What problem does it solve? Coordinating multiple AI agents (DEV, QC, orchestrator) on GitHub issues without a message bus leads to lost state, clobbered writes, and missing audit trails. This Skill establishes a single authoritative protocol so every agent reads and writes workflow state through the same three artifacts: the Projects v2 Status field, prefixed issue comments, and an AGENTFLOW-STATE section in the issue body. ## Core Features & Use Cases - Six-column state machine: Defines the canonical board columns (Inbox, Ready for Dev, In Progress, In QC, Ready for Review, Done) with ownership rules, a rework loop, and escalation after more than two consecutive QC failures. - Strict read/write ordering: Enforces body-first, comment-second, aux-label-third, Status-write-last ordering with compare-then-write protection against human mid-run interference. - Config and secret discipline: Specifies agentflow.yaml schema gating, board URL parsing, forbidden paths, QC tiers, and single-destination secret storage in .claude/settings.local.json. - Use Case: A team runs autonomous DEV and QC agents against a GitHub Projects board; this protocol ensures a QC rejection always lands with a rework label before the Status moves back to Ready for Dev, so DEV reads the rejection before touching code. ## Quick Start Read this protocol before touching any board artifact, issue, config, or external service in an AgentFlow-managed repository.
