drive-opencode

name	drive-opencode
description	Use when delegating a coding task to the OpenCode CLI (`opencode`) from Claude Code. Covers one-shot `opencode run`, interactive TUI via tmux, sessions, PR review, and parallel work. NOTE opencode is not installed by default on this machine. Load delegate-to-agents first.

Drive OpenCode CLI

Delegate to OpenCode, a provider-agnostic open-source coding agent with a TUI and CLI. Read delegate-to-agents first.

Not installed by default here. Preflight will tell you. Install with npm i -g opencode-ai@latest or brew install anomalyco/tap/opencode.

Preflight

command -v opencode || echo "opencode NOT installed — npm i -g opencode-ai@latest"
opencode --version
which -a opencode          # PATH can resolve the wrong binary; pin a path if so
opencode auth list         # should show ≥1 provider

If not installed/authed, install it or fall back to a different agent — don't treat this as a hard failure.

One-shot (PREFERRED): opencode run

opencode run 'prompt' is bounded and non-interactive — no PTY needed, so a plain Bash call works.

# bounded task
Bash(command="cd /repo && opencode run 'Add retry logic to API calls and update tests'", timeout=600000)

# attach context files
Bash(command="cd /repo && opencode run 'Review this config for security issues' -f config.yaml -f .env.example", timeout=300000)

# force a model
Bash(command="cd /repo && opencode run 'Refactor auth module' --model openrouter/anthropic/claude-sonnet-4", timeout=600000)

# machine-readable output
Bash(command="cd /repo && opencode run 'List all API routes' --format json", timeout=300000)

# long task: detach + log
Bash(command="cd /repo && opencode run 'Migrate to the new ORM' > /tmp/agent-opencode.log 2>&1", run_in_background=true)

Common flags

Flag	Use
run '<prompt>'	One-shot execution and exit (no PTY)
-c, --continue	Continue the last session
-s, --session <id>	Continue a specific session
--agent <name>	Choose agent (build or plan)
--model provider/model	Force a model
--format json	Machine-readable output/events
-f, --file <path>	Attach file(s)
--thinking	Show model thinking
--variant <level>	Reasoning effort (high, max, minimal)
--title <name>	Name the session

Interactive (tmux)

The bare opencode TUI needs a real PTY — drive via tmux.

Bash(command="tmux new-session -d -s oco -x 200 -y 50")
Bash(command="tmux send-keys -t oco 'cd /repo && opencode' Enter")
Bash(command="sleep 6 && tmux capture-pane -t oco -p -S -40")
Bash(command="tmux send-keys -t oco 'Implement OAuth refresh flow and add tests' Enter")
Bash(command="sleep 2 && tmux send-keys -t oco Enter")   # Enter may need pressing twice to submit
Bash(command="sleep 25 && tmux capture-pane -t oco -p -S -80")
Bash(command="tmux send-keys -t oco C-c")                # EXIT with Ctrl+C — never /exit
Bash(command="tmux kill-session -t oco")

Gotchas:

• /exit is NOT valid — it opens an agent selector. Exit with Ctrl+C (C-c).
• Enter may need pressing twice in the TUI (finalize text, then send).
• PATH mismatch can select the wrong binary/model config — which -a opencode.

Sessions / cost

opencode session list
opencode stats
opencode stats --days 7

PR review

# built-in PR command
Bash(command="cd /repo && opencode run --agent plan 'Review PR #42 vs main: bugs, security, test gaps'", timeout=300000)

# or in an isolated clone
Bash(command="REVIEW=$(mktemp -d) && git clone https://github.com/USER/REPO.git $REVIEW && cd $REVIEW && gh pr checkout 42 && opencode run 'Review this PR vs main. Report bugs, security risks, test gaps, style.'", timeout=600000)

Parallel work

Bash(command="git -C /repo worktree add -b fix/101 /tmp/iss-101 main")
Bash(command="cd /tmp/iss-101 && opencode run 'Fix issue #101 and commit' > /tmp/oco-101.log 2>&1", run_in_background=true)

Rules

1. Prefer opencode run for automation — simpler, no PTY.
2. Interactive only when iterating — and exit with Ctrl+C, never /exit.
3. Scope each session to one repo/worktree; never share a workdir.
4. which -a opencode if behavior looks wrong (binary/PATH mismatch).
5. Clean up tmux sessions, worktrees, and logs.
6. Review the diff and re-run tests yourself.