Parallel Task Executor (tmux Live Workers)

Run dependency-ordered plan execution like parallel-task, but route each worker through codex exec inside tmux panes so progress is visible in real time.

Use this skill when you want:

• Live visibility into each running worker
• A tiled worker view that expands as tasks launch
• Automatic pane cleanup that shrinks the layout as tasks finish

Execution Model

1. Parse the plan and compute unblocked tasks by depends_on.
2. Build one worker prompt file per task.
3. Launch one tmux pane per task via scripts/tmux_spawn_worker.sh.
4. Monitor status/logs and reap completed panes via scripts/tmux_reap_completed.sh.
5. Validate completed tasks before advancing to the next dependency wave.
6. Repeat until all requested tasks are complete.

Preflight

Run these checks before launching workers:

command -v tmux >/dev/null
command -v codex >/dev/null

Resolve the skill directory once so helper script paths are stable:

SKILL_ROOT="$(fd -td -p 'parallel-task-tmux' "$PWD" "$HOME/.codex/skills" "$HOME/.agents/skills" 2>/dev/null | head -n 1)"

Fail fast if SKILL_ROOT is empty.

Initialize one run directory per execution:

PLAN_FILE="./plan.md"
PLAN_BASE="$(basename "$PLAN_FILE" .md)"
RUN_ID="$(date +%Y%m%d-%H%M%S)"
SESSION_NAME="swarms-${PLAN_BASE}-${RUN_ID}"
RUN_ROOT=".swarms/tmux/${SESSION_NAME}"
mkdir -p "$RUN_ROOT/prompts" "$RUN_ROOT/logs"
: > "$RUN_ROOT/status.tsv"
: > "$RUN_ROOT/task_map.tsv"
: > "$RUN_ROOT/reaped.tsv"

Plan Parsing and Scheduling

Mirror parallel-task behavior:

• Parse task sections (for example ### T1:).
• Extract task id, title, depends_on, location, description, acceptance criteria, validation.
• Filter to requested subset and required dependencies if the user passes a subset.
• Launch only tasks whose dependencies are complete.

Worker Prompt Contract

Create one prompt file per task at:

• $RUN_ROOT/prompts/<TASK_ID>.md

Use the same task context and completion requirements as parallel-task:

• Read the working plan and relevant files before coding.
• Default to TDD RED phase first using a tdd_test_writer subagent, or explicitly record reason_not_testable with an alternative verification contract.
• Treat RED-phase tests (or approved non-testable verification plan) as the implementation contract.
• Implement acceptance criteria.
• Keep edits atomic.
• Run the exact new/updated test commands until GREEN, or run the documented alternative verification and capture evidence.
• Commit only task-scoped files.
• Update the plan status/log/files after the commit.
• Return modified files, criteria coverage, and RED -> GREEN or non-testable verification evidence.

Launching tmux Workers

When running inside tmux, prefer in-place splits so the user can keep chatting in the current Codex pane while workers appear beside it.

In-place mode (recommended inside tmux):

"$SKILL_ROOT/scripts/tmux_spawn_worker.sh" \
  --split-current \
  --workspace "$PWD" \
  --task-id "$TASK_ID" \
  --prompt-file "$RUN_ROOT/prompts/${TASK_ID}.md" \
  --log-file "$RUN_ROOT/logs/${TASK_ID}.log" \
  --status-file "$RUN_ROOT/status.tsv" \
  --map-file "$RUN_ROOT/task_map.tsv"

Detached session mode (for non-tmux shells):

"$SKILL_ROOT/scripts/tmux_spawn_worker.sh" \
  --session "$SESSION_NAME" \
  --workspace "$PWD" \
  --task-id "$TASK_ID" \
  --prompt-file "$RUN_ROOT/prompts/${TASK_ID}.md" \
  --log-file "$RUN_ROOT/logs/${TASK_ID}.log" \
  --status-file "$RUN_ROOT/status.tsv" \
  --map-file "$RUN_ROOT/task_map.tsv"

Behavior:

• In split-current mode, launches workers as new panes in the current tmux window.
• In detached mode, creates tmux session/window if missing.
• Re-tiles panes after each launch so the grid expands live.
• Streams codex exec output live in each pane while also writing logs to logs/<TASK_ID>.log.
• Tracks pane ids in task_map.tsv for deterministic cleanup.

In detached mode, if not already attached, print:

tmux attach -t "$SESSION_NAME"

Monitoring and Pane Reaping

During each wave, monitor active panes:

tmux list-panes -t "${SESSION_NAME}:workers" -F '#{pane_id} #{pane_current_command} #{pane_dead}'

Collect completed workers and close their panes:

"$SKILL_ROOT/scripts/tmux_reap_completed.sh" \
  --session "$SESSION_NAME" \
  --status-file "$RUN_ROOT/status.tsv" \
  --map-file "$RUN_ROOT/task_map.tsv" \
  --seen-file "$RUN_ROOT/reaped.tsv"

This is what gives the expand/shrink behavior:

• Launching tasks adds panes and expands the tile grid.
• Reaping completed tasks kills panes and shrinks the grid.

Inspect logs before accepting a task:

tail -n 80 "$RUN_ROOT/logs/${TASK_ID}.log"

Validation and Wave Control

For each completed task:

1. Verify files changed as expected.
2. Verify plan status/log/files fields were updated.
3. Verify RED -> GREEN test evidence or explicit reason_not_testable plus concrete alternative verification output.
4. Mark task complete in orchestrator state only after verification passes.
5. Retry failed tasks with corrected prompts when needed.

Launch the next wave only when all tasks in the current wave are verified complete.

Finalization

After all tasks are complete:

1. Run final validation commands from the plan.
2. Confirm no pending tasks remain.
3. Provide an execution summary with completed, failed, retried tasks, modified files, and verification evidence.

Leave the tmux session alive by default for auditability. Only kill it if the user asks:

tmux kill-session -t "$SESSION_NAME"

Error Handling

• Missing task ids in subset: list valid task ids from the parsed plan.
• tmux unavailable: stop and instruct user to install tmux.
• Worker exits non-zero: inspect log, repair prompt/context, relaunch task.
• No status updates: verify pane is still running; if dead without status line, relaunch.
• Parse failure: report parser assumptions and request corrected plan format.

Example Usage

/parallel-task-tmux plan.md
/parallel-task-tmux ./plans/auth-plan.md T1 T2 T4
/parallel-task-tmux user-profile-plan.md --tasks T3 T7