[idea] Resume agents with their original launch flags (--dangerously-skip-permissions, --model, …) after a restart

[iosmanthus]

idea / problem

When herdr restarts and resumes a native agent session, it rebuilds the command as just <agent> --resume <id> and drops every flag the pane was launched with. A pane I started as claude --dangerously-skip-permissions --model opus comes back as plain claude --resume <id> — so I get re-prompted for permissions and lose --model. herdr already records the original command in launch_argv and persists it in the session snapshot; the resume path just doesn't use it.

requested change

When a pane has a recorded launch_argv, replay that original command and append the agent's resume reference, instead of rebuilding from scratch:

claude --dangerously-skip-permissions --model opus --resume <id>

How it works:

• Strip any session selector the recorded command already carries before appending the new one — --resume/-r, --continue/-c, codex's resume subcommand — so the replayed command can never carry two conflicting selectors. (codex's -c/--config is intentionally left alone — it isn't --continue.)
• Fall back to today's minimal <agent> --resume <id> when there's no recorded launch command (e.g. agents detected from a hand-typed shell).
• Carry launch_argv onto the resumed terminal so the flags survive repeated restarts, not just the first.

Like the auto-hibernation idea, this is mostly an automatic use of primitives herdr already has (launch_argv capture + deferred resume) rather than new machinery.

I have a working implementation (src/agent_resume.rs + src/persist/restore.rs, with unit tests across the supported agents) and can open a PR if this direction is welcome.

[iRonin]

+1 — and this also fixes harness forks, not just flags. A fork/renamed build that reports as its parent agent (e.g. ronin, a fork of pi — same sessions under ~/.pi/agent, reports agent="pi") currently resumes as the parent binary pi, because plan() rebuilds from the hardcoded recipe. Replaying launch_argv means a pane started as ronin --session … comes back as ronin — which removes the only remaining reason such forks impersonate + PATH-shim the parent binary.

So this generalizes cleanly from "preserve flags" to "preserve the actual binary," and covers the fork case for free. Strong +1 — happy to test the fork path against your implementation.

[Fanduzi]

+1 — hit exactly this. I launch claude --dangerously-skip-permissions --disallowedTools "Bash(rm:*)" ...; after a reboot the resumed pane drops back to default permission mode and loses the dangerous-command denylist — a real safety regression, not just convenience (losing --model). Replaying launch_argv fixes both. Today's workaround is moving the rules into ~/.claude/settings.json, which works but is a second place to manage and easy to forget on resume. Happy to test the PR on macOS / Claude Code.

[iosmanthus]

Try this branch if you’re interested in this feature: https://github.com/iosmanthus/herdr/tree/release-20260622

[pranaygp]

Another data point, a workaround that works today, and a possibly even smaller native shape than replaying launch_argv.

I run Claude Code panes as claude --dangerously-skip-permissions (and codex as codex --dangerously-bypass-approvals-and-sandbox); every restore downgrades them to default permission mode. Since herdr currently types the resume line into the pane's interactive shell (start_pending_agent_resume in src/app/agent_resume.rs), this zsh function repairs it today:

# herdr types `claude --resume <id>` into the pane shell on restore
claude() {
  if [[ -n "$HERDR_ENV" && "$1" == "--resume" ]]; then
    command claude --dangerously-skip-permissions "$@"
  else
    command claude "$@"
  fi
}

It works, but it's zsh-only, copy-pasted per agent, and it leans on resume being typed into a shell — if that ever becomes a direct spawn, it breaks silently.

Given #1080 drew the boundary at "no process/launch tracking", maybe the minimal native shape is neither replay nor derivation, but a static config override of the per-agent resume template in config.toml, falling back to today's built-ins:

[session.resume_commands]
claude = ["claude", "--dangerously-skip-permissions", "--resume", "{session_id}"]
codex  = ["codex", "--dangerously-bypass-approvals-and-sandbox", "resume", "{session_id}"]

Still fully declarative: nothing tracked or parsed, the session ref stays data substituted into one argv slot, and agents without an override behave exactly as today. It also covers the fork-binary case above (point the template at ronin) and profile dirs via ["env", "CLAUDE_CONFIG_DIR=/x", "claude", ...]. It composes fine with the session-file derivation idea from #1080 too — explicit config wins when set, derivation as the smart default.