ReqPlan Orchestrator

Problem

ReqPlan Orchestrator addresses the practical failure modes of agent-driven engineering tasks: skipped analysis, fake completion, stage jumps, lost context across sessions, and unverified outputs being treated as deliverables. It turns development, debugging, and analysis requests into a checked execution path.

How It Works

• Activation triggers execution: when development, repair, or review intent is detected, it enters a fixed state machine instead of waiting for extra commands.
• Hard checkpoints: START, ANALYZE, CONFIRM, DESIGN, IMPLEMENT, VERIFY, and JUDGE proceed in order; missing prerequisite artifacts block the next stage.
• Persistent baton: state and progress are written to .agent/harness/_baton.md, supporting recovery after interruption and continuation across sessions without relying on chat memory.
• Artifacts and blind audit: it produces _analysis.md, _design.md, _implementation.md, and _verification.md, while an independent quality auditor checks analysis, design, implementation, and verification quality.
• Safety boundaries: it works only with project code and user-provided text, stores outputs under local .agent/harness/, and restricts sensitive credentials or private data from artifacts.

Scope

It fits software tasks with a clear project path, such as new features, bug fixes, test optimization, documentation, and architecture refactoring. Pure non-software tasks can reuse only parts of the analysis and design stages. Note that CONFIRM requires human approval, failed audits cannot be bypassed, and it does not replace real tests, CI, or security audits.