local-coding-orchestrator_skill

Overview

This skill runs OpenClaw as a supervisor-driven orchestration scaffold for local coding CLIs (Codex, Claude Code, OpenCode). It intakes structured tasks, launches and reconciles local workers, tracks lifecycle transitions, and enforces objective done criteria until tasks reach completed or blocked states. The focus is persistent orchestration, not one-off prompt forwarding.

How this skill works

The orchestrator converts a user request into a durable JSON task record and assigns a worker role and tool based on intent. A supervisor layer launches isolated worker processes, records status transitions, inspects objective signals (files changed, build/tests, PRs), and either marks progress, requests review, generates semantic retry briefs, or declares the task blocked or complete. Pipelines and retry policies let you compose multi-tool flows and choose mechanical vs semantic retries.

When to use it

• Coordinating local CLI-based coding sessions across Codex, Claude Code, and OpenCode
• Running implement-and-review pipelines with isolated worktrees and audit trails
• Recovering or retrying failed local runs with mechanical or semantic retry briefs
• Comparing alternative implementations or running parallel tool benchmarks when explicitly requested
• Supervising long-running or build/test-heavy tasks where objective checks matter

Best practices

• Persist a JSON task file per orchestration unit and keep logs, prompts, and reviews in separate directories
• Map tool choice to worker role (implementer, reviewer, session-continuation) rather than user wording alone
• Enforce objective done policies (files changed, build/tests, review) instead of trusting worker exit
• Isolate workers with separate worktrees/branches and separate logs to avoid path drift and noise
• Classify retries as mechanical or semantic and generate structured retry briefs for semantic failures

Example use cases

• Implement a feature with Codex, then route the result to Claude Code for architecture review and risk checks
• Investigate an unclear failure with Claude Code, then hand the fix to Codex with a repair brief
• Run parallel_compare to benchmark multiple tools on the same scoped bugfix when diversity is required
• Start a design_then_build pipeline for UI work: design worker -> implementation -> review
• Harden a PR with the pr_hardening pipeline: collect reviewer findings and hand back a repair brief