Indaba: a TypeScript DAG runner that stops parallel coding agents from colliding
yaotzin1 · reddit · 2026-11-21
After months of shipping client code with CLI coding agents, the author found the bottleneck wasn't model intelligence but orchestration: single chat loops fall apart on multi-file tasks, parallel sessions overwrite each other's edits, and self-review is confirmation bias. He extracted his manual flow into Indaba (@indaba/engine), an MIT-licensed TypeScript engine.
- Define a deterministic DAG in workflow.ai.yml instead of an autonomous loop
- Parallel nodes run in isolated git worktree directories
- Wire different runners (Claude Code/Codex CLIs, ACP agents, raw APIs) into review/consensus steps before merging, with configurable consensus rounds
- Dogfooded to build a zero-dependency React data grid (@apsw/gridwright) now running in production for a client
Early alpha; the author is soliciting architectural feedback on where strict DAGs break down in real coding workflows.
More from coding & agent
- Supericons MCP lets coding agents search and retrieve 20,000+ SVG icons semantically — modelcontextprotocol · 2026-11-21
- Dev Reflects: My Coding Agent Now Tests Against 46 Browsers by Default — annetgriffin · 2026-11-21
- quantledger-mcp: 73 deterministic GAAP/tax solvers replace LLM arithmetic — SideInitial3961 · 2026-11-21
- Antigravity's /btw now forks, explores, and promotes side threads without losing your place — rseroter · 2026-11-21
- earshot v1: let your coding agent generate sound effects from a prompt — lyraxana · 2026-11-21
- Vercel Board skill merges per-harness tool mappings so Codex and Claude Code interoperate — cramforce · 2026-11-21