fix(spawn): write launch command to a file instead of typing it whole

.agents/skills/harness-adapters/references/harness/rovo.md

@@ -36,7 +36,7 @@ rovo launches BARE (`rovo run --yolo`, plus any `--model`/`--config-override` fl
 2. **Typed pointer**: `Read the brief at <absolute-path> and follow it exactly.`, submitted through `fm_backend_send_text_submit` (the exact wording and mechanism kimi uses).
 3. **Delivery gate** (`rovo_wait_for_delivery`): composer empty AND either the echoed pointer text (`Read the brief at`) has scrolled into view or rovo's `Context:` footer percentage has advanced off zero. rovo's real footer is `Context: <bar> N.N% NN.NK/NNNK` (e.g. `Context: ▎ 3.3% 30.1K/922K`); the delivery regex tolerates the bar glyph and arbitrary spacing but anchors to the digits before the `%`, so the always-nonzero denominator (`.../922K`) can never masquerade as usage.
 
-A positional brief is dead-on-arrival: `rovo run --yolo "<brief>"` loads, never enters a working state, and drops back to an idle shell within about 10-15 seconds - confirmed independently four times over a raw PTY and once under real tmux 3.6a with the exact `fm-spawn.sh` send-keys shape. `--startup-receipt` cannot rescue that shape either: it requires "prompt-free interactive mode" (`Invalid value: --startup-receipt requires prompt-free interactive mode in a terminal`), so it cannot gate a launch that will have a message typed into it. The launch-then-send shape, by contrast, is confirmed live end to end (bare launch -> `Welcome to Rovo!` -> typed pointer -> `Rovo is thinking` for a real bash tool call -> clean `/exit`); see `../../../../docs/verification/rovo.md`.
+A positional brief is dead-on-arrival: `rovo run --yolo "<brief>"` loads, never enters a working state, and drops back to an idle shell within about 10-15 seconds - confirmed independently four times over a raw PTY and once under real tmux 3.6a with `fm-spawn.sh`'s then-current send-keys shape (typing the full launch line; `fm-spawn.sh` now types a short line sourcing a launch file instead, see its header). `--startup-receipt` cannot rescue that shape either: it requires "prompt-free interactive mode" (`Invalid value: --startup-receipt requires prompt-free interactive mode in a terminal`), so it cannot gate a launch that will have a message typed into it. The launch-then-send shape, by contrast, is confirmed live end to end (bare launch -> `Welcome to Rovo!` -> typed pointer -> `Rovo is thinking` for a real bash tool call -> clean `/exit`); see `../../../../docs/verification/rovo.md`.
 rovo leaves no worktree-resident artifact and no firstmate-owned sidecar at all, and has no readiness receipt or session-id to record.
 
 ## Composer ghost text: a known, unfixed gap

bin/fm-spawn.sh

@@ -81,10 +81,17 @@
 #   Before typing anything, a relaunch clears whatever the adopted shell still
 #   holds with Ctrl+C, so a previous launch that never landed - a partial line
 #   or an open quote - cannot swallow the replacement's.
+# Launch file: the launch command is never typed whole. Typed text that
+#   arrives while the pane shell is not in its line editor - still starting, or
+#   running a pre-prompt hook such as mise's after each typed export - is held
+#   by the terminal's canonical input, which keeps only its first 1024 bytes on
+#   macOS and drops the rest along with the Enter behind it. So the command is
+#   written to a private launch.sh in the per-task temp root and the pane is
+#   typed only a short line sourcing it, which evaluates the command in the pane
+#   shell exactly as typing it would.
 # Launch confirmation: a typed launch is never its own proof that an agent
-#   started. A line longer than the terminal's canonical-input limit (1024
-#   bytes on macOS) that arrives before the pane shell's line editor is running
-#   is cut short, which can leave the shell at a continuation prompt with no
+#   started. A launch line can still be cut short or garbled on its way into
+#   the pane, which can leave the shell at a continuation prompt with no
 #   agent at all. On a backend whose agent-state classifier proves a launched
 #   agent from the pane's own processes (bin/fm-backend.sh's
 #   fm_backend_launch_confirmable: tmux) every launch, fresh or relaunch,
@@ -3765,7 +3772,7 @@ spawn_clear_shell_input() {
 }
 
 spawn_type_launch() {
-  spawn_send_literal "$T" "$LAUNCH"
+  spawn_send_literal "$T" "$LAUNCH_LINE"
   sleep 0.3
   spawn_send_key "$T" Enter
 }
@@ -4809,8 +4816,24 @@ if [ "$LAUNCH_ENV_ENABLED" = 1 ]; then
   fi
   LAUNCH="$LAUNCH_ENV_PREFIX /bin/sh -c $(shell_quote "$LAUNCH")"
 fi
+# The header's "Launch file" paragraph owns why only this short line is typed.
+# The temp root must be this user's own real directory, since the pane shell
+# runs whatever launch.sh holds.
+if [ -L "$TASK_TMP" ] || [ ! -d "$TASK_TMP" ] || [ ! -O "$TASK_TMP" ]; then
+  echo "error: per-task temp root $TASK_TMP is not a directory owned by this user; refusing to write a launch file there; inspect window $T" >&2
+  exit 1
+fi
+LAUNCH_FILE="$TASK_TMP/launch.sh"
+if ! LAUNCH_FILE_TMP=$(mktemp "$TASK_TMP/.launch.sh.XXXXXX") ||
+  ! printf '%s\n' "$LAUNCH" >"$LAUNCH_FILE_TMP" ||
+  ! mv -f "$LAUNCH_FILE_TMP" "$LAUNCH_FILE"; then
+  rm -f "${LAUNCH_FILE_TMP:-}" 2>/dev/null || true
+  echo "error: could not write the launch file $LAUNCH_FILE; inspect window $T" >&2
+  exit 1
+fi
+LAUNCH_LINE=". $(shell_quote "$LAUNCH_FILE")"
 sleep 0.3
-spawn_send_literal "$T" "$LAUNCH"
+spawn_send_literal "$T" "$LAUNCH_LINE"
 sleep 0.3
 if [ "${HERDR_PROJECTED:-0}" -eq 1 ]; then
   HERDR_PROJECTION_ABORT_CLEANUP=0

docs/verification/rovo.md

@@ -38,7 +38,7 @@ Each gate fails the spawn loudly (a `failed:` line in the task status file) if i
 
 ### Why not a positional brief
 
-A positional brief is dead-on-arrival. `rovo run --yolo "<brief>"` loads a spinner, never enters a working state, prints no reply, and drops back to a bare idle shell prompt within about 10-15 seconds. This was reproduced independently four times over a raw PTY (varying `TERM`, window size, workspace, and 60-150s windows) and once more under real tmux 3.6a driven with the exact `fm-spawn.sh` send-keys shape (new window, `send-keys -l` the full launch line, then `Enter`). `--startup-receipt` cannot rescue that shape either - it is rejected before start alongside any message:
+A positional brief is dead-on-arrival. `rovo run --yolo "<brief>"` loads a spinner, never enters a working state, prints no reply, and drops back to a bare idle shell prompt within about 10-15 seconds. This was reproduced independently four times over a raw PTY (varying `TERM`, window size, workspace, and 60-150s windows) and once more under real tmux 3.6a driven with `fm-spawn.sh`'s then-current send-keys shape (new window, `send-keys -l` the full launch line, then `Enter`; `fm-spawn.sh` now types a short line sourcing a launch file instead, see its header). `--startup-receipt` cannot rescue that shape either - it is rejected before start alongside any message:
 
 ```
 $ rovo run --startup-receipt receipt.json --yolo "Reply with PONG"

tests/fixtures.sh

@@ -134,7 +134,7 @@ case "${1:-}" in
       prev=
       for a in "$@"; do
         if [ "$prev" = "-l" ]; then
-          printf '%s\n' "$a" >> "$FM_FAKE_LAUNCH_LOG"
+          printf '%s\n' "$(fm_fake_sourced_launch "$a")" >> "$FM_FAKE_LAUNCH_LOG"
         fi
         prev=$a
       done
@@ -144,9 +144,38 @@ case "${1:-}" in
 esac
 exit 0
 SH
+  fm_test_fake_sourced_launch_fn "$fakebin/tmux"
   chmod +x "$fakebin/tmux"
 }
 
+# fm_test_fake_sourced_launch_fn <fake-script>
+# fm-spawn.sh types a launch as a short line sourcing a launch file, which the
+# pane shell evaluates as the command the file holds. Inserts, right after the
+# fake's shebang line, fm_fake_sourced_launch <literal>: it prints the file's
+# command for such a line and the literal unchanged otherwise, so a fake's
+# launch log keeps reading as the command the agent was started with.
+fm_test_fake_sourced_launch_fn() {
+  local script=$1 tmp
+  tmp=$(mktemp "$script.XXXXXX") || return 1
+  {
+    head -n 1 "$script"
+    cat <<'SH'
+fm_fake_sourced_launch() {
+  local f
+  case "$1" in
+    ". '"*"'")
+      f=${1#". '"}
+      f=${f%"'"}
+      if [ -f "$f" ]; then cat "$f"; return 0; fi
+      ;;
+  esac
+  printf '%s' "$1"
+}
+SH
+    tail -n +2 "$script"
+  } > "$tmp" && mv "$tmp" "$script"
+}
+
 # fm_test_fake_tmux_send <fakebin>
 # Send-world tmux: logs send-keys -l payloads to FM_SEND_LOG, reports a numeric
 # cursor_y, and renders an empty bordered composer so the submit path reads

tests/fm-agy-harness.test.sh

@@ -34,8 +34,8 @@
 #      and nothing short of that shared proof flips an agy pane to agent-free.
 set -u
 
-# shellcheck source=tests/lib.sh
-. "$(dirname "${BASH_SOURCE[0]}")/lib.sh"
+# shellcheck source=tests/fixtures.sh
+. "$(dirname "${BASH_SOURCE[0]}")/fixtures.sh"
 
 # bin/fm-harness.sh checks verified ENV markers before ancestry. A suite run
 # from inside another harness inherits those markers, which outrank the fake
@@ -498,7 +498,7 @@ case "${1:-}" in
     literal=
     prev=
     for arg in "$@"; do
-      if [ "$prev" = -l ]; then literal=$arg; break; fi
+      if [ "$prev" = -l ]; then literal=$(fm_fake_sourced_launch "$arg"); break; fi
       prev=$arg
     done
     if [ -n "$literal" ]; then
@@ -536,6 +536,7 @@ case "${1:-}" in
 esac
 exit 0
 SH
+  fm_test_fake_sourced_launch_fn "$fakebin/tmux"
   chmod +x "$fakebin/tmux"
   cat > "$fakebin/agy" <<'SH'
 #!/usr/bin/env bash

tests/fm-backend-orca.test.sh

@@ -3,8 +3,8 @@
 # terminal adapter primitives in bin/backends/orca.sh.
 set -u
 
-# shellcheck source=tests/lib.sh
-. "$(dirname "${BASH_SOURCE[0]}")/lib.sh"
+# shellcheck source=tests/fixtures.sh
+. "$(dirname "${BASH_SOURCE[0]}")/fixtures.sh"
 
 TMP_ROOT=$(fm_test_tmproot fm-backend-orca-tests)
 # A claude spawn writes workspace trust into the launching user's own store,
@@ -39,7 +39,12 @@ COUNT_FILE="$RESP/.count"
 next=$(( $(cat "$COUNT_FILE" 2>/dev/null || echo 0) + 1 ))
 {
   printf 'orca'
-  for a in "$@"; do printf '\x1f%s' "$a"; done
+  prev=
+  for a in "$@"; do
+    if [ "$prev" = --text ]; then a=$(fm_fake_sourced_launch "$a"); fi
+    printf '\x1f%s' "$a"
+    prev=$a
+  done
   printf '\n'
 } >> "$LOG"
 if [ "${1:-}" = status ] && [ "${FM_ORCA_STATUS_RESPONSE:-ready}" != sequence ]; then
@@ -54,6 +59,7 @@ fi
 [ -f "$RESP/$n.out" ] && cat "$RESP/$n.out"
 exit 0
 SH
+  fm_test_fake_sourced_launch_fn "$fb/orca"
   chmod +x "$fb/orca"
   fm_fake_harness_clis "$fb"
   printf '%s\n' "$fb"

tests/fm-control-relaunch.test.sh

@@ -19,8 +19,8 @@
 #      agent exited.
 set -u
 
-# shellcheck source=tests/lib.sh
-. "$(dirname "${BASH_SOURCE[0]}")/lib.sh"
+# shellcheck source=tests/fixtures.sh
+. "$(dirname "${BASH_SOURCE[0]}")/fixtures.sh"
 # shellcheck source=/dev/null
 . "$ROOT/bin/fm-control-lib.sh"
 # shellcheck source=/dev/null
@@ -72,6 +72,7 @@ case "${1:-}" in
     done
     payload=${1:-}
     if [ "$literal" = 1 ]; then
+      payload=$(fm_fake_sourced_launch "$payload")
       printf '%s\n' "$payload" >> "$D/literal"
       case "$payload" in
         /exit|/quit)
@@ -124,6 +125,7 @@ case "${1:-}" in
 esac
 exit 0
 SH
+  fm_test_fake_sourced_launch_fn "$fb/tmux"
   chmod +x "$fb/tmux"
   cat > "$fb/sleep" <<'SH'
 #!/usr/bin/env bash

tests/fm-control.test.sh

@@ -19,8 +19,8 @@
 #      while fm-send's marking of the same task is untouched.
 set -u
 
-# shellcheck source=tests/lib.sh
-. "$(dirname "${BASH_SOURCE[0]}")/lib.sh"
+# shellcheck source=tests/fixtures.sh
+. "$(dirname "${BASH_SOURCE[0]}")/fixtures.sh"
 # shellcheck source=/dev/null
 . "$ROOT/bin/fm-control-lib.sh"
 # shellcheck source=/dev/null
@@ -93,6 +93,7 @@ case "${1:-}" in
     done
     payload=${1:-}
     if [ "$literal" = 1 ]; then
+      payload=$(fm_fake_sourced_launch "$payload")
       printf '%s\n' "$payload" >> "$D/literal"
       if [ -z "${FM_FAKE_NEVER_DIES:-}" ] \
          && { [ "$payload" = /exit ] || [ "$payload" = /quit ]; }; then
@@ -134,6 +135,7 @@ case "${1:-}" in
 esac
 exit 0
 SH
+  fm_test_fake_sourced_launch_fn "$fb/tmux"
   chmod +x "$fb/tmux"
   cat > "$fb/sleep" <<'SH'
 #!/usr/bin/env bash

tests/fm-kimi-harness.test.sh

@@ -2,8 +2,8 @@
 # Behavior tests for the verified Kimi Code CLI crewmate adapter.
 set -u
 
-# shellcheck source=tests/lib.sh
-. "$(dirname "${BASH_SOURCE[0]}")/lib.sh"
+# shellcheck source=tests/fixtures.sh
+. "$(dirname "${BASH_SOURCE[0]}")/fixtures.sh"
 
 # bin/fm-harness.sh answers from environment markers and process ancestry. A
 # suite run from inside Cursor, Claude, Pi, or Grok inherits those markers and
@@ -79,7 +79,7 @@ case "${1:-}" in
     prev=
     literal=
     for arg in "$@"; do
-      if [ "$prev" = -l ]; then literal=$arg; break; fi
+      if [ "$prev" = -l ]; then literal=$(fm_fake_sourced_launch "$arg"); break; fi
       prev=$arg
     done
     if [ -n "$literal" ]; then
@@ -139,6 +139,7 @@ case "${1:-}" in
 esac
 exit 0
 SH
+  fm_test_fake_sourced_launch_fn "$fakebin/tmux"
   chmod +x "$fakebin/tmux"
   fm_fake_exit0 "$fakebin" treehouse gh-axi gh
   fm_fake_exit0 "$fakebin" kimi

tests/fm-muse-harness.test.sh

@@ -10,8 +10,8 @@
 # fixture without it would let a naive implementation pass.
 set -u
 
-# shellcheck source=tests/lib.sh
-. "$(dirname "${BASH_SOURCE[0]}")/lib.sh"
+# shellcheck source=tests/fixtures.sh
+. "$(dirname "${BASH_SOURCE[0]}")/fixtures.sh"
 
 # bin/fm-harness.sh checks verified ENV markers before ancestry. Muse is
 # markerless, so an inherited Cursor/Claude/Pi/Grok marker would outrank the
@@ -96,6 +96,7 @@ case "${1:-}" in
     prev=
     for arg in "$@"; do
       if [ "$prev" = -l ]; then
+        arg=$(fm_fake_sourced_launch "$arg")
         printf '%s\n' "$arg" >> "$FM_FAKE_LAUNCH_LOG"
         if [ "${FM_FAKE_EXECUTE_MUSE_LAUNCH:-}" = 1 ]; then
           case "$arg" in
@@ -111,6 +112,7 @@ case "${1:-}" in
 esac
 exit 0
 SH
+  fm_test_fake_sourced_launch_fn "$fakebin/tmux"
   chmod +x "$fakebin/tmux"
   cp "$(command -v bash)" "$fakebin/muse-bin-test-version"
   cat > "$fakebin/muse" <<'SH'

tests/fm-rovo-harness.test.sh

@@ -2,8 +2,8 @@
 # Behavior tests for the verified Rovo CLI crewmate/scout adapter.
 set -u
 
-# shellcheck source=tests/lib.sh
-. "$(dirname "${BASH_SOURCE[0]}")/lib.sh"
+# shellcheck source=tests/fixtures.sh
+. "$(dirname "${BASH_SOURCE[0]}")/fixtures.sh"
 
 # bin/fm-harness.sh checks verified ENV markers before ancestry, but that
 # ordering settles the marker layer only: a structural (comm-strength)
@@ -75,7 +75,7 @@ case "${1:-}" in
     prev=
     literal=
     for arg in "$@"; do
-      if [ "$prev" = -l ]; then literal=$arg; break; fi
+      if [ "$prev" = -l ]; then literal=$(fm_fake_sourced_launch "$arg"); break; fi
       prev=$arg
     done
     if [ -n "$literal" ]; then
@@ -130,6 +130,7 @@ case "${1:-}" in
 esac
 exit 0
 SH
+  fm_test_fake_sourced_launch_fn "$fakebin/tmux"
   chmod +x "$fakebin/tmux"
   fm_fake_exit0 "$fakebin" treehouse gh-axi gh
   fm_fake_exit0 "$fakebin" rovo

tests/fm-secondmate-harness.test.sh

@@ -43,8 +43,8 @@
 #      flags still win.
 set -u
 
-# shellcheck source=tests/lib.sh
-. "$(dirname "${BASH_SOURCE[0]}")/lib.sh"
+# shellcheck source=tests/fixtures.sh
+. "$(dirname "${BASH_SOURCE[0]}")/fixtures.sh"
 # shellcheck source=/dev/null
 . "$ROOT/bin/fm-ff-lib.sh"
 # shellcheck source=/dev/null
@@ -692,7 +692,7 @@ case "${1:-}" in
       prev=
       for a in "$@"; do
         if [ "$prev" = "-l" ]; then
-          printf '%s\n' "$a" >> "$FM_FAKE_LAUNCH_LOG"
+          printf '%s\n' "$(fm_fake_sourced_launch "$a")" >> "$FM_FAKE_LAUNCH_LOG"
         fi
         prev=$a
       done
@@ -702,6 +702,7 @@ case "${1:-}" in
 esac
 exit 0
 SH
+  fm_test_fake_sourced_launch_fn "$fakebin/tmux"
   chmod +x "$fakebin/tmux"
   fm_fake_exit0 "$fakebin" pi
   # BASE_PATH deliberately omits the developer's node, which the trust

tests/fm-secondmate-restart.test.sh

@@ -24,8 +24,8 @@
 #      its agent left running.
 set -u
 
-# shellcheck source=tests/lib.sh
-. "$(dirname "${BASH_SOURCE[0]}")/lib.sh"
+# shellcheck source=tests/fixtures.sh
+. "$(dirname "${BASH_SOURCE[0]}")/fixtures.sh"
 
 RESTART="$ROOT/bin/fm-secondmate-restart.sh"
 
@@ -63,6 +63,7 @@ case "${1:-}" in
     done
     payload=${1:-}
     if [ "$literal" = 1 ]; then
+      payload=$(fm_fake_sourced_launch "$payload")
       printf '%s\n' "$payload" >> "$D/literal"
       case "$payload" in
         /exit|/quit)
@@ -115,6 +116,7 @@ case "${1:-}" in
 esac
 exit 0
 SH
+  fm_test_fake_sourced_launch_fn "$fb/tmux"
   chmod +x "$fb/tmux"
   cat > "$fb/sleep" <<'SH'
 #!/usr/bin/env bash