npx @cardor/agent-harness-kit init scaffolds a full multi-agent TypeScript harness with SQLite state, MCP tools, and four specialized roles in three prompts.
Key Takeaways
Four fixed roles (Lead, Explorer, Builder, Reviewer) have explicit permission boundaries: Explorer is read-only, Builder writes only to src/ and tests/.
SQLite is the single source of truth for agent state; a Markdown fallback exists when MCP is unavailable.
Ships with a built-in MCP server, ahk dashboard web UI, and health.sh; OpenTelemetry integration is in progress.
Provider-agnostic at launch: supports Claude Code and OpenCode, with Jira, Linear, and GitHub Issues adapters on the roadmap.
MIT licensed, 2,343 downloads/month at v0.18.0.
Hacker News Comment Review
The only substantive comment raises core workflow flexibility questions: whether the hardcoded Lead-Explorer-Builder-Reviewer pipeline can support branching logic (e.g., loop from end-to-end test failure back to Builder, or per-finding sub-processes).
Notable Comments
@philipp-gayret: asks whether harness supports dynamic process graphs beyond linear build-review chains, and how conditional loops on review findings are handled.
