coding-agent

• Building or creating new features and apps — let a coding agent explore the codebase and implement end-to-end, rather than patching inline.
• Reviewing pull requests — spawn an agent in a temporary clone or git worktree so it can diff, analyze, and summarize changes safely.
• Refactoring large codebases — multi-file changes that require context across many files benefit from an agent that can traverse the project tree.
• Parallel issue fixing — use git worktrees to run multiple agents simultaneously, one per issue, then create PRs from each branch.

• Simple one-liner fixes — if the change is obvious and localized, just edit the file directly; spawning an agent adds unnecessary overhead.
• Read-only code exploration — use a read/grep tool to inspect code rather than spinning up a full coding agent.
• Work inside ~/clawd or the live OpenClaw workspace — agents must never be spawned in the orchestrator's own working directory.
• Thread-bound ACP harness requests — if you need to spawn or run Codex/Claude Code inside a Discord thread or similar ACP context, use sessions_spawn with runtime:"acp" instead.

Supports four distinct coding agents with correct invocation patterns for each. Codex and Pi require pty:true for interactive terminal behavior; Claude Code uses --print --permission-mode bypassPermissions for clean non-PTY execution.

Long-running tasks can be launched with background:true, which returns a sessionId. You can then poll status, stream logs, send stdin input, or kill the session using process tool actions — keeping the user informed without blocking.

Setting workdir ensures the agent wakes up in the correct project folder and doesn't wander into unrelated files. This is critical for PR reviews and parallel issue fixing.

Create isolated git worktrees per issue, launch a dedicated agent in each, monitor them concurrently, and push separate branches for PR creation — all without touching the main branch.

Append an openclaw system event command to the agent's prompt so the orchestrator receives an immediate wake event when the task finishes, instead of waiting for the next heartbeat poll.

Reviews are always performed in a temporary clone or git worktree — never in the live project folder — to prevent accidental branch checkouts or mutations to the active codebase.

Point a Codex or Claude Code agent at your project directory with workdir and a clear feature prompt. The agent explores the codebase, implements the feature across multiple files, and commits — while you monitor via process action:log.

Fetch all open PR refs, spin up one Codex process per PR in background mode, collect their reviews, and post results as GitHub comments via the gh CLI — all running in parallel.

Create a git worktree for each open issue, launch an agent in each worktree, monitor progress, then push branches and open PRs for human review — cutting fix time dramatically.

For quick one-shot prompts with no existing project, create a temp directory, run git init, and let Codex execute a self-contained task — then discard the directory when done.