Resume prompted Hermes sessions literally

Hermes TUI startup queries execute slash, shell, interpolation, and multiline syntax before reaching the model. Run the prompt through literal one-shot mode, read its exact session ID from a private usage report, and resume that session in the TUI so arbitrary prompt text stays data while the conversation remains interactive.

Co-Authored-By: Codex XHigh <noreply@openai.com>

bin/omarchy-agent

@@ -88,8 +88,7 @@ codex)
   ;;
 hermes)
   if [[ -n ${prompt:-} ]]; then
-    # Keep inherited flags after chat so older Hermes subparsers do not clear them.
-    command=(hermes chat --yolo --tui "--query=$prompt")
+    command=(omarchy-agent-hermes "$prompt")
   else
     command=(hermes --yolo)
   fi

bin/omarchy-agent-hermes

@@ -0,0 +1,21 @@
+#!/bin/bash
+
+# omarchy:summary=Seed Hermes literally and resume the resulting interactive session
+# omarchy:args=<prompt>
+# omarchy:hidden=true
+
+set -euo pipefail
+
+prompt=${1:?usage: omarchy-agent-hermes <prompt>}
+usage=$(mktemp)
+trap 'rm -f "$usage"' EXIT
+
+# TUI startup queries pass through Hermes' slash, shell, and interpolation
+# dispatcher. One-shot treats the prompt literally and records its session ID.
+HERMES_SESSION_SOURCE=tui hermes --yolo --usage-file "$usage" --oneshot "$prompt"
+session_id=$(jq -er '.session_id | strings | select(length > 0)' "$usage")
+
+rm -f "$usage"
+trap - EXIT
+
+exec hermes chat --yolo --tui --resume "$session_id"

test/shell.d/default-agent-test.sh

@@ -462,7 +462,7 @@ assert_launch claude claude --permission-mode auto -- "Review this project"
 assert_launch codex codex --approve-for-me -- "Review this project"
 assert_launch crush crush run "Review this project"
 assert_launch grok grok --permission-mode bypassPermissions -- "Review this project"
-assert_launch hermes hermes chat --yolo --tui "--query=Review this project"
+assert_launch hermes omarchy-agent-hermes "Review this project"
 assert_launch agy agy --dangerously-skip-permissions --prompt-interactive "Review this project"
 assert_launch copilot copilot --allow-all --interactive "Review this project"
 pass "agent launcher adapts initial prompts for every supported agent"

test/shell.d/hermes-agent-test.sh

@@ -0,0 +1,79 @@
+#!/bin/bash
+
+set -euo pipefail
+
+source "$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" && pwd)/base-test.sh"
+
+test_tmp=$(mktemp -d)
+trap 'rm -rf "$test_tmp"' EXIT
+
+mock_bin="$test_tmp/bin"
+oneshot_log="$test_tmp/oneshot"
+resume_log="$test_tmp/resume"
+source_log="$test_tmp/source"
+mkdir -p "$mock_bin"
+
+cat >"$mock_bin/hermes" <<'SH'
+#!/bin/bash
+
+if [[ " $* " == *" --oneshot "* ]]; then
+  printf '%s\0' "$@" >"$HERMES_TEST_ONESHOT_LOG"
+  printf '%s' "${HERMES_SESSION_SOURCE:-}" >"$HERMES_TEST_SOURCE_LOG"
+
+  while (( $# )); do
+    if [[ $1 == "--usage-file" ]]; then
+      usage=$2
+      break
+    fi
+    shift
+  done
+
+  [[ ${HERMES_TEST_ONESHOT_FAIL:-false} == "false" ]] || exit 42
+  [[ ${HERMES_TEST_USAGE_FAIL:-false} == "false" ]] && printf '{"session_id":"session-123"}\n' >"$usage"
+  printf '%s\n' response
+  exit
+fi
+
+printf '%s\0' "$@" >"$HERMES_TEST_RESUME_LOG"
+SH
+
+chmod +x "$mock_bin/hermes"
+
+export PATH="$mock_bin:$PATH"
+export HERMES_TEST_ONESHOT_LOG="$oneshot_log"
+export HERMES_TEST_RESUME_LOG="$resume_log"
+export HERMES_TEST_SOURCE_LOG="$source_log"
+
+sentinel="$test_tmp/hermes-seed-must-stay-literal"
+prompt="!Crash /quit {!touch $sentinel}"$'\ntrailing\\'
+"$ROOT/bin/omarchy-agent-hermes" "$prompt" >/dev/null
+
+mapfile -d '' -t oneshot_args <"$oneshot_log"
+(( ${#oneshot_args[@]} == 5 )) || fail "Hermes literal seed has five one-shot arguments"
+[[ ${oneshot_args[0]} == "--yolo" ]] || fail "Hermes literal seed enables yolo mode"
+[[ ${oneshot_args[1]} == "--usage-file" ]] || fail "Hermes literal seed requests the session report"
+usage_file=${oneshot_args[2]}
+[[ ${oneshot_args[3]} == "--oneshot" && ${oneshot_args[4]} == "$prompt" ]] ||
+  fail "Hermes literal seed remains one argument"
+[[ ! -e $usage_file ]] || fail "Hermes literal seed removes its session report"
+[[ ! -e $sentinel ]] || fail "Hermes literal seed never executes prompt interpolation"
+[[ $(<"$source_log") == "tui" ]] || fail "Hermes literal seed records an interactive session"
+
+mapfile -d '' -t resume_args <"$resume_log"
+[[ ${resume_args[*]} == "chat --yolo --tui --resume session-123" ]] ||
+  fail "Hermes literal seed resumes the exact completed session"
+pass "Hermes sends initial prompts literally and resumes their exact session"
+
+: >"$resume_log"
+if HERMES_TEST_ONESHOT_FAIL=true "$ROOT/bin/omarchy-agent-hermes" failure >/dev/null 2>&1; then
+  fail "Hermes literal seed reports a failed initial turn"
+fi
+[[ ! -s $resume_log ]] || fail "Hermes literal seed does not resume a failed initial turn"
+pass "Hermes does not resume after a failed initial turn"
+
+: >"$resume_log"
+if HERMES_TEST_USAGE_FAIL=true "$ROOT/bin/omarchy-agent-hermes" missing-session >/dev/null 2>&1; then
+  fail "Hermes literal seed requires a recorded session ID"
+fi
+[[ ! -s $resume_log ]] || fail "Hermes literal seed does not guess which session to resume"
+pass "Hermes resumes only the session recorded by the initial turn"