Coding Agent

Orchestrate Codex CLI and Claude Code as background assistants via tmux.

Spawn agents liberally — if a task can be done independently by another agent, delegate it.

Session Management

Use remain-on-exit on so the pane persists after the process exits — output stays capturable.

Use descriptive session names: codex-auth-refactor, claude-fix-78. Kill sessions after capturing output to avoid sprawl.

Waiting for completion

All agents in this skill run as one-shot processes (codex exec, claude -p) that exit when done. Poll pane_dead to detect completion without wasted time:

while [ "$(tmux display-message -t NAME -p '#{pane_dead}')" != "1" ]; do sleep 1; done
# Now read: tmux capture-pane -t NAME -p -S -
# Exit code: tmux display-message -t NAME -p '#{pane_dead_status}'

For multi-turn conversations, prefer resume with an explicit session ID over keeping an interactive TUI session alive — each turn is a clean one-shot process, so the same pane_dead polling works every time. Always pass the session ID explicitly (exec resume <ID>, --resume <ID>) rather than using --last or -c, which can pick the wrong session when multiple agents run in parallel. See multi-turn examples below for how to capture and reuse the session ID.

Codex CLI

Default model is configured in ~/.codex/config.toml. Requires a git repository — use --skip-git-repo-check to override, or DIR=$(mktemp -d) && git -C "$DIR" init for scratch work.

Progress streams to stderr, final result to stdout — enables piping: codex exec "..." | tee result.md. Use - to pipe prompt from stdin: cat spec.md | codex exec -.

Flags

Top-level subcommands (not exec flags):

--full-auto vs --yolo: --full-auto runs in a workspace-write sandbox with model-driven approval — safe for most building tasks. --yolo removes all guardrails (no sandbox, no approvals). Prefer --full-auto; only use --yolo when the agent needs unrestricted system access (e.g., installing packages, modifying system files).

# One-shot
tmux new-session -d -s codex-task -c ~/project \
  "codex exec --full-auto 'Add error handling to API calls'" \; \
  set remain-on-exit on

# Read-only research (safe, no writes)
tmux new-session -d -s codex-research -c ~/project \
  "codex exec --sandbox read-only -o /tmp/codex-research.txt 'Analyze how auth tokens flow through this codebase'" \; \
  set remain-on-exit on

# Multi-directory access
tmux new-session -d -s codex-task -c ~/project \
  "codex exec --full-auto --add-dir ../shared-lib 'Update the API client to use the new shared auth module'" \; \
  set remain-on-exit on

# Multi-turn: extract session ID from pane, then resume with follow-up
tmux new-session -d -s codex-review -c ~/project \
  "codex exec --full-auto 'Review the auth module'" \; \
  set remain-on-exit on
# Wait, extract session ID from the pane header, then resume:
while [ "$(tmux display-message -t codex-review -p '#{pane_dead}')" != "1" ]; do sleep 1; done
SESSION_ID=$(tmux capture-pane -t codex-review -p -S - | grep "session id:" | awk '{print $NF}')
tmux kill-session -t codex-review
tmux new-session -d -s codex-review-2 -c ~/project \
  "codex exec resume --full-auto $SESSION_ID 'Now fix the issues you found'" \; \
  set remain-on-exit on

# Monitor
tmux capture-pane -t codex-task -p -S -

Claude Code

Flags

Variadic flag gotcha: --allowedTools and --tools are variadic — they consume all following positional arguments, including the prompt. When using them, place -- before the prompt to terminate option parsing:

# WRONG — prompt gets swallowed as a tool name:
claude -p --allowedTools Bash Read Edit "Fix the bug"

# RIGHT — use -- to separate flags from the prompt:
claude -p --allowedTools Bash Read Edit -- "Fix the bug"

# ALSO RIGHT — pipe prompt via stdin:
echo "Fix the bug" | claude -p --allowedTools Bash Read Edit

--allowedTools vs --tools: --tools controls what's available. --allowedTools controls what's auto-approved. Use both for tight scoping: --tools Bash Read --allowedTools Read makes Bash available but still prompts, while Read is auto-approved.

Stdin is piped as context: gh pr diff 130 | claude -p "Review this diff".

# One-shot (print mode)
tmux new-session -d -s claude-task -c ~/project \
  "claude -p 'Refactor the auth module'" \; \
  set remain-on-exit on

# Research with haiku
tmux new-session -d -s claude-research -c ~/project \
  "claude -p --model haiku 'Summarize how auth works in this codebase'" \; \
  set remain-on-exit on

# With allowedTools — note the -- before the prompt
tmux new-session -d -s claude-task -c ~/project \
  "claude -p --allowedTools Bash Read Edit -- 'Process src/auth/'" \; \
  set remain-on-exit on

# Batch pattern: instructions from file
tmux new-session -d -s claude-task -c ~/project \
  "claude -p --append-system-prompt-file /tmp/instructions.md 'Process src/auth/'" \; \
  set remain-on-exit on

# Multi-turn: capture session ID with --output-format json, then resume
# (Claude doesn't print session ID in its pane output, so JSON redirect is needed)
tmux new-session -d -s claude-review -c ~/project \
  "claude -p --output-format json 'Review the auth module' > /tmp/claude-review.json" \; \
  set remain-on-exit on
# Wait, extract session ID and readable result, then resume:
while [ "$(tmux display-message -t claude-review -p '#{pane_dead}')" != "1" ]; do sleep 1; done
SESSION_ID=$(jq -r '.session_id' /tmp/claude-review.json)  # requires jq
# Read the result: jq -r '.result' /tmp/claude-review.json
tmux kill-session -t claude-review
tmux new-session -d -s claude-review-2 -c ~/project \
  "claude -p --resume $SESSION_ID 'Now fix the issues you found'" \; \
  set remain-on-exit on

# Monitor
tmux capture-pane -t claude-task -p -S -

Building with Coding Agents

For substantial features, tools, or scripts, use a phased pipeline. Each step is a separate agent session — you orchestrate the handoffs.

1. Research

Spawn an agent to investigate the problem space, explore relevant code, and gather context. Write findings to a file.

tmux new-session -d -s research -c ~/project \
  "codex exec --full-auto 'Research how auth tokens are handled in this codebase. Write findings to /tmp/research.md'" \; \
  set remain-on-exit on

2. Plan & spec

Feed the research to an agent tasked with producing a concrete plan and spec document.

tmux new-session -d -s plan -c ~/project \
  "claude -p --allowedTools Read Write -- 'Read /tmp/research.md. Create a detailed implementation plan and spec at /tmp/plan.md. Include phases, file changes per phase, and acceptance criteria.'" \; \
  set remain-on-exit on

3. Review the plan

Spawn a review agent (or use resume for a conversation) to challenge the plan, find gaps, and refine it.

tmux new-session -d -s plan-review -c ~/project \
  "claude -p --allowedTools Read Write -- 'Read /tmp/plan.md. Review it critically: are there gaps, risks, or missing edge cases? Update the plan with your suggestions.'" \; \
  set remain-on-exit on

4. Implement phase by phase

Split the plan into phases, then loop: implement → review → fix → next phase.

# Phase 1: implement
tmux new-session -d -s phase-1-impl -c ~/project \
  "codex exec --full-auto 'Read /tmp/plan.md. Implement Phase 1 only. Commit when done.'" \; \
  set remain-on-exit on

# Phase 1: review & fix
tmux new-session -d -s phase-1-review -c ~/project \
  "claude -p --allowedTools Bash Read Edit -- 'Read /tmp/plan.md. Review the Phase 1 implementation against the spec. Fix any issues. Commit when done.'" \; \
  set remain-on-exit on

# Repeat for Phase 2, 3, ...

5. Final review

Once all phases are complete, spawn a final agent to review the full implementation against the original spec.

tmux new-session -d -s final-review -c ~/project \
  "claude -p --allowedTools Bash Read -- 'Read /tmp/plan.md. Review the full implementation against this spec. Report what was completed, what diverged, and any remaining issues.'" \; \
  set remain-on-exit on

Wait for each step to complete before starting the next (see "Waiting for completion" above). Read results with tmux capture-pane -t NAME -p -S -.

Parallel Issue Fixing

Use git worktrees to fix multiple issues concurrently:

git worktree add -b fix/issue-78 /tmp/issue-78 main
git worktree add -b fix/issue-99 /tmp/issue-99 main

# --yolo here because pnpm install needs system writes outside the workspace sandbox.
# Use --full-auto when the task only needs workspace writes.
tmux new-session -d -s fix-78 -c /tmp/issue-78 \
  "pnpm install && codex exec --yolo 'Fix issue #78: <desc>. Commit and push.'" \; \
  set remain-on-exit on
tmux new-session -d -s fix-99 -c /tmp/issue-99 \
  "pnpm install && codex exec --yolo 'Fix issue #99: <desc>. Commit and push.'" \; \
  set remain-on-exit on

# After completion: push branches, create PRs, remove worktrees

Rules

1. Respect tool choice. Use the agent the user asks for. Do not silently substitute your own edits when an agent fails — respawn or ask the user.
2. Use --full-auto for building, vanilla for reviewing. No approval flags needed for code review.

Progress Updates

When spawning background agents, keep the user informed:

• 1 short message at launch (what's running, where).
• Update only on state changes: milestone reached, agent needs input, error hit, or agent finished.
• If you kill a session, say so immediately with the reason.