[idea] Follow the foreground process into a wrapper's inner PTY (uim-fep and similar FEPs)

[moezakura]

idea / problem

I use uim-fep, a Japanese input-method front-end. It owns the pane's PTY and runs my shell on an inner PTY, so herdr's foreground process is always uim-fep and nothing under it is detected. Wrapping the shell is just how a FEP is used — it's the normal setup for CJK input, not an exotic one.

HERDR_AGENT works, but it's only read from the pane's foreground process, so it has to sit on the FEP. When the FEP wraps a long-lived shell, the hint sticks for the whole life of the pane. Checked on an HERDR_AGENT=claude uim-fep pane:

• inner program is vim → still reported as claude / idle
• agent prompt on that pane → returns agent_prompted, and the text lands in the vim buffer

(herdr 0.7.5 stable, Linux, uim-fep 1.9.6)

requested change

When the pane's foreground process is a transparent PTY wrapper, run the same foreground-pgid resolution once more on its inner PTY instead of stopping at the wrapper.

Not a descendant scan — just the existing resolution applied one level down. It follows whatever is actually in the foreground inside, so running vim in the FEP means no agent is detected. A static HERDR_AGENT hint can never do that.

Gate it on a small manifest of known pass-through wrappers (uim-fep, skkfep, and similar) so nothing changes for anyone not using one.

Screen manifests shouldn't need to change. The pane buffer is what the FEP composited, and with each agent under uim-fep with its default status line I get real rule matches, not fallbacks:

• Claude Code → live_prompt_box ✅
• Codex → osc_title_working ✅
• Devin → workspace_trust_prompt ✅

Linux would be /proc-based; macOS and Windows would need their own implementation.

why you want this

Right now the choice is: keep the FEP on the shell and lose agent detection, or move it onto each agent and lose Japanese input in the shell. I took the second one, and this is all it takes:

#!/usr/bin/env bash
# uim-agent-exec <herdr-agent-kind> <command> [args...]
set -euo pipefail
kind=$1; shift

# outside a herdr pane, no uim-fep, or already inside one: run it as-is
if [ -z "${HERDR_PANE_ID:-}" ] || [ -n "${UIM_FEP_PID:-}" ] || ! command -v uim-fep >/dev/null 2>&1; then
  exec "$@"
fi

# HERDR_AGENT has to sit on uim-fep, since that is the pane's foreground process
exec env "HERDR_AGENT=$kind" uim-fep -f "$HERDR_PANE_ID" -e "$@"

exec uim-agent-exec claude claude "$@"    # one of these per agent
exec uim-fep -f "$HERDR_PANE_ID" -e fish  # and one to get an IM shell back, no HERDR_AGENT

Plus dropping the line in my shell rc that used to wrap every pane in uim-fep.

That works — all my agents are detected. What it costs is that a plain pane can't type Japanese any more: I run the second command first, and have to leave it again before starting an agent. A shell where I write commit messages in Japanese is where that shows up.

If you think the wrapper above is good enough and the answer is "wrap the agent, not the shell", close this with that — I'd rather have it written down somewhere than keep guessing.