project-orchestrator

Executes approved development contracts through deterministic Git worktree orchestration and provider-managed coding agents.

Updated Jun 11, 2026
One-click install
npx skills add https://github.com/coferlandia/coferlandia-skills --skill project-orchestrator-coferlandia
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: project-orchestrator
Source: https://github.com/coferlandia/coferlandia-skills/tree/main/skills/ops/project-orchestrator
Command: npx skills add https://github.com/coferlandia/coferlandia-skills --skill project-orchestrator-coferlandia

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes scripts (resource) and references (resource) components.

What problem does it solve? Coordinating multi-task software development with AI coding agents often leads to duplicated work, lost state, unreviewed merges, and untraceable changes. This Skill provides a deterministic controller that owns Git worktrees, commits, durable run state, provider retries, and GitHub traceability so approved development contracts execute safely and audibly from plan to merged PR. ## Core Features & Use Cases - Dual execution modes: Run a single detailed plan as one direct-plan unit, or execute an Analyst-produced DAG of atomic tasks serially in dependency order. - Immutable review lifecycle: Every task produces an additive candidate commit reviewed from a detached worktree at an exact SHA; fixes are additive and never amend reviewed commits. - Durable claims and recovery: Repository-wide Epic/task claims prevent duplicate execution across worktrees, survive process termination, and support resume, retry, cancel, and audited administrative release. - GitHub integration gates: Final integration requires exact-candidate CI gates, remote base validation, and an explicit integrate command with head-conditional squash merge. - Use Case: Point the controller at a GitHub Epic Issue; it materializes local contracts, claims the work, drives coding/review/fix agents through each task, opens one final PR, waits for green CI, and merges only after explicit approval. ## Quick Start Run the project-orchestrator CLI with run --spec plan.md --json to execute a detailed local plan through the full claim, coding, review, and integration lifecycle.

Frequently Asked Questions about project-orchestrator

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I execute a development plan with AI coding agents?▼

Run the CLI with run --spec plan.md to execute a detailed plan as one direct-plan unit, or use --manifest for a task DAG. The controller creates an Epic worktree, claims the work, invokes coding and review agents, and produces additive commits with durable evidence.

How do I orchestrate a GitHub Epic Issue into implemented code?▼

Use run --epic with the Issue reference. The controller materializes the Epic and task Issues into local contracts, executes tasks in dependency order, opens one final PR, and merges only after an explicit integrate command with passing CI gates.

What providers and tools does project-orchestrator require?▼

It requires Python 3.11+, Git with worktree support, and configured Codex and/or OpenCode CLIs as execution providers. GitHub-backed Epic execution additionally requires an authenticated gh CLI; local spec and manifest runs do not.

Can two runs execute the same Epic or task concurrently?▼

No. Durable claims stored under the Git common directory atomically block duplicate Epic or task execution across all worktrees sharing that repository. Claims survive crashes and are released only after verified delivery, cancellation, or audited administrative release.

Why does integration fail even when CI was green before?▼

Gates are evaluated only for the exact current integration candidate SHA; a green older SHA never satisfies the gate. Base movement, head changes, or missing required checks fail closed and require fresh review or reconciliation before merging.

When should I not use project-orchestrator?▼

Avoid it for one-off prompt snippets or unsupervised publishing; it activates only on explicit invocation and enforces review, claims, and explicit integration. It also does not provide distributed locking across independent clones or machines.