely. It should fail closed with a clear message.

Reproduce

./acq --backend msb run --name s2verify -- sh -lc 'echo hi'
# -> prints nothing, creates nothing; `msb ls` shows "No sandboxes found."

Root cause

In the run) dispatch (acq), args are split at -- into create_args and agent_args. first=$(first_positional "$create_args") skips flags and their values, so with only --name s2verify there is no positional and first is empty. That fails both is_known_agent "$first" and acq_backend_exists "$first", so control reaches the final else "pass through" branch:

else
  # Pass through.
  acq_backend_attach "${create_args[@]+\"${create_args[@]}\"}"

This calls acq_backend_attach --name s2verify — never provisions, silently discards agent_args, and attaching a nonexistent sandbox no-ops. Net: a silent no-op.

Expected

acq run is create-and-attach for a known agent (acq run opencode …). When invoked without a resolvable agent/sandbox, it should fail closed (non-zero) with an actionable message, e.g.:

acq: run: no agent specified. Usage: acq run <agent|sandbox-name> [create-args] [-- CMD].
     To create a non-agent sandbox use `acq create …`; to run a command in an
     existing sandbox use `acq exec <name> -- CMD`.

This aligns with AGENTS.md "fail closed on ambiguity — never guess / no silent failures."

Notes

• Discovered while writing a host-side verification for the ADR-0015 post-hoc port-publish path (PR feat(acq)!: sbx↔msb backend parity, incl. live openchamber-on-msb (omnibus #234) #233); the operator hit the silent no-op and it wasn't obvious why.
• Fix is small (guard the empty-first case in the run) branch), but it changes a user-facing error path, so it warrants its own PR + a test-acq case asserting non-zero exit + the message.

AI-assisted (OpenCode); human-reviewed.