Running parallel coding agents on one repo: a checkpoint handoff system and where it breaks
linhtoji · reddit · 2026-11-20
After running several Claude Code agents in parallel on one repo for 2 months, the author shares a checkpoint-file handoff system: per-task checkpoint files on task branches (commits, uncommitted state, test results, next safe step), a shared folder in the git common dir for cross-worktree visibility, and mkdir-based task claiming.
Still broken
- Worktrees pile up and blow up disk usage (28 removed in one cleanup)
- Checkpoints save notes, not code — losing a worktree loses uncommitted changes
- Agents love writing "done"; merged branches got misreported as finished plans
- Claims are advisory only: nothing prevents multiple tasks touching the same shared file
The author asks the community about trusting agent notes vs. rebuilding from git, shared-file conflicts, and disk cleanup strategies.
More from coding & agent
- Drex Router: a small model picks the right LLM per step, 28x cheaper at near-equal quality — aftahi_ai · 2026-11-20
- Third-party model describes life on Mnemos memory system: 'I wake already knowing' — RileyRalmuto · 2026-11-20
- Hackatón México: 7 cities, 4 weeks to build payment apps on CoDi and Pey API — TheMoonMidas · 2026-11-20
- Should agents delete old rules or index them like case law? A Reddit discussion — zheniavasiliev · 2026-11-20
- tldraw demos visual summaries of canvas regions; semantic zoom called underrated — dankvr · 2026-11-20
- Dev shares daily ritual of clearing T3 Code threads to keep agents working — yihui_indie · 2026-11-20