feat(sessions): kickoff prompts, resume-on-respawn, and full CLI parity

[defangdevs]

Motivation

From a main/interactive agent-box session (Claude or Codex, on the phone via Remote Control) you should be able to kick off work in sibling sessions — possibly on a different harness — and reach everything the settings UI can do, without opening the web UI or doing a rebuild. Sessions already had UI/CLI CRUD over the same on-disk sessions.json; this fills the gaps.

Base: stacked on #128 (fix/remote-control-name-empty-host). My supervisor rewrite references hostLabel/remoteControlHost introduced there, so this depends on #128. Rebase base to master once #128 merges.

What this adds

Kickoff prompt (first spawn only). New initialPrompt is handed to the harness as its positional prompt argument on the first spawn; the supervisor then consumes it (mark_started sets hasRun, records boxSessionId, clears initialPrompt) so a respawn never redoes the task.

Resume-on-respawn. Every later spawn — crash, reboot, Spot stop→restart, all of which keep the transcript because /home is the persistent root EBS volume — resumes the prior transcript instead of redoing or idling. boxSessionId is the stable id we own across respawns:

• Claude consumes it directly as --session-id (first spawn) / --resume (respawn, when a transcript exists; otherwise a fresh --session-id) — exact, so concurrent sessions never cross.
• Codex mints its own id, so we stamp [agent-box session <id>] into the kickoff prompt and locate the exact rollout by that marker (grep → newest → trailing-36-char UUID) → codex resume <uuid>. No match → start fresh; never resume --last, which could grab a sibling session's transcript in a shared cwd.

resumePrompt overrides the built-in steer (continue unfinished work, or stop if already done).

CLI parity (agent-box-session):

• add … --prompt "TASK" [--resume-prompt "…"]
• env ls | set KEY VALUE | rm KEY — writes the same ~/.config/agent-box/env the settings page manages and the env-exec wrapper reads (ls shows keys only, never values)
• restart NAME | --all

Settings page: add-session forms gain an optional kickoff-prompt field wired to initialPrompt.

Discoverability: seeded AGENTS.md documents --prompt, resume semantics, env, and restart --all.

User-visible / security effects

• No new network surface. env writes the existing user-owned 0600 file with the same KEY_RE charset as the settings daemon and the env-exec wrapper; values are never surfaced by ls.
• Kickoff/resume prompts are passed via printf %q, same injection-safe path as extraArgs.
• No AWS cost / IAM / networking / migration impact.

Checks run

• nix-instantiate --parse modules/agent-box.nix — OK
• nix build .#checks.x86_64-linux.module-single-file and .multi-user — evaluate cleanly (.drv produced; x86_64 build and VM tests couldn't run on the aarch64 dev box without /dev/kvm).
• Validated the non-trivial bash/jq in real bash+jq: rollout-UUID extraction, concurrent-session marker disambiguation, both jq filters, env_rewrite set -e safety.
• Verified flag composition against the installed CLIs (Claude --session-id/--resume/positional prompt; Codex resume [id] [PROMPT] + bypass flag on the subcommand).

Needs CI (KVM): the new tests/sessions.nix subtests (kickoff deliver-once + resume state, env, restart --all, web-form prompt).

Not verified headlessly: that a positional prompt injects as a real first turn in a live --remote-control interactive launch, and that --session-id composes with --remote-control — needs a real attach on a deployed box.

🤖 Generated with Claude Code